THE WORLD BANK ECONOMIC REVIEW 73275 <..:orpornl .. Gov,",n:"~ tI the World Ib.nk and m., Dilemma of (;lob~ 1 WVl."rn.lInce A.hw;n Kt.f" fJ"d enc W..rk~r M"-"'SUJ"ing HOU5Coold Uugt of Financial SC'n-iccs: I'>oe~ it Matter How or Who'" You Ask? Robm Cull "nd ""'non Srotl Bankmg on Polill(s: Wj,tn Fortner I I,gh-rankmg Polj{ki~ns Become Bank l)11e<;101'$ M"tras Braun ,mil CI"ud,o Raddalt Narum] 1)'5as.tcn and l1uman up',,,! Accumulation fItS .. , Crnpo c...,'ItSHU M.,I.Ig.1lg for Rrsulu In Prlnlary IAIk;"' ..... in M3it Bourguignon, Ami ~ Jo","", F"""u Willi.m F. M:oloo.cy. World B""i Luc Chriotiacnocn, U"ikd N"timu n'M McKenzie, W",lJ B"d u,,;.,.mly. WIDER. P;"/i,,,d Jacque. Mori.oo:t, /Vorl"!lad Stijn Clocssen., ]n/ Plalrovic, WorM Bad EIim. La Fc!'f'll'>., Uni_,ilil B"",,,~ IfRly Mattin Ravcallian, WorlJ &mlt Augu.tin Kw ..i Fosu, Ut.;fet/ NaNO", Eli ••beth Sadoulel, Unjwnity h Stiglitz, CtJ~",bi. U"iW17ily. USA Kula Hoff, W",.U BanJ. JOfl~thUl Temple, U"iwniry ofB.m.t. UK Emmrnud Jimenez, World &,u Romain W=iotg, U"iwnily rfc.difornu., Fliubtth M. King, World &mit Ltn A"piu. USA Aut Knay, Wodd &mit Ruolan Yemtom', World B""J. Justin ym.. Lin, World &mit Yaohui Zhao, CCER, P.J.i~K Utoi"",u,. a,m., &""""'"' 1&.0."" is • profas.ionol journol for the diso, univpmc:nt policy anolyoio, emphati..mg policy "'levance and optntional aspecto of economic., nth.r than p.organdattheWorld funk U www.vrorhh.nk.orgIn:ocan:hI"jUU!1I:llo. lnotructior.o fur authon wishmg to submit articles ott available online at www.wbcr.oxfi>:< oil editoriol com:spondo:nce to the Editn< at ~ldbank.oq:. THE WORLD BANK ECONOMIC REVIEW Volume 24 . 2010 . Number 2 Corporate Governance at the World Bank and the Dilemma of Global Governance 171 Ashwin Ka;a and Eric Werker Measuring Household Usage of Financial Services: Does it Matter How or Whom You Ask? 199 Robert Cull and Kinnon Scott Banking on Politics: When Former High-ranking Politicians Become Bank Directors 234 Matias Braun and Claudio Raddatz Natural Disasters and Human Capital Accumulation 280 Jesus Crespo Cuaresma Managing for Results in Primary Education in Madagascar: Evaluating the Impact of Selected Workflow Interventions 303 Gerard Lassibille, Jee-Peng Tan, Cornelia Jesse, and Trang Van Nguyen Potential Implications of a Special Safeguard Mechanism in the World Trade Organization: the Case of Wheat 330 Thomas W. Hertel, Will Martin, and Amanda M. Leister Corporate Governance at the World Bank and the Dilemma of Global Governance Ashwin Kaja and Eric Werker Most major decisions at the World Bank are made by its Board of Executive Directors. While some countries enjoy the opportunity to serve on this powerful body, most countries rarely, if ever, get that chance. This gives rise to the question: Does board membership lead to higher funding from the World Bank’s two main develop- ment �?nancing institutions, the International Bank for Reconstruction and Development (IBRD) and the International Development Association (IDA). Empirical analysis shows that developing countries serving on the board can expect more than double the funding from the IBRD as countries not on the board. In absolute terms, countries on the board receive an average $60 million “bonus�? in IBRD loans, an amount that rises in years when IBRD loans are in high demand, particularly for countries in the most inﬂuential seats. This effect is more likely driven by informal rules and norms in the boardroom than by the power of the vote itself. No signi�?cant effect is found in IDA funding. These results point to challenges of global governance through representative institutions. JEL codes: F34, F35, F53, G34 Understanding that real-time decisions on global governance cannot be made by a consensus of all countries, founders of international institutions often grant decisionmaking powers to a smaller, nimbler body. But country represen- tatives on these bodies face an inevitable tension between promoting their national interests and those of the larger international community. A similar dilemma arises in U.S. politics and business. Since the seminal work of Ferejohn (1974), scholars have found that membership on powerful committees allows members of Congress to bring home the pork to their local Ashwin Kaja (kaja@post.harvard.edu.) is a juris doctor candidate at Harvard Law School. Eric Werker (corresponding author, ewerker@hbs.edu) is associate professor at Harvard Business School. The authors thank Tiffany Chan, Axel Dreher, Jeff Frieden, Arif Lakhani, Tracy Li, Linda Liu, Kenneth Mirkin, James Vreeland, Matthew Young, and seminar participants at the Political Economy of International Organizations, the International Political Economy Society meetings, and the World Bank for insightful conversations and feedback; Byron Hussie, Michael Sorell, and James Zeitler for excellent research assistance; and the journal editor and three anonymous referees for useful comments and suggestions. Werker acknowledges �?nancial support from the Harvard Business School Division of Research and Faculty Development. A supplemental appendix to this article is available at http://wber. oxfordjournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 171 –198 doi:10.1093/wber/lhq006 Advance Access Publication June 14, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 171 172 THE WORLD BANK ECONOMIC REVIEW constituencies (Ray 1981; Rundquist, Lee, and Rhee 1996; Carsey and Rundquist 1999; Rundquist and Carsey 2002). A parallel, though surprisingly thin, literature on corporate �?nance and law examines how corporate board members bene�?t from their positions at the expense of the larger company (Bebchuk and Fried 2004; Brick, Palmon, and Wald 2006). Countries—unlike states in Congress—do not have equal access to the most powerful international bodies, such as the Group of Seven industrial countries, the Group of 20 developed and developing countries, and the U.N. Security Council. This article investigates the consequences of unequal access to deci- sionmaking at the international level, looking at the politics of corporate gov- ernance in the world’s largest appropriations committee, the World Bank’s Board of Executive Directors. In 2008, the World Bank’s two primary component institutions—the International Bank for Reconstruction and Development (IBRD) and the International Development Association (IDA)—committed nearly $25 billion in loans and grants through more than 300 development projects around the globe. The IBRD offers low-interest loans to middle- and low-income develop- ing countries. The IDA focuses exclusively on loans and grants to the world’s poorest and neediest countries. Each institution is overseen by a Board of Executive Directors, composed of representatives of shareholding countries, which approves all projects and policies. The 5 largest shareholding countries appoint 5 of the 24 country members of the board, while the remaining 19 are elected for renewable two-year terms. The elected members typically vote on behalf of a handful of other countries. Thus, the vast majority of member states �?nd their interests represented by another country. The article tests whether members of the board bring more Bank funds to their home countries. It �?nds a strong effect for the IBRD. Yet a simple corre- lation may not be particularly illuminating. After all, a country with board membership may be overseeing Bank loans granted previously. Or a board seat may reﬂect a country’s rise in international prestige, which could independently bring about more World Bank projects. If true, these explanations may not be entirely troubling. The article argues, however, that the data are better explained by self-serving behavior in which membership on the board is used as a platform to channel more or larger Bank loans and grants to the directors’ home countries. The analysis does not rule out the possibility that World Bank staff prefer to lend to countries with a seat on the board, but the distributional implications are the same. The results are stark: countries receive a large increase in IBRD loans during years when they have a seat on the board. Speci�?cally, developing countries serving on the Board of Executive Directors can expect more than a doubling of funds on average from the IBRD. Countries serving on the board are rewarded with an average $60 million “bonus�? in IBRD loans. Only the time on the board, not the years before or after, is associated with increased com- mitments. Developing countries representing seats with a higher “effective Ashwin Kaja and Eric Werker 173 vote�?—seats shared by richer countries that are themselves uninterested in IBRD loans—tend to get larger increases. Moreover, countries with directors on the board receive the largest increases during years in which IBRD funding is in greatest demand—when the value of a seat is the highest. The data do not yield the same results for IDA funding: no signi�?cant association is found between board membership and IDA loan and grant com- mitments. The difference in IBRD and IDA results may be explained by differ- ences in their missions and funding policies. The IDA has allocated funding using a performance- and poverty-based formula since 1977 (IDA 2004), while GDP per capita and regional equity have been of central concern since the organization’s early years (Kapur, Lewis, and Webb 1997). While the �?ndings for the IBRD are extremely robust, there are hints that executive board power may come largely from having a seat at the table rather than from the mathematical allotment of voting power itself. Alternate board members—who are entitled to vote only when their executive director is absent—receive similar increases in commitments. (It turns out that executive directors are absent quite frequently.) With this lack of distinction, it may be that the informal workings of the IBRD Board of Executive Directors or the informal networks between Bank staff and board members/alternates reward insiders without distinguishing greatly between them. The article is organized as follows. Section I describes the World Bank’s decisionmaking structures, particularly the Board of Executive Directors. Section II introduces the data and explains the empirical methodology under- lying the analysis. Section III presents the results, while section IV examines whether the IBRD bonus varies by other factors. Section V offers some con- cluding remarks. I. DECISIONMAKING AT THE WORLD BANK The World Bank has 186 members.1 Its two main branches, the IBRD and the IDA, perform different functions that contribute to its broader mission. The IBRD, the historical core of the Bank’s operations, now directs credit mainly to middle-income and creditworthy poorer countries (World Bank 2007). Loans �?nance speci�?c projects and programs. An undisclosed methodology deter- mines the amount of IBRD loans that can be made to a country. Most countries are well below this lending ceiling. The scarcity of loans is assured by the limited staff capacity to evaluate project viability and structure loans. The IDA focuses on the world’s neediest countries—countries that fall below a certain income threshold, have poor credit ratings, or in some other way require special assistance. The IDA is more responsive to short-term disasters and emergencies and has the power to negotiate the income ceiling under 1. Unless otherwise cited, everything in this section is drawn from the World Bank’s Articles of Agreement (World Bank 2007) or its current website (www.worldbank.org). 174 THE WORLD BANK ECONOMIC REVIEW special circumstances, although a strong norm for allocative guidelines has been around since at least 1964 (Kapur, Lewis, and Webb 1997). Since 1977, an explicit formula, the Performance-Based Allocation System, has been the basis for distributing IDA funds (IDA 2004). Membership and Voting Power The World Bank is structured like many major corporations and banks. However, it is solely owned by countries, which serve as its shareholders. The IBRD currently has 186 shareholding member countries while the IDA has 169 (most countries are “blend�? countries, belonging to both groups). Each member country is required to purchase a certain number of shares based on a formula that accounts for its weight in the world economy (Woods 2001). The shareholders are technically the ultimate authority in Bank decisions. Each country is assigned a certain number of votes in broad, high-level Bank deci- sionmaking that is related to the number of shares it owns. These votes are an explicit valuation of a country’s power within the institution. The Articles of Agreement allocate 250 basic votes to each country plus one additional vote for each share of stock held. While the 250 basic votes are a concession to the principle of equality, tremendous growth in the total number of shares has marginalized their value. At the peak in 1955, the basic votes accounted for 14 percent of votes at the Bank; by 2001, they accounted for just 3 percent (Woods 2001). Decisionmaking and Election of Executive Directors Each World Bank member country appoints a governor and an alternate gover- nor to serve a �?ve-year term on the IBRD’s Board of Governors. Usually, �?nance ministers or ministers of development are chosen as governors. If the country is also a member of the IDA, the governor serves ex of�?cio on the IDA Board of Governors as well. While of�?cially the highest authority at the Bank, the Board of Governors meets only once a year. Governors “admit or suspend members, increase or decrease the authorized capital stock, determine the dis- tribution of net income, review �?nancial statements and budgets, and exercise other powers that they have not delegated to the Executive Directors�? (World Bank 2010). The Board of Governors delegates all powers not expressly reserved for the governors in the Articles of Agreement to the Board of Executive Directors. Thus, the Board of Executive Directors is responsible for the general operations of the Bank and makes important day-to-day decisions. The board meets regularly and is responsible for approving Bank loan and grant proposals put forth by management. Executive directors report to the Board of Governors on Bank operations, accounts, and other matters during the Bank’s annual meetings. Because having a board that includes all member countries would be unwieldy, the Articles of Agreement establish a procedure for having multiple Ashwin Kaja and Eric Werker 175 countries represented by one executive director. Executive directors are gener- ally elected every two years at the Bank’s annual meetings. Each member country casts all the votes allotted to that country for a single candidate. Additional election rules, which must be adopted by the Board of Governors before each election, customarily help ensure geographic diversity. As with gov- ernors, each IBRD executive director whose country is also a member of the IDA serves as an ex of�?cio member of the IDA Board of Executive Directors. In the conduct of regular business, executive directors cast votes, as a unit, for all the countries that they represent. Executive directors may appoint an alternate to assume full power and responsibilities in their absence. When executive directors are present at meetings, alternates may participate but cannot vote. In general, matters before the board are decided by majority vote. It is this decisionmaking structure that motivates the empirical analysis. At any given time, most countries do not serve on the Board of Executive Directors, making executive directors responsible for representing the varied interests of the Bank’s diverse membership in important decisions. Since execu- tive directors are expected to represent the interests of the whole, they are not to use their temporary inﬂuence to further their own countries’ agendas. Moreover, since many countries rarely or never serve on the board (tables 1 and 2), providing higher Bank funding for the countries that do would clearly be an unfair advantage. The seats are quite heterogeneous. Currently, eight seats are occupied by individual countries: China, France, Germany, Japan, Russia, Saudi Arabia, the United Kingdom, and the United States. Most of the others are shared by devel- oped and developing countries—for example, Canada and Ireland share a seat with a handful of Caribbean countries, and Australia and New Zealand share with a number of Oceanic and Southeast Asian countries. Some seats represent exclusively developing countries—for example, Sub-Saharan African countries have two seats. These groupings have changed over time: India once had its own seat, for example, while Saudi Arabia was once grouped with other predo- minantly Muslim countries. A seat may rotate among members, or a single developed country or regional hegemon may retain the directorship. Nearly all developing countries that have served as board members (and most that have not) have received funding from the World Bank. A handful of poor countries that have served on the board have received only IDA funding. Namibia is the only developing country in the sample that has served on the board but has received no World Bank funding. While there are no published analyses of the inner workings of the Board of Executive Directors,2 the authors were able to speak in con�?dence with former directors. Their descriptions suggest that each seat on the board is run differ- ently, with a different process for selecting directors and alternates, when there 2. Momani (2007) and Malone (2000) offer insightful accounts on the politics of accession to the International Monetary Fund’s executive board and the UN Security Council. 176 T A B L E 1 . Years of Service on the International Bank for Reconstruction and Development Executive Board of Directors, by Country, 1947–2005 Country Years Country Years Country Years Country Years Afghanistan 0 Dominica 0 Luxembourg 0 Singapore 0 Albania 0 Dominican Republic 0 Macedonia, FYR 0 Slovak Republic 0 Algeria 22 Ecuador 2 Madagascar 7 Slovenia 0 Angola 0 Egypt, Arab Rep. 7 Malawi 4 Solomon Islands 0 Antigua and Barbuda 0 El Salvador 3 Malaya 2 Somalia 0 Argentina 23 Equatorial Guinea 0 Malaysia 15 South Africa 0 THE WORLD BANK ECONOMIC REVIEW Armenia 0 Eritrea 2 Maldives 0 Spain 19 Australia 37 Estonia 0 Mali 4 Sri Lanka 0 Austria 8 Ethiopia 4 Malta 0 St. Kitts and Nevis 0 Azerbaijan 0 Fiji 0 Marshall Islands 0 St. Lucia 0 Bahamas 0 Finland 9 Mauritania 10 St. Vincent & the Grenadines 0 Bahrain 0 France 57 Mauritius 0 Sudan 3 Bangladesh 0 Gabon 1 Mexico 10 Suriname 0 Barbados 0 Gambia, The 2 Micronesia, Fed. Sts. 0 Swaziland 0 Belarus 0 Georgia 0 Moldova 0 Sweden 12 Belgium 50 Germany 51 Mongolia 0 Switzerland 13 Belize 0 Ghana 0 Montenegro 0 Syrian Arab Republic 3 Benin 3 Greece 2 Morocco 12 Tajikistan 0 Bhutan 0 Grenada 0 Mozambique 2 Tanzania 2 Bolivia 6 Guatemala 0 Myanmar 0 Thailand 14 Bosnia and Herzegovina 0 Guinea 0 Namibia 2 Timor-Leste 0 Botswana 2 Guinea-Bissau 3 Nepal 0 Togo 0 Brazil 11 Guyana 0 Netherlands 55 Tonga 0 Brunei Darussalam 0 Haiti 0 New Zealand 15 Trinidad and Tobago 0 Bulgaria 0 Honduras 0 Nicaragua 2 Tunisia 4 Burkina Faso/Upper Volta 0 Hungary 0 Niger 0 Turkey 6 Burundi 5 Iceland 11 Nigeria 5 Turkmenistan 0 Cambodia 0 India 58 Norway 8 Uganda 2 Cameroon 0 Indonesia 8 Oman 0 Ukraine 0 Canada 58 Iran, Islamic Rep. 0 Pakistan 29 United Arab Emirates 0 Cape Verde 0 Iraq 0 Palau 0 United Kingdom 58 Central African Republic 4 Ireland 0 Panama 2 United States 58 Chad 0 Israel 0 Papua New Guinea 0 Uruguay 7 Chile 10 Italy 42 Paraguay 3 Uzbekistan 0 China 50 Jamaica 0 Peru 5 Vanuatu 0 Colombia 33 Japan 51 Philippines 8 Venezuela 12 Comoros 4 Jordan 0 Poland 3 Vietnam 0 Congo, Dem. Rep. 0 Kazakhstan 0 Portugal 0 Yemen 0 Congo, Rep. 4 Kenya 0 Qatar 0 Yemen, Rep. 0 Costa Rica 1 Kiribati 0 Romania 0 Yugoslavia 0 Co ˆ te d’Ivoire 0 Korea, Rep. 5 Russian Federation 13 Zambia 0 Croatia 0 Kuwait 20 Rwanda 0 Zimbabwe 0 Cuba 18 Kyrgyz Republic 0 Samoa 0 Cyprus 0 Lao PDR 0 San Marino 0 Czech Republic 0 Latvia 0 Sa ˜ o Tome ´ncipe ´ and Prı 0 Czechoslovakia 0 Lebanon 0 Saudi Arabia 19 Dahomey 0 Lesotho 1 Senegal 0 Denmark 11 Liberia 3 Serbia 0 Djibouti 0 Libya 0 Seychelles 0 Lithuania 0 Sierra Leone 2 Source: Authors’ analysis based on data described in the text. Ashwin Kaja and Eric Werker 177 178 T A B L E 2 . Years of Service on the International Development Association Board of Executive Directors, by Country, 1961–2005 Country Years Country Years Country Years Country Years Afghanistan 0 Dominica 0 Luxembourg 0 Singapore 0 Albania 0 Dominican Republic 0 Macedonia, FYR 0 Slovak Republic 0 Algeria 22 Ecuador 0 Madagascar 7 Slovenia 0 Angola 0 Egypt, Arab Rep. 7 Malawi 4 Solomon Islands 0 Antigua and Barbuda 0 El Salvador 3 Malaya 0 Somalia 0 Argentina 21 Equatorial Guinea 0 Malaysia 15 South Africa 0 THE WORLD BANK ECONOMIC REVIEW Armenia 0 Eritrea 2 Maldives 0 Spain 17 Australia 25 Estonia 0 Mali 4 Sri Lanka 0 Austria 8 Ethiopia 4 Malta 0 St. Kitts And Nevis 0 Azerbaijan 0 Fiji 0 Marshall Islands 0 St. Lucia 0 Bahamas 0 Finland 9 Mauritania 10 St. Vincent & the Grenadines 0 Bahrain 0 France 44 Mauritius 0 Sudan 2 Bangladesh 0 Gabon 1 Mexico 10 Suriname 0 Barbados 0 Gambia, The 2 Micronesia, Fed. Sts. 0 Swaziland 0 Belarus 0 Georgia 0 Moldova 0 Sweden 9 Belgium 36 Germany 44 Mongolia 0 Switzerland 13 Belize 0 Ghana 0 Montenegro 0 Syrian Arab Republic 3 Benin 3 Greece 0 Morocco 12 Tajikistan 0 Bhutan 0 Grenada 0 Mozambique 2 Tanzania 2 Bolivia 6 Guatemala 0 Myanmar 0 Thailand 14 Bosnia and Herzegovina 0 Guinea 0 Namibia 2 Timor-Leste 0 Botswana 2 Guinea-Bissau 3 Nepal 0 Togo 0 Brazil 11 Guyana 0 Netherlands 44 Tonga 0 Brunei Darussalam 0 Haiti 0 New Zealand 15 Trinidad and Tobago 0 Bulgaria 0 Honduras 0 Nicaragua 0 Tunisia 4 Burkina Faso/Upper Volta 0 Hungary 0 Niger 0 Turkey 0 Burundi 5 Iceland 10 Nigeria 5 Turkmenistan 0 Cambodia 0 India 45 Norway 6 Uganda 2 Cameroon 0 Indonesia 5 Oman 0 Ukraine 0 Canada 45 Iran, Islamic Rep. 0 Pakistan 20 United Arab Emirates 0 Cape Verde 0 Iraq 0 Palau 0 United Kingdom 44 Central African Republic 4 Ireland 0 Panama 0 United States 44 Chad 0 Israel 0 Papua New Guinea 0 Uruguay 7 Chile 6 Italy 37 Paraguay 1 Uzbekistan 0 China 37 Jamaica 0 Peru 3 Vanuatu 0 Colombia 26 Japan 43 Philippines 8 Venezuela 12 Comoros 4 Jordan 0 Poland 0 Vietnam 0 Congo, Dem. Rep. 0 Kazakhstan 0 Portugal 0 Yemen 0 Congo, Rep. 4 Kenya 0 Qatar 0 Yemen, Rep. 0 Costa Rica 1 Kiribati 0 Romania 0 Yugoslavia 0 Co ˆ te d’Ivoire 0 Korea, Rep. 5 Russian Federation 13 Zambia 0 Croatia 0 Kuwait 20 Rwanda 0 Zimbabwe 0 Cuba 9 Kyrgyz Republic 0 Samoa 0 Cyprus 0 Lao PDR 0 San Marino 0 Czech Republic 0 Latvia 0 Sa ˜ o Tome ´ncipe ´ and Prı 0 Czechoslovakia 0 Lebanon 0 Saudi Arabia 19 Dahomey 0 Lesotho 1 Senegal 0 Denmark 9 Liberia 3 Serbia 0 Djibouti 0 Libya 0 Seychelles 0 Lithuania 0 Sierra Leone 2 Source: Authors’ analysis based on data described in the text. Ashwin Kaja and Eric Werker 179 180 THE WORLD BANK ECONOMIC REVIEW is any rotation at all. In board meetings, loan requests are almost always granted, and according to these sources, a norm exists that restrains directors from developing countries from voting against each other’s requests. However, before meetings, the directors—most of them politically important �?gures from their home countries—often have a chance to make their views known on potential projects. Bank staff, meanwhile, are aware of the board status of the countries to which they are lending. I I . D ATA AND M E T H O D O LO GY The empirical strategy for determining whether countries serving on the World Bank’s Board of Executive Directors are able to use this position of inﬂuence to bring more Bank funding to their countries is to observe how approval of World Bank loan commitments varies as a function of having a seat on the board or not. A simple correlation between board membership and loan com- mitments is not, in itself, necessarily illuminating. Factors that affect a coun- try’s likelihood of serving on the Board of Executive Directors and its likelihood of receiving World Bank funding could bias this result. The methods described here provide empirical support for the hypothesis that board mem- bership itself, rather than alternative explanations, drives the positive associ- ation between Bank funding commitments and board membership—the pork-barrel hypothesis. A panel dataset was constructed featuring Part II IDA countries (mainly low- and middle-income countries) that have been members of the World Bank at any time since its founding. The main dependent variables are approved loan commitments from the IBRD and approved loan and grant commitments from the IDA. Data on all World Bank development projects since 1946 are readily available on its website, but the sample is restricted to projects after 1961, when the Part-II IDA de�?nition �?rst came into use. Funding commit- ments for each country in a given year are summed and converted to 1996 dollar values.3 In analyzing the characteristics of World Bank lending, this article is in the tradition of Andersen, Hansen, and Markussen (2006) and Kilby (2009). The World Bank’s annual reports contain a wealth of information, if not in a readily usable format (World Bank various years). The information is used to construct three key variables. The �?rst is a dummy variable (Board member) representing whether a country served on the Board of Executive Directors in a given year. The number of times each country in the dataset served on the IBRD and the IDA boards is shown in tables 1 and 2. Around half the 3. Since the project database contains only approved loan commitments, countries eligible but not receiving funding are omitted. So that these country-years are included in the dataset, they are assigned a value of zero. In speci�?cations using the logarithm of Bank commitments as the dependent variable, these values are set to a negligible $1 (since ln(1) ¼ 0). Ashwin Kaja and Eric Werker 181 countries that were members of the World Bank at some point since its founding have never served on the board—countries from Afghanistan and Albania to Zambia and Zimbabwe. Other countries, including Colombia, India, and Pakistan, have served many times. Because terms on the board begin and end in the middle of the calendar year while all other data are on a calen- dar year basis, the half-year lag effect must be accounted for when interpreting the results of the analysis. A similar variable reﬂects the same information for alternate directors (Alternate board). Summary statistics (reported in table S1 in the supplemental appendix, available at http://wber.oxfordjournals.org) are shown in table 3 for full board members, alternate board members, and other countries. A second variable indicates the amount of voting power assigned to each Bank member country based on its number of World Bank shares (Bank voting power). The amount of voting power may indicate a country’s pull within the institution. A third variable, drawn from data in the annual reports, reﬂects the aggregate voting power each board member wields in making board decisions (Board voting power). It is calculated by dividing each executive director’s votes (equal to the sum of the general Bank votes of each country that execu- tive director represents) by the total votes available that year. While Board member is a dummy variable that identi�?es whether a country is serving on the board in a given year, this scalar variable takes into account the fact that all board seats are not created equal by scaling membership by each seat’s aggre- gate voting power. Board voting power for developing countries on the IBRD board, for instance, varies from less than 1.6 percent (Comoros in 1995) to nearly 6 percent (India in 1982). The pork-barrel hypothesis is tested using two types of speci�?cations. The main model is a �?xed effects regression, and the second is an event-time speci�?- cation that continues to use �?xed effects but analyzes trends before and after membership. Robustness checks and additional tests show whether some types of board members are systematically more effective in bringing home additional aid. The main speci�?cations use the logarithm of World Bank funding commitments as the dependent variable, while alternate speci�?cations also use absolute levels of commitments and a logit model. Additionally, loan receipts are compared for countries represented by executive directors and those represented by alternate directors to examine whether the vote itself or a more complex institutional factor explains increased loans. All speci�?cations include standard errors clustered at the country level. In addition to the Bank voting power variable described above, the analysis controls for several other factors. Data from the Penn World Tables (Heston, Summers, and Aten 2009) are used to control for the log of real per capita GDP and population—both factors that might inﬂuence the funding commit- ments a country receives as well as its likelihood of being elected to the board. The analysis also controls for two political variables that could have a signi�?- cant impact on World Bank lending decisions or board election. The �?rst is the 182 THE WORLD BANK ECONOMIC REVIEW T A B L E 3 . Summary Statistics for Members of the World Bank Board of Executive Directors, Alternates, and Nonmembers, 1961–2004 Number of Variable observations Mean Standard deviation International Bank for Reconstruction and Development Statistics for board members Bank voting power (% total) 420 1.193 1.179 Per capita real GDP (1995 US$) 407 4751 4242 Population (thousands) 429 187734 344052 Major war ( . 1000 deaths, dummy variable) 429 0.091 0.288 Democracy (scale – 1 to 1) 413 0.123 7.386 IBRD commitments (1996 US$) 429 387.3 637.4 Statistics for alternate board members Bank voting power (% total) 489 0.438 0.411 Per capita real GDP (1995 US$) 420 5613 5302 Population (thousands) 447 31693 40091 Major war ( . 1000 deaths, dummy variable) 489 0.070 0.255 Democracy (scale – 1 to 1) 436 2 0.661 7.254 IBRD commitments (1996 US$) 489 165.9 367.9 Statistics for board nonmembers Bank voting power (% total) 4,277 0.204 0.262 Per capita real GDP (1995 US$) 3,924 5021 6206 Population (thousands) 4,533 13442 42881 Major war ( . 1000 deaths, dummy variable) 4,727 0.069 0.253 Democracy (scale – 1 to 1) 3,607 2 1.335 7.411 IBRD commitments (1996 US$) 4,727 57.85 217.6 International Development Association Statistics for board members Bank voting power (% total) 391 1.154 1.152 Per capita real GDP (1995 US$) 406 4760 4244 Population (thousands) 428 188133 344356 Major war ( . 1000 deaths, dummy variable) 428 0.091 0.288 Democracy (scale – 1 to 1) 412 0.141 7.387 IBRD commitments (1996 US$) 428 172.3 435.9 Statistics for alternate board members Bank voting power (% total) 427 0.528 0.511 Per capita real GDP (1995 US$) 419 5624 5303 Population (thousands) 446 31762 40110 Major war ( . 1000 deaths, dummy variable) 488 0.070 0.255 Democracy (scale – 1 to 1) 435 2 0.646 7.256 IBRD commitments (1996 US$) 488 49.37 146.75 Statistics for board nonmembers Bank voting power (% total) 3,757 0.245 0.249 Per capita real GDP (1995 US$) 3,926 5019 6205 Population (thousands) 4,535 13441 42872 Major war ( . 1000 deaths, dummy variable) 4,729 0.069 0.253 Democracy (scale – 1 to 1) 3,609 2 1.338 7.410 IBRD commitments (1996 US$) 4,729 23.13 62.45 Source: Authors’ analysis based on data described in the text. Ashwin Kaja and Eric Werker 183 occurrence of a war with at least 1,000 battle deaths in a given country-year, using data from the Department of Peace and Conﬂict Research at Uppsala University and the International Peace Research Institute in Oslo (Uppsala University and PRIO 2007). The second is the political climate in a country (whether its government can be characterized as a democracy, autocracy, or something in between), based on the Polity 2 variable developed by the University of Maryland’s Center for International Development and Conﬂict Management in its Polity IV data set (Marshall and Jaggers 2007). This vari- able, coded as a score from –10 ( perfect autocracy) to þ 10 ( perfect democ- racy), is referred to as the Democracy variable. Primary Fixed Effects Speci�?cation Even after controlling for the variables just described, there are other omitted effects that could bias estimates of the value of a seat on the Board of Executive Directors in World Bank funding commitments. Country and year �?xed effects in the main speci�?cation are used to account for such bias. Trends in World Bank funding over time will be absorbed by the year �?xed effects, while omitted variables that affect individual countries’ average loan receipts will be absorbed by the country �?xed effects. The primary logarithmic �?xed effects regression, following Alesina and Dollar (2000) and Kuziemko and Werker (2006), is as follows: lnðLoan commitmentsÞit ¼ b0 þ b1 ðBoard memberÞit þ b2 ðBank Voting PowerÞit þ b3 ðXÞit þ gt þ di þ 1it ð1Þ where Loan commitments represents the amount of money committed by either the IBRD or the IDA to country i in year t; Board member is a dummy variable for whether a country has a seat on the Board of Executive Directors; X is a vector of time-varying World Bank, political, and economic controls for each country; g is a vector of year �?xed effects; and d is a vector of country �?xed effects. Note that the half-year lag due to board terms beginning and ending in the middle of a calendar year (so that a country’s last year serving as an executive director before leaving the board is actually only half a year) may bias estimates of b1 downward. In another speci�?cation, the dummy variable for board membership is replaced by the scalar variable Board voting power. Event-Time Speci�?cation If countries use board membership to create awareness of their legitimate development needs, an increase in funding may not be entirely troubling. An event-time speci�?cation, similar to that used by Kuziemko and Werker (2006), helps rule out this and other alternative explanations. The event-time regression 184 THE WORLD BANK ECONOMIC REVIEW is as follows: lnðLoan commitmentsÞit ¼ b0 þ b1 ðT À 3Þit þ b2 ðT À 2Þit þ b3 ðT À 1Þit þ b4 ðBoard memberÞit þ b5 ðT þ 1Þit þ b6 ðT þ 2Þit þ b2 ðBank voting powerÞit þ b3 ðXÞit þ gt þ di þ 1it ð2Þ where T – x is a dummy variable indicating that the year is x full calendar years before a country begins its term on the Board of Executive Directors and T þ x is a dummy variable indicating that the year is x full calendar years after the country has completed its term on the board. The time dummy variables are extended to T – 3 years to account for the lag caused by executive director terms beginning in the middle of a calendar year. Because of this effect, T – 1 includes half a year of board service. As with the primary �?xed effects speci�?- cation, this lag might bias the estimate of b4 downward. This speci�?cation enables identifying the effect of serving on the board by comparing a country’s loan commitments during years of board membership with those in the years immediately before its term begins and the years immediately after its term ends. (It also enables identifying any lag structure between project conceptualization and board-approved commitments. An infra- structure project may require three or more years to iron out all the details, but other categories—like budget support—can be approved much faster.) A sharp increase in loan commitments during a country’s term compared with the years immediately before and after the term would help to rule out alternative expla- nations for a positive association between board membership and Bank funding and lend credence to the pork-barrel hypothesis. I II. RE S U LTS The analysis �?nds a striking contrast in the returns to board membership between the IDA and the IBRD. International Bank for Reconstruction and Development The results of the main speci�?cations (regressions 1, 2, and 3) for IBRD com- mitments are presented in table 4. The �?rst speci�?cation regresses the logarithm of IBRD funding commitments on the board membership dummy variable including country and year �?xed effects but excluding control variables. The results show a statistically (at the 5 percent level) and quantitatively signi�?cant estimate of the coef�?cient of the board membership variable. The addition of control variables in regression 2 has a negligible effect on that estimate, which remains signi�?cant (at the 10 percent level). Board membership in a given year is associated with a 138 log-point, or roughly 300 percent, increase in World Bank loans to the country (e1.38 ¼ 2.97). While there are no signi�?cant Ashwin Kaja and Eric Werker 185 T A B L E 4 . Ordinary Least Squares Regressions of Logarithm of International Bank for Reconstruction and Development Commitments on Board Membership (1996 U.S. dollars) Variable Regression (1) Regression (2) Regression (3) Regression (4) Board member 1.464 1.382 1.509 (1.99)** (1.97)* (1.84)* Board voting power 0.539 (2.49)** World Bank voting power 0.967 0.781 0.93 (0.88) (0.71) (0.83) ln(per capita real GDP) 1.982 1.998 1.96 (1.6) (1.6) (1.57) ln(population) 0.368 0.393 0.331 (0.12) (0.12) (0.1) Major war 2 1.881 2 1.882 2 1.915 (2.51)** (2.51)** (2.55)** Democracy 2 0.01 2 0.011 2 0.011 (0.26) (0.27) (0.27) Board entry 2 3 years 2 0.835 (1.15) Board entry 2 2 years 0.503 (0.65) Board entry 2 1 year 0.054 (0.07) Board exit þ 1 year 2 0.096 (0.1) Board exit þ 2 years 0.912 (1.02) Number of observations 5645 4061 4061 4061 Number of countries 173 135 135 135 R squared 0.04 0.08 0.08 0.08 *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Numbers in parentheses are the absolute values of robust t-statistics. All regressions include country and year �?xed effects; standard errors are clustered at the country level. Source: Authors’ analysis based on data described in the text. coef�?cients on the per capita income, population, or democracy controls in any of the main IBRD speci�?cations, the occurrence of a major war in a country has a statistically signi�?cant negative effect on World Bank loans in all these speci�?cations. The positive coef�?cient on per capita income indicates that a higher volume of IBRD loans ﬂows to the middle-income countries in the sample. Speci�?cation 3 replaces the dummy variable for board membership with a scalar representing a country’s voting power on the board. Countries not serving on the board receive a value of zero. The positive coef�?cient on board voting power, signi�?cant at the 5 percent level, indicates that a 1 percentage point increase in board voting power (measured as a percentage of the total 186 THE WORLD BANK ECONOMIC REVIEW amount of available voting power) is associated with a 71 percent increase in IBRD loans. Speci�?cation 4, the event-time speci�?cation, adds dummy variables for the years immediately before and after board service. The coef�?cient on board membership remains essentially unchanged (it increases slightly) and remains statistically signi�?cant at the 10 percent level. None of the dummy variables is statistically signi�?cant or shows any clear pattern, suggesting that there is an increase in Bank loans to a country during years when the country is serving on the board but not before its entry onto the board or after its exit. If omitted characteristics such as changes in a country’s reputation within the World Bank or changes in a country’s development needs are inﬂuencing both a coun- try’s election to the Board of Executive Directors and its ability to attract Bank funding, a rise in funding would be expected not just during years of service but in the years before and after service as well. If countries use board member- ship as a platform to draw attention to their legitimate development needs, that increased awareness would not be expected to disappear after the country completes its board term. By helping rule out several alternative hypotheses, the event-time speci�?cation lends credence to the hypothesis that increases in IBRD loans are closely tied to an insider bonus for countries serving on the board. ALTERNATIVE SPECIFICATIONS. Regressions were also run with several alternative speci�?cations (table 5). First, the dependent variable, the logarithm of IBRD commitments, is replaced with absolute commitment levels. The absolute regressions are included because there is no obvious ex ante reason to believe that loan bonuses work as a proportion of existing loans. Speci�?cation 1 shows that the board membership variable is signi�?cant at the 10 percent level, indi- cating an approximately $60 million bonus from board membership. While the major war control variable ceases to be signi�?cant in this regression, a positive coef�?cient on per capita GDP, signi�?cant at the 10 percent level, suggests that wealthier developing countries may see larger Bank loans in a given year. Event-time speci�?cation 2, using absolute commitment values as the dependent variable, shows higher Bank loans only in years of board service and the year immediately before board service. (Since the analysis for the year immediate before board service includes the �?rst half-year of board service due to the lag caused by the Bank’s election schedule, as described in section II, this result also con�?rms the �?ndings in the main speci�?cations.) Speci�?cations 3 and 4 present the results of a logit model used to determine whether board membership can explain a country’s receipt of World Bank loans in a given country-year.4 These results also show a positive coef�?cient on the board membership variable, suggesting a story on the extensive margin that 4. Tobit and Heckman models failed to converge with country and year �?xed effects. The logit sample size is reduced since observations that never change board member status are dropped. T A B L E 5 . Alternative Speci�?cations: Ordinary Least Squares Regressions of International Bank for Reconstruction and Development Commitments on Board Membership (1996 U.S. dollars) Variable Absolute (1) Absolute (2) Logit (3) Logit (4) Log (5) Log (6) Board member 59.758 72.256 0.513 0.575 1.835 1.739 (1.90)* (2.08)** (1.64) (1.55) (2.38)** (2.02)** Alternate board member 1.953 2.468 (3.12)*** (3.33)*** Bank voting power 62.687 55.917 0.521 0.515 0.995 1.135 (1.2) (1.06) (0.88) (0.86) (0.95) (1.1) ln(per capita real GDP) 113.717 111.489 1.017 1.008 1.867 1.768 (1.73)* (1.67)* (1.91)* (1.89)* (1.5) (1.41) ln(population) 2 67.594 2 68.569 2 0.277 2 0.292 0.561 0.582 (0.81) (0.82) (0.19) (0.2) (0.18) (0.19) Major war 2 34.029 2 35.204 2 1.06 2 1.093 2 1.825 2 1.86 (1.34) (1.39) (2.89)*** (3.01)*** (2.46)** (2.53)** Political climate 2 1.734 2 1.839 0.005 0.005 2 0.008 2 0.005 (1.13) (1.24) (0.22) (0.22) (0.21) (0.14) Executive director entry 2 3 years 2 7.76 2 0.418 2 1.676 (0.29) (1.42) (2.09)** Executive director entry 2 2 years 6.309 0.253 2 0.24 (0.22) (0.81) (0.3) Executive director entry 2 1 year 74.235 2 0.121 2 0.654 (1.75)* (0.39) (0.82) Executive director exit þ 1 year 2 12.589 2 0.09 2 0.038 (0.52) (0.23) (0.04) Executive director exit þ 2 years 40.952 0.491 1.033 (0.92) (1.28) (1.15) Alternate entry 2 3 years 0.366 (0.55) Ashwin Kaja and Eric Werker Alternate entry 2 2 years 1.763 (2.74)*** 187 (Continued ) 188 TABLE 5. Continued Variable Absolute (1) Absolute (2) Logit (3) Logit (4) Log (5) Log (6) Alternate entry 2 1 year 2.38 (3.17)*** Alternate exit þ 1 year 0.483 (0.65) Alternate exit þ 2 years 0.398 THE WORLD BANK ECONOMIC REVIEW (0.5) Number of observations 4,061 4,061 3,148 3,148 4,061 4,061 Number of countries 135 135 135 135 R squared 0.1 0.1 0.08 0.09 *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Numbers in parentheses are the absolute values of robust t-statistics. All regressions include country and year �?xed effects; standard errors are clustered at the country level. Source: Authors’ analysis based on data described in the text. Ashwin Kaja and Eric Werker 189 is similar to that found in the logarithmic and absolute speci�?cations, although the coef�?cients are not statistically signi�?cant at conventional levels. Finally, the IBRD analysis is extended in speci�?cation 5 by adding dummy variables representing alternate board membership to speci�?cation 1. While the coef�?cient estimate for board membership does not change much, the coef�?- cient on the alternate board membership variable is also signi�?cant and similar in magnitude. A t-test comparing the coef�?cients on board membership and alternate board membership shows no signi�?cant difference. The remaining variable coef�?cients, including that of the signi�?cant major war variable, show little change from the main speci�?cations without alternate membership. Since alternate board members may participate in board meetings but may vote only when the appointing executive director is absent, the similar positive and stat- istically signi�?cant effect on alternate and full board members suggests that the loan increases resulting from board membership may stem not just from formal power, but also from the informal norms governing the board, its relationship with Bank staff, and its exercise of power. An analysis of board minutes available on the World Bank’s website (minutes from the �?rst IBRD and IDA board meetings of each month since April 2005) suggests that there may be little difference in practice in the voting power of directors and alternates. The limited sample of minutes from these 58 meetings (results available on request) shows that an average of 10.4 execu- tive directors were present in their voting capacity and that an average of 13.6 of the 24 voting seats were occupied by alternate directors voting in place of their executive directors. Only 3.8 alternates, on average, were present as non- voting attendees. While this limited sample seems to con�?rm the �?nding that alternate board membership may have an effect similar to that of full board membership, the �?ndings on the inﬂuence of alternate directors are not as robust, as demonstrated in the event-time speci�?cation in column 6 of table 5. ROBUSTNESS CHECKS. A series of robustness checks (reported in table S2 of the online supplemental appendix) explore potential limitations of the main analy- sis, using various techniques to further validate the �?ndings. To correct for any potential case selection bias, all countries in the dataset that have never served on the board were dropped from the analysis. To ensure that the particularities of the logarithmic speci�?cation (when zero commitment values are not uncom- mon) are not driving the results, the regressions were rerun after raising all zero commitment values to 12.5, a value just below the lowest nonzero logar- ithmic commitment value in the data. The �?ndings remain, with the coef�?cients (mechanically) reduced. To ensure that no other particularities of the panel data format are driving the result, the board member variable was replaced with a placebo—board membership 10 years before—and no positive coef�?- cient was found on the placebo variable. For each group of countries in the World Bank whose interests are rep- resented by a single executive director, there are different expectations about 190 THE WORLD BANK ECONOMIC REVIEW who will represent the group. For most groups, member countries—at least the larger ones—take turns on the board. But for a handful of groups, such as the one that includes India, one country always maintains the seat. As an additional robustness check, groups that do not allow meaningful rotation are dropped from the analysis. The patterns remain the same. Another way to con- ceive of the empirical speci�?cation is to take advantage of this group data by adding group-year �?xed effects. This speci�?cation essentially compares countries that serve on the board with countries in their group that are not on the board. Even with the large reduction in degrees of freedom, the same pat- terns remain, though the coef�?cient on the absolute level of commitments loses more than half its magnitude. Finally, the results are robust to dropping countries with income below $800 per capita, as such countries are more likely to borrow through the IDA (available on request). International Development Agency The results of the main speci�?cations for the IDA are presented in table 6. In stark contrast to the �?ndings for the IBRD, the regression of IDA funding on board membership with country �?xed effects but without control variables shows no signi�?cant association between the two variables (column 1). The main speci�?cations again fail to yield a statistically signi�?cant coef�?cient esti- mates on the board membership variable (columns 2 –4). As with the main IBRD speci�?cations, the only control variable that is signi�?cant is that for the occurrence of a major war. Mirroring the IBRD case, countries experiencing a major ongoing war can expect a very large decrease in IDA funding. The coef�?- cient on per capita income, in contrast to the IBRD speci�?cation, is negative, indicating that IDA grants and loans ﬂow to the poorer countries in the sample. The alternate speci�?cations using absolute commitment levels as the dependent variable (columns 5 and 6) similarly fail to �?nd a signi�?cant link between board membership and funding.5 (While answering a different ques- tion, these results complement those of Andersen and others (2006), who �?nd that IDA allocation is correlated with U.S. strategic interest.) The vast difference in results for the IBRD and the IDA raises interesting questions. Why would two institutions, similarly structured, exhibit such differ- ent behavior for the association between board membership and funding? One plausible explanation is that the difference stems from their different missions. The IDA’s exclusive focus on the poorest, neediest countries might well alter 5. Since the IDA targets low-income countries, it may be prudent, as one referee suggested, to limit the sample to countries with especially low income per capita. The analysis revealed that the IDA regularly awarded projects to countries with GDP per capita upwards of $4,000 (in 1996 dollars). The speci�?cations in table 6 were rerun, alternately limiting the sample to countries with incomes of $8,000 and $800 per capita; the lack of signi�?cance for the board member variable remained (results available on request). The sample limited to $800, due partly to its smaller size, was particularly noisy and not robust to modi�?cations. However, the coef�?cients on board member were closer in magnitude to those in the IBRD sample. Ashwin Kaja and Eric Werker 191 T A B L E 6 . Ordinary Least Squares Regressions of International Development Association Commitments on Board Membership (1996 U.S. dollars) Absolute Absolute Variable Log (1) Log (2) Log (3) Log (4) (5) (6) Board member 2 0.007 0.328 0.408 9.03 8.585 (0.01) (0.47) (0.55) (0.5) (0.61) Board voting power 0.042 (0.24) World Bank voting power 2 0.672 2 0.641 2 0.71 2 32.368 2 32.268 (0.51) (0.49) (0.54) (0.86) (0.86) ln(per capita real GDP) 2 1.531 2 1.516 2 1.542 2 13.077 2 12.811 (1.12) (1.1) (1.12) (0.62) (0.61) ln(population) 2 0.09 2 0.064 2 0.088 25.799 25.775 (0.02) (0.02) (0.02) (0.92) (0.92) Major war 2 2.309 2 2.311 2 2.292 2 38.292 2 38.376 (3.19)*** (3.19)*** (3.16)*** (1.70)* (1.69)* Democracy 0.013 0.014 0.011 2 0.248 2 0.24 (0.38) (0.4) (0.33) (0.63) (0.61) Board entry 2 3 years 0.433 2 10.77 (0.59) (1.11) Board entry 2 2 years 1.068 8.04 (1.47) (0.83) Board entry 2 1 year 0.683 2 3.415 (0.95) (0.49) Board exit þ 1 year 2 0.452 2 3.009 (0.64) (0.58) Board exit þ 2 years 2 0.802 2 4.285 (1.54) (0.63) Number of observations 5,624 3,619 3,619 3,619 3,619 3,619 Number of countries 173 122 122 122 122 122 R squared 0.05 0.04 0.04 0.05 0.05 0.05 *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Numbers in parentheses are the absolute values of robust t-statistics. All regressions include country and year �?xed effects; standard errors are clustered at the country level. Source: Authors’ analysis based on data described in the text. the dynamics of board politics and decisionmaking, reducing the effect of its institutional structure on outcomes. A more speci�?c explanation centers on the strong role of easily observable factors such as per capita income and region in dictating IDA allocations since its early years. In fact, the IDA has used speci�?c formulas for funding allocation since 1977. A third explanation, suggested by World Bank staff, is that the IDA has an additional institutional structure of approval, effectively weakening the power of the board. Evidence for India has found that �?scal transfers became much less politically determined when they were transferred to an independent agency (Khemani 2007). Testing these hypotheses is beyond the scope of this article, as this norm has been present with the organization since its founding, even though the IDA did not adopt 192 THE WORLD BANK ECONOMIC REVIEW speci�?c formulas until 1977. That said, the correlation between IDA board membership and grants and loans—weak at all times—is weaker before 1977, suggesting that the norm rather than the formula is the driving factor (results available on request). I V. D I F F E R E N T I A L T R E A T M E N T To follow up on the �?nding that board membership does lead to higher IBRD commitments, do certain characteristics allow countries to exploit board mem- bership better than other countries do? The primary �?xed effects speci�?cation is rerun with various interaction effects: lnðLoan commitmentsÞit ¼b0 þ b1 ðBoard memberÞit þ b1 ðBoard member x hÞit þ b2 ðBank voting powerÞit þ b3 ðXÞit þ gt þ di þ 1it ð3Þ where h is the variable being interacted with the dummy variable Board member. The �?rst interaction term examines whether a country’s per capita GDP inﬂuences the bene�?t it derives from board membership. It is possible that economically stronger developing countries may command more respect or inﬂuence, making them more able to cash in on insider status. Second, the interaction between board membership and democracy deter- mines whether political climate (a country’s degree of autocracy or democracy) has an effect on a country’s ability to convert board membership into funding. For example, the board may be biased against more autocratic governments, limiting their ability to bene�?t from an insider position. Third, the analysis explores whether the effect of board membership on World Bank funding is signi�?cantly different in the years before and after the cold war by interacting board membership with a dummy variable indicating whether the year is after 1990. The end of the cold war altered the balance of power in the international system and, consequently, might have inﬂuenced the operations of international institutions in a noteworthy manner. Fourth, board membership is interacted with a dummy variable indicating whether a country has served on the board for more than 14 years—approxi- mately a third of the time covered by the data set. Countries that have served longer might have more experience and command more respect on the board, enabling them to take greater advantage of their board membership. On the other hand, countries that frequently serve on the board may be less eager to exploit the opportunity or may see the returns spread out over multiple board terms. Fifth, board membership is interacted with board voting power in a regression that pools the dummy and the scalar variables for board Ashwin Kaja and Eric Werker 193 membership and tests whether countries representing powerful groups achieve larger gains from board service. Sixth, board membership is interacted with a scaled measure of board voting power labeled Effective voting power. Since developed countries do not receive World Bank loans, board voting power is multiplied by the ratio of total votes to developing country votes. A developing country that shares its board seats with developed countries (which are not clamoring for loans) should have a larger effective vote than a developing country that shares its board seats with other developing countries. And �?nally, board membership is interacted with the total annual loan volume of the IBRD. Since IBRD loans are ostensibly set at market rates, the bene�?t of securing a loan may be limited during quiet years in international �?nance. (Of course, because sovereign debt markets were very poorly devel- oped before the 1980s, and because World Bank loans stretch well beyond the political horizon of most politicians, the IBRD has remained an attractive place to borrow.) The total annual volume of IBRD loans serves as a proxy for the value of a World Bank loan to a potential borrower since commercial loans are harder, if not impossible, to �?nd in years of crisis, making the otherwise “market�? rates of the IBRD relatively more attractive.6 This speci�?cation tests the hypothesis that when the value of a Bank loan is high, directors with a seat on the board will be more likely to exercise their inﬂuence to direct a loan to their home country. The results of these various interaction effects are reported in table 7. Speci�?cation 1 shows a positive but not statistically signi�?cant estimate of the coef�?cient of the interaction between per capita GDP and board membership. The interaction between board membership and democracy in speci�?cation 2 yields a negative coef�?cient, but one that is also not statistically signi�?cant. Speci�?cation 3 shows that the �?nancial bonus from board membership is substantially higher after 1990. There could be several explanations for this �?nding: the opening of international political space to developing countries fol- lowing the depoliticization of strategic aid, the inﬂuence of the new states emerging from the formerly Communist countries, or changes in the rules and norms at the World Bank. The estimate on the interaction term in speci�?cation 4 suggests that countries that have served on the board for more than 14 years see higher returns to board membership, though this coef�?cient estimate is not quite statistically signi�?cant. The results for speci�?cation 5 display a positive but not statistically signi�?- cant effect, indicating that it is hard to determine whether countries with greater board voting power convert their board membership into higher IBRD commitments. Speci�?cation 6 shows that statistically signi�?cant additional 6. Alternate indicators are unavailable because of the poorly developed sovereign bond market. Emerging market bond yields are available only for a subset of countries and beginning only in the 1980s. 194 T A B L E 7 . Differential Treatment of Board Membership: Ordinary Least Squares Regressions of International Bank for Reconstruction and Development Commitments on Board Membership (1996 U.S. dollars) Democracy Post-1990 . 14 years Board voting Effective Total IBRD Dependent variable and interaction terms GDP (1) (2) (3) on board (4) power (5) votea (6) loans (7) Board member 2 2.688 1.369 0.326 0.539 2 1.917 2 0.521 2 2.640 (0.53) (2.04)** (0.37) (0.76) (0.98) (0.56) (1.35) THE WORLD BANK ECONOMIC REVIEW Board voting powerb 0.485 (1.43) Effective voting powera,b 0.047 (0.4) Bank voting power 0.877 0.776 0.96 0.724 0.499 0.863 0.803 (0.81) (0.72) (0.84) (0.72) (0.43) (0.8) (0.76) ln(per capita real GDP) 1.932 1.999 1.76 2.001 1.282 1.22 1.088 (1.59) (1.62) (1.51) (1.62) (1.04) (0.98) (0.90) ln(population) 0.345 0.268 0.484 0.307 0.717 0.466 0.487 (0.11) (0.08) (0.15) (0.1) (0.22) (0.14) (0.15) Major war 2 1.894 2 1.866 2 2.024 2 1.903 2 1.69 2 1.697 2 1.743 (2.54)** (2.47)*** (2.75)*** (2.55)** (2.33)** (2.31)** (2.37) Democracy 2 0.009 2 0.002 2 0.008 2 0.007 0 0.001 0.002 (0.23) (0.05) (0.21) (0.19) (0.01) (0.02) (0.06) Board member*ln(per capita real GDP) 0.504 (0.78) Board member*Democracy 2 0.114 (1.12) Board member*Post-1990 3.126 (2.53)** Board member*( . 14 years on board) 2.294 (1.45) Board member*Board voting power 0.943 (1.52) Board member*Effective voting powerb 0.406 (2.28)** Board member*Total IBRD loans 0.000256 (1.94)* Number of observations 4,061 4,061 4,061 4,061 3,673 3,673 3,673 Number of countries 135 135 135 135 134 134 134 R squared 0.08 0.08 0.08 0.08 0.06 0.06 0.06 *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Numbers in parentheses are the absolute values of robust t-statistics. All regressions include country and year �?xed effects; standard errors are clustered at the country level. a. Effective vote is Board voting power multiplied by the total votes per board member divided by the total developing country votes. b. These values are assigned to all countries in the group regardless of their board status. Source: Authors’ analysis based on data described in the text. Ashwin Kaja and Eric Werker 195 196 THE WORLD BANK ECONOMIC REVIEW leverage is gained from representing the votes of developed countries that are not interested in extra loans for themselves. This suggests that developing countries sharing a seat with donor countries are better able to use their board seat to draw more loans. Finally, speci�?cation 7 indicates that when total loan volumes are higher, board members direct even more loans to their home countries. (Since the regression included year dummy variables, this means that board members fared even better than usual, relative to nonmembers.) This result is statistically signi�?cant at the 10 percent level. Recall that total loan volume is used as an indicator of the value of a Bank loan. The rise in bene�?ts to board membership in proportion to the value of a loan serves as a causality check on the pork- barrel hypothesis. Consistent with the lack of signi�?cant �?ndings in all the IDA speci�?cations described in section IV, the differential treatment of board membership also fails to reveal any statistically signi�?cant effect for the IDA board (reported in table S3 in the online supplemental appendix). V. C O N C L U S I O N S AND IMPLICATIONS The �?ndings of this analysis point to and quantify (for the World Bank) a dilemma of global governance: when the number of decisionmakers is limited, countries that are not part of the debate may be short-changed in the distri- bution of bene�?ts. A seat on the IBRD’s Board of Executive Directors is important not just for intangible reasons such as international prestige but also for the large increase in loan commitments that executive directors bring to their home countries. A developing country serving on the board can expect, on average, a more than doubling of its normal funding levels or, in absolute terms, a nearly $60 million bonus. Furthermore, board membership, rather than omitted trends or alternative explanations, appears to drive much of this striking effect. The evidence also suggests that the returns to board membership are higher for board members in the post –Cold War era, for developing countries whose voting power on the board also represents that of developed countries, and for countries fortunate enough to be on the board when IBRD loans are most sought after. Yet the story is not simply one of rules and abuse. Comparison of the inﬂuence of executive directors and their alternates shows no signi�?cant difference in additional loans received, even though the executive director wields much more formal power. If it were simply a matter of formal insti- tutional power, alternates should not have done as well as executive directors in bringing home loans. Instead, this analysis suggests that the reward to board membership may stem more from boardroom norms and informal rules and the relationship between board members and World Bank staff than simply from voting rights. In the U.S. Congress, pork-barrel politics and logrolling are tolerated as a cost Ashwin Kaja and Eric Werker 197 of the political process. But such behavior on international appropriations com- mittees deserves more skepticism, because power there is determined by a much less structured international system. If board membership were egalitar- ian, with all countries having the same opportunity to serve on a regular basis, the �?ndings reported here might not be troubling. However, a majority of World Bank member countries never or rarely get a seat at the table. An additional warning: research in corporate �?nance has shown that �?rms with overcompensated directors and weak shareholder rights underperform (Brick, Palmon, and Wald 2006; Gompers, Ishii, and Metrick 2003). While the analysis �?nds strong results for the IBRD, it �?nds no signi�?cant association between board membership and IDA funding. This stark contrast between two institutions with similar decisionmaking structures suggests that this institutional design may not always be problematic. The difference may lie in the IDA’s exclusive focus on the world’s poorest and neediest countries or its strong norm of using external information to drive credit allocations, suggesting that governance challenges can be overcome through a less discre- tionary mandate. REFERENCES Alesina, Alberto, and David Dollar. 2000. “Who Gives Aid to Whom and Why?�? Journal of Economic Growth 5 (1): 33 –63. Andersen, T.B., H. Hansen, and T. Markussen. 2006. “U.S. Politics and World Bank IDA Lending.�? Journal of Development Studies 42 (5): 772–94. Bebchuk, L., and J. Fried. 2004. Pay without Performance: The Unful�?lled Promise of Executive Compensation. Cambridge, Mass.: Harvard University Press. Brick, I.E., O. Palmon, and J.K. Wald. 2006. “CEO Compensation, Director Compensation, and Firm Performance: Evidence of Cronyism?�? Journal of Corporate Finance 12 (3): 403–23. Carsey, T.M., and B. Rundquist. 1999. “The Reciprocal Relationship between State Defense Interest and Committee Representation in Congress.�? Public Choice 99 (3 –4): 455–63. Ferejohn, J. 1974. Pork Barrel Politics: Rivers and Harbors Legislation, 1947–1968. Palo Alto, Calif.: Stanford University Press. Gompers, P.A., J.L. Ishii, and A. Metrick. 2003. “Corporate Governance and Equity Prices.�? Quarterly Journal of Economics 118 (1): 107– 55. Heston, A., R. Summers, and B. Aten. 2006. Penn World Table Version 6.2. University of Pennsylvania, Center for International Comparisons of Production, Income, and Prices, Philadelphia, Pa. Available at http://pwt.econ.upenn.edu. IDA (International Development Association). 2004. “IDA’s Performance-Based Allocation System: IDA Rating Disclosure and Fine-tuning the Governance Factor.�? International Development Agency, Washington, D.C. Available at http://siteresources.worldbank.org/IDA/Resources/PBAFINAL.pdf. Kapur, D., J.P. Lewis, and R. Webb. 1997. The World Bank: Its First Half Century. Washington, D.C.: Brookings Institution Press. Khemani, S. 2007. “Does Delegation of Fiscal Policy to an Independent Agency Make a Difference? Evidence from Intergovernmental Transfers in India.�? Journal of Development Economics 82 (2): 464–84. 198 THE WORLD BANK ECONOMIC REVIEW Kilby, C. 2009. “The Political Economy of Conditionality: An Empirical Analysis of World Bank Loan Disbursements.�? Journal of Development Economics 89 (1): 51– 61. Kuziemko, I., and E. Werker 2006. “How Much Is a Seat on the Security Council Worth? Foreign Aid and Bribery at the United Nations.�? Journal of Political Economy 114 (5): 905–30. Malone, D.M. 2000. “Eyes on the Prize: The Quest for Nonpermanent Seats on the U.N. Security Council.�? Global Governance 6 (1): 3– 24. Marshall, M., and K. Jaggers. 2007. Polity IV Project Data Set. University of Maryland, Center for International Development and Conﬂict Management, College Park, Md. Available at www.cidcm. umd.edu/polity. Momani, B. 2007. “Another Seat at the Board: Russia’s IMF Executive Director.�? International Journal 62 (4): 916 –39. Ray, B.A. 1981. “Military Committee Membership in the House of Representatives and the Allocation of Defense Department Outlays.�? The Western Political Quarterly 34 (2): 222 –34. Rundquist, B., J. Lee, and J. Rhee. 1996. “The Distributive Politics of Cold War Defense Spending: Some State-Level Evidence.�? Legislative Studies Quarterly 21 (2): 265– 81. Rundquist, B.S., and T.M. Carsey. 2002. Congress and Defense Spending: The Distributive Politics of Military Procurement. Norman, Okla.: University of Oklahoma Press. Uppsala University and PRIO (Peace Research Institute Oslo). 2007. UCDP Database. Uppsala Conﬂict Data Program. Uppsala University, Department of Peace and Conﬂict Research, Uppsala, Sweden, and Peace Research Institute Oslo. Available at www.ucdp.uu.se/database. Woods, N. 2001. “Making the IMF and World Bank More Accountable.�? International Affairs 77 (1): 83 –100. World Bank. 2007. Articles of Agreement. Washington, D.C.: World Bank. Available at http://go. worldbank.org/BAEZH92NH0. ———. 2010. “Boards of Governors.�? http://go.worldbank.org/YM7PNV31I0. ———. Various years. Annual Report. Washington, D.C.: World Bank. Available at http://go. worldbank.org/VLWFADE5O0. Measuring Household Usage of Financial Services: Does it Matter How or Whom You Ask? Robert Cull and Kinnon Scott In recent years, the number of surveys on access to and use of �?nancial services has multiplied, but little is known about whether the data generated are comparable across countries or within the same country over time. A randomized experiment in Ghana tested whether the identity of the respondent and the inclusion of product- speci�?c cues in questions affect reported rates of use of �?nancial services. Rates of household use are almost identical whether the head reports on behalf of the house- hold or whether the rate is tabulated from a full enumeration of household members. A less complete summary of household use of �?nancial services results when randomly selected informants (nonheads of household) provide the information. For credit from formal institutions, informal sources of savings, and insurance, reported use is higher when questions are asked about speci�?c �?nancial products rather than about the respondent’s dealings with types of �?nancial institutions. In short, who is asked the questions and how the questions are asked both matter. By now, the link between �?nancial sector depth and economic growth is well established.1 Most studies rely on aggregate measures of deposits and loans in the formal �?nancial system, predominantly through banks.2 Because aggregate measures, such as the ratio of credit extended to the private sector to GDP, do Robert Cull (corresponding author, rcull@worldbank.org is a lead economist in the Development Economics Development Research Group at the World Bank. Kinnon Scott (kscott1@worldbank.org is a senior economist in the Development Economics Development Research Group at the World Bank. The authors thank the Knowledge for Change program for providing funding for this project and Mircea Tranda�?r for excellent research assistance. They thank Asli Demirgu ¨c¸ -Kunt, Pete Lanjouw, Jonathan Morduch, and Colin Xu for insightful comments. They also owe a large debt of gratitude to their collaborators on the survey team at the Ghana Statistical Service, without whom this work could not have been done. Finally, they acknowledge that this project grew out of discussions between World Bank and FinMark staff about harmonizing survey approaches, and thus the authors thank Darrell Begin, Norman Bradburn, Anne Marie Chidzero, Bob Curran, Karen Ellis, Anjali Kumar, Mark Napier, Adam Parsons, and Lorraine Ronchi for setting out the issues. 1. See Beck, Levine, and Loayza 2000; Levine 2005; Levine, Loayza, and Beck 2000; Levine and Zervos 1998; and Rajan and Zingales 1998. 2. See Beck, Demirgu ¨c¸ -Kunt, and Levine (2000) for an overview of measures of �?nancial sector depth and their construction. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 199 –233 doi:10.1093/wber/lhq004 Advance Access Publication April 14, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 199 200 THE WORLD BANK ECONOMIC REVIEW not provide information about the average size of a loan (or deposit), they give an imperfect sense of the reach of the �?nancial sector. A highly concentrated banking sector, in which a small number of relatively wealthy depositors and borrowers are responsible for a large share of banking activity, could score comparatively well on �?nancial depth while having limited breadth of outreach. There are reasons to be concerned about breadth of outreach, especially in developing countries. Informational asymmetries, transaction costs, and con- tract enforcement costs lead to market imperfections that disproportionately disadvantage the poor, who tend to lack collateral, credit histories, and connec- tions (Levine 2005; World Bank 2007). And recent studies have established a link between �?nancial sector development and poverty alleviation (Beck, Demirgu ¸ -Kunt, and Levine, 2007; Clarke, Xu, and Zou, 2006; Honohan ¨c 2004). Perhaps the major reason why �?nancial sector breadth has been understu- died is the dif�?culty collecting data.3 Whereas measures of �?nancial depth can be derived from the balance sheets of �?nancial institutions that already furnish this information to supervisors such as central banks, the same information is not readily available for �?nancial sector breadth and certainly not in a consist- ent format across countries. Recent attempts to collect data on �?nancial sector breadth have pushed beyond balance sheet information, using both demand- and supply-side approaches. On the supply side, measures of �?nancial sector outreach often focus on the number of accounts of providers of �?nancial services. For example, Beck, Demirgu ¸ -Kunt, and Martinez Peria (2007) collected information on the aggre- ¨c gate number of deposit and loan accounts from bank regulators in 99 countries. They also collected information on the number of bank branches and automated teller machines (ATMs) in each country as a proxy for physical access to �?nan- cial services, even among those who do not use them. A limitation of those data is that they are derived only from information about banks, which, while impor- tant or even dominant providers of �?nancial services in many countries, are not the full story. Honohan (2008) combines the commercial bank accounts from Beck, Demirgu ¸ -Kunt, and Martinez Peria with accounts at micro�?nance insti- ¨c tutions (from Christen, Jayadeva, and Rosenberg 2004) and at savings banks that are members of World Savings Bank Institute (from Peachey and Roe 2006) to produce the most comprehensive, though admittedly still rough, accounts- based estimates of access to date. While this represents a step forward, the accounts-based approach provides little information about the account holders and thus about �?nancial exclusion in a given country. A more satisfying, but costlier, approach is to interview users and potential users of �?nancial services through surveys of individuals and households. Broadly speaking, there are two approaches: stand-alone surveys on access to 3. See World Bank (2007) for a discussion. Robert Cull and Kinnon Scott 201 �?nancial services, which tend to be relatively expensive but produce rich data sets and a detailed portrait of access, and a small module of questions on �?nan- cial access and use that is embedded within a larger survey designed to cover another topic (such as surveys of household expenditures or labor market par- ticipation) or multiple topics (as in the Living Standards Measurement Study (LSMS) surveys). The marginal cost of the modules is much lower than that of stand-alone surveys, but they yield data that are much less rich. Neither approach has produced comparable �?nancial use data at regular intervals that could be used to monitor the situation in a given country over time or to compare outreach across many countries. Because the stand-alone surveys are costly, they tend not to be repeated at regular intervals, and when they are eventually repeated, the sampling frame and questions may differ or a different organization may conduct the survey. In surveys designed for a differ- ent purpose, modules of �?nancial questions tend not to be given high priority, and comparability of data across surveys occurs largely by chance. A recent summary of the �?nancial information generated in the LSMS shows that only a handful of basic questions about accounts and loans are asked in most modules, and those are often asked in different ways, making the validity of comparisons across surveys dubious (Gasparini and others 2004). While the accounts-based and survey-based measures of use of �?nancial ser- vices are not substitutes, recent research has found a robust statistical link between them (Beck, Demirgu ¸ -Kunt, and Martinez Peria 2007; Honohan ¨c 2008). Thus, a regression model constructed from the more readily available accounts-based information can be used to generate reasonably accurate esti- mates of the harder to collect survey-based data. Still, the �?t of these regressions is not perfect. For example, Honohan (2008) estimates that 16 percent of Ghanaians have an account, whereas the information derived from the survey described below places that �?gure at 25 percent. At best, it would appear that the estimates derived from accounts-based information could be used to monitor access between surveys of users. Scaling up data collection on use of �?nancial services to ensure accuracy and comparability across countries and over time would therefore require a survey- based approach. While there have been other stand-alone efforts to measure use, the most advanced current one is that by the FinMark Trust, which has deployed its FinScope survey (www.�?nscope.co.za) in several developing countries, primarily in Africa.4 FinScope surveys are designed to provide nationally representative information on individuals’ use of �?nancial services. The questions are similar to those that might be found in a marketing study, including detailed inquiries about speci�?c types of �?nancial products. These questions are supplemented by others on respondents’ attitudes toward �?nan- cial institutions, risk, and coping strategies in times of economic hardship, among other issues. 4. The FinScope website lists ongoing or completed surveys for 14 African countries and Pakistan. 202 THE WORLD BANK ECONOMIC REVIEW By contrast, the most comprehensive effort to use the modular approach to measure use, the LSMS, tends to ask broad, generic questions about “credit�? or “accounts�? or dealings with types of �?nancial institutions. Another impor- tant difference between the FinScope and LSMS approaches is that the LSMS �?nance modules track household use of �?nancial services, whereas FinScope randomly selects individuals from the population to provide information only on their own use. In light of these differences, a randomized experiment was devised to test whether measured use of �?nancial services is similar when respondents are asked detailed product-based questions (the FinScope approach) or are asked more generic, institution-based questions (the LSMS approach). The two approaches are found to yield similar estimates for basic products such as savings accounts with banks or other formal providers but not for others such as insurance or credit provided by banks and other institutions. These comparisons are potentially important because the expense of stand-alone surveys makes it unlikely that they will be rolled out throughout developing countries any time soon. The results in this article provide guidance on the product- and institution-based questions that yield similar estimates of use, and they suggest ways that generic, institution-based questions used in �?nance modules could be modi�?ed to produce similar estimates of use for pro- ducts such as insurance and formal credit. For household use of �?nancial services, an important consideration is whether the identity of the survey respondent affects the accuracy of the infor- mation received. The most comprehensive approach to measuring household use is a full enumeration: each member of the household reports on personal use of �?nancial services, and individual responses are then aggregated to the household level. Other approaches use an informant to provide information on the use of �?nancial services by all members of the household, typically either the head of household or a randomly selected adult. Another part of the exper- iment, therefore, tests whether the household �?nancial use information pro- vided by the household head or a randomly selected informant is as accurate as that provided by a full enumeration. Because a full enumeration is more time consuming, these results can indicate the services for which informants can provide reliable, cost-effective information. Section I describes the experimental design, and section II compares the characteristics of the sample with that of the full Ghana LSMS—only a subset of households were re-visited, though the sample was designed to be nationally representative. Sample characteristics are also compared across treatment groups. Section III reports rates of use across �?nancial products for product- and institution-based questions and household use rates provided through full enumeration and through a randomly selected informant. Section IV introduces regressions to test whether certain types of individuals and households are responsible for the under-reporting of access found for some questionnaire formats. Section V offers concluding remarks. Robert Cull and Kinnon Scott 203 I. THE DESIGN OF THE EXPERIMENT Household surveys vary across multiple dimensions as tradeoffs are made among respondents, data quality, and cost. Choice of Respondents For �?nancial (and other) surveys, an important dimension is the choice of respondents. Heads of household, however de�?ned, are often selected because they are considered knowledgeable, other elements of the survey require their personal information, and selecting one person to provide data at the house- hold level saves time and resources. LSMS surveys had traditionally collected information this way. However, a review of the surveys suggested that the head of household may not be aware of all the services used by household members and that relying on this one informant could lead to underestimation of some types of service use and of overall household use (Kochar 2000; Scott 2000). More recent LSMS surveys have moved to direct informants (full enumeration) for �?nancial information. A third option is to randomly select one adult per household, the sampling technique used in the FinScope surveys, whose goal is to achieve a probability selection of adults in the country. Logistically, it would be easiest to have this informant provide the household-level data if such data were desired. The ques- tion is whether this strategy would provide data of similar quality to that from full enumeration of adults or from the head of household. It is not clear, a priori, that every individual in the household will be equally well informed about other members’ �?nancial sector involvement. Aggregation of Questions A second key dimension on which surveys vary is the level of aggregation of questions. A short set of highly aggregated questions can reduce costs, simplify �?eldwork, and lessen the burden on respondents. The LSMS surveys have used this method, asking about �?nancial service use at an aggregate level with a greater focus on relationships with types of �?nancial service providers than on products used. There is a concern that some services might be missed using this approach, however. Research in other areas has shown that such aggregation may lead to accidental omissions or memory lapses, thus lowering reported incidence or use. Experiments in measuring household consumption have cer- tainly shown this to be the case (Joliffe 2001; Pradhan 2001; Steele 1998; STATIN 1994). The opposite approach is to ask respondents about each �?nancial service or product available. This approach, taken in the FinScope surveys, should prevent accidental omission of service use. It does, however, increase the burden of the interview, which can lead to lower data quality. It also may pre- clude multitopic surveys from addressing �?nancial service use as there simply is not space or time for so many questions. 204 THE WORLD BANK ECONOMIC REVIEW The Ghana Experiment The experiment carried out in Ghana explicitly tests the effect of changing the respondent and changing the set of questions on �?nancial service use. The Ghana Statistical Service (GSS) collaborated with the authors in developing and administering a �?nancial services survey to a subsample of households in the Ghana Living Standards Study (GLSS5) survey. To augment GSS’s own survey experience, the GSS called on experts and other sources in the country to compile a comprehensive list of �?nancial services and service providers. The GSS staff also determined the best terminology and strategies for minimizing translation problems, prepared training materials, and trained interviewers. The �?nancial services survey, by revisiting GLSS5 households, was able to take advantage of the rich data already collected from those households. This released many constraints on the experiment survey and allowed a more complex design: interviewers needed only to be trained on �?nancial questions and data collection. And more risks could be taken because the government’s national survey was in no way at risk from the work (since households were visited after the GLSS5 was �?nished). The original framework for the experiment was a three by two matrix— three types of respondents (head, randomly selected adult, and full enumer- ation) and two types of questionnaires ( product-based and institution-based). This framework was determined to be too complex for ensuring quality in the �?eldwork. A simpli�?ed, feasible design was drawn up that still allowed com- parison on the two main issues of interest: the quality of household use infor- mation provided by informants compared with a full enumeration and the quality of data obtained using a product-based questionnaire compared with an institution-based one. Physically, three different questionnaires were �?elded, with the second and third questionnaires containing more than one treatment (table 1). Households were randomly preassigned to one of three groups with each group being admi- nistered a different questionnaire. In households where one of the treatments was for a randomly selected adult to be interviewed, interviewers used Kish tables to make that selection in the �?eld.5 Only individuals ages 15 or older were surveyed. The same information was not obtained across all households. Collecting �?nancial services use information for individuals and households differs. An individual respondent can provide information about other household members only insofar as the respondent knows about their �?nancial activities. 5. Use of a Kish table enables interviewers with a sample of household addresses (in this case the 15 houses in each enumeration area) to randomly sample individuals on the doorstep by following a simple rule for selecting one household resident to interview. A list of eligible individuals at a particular address is ordered by age, and then one person is selected according to the serial number of the address. All individuals in a household have an equal chance of selection, resulting in a representative sample of all individuals in a population. Robert Cull and Kinnon Scott 205 Knowledge of one’s own use is more a matter of straight reporting and thus offered a cleaner test of whether including product cues in questions yields higher use rates than using institutional questions alone. Overlaying the product/institutional treatments on the household use experiment would have made it more dif�?cult to distinguish the effects of question format from those of the quality of the respondent’s knowledge. Once the decision was made to separate the household use and individual use experiments, issues arose about the optimal sequencing of treatments within the same visit to a household. One issue was repetition. In general, it was preferable not to have the same individual respond �?rst to institutional and then to product-based questions about their personal use of �?nancial ser- vices, for two reasons. Respondents might grow impatient at the repetition, and their answers about �?nancial use under one question format might inﬂu- ence their subsequent responses under the other format.6 The time spent inter- viewing a household also affected the design of the experiment. For households in group 1, in which all members answered the longer, product-based ques- tions, visit length was a concern. Also, multiple visits to the same household were often necessary to collect information from all members. For those reasons, institution-based questions were not added to the group 1 question- naire. This also would yield a group of responses for which the sequencing concerns described above would not be relevant. The GLSS5 households included in the �?nancial services survey were taken from the last two GLSS5 interviewing cycles, which were closest in time to the �?elding of the experiment. This was done to minimize the chances that a household might have changed signi�?cantly between surveys, since the study relies on the GLSS5 data for non�?nancial information on households and indi- viduals. The selected enumeration areas (and households) were distributed throughout the country (table 2). The instruments were piloted and revised, and the survey took place over October–December 2006. 6. The only time such repetition occurred was for households in group 2 (see table 1), in which all household members were �?rst asked about their own use of �?nancial services, using the institution-based questions. Then a member of the household was randomly selected to answer the more detailed product-based questions. Members were told that because the product-based questions were more time consuming, it made sense to have only one person answer them. While this seemed to be a natural transition, concerns remained that answering the institutional questions �?rst might inﬂuence the selected member’s product-based responses. This could be checked by comparing responses with those from group 1, which asked all household members only product-based questions. The product/ institutional comparisons are very similar whether the group 2 product-based responses from the randomly selected household members are included or not. Thus, the results for product-based use questionnaires are reported for both groups 1 and 2. The product-based information from group 3 was excluded, however, in constructing the tests because the sampling procedure within households was not random: �?rst the household head was interviewed, and then a randomly selected nonhead was interviewed. Again, however, inclusion of the group 3 observations does not greatly affect the comparisons between the product and institutional question formats. 206 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 . Questionnaire Treatments Questionnaire administered Respondent Product Institutional Head of household Group 3 (n ¼ 659 households) Answers institution-based questions about household use in �?rst section; answers product-based questions about individual use in second section. Randomly selected adult Group 2 Group 3 (n ¼ 643 households) (n ¼ 659 households) Answers product-based Answers institution-based questions about individual questions about household use. use in �?rst section; answers product-based questions about individual use in second section. All adults (15 and older) Group 1 Group 2 (n ¼ 653 households, 1,570 (n ¼ 643 households, 1,568 individuals) individuals) All household members answer All household members answer product-based questions institution-based questions about their own use. about their own use. Note: Each group represents a different questionnaire. T A B L E 2 . Enumeration Areas from the Ghana Living Standards Study 5 Used for Financial Service Survey Sample Ghana Living Standards Study 5, Ghana Living Standards Study 5, Cycle10 Cycle 11 Urban Rural Urban Rural Region enumeration area enumeration area enumeration area enumeration area Total Northern 1 5 0 6 12 Upper East 1 2 0 3 6 Upper West 0 3 0 3 6 Ashanti 3 6 3 6 18 Eastern 2 4 2 4 12 Brong Ahafo 2 4 0 6 12 Volta 0 6 0 6 12 Western 4 2 3 3 12 Central 6 0 4 2 12 Greater Accra 6 3 8 1 18 Total 25 35 20 40 120 Source: Table prepared by Ghana Statistical Service based on GSS (2006). Robert Cull and Kinnon Scott 207 II. DESCRIPTION OF THE D ATA The �?nal sample contained 1,955 households. Efforts were made to ensure that the experiment’s results could be extrapolated to the entire population and did not apply only to the households included in the survey. The sample of house- holds for the �?nancial services survey was a random subsample of the GLSS5 sample, itself a probability sample (GSS 2006). This simple random sample of enumeration areas, however, failed to take into account the larger population in urban enumeration areas that was captured in the original probability pro- portional to size sample. This shows up in variables related to location: the �?nancial services survey sample deviates slightly from the GLSS sample in being more rural, having more households engaged in agriculture, and being more likely to be located in the Coastal and Forest Zones of the country (table 3). The effects are not strong, but it should be remembered that the sample somewhat overrepresents the rural population. Sample attrition was a second potential area of bias. The �?nancial survey is essentially a panel survey. Of the 2,291 households revisited, 336 could not be reinterviewed.7 This is a problem only if there are systematic differences between the households that could and those that could not be reinterviewed. A probit model in which the dependent variable takes a value of one if the household was not reinterviewed and zero otherwise found that rural house- holds and households with older heads were less likely to be lost between rounds. The sample may therefore underrepresent more mobile households, and the attrition reinforces the slight bias toward rural households arising from the original selection of enumeration areas. These tendencies also need to be kept in mind when drawing conclusions from the data. Finally, the allocation of households to questionnaire groups could be a concern. Comparing the means of key variables across the three groups is reas- suring on this point (table 4). The only problem area might be that households in group 2 are slightly smaller than households in the other two groups, on the order of 0.25 fewer people per household. While the difference is statistically signi�?cant, it is small and unlikely to have any effect on the results. III. BASIC COMPARISONS ACROSS TR EATM E NTS For the institution-based questions asked of a household informant (either the head of household or a randomly selected adult that is not the head), seven indicators of the use of �?nancial services were calculated, listed here with the survey questions from which they are derived: 7. Fifty percent of the nonresponses were due to vacant dwellings (either permanent or temporary) and 40 percent to households that had moved. Refusals to participate represented less than 3 percent of nonresponse. 208 THE WORLD BANK ECONOMIC REVIEW (1) Banked: Some people like to keep their money in an account with a bank. Do you or any member of your household have a bank account?8 (2) Indirect access to an account: Do you or other members of your house- hold perform banking transactions using someone else’s account?9 (3) Formal nonbank savings: Now think of all the ways that you and members of your household save money. We are not talking about investing in a business or buying land, but only about where you or other household members put their money to use later. Have you or anyone in your household used an institution such as a credit union or a savings association to save money in the past 12 months? (4) Formal credit: Many people borrow money to buy things on credit. Have you or any other member of your household used an institution such as a credit union, savings association, or bank to borrow money or to buy on credit in the past 12 months? (5) Informal savings: Have you or any other household member used a Susu,10 welfare scheme, or other savings club to save money in the past 12 months? (6) Informal credit: Have you or any member of your household used a Susu, welfare scheme, or savings club to borrow money in the past 12 months? (7) Insurance: Many people insure themselves and their possessions against unexpected circumstances. Have you or any member of your household used an institution to insure yourselves (life, health) or property (house- hold goods, house, vehicle, and the like) in the past 12 months? That is, do you or anyone in the household have any long- or short-term insur- ance policies with any institution? The same indicators were calculated and the same questions were asked for the full enumeration treatments but for individual use. For example, for the 8. To help respondents distinguish banks from other �?nancial service providers, interviewers received a list of the banks operating in Ghana and a glossary of de�?nitions of �?nancial terms, including for example, “micro�?nance: small-scale loans typically given to owners of microenterprises to cover business expenses including small-scale investments, though the loan proceeds can be used for nonbusiness purposes including consumption. Liability for loan repayment can apply only to the borrower (individual-based) or to a solidarity lending group to which the borrower belongs. Under solidarity group lending, group members have strong incentives to ensure that fellow group members repay their loans. Some, but not all, micro�?nance institutions in Ghana also provide savings services to their members.�? Observations of interviews in the �?eld indicated that respondents had little trouble identifying banks. 9. “Someone else�? could be either a family member or a nonfamily member. In either case, the household would be considered “banked.�? If the response was that neither the respondent nor any member of the household had a bank account but that the respondent (or another household member) did banking transactions through someone else’s account, it could be inferred that the “someone else�? was not a family member. This occurred rarely. 10. For a small fee, Susu collectors provide a secure, informal means for Ghanaians to save and access their own money and to gain some limited access to microcredit. Robert Cull and Kinnon Scott 209 T A B L E 3 . Descriptive Statistics: Full Ghana Living Standards Study 5 Sample and Subsample Used in the Financial Services Survey Ghana Living Financial Services t-test of Standards Study 5 Survey subsample equivalence Variable full sample mean mean of means Region Coastal 29.65 33.12 3.21 (45.68) (47.07) (0.00) Forest 40.83 38.87 1.70 (49.15) (48.76) (0.09) Savannah 29.52 28.01 1.41 (45.61) (44.91) (0.16) Rural 58.35 62.35 3.46 (49.30) (48.46) (0.00) Household characteristics Female household head 27.88 28.62 0.70 (44.84) (45.21) (0.48) Head of household literate 47.79 44.35 2.94 (49.95) (49.69) (0.00) Head of household numerate 64.24 64.16 0.06 (47.93) (47.96) (0.95) Age of head of household 45.34 45.51 0.46 (15.63) (15.64) (0.65) Extended family 26.82 27.97 1.10 (44.31) (44.89) (0.27) Household size 4.20 4.22 0.24 (2.83) (2.87) (0.81) Household has agricultural worker(s) 65.10 69.50 3.96 (47.67) (46.05) (0.00) Household has self-employed worker(s) 69.59 70.72 1.06 (46.01) (45.51) (0.29) Household has employed worker(s) 23.56 23.56 0.00 (42.44) (42.45) (1.00) Individual characteristics Age 19.62 24.04 0.66 (19.56) (24.19) (0.51) Male 48.69 49.31 1.08 (49.98) (50.00) (0.28) Note: Numbers in parentheses are standard deviations for sample means and p-values for t-statistics. Source: Authors’ analysis of survey data. banked indicator, the question was simply: “Do you have a bank account?�? Reponses are aggregated across all members of the household to arrive at the measure of household use. In other words, if one member of the household reports having a bank account, then the whole household is considered banked for the full enumeration treatments. An element of subjectivity went into the crafting of these questions, and one might worry that slight tinkering with the institution-based questions could 210 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Region and Household and Individual Characteristics across Treatment Groups Means t-tests of equivalence of means Group Group Group Groups 1 Groups 2 Groups 3 Treatment group 1 2 3 and 2 and 3 and 1 Region Coastal 30.5 29.6 29.3 0.37 0.12 0.49 (46.1) (45.7) (45.5) (0.71) (0.90) (0.62) Forest 40.6 41.4 40.2 0.30 0.45 0.15 (49.1) (49.3) (49.1) (0.76) (0.65) (0.88) Savannah 28.9 29.0 30.5 0.05 0.60 0.65 (45.3) (45.4) (46.1) (0.96) (0.55) (0.51) Rural 64.4 650 65.7 0.20 0.28 0.49 (47.9) (47.7) (47.5) (0.84) (0.78) (0.63) Household characteristics Female household head 26.4 28.8 29.7 0.97 0.37 1.34 (44.1) (45.3) (45.7) (0.33) (0.71) (0.18) Head of household literate 41.8 41.4 43.9 0.15 0.88 0.74 (49.4) (49.3) (49.7) (0.88) (0.38) (0.46) Head of household numerate 63.5 60.7 62.8 1.03 0.77 0.26 (48.2) (48.9) (48.4) (0.31) (0.44) (0.80) Age of head of household 46.0 46.4 46.5 0.39 0.17 0.57 (15.5) (15.7) (15.1) (0.70) (0.87) (0.57) Extended family 29.6 28.0 30.8 0.63 1.09 0.47 (45.7) (45.0) (46.2) (0.53) (0.27) (0.64) Household size 4.50 4.24 4.59 1.63 2.15 0.56 (2.86) (2.92) (2.98) (0.10) (0.03) (0.57) Household has agricultural worker(s) 72.7 71.2 71.9 0.60 0.30 0.30 (44.6) (45.3) (45.0) (0.55) (0.77) (0.76) Household has self-employed worker(s) 73.3 75.1 74.4 0.74 0.30 0.44 (44.6) (43.3) (43.7) (0.46) (0.76) (0.66) Household has employed worker(s) 22.6 21.7 21.4 0.41 0.11 0.52 (41.9) (41.2) (41.0) (0.68) (0.91) (0.60) Individual characteristics Age 23.7 24.3 23.7 1.10 1.13 0.02 (19.6) (20.0) (19.6) (0.27) (0.26) (0.98) Male 49.4 49.1 49.2 0.19 0.04 0.15 (50.0) (50.0) (50.0) (0.84) (0.97) (0.88) Note: Numbers in parentheses are standard deviations for sample means and p-values for t-statistics. Source: Authors’ analysis of survey data. increase reported levels of use. Considerable work with GSS staff to adapt these questions to the country context and extensive piloting of the questions (both at GSS and in the �?eld) helped ensure that the questions were well under- stood by respondents. Thus con�?dence is high that these questions represent a reasonable and fair attempt to gather data on �?nancial service use in Ghana Robert Cull and Kinnon Scott 211 T A B L E 5 . Percentage of Households That Use Financial Services Formal Indirect nonbank Formal Informal Informal Survey type Banked access saving credit savings credit Insurance Sample means (percent) Head of household 26.5 6.4 3.0 3.3 19.7 4.2 11.3 (n ¼ 638) (1.7) (1.0) (0.7) (0.7) (1.6) (0.8) (1.3) Random household 10.0 3.3 1.7 1.5 17.7 4.2 10.6 member (n ¼ 480) (1.4) (0.8) (0.6) (0.5) (1.7) (0.9) (1.4) Full enumeration 25.5 5.1 2.5 1.9 17.3 4.2 7.9 (n ¼ 643) (1.7) (0.9) (0.6) (0.5) (1.5) (0.8) (1.1) t-tests of equivalence of means Head or random 7.05 2.33 1.41 1.94 0.86 0.05 0.35 household member (0.00) (0.02) (0.16) (0.05) (0.39) (0.96) (0.73) Head or full enumeration 0.40 0.99 0.54 1.61 1.15 0.03 2.04 (0.69) (0.32) (0.59) (0.11) (0.25) (0.98) (0.04) Full enumeration or 6.69 1.46 0.94 0.52 0.19 0.03 1.55 random household member (0.00) (0.14) (0.35) (0.60) (0.85) (0.98) (0.12) Note: Numbers in parentheses are standard deviations for sample means and p-values for t-statistics. Source: Authors’ analysis of survey data. through institution-based questions. Moreover, in the �?rst set of comparisons between full enumerations and informants, all respondents were asked the same questions. While use levels might be affected by the speci�?cs of those questions, any differences in the data generated by a full enumeration and that gathered from informants are much less likely to be affected. For �?ve of the seven indicators—banked, indirect access, formal nonbank savings, informal savings, and informal credit—household use rates are almost identical when the head of household is the informant and when a full enumer- ation is undertaken (table 5). For formal credit, use rates reported by the head of household are slightly higher than those from the full enumeration treat- ments, though the hypothesis that the two rates are equal to one another cannot be rejected. Overall, the head of household reports information that is very similar to that generated by full enumeration. 212 THE WORLD BANK ECONOMIC REVIEW This is good news. Interviewing the head only is much cheaper than inter- viewing all adult members of a household, an issue returned to below. However, some surveys with a �?nancial services module, such as labor force participation surveys, are designed to interview all members of a household. The results are good news in those cases, too. The information generated through the full enumeration appears to be a reasonable substitute for that gen- erated by the head of household. Because the household use rates calculated from responses to institution-based questions are comparable using either method, valid comparisons could be made across a much broader set of countries. In contrast, a randomly selected adult from the household (who is not the head) does not provide information that is comparable to that generated by the head or by full enumeration. Randomly selected informants produce use rates that are lower than those for the other two methods and signi�?cantly lower for banked, indirect access, and formal credit. This pattern suggests that the random informant has substantially less knowledge about household use of �?nancial services than does the head of household. Disparities are greatest for services provided by formal institutions. For both informal savings and infor- mal credit, the use rates produced by random informants are almost identical to those produced by the head of household or through full enumeration. This could be because many informal savings and credit arrangements involve social activities (meetings) that all household members know about. Although the head of household respondents and the full enumeration tend to yield very similar use rates, insurance is an exception. One would expect the full enumeration to provide the most complete information and thus produce the highest use levels. And yet the percentage of households that have insurance is reported at 11.3 percent when information is provided by the head of house- hold and 7.9 percent when it is collected through a full enumeration of individ- ual use. It is conceivable that the head of household has purchased insurance for other household members of which those members are not aware. Another issue, turned to in more detail below, is that the institution-based question is a poor method of collecting information on insurance use, and therefore that none of the estimates for that indicator reported in table 5 is reliable.11 Comparisons of use rates calculated from product- and institution-based questions also reveal stark differences across indicators. The product-based questions are similar to those used in FinScope surveys. For example, a respon- dent who answered yes to any of the following questions was considered banked: 11. Recall that the head of household is asked only about his or her own personal use of insurance products in the full enumerations, and thus it is possible that the full enumeration could yield a smaller average use rate than when the head responds on behalf of the household, for the reason mentioned. However, observations of �?eld training suggest that an institution-based question is simply not a good method for collecting reliable information about insurance use. Robert Cull and Kinnon Scott 213 (1) Do you currently have an ATM card? (2) Do you currently have a debit card? (3) Do you currently have a Savings Plus account?12 (4) Do you currently have a current account (checking)? (5) Do you currently have a savings account at a bank? (6) Do you currently have a PostBank account or a post of�?ce savings account? (7) Do you currently have a bank loan? (8) Do you currently have a bank overdraft facility? The questions underlying each indicator appear in the appendix. Note that there is no product-based indicator for indirect access since there was only one question on that topic and it was asked in the same way in both the product- based and institution-based questionnaires. That indicator is therefore dropped from subsequent tables. As noted, the focus is on individual use of �?nancial services so as not to con- ﬂate the effects of the method of eliciting household use information (infor- mant or full enumeration) with the effects of asking product- or institution-based questions. Again, while some degree of subjectivity entered into the selection of questions underlying the product-based indicators of �?nan- cial services use, care was taken to adapt those questions to the country context. And many questions were selected from those used in past FinScope surveys. This should therefore constitute a fair test of the importance of asking product-based questions in the sense that it well represents the most advanced surveys undertaken to date. Product- and institution-based questions produce very similar use rates for basic services, such as banked and formal saving (banks þ nonbanks; table 6). By contrast, the product-based questions yield much higher use rates than do the institution-based questions for formal credit (2.8 percent and 0.8 percent), informal savings (18.8 percent and 8.9 percent), and insurance (16.3 percent and 5.7 percent), and all of the differences are statistically signi�?cant. For these arguably more complex �?nancial services, product-related cues appear to produce a much more complete picture of use. A surprising result is that reported use of informal credit is higher for insti- tutional than for product-based questions. This is because the product-based question on informal credit was poorly designed. It explicitly mentioned Susu’s, welfare schemes, and savings clubs, which mirrors the institutional question. The single institutional question asked about the past year, while the product-based questions asked about current use and whether such services had ever been used. To be consistent across the product-based indicators, only current use should be considered. However, because the institutional question 12. This is the brand name of a specialized savings account offered by some Ghanaian banks with additional features such as limited checking. 214 THE WORLD BANK ECONOMIC REVIEW T A B L E 6 . Percentage of Individuals Who Use Financial Services, by Product and Institutional Questions Formal saving (banks þ Formal Informal Informal Survey type Banked nonbanks) credit savings credit Insurance Sample means (percent) Questions on use of products 14.3 14.2 2.8 18.8 0.7 16.3 (n ¼ 2,201) (0.7) (0.7) (0.4) (0.8) (0.2) (0.8) Questions on use of institutions 13.3 13.8 0.8 8.9 2.2 5.7 (n ¼ 1,568) (0.9) (0.9) (0.2) (0.7) (0.4) (0.6) t-tests of equivalence of means Products or institutions 0.88 0.39 4.32 8.49 3.97 10.04 (0.38) (0.70) (0.00) (0.00) (0.00) (0.00) Note: Numbers in parentheses are standard deviations for sample means and p-values for t-statistics. Source: Authors’ analysis of survey data. asks about the past year and because users switch in and out of these services regularly, the product-based question produces a lower use rate than the insti- tutional question, which is misleading. The construction of the questions on informal credit does not therefore permit meaningfully comparing product and institutional questions.13 By de�?nition, the level of individual use of �?nancial services would not be expected to exceed the level of household use. The results show that this is true for all services except for insurance (compare tables 5 and 6). For that indi- cator, the individual use rate based on product-related questions far exceeds the household use rate calculated from the institution-based question. This shows that the institutional insurance question is not a good substitute for a series of product-related questions. 13. At the same time, use rates for semiformal and other informal credit services reveal some interesting patterns. First, 4.7 percent of respondents said that they were currently using a hire purchase or installment credit plan, while an additional 8.1 percent reported that they were using credit facilities other than bank loans, credit cards, hire purchase, or installment plans. These are sizable fractions in a country where only about a quarter of households are banked. It suggests that if institutional questions had targeted the providers of such facilities, a meaningful institutional–product comparison could have been made, one that likely would have tipped in favor of product-based questions. Still, it is hard to identify the providers of such facilities for a survey respondent without also de�?ning what those facilities are. Indeed, �?eld tests showed that interviewers needed to explain some of these concepts in depth to respondents. For informal credit, therefore, it might not be possible to separate institution- and product-based descriptions suf�?ciently to construct a test. Robert Cull and Kinnon Scott 215 T A B L E 7 . Percentage of Individuals Who Use Financial Services, by Product and Institutional Questions and Respondent Type Formal saving (banks þ Formal Informal Informal Survey type Banked nonbanks) credit saving credit Insurance Sample means (percent) Household heads Questions on use of products 22.8 22.7 4.6 21.9 0.7 17.8 (n ¼ 978) (1.3) (1.3) (0.7) (1.3) (0.3) (1.2) Questions on use of institutions 23.8 24.5 1.4 12.7 2.7 7.5 (n ¼ 638) (1.7) (1.7) (0.5) (1.3) (0.6) (1.0) t-tests of equivalence of means Products or institutions 0.48 0.81 3.50 4.70 3.17 5.92 (0.63) (0.42) (0.00) (0.00) (0.00) (0.00) Sample means (percent) Nonhousehold heads Questions on use of products 7.4 7.4 1.4 16.3 0.7 15.0 (n ¼ 1,223) (0.8) (0.8) (0.3) (1.1) (0.2) (1.0) Questions on use of institutions 6.0 6.5 0.4 6.3 1.8 4.4 (n ¼ 930) (0.8) (0.8) (0.2) (0.8) (0.4) (0.7) t-tests of equivalence of means Products or institutions 1.29 0.89 2.25 7.10 2.52 8.11 (0.20) (0.37) (0.02) (0.00) (0.01) (0.00) Note: Numbers in parentheses are standard deviations for sample means and p-values for t-statistics. Source: Authors’ analysis of survey data. Nor does the problem appear to stem from the �?nancial knowledge of the respondent. One would expect the head of household to be the most �?nancially knowledgeable member of the household, but even when the head is asked about personal use of insurance products, the product-based use rate is much higher than the institution-based measure (table 7). A similar pattern holds for formal credit and informal savings, for both household heads and nonheads, and the differences between the product- and institution-based use rates are statistically signi�?cant. The evidence points to across-the-board dif�?culties for 216 THE WORLD BANK ECONOMIC REVIEW all respondents in using institution-based questions to gather information on formal credit, informal savings, and insurance. In summary, the preliminary comparisons across treatment groups indicate that the identity of the respondent and the way questions are asked affect reported use of some �?nancial services. Full enumerations of all household members produce use rates similar to those reported by the head of household, while interviewing a randomly selected nonhead produces lower levels of household use. Product-related cues appear to be important to fully understand the use of insurance, formal credit, and informal savings but do not appear necessary for more basic services such as bank accounts and formal savings. I V. R E G R E S S I O N S This section reports on tests of whether the differences across treatments described in the previous section hold up in regressions after controlling for other factors that could affect use. Some regressions are also designed to ident- ify the characteristics of the individuals and households that reported lower levels of use on institution-based questions than on product-based questions. Another set of regressions examines the household characteristics of the ran- domly selected informants who reported lower household use rates than those obtained from the head of household or the full enumeration. The hope is to identify the types of respondents who have dif�?culty with certain question formats. Household Use: Full Enumeration or Informants To describe household use of �?nancial services, the following speci�?cation was estimated in a probit regression model: Financei ¼ a þ b1 agei þ b2 rurali þ b3 sizei þ b4 dependent sharei þ b5 female headi þ b6 age of headi þ b7 education of headi þ b8 head numeratei þ b9 share in agriculturei þ b10 share employedi þ b11 share self -employedi þ b12 informant is headi þ b13 random informanti þ ei where �?nance is one of the seven indicators of household use of �?nancial ser- vices described in section III (banked, indirect access, nonbank saving, informal saving, formal credit, informal credit, and insurance). All those indicators are dummy variables equal to one if any member of household i uses that service. Four variables control for the composition and location of the household. Positive coef�?cients are expected for the average age of household members and household size because larger households with older members are more likely to have an individual who uses �?nancial services. For the same reason, households with a high dependent share are expected to use fewer �?nancial Robert Cull and Kinnon Scott 217 services. Use is expected to be lower in rural areas because �?nancial services are less available. Variables for gender, age, education, and numeracy control for characteristics of the head of household. The dummy variable indicating whether the head is female is expected to be negatively linked to use of �?nan- cial services if providers exhibit biases against women or perhaps for broader cultural reasons. Age, education, and numeracy are expected to be positively associated with use of �?nancial services. Education is controlled for using two dummy variables: one indicating whether the head attended primary school and another indicating whether the head attended upper secondary school.14 Three variables control for the employment composition of the household: share in agriculture, share employed, and share self-employed. Those who are employed are expected to have greater need for �?nancial services. Agricultural workers and the self-employed might have different needs or �?nd it more dif�?cult to obtain �?nancial services. Informant is head and random informant are dummy variables that describe the identity of the survey respondent. The informant dummy variables therefore capture the effects on reported household use rates relative to the omitted treatment category, a full enumeration of all adult household members’ individual use of �?nancial services. De�?nitions and summary statistics for the variables used in the analysis are in table 8; the correlations between variables appear in table 9. The correlations indicate that many household characteristics �?t together in predictable ways. For example, rural households tend to be larger and more focused on agricultural activities. The summary statistics and correlations are calculated for the 3,630 obser- vations that enter the regressions that summarize individual use. Very similar summary statistics and correlations are found for the 1,734 observations that enter the household use regressions. To conserve space, only the information from the larger sample of individual use is reported here. The regression results for household use of �?nancial services appear in table 10. In the regression with banked as the dependent variable (column 1), many of the control variables are signi�?cant and of the expected sign. In par- ticular, household size, age of the head of household, and attendance in upper secondary school (or beyond) are all signi�?cantly positively linked to being banked. Rural location, female headship, the share of dependents, and the share of self-employed workers are all negatively linked to being banked. The control variables do a better job of explaining variation in the banked indicator than in the other indicators, as reﬂected in both the overall �?t of the regressions and the number of signi�?cant variables. There is also a general tendency for the control variables to explain more variation in the use of services from 14. These dummy variables were chosen because they provide a reasonably large number of respondents in the lowest (no formal schooling) and highest (upper secondary school and beyond) categories. Note also that both dummy variables are equal to one for respondents that attended upper secondary school and beyond. To measure the effects of education on �?nancial usage for those respondents, the coef�?cients on both of the dummy variables must be summed. 218 THE WORLD BANK ECONOMIC REVIEW T A B L E 8 . Variable Descriptions and Summary Statistics Variable De�?nition Mean Minimum Maximum Financial use variables Banked Equals 1 if any member of the 0.140 0 1 household has an account with a bank Formal savings Equals 1 if household has 0.142 0 1 formal non-bank savings Informal savings Equals 1 if household has 0.150 0 1 informal savings Formal credit Equals 1 if household has credit 0.021 0 1 from a formal provider of �?nancial services Informal credit Equals 1 if household has credit 0.013 0 1 from informal sources Insurance Equals1 if household has 0.120 0 1 insurance product from a formal provider Household characteristics Age Average age of household 36.864 8 98 members Rural Equals 1 if rural 0.679 0 1 Household size Number of household members 5.453 1 23 Share dependents Percentage of dependents in 0.540 0 1 household Female household head Equals 1 if household head is 0.225 0 1 female Age of household head Age of household head in years 47.436 16 98 Household head attended Equals 1 if household head has 0.604 0 1 primary school attended primary school Household head attended Equals 1if household head has 0.339 0 1 upper-secondary school attended upper secondary school Household head Equals 1 if household head can 0.591 0 1 numerate do written calculations Share agricultural Percentage of agricultural 0.398 0 1 workers workers in household Share employed Percentage of employed 0.081 0 1 members of household Share self-employed Percentage of self-employed 0.246 0 1 members of household Attended primary school Equals 1 if household member 0.614 0 1 has attended primary school Attended Equals 1 if household member 0.242 0 1 upper-secondary school has attended upper-secondary school Numerate Equals 1 if household member 0.594 0 1 can do written calculations Source: Authors’ analysis of survey data. T A B L E 9 . Correlations between Variables Female Age of Formal Informal Formal Informal Household Dependent. household household Variable Banked savings savings credit credit Insurance Age Rural size share head head Banked 1 Formal savings 0.935*** 1 Informal savings 0.132*** 0.140*** 1 Formal credit 0.231*** 0.223*** 0.134*** 1 Informal credit 0.035** 0.034** 0.191*** 0.101*** 1 Insurance 0.255*** 0.227*** 0.146*** 0.143*** 0.0228 1 Age 0.144*** 0.144*** 0.033* 0.058*** 0.0054 0.060*** 1 Rural – 0.191*** – 0.186*** –0.132*** –0.016 –0.053*** – 0.121*** 0.034** 1 Household size – 0.102*** – 0.108*** –0.092*** –0.010 –0.017 – 0.034** – 0.166*** 0.224*** 1 Dependent share – 0.132*** – 0.136*** –0.092*** –0.036** –0.007 – 0.019 0.226*** 0.202*** 0.314*** 1 Female household head – 0.072*** – 0.070*** 0.047*** –0.032* 0.011 – 0.003 0.071*** –0.151*** –0.250*** 0.125*** 1 Age of household head – 0.021 – 0.033** –0.108*** –0.022 –0.016 0.020 0.458*** 0.009 0.064*** 0.391*** 0.104*** 1 Household head attended 0.190*** 0.195*** 0.145*** 0.058*** –0.008 0.145*** – 0.180*** –0.267*** –0.143*** – 0.293*** –0.107*** –0.326*** primary school Household head attended 0.274*** 0.270*** 0.107*** 0.084*** 0.002 0.186*** – 0.057*** –0.308*** –0.097*** – 0.216*** –0.122*** –0.038** upper-secondary school Household head 0.217*** 0.217*** 0.140*** 0.050*** 0.010 0.129*** – 0.147*** –0.296*** –0.150*** – 0.245*** –0.089*** –0.256*** numerate Share agricultural – 0.166*** – 0.156*** –0.095*** –0.027 –0.016 – 0.177*** 0.112*** 0.403*** –0.022 – 0.038** –0.180*** 0.070*** workers Share employed 0.245*** 0.244*** 0.105*** 0.087*** –0.007 0.094*** – 0.036** –0.339*** –0.247*** – 0.380*** –0.050*** –0.126*** Share self-employed – 0.088*** – 0.082*** –0.017 –0.044*** –0.037** – 0.060*** 0.143*** 0.0903*** –0.345*** – 0.123*** 0.047*** 0.021 Attended primary school 0.186*** 0.184*** 0.099*** 0.048*** –0.010 0.134*** – 0.336*** –0.296*** –0.159*** – 0.235*** 0.031* –0.156*** Attended 0.357*** 0.348*** 0.118*** 0.108*** 0.012 0.201*** 0.068*** –0.286*** –0.144*** – 0.226*** –0.023 –0.018 upper-secondary school Robert Cull and Kinnon Scott Numerate 0.208*** 0.207*** 0.106*** 0.0452*** 0.014 0.128*** – 0.289*** –0.319*** –0.167*** – 0.206*** 0.039** –0.122*** Employed 0.067*** 0.074*** 0.110*** 0.073*** 0.033** – 0.050*** 0.152*** 0.163*** –0.028* – 0.105*** –0.121*** –0.183*** 219 (Continued) 220 T A B L E 9 . Continued Household head has Household head some has some Household Share Some Some primary upper-secondary head agricultural Share Share primary upper-secondary education education numerate workers employed self-employed education education Numerate Employed Head has some primary 1 education Head has some 0.579*** 1 upper-secondary education THE WORLD BANK ECONOMIC REVIEW Household head numerate 0.730*** 0.573*** 1 Share agricultural workers – 0.281*** –0.338*** – 0.304*** 1 Share employed 0.210*** 0.275*** 0.204*** – 0.212*** 1 Share self-employed – 0.027 –0.110*** – 0.034** 0.412*** –0.297*** 1 Attended primary school 0.651*** 0.436*** 0.533*** – 0.303*** 0.197*** –0.037** 1 Attended upper-secondary 0.394*** 0.641*** 0.389*** – 0.268*** 0.274*** –0.080*** 0.448*** 1 school Numerate 0.511*** 0.445*** 0.691*** – 0.305*** 0.198*** –0.049*** 0.767*** 0.447*** 1 Employed – 0.040** –0.114*** – 0.060*** 0.394*** 0.060*** 0.253*** –0.153*** –0.027 – 0.157*** 1 *Signi�?cant at the 10 percent level;**signi�?cant at the 5 percent level;*** signi�?cant at the 1 percent level. Source: Authors’ analysis of survey data. T A B L E 1 0 . Household Financial Services Use Rate Regressions, Marginal Effects Indirect Formal access to nonbank Informal Expected Banked account savings savings Formal credit Informal credit Insurance Variable signa (1) (2) (3) (4) (5) (6) (7) Average age of household þ 2 0.0015 0.0005 0.0002 0.0001 0.0005* 2 0.0001 0.0004 members (0.0011) (0.0005) (0.0003) (0.0009) (0.0003) (0.0004) (0.0006) Rural 2 2 0.0734*** 0.0229** 2 0.0060 0.0461** 0.0020 2 0.0271*** 0.0187 (0.0223) (0.0116) (0.0057) (0.0216) (0.0059) (0.0102) (0.0153) Household size þ 0.0040*** 0.0020 0.0003 0.0049 0.0028*** 0.0024 0.0044 (0.0040) (0.0021) (0.0010) (0.0040) (0.0009) (0.0017) (0.0027) Share dependents 2 2 0.1324*** 2 0.0063 0.0157 0.0425 2 0.0287** 0.0135 0.0121 (0.0459) (0.0226) (0.0129) (0.0446) (0.0126) (0.0217) (0.0313) Female household head 2 2 0.1243*** 0.0042 2 0.0052 0.0566** 2 0.0092 0.0143 2 0.0028 (0.0201) (0.0131) (0.0060) (0.0256) (0.0057) (0.0125) (0.0170) Age of household head þ 0.0034*** 2 0.0002 2 0.0004 0.0009 2 0.0003 2 0.0003 0.0012* (0.0011) (0.0005) (0.0003) (0.0010) (0.0003) (0.0004) (0.0007) Household head attended þ 2 0.0520 0.0218 2 0.0035 0.0869*** 0.0064 2 0.0070 0.0011 primary school (0.0346) (0.0174) (0.0102) (0.0309) (0.0097) (0.0141) (0.0247) Household head attended þ 0.1767*** 0.0287** 0.0125* 0.0251 0.0056 2 0.0142 0.0616*** upper-secondary school (0.0258) (0.0125) (0.0068) (0.0239) (0.0065) (0.0116) (0.0176) Household head numerate þ 0.0948*** 0.0103 0.0196** 0.0329 0.0038 0.0247* 0.0494** (0.0299) (0.0159) (0.0087) (0.0286) (0.0089) (0.0121) (0.0212) Share agricultural workers 2 – 0.0552 2 0.0013 0.0073 0.0116 2 0.0068 0.0178 2 0.1015*** (0.0368) (0.0189) (0.0107) (0.0355) (0.0084) (0.0168) (0.0273) Share employed þ 0.0420 0.0482* 0.0125 0.0406 0.0157 2 0.0046 0.0275 Robert Cull and Kinnon Scott (0.0508) (0.0251) (0.0127) (0.0527) (0.0129) (0.0240) (0.0381) Share self-employed 2 2 0.1096** 0.0301 0.0146 0.0380 0.0051 2 0.0630** 0.0562* 221 (Continued ) 222 TABLE 10. Continued Indirect Formal access to nonbank Informal Expected Banked account savings savings Formal credit Informal credit Insurance Variable signa (1) (2) (3) (4) (5) (6) (7) (0.0474) (0.0234) (0.0144) (0.0452) (0.0126) (0.0238) (0.0334) Informant is head of household ? 0.0233 0.0071 0.0015 0.0197 0.0055 2 0.0004 0.0257 (0.0235) (0.0117) (0.0061) (0.0229) (0.0065) (0.0107) (0.0175) Informant is randomly selected 2 2 0.1655*** 2 0.0149 0.0052 0.0044 2 0.0043 2 0.0015 0.0259 nonhead THE WORLD BANK ECONOMIC REVIEW (0.0190) (0.0120) (0.0061) (0.0250) (0.0063) (0.0113) (0.0197) Number of observations 1734 1734 1734 1734 1734 1734 1734 Log likelihood 2 738.6819 2 329.6634 179.6205 788.4633 2 165.9691 2 287.8245 2 516.1416 Pseudo R2 0.1910 0.0607 0.1083 0.0457 0.1103 0.0491 0.0829 Chi2 head random 47.05*** 2.38 0.89 0.30 1.87 0.01 0.00 p head random 0.0000 0.1231 0.3442 0.5838 0.1712 0.9310 0.9845 *Signi�?cant at the 10 percent level;**signi�?cant at the 5 percent level;*** signi�?cant at the 1 percent level. a. Expected signs are for access to formal �?nancial services, especially for being banked. For other services, especially those that are informally pro- vided, these hypothesized relationships might not hold as well, or at all. Note: Numbers in parentheses are standard errors. Source: Authors’ analysis of survey data. Robert Cull and Kinnon Scott 223 formal providers (banked, formal nonbank saving, and formal credit) than from informal providers (informal credit and informal savings). When household characteristics are controlled for in the regressions, the comparisons across the treatment groups remain similar to those in the summary statistics in table 5. There are no signi�?cant differences between the head of household as the respondent and a full enumeration, as reﬂected in the insigni�?cant coef�?cients for the informant ¼ head variable for all indi- cators. By contrast, the tests of whether coef�?cients for the two informants (head and random) are equal, reported at the bottom of the table, reveal signi�?- cant differences for the banked indicator. There is no longer a signi�?cant differ- ence for formal credit or indirect access, as there was in the summary comparisons, but that could be because there is so little use of those services in the sample. In countries where formal credit or indirect access is more preva- lent, signi�?cant differences might emerge.15 In addition, the coef�?cient for a randomly selected informant is negative and highly signi�?cant for being banked, indicating that the random informant pro- vides less complete information on household use of banking services than does a full enumeration. In short, though the signi�?cance levels fall when con- trolling for additional factors that affect use, the same qualitative patterns emerge: the head of household and a full enumeration produce similar house- hold use rates, but a randomly selected (nonhead) informant produces lower use rates for services from formal providers. To get a better understanding of whether particular household character- istics drive the lower use rates reported by random informants, the control vari- ables are interacted with the treatment variables. That is, the explanatory variables are multiplied by head ¼ informant to derive a second set of explana- tory variables and multiplied by random informant to derive a third set. The two new sets of explanatory variables are included in the original regressions in the full-interaction speci�?cations: X 3 Financei ¼ ð@t þ bt Xi Þ þ 1i t ¼1 where t refers to the three treatment categories (full enumeration, head of household informant, and random informant) and X is the set covariates from the original regression. This allows the control variables to affect reported use in different ways across treatment categories. For the most part, the determinants of banked, indirect access, and formal credit are similar across the three treatments, as indicated by insigni�?cant 15. The signi�?cant differences for formal credit and indirect access found in table 5 might reﬂect the problem of small numbers (where small deviations seem to imply highly signi�?cant differences since so few respondents use these services). In any event, signi�?cance disappears after controlling for other relevant factors through regression. 224 THE WORLD BANK ECONOMIC REVIEW coef�?cients on the interaction variables. To conserve space, these results are not presented here. There are some exceptions for banked that are worth noting, however. For the trials that use a random informant, the share of dependents has a strong negative association with being banked. Note that for the survey, qualifying adults were all household members age 15 or older. Thus, a number of the randomly selected informants were dependents under the de�?nition used to construct the dependent share control variable. The nega- tive signi�?cant coef�?cient for dependent share reﬂects, at least in part, the dif�?- culties that young adults face in responding to institution-based questions about household use of banking services.16 Since the determinants of household use of banking services are similar whether use is reported by a random informant or calculated from a full enu- meration of individuals’ use, and since the constant is not statistically different for those two treatment categories in the full-interaction speci�?cation, it appears that younger, poorly informed household members were largely responsible for the relatively low use of banking services reported by random informants and shown in the summary statistics in table 5 and the basic regressions in table 10. The great majority of coef�?cients for heads of household are insigni�?cant, indicating that the determinants of use are similar for those treatments and the ones that used full enumeration, but again the exceptions are instructive. The �?rst is that in households where the head is numerate, use of banking services is signi�?cantly greater under full enumeration but not when the head reports on household use. This suggests that numerate heads pass on knowledge to other household members about banking services that increases others’ per- sonal use, but both numerate and innumerate heads have a reasonable grasp of household use of banking services when they are asked. A second difference is that the share of employed household members is positive and signi�?cant in the full enumeration speci�?cations, presumably because the employed have greater need of banking services, but insigni�?cant when the head reports on household use. Like the insigni�?cant result for numerate heads of household, the one for share of employed household members suggests that household heads know about the use of banking services by the employed members of their household and are able to report on it when asked the institution-based question. Product or Institutional Questions For the regressions that describe individual use of �?nancial services and compare product- and institution-based questions, the following individual 16. When a dummy variable is included indicating that the random respondent is 15– 18 years old, it is negative and signi�?cant while the dependent share variable is no longer signi�?cant. This provides additional evidence that it is younger respondents who have dif�?culty providing accurate information about household use of banking services. Robert Cull and Kinnon Scott 225 characteristics are added to the household characteristics used in the regressions in the previous section: education level of the respondent (dummy variables for attended primary and attended upper secondary school) and dummy variables indicating whether the respondent is numerate and employed. Education level, numeracy, and employment are expected to be positively linked to personal use of �?nancial services. The informant dummy variables are replaced with a dummy variable indicating whether the respondent was asked product-based questions. The coef�?cient on that variable therefore measures reported use rates relative to the omitted category, respondents who answered institution- based questions. The regressions results appear in table 11.17 The individual characteristics are almost all positive and signi�?cant for banked and formal savings (banks þ nonbanks). Employed respondents are signi�?cantly more likely to use all types of �?nancial services except for insurance, and the marginal effects for the non- insurance indicators are large relative to the average individual use rates reported in table 6. The education level of the respondent is associated with greater use of insurance, however. In all, the individual characteristics explain substantial variation in the �?nancial use indicators. That said, household characteristics also explain substantial variation in individual use. As in the household use regressions, average age in the house- hold is positively associated with the indicators. Rural location, female head- ship, and the shares of agricultural workers and self-employed workers are signi�?cantly negatively associated with the indicators.18 The overall �?t is also better in the individual use regressions than in the household use regressions, as reﬂected in the pseudo-R2 values. Most important, the dummy variable indicating whether the respondent answered product-based questions is positive and signi�?cant for informal savings, formal credit, and insurance, as was true for the summary compari- sons in table 6. The marginal effects of the product-based questions variable are also large in those regressions relative to the levels of personal use of those services reported in table 6. Moreover, the coef�?cient for the product-based question format implies a disparity between treatments similar to that implied by the simple bivariate comparisons. For example, table 6 indicates that 17. For households in group 2, all members were �?rst asked about their own use of �?nancial services using the institution-based questions, and then a member of the household was randomly selected to answer the more detailed product-based questions. In this way, the random respondent provides an observation under both question formats. Although this could have implications for the standard errors, this does not appear to be a cause for major concern: product– institutional comparisons are very similar whether the group 2 product-based responses from the randomly selected household members are included or not. Results are therefore reported for product-based questions for both groups 1 and 2. 18. The age of the head of household is negatively associated with indicators of individual use, whereas it was positively associated with household use. This is because the age of the household head competes with the average age of all household members in the individual use regressions. When one of those variables is dropped, the other is positive and signi�?cant in the regressions in table 11. 226 T A B L E 1 1 . Individual Financial Services Use Rate Regressions, Marginal Effects of Product versus Institutional Questions Expected Formal saving signa Banked (banks þ nonbanks) Informal savings Formal credit Informal credit Insurance Variable (1) (2) (3) (4) (5) (6) Average age of household members þ 0.0042*** 0.0045*** 0.0019*** 0.0005*** 2 0.00001 0.0014*** (0.0004) (0.0004) (0.0005) (0.0001) (0.0001) (0.0004) Rural 2 2 0.0386*** 2 0.0385*** 2 0.0533*** 0.0018 2 0.0128*** 2 0.0064 THE WORLD BANK ECONOMIC REVIEW (0.0124) (0.0126) (0.0148) (0.0023) (0.0049) (0.0113) Household size ? 2 0.0021 2 0.0026 2 0.0035 0.0003 – 0.0012** 0.0012 (0.0020) (0.0021) (0.0023) (0.0004) (0.0005) (0.0018) Share dependents 2 2 0.0371 2 0.0368 2 0.0305 2 0.0030 2 0.0036 0.0280 (0.0226) (0.0228) (0.0270) (0.0051) (0.0054) (0.0232) Female household head 2 2 0.0543*** 2 0.0524*** 0.0391*** 2 0.0035 2 0.0021 2 0.0168 (0.0099) (0.0102) (0.0158) (0.0024) (0.0025) (0.0114) Age of household head þ 2 0.0018*** 2 0.0022*** 2 0.0021*** 2 0.0003*** 2 0.00005 0.0002 (0.0005) (0.0005) (0.0005) (0.0001) (0.0001) (0.0004) Household head attended primary þ 2 0.0479** 2 0.0270 0.0482** 0.0039 2 0.0021 0.0449** school (0.0253) (0.0243) (0.0215) (0.0043) (0.0046) (0.0189) Household head attended þ 2 0.0137 0.0172 0.0094 0.0046 2 0.0004 0.0272* upper-secondary school (0.0174) (0.0176) (0.0184) (0.0050) (0.0039) (0.0162) Household head numerate þ 0.0343 0.0200 0.0120 2 0.0076 2 0.0007 2 0.0283 (0.0222) (0.0231) (0.0229) (0.0068) (0.0047) (0.0214) Share agricultural workers 2 2 0.0713*** 2 0.0630*** 2 0.0606*** 2 0.0015 2 0.0026 2 0.1295*** (0.0196) (0.0197) (0.0226) (0.0041) (0.0047) (0.0205) Share employed þ 0.0490* 0.0497* 2 0.0335 0.0025 2 0.0290*** 0.0284 (0.0259) (0.0262) (0.0316) (0.0053) (0.0087) (0.0276) Share self-employed 2 2 0.0754*** 2 0.0754*** 2 0.0684** 2 0.0147** 2 0.0292*** 0.0211 (0.0265) (0.0266) (0.0296) (0.0069) (0.0074) (0.0268) Respondent attended primary þ 0.0458** 0.0350*** 2 0.0081 0.0004 2 0.0076 0.0151 school (0.0215) (0.0220) (0.0233) (0.0048) (0.0059) (0.0203) Respondent attended þ 0.1109*** 0.1008*** 0.0082 0.0048 0.0020 0.0467*** upper-secondary school (0.0230) (0.0225) (0.0187) (0.0053) (0.0044) (0.0175) Respondent numerate þ 0.03700 0.0488** 0.0241 0.0057 0.0046 0.0219 (0.0223) (0.0223) (0.0230) (0.0051) (0.0045) (0.0203) Respondent employed þ 0.0509*** 0.0527*** 0.0991*** 0.0122*** 0.0093*** 2 0.0046 (0.0106) (0.0107) (0.0114) (0.0027) (0.0022) (0.0126) Product þ 0.0122 0.0067 0.0913*** 0.0100*** 2 0.0112*** 0.1019*** (0.0095) (0.0096) (0.0106) (0.0028) (0.0022) (0.0091) Number of observations 3630 3630 3630 3630 3630 3630 Log likelihood 2 1126.0058 2 1145.0915 2 1371.3258 2 303.0288 2 227.5562 2 1152.5292 Pseudo R2 0.2357 0.2294 0.1068 0.1702 0.1235 0.1365 *Signi�?cant at the 10 percent level;**signi�?cant at the 5 percent level;*** signi�?cant at the 1 percent level. a. Expected signs are for access to formal �?nancial services, especially for being banked. For other services, especially those that are informally pro- vided, these hypothesized relationships might not hold as well, or at all. Note. Numbers in parentheses are standard errors. Source: Authors’ analysis of survey data. Robert Cull and Kinnon Scott 227 228 THE WORLD BANK ECONOMIC REVIEW product-based questions produce an insurance use rate 10.6 percentage points higher than institution-based questions, while the product-based coef�?cient from the regression indicates a 10.2 percentage point difference between treat- ments.19 The regression results therefore reinforce the conclusion that product- based cues help respondents provide a more complete picture of their use of those three �?nancial services. For banked and formal savings (banks þ non- banks), the product-based questions dummy variable is insigni�?cant, indicating again that product-based cues are less important for those services. For informal credit, the product-based dummy variable is negative and signi�?cant, but again the way that product-based question was constructed led to a test that is not very meaningful. To better identify the types of individuals who bene�?t most from product-related cues, the explanatory variables in table 11 were interacted with the dummy variable for product-based questions for the three services for which a signi�?cant increase in use rates was found for product-based questions compared with institution-based questions. Almost all the coef�?cients on the interaction terms are insigni�?cant, indicating that the determinants of reported use are similar for the two question formats and suggesting that all respondents bene�?t from product-related cues regarding formal credit, informal savings, and insurance. Therefore, those results are not presented. These �?ndings reinforce the conclusions drawn from the simple sample breakdown in table 7. Controlling for Supply-Side Effects To ensure that supply-side effects—the presence of providers of �?nancial services—are not driving the differences in use across treatments that were reported above, additional regressions were run to control for travel time (in minutes) to the nearest bank as a measure of the local availability of �?nancial services and others that include regional or local dummy variables to capture these effects. As in the base regressions, household use is very similar for full enumeration and when the head of household is the informant, and reported use of banking services is signi�?cantly lower for the random informant. Individual use rates are signi�?cantly higher for product-based questions for informal savings, formal credit, and insurance.20 In short, it seems unlikely that the omission of supply-side variables from the base regressions could be driving the results. 19. The same is true of the household use regressions. For example, the regressions imply that a random respondent is 16.6 percentage points less likely to report that someone in the household is banked than is revealed through a full enumeration. The difference between those two treatments in the bivariate comparison is 15.5 percentage points. 20. The working paper version of this article provides more details about these regressions, including the results (Cull and Scott 2009). Robert Cull and Kinnon Scott 229 V. C O N C L U S I O N S AND I M P L I CAT I O N S FOR FUTURE SURVEYS Measuring the breadth of outreach of �?nancial sectors in developing countries remains a challenge, but one that must be met to better understand how �?nan- cial services (or their absence) affect the livelihoods of the poor. Surveys of individuals and households about their use of �?nancial services hold the most promise for measuring outreach well, but their cost and other logistical hurdles have made it dif�?cult to develop a standard method of questioning that would generate comparable �?nancial use data across countries and within countries over time. This experimental analysis was designed to contribute to an under- standing of the comparability of �?nancial use data generated under different question formats. The main �?ndings are straightforward, intuitive, and should be useful for future data gathering efforts. Rates of household use of �?nancial services are similar when the head reports on behalf of the household and when the rate is tabulated from a full enumeration of individual use. By contrast, randomly selected informants provide a less complete picture of household use of �?nan- cial services than do the other two methods. The comparability of data for the head of household and the full enumeration is potentially important because interviewing only the head is much less costly than interviewing all household members. At the same time, some surveys, such as those measuring labor force participation, are designed to be full enumerations. Using the head of house- hold when possible and a full enumeration when dictated for other reasons should increase the number of countries for which comparable data can be generated. For formal credit, informal savings, and insurance, reported use is higher when questions are asked about speci�?c �?nancial products rather than about the respondents’ dealings with types of �?nancial institutions. Product-related cues for these services appear to be important for all respondents, not just those who might be expected to be less knowledgeable about �?nancial matters. The results are therefore similar to those from experiments on measuring household consumption, where inclusion of a longer list of items leads to higher reported consumption than does a shorter list of broader consumption categories. If treatments that yield higher use rates are presumed to be a more accurate depiction of reality (as seems likely), then using techniques such as random respondents and institution-based questions, though certainly less costly, will capture only a fraction of the use of some �?nancial services (only half for some services in Ghana). This could make it dif�?cult to design appropriate policies to foster �?nancial inclusion and to measure their effectiveness. Although product-based and institution-based use were tested only for per- sonal use of �?nancial services, it seems likely that product-related cues would also bene�?t respondents informing about household use of those services. That implies adapting the institution-based questions used in the �?nancial modules 230 THE WORLD BANK ECONOMIC REVIEW T A B L E 1 2 . Time Costs of Administering Financial Services Survey (minutes) Questionnaire type Services and response 1 2 3 Total Banked No 41.52 32.15 35.92 36.50 (22.26) (14.21) (15.61) (18.08) [n ¼ 468] [n ¼ 475] [n ¼ 479] [n ¼ 1,422] Yes 52.75 36.09 42.76 44.32 (34.89) (15.97) (20.50) (26.45) [n ¼ 187] [n ¼ 158] [n ¼ 178] [n ¼ 523] Formal credit No 44.12 33.15 37.50 38.23 (26.30) (14.79) (17.27) (20.51) [n ¼ 613] [n ¼ 621] [n ¼ 634] [n ¼ 1,868] Yes 53.50 32.08 45.22 47.69 (34.20) (13.51) (17.68) (28.41) [n ¼ 42] [n ¼ 12] [n ¼ 23] [n ¼ 77] Insurance No 41.23 32.95 37.02 36.85 (22.46) (14.74) (17.48) (18.57) [n ¼ 495] [n ¼ 582] [n ¼ 579] [n ¼ 1,656] Yes 55.55 35.18 43.32 48.65 (35.51) (14.99) (15.09) (29.37) [n ¼ 160] [n ¼ 51] [n ¼ 78] [n ¼ 289] Total 44.73 33.13 37.77 38.60 (26.94) (14.76) (17.32) (20.95) [n ¼ 655] [n ¼ 633] [n ¼ 657] [n ¼ 1,945] Notes: Numbers in parentheses are standard errors. Source: Authors’ analysis of survey data. of larger, multipurpose surveys to include product-based cues that are appro- priate to the country context. Decisions on future questionnaires will also need to consider the relative costs (in interview time) of implementing the different treatments, which conform to expectations.21 The full enumeration using the product list takes the longest to administer. But full enumeration itself, using either the product- based or institution-based questionnaire, adds considerable time to the inter- views compared with use of a proxy respondent for the household (see results for group 1 in table 12). In other words, the �?nding that the head of household is able to provide similar data to that obtained from full enumeration for most products has positive implications for the feasibility of expanding data 21. The time data collected in this survey are, at best, rough approximations of the actual time required. No effort was made to record time at the level of the speci�?c product or institution modules. Only a total for the entire household interview, which includes a roster and further questions on attitudes and knowledge of �?nance, is available. Also, as groups 2 and 3 contain two different treatments, it is not possible to separate the time costs associated with each one. Robert Cull and Kinnon Scott 231 collection on �?nancial service use to other countries. Finally, for survey designers in countries that may have higher levels of �?nancial service use, it is important to note how much average interview time rises when household use of �?nancial services is higher. For example, the full enumeration product-based format in questionnaire 1 took 20 –30 percent more time to administer when members of the household used banking or insurance services than when they did not (see table 12). Discernible throughout this article is a concern about the ability to general- ize beyond Ghana. While there is a strong undercurrent of common sense to the main �?ndings, which are thus likely to be relevant in other countries as well, the article is speci�?cally about Ghana. And while Ghana might be an ade- quate reﬂection of low-income countries in much of Sub-Saharan Africa, it is unlikely to be reﬂective of the whole developing world. The best that can be done in the context of this article is simply to acknowledge this limitation. In future, however, this type of experiment can be repeated in other countries. A similar study in Timor Leste, where very few respondents use any �?nancial ser- vices, found that the differences across treatments were not signi�?cant, suggesting that the concerns raised in this analysis are of second order impor- tance in the most �?nancially underdeveloped countries. We live in a world of rough approximation when it comes to measuring the outreach of the �?nancial systems of developing countries. The reliability of esti- mates from accounts-based approaches and approaches that meld accounts- based and survey-based information through regressions is dif�?cult to assess. The hope is that the results here provide some practical guidance on how to generate comparable �?nancial use data across countries through surveys, which appear to represent the best vehicle for generating accurate data. APPENDIX. CONSTRUCTION O F I N D I CATO R S FROM PRODUCT-LEVEL QUESTIONS Banked: Q2 ATM card Q4 Debit card Q6 Savings Plus account Q8 Current account Q10 Savings account at bank Q12 PostBank account, post of�?ce savings account Q36 Bank loan Q54 Bank overdraft facility Indirect: Q16 Use of someone else’s account 232 THE WORLD BANK ECONOMIC REVIEW Formal savings: Q6 Savings Plus account Q10 Savings account at bank Q12 PostBank account, post of�?ce savings account Q14 CDs, treasury bills, notes, money market funds Q22 Savings with regulated micro�?nance institution Q24 Savings with credit union Q30 Shares, investment funds Q32 Provident fund Q34 Pensions fund Informal savings: Q26 Susu scheme Q28 Welfare scheme, other savings club (e.g., with religious organization). Formal credit: Q36 Bank loan Q38 Loan from government Q40 Loan from credit union Q42 Loan from micro�?nance institution Q44 Loan from employer Informal credit: Q48 Welfare scheme, Susu, savings club Insurance: Q60 Vehicle Q62 Property Q64 Homeowners Q66 Debts Q68 Travel Q70 Life Q72 Debts if you die Q74 Disability from employer Q76 Other disability Q78 Professional Q80 Funeral policy with institution Q84 Health/medical Q86 Children’s education REFERENCES ¨c Beck, Thorsten, Asli Demirgu ¸ -Kunt, and Ross Levine. 2000. “A New Database on the Structure and Development of the Financial Sector.�? World Bank Economic Review 14 (3): 597– 605. ———. 2007. “Finance, Inequality, and the Poor.�? Journal of Economic Growth 12(1): 27– 49. Beck, Thorsten, Ross Levine, and Norman Loayza. 2000. “Finance and the Sources of Growth.�? Journal of Financial Economics 58 (1 –2): 261–300. Robert Cull and Kinnon Scott 233 ¨c Beck, Thorsten, Asli Demirgu ¸ -Kunt, and Maria Soledad Martinez Peria. 2007. “Reaching Out: Access to and Use of Banking Services across Countries.�? Journal of Financial Economics 85(1): 234–66. Clarke, George R.G., L. Colin Xu, and Heng-Fu Zou. 2006. “Finance and Income Inequality: What Do the Data Tell Us?�? Southern Economic Journal 72(3): 578–96. Cull, Robert, and Kinnon Scott. 2009. “Measuring Household Usage of Financial Services : Does It Matter How or Whom You Ask?�? Policy Research Working Paper 5048. World Bank, Washington, D.C. Gasparini, Leonardo, Federico Gutierrez, A Tamola, L. Tornarolli, and Guido Porto. 2004. “Finance and Credit Variables in Household Surveys of Developing Countries.�? Centro de Estudios Distributivos, Laborales, y Sociales, Universidad Nacional de La Plata, Argentina, and World Bank Development Economics Research Group, Washington, D.C. GSS (Ghana Statistical Service). 2006. “Sample Design for Round Five of the Ghana Living Standards Survey (GLSS 5).�? Accra. Honohan, Patrick. 2004. “Financial Development, Growth, and Poverty: How Close Are the Links?�? In Financial Development and Economic Growth: Explaining the Links, ed. C. Goodhart. London: Palgrave. ———. 2008. “Cross-Country Variation in Household Access to Financial Services.�? Journal of Banking and Finance 32(11): 2493– 500. Jolliffe, Dean. 2001. “Measuring Absolute and Relative Poverty: The Sensitivity of Estimated Household Consumption to Survey Design.�? Journal of Economic and Social Measurement 27(1): 1–23. Kochar, Anjali. 2000. “Savings.�? In Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Measurement Study, ed. Margaret Grosh, and Paul Glewwe. Washington, D.C.: World Bank. Christen, Robert P., Veena Jayadeva, and Richard Rosenberg. 2004. “Financial Institutions with a Double Bottom Line: Implications for the Future of Micro�?nance.�? CGAP Occasional Paper 8. Consultative Group to Assist the Poor, Washington, D.C. Levine, Ross. 2005. “Finance and Growth: Theory and Evidence.�? In Handbook of Economic Growth, ed. Philippe Aghion, and Steven Durlauf. Amsterdam: Elsevier Science. Levine, Ross, Norman Loayza, and Thorsten Beck. 2000. “Financial Intermediation and Growth: Causality and Causes.�? Journal of Monetary Economics 46 (1): 31– 77. Levine, Ross, and Sara Zervos. 1998. “Stock Markets, Banks, and Economic Growth.�? American Economic Review 88 (3): 537– 58. Peachey, Stephen, and Allan Roe. 2006. “Access to Finance: Measuring the Contributions of Savings Banks.�? Working Paper. World Savings Banks Institute, Brussels. Pradhan, Menno. 2001. “Welfare Analysis with a Proxy Consumption Measure: Evidence from a Repeated Experiment In Indonesia.�? Cornell Food and Nutrition Policy Program Working Paper 126. Cornell University, Ithaca, N.Y. Rajan, Raghuram G., and Luigi Zingales.1998. “Financial Dependence and Growth.�? American Economic Review 88 (3): 559– 86. Scott, Kinnon. 2000. “Credit.�? In Designing Household Survey Questionnaires for Developing Countries: Lessons from 15 Years of the Living Standards Measurement Study, ed. Margaret Grosh, and Paul Glewwe. Washington, D.C.: World Bank. STATIN (Statistical Institute and Planning Institute of Jamaica). 1994. “Survey of Living Conditions, 1994.�? Kingston. Steele, Diane, 1998. “Ecuador Consumption Items.�? World Bank, Development Economics Research Group, Washington, D.C. World Bank. 2007. Finance for All? Policies and Pitfalls in Expanding Access. Policy Research Report. Washington, D.C.: World Bank. Banking on Politics: When Former High-ranking Politicians Become Bank Directors ´as Braun and Claudio Raddatz Matı New data are presented for a large number of countries on how frequently former high-ranking politicians become bank directors. Politician-banker connections at this level are relatively rare, but their frequency is robustly correlated with many important characteristics of banks and institutions. At the micro level, banks that are politically connected are larger and more pro�?table than other banks, despite being less lever- aged and having less risk. At the country level, this connectedness is strongly nega- tively related to economic development. Controlling for this, the analysis �?nds that the phenomenon is more prevalent where institutions are weaker and governments more powerful but less accountable. Bank regulation tends to be more pro-banker and the banking system less developed where connectedness is higher. A benign, public- interest view is hard to reconcile with these patterns. Banking sector development, institutions, political economy. JEL codes: G15, G21, P16 There is ample evidence that access to external �?nancing is critical for the level and ef�?ciency of investment, productivity, and economic growth at the �?rm and the aggregate level. Yet �?rms in different countries do not have the same access to �?nance.1 This raises two important questions: Why do some countries lack a well developed �?nancial system if it is so bene�?cial? And how do �?rms react to �?nancial sector underdevelopment? A recent strand of �?nancial devel- opment literature aims at answering both questions from a political economy standpoint. On the �?rst question, the literature complements theories of �?nancial devel- opment based on stable and largely predetermined factors (such as the origins of a country’s legal system, pattern of colonization, religion and culture, and social capital endowment) with a role for dynamic political economy Matı´as Braun (matias.braun@uai.c) is director of strategy and partner at IM Trust and professor of economics and �?nance at Universidad Iban ˜ ez. Claudio Raddatz (corresponding author; craddatz@ worldbank.org) is a senior economist at the World Bank in the Macroeconomics and Growth Unit of the Development Economics Research Group. Braun gratefully acknowledges �?nancial support from Fondecyt Chile [grant number1060015.] 1. See Levine (2006) for an extensive review of the literature on the subject. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 234– 279 doi:10.1093/wber/lhq007 Advance Access Publication June 23, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 234 ´as Braun and Claudio Raddatz Matı 235 considerations (LaPorta and others 1997, 1998; Acemoglu and Johnson 2005, Stulz and Williamson 2003, and Guiso, Sapienza, and Zingales 2004). Private interests and politics appear to be relevant determinants of �?nancial develop- ment, as suggested, for instance, by Rajan and Zingales (2003), Pagano and Volpin (2001, 2005), and Braun and Raddatz (2007, 2008). One channel through which this could occur is the regulatory effect of the interaction between politicians and �?nancial sector �?rms.2 That regulators come from or end up in the regulated industry—the revolving door phenomenon—has long been recognized as a potential determinant of regulation.3 And indeed, the empirical work, although still scarce, points to its having large social costs (see Kwhaja and Mian 2005 and Dal Bo ´ and Rossi 2006.) As for how �?rms react to �?nancial sector underdevelopment, several recent papers have documented that politically connected �?rms seem to get preferen- tial access to credit (Cull and Xu 2005; Khwaja and Mian 2005) and better treatment by the government. These links between politics and business seem quite widespread (Faccio 2006) and seem to add considerable value to �?rms (Fisman 2001). This article focuses on banks. Because of their critical role in allocating credit, the behavior of banks, unlike that of most other types of �?rms, affects the entire economy. A new dataset linking more than 10,000 politicians (cabinet members, �?nancial sector regulators, and central bank governors) and some 60,000 members of bank boards in a large number of countries is used to compare the names of bankers and of politicians to search for matches. The frequency of these matches is then used to compute measures of the connection between politicians and bankers to explore the role of political connectedness. Banks, like any other �?rm, may use these connections to improve their pos- ition, perhaps by affecting banking regulation. This would be more likely to happen where institutions are weak and the government is relatively powerful yet less accountable. It may also carry large social costs, through more restricted access to credit. The article examines the extent to which banks are politically connected, where this connectedness is more prevalent, and whether it is associated with better outcomes for banks. This private-interest view of the presence of former politicians on bank boards is, of course, not the only possibility. Links between politicians and bankers may be a way of fruitfully sharing ability, knowledge, and experi- ence between the public and private sectors. These links could imply better outcomes for the �?rm without negative social effects. Banks could simply be lobbying to make a legitimate case to government of�?cials or could consider these links more as consumption than as investment (see, for instance, Ansolabehere, de Figueiredo, and Snyder 2003). The merit of these two 2. Financial sector incumbents are de�?ned as the people and �?rms that form part of the �?nancial sector at a given time, as opposed to those interested in entering the sector. 3. See Dal Bo´ (2006) for a review of regulatory capture. 236 THE WORLD BANK ECONOMIC REVIEW perspectives is ultimately an empirical question. In that sense, the stylized facts provided in this article may shed some light on which interpretation is more likely. Several stylized facts stand out. At the micro level, politically connected banks are different from unconnected banks: they are larger, more pro�?table, less leveraged, and less risky. When aggregated at the country level in various ways, bank connectedness is found to be strongly negatively related to GDP per capita. After controlling for this and for other traditional elements, countries where banks are more connected are shown to rank higher on cor- ruption and government regulatory power and lower on accountability. Overall regulation is less market friendly, bank regulation is generally more pro-banker, and the �?nancial system is less developed. This article is closely related to the recent literature showing that politically connected �?rms appear to fare better than the rest (see, for example, Faccio 2006; and Faccio, Masulis, and McConnell 2005.) This article adds to this work in three main ways. It focuses on banks, an important contribution because of the likely effect bank connectedness may have on the entire economy through credit allocation. Rather than determining whether political connections improve outcomes for the connected �?rms, it delves deeper into the country characteristics and policy choices associated with these kinds of connections. And it looks at former politicians as well as incumbents. The article is also related to the literature on the search for political experi- ence by boards of directors (see, for instance, Agrawal and Knoeber 2001 and Goldman and others 2009). Similarly, it is related to recent work on the relationship between connections and development, including banking sector development from a historical perspective (Haber 1991; Maurer 2002; Maurer and Gomberg 2005; Milanovic, Hoff, and Horowitz forthcoming; and Razo forthcoming). In this article, the assembly of the new dataset has allowed con- sistent exploration of the issue across a large number of countries. The article compares politically connected banks to banks that are not con- nected and correlates several country-level measures of connectedness with variables capturing the quality of institutions, bank regulation, and �?nancial development. Section I describes the data and the matching procedure used to identify banker-politicians. It also discusses ways of aggregating the results into a country-level connectedness variable. Section II shows how connected banks differ from unconnected ones and explores the characteristics of countries where the phenomenon is more frequent. Section III presents conclusions and implications. I. MEASURING THE CONNECTION BETWEEN BANKERS AND POLITICIANS This section describes the methodology used to measure the connection between bankers and politicians, presents summary statistics from the resulting ´as Braun and Claudio Raddatz Matı 237 dataset, and introduces aggregate measures of the degree of connection across countries. Building the Data The data on names of politicians came from the Country Reports of the Economist Intelligence Unit, which were revised twice yearly for each country for 1996–2005. This review yielded 72,769 names of cabinet members and central bank governors. These names were complemented by a smaller set of 593 names of �?nancial sector supervisors obtained from the 2000, 2002, 2003, and 2004 editions of How Countries Supervise their Banks, Insurers, and Securities Markets (Central Bank Publications, various years). These two data sets together provide extensive coverage for cabinet members and �?nancial sector supervisors in 154 countries over 10 years (see supplemental appendix table S1, column 3, available at http://wber.oxfordjournals.org/). Once cleaned (as explained below), the data yielded an average of 72 politicians in each country, which is around 7 a year. There is some variation across countries, but it is small: 40 –100 names of politicians were found for 70 percent of the countries. The names of bank board members are from Bankscope (Bureau van Dijk 2006), which has data on the most recent board composition of both listed and unlisted banks in nearly all countries. The data were collected for 2006, so the board composition is typically from December 2005. Once duplicates were identi�?ed among the 109,645 board member names found for 4,618 banks, 64,169 unique board member names remained. Although Bankscope is the most comprehensive source of bank data around the world, its coverage is not necessarily complete. It is close to universal, however, as evidenced by the cor- respondence between the average number of banks with board composition data in Bankscope in 2001 and the total number of commercial banks reported by Barth, Caprio, and Levine (2003) for the same year (see column 5 in sup- plemental appendix table S1). Although there is some variation across countries, the difference between the number of banks in the two datasets falls within a 20 percent range in about 70 percent of countries. The banks for which board data are available account, on average, for 72 percent of the assets in each country; in only about a fourth of the countries is the fraction below 60 percent (see column 6). Because data on bank directors are from the December 2005 issue of Bankscope and data on politicians cover the period 1996–2005, matches between the two datasets typically consist of former politicians who later sit on a bank board. This is the convention followed in the rest of the article, which refers to these individuals as “former politicians.�? There are a few caveats with this terminology. First, the entire history of each individual is unknown. Thus, some of them may have been bankers before 1996 (the �?rst observation of the politician dataset). Second, how long a director has been sitting on the board is 238 THE WORLD BANK ECONOMIC REVIEW also unknown. For instance, the data do not show whether a politician who is on a board in 2005 and whose term in government ended in 2004 was already sitting on the board in 2003. Third, matches between politicians who are in their political positions in 2005 correspond to cases where the politicians sim- ultaneously hold both positions. Fourth, a given issue of Bankscope reports the latest director data available. In more than two-thirds of cases, this corresponds to December 2005, but in a few cases the data are from earlier years. Thus, to be more precise, “former politicians�? refer to individuals who were politicians at some point during 1996–2005 and who were on a bank board in December 2005. Finding matches between politicians’ and bankers’ names involved four steps. First, the strings containing the names were standardized by converting them to lowercase and removing punctuations and titles (Sir, PhD, and so on). Second, duplicate entries were removed by identifying observations that were simply different spellings of the same name (for instance, with and without the middle initial). Third, the datasets containing names of politicians were pooled and duplicate observations across the datasets were identi�?ed. Once the names had been cleaned in this way, the names in the politician and banker datasets were compared to obtain the matching observations. At each step, a record-linkage algorithm was used to �?nd matching names. The algorithm forms all possible pairs of names within each country and ranks the pairs on three standard measures of string similarity from the record-linkage literature: bigram, Levenshtein, and longest common subse- quence.4 The bigram metric counts the number of consecutive matching pairs of characters between two strings. The Levenshtein measure counts the minimal number of edits required to convert one string into the other. Allowable edit operations are the deletion of a single character, the insertion of a single character, and the substitution of one character for another. The longest common subsequence counts the number of consecutive characters that are present in two strings, and keeps the largest number. All three methods are based on the way names are written. If the difference between the way a name sounds and the way it is written varies across countries, so that mistakes are more prevalent in some countries than in others, these methods could be differentially effective and could potentially induce bias. For these reasons, the algorithms were used only to restrict the sample of potential matches, as described below. Ultimately, the matches were visually identi�?ed. When two strings containing names are compared, each of these criteria results in a value between 0 and 1 that measures the likeliness of the two names being the same. All pairs with a minimum value of 0.8 in at least one of the three methods were retained and visually checked to determine 4. The record linkage software used was Merge Toolbox, a Java-based tool created by the members of the Safelink project (see Schnell, Bachteler, and Bender 2004). ´as Braun and Claudio Raddatz Matı 239 whether they matched. While alternative ways could have been used to restrict the set of pairs to be visually checked, this relatively restrictive way was chosen so as to err more on the side of failing to �?nd true matches than of falsely identifying matches. This was also the basic principle used for the visual veri�?cation. After step two, the data contained 10,829 politicians and 62,981 bankers in 146 countries. Step three yielded 218 matching names across these two lists (see column 4 in table S1). The mean number of matches per country was 1.4, and the median was 1.0. At 0.34 percent, the share of bankers who are politicians is quite small and unimpressive. The dearth of matches reﬂects in part the restrictive way that the matches were identi�?ed. On the other hand, the fraction of politician-bankers does not seem as small in the context of the size of the populations from where they were drawn (see below). Having high-ranking politicians on the board of banks is not the only way banks can be politically connected. Non-cabinet level politicians can also play an important role connecting banks. And there are more subtle forms of con- nection: a politician can be connected to a bank by having relatives or associ- ates on the board (Faccio 2006) or by supporting the appointment of directors or chief executive of�?cers. There are also less subtle ways, such as outright bribery and corruption. Politicians sitting on bank boards seem to be a rela- tively rare form of connection compared with some other channels, to judge by country case studies and anecdotal evidence.5 However, these other types of connections are much more dif�?cult to document systematically across countries. Rather than arguing that a direct presence on the board is the only or the most important way politicians and bankers relate, the article considers the presence of high-level politicians on bank boards as a proxy for the general connection between politicians and bankers. As long as people do not completely specialize in one particular form of connection, the different ways of connecting are likely to be positively correlated. Since the analysis here looks only at the top posts in both politics and banking, the results are likely just the tip of the iceberg. Instead of focusing on absolute magnitudes, the article looks at how vari- ations in the importance of politicians sitting on bank boards links to several bank and country characteristics. There are two sources of variation in the data: variation between countries with matches and those without (the exten- sive margin) and variation in the number of matches for the countries with at least one match (the intensive margin). The 72 of 154 countries for which no matches were found were dropped from the sample for most of the analyses, for several reasons. Most important is concern about the reliability of the data for many countries with zero 5. See, for instance, Fisman (2001) for an account of Suharto’s Indonesia. 240 THE WORLD BANK ECONOMIC REVIEW matches. For instance, while 60 percent of countries with some matches meet the International Monetary Fund’s Special Data Dissemination Standard (IMF 2009), only 20 percent on those with zero matches do (many of these countries are not generally included in systematic cross-country analyses).6 Second, many countries with zero matches have very few banks. A third had fewer than three banks in Bankscope in 2005, compared with just 4 percent among countries with matches. And the median number of banks with data is 5 in the no-matches group but 16 in the group with at least one match. Third, the zeroes give little information on whether the selection of bankers is biased toward former politicians. Under reasonable assumptions, the probability of �?nding zero matches between bankers and politicians is high even if banker selection is seriously biased toward picking politicians.7 In contrast, �?nding even one match provides considerable information on the likely bias of the selection, since a match is typically a low probability event under the null hypothesis of unbiased matching. Nevertheless, results are also presented for analyses that include countries with zero matches but more than two banks (as an arbitrary cutoff for considering the zero as reliable), and many of the corre- lations documented below remain unaffected. Of course, this argument could be stretched to restrict the sample to countries with more than one or two matches because �?nding a small number of matches may simply occur by chance, something that is less likely if a more substantial number of matches are found. Although �?nding a single match is a very low probability event that is unlikely to occur by chance in most countries, the article returns to this issue below to show that, even though the sample size drops quite rapidly, the results are not very different when the sample is further restricted. Measuring Connectedness at the Aggregate Level There are several ways to aggregate the information on individual matches to measure and compare the connectedness between banking and politics in different countries. Each method has pros and cons and is more or less appro- priate under different assumptions about the process that generates the matches between politicians and bankers. Instead of focusing on a single measure, the analysis is conducted with �?ve different metrics (table 1). Three measures are straightforward, and two are more elaborate because they address some short- comings of the other three. The �?ve measures are computed twice: once for matches found for all Bankscope banks ( public, private, and mixed), and once only for matches for fully private banks. 6. For instance, 63 percent of these countries were not included in the cross-country analysis of bank regulation by Barth, Caprio, and Levine (2003). Some were included in later rounds of the survey, but coverage is incomplete. 7. See the appendix for a description of the distribution of matches under an unbiased selection process. T A B L E 1 . Measures of the Degree of Connectedness across Countries between Banks and Politicians All Banks Fully Private Banks Only World Bank Country FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE Country Code (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Gabon GAB 100 100 10 7.75 0.1 - - - - - Georgia GEO 100 100 8 8.48 0.08 100 100 8 8.48 0.08 Myanmar MMR 100 - 19 11.93 - 100 - 19 11.93 0 Angola AGO 67 66 8 9.76 0.02 - - - - - Burundi BDI 67 64 8 8.61 0.03 33 9 3 7.71 0.8 Madagascar MDG 67 68 6 9.23 0.04 100 29 14 10.05 0.08 Cameroon CMR 50 84 4 8.86 0.27 50 84 4 8.86 0.27 Malta MLT 50 53 5 6.16 0.94 0 0 0 - - Rwanda RWA 50 52 5 8.51 0.1 50 52 4 8.37 0.4 Belarus BLR 45 84 4 8.59 0.04 38 62 4 8.52 0.04 Qatar QAT 43 61 5 6.41 0.4 33 11 2 5.76 1 Uzbekistan UZB 40 89 5 9.73 0.02 40 89 5 9.73 0.02 Peru PER 38 29 2 8.14 0.09 29 12 1 7.88 0.13 Bangladesh BGD 35 23 2 10.44 0 35 11 2 10.27 0.01 Morocco MAR 33 17 3 9.39 0.02 50 17 4 9.68 0.01 Sierra Leone SLE 33 12 7 7.94 0.18 50 12 11 8.34 0.11 United Arab ARE 32 44 3 7.77 0.06 50 23 7 8.55 0.04 Emirates Kuwait KWT 27 20 2 6.46 0.51 43 20 3 6.96 0.31 Matı Tunisia TUN 27 44 3 8.3 0.04 0 0 0 - - Burkina Faso BFA 25 16 2 7.98 0.64 50 16 5 8.97 0.21 El Salvador MRT 25 37 2 7.45 1.09 0 0 0 - - Mauritania OMN 25 32 2 6.42 0.58 33 32 3 6.67 2 Oman SLV 25 - 2 6.62 2.15 50 - 5 7.63 0.9 Chile CHL 23 30 1 8.09 0.09 25 30 2 8.11 0.09 Jordan JOR 22 5 2 6.64 0.64 22 5 2 6.64 0.64 Hungary HUN 21 37 2 7.64 0.09 27 37 2 7.85 0.09 Colombia COL 20 14 1 8.58 0.06 0 0 0 - - Serbia & SDN 20 - 1 7.19 - 22 - 2 7.22 0.3 Montenegro Sudan YUG 20 23 1 8.7 0.05 25 23 2 8.87 0.05 Philippines PHL 18 34 2 9.37 0.01 20 34 2 9.52 0.01 ´as Braun and Claudio Raddatz Zambia ZMB 18 17 3 8.06 0.15 22 17 3 8.24 0.1 Armenia ARM 17 11 3 7.15 1.47 20 11 4 7.36 1.17 (Continued ) 241 TABLE 1. Continued 242 All Banks Fully Private Banks Only World Bank Country FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE Country Code (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Lebanon LBN 17 23 1 6.25 1 17 23 1 6.25 1 Moldova, Rep. MDA 17 5 4 7.37 0.23 25 5 6 7.82 0.12 Yemen YEM 17 20 4 9.2 0.1 33 20 14 10.43 0.05 Iceland ISL 15 60 2 5.52 1.63 14 37 2 5.31 9.21 Dominican DOM 14 25 1 7.29 1.24 0 0 0 - - Republic Lithuania LTU 14 7 3 6.96 1.61 14 7 3 6.96 1.61 Macedonia, MKD 14 37 1 5.57 4.96 20 37 2 5.94 5 FYR Korea, Rep. KOR 12 21 1 8.18 0.06 9 4 0 7.15 0.93 Romania ROM 12 42 2 8.1 0.1 0 0 0 - - Croatia HRV 11 23 3 7.17 0.13 0 0 0 - - Hong Kong, HKG 11 16 1 6.81 0.12 10 14 1 6.87 0.12 China THE WORLD BANK ECONOMIC REVIEW Latvia LVA 11 24 2 6.19 0.6 7 20 1 6.05 1.21 South Africa ZAF 11 2 1 8.34 0.05 12 2 1 8.46 0.06 Taiwan, China TWN 11 9 0 - - 4 5 0 - - Thailand THA 11 24 1 8.13 0.08 8 23 1 8.11 0.39 Nepal NPL 9 15 1 7.61 0.94 10 15 1 7.68 0.88 Belgium BEL 8 10 1 7.42 0.13 7 10 1 7.01 0.29 Canada CAN 8 1 0 7.13 0.41 10 1 0 7.37 0.32 Egypt, Arab EGY 8 6 1 9.15 0.05 8 2 1 9.36 0.14 Rep. Israel ISR 8 1 1 5.88 5.21 14 1 1 6.48 2.61 Turkey TUR 8 25 1 7.72 0.1 10 25 1 7.9 0.1 Brazil BRA 7 15 1 9.4 0.02 6 2 1 8.96 0.05 Cyprus CYP 7 3 1 4.65 11.57 8 3 1 4.87 10 Nigeria NGA 7 6 0 8.49 0.1 4 1 0 7.91 0.6 Russian RUS 7 58 1 9.51 0.01 6 7 1 9.17 0.02 Federation Uganda UGA 7 27 1 8.31 0.36 7 27 1 8.31 0.4 China CHN 6 16 0 11.58 0 0 0 0 - - Finland FIN 6 - 0 6.29 3.34 8 - 1 6.45 2.51 Austria AUT 5 13 0 6.24 0.57 5 13 0 6.28 0.57 Netherlands NLD 5 49 0 7.37 0.16 4 36 0 7.12 0.32 India IND 4 22 0 10.2 0.01 4 1 0 10.29 0.02 Mexico MEX 4 3 2 10.19 0.01 0 0 0 - - Pakistan PAK 4 1 0 8.73 0.29 0 0 0 - - Portugal PRT 4 1 0 5.68 6.13 4 1 0 5.75 5 Denmark DNK 3 10 0 5.56 1.55 3 10 0 5.61 1.55 Indonesia IDN 3 0 0 8.34 0.31 3 0 0 8.58 0.31 Luxembourg LUX 3 11 0 3.54 14.88 1 4 0 2.91 - Norway NOR 3 0 0 5.1 11.46 0 0 0 - - Poland POL 3 5 0 6.86 1.94 4 5 0 7.18 1.16 Singapore SGP 3 - 0 5.88 4.93 3 - 0 5.93 4.93 Australia AUS 2 2 0 6.73 2.16 0 0 0 - - Argentina ARG 1 0 0 7.09 1.33 0 0 0 - - France FRA 1 4 0 5.4 8.3 1 4 0 5.45 7.61 Germany DEU 1 3 0 6.56 0.26 1 1 0 5.69 1.31 Italy ITA 1 8 0 6.47 0.34 1 2 0 5.65 1.7 Japan JPN 1 0 0 5.95 4.57 1 0 0 5.98 4.57 Spain ESP 1 15 0 5.67 6.11 1 15 0 5.75 5 Switzerland CHE 1 0 0 4.77 12.11 1 0 0 4.96 10 United GBR 1 0 0 6.9 0.28 2 0 0 6.98 0.3 Kingdom United States USA 1 8 0 7.82 0.07 1 7 0 7.32 0.1 Albania ALB 0 0 0 - - 0 0 0 - - Algeria DZA 0 0 0 - - 0 0 0 - - Andorra ADO 0 0 0 - - 0 0 - - - Antigua & ATG 0 0 0 - - 0 0 - - - Barbuda Aruba ABW 0 0 0 - - 0 0 0 - - Matı Azerbaijan AZE 0 0 0 - - 0 0 0 - - Bahamas BHS 0 0 0 - - 0 0 0 - - Bahrain BHR 0 0 0 - - 0 0 0 - - Barbados BRB 0 0 0 - - 0 0 0 - - Benin BEN 0 0 0 - - 0 0 0 - - Bermuda BMU 0 0 0 - - 0 0 0 - - Bolivia BOL 0 0 0 - - 0 0 0 - - Botswana BWA 0 0 0 - - 0 0 0 - - Brunei BRN 0 0 0 - - 0 0 0 - - Darussalam Bulgaria BGR 0 0 0 - - 0 0 0 - - Cambodia KHM 0 0 0 - - 0 0 0 - - Cape Verde CPV 0 0 0 - - 0 0 0 - - Cayman CYM 0 0 0 - - 0 0 0 - - ´as Braun and Claudio Raddatz Islands Costa Rica CRI 0 0 0 - - - - - - - 243 (Continued ) TABLE 1. Continued 244 All Banks Fully Private Banks Only World Bank Country FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE Country Code (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Coˆ te d’Ivoire CIV 0 0 0 - - 0 0 0 - - Cuba CUB 0 0 0 - - 0 0 0 - - Czech Republic CZE 0 0 0 - - 0 0 0 - - Djibouti DJI 0 0 0 - - 0 0 0 - - Ecuador ECU 0 0 0 - - 0 0 0 - - Estonia EST 0 0 0 - - 0 0 0 - - Ethiopia ETH 0 0 0 - - 0 0 0 - - Gambia, The GMB 0 0 0 - - 0 0 0 - - Ghana GHA 0 0 0 - - 0 0 0 - - Gibraltar GIB 0 0 0 - - 0 0 - - - Greece GRC 0 0 0 - - 0 0 0 - - Grenada GRD 0 0 0 - - 0 0 - - - Guatemala GTM 0 0 0 - - 0 0 0 - - Guyana GUY 0 0 0 - - 0 0 0 - - THE WORLD BANK ECONOMIC REVIEW Haiti HTI 0 0 0 - - 0 0 0 - - Honduras HND 0 0 0 - - 0 0 0 - - Iran, Islamic IRN 0 0 0 - - 0 0 0 - - Rep. Ireland IRL 0 0 0 - - 0 0 0 - - Jamaica JAM 0 0 0 - - 0 0 0 - - Kazakhstan KAZ 0 0 0 - - 0 0 0 - - Kenya KEN 0 0 0 - - 0 0 0 - - Kyrgyz KGZ 0 0 0 - - 0 0 0 - - Republic Lesotho LSO 0 0 0 - - 0 0 0 - - Libya LBY 0 0 0 - - - - - - - Liechtenstein LIE 0 0 0 - - 0 0 0 - - Macao MAC 0 0 0 - - 0 0 0 - - Malawi MWI 0 0 0 - - 0 0 0 - - Malaysia MYS 0 0 0 - - 0 0 0 - - Maldives MDV 0 0 0 - - - - - - - Mali MLI 0 0 0 - - 0 0 0 - - Mauritius MUS 0 0 0 - - 0 0 0 - - Monaco MCO 0 0 0 - - 0 0 - - - Mongolia MNG 0 0 0 - - 0 0 0 - - Mozambique MOZ 0 0 0 - - 0 0 0 - - Namibia NAM 0 0 0 - - 0 0 0 - - Netherlands ANT 0 0 0 - - 0 0 - - - Antilles New Zealand NZL 0 0 0 - - 0 0 0 - - Niger NER 0 0 0 - - 0 0 0 - - Panama PAN 0 0 0 - - 0 0 0 - - Papua New PNG 0 0 0 - - 0 0 0 - - Guinea Paraguay PRY 0 0 0 - - 0 0 0 - - Saint Kitts & KNA 0 0 0 - - 0 0 - - - Nevis Saint Lucia LCA 0 0 0 - - - - - - - Samoa WSM 0 0 0 - - 0 0 0 - - San Marino SMR 0 0 0 - - 0 0 - - - Saudi Arabia SAU 0 0 0 - - 0 0 0 - - Senegal SEN 0 0 0 - - 0 0 0 - - Seychelles SYC 0 0 0 - - 0 0 0 - - Slovakia SVK 0 0 0 - - 0 0 0 - - Slovenia SVN 0 0 0 - - 0 0 0 - - Sri Lanka LKA 0 0 0 - - 0 0 0 - - Suriname SUR 0 0 0 - - - - - - - Swaziland SWZ 0 0 0 - - 0 0 0 - - Sweden SWE 0 0 0 - - 0 0 0 - - Syrian Arab SYR 0 0 0 - - - - - - - Rep. Tanzania TZA 0 0 0 - - 0 0 0 - - Matı Togo TGO 0 0 0 - - 0 0 0 - - Trinidad & TTO 0 0 0 - - 0 0 0 - - Tobago Ukraine UKR 0 0 0 - - 0 0 0 - - Uruguay URY 0 0 0 - - 0 0 0 - - Venezuela, RB VEN 0 0 0 - - 0 0 0 - - Vietnam VNM 0 0 0 - - 0 0 0 - - Zimbabwe ZWE 0 0 0 - - 0 0 0 - - Total 10 12 1 7.58 1.54 9 7 1 7.52 1.4 Note: Countries are sorted in decreasing order by fraction of banks with data on board of directors that had a former politician on their boards; countries with the same values are listed alphabetically. FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole popu- ´as Braun and Claudio Raddatz lation of a country. MAXSHARE is the largest fraction of a country’s population from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. Not available is coded as “-�?. Source: Authors’ analysis based on data described in the text. 245 246 THE WORLD BANK ECONOMIC REVIEW FRACTION OF CONNECTED BANKS. The fraction of connected banks (FRACBANKS) is the number of banks with at least one former politician on the board of directors divided by the number of banks for which there are data on board members. The mean fractions of connected banks of 10 percent for all kinds of banks and 9 percent for private banks are much larger than the fraction of matches among individuals documented above. Indeed, when only countries with at least one match are considered, the average share increases to about a �?fth of the banks. There is interesting variation across countries. The countries with fewest connected banks are Germany, the United States, Italy, Japan, and Switzerland, all with less than 2 percent of banks connected in this way. In contrast, more than two-thirds of the banks are connected in Gabon, Georgia, Myanmar, Angola, Burundi, and Madagascar. The picture is gener- ally the same whether considering all banks or just private banks; the corre- lation between the two groups is 0.86. The rationale behind this �?rst aggregation is that what determines a signi�?- cant political link for a bank is whether the bank has at least one politician on its board. The higher the fraction of the banks in the system that are con- nected in this way, the larger the degree of connectedness between banking and politics. The issue is not about having a large number of people in both worlds but rather about having people in the right place, even if their number is small. In this sense, FRACBANKS is more naturally interpreted as a measure of the institutional connection between banking and politics, rather than a personal matter related perhaps to the existence of a common set of skills. SHARE OF ASSETS OF CONNECTED BANKS. A simple variation on the FRACBANKS measure consists of computing the share of total banking system assets in banks that have a politician on their board. This metric, the share of assets of connected banks (SHAREASSETS), has the advantage of acknowledging that larger banks might differ from smaller ones in their need or ability to connect to politics. Smaller banks may �?nd free-riding on the connections of large banks more pro�?table than establishing their own connections. Also, this measure would probably be more relevant when looking at the likely effects of connectedness since it would be a measure of the amount of credit that is subject to these links. This metric is then more likely a proxy for the extent of power—both political and economic—that these relationships might entail. On a more technical note, giving a higher weight to larger banks minimizes the potential problems induced by the smaller coverage for small banks. SHAREASSETS is strongly and signi�?cantly correlated with FRACBANKS, both for all banks and for private banks (table 2). For countries with at least one match, the mean share is 25 percent for all banks and 18 percent for private banks. The groups of countries that rank very high and very low are similar to those for the FRACBANKS measure. These results suggest that the difference between large and small banks might not be very relevant. The T A B L E 2 . Correlation among Measures of Connectedness All Banks Fully Private Banks Only Measures FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE FRACBANKS SHAREASSETS FRACBANKERS PREVALENCE MAXSHARE FRACBANKS 1.00 1.00 SHAREASSETS 0.88*** 1.00 0.75*** 1.00 FRACBANKERS 0.92*** 0.82*** 1.00 0.87*** 0.55*** 1.00 PREVALENCE 0.40*** 0.30*** 0.43*** 1.00 0.56*** 0.30 0.57*** 1.00 MAXSHARE 2 0.30*** 2 0.30*** 2 0.31*** 2 0.70*** 1.00 2 0.32** 2 0.17** 2 0.25** 2 0.66*** 1.00*** **Signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Correlations are computed including the countries with zero matches (for those measures that can take the value zero). FRACBANKS is the frac- Matı tion of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a country’s population from which politicians and bankers would have to be selected so that the hypoth- esis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. Source: Authors’ analysis based on data described in the text. 247´as Braun and Claudio Raddatz 248 THE WORLD BANK ECONOMIC REVIEW correlation between the measures computed over all banks and over private banks is also quite high (0.79). FRACTION OF CONNECTED BANKERS. The third measure, fraction of connected bankers (FRACBANKERS), is the ratio of the number of matches to the number of bank directors in the dataset. This metric looks at the extent to which politicians populate bank boards. The average fraction of connected bankers across all countries is around 1 percent and is close to 2 percent for countries with more than one match. These numbers suggest that the phenom- enon is not particularly frequent. The correlation with the �?rst two measures is small (0.34 for FRACBANKS and 0.38 for SHAREASSETS for all banks) but statistically signi�?cant. Furthermore, the countries at both tails of the measure are similar to those at the tails of the previous two measures. Thus, despite the low level of the variable, its cross-country variation captures a similar concept to the previous two. PREVALENCE. The �?rst three measures of connectedness are easy to compute and natural in their interpretation. But they do not take into account that the expected number of banker-politicians may differ across countries even if the selection of bankers is not biased toward former politicians. In particular, countries with more matches might simply be countries with fewer people from which both bankers and politicians are selected. For greater precision, the probability of obtaining a given number of matches was derived under the assumption that the people needed to �?ll the politician and banker posts are selected randomly with replacement (at the sample level) from a common pool (see the appendix). Everyone in the pool has the same probability of being selected for either position, and there is no bias in favor of politicians in the selection of bankers. This probability is then used to compute the expected number of matches assuming that the common pool is the entire population of each country (more on this below). This ratio of actual to expected matches (in logs) is called PREVALENCE. The correlation of this metric with the previous ones is not as strong as for the others, particularly with FRACBANKERS, but it is still positive. The countries that rank highest in this connectedness measure are Myanmar, China, Bangladesh, India, and Mexico (see table 1). The countries where the phenomenon is least prevalent include Luxembourg, France, Switzerland, and Norway. For most countries the actual number of matches is many times larger than the expected number because of the assumption that the pool from which directors are selected is the total population of a country. Since it is highly unli- kely that every person has the same probability of being chosen as a politician or a banker, the results for this measure are exaggerated. Nevertheless, the cross-country variation in this measure is the same as it would be if it were assumed that the selection pool for bankers and politicians is a �?xed fraction of a country’s population. In fact, it can be shown that the expected number of ´as Braun and Claudio Raddatz Matı 249 matches is proportional to the size of the pool. Therefore, ELITE ð1Þ PREVALENCE ¼ PREVALENCEðELITEÞ þ log POP where PREVALENCE(ELITE) is the log ratio of actual to expected matches ELITE considering the true size of the elite, and log is the log ratio of the POP size of the elite as a fraction of the population. Thus, as long as the elite are a �?xed fraction of the population across countries, the PREVALENCE measure and true prevalence would differ only in a constant. The measure will be incorrect, however, if there is systematic variation across countries in the elite as a share of population. This could happen if the number of elite is relatively �?xed in all countries, so that the elite decline as a fraction of the population from smaller to larger countries. The analysis con- siders the size of the population in each country to control for this possibility. MAXIMUM SHARE OF POPULATION FOR RANDOMNESS. Another possibility is that the size of the elite is related to the educated portion of the population. If one assumes that the pool is the number of people with a tertiary education, the expected �?gures are closer to the actual number of matches. This correction incorporates the possibility that PREVALENCE is highest in some countries simply because there are so few people in those countries who are capable of assuming these posts. The correction, however, is not free of problems because it is not obvious that the relevant pool is the group of highly educated people. On the one hand, the pool may be too narrowly de�?ned since not all the bankers and politicians have a tertiary education.8 On the other, the pool might not be suf�?- ciently small if a certain kind of economic or �?nancial skill is shared between politicians working in economic spheres within the government and bankers. Most important, such a correction might confound the interpretation of the results because one variable mixes two concepts—availability of human capital and connectedness—that may have independent (and opposite) effects on many country characteristics (such as real GDP per capita). The �?nal measure, maximum share of population for randomness (MAXSHARE), takes into consideration the uncertainty about the size of the pool of individuals from which bankers and politicians are selected. MAXSHARE corresponds to the largest pool (as a fraction of the population) from which bankers and politicians are selected so that the hypothesis that the selection is random could not be rejected at the 5 percent level (for the number of matches found in the data). For most countries, in order not to reject this hypothesis, the size of the pool turns out to be a very small fraction of the 8. See Dreher and others (2009) for data on the educational attainment of presidents. These data show that 30 percent of presidents worldwide since 1975 did not receive a higher education. 250 THE WORLD BANK ECONOMIC REVIEW population. As expected, this variable is negatively correlated with the previous ones because it measures the inverse of the underlying concept. The usual groups of countries are at both extremes of the metric. CONNECTEDNESS AND COUNTRY CHARACTERISTICS. Overall, the different measures are signi�?cantly correlated, suggesting that they are likely to be different proxies for the same general concept. It is also clear that considering links solely to private banks makes little difference, suggesting that politicians sitting on the boards of state-owned banks do not drive the �?ndings for the various measures. Countries that rank highest on the connectedness measures9 (Bangladesh, China, Mexico, India, and the Russian Federation) and those that rank lowest (Luxembourg, Switzerland, Cyprus, Norway, and France ) clearly differ in many other respects as well. The most obvious is economic development. Countries where connectedness is more prevalent are signi�?cantly poorer than countries where it is less prevalent. Mean GDP per capita is $3,944 for countries with higher than the median share of connected banks, and $18,958 for the others. The share of connected banks in countries with lower than median per capita GDP (28.2 percent) is two and a half times larger than the share in more developed countries (11.4 percent). The picture is about the same for the other measures and when only private banks are considered. The second distinctive feature is that countries where the prevalence of con- nectedness is higher also appear to have less developed institutions. For instance, countries with lower than median connectedness have control of corruption indi- cators (de�?ned below) that are one standard deviation higher than countries with higher prevalence. While the share of connected banks is 15.1 percent in countries with higher than median control of corruption, it is 26.5 percent in the rest. Finally, banking sector development differs considerably across the two groups of countries. The ratio of private credit to GDP (from Beck, Demirguc-Kunt, and Levine 2000) is 3 times higher where connectedness measures are lower (76 percent) than where they are higher (25 percent), while the share of connected banks is almost twice as high in countries where banking sector development is low (26.5 percent) than in those where it is high (15.1 percent). Connectedness, then, does not seem to be equally distributed across countries but rather to cluster in countries where things do not work very well. In particu- lar, connectedness is higher where economic development is low and where institutions and the �?nancial system are underdeveloped. These are some of the relationships examined more deeply in the following section. I I . T H E CO R R E L AT E S OF CONNECTEDNESS This section explores the correlates of connectedness �?rst at the bank level and then at the cross-country level. It shows that the measures of connectedness 9. Giving equal weight to each of the �?ve different connectedness measures. ´as Braun and Claudio Raddatz Matı 251 introduced above are robustly correlated to important bank and country characteristics and also to policy choices. Bank Characteristics Connected and unconnected banks can be compared on several characteristics. Here, they are compared on measures of size, pro�?tability, leverage, and riski- ness, which were constructed from Bankscope data using bank statements at the end of 2004. Table 3 shows averages for these characteristics for connected and uncon- nected banks, their differences, and whether the differences are statistically sig- ni�?cant according to a simple test of means. Clearly, connected banks are larger, more pro�?table, and less leveraged than are unconnected banks. They also have a smaller share of net charge-offs to gross loans, suggesting that they take less risk than unconnected banks, although on a worldwide comparison, the difference is not signi�?cant. The sign and signi�?cance of these differences remain unchanged when only fully private banks are considered. The regressions in table 4 further test whether these correlations hold when connected and unconnected banks are compared within a country. The param- eters of the following parsimonious speci�?cation are estimated: ð2Þ Yi;c ¼ a þ b Â CONNECTEDi;c þ gSIZEi;c þ uc þ 1i;c where Yi,c corresponds to the �?nancial characteristics of bank i in country c, which include measures of size, pro�?tability, riskiness, liquidity, and leverage; CONNECTEDi,c is a dummy variable that takes a value of 1 if at least one of the bank’s directors has been a politician or bank regulator, and 0 otherwise; SIZEi,c controls for (log) total assets (except when the left-side variable is itself T A B L E 3 . Differences between Connected and Unconnected Banks, Worldwide Comparison of Average Bank Characteristics (tests of equality of means) Bank Characteristics Connected Unconnected Difference All Bankscope Banks Total assets 9.72 8.60 1.12*** Return on average assets 2.40 1.26 1.14*** Equity/Total assets 14.23 11.44 2.79*** Net charge-off/Average gross loans 0.70 1.24 2 0.54 Fully Private Banks Only Total assets 9.58 8.44 1.14*** Return on average assets 2.46 1.19 1.27*** Equity/Total assets 15.20 11.17 4.02*** Net charge-off/Average gross loans 0.66 1.11 2 0.45 ***Signi�?cant at the 1 percent level on a simple test of means. Source: Authors’ analysis based on data described in the text. 252 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Differences between Connected and Unconnected Banks, Within-Country Comparison of Bank Characteristics (regression analysis) Total Return on Equity/Total Net Charge-off/Average Assets (1) Average Assets (2) Assets (3) Gross Loans (4) All Bankscope Banks Connected 0.3358** 0.0062** 0.0225** 2 0.0054** (0.1349) (0.0025) (0.0105) (0.0023) Number of observations 3,312 3,285 3,311 1,176 R2 0.635 0.150 0.329 0.294 Fully Private Banks Only Connected 0.3131* 0.0079** 0.0284*** 2 0.0050* (0.1600) (0.0031) (0.0108) (0.0026) Number of observations 2,845 2,819 2,845 1,016 R2 0.611 0.145 0.324 0.239 *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Number in parentheses are robust standard errors. All dependent variables are in logs. Ratios that can take negative values are measured as the log of one plus the corresponding ratio. Connected is a dummy variable that takes a value of 1 if a bank has at least one former politician among its board members and 0 otherwise. All regressions included a country �?xed effect, and all regressions except for total assets also control for (log) total assets. Source: Authors’ analysis based on data described in the text. a measure of size); uc is a country �?xed-effect that controls for cross-country differences in bank characteristics, and 1i,c is a residual term. Since these regressions exploit only within-country differences between connected and unconnected banks, and bank-level data are notoriously noisy, all variables are measured in logarithms to reduce the inﬂuence of outliers (variables corre- sponding to ratios that can plausibly take negative values are expressed as the logarithm of one plus the variable).10 As in table 3, the parameters of the benchmark model are estimated separately for all banks and for banks with no public ownership. The coef�?cients con�?rm that connected banks tend to be the largest banks in a country, with total assets about 34 percent larger than those of the average unconnected bank (see table 4, column 1). Similar results are obtained for other measures of size, such as loans and country ranking (not reported). Connected banks also tend to be more pro�?table and to have a return on average assets 0.6–0.8 percent higher than the average unconnected bank (column 2). Leverage is signi�?cantly lower among connected banks; the ratio of equity to total assets is 2 percent higher in connected banks than in the average bank, a difference that increases to 3 percent in the sample of privately owned banks (column 3). Connected banks also tend to have a lower 10. This is not a major issue in the overall comparisons in table 3, which compute the average of each characteristic across all connected and all unconnected banks. In contrast, these regressions compare connected and unconnected banks within a country. ´as Braun and Claudio Raddatz Matı 253 proportion of write-offs and impaired loans relative to average gross loans and reserves, suggesting that they take on less risk (column 4). Overall, the results across and within countries show that connected banks are larger, more pro�?table, less leveraged, and less risky than unconnected banks, regardless of whether there is any government ownership.11 In addition, to see whether bank characteristics are correlated with the share of former poli- ticians on a bank’s board, equation (2) was reestimated using that share (a measure of the intensity of banks’ political connections) instead of the dummy variable described previously. While the results are similar to those reported in table 4, they are weaker in statistical and economic terms (not reported). Thus, desirable bank characteristics are more strongly correlated with whether a bank has a former politician on its board than with the number of former poli- ticians. It does not seem that politicians cluster in banks with desirable characteristics. Country Characteristics As discussed in section I, a simple look at the data suggested that banks were less politically connected in richer, more �?nancially developed countries. The results reported here systematically test whether the degree of connectedness of banks is robustly correlated with important country characteristics and whether those correlations survive when controlling for several straightforward omitted variables in a multivariate setting. Country characteristics, such as development level, institutional quality, extent of pro-banker regulation, and banking sector development, were related to the �?ve measures of connectedness by estimating the parameters of the following speci�?cation: ð3Þ Yc ¼ a þ b Â CONNECTEDNESSc þ g 0 Xc þ 1c where Y is a measure of any of the country characteristics described above for country c, and CONNECTEDNESS is any of the �?ve measures of connected- ness discussed in section I: the fraction of connected banks (FRACBANKS), the share of assets of connected banks (SHAREASSETS), the fraction of connected bankers (FRACBANKERS), the (log) of actual to expected number of matches of bankers-politicians (PREVALENCE), and the maximum share of the popu- lation from which bankers and politicians are selected so that the null of random selection cannot be rejected at a 5 percent level of signi�?cance (MAXSHARE). The variables in X control for other country characteristics that may be simultaneously related to both Y and CONNECTEDNESS. 11. These �?ndings are robust to using the standard Heckman (1979) two-step estimator to control for possible sample selection issues in the set of banks with information on directors (not reported). 254 THE WORLD BANK ECONOMIC REVIEW Economic Development The results show a strong negative correlation between the degree of connect- edness and GDP per capita, whether considering all banks or only those that are fully private (table 5). The correlation is particularly strong when no T A B L E 5 . Connectedness and Development Controls: Log Population, Log Fraction of Population with No Controls Tertiary Education Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) All Bankscope Banks FRACBANKS 2 2.044*** 79 0.136 2 0.450 78 0.794 (0.721) (0.376) SHAREASSETS 2 0.994** 76 0.047 0.146 75 0.788 (0.479) (0.250) FRACPOLITICIANS 2 23.35*** 79 0.192 2 5.644 78 0.796 (6.435) (3.742) PREVALENCE 2 0.481*** 79 0.383 2 0.157** 78 0.805 (0.0588) (0.0600) MAXSHARE 0.163*** 79 0.184 0.0319* 78 0.795 (0.0214) (0.0179) Fully Private Banks Only FRACBANKS 2 2.673*** 64 0.215 2 0.848* 63 0.814 (0.678) (0.433) SHAREASSETS 2 1.425*** 61 0.061 0.167 60 0.796 (0.490) (0.271) FRACPOLITICIANS 2 20.72*** 64 0.26 2 8.004*** 63 0.827 (3.230) (2.195) PREVALENCE 2 0.534*** 64 0.436 2 0.203*** 63 0.829 (0.0530) (0.0717) MAXSHARE 0.197*** 63 0.153 0.0562** 62 0.801 (0.0340) (0.0266) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: The dependent variable is the average 1995– 2005 log real GDP per capita (from Heston, Summers, and Aten 2006). Standard errors (in parentheses) are robust to heteroskedasti- city. FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a country’s population from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. Source: Authors’ analysis based on data described in the text and table. ´as Braun and Claudio Raddatz Matı 255 F I G U R E 1. Connectedness and Development Note: Figure shows the relation between (log) average real 1995– 2005 GDP per capita (from Heston, Summers, and Aten 2006) and (log) ratio of actual to expected number of matches between bankers and politicians (PREVALENCE), controlling for (log) fraction of population over age 25 with a tertiary education and log population. The displayed coef�?cient is the value for the PREVALENCE measure of connectedness in the multivariate regression against log real GDP per capita. Country observations are labeled according to the World Bank’s codi�?cation system (see Table 1.) Source: Authors’ analysis based on data described in the text. additional controls are included (columns 1–3), but it survives after controlling for log population and for the fraction of the population with tertiary edu- cation (columns 4 –6), especially when focusing on fully private banks. Educational attainment is particularly relevant as a standard measure of a country’s stock of human capital (which most theories relate to a country’s per capita GDP), but it is also important as a proxy for the size of the pool of elite from which politicians and bankers are selected (see section I). The results are statistically stronger for the more complex measures of connectedness: PREVALENCE and MAXSHARE. This suggests that these measures have greater economic content than the simpler ones. Nonetheless, results are quali- tatively similar, whatever the measure. Furthermore, �?gure 1 shows that the negative correlation between connected- ness and development is not driven by a few outliers but reﬂects a robust pattern of the data. The relation between connectedness and GDP per capita (from Heston, Summers, and Aten 2006) is economically large. For instance, the differ- ence in (log) GDP per capita between Morocco and France is commensurate with their difference in PREVALENCE. Although causality cannot be attributed 256 THE WORLD BANK ECONOMIC REVIEW to this strong cross-country correlation without a good instrument, it is clear that the degree of connectedness is not neutral but is associated with a country’s level of development. The regressions discussed below show that connectedness is also associated with other country characteristics that have been causally related in the literature to level of development, even after controlling for the direct link between development and connectedness documented here. Institutions Correlating the �?ve measures of connectedness with cross-country measures of institutional quality (from Kaufmann, Kraay, and Mastruzzi 2004) shows that connectedness is signi�?cantly higher in countries with less developed insti- tutions for preventing corruption and limiting the power of the government over its citizens (voice and accountability; table 6). The relation holds for all banks and for private banks. For all measures, the relation between connected- ness and institutional quality is signi�?cantly negative even after controlling for GDP per capita and population size (columns 4 –6 and 10 –12). This is reassur- ing because of the widely documented link between institutions and develop- ment and because the measures of connectedness may be correlated with population size. With these correlations, it is not surprising that the estimated coef�?cient changes according to the unconditional speci�?cation. However, all the coef�?cients maintain their sign and statistical signi�?cance, which shows that the relations between connectedness and a country’s level of development and population size are not qualitatively driving the �?ndings.12 It is even clearer than for the case of overall development that a few outliers do not drive the relations between connectedness and institutional quality (�?gure 2). The magnitude of the estimated coef�?cient is also economically rel- evant: a one standard deviation increase in SHAREASSETS (equivalent to the difference between Luxembourg and Philippines) is associated with a 0.4 decline (around half a standard deviation) in the control of corruption, corre- sponding to 25 percent of the difference in control of corruption between the two countries. Also, as shown in the bottom panel of �?gure 2, the difference in control of corruption between Angola and Spain is commensurate with their difference in PREVALENCE. Regulation The results so far have shown that prevalence is systematically related to under- development and weak institutions. The next test is for a systematic relation between connectedness and banking sector regulation. As discussed in the introduction, the political economy literature typically associates the links 12. To capture nonlinearities, speci�?cations were also estimated that included a quadratic term for log GDP per capita in addition to log GDP per capita and log population. The results, available on request, are qualitatively and quantitatively similar to those obtained with the baseline control set. This exercise was repeated for all the regressions reported in this section, with similar results. T A B L E 6 . Connectedness and Institutions Control of Corruptiona Voice and Accountabilitya No Controls Controls: Log Real GDP, Log Population No Controls Controls: Log Real GDP, Log Population Number of Number of Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) All Bankscope Banks FRACBANKS 2 2.377*** 79 0.21 2 1.230*** 79 0.72 2 2.168*** 79 0.23 2 1.264*** 79 0.58 (0.435) (0.371) (0.556) (0.440) SHAREASSETS 2 1.575*** 76 0.15 2 1.012*** 76 0.73 2 1.539*** 76 0.17 2 1.076*** 76 0.62 (0.379) (0.285) (0.368) (0.304) FRACPOLITICIANS 2 25.19*** 79 0.26 2 13.30*** 79 0.72 2 22.42*** 79 0.27 2 13.19*** 79 0.58 (3.691) (3.897) (4.308) (3.962) PREVALENCE 2 0.473*** 79 0.43 2 0.263*** 79 0.73 2 0.393*** 79 0.38 2 0.330*** 79 0.63 (0.0575) (0.0636) (0.0491) (0.0718) MAXSHARE 0.174*** 79 0.24 0.0613*** 79 0.71 0.143*** 79 0.21 0.0644*** 79 0.56 (0.0242) (0.0201) (0.0203) (0.0222) Fully Private Banks Only FRACBANKS 2 2.317*** 64 0.20 2 0.573 64 0.73 2 2.335*** 64 0.28 2 1.005 64 0.56 (0.548) (0.530) (0.712) (0.746) SHAREASSETS 2 1.691*** 61 0.12 2 0.790*** 61 0.74 2 1.734*** 61 0.15 2 0.926** 61 0.59 (0.404) (0.283) (0.458) (0.394) FRACPOLITICIANS 2 15.60*** 64 0.19 2 0.782 64 0.72 2 14.51*** 64 0.22 2 2.784 64 0.54 (3.943) (3.699) (4.448) (4.916) Matı PREVALENCE 2 0.474*** 64 0.43 2 0.146** 64 0.74 2 0.367*** 64 0.35 2 0.250*** 64 0.60 (0.0543) (0.0673) (0.0555) (0.0814) MAXSHARE 0.207*** 63 0.21 0.0717** 63 0.73 0.171*** 63 0.19 0.103*** 63 0.58 (0.0377) (0.0290) (0.0265) (0.0313) **Signi�?cant at the 5 percent level; ***Signi�?cant at the 5 percent level FRACBANKS is the fraction of banks with Bankscope data on board of direc- tors that had a former politician on their boards. Note: Standard errors (in parentheses) are robust to heteroskedasticity. FRACBANKS is the fraction of banks with Bankscope data on board of direc- tors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a ´as Braun and Claudio Raddatz country’s population from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. a. Average for 1996– 2002, from Kaufmann, Kraay, and Mastruzzi (2004). 257 Source: Authors’ analysis based on data described in the text and table. 258 THE WORLD BANK ECONOMIC REVIEW F I G U R E 2. Connectedness and Institutions Note: Figures show the relation between control of corruption (average 1996– 2002 from Kaufmann, Kraay, and Mastruzzi 2004) and the fraction of total banking system assets owned by connected banks (SHAREASSETS; top panel) and the (log) ratio of actual to expected number of matches between bankers and politicians (PREVALENCE; bottom panel), controlling for (log) real GDP per capita (adjusted for purchasing power parity) and log population. The displayed coef�?cients are the values for the two connectedness measures in the multivariate regression against control of corruption. Source: Authors’ analysis based on data described in the text. ´as Braun and Claudio Raddatz Matı 259 between regulators and regulated �?rms with private interests that depend criti- cally on both parties having something to gain from colluding. Regulation that favors incumbents in the banking system is the obvious service that politicians can exchange for a seat on a bank’s board. Barth, Caprio, and Levine (2003) use �?ve dimensions of �?nancial regulation to show how countries regulate their �?nancial systems: restrictions on bank activities, entry regulation, supervisory powers, private monitoring and self- regulation, and capital requirements. They assign an index to each of these broad ways of regulating banks that corresponds to the �?rst principal com- ponent of the answers to surveys conducted by regulators in each country. To address the ambiguity inherent in some of these dimensions, these indexes were used to construct an overall measure of the pro-banker leaning of �?nancial regulations across countries. For instance, it is unclear whether restric- tions on bank activities are pro- or anti-incumbents. On the one hand, restric- tions constrain the ability of banking incumbents to expand into new lines of business. On the other hand, restrictions constrain other institutions from expanding into the banking business. Similarly, whether giving responsibility for supervision and monitoring to the public or private sector is pro- or anti- bankers depends on what type of monitors are more easily captured. Instead of taking an arbitrary stance on whether each of these �?ve dimensions of �?nancial regulation is pro- or anti-banking incumbents, cross-country data on the degree of rents in a country’s banking sector (measured as the average net interest margin, also from Barth, Caprio, and Levine 2003) are used to build a de facto index by regressing these rents on the �?ve individual indexes. Burnside and Dollar’s (2000) methodology was used to construct an index of the pro- banker intensity of regulation by weighting each index by its estimated elasticity to rents. The intention is to let the data speak: if a given dimension of regulation is more pro-banker, an increase in its index should be associated with higher rents (and vice-versa). The regression yields the following result NIM ¼ :30 ÂENT À :32 ÂCAP þ :51 ÂACT À 1:0 ÂPRIV À :05 ÂOSP; ð:20Þ ð:31Þ ð:33Þ ð:38Þ ð:24Þ ð4Þ 2 R ¼ 0:28 where NIM is a country’s banking sector average net interest margin, and ENT, CAP, ACT, PRIV, and OSP are the �?ve principal component indexes as described above: entry restrictions, capital requirements, activities restrictions, private monitoring, and overall supervisory power (all standardized to have zero mean and unit variance so that the magnitude of the coef�?cients reveal the rela- tive importance of each dimension). According to the regression, average net interest margins are positively correlated with restrictions on entry and activity and negatively correlated with capital requirements, the extent of private moni- toring, and the power of the supervisor. In terms of magnitude and signi�?cance, private monitoring has the largest correlation with margins, followed by 260 THE WORLD BANK ECONOMIC REVIEW restrictions on activities, capital requirements, and entry restrictions. Surprisingly, the index of supervisory power has a negligible correlation with margins, in both magnitude and signi�?cance. In addition to this index, the Kaufmann, Kraay, and Mastruzzi (2004) index of regulatory quality was also used (the index measures the incidence of market-unfriendly policies such as price controls or inadequate bank supervi- sion), and the correlation between connectedness and each of the �?ve individ- ual dimensions of regulation was checked (see table A1 in the appendix). Table 7 shows the relation between the measures of connectedness and the index of pro-banker regulation (columns 1 –6) and the index of overall regulat- ory quality (columns 7–12) for all banks and for private banks only, both unconditionally and after controlling for log real GDP per capita and log popu- lation. With a few exceptions, there is a positive relation between connected- ness, however measured, and the index of pro-banker regulation. There is also a strong negative correlation between connectedness and the index of regulat- ory quality (the correlations with MAXSHARE have the opposite sign, as expected). The results are especially strong when connectedness is measured among private banks only, demonstrating again that politicians sitting on boards in public banks do not drive the �?ndings. Again, the economic magni- tude of the effect is large: moving from the 10th to the 90th percentile of PREVALENCE is associated with a one standard deviation increase in the index of pro-bank regulation, an increase roughly commensurate with the difference between the index in Lithuania and Spain. Similarly, the same increase in PREVALENCE is associated with more than a one standard devi- ation decline in the index of regulatory quality, commensurate with the differ- ence between Egypt and Japan. The correlations with the regulatory index are not driven by a few outliers (�?gure 3), although the relation is not as strong as that with country character- istics. This is due partly to the smaller sample for regulatory variables, but also to the dif�?culty of aggregating the indicators into a measure of pro-banker regulation. To check the robustness of the results, the pro-banker index was also built using the simpler indexes reported by Barth, Caprio, and Levine (2003) for each dimension of regulation instead of the principal component indexes. The results are qualitatively similar, but signi�?cance is lost in several cases. Finally, the results were checked using data from Barth, Caprio, and Levine (2006) to construct a simple index (rather than a principal components index) based on surveys in 2001 and 2003, which increases the cross-sectional dimension of the data. As before, the results are qualitatively similar, but the signi�?cance is lost except in the unconditional regressions and the conditional regressions using MAXSHARE and PREVALENCE.13 13. Results are available on request. T A B L E 7 . Connectedness and Regulation Pro-Banker Regulation Indexa Regulatory Qualityb Controls: Log Real GDP, Log Controls: Log Real GDP, Log No Controls Population No Controls Population Number of Number of Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) All Bankscope Banks FRACBANKS 5.055*** 51 0.25 1.733 51 0.49 2 2.175*** 79 0.29 2 1.401*** 79 0.68 (1.456) (2.142) (0.456) (0.362) SHAREASSETS 3.818*** 48 0.26 2.360** 48 0.57 2 1.593*** 76 0.24 2 1.190*** 76 0.70 (0.888) (0.963) (0.339) (0.332) FRACPOLITICIANS 54.51*** 51 0.33 28.71 51 0.52 2 23.82*** 79 0.38 2 17.35*** 79 0.72 Matı (18.31) (25.83) (3.833) (3.721) PREVALENCE 0.491*** 51 0.25 0.362** 51 0.53 2 0.349*** 79 0.38 2 0.241*** 79 0.67 (0.0968) (0.170) (0.0475) (0.0739) MAXSHARE 2 0.216*** 51 0.14 2 0.0709 51 0.49 0.116*** 79 0.18 0.0352** 79 0.61 (0.0501) (0.0435) (0.0178) (0.0157) Fully Private Banks Only FRACBANKS 4.444** 46 0.20 1.877 46 0.50 2 2.170*** 64 0.28 2 1.004** 64 0.64 (1.729) (1.586) (0.485) (0.501) SHAREASSETS 4.561*** 43 0.21 3.497*** 43 0.60 2 1.716*** 61 0.19 2 1.048** 61 0.65 (1.463) (1.219) (0.418) (0.469) FRACPOLITICIANS 55.28*** 46 0.31 39.28** 46 0.55 2 14.65*** 64 0.26 2 4.803 64 0.62 ´as Braun and Claudio Raddatz (14.28) (18.33) (3.581) (3.765) (Continued ) 261 262 TABLE 7. Continued Pro-Banker Regulation Indexa Regulatory Qualityb Controls: Log Real GDP, Log Controls: Log Real GDP, Log No Controls Population No Controls Population Number of Number of Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) PREVALENCE 0.612*** 46 0.32 0.389** 46 0.54 2 0.360*** 64 0.40 2 0.178** 64 0.64 (0.108) (0.170) (0.0437) (0.0787) THE WORLD BANK ECONOMIC REVIEW MAXSHARE 2 0.267*** 46 0.16 2 0.0997* 46 0.50 0.135*** 63 0.14 0.0407 63 0.60 (0.0555) (0.0511) (0.0286) (0.0268) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: A higher number indicates more pro-banker regulation or better regulatory quality. Standard errors (in parentheses) are robust to heteroskedasti- city. FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a country’s population from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. a. Index built from Barth, Caprio, and Levine (2003) using the methodology of Burnside and Dollar (2000). b. Average for 1996– 2002 from Kaufmann, Kraay, and Mastruzzi (2004). Source: Authors’ analysis based on data described in the text and table. ´as Braun and Claudio Raddatz Matı 263 F I G U R E 3. Connectedness and Pro-Banker Regulation Note: Figures show the relation between the index of pro-banker regulation and the fraction of total banking system assets owned by connected banks (SHAREASSETS; top panel) and the (log) ratio of actual to expected number of matches between bankers and politicians (PREVALENCE; bottom panel), controlling for (log) real GDP per capita (adjusted for purchasing power parity) and log population. The displayed coef�?cients are values for the two connectedness measures in the multivariate regression against pro-banker regulation. Source: Authors’ analysis based on data described in the text. 264 THE WORLD BANK ECONOMIC REVIEW Financial Development The evidence presented above suggests that the connectedness of bankers and politicians is signi�?cantly and robustly correlated with how the banking sector operates and is regulated. Insofar as these differences have no impact on the ef�?ciency of the �?nancial system, the issue would simply be a matter of diverse preferences across countries. The importance rises, however, if the connection between bankers and politicians is correlated with the ability of the system to allocate funds ef�?ciently. This section tests whether connectedness is related to the degree of development of the banking system. The speci�?cation is the same as above, with Y now being each country’s log ratio of bank credit to the private sector to GDP. Also as before, univariate and multivariate regressions are presented that control for per capita GDP and population size and for other standard determinants of �?nancial development. The coef�?cient of all measures of connectedness is negative (except, of course, for MAXSHARE, which is an inverse measure of connectedness) and almost always signi�?cant in univariate and multivariate regressions regardless of whether connectedness is measured over all banks or private banks only (table 8). In fact, as before, the results are stronger when connectedness is measured over private banks only. Thus, connectedness is associated with a lower degree of banking sector development. The relation is large in economic terms: moving from the 10th to the 90th percentile of prevalence is associated with a ratio of private credit to GDP 45 percentage points higher, an increase roughly commensurate with the difference between Philippines and Japan. Again, a few outliers do not drive the results (�?gure 4). The negative correlation between the measures of connectedness and �?nan- cial development is not driven by the traditional measures used to explain �?nancial development across countries, such as the degree of protection of creditors, the quality of accounting practices, and investment opportunities measured using the decade’s effective GDP growth rate (columns 7–9).14 Both creditor rights and accounting quality enter positively as expected (although not signi�?cantly). Robustness The results have shown that the connectedness of banks, however measured, is negatively correlated with economic development, the existence of less corrupt and more accountable institutions, and development of the banking sector and is positively correlated with the extent to which regulation favors bank incum- bents. As mentioned, without a good instrument for connectedness, causal inferences cannot be made, but it has been shown that these reduced form relations are not trivially driven by some obvious third variables that may be 14. When the decadal growth rate of per capita GDP is included, log real per capita GDP is dropped. T A B L E 8 . Connectedness and Financial Development Controls: Log Population, Creditor Controls: Log Population, Log Real Rights, Accounting Standards, per No Controls GDP per Capita Capita GDP Growth Number of Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) All Bankscope Banks FRACBANKS 2 2.905*** 70 0.276 2 0.844 70 0.63 2 3.275*** 59 0.382 (0.512) (0.526) (0.575) SHAREASSETS 2 2.189*** 67 0.219 2 1.039** 67 0.65 2 1.961*** 56 0.333 (0.381) (0.404) (0.544) FRACPOLITICIANS 2 33.95*** 70 0.419 2 15.13** 70 0.657 2 34.57*** 59 0.421 Matı (5.164) (6.581) (7.436) PREVALENCE 2 0.412*** 70 0.268 2 0.229** 70 0.651 2 0.466*** 59 0.413 (0.0703) (0.0870) (0.0849) MAXSHARE 0.150*** 70 0.128 0.0358 70 0.621 0.125*** 59 0.262 (0.0292) (0.0226) (0.0288) Fully Private Banks Only FRACBANKS 2 2.594*** 58 0.301 2 0.857* 58 0.69 2 2.958*** 49 0.528 (0.466) (0.441) (0.437) SHAREASSETS 2 2.216*** 55 0.175 2 1.155*** 55 0.705 2 1.677** 46 0.319 FRACPOLITICIANS 2 21.18*** 58 0.408 2 9.147*** 58 0.715 2 23.94*** 49 0.558 ´as Braun and Claudio Raddatz (3.467) (3.323) (4.027) (Continued ) 265 266 TABLE 8. Continued Controls: Log Population, Creditor Controls: Log Population, Log Real Rights, Accounting Standards, per No Controls GDP per Capita Capita GDP Growth Number of Number of Number of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) PREVALENCE 2 0.441*** 58 0.327 2 0.193** 58 0.701 2 0.416*** 49 0.506 (0.0744) (0.0813) (0.0910) MAXSHARE 0.195*** 58 0.138 0.0617** 58 0.681 0.154*** 49 0.344 THE WORLD BANK ECONOMIC REVIEW (0.0428) (0.0306) (0.0433) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Standard errors (in parentheses) are robust to heteroskedasticity. The dependent variable is the (log) ratio of average 1995– 2005 private credit to GDP (from Beck, Demirguc-Kunt, and Levine 2000). FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a country’s popu- lation from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. Source: Authors’ analysis based on data described in the text and table. ´as Braun and Claudio Raddatz Matı 267 F I G U R E 4. Connectedness and Financial Development Note: The �?gures show the relation between the ratio of average 1995– 2005 private credit to GDP and the fraction of total banking system assets owned by connected banks (SHAREASSETS; top panel) and the (log) ratio of actual to expected number of matches between bankers and politicians (PREVALENCE; bottom panel), controlling for (log) real GDP per capita (adjusted for purchasing power parity) and log population. The displayed coef�?cients are the values for the two connectedness measures in the multivariate regression against private credit to GDP. Source: Authors’ analysis based on data described in the text. 268 THE WORLD BANK ECONOMIC REVIEW simultaneously related to the connectedness measures and any of the country characteristics analyzed, such as a overall development or population size. The regressions reported here address some further robustness concerns. As discussed in section I, although the PREVALENCE measure takes no stance on the share of the population from which bankers and politicians are selected, it assumes that the share is constant across countries. This is a reason- able assumption, but it may be that the elite are not a �?xed share of the popu- lation but rather a �?xed number of people. If so, PREVALENCE, one of the most robust measures, could simply be capturing the relation between cross- country differences in the size of the elite as a fraction of the population over several country characteristics. This is partially controlled by including the log population in the speci�?cations, which does not eliminate the �?ndings of the unconditional regressions. Nevertheless, it is also possible that the size of the elite is not �?xed but proportional to the share of the highly educated popu- lation. To check for this possibility, the log share of the population with ter- tiary education was added to each speci�?cation.15 The regressions show that differences in the size of the elite as a fraction of the population do not drive the documented negative correlation of connectedness with institutions and �?nancial development or the positive correlation with pro-banker regulations. Although this is mainly a concern for the PREVALENCE measure, results are also reported using the share of assets of connected banks to show that control- ling for this additional variable does not change these results either. Results for other variables are similar and available on request. Another concern with the connectedness measures is that, empirically, they are negatively correlated with the number of banks reporting to Bankscope. This number is an endogenous variable that may clearly be correlated with banking sector development, but since the measures of connectedness may be mechanically related to this number by construction, the documented corre- lations could be spurious. To check for this possibility, the measures were recomputed using only the 10 largest banks in a country as measured by total assets at the end of 2004 (columns 4–6 in table 9). All banks were included for countries with fewer than 10 reporting banks. This reduced by two orders of magnitude the cross-country variance in the number of banks used in calcu- lating the measures of connectedness, and the resulting measures are not signi�?- cantly correlated with the number of banks. Nevertheless, the results obtained with these measures are quantitatively and qualitatively similar to those obtained when all banks are included. Thus the signi�?cantly larger number of banks reporting in richer and more developed countries is not behind the docu- mented correlations. 15. As shown in section I, PREVALENCE computed using the total population equals PREVALENCE computed using only the elite plus the log of elite share of the population. Assuming that this log share is proportional to the share of the population with tertiary education, true PREVALENCE would be PREVALENCE with all population less the fraction of the population with a tertiary education. T A B L E 9 . Robustness Exercises Controls: Tertiary Education, Computing Connectedness on 10 Dropping Countries with Fewer Using Robust Regression; Controls: Log Real GDP Per Log Real GDP Per Capita, Log Largest Banks only; Controls: than Two Matches; Controls: Controls: Log Real GDP Per Capita, log Population, and Population Log real GDP, Log Population Log Real GDP, Log Population Capita and Log Population Former Socialist Countries Dependent Number of Number of Number of Number of Number of Variable and Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) Dependent variable: control of corruptiona SHAREASSETS 2 0.988*** 75 0.73 2 0.956*** 65 0.71 2 1.066*** 52 0.72 2 1.119*** 76 0.72 2 0.828*** 73 0.76 (0.318) (0.283) (0.330) (0.287) (0.301) PREVALENCE 2 0.251*** 78 0.74 2 0.234** 65 0.68 2 0.335*** 52 0.73 2 0.271*** 79 0.70 2 0.243*** 76 0.77 (0.0661) (0.105) 0.711 (0.0890) (0.0767) (0.0643) Dependent variable: pro-banker regulation indexb SHAREASSETS 2.193** 48 0.57 2.081* 39 0.52 2.725** 36 0.55 2.427*** 48 0.52 1.779** 48 0.62 (0.952) (1.073) (1.024) (0.896) (0.871) PREVALENCE 0.347** 51 0.55 0.599* 39 0.50 0.634** 36 0.53 0.322* 51 0.47 0.256* 51 0.59 (0.164) (0.300) (0.234) (0.174) (0.149) Dependent variable: �?nancial developmentc SHAREASSETS 2 1.035** 67 0.65 2 1.082** 56 0.62 2 1.347** 44 0.67 2 1.122*** 67 0.62 2 0.697* 64 0.72 Matı (0.408) (0.423) (0.500) (0.419) (0.353) PREVALENCE 2 0.228** 70 0.65 2 0.425*** 56 0.64 2 0.311** 44 0.64 2 0.216** 70 0.62 2 0.182)** 67 0.73 (0.0870) (0.125) (0.134) (0.0947) (0.0792) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Standard errors (in parentheses) are robust to heteroskedasticity. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. a. Average for 1996– 2002, from Kauffman, Kraay, and Mastruzzi (2004). b. Index built from Barth, Caprio, and Levine (2003) using the methodology of Burnside and Dollar (2000). ´as Braun and Claudio Raddatz c. Log ratio of average 1996– 2002 private credit to GDP (from Beck, Demirguc-Kunt, and Levine 2000). Source: Authors’ analysis based on data described in the text. 269 270 THE WORLD BANK ECONOMIC REVIEW The analysis was restricted to countries with at least one match, but it could be argued that that was not restrictive enough and that the �?nding of one or two matches might be an overinterpretation. To check this, the analysis was restricted to countries with at least two matches (making two matches the base- line). The results follow the same pattern as before (columns 7 –9), indicating that countries with more than one match drive the correlations. Further restricting the sample to include only countries with at least three matches (31 countries) yields qualitatively similar results, but some of the coef�?cients are not signi�?cant at a 10 percent level because of the reduction in the sample size (31 countries; not reported). The regressions reported in columns 10–15 check for the inﬂuence of out- liers on the results. Columns 10–12 take an agnostic approach and simply use a robust regression technique to reduce the inﬂuence of outliers.16 As before, there is no important change in the results. Columns 13 –15 control for the potential inﬂuence of socialist countries. Although �?gures 1 –4 and the regressions reported in columns 10–12 show that a few countries do not drive the correlations, they also show that the group of formerly socialist countries tends to be at the extreme of the distribution of connectedness. Thus, the corre- lations reported may come from the difference between former socialist countries and the rest of the sample. To check for this without unnecessarily reducing the sample, a dummy variable was added that takes a value of 1 for formerly socialist countries and 0 otherwise. Reassuringly, the sign and magni- tude of all the reduced-form coef�?cients remain unaffected (the dummy for for- merly socialist countries is typically signi�?cant and in the expected direction, for example, with lower �?nancial development). Finally, because the quality of the information in many countries with zero matches cannot be trusted and the �?nding of a zero match provides very little information on the process driving the selection of bankers and politicians, the analyses were conducted again after dropping countries with zero matches. Countries with zero matches are very heterogeneous, and there is no good way of separating the zeroes resulting from data quality from the true zeroes. While this seems to be the right approach, it would be troubling if the pattern of results changed qualitatively or was even reversed when the zeroes were con- sidered. That was not the case (table 10). As a mild way of separating zeroes resulting from poor data from true zeroes, only the countries with zero matches and more than two banks were included. The unconditional regressions always result in signi�?cant coef�?cients of the same sign as those reported previously, and the regressions controlling for log real GDP per capita and log population size also show a similar pattern to those previously reported. The only major difference is that the coef�?cients for the degree of pro-banker regulation are no longer statistically signi�?cant for any measure. This is not so surprising considering that the relation with 16. Stata command rreg. T A B L E 1 0 . Robustness Exercise Including Countries with Zero Matches Control of Corruptiona Pro-banker Regulation Indexb Financial Developmentc Number of Number of Number of Measure Coef�?cient (1) Observations (2) R2 (3) Coef�?cient (4) Observations (5) R2 (6) Coef�?cient (7) Observations (8) R2 (9) FRACBANKS 2 1.429*** 131 0.08 2.806** 74 0.07 2 1.788*** 110 0.10 (0.256) (1.345) (0.407) SHAREASSETS 2 0.819** 126 0.04 2.051** 71 0.07 2 1.247*** 107 0.07 (0.317) (0.928) (0.362) FRACBANKERS 2 12.56*** 131 0.10 33.95** 74 0.11 2 22.51*** 110 0.17 (2.448) (13.11) (4.278) PREVALENCE 2 0.301*** 130 0.09 0.240* 73 0.03 2 0.211** 109 0.04 (0.0585) (0.124) (0.0929) FRACBANKS 2 0.633** 126 0.64 0.0525 72 0.54 2 0.734* 108 0.58 (0.300) (1.380) (0.404) SHAREASSETS 2 0.552** 123 0.64 1.207 69 0.56 2 0.815** 105 0.59 (0.254) (0.976) (0.327) FRACBANKERS 2 6.257** 126 0.64 5.338 72 0.54 2 10.12** 108 0.60 (2.956) (13.92) (4.425) PREVALENCE 2 0.118* 126 0.64 0.0550 72 0.54 2 0.145 108 0.58 Matı (0.0639) (0.158) (0.0891) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Standard errors (in parentheses) are robust to heteroskedasticity. All regressions include the observations with zero matches between bankers and politicians in countries with more than two banks, and all regressions control for log real GDP per capita and log population. FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. a. Index of control of corruption, average 1996– 2002, from Kauffman, Kraay, and Mastruzzi (2004). ´as Braun and Claudio Raddatz b. Built by the authors using data from Barth and others (2003). c. Log ratio of average 1996– 2002. Private credit to GDP (from Beck, Demirguc-Kunt, and Levine 2000). Source: Authors’ analysis based on data described in the text. 271 272 THE WORLD BANK ECONOMIC REVIEW regulation is the most dif�?cult to pin down and was the weakest among those reported in the baseline results. Including many diverse countries with the same value of connectedness (zero) clearly reduces the variance of the explana- tory variable and its ability to account for this country characteristic. III. CONCLUDING REMARKS This article builds an extensive dataset to measure the extent to which banks are politically connected across countries. The measure is based on the fact that some high-ranking politicians end up on bank boards of directors. Of course, this represents just one way of establishing relationships between bankers and politicians. It may not even be the most important one, but it is likely to be correlated with other forms. Although formal tests are not pre- sented and causality is not established, the article presents pieces of reduced- form evidence that hold together better as a private interest story than as a public interest story. First, connected banks do better than unconnected ones: they are larger and more pro�?table, and these characteristics are not related to higher risk taking. These results are consistent with those for nonbank �?rms documented in the political economy literature. While a public interest view is still possible (say if politicians were attracted to good banks), in that case poli- ticians would be expected to cluster in the best banks, which should result in a strong relation between the share of politicians on a bank’s board and the bank’s performance. But no such relationship was found: once a bank is con- nected, having more politicians on the board is not associated with better performance. Second, connectedness is more prevalent where deals between bankers and politicians are likely to be less costly and more inﬂuential. Connectedness corre- lates positively with corruption but negatively with government accountability. Third, these politician-banker relationships are associated with poorer out- comes for society in the form of lower overall and �?nancial development. A likely mechanism is regulatory capture, a conjecture supported by the �?nding that bank regulation is more pro-banker and of lower quality where these links are more important. If that is the direction of causality, a permissive insti- tutional context allows banks to achieve better regulatory treatment by con- necting themselves to politicians. These links allow banks to achieve higher pro�?ts without taking more risk or boosting ef�?ciency, in the process incurring high social costs in the form of inhibited �?nancial sector development and reduced access to �?nancing for many �?rms. Restricting these types of connec- tions could limit the ability of incumbent �?nanciers to tilt regulations in their favor and impede �?nancial sector development. It is important, however, not to draw direct, partial equilibrium policy conclusions from this exercise. If this particular avenue of connection is absent, incumbents might instead pressure regulators some other way, such as through outright bribes, that could be even more detrimental to the institutional framework. ´as Braun and Claudio Raddatz Matı 273 REFERENCES Acemoglu, Daron, and Simon Johnson. 2005. “Unbundling Institutions.�? Journal of Political Economy 113 (5): 949–95. Agrawal, Anup, and Charles Knoeber. 2001. “Do some Outside Directors Play a political Role?�? Journal of Law and Economics 44: 179– 98. Ansolabehere, Stephen, John de Figueiredo, and James Snyder, 2003. “Why Is There So Little Money in US Politics?�? Journal of Economic Perspectives 17: 105 –30. Barth, James, Gerard Caprio, and Ross Levine. 2003. “Bank Supervision and Regulation: What Works Best?�? Journal of Financial Intermediation 13: 205–48. ———. 2006. Rethinking Bank Regulation: Till Angels Govern. Cambridge, U.K.: Cambridge University Press. Beck, Thorsten, Asli Demirguc-Kunt, and Ross Levine. 2000. “A New Database on Financial Development and Structure.�? World Bank Economic Review 14 (3): 597–605. Updated and accessed through worldbank.org. Braun, Matias, and Claudio Raddatz. 2007. Trade Liberalization, Capital Account Liberalization, And The Real Effects Of Financial Development. Journal of International Money and Finance 26 (5): 730–61. ———. 2008. “The Politics of Financial Development: Evidence from Trade Liberalization.�? Journal of Finance 63 (3): 1469–508. Bureau van Dijk. 2006. Bankscope. www.bvdinfo.com/Products/Company-Information/International/ BANKSCOPE.aspx Burnside, Craig, and David Dollar. 2000. “Aid, Policies, and Growth.�? American Economic Review 90 (4): 847– 68. How Countries Supervise Their Banks, Insurers and Securities Markets 2000. London: Central Banking Publications, 1999. How Countries Supervise Their Banks, Insurers and Securities Markets 2002. London: Central Banking Publications, 2001. How Countries Supervise Their Banks, Insurers and Securities Markets 2003. London: Central Banking Publications, 2002. Cull, Robert, and Colin Xu. 2005. “Institutions, Ownership, and Finance: The Determinants of Pro�?t Reinvestment among Chinese Firms.�? Journal of Financial Economics 77: 117–46. ´ , Ernesto. 2006. “Regulatory Capture: A Review.�? Oxford Review of Economic Policy 22 (2): Dal Bo 203–25. Dreher, Axel, Michael Lamla, Sarah Lein, and Frank Somogoyi. 2009. “The Impact of Political Leaders’ Profession and Education on Reforms.�? Journal of Comparative Economics 37: 169– 93. Economist Intelligence Unit. Various years. Country Reports, 1996– 2005. London: Economist Intelligence Unit. http://countryanalysis.eiu.com/country_reports.html. Faccio, Mara. 2006. “Politically Connected Firms.�? American Economic Review 96 (1): 369– 86. Faccio, Mara, Ronald Masulis, and John J. McConnell. 2005. “Political Connections and Corporate Bailouts.�? Journal of Finance 61 (6): 2597– 635. Fisman, Raymond. 2001. “Estimating the Value of Political Connections.�? American Economic Review 91 (4): 1095–102. ¨ rg Rocholl, and Jongil So. 2009. “Do Politically Connected Boards Affect Firm Goldman, Eitan, Jo Value?�? Review of Financial Studies 22 (6): 2331–60 Guiso, Luigi, Paola Sapienza, and Luigi Zingales. 2004. “The Role of Social capital in Financial Development.�? American Economic Review 94 (3): 526– 56. Haber, Stephen H. 1991. “Industrial Concentration and the Capital Markets: A Comparative Study of Brazil, Mexico, and the United States,1830–1930.�? Journal of Economic History 51: 559– 80. 274 THE WORLD BANK ECONOMIC REVIEW Heckman, James. 1979. “Sample Selection Bias As a Speci�?cation Error.�? Econometrica 47: 153– 61. Heston, A., R. Summers, and B. Aten. 2006. Penn World Table Version 6.2. University of Pennsylvania, Center for International Comparisons of Production, Income, and Prices, Philadelphia, Pa. Available at http://pwt.econ.upenn.edu. IMF (International Monetary Fund). 2009. Special Data Dissemination Standard. http://dsbb.imf.org/ Pages/SDDS/Home.aspx. Kaufmann, Daniel, Aart Kraay, and Massimo Mastruzzi. 2004. “Governance Matters III: Governance Indicators for 1996, 1998, 2000, and 2002.�? World Bank Economic Review 18 (2): 253– 87. Khwaja, Asim Ijaz, and Atif Mian. 2005. “Do Lenders Favor Politically Connected Firms: Rent Provision in an Emerging Financial Market.�? Quarterly Journal of Economics 120 (4): 1371–411. ´ pez-de-Silanes, Andrei Shleifer, and Robert Vishny. 1997. “Legal determi- La Porta, Rafael, Florencio Lo nants of external �?nance.�? Journal of Finance 52 (3): 1131–50. ———. 1998. “Law and �?nance.�? Journal of Political Economy 106 (6): 1113– 55. Levine, Ross. 2006. “Finance and Growth: Theory and Evidence.�? In Handbook of Economic Growth, Vol. 1A., ed. Philippe Aghion, and Steven Durlauf. Amsterdam: North Holland. Maurer, Noel. 2002. The Power and the Money: The Mexican Financial System, 1876–1932. Palo Alto, CA: Stanford University Press. Maurer, Noel, and Andrei Gomberg. 2005. “When the State Is Untrustworthy: Public Finance and Private Banking in Por�?rian, Mexico.�? Journal of Economic History 64 (4): 1087– 107. Milanovic, Branko, Karla Hoff, and Shale Horowitz. Forthcoming. “Turnover in Power As a Restraint on Investing in Inﬂuence: Evidence from the Post-Communist Transition.�? Economics and Politics. Pagano, Marco, and Paolo Volpin. 2001. “The Political Economy of Finance.�? Oxford Review of Economic Policy 17 (4): 502– 519. ———. 2005. “The Political Economy of Corporate Governance.�? American Economic Review 95 (4): 1005– 30. Rajan, Raghuram, and Luigi Zingales. 2003. “The Great Reversals: The Politics of Financial Development in the Twentieth Century.�? Journal of Financial Economics 69 (1): 5–50. Razo, Armando. Forthcoming. Social Foundations of Limited Dictatorship: Networks and Private Protection during Mexico’s Early Industrialization. Palo Alto, CA: Stanford University Press. Schnell, Rainer, Tobias Bachteler, and Stefan Bender. 2004. “A Toolbox for Record Linkage.�? Austrian Journal of Statistics 33 (1 –2):125– 33. Stulz, Rene, and Rohan Williamson. 2003. “Culture, Openness, and Finance.�? Journal of Financial Economics 70: 313–49. APPENDIX. DISTRIBUTION OF THE NUMBER OF MATCHES UNDER R A N D O M D R AW S Consider a population with NP politicians and NB bankers. The intersection of the two groups consists of NPB banker-politicians. Two samples are taken con- secutively and matched from the population of bankers and politicians with replacement at the sample level,17 the �?rst consisting of nB NB bankers and the second of nP NP politicians. Let X be a random variable that counts the 17. This means that all individuals from the �?rst sample are replaced in the population before taking the second sample, so that an individual from the intersection of the two samples can be drawn twice. ´as Braun and Claudio Raddatz Matı 275 number of matches and is distributed according to: NPB PNPB Àk NPB À k NB À NPB NP À k À i k i¼0 i nB À k À i nP À k PðX ¼ kÞ ¼ NP NB nP nB The denominator corresponds to the ways the two samples of sizes nP and nB can be chosen from populations of sizes NP and NB. The numerator has various components. The �?rst term corresponds to the number of ways in which the k common elements can be chosen among the NPB members of the intersection. The summation that follows counts the number of ways to choose the remaining nP 2 k and nB 2 k terms. The �?rst term counts the ways to choose i of those elements from among the rest of the members of the intersection. If the i terms are chosen in this way, they can be in only one of the samples. For instance, assume that one of the remaining nB 2 k components of nB also NPB À k belongs to NPB. This one term can be chosen in ways, and the 1 remaining nB 2 k terms, which are bankers only, can be chosen in NB À NPB ways. Given that one of the terms in nB 2 k belongs to the inter- nB À k À 1 section, it cannot be selected in the remaining nP 2 k draws from NP, so those NP À k À 1 terms can be chosen in only. nP À k This distribution is used to estimate the expected number of matches in a country considering the actual size of the samples of bankers and politicians available from the data, which pin down nP and nB and assuming that both are drawn from a common pool corresponding to a country’s total population. In the notation above, the assumption of a common pool corresponds to assuming that NP ¼ NB ¼ NPB. In this case the probability of �?nding k matches simpli�?es to N NÀk N À nP k n Àk n Àk PðX ¼ kÞ ¼ P b N N nP nB T A B L E A 1 . Connectedness and Detailed Regulation 276 Entry Requirements Capital Requirements Controls: Log Real GDP, Log Controls: Log Real GDP, Log No Controls Population No Controls Population Number of No. of No. of No. of Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Coef�?cient Observations R2 Measure (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) All Banks FRACBANKS 2 0.334 52 0.002 2 2.293 52 0.06 2 0.842 52 0.017 0.00387 52 0.129 (1.059) (1.646) (0.725) (1.010) SHAREASSETS 0.0731 49 0 2 0.276 49 0.026 2 0.879 49 0.032 2 0.542 49 0.133 (0.689) (0.792) (0.609) (0.606) FRACBANKERS 3.187 52 0.002 2 12.06 52 0.033 2 8.900 52 0.021 2 1.190 52 0.13 THE WORLD BANK ECONOMIC REVIEW (10.05) (13.12) (5.792) (8.839) CONNECTEDNESS 0.0203 52 0.001 2 0.0795 52 0.026 2 0.222** 52 0.127 2 0.149 52 0.15 (0.0930) (0.187) (0.0847) (0.130) PREVALENCE 2 0.0273 52 0.003 2 0.00193 52 0.023 0.112*** 52 0.09 0.0650* 52 0.154 (0.0589) (0.0664) (0.0323) (0.0367) Fully Private Banks Only FRACBANKS 2 0.178 46 0.000 2 1.485 46 0.044 2 0.180 46 0.001 0.269 46 0.109 (1.171) (1.634) (0.766) (0.997) SHAREASSETS 0.106 43 0.000 2 0.379 43 0.027 2 0.358 43 0.004 2 0.380 43 0.103 (1.268) (1.409) (0.675) (0.702) FRACPOLITICIANS 8.438 46 0.009 0.414 46 0.026 2 6.641 46 0.014 2 7.130 46 0.115 (8.671) (14.16) (7.420) (10.38) CONNECTEDNESS 0.0331 46 0.001 2 0.0972 46 0.031 2 0.238** 46 0.148 2 0.194* 46 0.153 (0.121) (0.218) (0.0947) (0.114) MAXSHARE 2 0.0283 46 0.002 0.00911 46 0.026 0.128*** 46 0.109 0.0914** 46 0.153 (0.0683) (0.0791) (0.0355) (0.0420) Measure Activities Restrictions Private Monitoring No Controls Controls: Log Real GDP, Log No Controls Controls: Log Real GDP, Log Population Population Coef�?cient No. of R2 Coef�?cient No. of R2 Coef�?cient No. of R2 Coef�?cient No. of R2 Observations Observations Observations Observations (13) (14) (15) (16) (17) (18) (19) (20) (21) (22) (23) (24) All Banks FRACBANKS 2.812*** 51 0.153 2.655** 51 0.271 2 3.227*** 52 0.226 2 0.994 52 0.391 (0.747) (1.072) (1.018) (1.470) SHAREASSETS 1.211 48 0.05 0.665 48 0.23 2 2.645*** 49 0.265 2 1.724** 49 0.478 (0.751) (0.842) (0.602) (0.662) FRACBANKERS 25.04*** 51 0.137 24.68** 51 0.258 2 35.31*** 52 0.301 2 17.65 52 0.414 (8.104) (11.83) (12.58) (17.24) CONNECTEDNESS 0.325*** 51 0.221 0.298* 51 0.266 2 0.233*** 52 0.125 2 0.169 52 0.403 (0.0787) (0.151) (0.0865) (0.129) PREVALENCE 2 0.120** 51 0.085 2 0.0478 51 0.209 0.103*** 52 0.068 0.0202 52 0.382 (0.0451) (0.0577) (0.0339) (0.0293) Fully Private Banks Only FRACBANKS 3.276*** 46 0.230 3.448*** 46 0.331 2 2.630* 46 0.150 2 0.691 46 0.402 (0.731) (0.719) (1.348) (1.248) Matı SHAREASSETS 1.309 43 0.036 1.086 43 0.210 2 3.480*** 43 0.252 2 2.731*** 43 0.529 (1.051) (1.129) (0.976) (0.702) FRACPOLITICIANS 27.26** 46 0.157 30.72* 46 0.270 2 34.50*** 46 0.254 2 20.47 46 0.438 (12.08) (15.92) (11.71) (15.64) CONNECTEDNESS 0.328*** 46 0.191 0.273* 46 0.237 2 0.340*** 46 0.207 2 0.202 46 0.430 (0.0975) (0.150) (0.0744) (0.130) MAXSHARE 2 0.149*** 46 0.101 2 0.0825 46 0.200 0.132*** 46 0.081 0.0268 46 0.398 (0.0495) (0.0636) (0.0419) (0.0372) 277´as Braun and Claudio Raddatz 278 Overall Supervisory Power No Controls Controls: Log Real GDP, Log Population Coef�?cient No. of Observations R2 Coef�?cient No. of Observations R2 THE WORLD BANK ECONOMIC REVIEW Measure (25) (26) (27) (28) (29) (30) All Banks FRACBANKS 0.310 52 0.002 0.360 52 0.059 (1.022) (1.511) SHAREASSETS 2 0.607 49 0.013 2 0.741 49 0.052 (0.909) (0.924) FRACBANKERS 2 0.842 52 0 2 4.211 52 0.059 (8.829) (13.97) CONNECTEDNESS 0.131 52 0.031 2 0.0293 52 0.058 (0.0942) (0.158) PREVALENCE 2 0.00670 52 0 0.0452 52 0.066 (0.0678) (0.0720) Fully Private Banks Only FRACBANKS 1.119 46 0.021 1.901 46 0.103 (1.011) (1.354) SHAREASSETS 2 0.175 43 0.001 0.0633 43 0.041 (1.509) (1.621) FRACPOLITICIANS 6.595 46 0.007 16.86 46 0.088 (11.08) (19.25) CONNECTEDNESS 0.188 46 0.048 0.0687 46 0.069 (0.121) (0.188) MAXSHARE 2 0.0385 46 0.005 0.0142 46 0.067 (0.0948) (0.101) *Signi�?cant at the 10 percent level; **signi�?cant at the 5 percent level; ***signi�?cant at the 1 percent level. Note: Standard errors (in parentheses) are robust to heteroskedasticity. The dependent variable are the Barth, Caprio, and Levine (2003) principal component indexes of �?ve dimensions of bank regulation: the degree of restrictions to entry, the magnitude of capital requirements, the extent of restric- tions to cross activities, the reliance of self monitoring, and the overall authority of the regulator. All regressions include the observations with zero matches between bankers and politicians in countries with more than two banks, and all regressions control for log real GDP per capita and log popu- lation. FRACBANKS is the fraction of banks with Bankscope data on board of directors that had a former politician on their boards. SHAREASSETS is the share of the total assets of banks with Bankscope data on board of directors that is represented by connected banks. FRACBANKERS is the fraction of bank directors that had a previous political position. PREVALENCE is the (log) ratio of the actual to the expected number of matches, where the expected number is computed assuming no bias toward politicians in the selection of bankers and assuming that both bankers and politicians are selected from the whole population of a country. MAXSHARE is the largest fraction of a country’s population from which politicians and bankers would have to be selected so that the hypothesis that the selection of bankers is not biased toward politicians could not be rejected at the 5 percent level. Source: Authors’ analysis based on data described in the text and table. Matı 279´as Braun and Claudio Raddatz Natural Disasters and Human Capital Accumulation Jesus Crespo Cuaresma The empirical literature on the relationship between natural disaster risk and invest- ment in education is inconclusive. Model averaging methods in a framework of cross- country and panel regressions show an extremely robust negative partial correlation between secondary school enrollment and natural disaster risk. This result is driven exclusively by geologic disasters. Exposure to natural disaster risk is a robust determi- nant of differences in secondary school enrollment between countries but not necess- arily within countries Natural disasters, human capital, education, school enrollment, Bayesian model averaging. JEL codes: Q54, I20, E24, C11. This article quanti�?es the effect of natural disaster risk on investments in edu- cation by exploiting both cross-country and time differences in school enroll- ment. Because of the large number of theories explaining differences in the rate of human capital accumulation across countries, model averaging techniques are used to explicitly take into account model uncertainty in extracting the effect of catastrophic risk on school enrollment. The empirical literature on the economic effects of natural disasters has traditionally concentrated on the short-run effects of catastrophic events (for example, Dacy and Kunreuther 1969; Albala-Bertrand 1993a, b; Tol and Leek 1999; Rasmussen 2004; and Noy 2009). In contrast, Skidmore and Toya (2002) and Crespo Cuaresma, Hlouskova, and Obersteiner (2008) Jesus Crespo Cuaresma ( jcrespo@wu.ac.at) is a professor in the Department of Economics, Vienna University of Economics and Business; a research scholar at the World Population Program, International Institute of Applied Systems Analysis; and a consultant at the Austrian Institute for Economic Research and the World Bank. This study was prepared as a background paper for the joint World Bank– United Nations Assessment on the Economics of Disaster Risk Reduction. The work was supported by the Global Facility for Disaster Reduction and Recovery. The author would like to thank Apurva Sanghi, the coordinator of the assessment, for many intellectually challenging discussions that helped shape the article. The article also pro�?ted from helpful comments by three anonymous referees, ´ Miguel Albala-Bertrand, Jed Friedman, Samir KC, Reinhard Mechler, Norman Loayza, Jose Paul Raschky, Gallina Vincelette, and participants at the Brown Bag Lunch Seminar at the World Bank. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 280– 302 doi:10.1093/wber/lhq008 Advance Access Publication July 9, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 280 Crespo Cuaresma 281 concentrate on long-run effects of disaster risk on the macroeconomy.1 With the exception of some results in Skidmore and Toya, there has been no fully ﬂedged empirical investigation of the effects of natural disasters on human capital accumulation across countries. This study aims to �?ll that gap. Case studies of individual economies have, however, examined the effect of natural disasters on educational attainment. Recently, Kim (2008) used data from the Demographic and Health Surveys and the Living Standard Measurement Study to examine empirically the effects of climate shocks on educational attainment in Burkina Faso, Cameroon, and Mongolia. Kim tends to �?nd negative effects of disaster risk on secondary school completion. From a theoretical perspective, the effect of natural disaster risk on edu- cational investments is ambiguous. Skidmore and Toya (2002) argue that to the extent that natural catastrophes reduce the expected return to physical capital, rational individuals would shift their investment toward human capital.2 But this is just one of the possible effects of natural disasters on human capital. One could also argue that, in a framework of models of agents with �?nite lives, the potential effect of natural disaster risk on mortality would lower education investment in disaster-prone countries. Checchi and Garcı´a-Pen ˜ alosa (2004) present a simple theoretical model assessing the effect of production risk on education in which aggregate production risk determines the average level of education and its distribution. Checchi and Garcı´a-Pen ˜ alosa show both theoretically and empirically that higher output volatility leads to lower educational attainment. If natural disaster risk is inter- preted as a component of aggregate production risk in the economy, countries that are more affected by disasters should also exhibit lower levels of human capital accumulation, ceteris paribus. These types of arguments stem from theoretical models and aim at unveil- ing the role of natural disaster risk as a determinant of cross-country differ- ences. In this sense, these theoretical explanations refer to the long-run effects of natural disasters on education investments. Short-run effects on human capital accumulation associated with the actual occurrence of the dis- aster could be extremely important as well. Consider the 2005 earthquake in Pakistan. The Asian Development Bank and the World Bank (2005), esti- mated that 853 teachers and 18,095 students lost their lives in the disaster. More than 7,500 schools were affected by the earthquake, and the estimated reconstruction costs for education were the second highest by sector, after private housing. To the extent that reconstruction efforts are unable to restore capacity and education infrastructure after a disaster, long-run effects 1. See Okuyama (2009) for a review of the literature on assessing and measuring the economic effects of natural disasters. 2. Skidmore (2001) studies investment decisions under catastrophic risk, but the empirical results are based on a very reduced dataset. 282 THE WORLD BANK ECONOMIC REVIEW may also emanate directly from the losses caused by the disaster. Natural disasters may also affect educational attainment through the effect of evacua- tions and school switching on the dropout rate and academic performance, as Sacerdote (2008) recently investigated using data from evacuations following hurricanes Katrina and Rita in New Orleans (see also Hanushek, Kain, and Rivkin 2004). The literature has also highlighted the effects on human capital investment related to loss of parents and to child labor decisions. Ultimately, the question of how natural disaster risk affects human capital accumulation is an empirical one. Because a single theoretical framework cannot be relied on for explaining the link, an explicit assessment of model uncertainty is called for when quantifying the effect of natural disasters on edu- cation investments. This article uses Bayesian model averaging (BMA) to obtain robust estimates of the effect of disaster risk on secondary school enroll- ment rates (see Raftery 1995 and Clyde and George 2004, for general discus- sions of BMA and Ferna ´ ndez, Ley, and Steel 2001b and Sala-i-Martin, Doppelhofer, and Miller 2004 for applications to the identi�?cation of robust determinants of economic growth). Model averaging ensures that the results are not speci�?c to the choice of model and take into the account not only the uncertainty of the estimates for a given model, but also the uncertainty in the choice of speci�?cation. The results indicate that geologic disaster risk is a robust variable for explaining differences in secondary school enrollment rates across countries. The effect is sizable and well estimated. The school enrollment effect corre- sponding to the mean geologic disaster risk is around 1.65 percentage points in secondary school enrollment compared with a country with zero disaster risk. The maximum disaster risk-driven effect in the dataset implies approximately a 20 percentage point decrease in secondary school enrollment. The article is structured as follows. Section I describes the empirical relation- ship between disaster risk and educational attainment. Section II describes BMA exercises to assess the robust effect of natural disaster risk as a determi- nant of differences in school enrollment rates in both a cross-section and a panel of countries. It also explicitly assesses subsample heterogeneity in the response of human capital accumulation to disaster risk. Section III summarizes the key �?ndings. I . A F I R S T LO O K AT E D U CAT I O N AND DISASTERS This section explores the relationship between natural disasters and human capital accumulation. Figure 1 presents scatterplots of average secondary school enrollment in 1980–2000 (after controlling for income per capita and geographic dummy variables based on world regions) against geologic and climate disasters and for all disasters combined for the 80 countries in the Crespo Cuaresma 283 F I G U R E 1. Natural Disaster Risk and Secondary School Enrollment (unexplained part) Source: Author’s analysis based on data described in the text. 284 THE WORLD BANK ECONOMIC REVIEW empirical analysis.3 Climate-related catastrophes include ﬂoods, cyclones, hur- ricanes, ice storms, snow storms, tornadoes, typhoons, storms, wild �?re, drought, and cold waves; geologic disasters include volcanic eruptions, natural explosions, avalanches, landslides, earthquakes, and wave surges. Following Skidmore and Toya (2002), �?gure 1 concentrates on a simple measure of natural disaster risk based on average disaster occurrence, here normalized by 1 million people. Disaster risk is thus measured as4 di ¼ log½1 þ ðNumber of disasters in country i=Population of country i in millionsÞ: ð1Þ Existing data on quanti�?ed losses and received aid are not used, since such measures are known to be plagued by endogeneity and other measurement pro- blems. On the one hand, to the extent that disaster aid decisions are inﬂuenced by reported losses or number of people affected, governments would have an incen- tive to overreport these �?gures. On the other hand, a country’s income level (which is highly correlated with human capital accumulation) is a basic determi- nant of the effectiveness of natural disaster risk management. Since successful risk management mechanisms will reduce the negative macroeconomic effects of disas- ters, using estimated losses could lead to a spurious negative correlation between disaster risk and education when the real correlation is between education and the reduction in natural disaster loss. Skidmore and Toya (2007), for instance, show that higher levels of education reduce the losses from natural disasters. The pro- blems related to the use of reported losses from natural disasters have been noted in the recent comparative literature. Guha-Sapir and Below (2002) highlight some of these problems and conclude that existing datasets on the socioeconomic impact of disasters are unsatisfactorily de�?ned and incomplete. Vulnerability to natural disasters can be thought of as comprising risk exposure as well as the ability to cope with disaster shocks. The disaster variable used in this analysis proxies exclusively the �?rst vulnerability component and thus is free of information on the ability to resist and recover from a natural dis- aster. Variables such as total estimated loss as a share of GDP or number of people injured or killed combine aspects of both vulnerability components. In this context, it would be dif�?cult to argue that human capital does not affect the second component, the ability to cope with disaster shock. This would raise 3. The choice of countries is determined exclusively by data availability. The 80 countries in the scatterplot are those for which data on all variables used in the Bayesian model averaging analysis are available. 4. The source of disaster data is the Emergency Events Data Base (EM-DAT), which reports on catastrophic events that meet at least one of the following criteria: 10 or more people reported killed, 100 people reported affected, a call for international assistance was issued, or a state of emergency was declared (CRED 2004). Crespo Cuaresma 285 serious doubts about the empirical study unless good instruments were found to identify the exogenous component of disaster risk, a task that is extremely dif�?- cult in practice. Instead, the analysis concentrates on measures based on the fre- quency of disaster occurrence that do not contain information on the magnitude of the disaster, thus ful�?lling the necessary condition of exogeneity. Figure 1 shows a weak positive relationship between disaster risk and the education variable for total disasters, which disappears when the data are dis- aggregated into subgroups of climate and geologic disasters. Although the relationship is not statistically signi�?cant in any of the three cases reported in �?gure 1, this �?rst glimpse at the relationship of interest seems to support the conclusions in Skidmore and Toya (2002) for the aggregated data. To extract the pure effect of disaster risk on education investment, however, other variables that independently affect differences in educational attainment across countries must be controlled for. Learning about the pure impact of natural disasters on education implies formulating a potentially large model that hypothesizes that a human capital accumulation measure depends on a set of determinants and natural disaster risk. Obviously, the choice of extra con- trols for a model linking disaster risk to human capital accumulation depends on the theoretical setting. The literature presents many competing theories and effects to explain cross-country differences in educational attainment when assessing empirically the determinants of human capital accumulation. So that the empirical results do not depend on a speci�?c theoretical (and thus econo- metric) speci�?cation or a particular choice of controls, BMA methods are used to investigate the robustness of disaster risk as a determinant of educational attainment in the framework of model uncertainty. Model averaging methods present a consistent framework to quantitatively assess model uncertainty when studying problems too ambiguous or theoretical complex to be convin- cingly represented by a single speci�?cation. II . A N E M PI R I CA L A N A LY S I S O F TH E EF FE C T O F DI S A S T E R R I S K ON H U M A N CA P I TA L AC C U M U L AT I O N This section assesses natural disaster risk as a determinant of differences in school enrollment in both a cross-section and a panel of countries using BMA. It also assesses subsample heterogeneity in the response of human capital accumulation to disaster risk. Model Uncertainty The effect of catastrophic risk on human capital accumulation is estimated using linear econometric models of the type: X K ð2Þ ei ¼ a þ bdi þ gj xj þ 1i; ; j ¼1 286 THE WORLD BANK ECONOMIC REVIEW where ei is a proxy for educational attainment, di is the disaster risk variable, X ¼ (x1 . . . xK) are other explanatory variables, and 1 is a zero-mean error term with variance equal to s2. In Skidmore and Toya (2002), for instance, the initial level of the educational variable and income per capita are the only vari- ables in the X set. Because numerous variables affect educational attainment, the aim is to obtain a measure that summarizes the effect of natural disaster risk on human capital accumulation after taking into account the degree of uncertainty embodied in speci�?cation (2) when the size of the model and the nature of the variables in X that belong to the model are unknown. BMA presents a consistent framework for assessing the dimension of model uncertainty highlighted above.5 Consider a set of K variables, X of which are potential determinants of educational attainment in a cross-country regression framework, so that the stylized speci�?cation considered is given by equation (2) ¯ for K K. In this situation, there are 2K possible combinations of regressors, each de�?ning a model Mk. The Bayesian approach implies considering model speci�?cation itself as a quantity to be estimated. In this sense, it follows immediately that, by Bayes’s theorem, PðYjMk ÞPðMk Þ ð3Þ PðMk jYÞ ¼ ; P2K PðYjMm ÞPðMm Þ m¼1 which indicates that the posterior probability of model Mk (the probability that the model is the true one given data Y) is related to its marginal likelihood, P(Y j Mk), and its prior probability, P(Mk), as compared with the other models in the model space. Following Ferna ´ ndez, Ley, and Steel (2001a), an improper diffuse prior is set on a and s, coupled with Zellner’s (1986) g-prior on the parameter vector, which implies that 1 ð4Þ Pða; b; gj ; sjMk Þ1 Nkþ1 ð0; s2 ðgX0j Xj ÞÀ1 Þ s where Nkþ1 is a multivariate normal distribution of dimension k þ 1, and Xj is a matrix whose columns are given by the independent regressors in model Mk. This setting implies that the Bayes factor (ratio of marginal likelihoods) for two competing models, M0 and M1, is given by ðk1 Àk0 Þ=2 ÀðNÀ1Þ=2 PðY jM1 Þ g 1 þ g À R2 1 ð5Þ B1;0 ¼ ¼ PðY jM0 Þ gþ1 1 þ g À R2 0 Where N is the sample size, kj is the dimension of model j, and R2 j is the stan- dard coef�?cient of determination for model j. Some particular values of g, the 5. Raftery (1995) and Clyde and George (2004) present general discussions of the use of BMA in linear regressions. Crespo Cuaresma 287 hyperparameter governing the prior over the slopes, have been systematically used in the literature. For g ¼ 1/N (the unit information prior), the Bayesian information criterion should be used in forming Bayes factors (see, for example, Kass and Wasserman 1995 and Kass and Raftery 1995), and thus BMA weights, while the risk inﬂation criterion (Foster and George, 1994) sets g ¼ 1/K 2.6 P(MkjY) can be used to build an estimate of the quantity of interest as, say, a weighted average of all estimates of b, where the weights are given by the posterior probability of each model from which the estimate was obtained, X ð6Þ EðbjY Þ ¼ EðbjY ; Mk ÞPðMk jY Þ: k Similarly, model averaged estimates of the posterior variance of b can be com- puted from the model averaged variance of the estimate, which in this setting summarizes information about the precision not only for a given model, but also across models. While the method has been put forward in the setting of a cross-sectional dataset, it can be generalized to panel data in a straightforward fashion using the Frisch-Waugh-Lovell theorem. In particular, in models with cross-sectional �?xed effects, the method can be applied to deviations of the mean for each cross-section (the within transformation) or to the mean of the cross-sections for each period when �?xed period effects are assumed. The method is used here to estimate the effects of natural disaster risk on secondary school enroll- ment, which are robust to model uncertainty. In addition to the distribution of the estimated parameter, also of interest here is whether the data support the inclusion of natural disaster risk in speci�?cations explaining differences in sec- ondary school enrollment. This characteristic can be estimated by summing the posterior probability of the models containing the natural disaster variable, a statistic referred to as the posterior inclusion probability of the variable. The Empirical Setting A group of variables identi�?ed in the literature as important determinants of differences in human capital accumulation across countries are added as poten- tial regressors in speci�?cation (2). The focus is on secondary school enrollment as the variable of interest, and thus the analysis aims to explain the ﬂow of human capital (its accumulation) rather than its stock (which is usually measured by mean years of schooling). This focus is justi�?ed because primary schooling is compulsory in most countries in the sample and because the most important results for the issue under study were obtained using gross secondary school enrollment as the human capital variable (Skidmore and Toya 2002). 6. Ferna´ ndez, Ley, and Steel (2001a) recommend using a benchmark prior based on the size of the group of potential regressors compared with sample size, so that g ¼ 1/max(K2,N). 288 THE WORLD BANK ECONOMIC REVIEW Table 1 presents the regressors used in the BMA exercise. As potential expla- natory variables, proxies for initial income ( y0) and initial school enrollment (e0) account for wealth-induced human capital accumulation effects and for the observed persistence of human capital accumulation variables across and within countries and for their potential convergence across countries. National Gini coef�?cients capture differences in income distribution across economies, and the standard deviation of annual GDP growth rates is used as a proxy in analyzing the potential effect of macroeconomic instability. Life expectancy at birth in the initial period controls for differences in health. Credit constraints are included in the model using domestic credit to the private sector as a per- centage of GDP as a proxy for �?nancial depth. The quality of political institutions is controlled for with the help of the Polity IV database, which offers a score variable ( polity2) that quanti�?es a country’s political system based on competitiveness and openness of executive recruitment, constraints on the chief executive, regulation, and competitiveness of participation. The polity2 measure ranges from –10 to þ 10, where –10 implies a strongly autocratic regime and þ 10 a strongly democratic regime. The models also control for war in a given country during the period under study. T A B L E 1 . Variables and De�?nitions Variable Description Source e Gross secondary school enrollment, average World Bank 2006 1980– 2000 e0 Initial gross secondary school enrollment, 1980 World Bank 2006 y0 Initial level of GDP per capita, 1980 World Bank 2006 gini Gini index for income World Bank 2006 life0 Life expectancy, 1980 World Bank 2006 vol Volatility of GDP per capita growth World Bank 2006 polity Polity 2 indicator Marshall and Jaggers 1995 pavroad Percentage of paved roads World Bank 2006 cred Credit to private sector (percent of GDP) World Bank 2006 area Land area World Bank 2006 popdens Population density, 1980 World Bank 2006 inv Investment in physical capital, 1980 Heston, Summers, and Aten 2006 war Dummy variable for occurrence of war — laam Dummy variable for Latin America and Caribbean — asia Dummy variable for Asia and Paci�?c — safrica Dummy variable for Sub-Saharan Africa — nafrica Dummy variable for North Africa and Middle East — Disaster risk, based on total disasters per million CRED 2004 inhabitants Disaster risk, based on climate disasters per million CRED 2004 inhabitants Disaster risk, based on geologic disasters per million CRED 2004 inhabitants Crespo Cuaresma 289 To control for the effect of country characteristics other than disaster risk on human capital investment, variables measuring total area and population density are included. Physical investment as a percentage of GDP is also con- sidered as a potential determinant of human capital accumulation, to capture the complementarity or substitutability effects of physical and human capital. Regional dummy variables (for Asia and Paci�?c, Latin America and the Caribbean, North Africa and the Middle East, and Sub-Saharan Africa are added to the set of potential determinants of enrollment rates. The cross- country dataset contains data on all 80 countries for which all variables in table 1 are available.7 Table 2 presents descriptive statistics for all variables over 1980–2000, as well as for the dataset divided into 10- and 5-year periods. Several empirical studies of the determinants of schooling have used these variables in econometric models. Flug, Spilimbergo, Wachtenheim (1998), for instance, assess the effect of macroeconomic volatility on investment in edu- cation and present models that control for income inequality, credit market development, initial per capita income, and initial education levels. Some studies have noted the importance of social and political institutions as factors affecting human capital accumulation (Stijns 2006). The results of the BMA exercise, obtained by averaging over the full model space, are presented in table 3.8 Before the analysis, the variables were standar- dized by subtracting the mean and dividing by the standard deviation, so the resulting parameter estimates should be interpreted as the effect of increasing the variable by one standard deviation. The table reports the posterior inclusion probability of each variable computed as the sum of the posterior probability of the models including that variable plus the mean of the posterior distribution of the parameter attached to the variable and its standard devi- ation. The posterior inclusion probability can be interpreted as the probability that a given variable belongs to the true model. Explanatory variables are classi�?ed as robust if the probability that the variable belongs to the model increases is higher than the prior inclusion probability of the variable. For the BMA results in table 3, a diffuse prior was imposed over the model space, so 7. The countries in the sample are Algeria, Australia, Austria, Bangladesh, Belgium, Bolivia, Botswana, Brazil, Burkina Faso, Burundi, Cameroon, Canada, Central African Republic, China, Colombia, Costa Rica, Co ˆ te d’Ivore, Denmark, Dominican Republic, Ecuador, Egypt, El Salvador, Finland, France, Gambia, Ghana, Greece, Guatemala, Honduras, India, Indonesia, Iran, Ireland, Israel, Italy, Jamaica, Japan, Jordan, Kenya, Republic of Korea, Lesotho, Malawi, Malaysia, Mali, Mauritania, Mexico, Morocco, Nepal, Netherlands, New Zealand, Nicaragua, Niger, Nigeria, Norway, Pakistan, Panama, Papua New Guinea, Paraguay, Peru, Philippines, Portugal, Rwanda, Senegal, Sierra Leone, Singapore, Spain, Sri Lanka, Sweden, Switzerland, Thailand, Trinidad and Tobago, Tunisia, Turkey, Uganda, United Kingdom, United States, Uruguay, Repu ´ blica Bolivariana de Venezuela, Zambia, and Zimbabwe. 8. In many other applications, the size of the model space renders the computation of all models intractable, and Markov Chain Monte Carlo methods tend to be used to reduce the number of models to be estimated. 290 T A B L E 2 . Descriptive Statistics Cross-country data Ten-year panel Five-year panel Standard Standard Standard Variable Mean Median Maximum Minimum deviation Mean Median Maximum Minimum deviation Mean Median Maximum Minimum deviation e 58.721 54.217 125.871 6.226 35.124 61.660 59.396 160.763 5.618 35.913 59.146 55.828 160.763 3.572 35.882 e0 44.919 38.777 104.812 2.697 30.285 49.592 43.370 119.509 2.697 31.483 52.942 47.527 142.488 2.697 33.520 y0 8.333 8.334 10.088 6.579 1.065 8.406 8.447 10.223 6.526 1.091 8.421 8.466 10.284 6.522 1.102 life0 61.118 61.920 76.092 35.403 11.138 62.378 65.614 78.837 0.000 12.532 63.481 66.417 79.531 35.196 11.401 vol 3.771 3.657 15.327 1.160 2.086 2.990 2.492 8.706 0.581 1.744 3.110 2.389 27.554 0.258 2.490 THE WORLD BANK ECONOMIC REVIEW polity 1.088 1.500 10.000 2 10.000 7.736 2.145 6.000 10.000 2 10.000 7.588 2.671 6.000 10.000 2 9.000 7.343 pavroad 45.039 39.567 100.000 4.657 27.264 45.568 45.808 100.000 4.300 26.471 45.504 45.808 100.000 4.300 26.095 gini 42.205 40.555 63.010 24.700 10.296 41.957 40.150 63.010 24.700 9.910 41.958 40.270 63.010 24.700 9.939 cred 35.935 29.135 122.146 0.965 25.990 41.100 30.873 175.731 0.000 33.361 41.872 30.683 180.509 0.965 34.805 area 0.979 0.296 9.327 0.001 2.100 0.930 0.294 9.327 0.001 1.997 0.972 0.296 9.327 0.001 2.063 popdens 0.132 0.045 3.603 0.002 0.408 0.131 0.054 4.084 0.002 0.363 0.139 0.052 4.548 0.002 0.423 inv 23.113 23.290 56.490 1.470 10.898 20.546 19.705 56.100 3.030 10.697 21.307 21.320 56.490 1.470 10.840 dt 1.004 0.912 2.394 0.000 0.577 0.081 0.048 0.395 0.000 0.080 0.084 0.054 0.477 0.000 0.087 dc 0.875 0.739 2.334 0.000 0.526 0.066 0.041 0.377 0.000 0.067 0.068 0.044 0.445 0.000 0.075 dg 0.279 0.067 1.781 0.000 0.410 0.017 0.002 0.195 0.000 0.034 0.018 0.000 0.322 0.000 0.039 Source: Author’s analysis based on data described in the text. T A B L E 3 . Bayesian Model Averaging Results for Cross-section of Countries Total disasters Climate disasters Geologic disasters Variable PIP PM PSD PIP PM PSD PIP PM PSD eo 0.999 0.609 0.074 0.999 0.610 0.076 0.999 0.615 0.077 y0 0.958 0.259 0.096 0.859 0.191 0.106 0.880 0.202 0.106 life0 0.716 0.15 0.12 0.934 0.237 0.102 0.920 0.228 0.104 vol 0.195 2 0.00 0.022 0.184 2 0.00 0.022 0.201 2 0.00 0.024 polity 0.158 0.005 0.021 0.108 0.000 0.013 0.108 0.000 0.013 pavroad 0.150 2 0.00 0.014 0.145 2 0.00 0.014 0.140 2 0.00 0.014 gini 0.127 0.002 0.018 0.117 0.001 0.016 0.122 0.002 0.017 cred 0.891 2 0.10 0.053 0.815 2 0.09 0.057 0.840 2 0.09 0.057 war 0.149 0.004 0.016 0.122 0.002 0.012 0.119 0.001 0.012 area 0.13 0.002 0.012 0.153 0.004 0.015 0.148 0.003 0.014 popdens 0.35 2 0.01 0.029 0.24 2 0.01 0.023 0.274 2 0.01 0.026 inv 0.138 0.003 0.014 0.136 0.003 0.014 0.136 0.003 0.014 safr 0.567 2 0.07 0.080 0.265 2 0.02 0.061 0.264 2 0.02 0.060 nafr 0.174 0.004 0.020 0.170 0.004 0.019 0.164 0.003 0.018 asia 0.224 2 0.01 0.037 0.250 2 0.01 0.042 0.234 2 0.01 0.041 laam 0.945 2 0.12 0.051 0.996 2 0.15 0.043 0.987 2 0.14 0.047 total disasters, dt 0.318 2 0.01 0.033 — — — — — — clim. disasters, dc — — — 0.134 2 0.00 0.015 — — — geol. disasters dg — — — — — — 0.868 2 0.08 0.049 g-prior BIC BIC BIC Prior model size 8.5 8.5 8.5 Number of observations 80 80 80 Number of models 131,072 131,072 131,072 PIP is posterior inclusion probability, PM is posterior mean, PSD is posterior standard deviation, and BIC is Bayesian information criterion. Crespo Cuaresma Note: Values in italics have a PIP higher than 0.5. Source: Author’s analysis based on data described in the text. 291 292 THE WORLD BANK ECONOMIC REVIEW ¯ that P(Mf ) ¼ 1/2K for all f, implying an average prior model size of K ¯ /2 and a prior inclusion probability of 0.5 for each regressor. The results in table 3 include a group of regressors with a different natural disaster risk proxy in each set of covariates (all disasters, climate disasters, and geologic disasters). The initial educational attainment variable and the initial per capita income are highly robust in explaining secondary education enroll- ment. The parameter attached to initial educational attainment is estimated very precisely, and its posterior distribution has a mean below unity, implying (conditional) convergence in secondary school enrollment levels across countries. Initial income level and life expectancy also appear as robust deter- minants of school enrollment, with positive effects that are estimated with good precision. The regional dummy variables for Latin American countries is robust and negatively related to school enrollment, while that for Sub-Saharan Africa is marginally robust in one of the two settings. The results for the credit variable, which are robust but negatively related to school enrollment, are sur- prising and counterintuitive; they seem to be caused by the credit variable’s high correlation with the regional dummy variables. In other settings that excluded the regional dummy variables and single variables, the credit variable was no longer robust, while all other results were unchanged. The effects of the other nondisaster variables were neither robust in posterior inclusion prob- ability nor estimated with precision. The results for the natural disaster risk variables shed light on the channels between human capital accumulation and catastrophic risk. When data on all disasters or climate disasters are used, the implied risk levels do not appear to be robustly linked to school enrollment. For geologic disasters, however, the risk variable is robust and negatively linked to educational attainment, and the effect is well estimated, with a ratio of posterior mean to posterior standard deviation of around 1.7.9 The results imply that the decline in secondary school enrollment for the mean country associated with geologic disaster risk is around 2.13 percentage points higher than that for a country with zero disaster risk. The maximum disaster risk–driven effect implies an approximately 13.6 percentage point decline in secondary school enrollment. Table 4 shows the results for the estimated models with the highest posterior probability. The setting with geologic disasters as the natural disaster risk vari- able belongs to the best model for posterior probability, which would have been the chosen speci�?cation had model selection been used instead of model 9. Although the ratio of the posterior mean to the posterior standard deviation is often used as a measure of precision in estimating the effect of an independent variable on the dependent variable, the usefulness of this statistic depends on the shape of the posterior distribution of the corresponding parameter. This is more the case if posterior distributions based on the full model space (and thus with a mass point at zero) are used instead of those computed using only models that include a given variable. Results that concentrate only on models including a given variable are not qualitatively different from those presented here (available from the author on request). Crespo Cuaresma 293 T A B L E 4 . Single Speci�?cations with Highest Posterior Probability Variable Best model 1a Best model 2b Intercept 0.000 (0.028) 0.000 (0.027) e0 0.598*** (0.068) 0.594*** (0.066) y0 0.227*** (0.069) 0.261*** (0.068) life 0.257*** (0.074) 0.240*** (0.072) cred 2 0.11*** (0.040) 2 0.12*** (0.039) laam 2 0.15*** (0.031) 2 0.11*** (0.034) Geolog. disasters — 2 0.07** (0.032) Adjusted R2 0.936 0.940 Obs. 80 80 *** Signi�?cance at the 1 percent level; **signi�?cant at the 5 percent level. Note: Numbers in parentheses are standard errors. a. Model with the highest posterior probability in the Bayesian model averaging (BMA) setting corresponding to columns 1 and 2 in table 2. b. Model with the highest posterior probability in the BMA setting corresponding to column 3 in table 2. Source: Author’s analysis based on data described in the text. averaging. In this speci�?cation, the effect of geologic disaster risk on enrollment is negative and signi�?cant. Climate and geologic disasters have several differential characteristics that can be helpful in understanding and interpreting the results of the BMA analy- sis. Climate disasters, which tend to occur at regular intervals, are more pre- dictable than geologic disasters, and their damage tends to be linked to physical capital, whereas geologic disasters affect primarily human lives.10 While economists have traditionally discussed the economic impact of natural disaster risk in terms of behavioral effects (related to the discounting of future utility or income) in the framework of theoretical models, several other chan- nels link natural disaster risk to educational attainment on both the supply and demand sides. Damage to schools and other infrastructure, and teacher casual- ties, are obvious factors affecting the supply of education in the aftermath of a natural disaster. On the demand side, in addition to the potential indirect chan- nels linking natural disaster risk with educational attainment through income, several studies show that children who lose a parent tend to have lower invest- ment in human capital, after controlling for other differences (see Gertler, Levine, and Ames 2004). In this sense, the results can be interpreted as sup- porting the belief that the effects of natural disasters on human capital accumu- lation work through increased mortality risk. Apart from the fact that human losses affect educational attainment at the aggregate level through the increased mortality of educated individuals in disaster-prone countries, human losses also 10. Skidmore and Toya (2002) interpret the climate disaster group as proxying risks related to physical capital and the geologic disaster groups as proxying risks related to human life. 294 THE WORLD BANK ECONOMIC REVIEW have an effect on child labor decisions, in particular since empirical results show that child labor is used to counteract short-run income shocks to the household (see Duryea, Lamb, and Levison 2007 for evidence from Brazil). Education and Disasters: Panel Setting The results indicate that natural disaster risk is a robust variable for explaining differences in secondary school enrollment across countries. The question natu- rally arises whether these effects are also observable within countries. Does the occurrence of a natural disaster reduce schooling rates immediately, so that the effect captured in the econometric analysis is a direct consequence of the disas- ter? Variation in disaster risk within countries could provide information on the direct effect of disasters instead of the effect of ex ante disaster risk. Thus, a clearer picture of the differential effect of disaster risk and disaster incidence might be obtained by complementing the cross-country results with time vari- ation in disaster incidence. To assess this possibility, the analysis was conducted again, this time using two panels based on 5- and 10-year subperiods. Because of the dynamic nature of the speci�?cation (the lagged dependent variable is potentially part of the model), estimation using country �?xed effects would lead to biased estimates. Instead, the model is estimated based on the pooled dataset using period �?xed effects. X K ð7Þ eit ¼ a þ bdit þ gj x jt þ 1it; ; j¼1 ð8Þ 1it ¼ lt þ nit ; where the error term 1it, can now be decomposed into a �?xed time effect common to all countries (lt), which summarizes common shocks to the edu- cation variable, and the usual error term with constant variance (vit,). The results reveal that the robust negative effect of natural disaster risk on human capital accumulation found in the cross-country regressions disappears when the focus is exclusively on shorter run variation in school enrollment (table 5). Although the sign of the parameter for geologic disasters remains negative, it is estimated with low precision and has an inclusion probability below 0.5. The inclusion probability of the disaster variables, particularly the geologic disaster variable, increases as the horizon under consideration moves toward long-run comparisons. These results provide an interesting insight into the determinants of human capital accumulation in the short and medium runs. The posterior inclusion probabilities of the variables for the 5-year panel show that, apart from the natural persistence of human capital accumulation variables, only income is an important determinant of secondary school enroll- ment rate differences. For the 10-year panel, life expectancy appears as an additional robust variable in explaining schooling differences. T A B L E 5 . Bayesian Model Averaging Results for Panel Setting Variable PIP PM PSD PIP PM PSD PIP PM PSD Five-year panel eo 0.999 0.844 0.039 0.999 0.848 0.038 0.999 0.848 0.038 y0 0.899 0.114 0.052 0.897 0.112 0.051 0.896 0.112 0.052 life0 0.177 0.012 0.033 0.173 0.011 0.032 0.174 0.012 0.032 vol 0.062 2 0.00 0.004 0.062 2 0.00 0.004 0.062 2 0.00 0.004 polity 0.262 0.011 0.022 0.218 0.008 0.019 0.221 0.008 0.019 pavroad 0.068 2 0.00 0.004 0.068 2 0.00 0.004 0.068 2 0.00 0.004 gini 0.067 2 0.00 0.005 0.069 2 0.00 0.005 0.068 2 0.00 0.005 cred 0.078 2 0.00 0.007 0.076 2 0.00 0.007 0.076 2 0.00 0.007 war 0.072 2 0.00 0.005 0.074 2 0.00 0.005 0.074 2 0.00 0.005 area 0.290 0.008 0.015 0.311 0.009 0.016 0.309 0.009 0.016 popdens 0.085 2 0.00 0.007 0.082 2 0.00 0.006 0.082 2 0.00 0.006 inv 0.115 0.002 0.010 0.117 0.002 0.010 0.117 0.002 0.010 safr 0.114 2 0.00 0.013 0.099 2 0.00 0.012 0.099 2 0.00 0.012 asia 0.077 0.000 0.006 0.075 0.000 0.006 0.075 0.000 0.006 laam 0.182 2 0.00 0.013 0.206 2 0.00 0.014 0.202 2 0.00 0.014 nafr 0.182 0.004 0.012 0.170 0.004 0.012 0.169 0.004 0.011 total disasters, dt 0.300 2 0.00 0.016 — — — — — — clim. disasters, dc — — — 0.061 0.000 0.004 — — — geol. disasters dg — — — — — — 0.081 2 0.00 0.005 g-prior BIC BIC BIC Prior model size 8.5 8.5 8.5 Number of observations 292 292 292 Number of models 131,072 131,072 131,072 Ten-year panel eo 0.999 0.610 0.082 0.999 0.609 0.082 0.999 0.610 0.082 y0 0.900 0.232 0.111 0.902 0.233 0.111 0.900 0.232 0.111 Crespo Cuaresma life0 0.554 0.077 0.084 0.550 0.076 0.084 0.554 0.077 0.084 (Continued) 295 296 TABLE 5. Continued Variable PIP PM PSD PIP PM PSD PIP PM PSD vol 0.104 2 0.00 0.018 0.104 2 0.00 0.018 0.104 2 0.00 0.018 polity 0.273 0.022 0.045 0.263 0.021 0.044 0.273 0.022 0.045 pavroad 0.124 2 0.00 0.014 0.124 2 0.00 0.015 0.124 2 0.00 0.014 gini 0.097 0.000 0.015 0.098 0.000 0.015 0.097 0.000 0.015 cred 0.098 2 0.00 0.016 0.098 2 0.00 0.016 0.098 2 0.00 0.016 war 0.107 2 0.00 0.013 0.106 2 0.00 0.013 0.107 2 0.00 0.013 area 0.339 0.021 0.035 0.345 0.021 0.035 0.339 0.021 0.035 popdens 0.095 2 0.00 0.016 0.094 2 0.00 0.016 0.095 2 0.00 0.016 inv 0.094 0.001 0.014 0.094 0.001 0.014 0.094 0.001 0.014 safr 0.191 2 0.01 0.043 0.192 2 0.01 0.043 0.191 2 0.01 0.043 THE WORLD BANK ECONOMIC REVIEW asia 0.130 0.001 0.021 0.131 0.001 0.021 0.130 0.001 0.021 laam 0.324 2 0.02 0.041 0.340 2 0.02 0.042 0.324 2 0.02 0.041 nafr 0.241 0.013 0.031 0.241 0.013 0.031 0.241 0.013 0.031 total disasters, dt 0.156 2 0.00 0.019 — — — — — — clim. disasters, dc — — — 0.088 0.000 0.010 — — — geol. disasters dg — — — — — — 0.156 2 0.00 0.019 g-prior BIC BIC BIC Prior model size 8.5 8.5 8.5 Number of observations 292 292 292 Number of models 131,072 131,072 131,072 PIP is posterior inclusion probability, PM is posterior mean, PSD is posterior standard deviation, and BIC is Bayesian information criterion. Note: Values in italics have a PIP higher than 0.5. All models include period �?xed effects. Source: Author’s analysis based on data described in the text. Crespo Cuaresma 297 The models were also estimated using country and period �?xed effects, but excluding the initial level of schooling from the pool of potential explanatory variables. The results are unchanged for natural disaster risk but differ for other explanatory variables. In particular, the BMA estimate of the effect of credit to the private sector is very robust and positively related to schooling, implying that credit constraints have a strong inﬂuence on medium-run human capital accumulation dynamics. A comparison of this result to the previous estimates implies that credit constraints are a robust determinant of schooling within countries but not necessarily across countries. These results complement those of Flug, Spilimbergo, and Wachtenheim (1998).11 Parameter Heterogeneity and Interaction Effects An important question is whether the effect of natural disaster risk on human capital accumulation depends on other country characteristics. Studies have found that the effects of natural disaster risk on several macroeconomic vari- ables are modulated by institutional and economic factors. Noy (2009) shows that the GDP costs depend on the strength of a country’s institutions, as well as on the level of income per capita. Similarly, Crespo Cuaresma, Hlouskova, and Obersteiner (2008) �?nd that the potential positive effects of disasters on technology imports exist only for more developed countries, not for poor econ- omies. The usual approach to assessing heterogeneity in elasticities is to include interaction terms. In this case, the class of models considered for the cross- country case is given by X K ð9Þ ei ¼ a þ bdi þ hdi zi þ gj xj þ 1i; ; j¼1 where variable z (in this case, z [ X, although that need not be so in all cases) is responsible for explaining differences in the elasticity of school enrollment to disaster risk. There is some debate in the literature on how to treat interaction terms in the framework of variable selection and BMA. While some analysts include the interaction as an extra linear covariate in the model, without setting any par- ticular prior structure on models including the product of variables (see Masanjala and Papageorgiou 2008), others provide special treatment to models with interaction terms (see Chipman 1996 for a general discussion and Crespo Cuaresma, Doppelhofer, and Feldkircher 2008 and Crespo Cuaresma forth- coming for applications). The main problem in interpreting BMA results when the interaction term is considered a standard variable and the model averages over all possible combi- nations of variables is that some estimates will be based on models that include the interaction terms but do not specify the main effect of the interacted 11. The detailed results are available from the author. 298 THE WORLD BANK ECONOMIC REVIEW variables (the “parent variables�?). This can lead to improper interpretation of the interaction effect, since the absence of the parent variables in the speci�?ca- tion implies that the interaction term may actually be capturing the direct effect of one or both of the parent variables. In this sense, if the aim is to ful�?ll Chipman’s (1996) strong heredity principle, only models that include both the interaction term and the parent terms should be considered. For instance, in a more general setting, with standard variables and an interaction term (consist- ing of variables from the former group), standard BMA would imply averaging over all possible combinations of these variables. But the strong heredity prin- ciple requires excluding model speci�?cations that include the interaction term without the parent variables, which means that 2K-1 þ 2K-3 models would be evaluated. Both approaches are applied to the dataset to evaluate the existence of sub- sample heterogeneity in the effects of natural disaster risk on human capital. Different model spaces are evaluated, each containing potential interactions of the disaster variable with the initial level of school enrollment, the level of income per capita, the political regime, and the degree of credit constraint. Thus, BMA estimates are alternatively obtained for model spaces de�?ned by the speci�?cation in equation (9) with the interaction variable z given by each one of these covariates. Table 6 presents the posterior inclusion probability, posterior mean, and posterior standard deviation for the interaction terms for model spaces comprising all combinations of all possible variables plus the interaction term and for model spaces respecting the strong heredity prin- ciple.12 Several interesting results emerge. There is little evidence for robust het- erogeneous effects of natural disasters on education. In the results obtained by imposing the strong heredity principle, the only interaction with a posterior inclusion probability higher than 0.5 is for the combined effect of geologic dis- asters and political regime ( polity) in the cross-section setting. The BMA esti- mate indicates that, ceteris paribus, school enrollment is more sensitive to natural disasters in democratic countries. A similar negative effect is found in the 10-year panel using the standard BMA prior across models instead of the strong heredity prior. Other Robustness Checks Other robustness checks were also performed to ensure that the results are not driven by the prior structure imposed on the BMA procedure. The results are robust to changing the parameter prior from the unit information prior to the risk inﬂation criterion as well as to the use of a hyperprior on model size as proposed by Ley and Steel (2009). For the cross-country setting, BMA was con- ducted on an alternative set of covariates, enlarging the group of explanatory 12. Complete results for all other variables are available from the author. The results presented in previous sections are not qualitatively affected by the inclusion of the interaction terms as extra variables. Crespo Cuaresma 299 T A B L E 6 . Bayesian Model Averaging Results for Interaction Terms Standard Bayesian model averaging Strong heredity priora Variable PIP PM PSD PIP PM PSD Cross-section of countries Total disasters* eo 0.314 2 0.02 0.048 0.048 2 0.00 0.023 Clim. disasters * eo 0.185 2 0.01 0.031 0.025 2 0.00 0.018 Geol. disasters * eo 0.266 2 0.01 0.039 0.095 0.000 0.020 Total disasters* y0 0.371 2 0.05 0.154 0.122 2 0.05 0.180 Clim. disasters * y0 0.155 2 0.01 0.069 0.031 2 0.00 0.072 Geol. disasters * y0 0.636 2 0.18 0.337 0.336 2 0.25 0.443 Total disasters* polity 0.475 2 0.05 0.080 0.140 2 0.01 0.044 Clim. disasters * polity 0.180 2 0.00 0.028 0.026 2 0.00 0.010 Geol. disasters * polity 0.951 2 0.11 0.045 0.736 2 0.07 0.059 Total disasters* cred 0.878 2 0.11 0.057 0.186 2 0.01 0.042 Clim. disasters * cred 0.624 2 0.06 0.062 0.467 2 0.06 0.076 Geol. disasters * cred 0.624 2 0.06 0.060 0.268 2 0.03 0.064 Five-year panel Total disasters* eo 0.069 2 0.00 0.005 0.005 0.000 0.002 Clim. disasters * eo 0.061 0.000 0.005 0.003 2 0.00 0.002 Geol. disasters * eo 0.179 2 0.00 0.012 0.018 2 0.00 0.004 Total disasters* y0 0.081 0.000 0.005 0.004 2 0.00 0.002 Clim. disasters * y0 0.061 2 0.00 0.009 0.003 2 0.00 0.008 Geol. disasters * y0 0.269 2 0.01 0.053 0.028 2 0.00 0.058 Total disasters* polity 0.102 2 0.00 0.009 0.003 2 0.00 0.002 Clim. disasters * polity 0.064 0.000 0.004 0.000 2 0.00 0.000 Geol. disasters * polity 0.483 2 0.02 0.026 0.044 2 0.00 0.017 Total disasters* cred 0.164 2 0.00 0.012 0.000 2 0.00 0.001 Clim. disasters * cred 0.107 2 0.00 0.008 0.000 2 0.00 0.001 Geol. disasters * cred 0.168 2 0.00 0.011 0.001 2 0.00 0.000 Ten-year panel Total disasters* eo 0.142 2 0.00 0.022 0.014 2 0.00 0.010 Clim. disasters * eo 0.093 2 0.00 0.015 0.009 2 0.00 0.009 Geol. disasters * eo 0.399 2 0.03 0.045 0.068 2 0.00 0.020 Total disasters* y0 0.153 2 0.00 0.015 0.014 2 0.00 0.009 Clim. disasters * y0 0.089 2 0.00 0.030 0.007 2 0.00 0.027 Geol. disasters * y0 0.527 2 0.14 0.338 0.182 2 0.16 0.433 Total disasters* polity 0.143 2 0.00 0.025 0.007 2 0.00 0.007 Clim. disasters * polity 0.090 0.000 0.013 0.002 2 0.00 0.002 Geol. disasters * polity 0.592 2 0.06 0.061 0.089 2 0.00 0.036 Total disasters* cred 0.223 2 0.01 0.028 0.002 2 0.00 0.004 Clim. disasters * cred 0.154 2 0.00 0.022 0.001 2 0.00 0.004 Geol. disasters * cred 0.205 2 0.01 0.027 0.006 2 0.00 0.003 PIP is posterior inclusion probability, PM is posterior mean, and PSD is posterior standard deviation. Note: Values in italics have a PIP higher than 0.5. a. Bayesian model averaging using only models that include the parent variables of the inter- action terms. Source: Author’s analysis based on data described in the text. 300 THE WORLD BANK ECONOMIC REVIEW variables in table 1 by an extra variable that measures the percentage of moun- tainous terrain in the countries. This variable controls for geographic and topo- graphic effects that may be correlated with the disaster risk variables but that exert an independent effect on human capital investment (for instance, by affecting the return of infrastructure in terms of providing access to schools and thus affecting school enrollment). The BMA results for the importance and size of the effect of geologic disaster risk were essentially unchanged, while the mountainous terrain variable achieved a low posterior inclusion probability. To assess the impact of inﬂuential observations, BMA parameters and inclusion probabilities were estimated based on subsamples. The results for the long-run effects of geologic disaster risk on secondary school enrollment rates are robust to the following changes in the dataset: † Excluding the observations for disasters with the highest ratio of affected individuals per square kilometer (so as not to reduce the estimation sample dramatically, the cut-point was set at percentiles of the distri- bution of affected people by area ranging from the 80th to the 95th). † Excluding the observations for the poorest countries in the sample (thresholds based on observed income levels ranging up to the 30th per- centile were tried). † Excluding the observations for zero disasters, so that the results are not driven exclusively by the differences between observations with zero dis- aster risk and those with a positive disaster risk. † Excluding the �?ve observations identi�?ed as outliers through inspection of the residuals of the speci�?cation that includes all potential variables. This change intensi�?es the effect of disasters on schooling, with the geo- logic disaster variable achieving even higher posterior inclusion prob- ability and a higher estimated effect in absolute value. † Allowing for differential effects in developed and developing countries. In this case, there is strong evidence of homogeneity of the effect across subsamples. III. CONCLUSIONS The effects of natural disaster risk on human capital accumulation have received little attention in the academic literature. This article offers a �?rst, fully ﬂedged empirical study of the effects of natural disasters on secondary school enrollment across countries. To avoid reaching conclusions that are driven by single speci�?cations, Bayesian model averaging techniques were used to assess the robustness and size of the effects of natural disaster risk on human capital accumulation. The results offer strong evidence of the negative effects of geologic natural disaster risk on secondary school enrollment rates and complement the case Crespo Cuaresma 301 study literature. The effects tend to be homogeneous across countries and do not depend on income or the degree of human capital accumulation within a country. The empirical results presented here are robust to numerous variations in setting. The evidence presented in this article unveils a negative effect of natural dis- aster risk that had hitherto been largely ignored in the academic literature. Further research on the issue should concentrate on isolating empirically the channels leading to the aggregate effect of disasters on educational attainment found in this analysis. REFERENCES Albala-Bertrand, J. 1993a. “Natural Disaster Situations and Growth: A Macroeconomic Model for Sudden Disaster Impacts.�? World Development 21: 1417–34. ———. 1993b. Political Economy of Large Natural Disasters. Oxford, UK: Clarendon Press. Asian Development Bank and World Bank. 2005. “Pakistan 2005 Earthquake Preliminary Damage and Needs Assessment.�? Asian Development Bank and World Bank, Islamabad. ´a-Pen Checchi, D., and C. Garcı ˜ alosa. 2004. “Risk and the Distribution of Human Capital.�? Economics Letters 82: 53 –61. Chipman, H.A. 1996. “Bayesian Variable Selection with Related Predictors.�? Canadian Journal of Statistics 24: 17 –36. Clyde, M.A., and E.I. George. 2004. “Model Uncertainty.�? Statistical Science 19: 81– 94. CRED (Collaborating Centre for Research on the Epidemiology of Disasters, a World Health Organization Collaborating Centre). 2004. Emergency Events Database, EM-DAT. Universite ´ Catholique de Louvain, Brussels. Available at http://www.emdat.be/. Crespo Cuaresma, J. Forthcoming. “How Different Is Africa?�?. Journal of Applied Econometrics. Crespo Cuaresma, J., G. Doppelhofer, and M. Feldkircher. 2008. “The Determinants of Economic Growth in European Regions.�? Working Paper 2008-26. Faculty of Economics and Statistics, University of Innsbruck. Crespo Cuaresma, J., J. Hlouskova, and M. Obersteiner. 2008. “Natural Disasters as Creative Destruction: Evidence from Developing Countries.�? Economic Inquiry 46: 214– 26. Dacy, D.C., and H.C. Kunreuther. 1969. The Economics of Natural Disasters. New York: Free Press. Duryea, S., D. Lamb, and D. Levison. 2007. “Effects of Economic Shocks on Children’s Employment and Schooling in Brazil.�? Journal of Development Economics 84: 188– 214. ´ ndez, C., E. Ley, and M.F. Steel. 2001a. “Benchmark Priors for Bayesian Model Averaging.�? Ferna Journal of Econometrics 100: 381 –427. ———. 2001b. “Model Uncertainty in Cross-Country Growth Regressions.�? Journal of Applied Econometrics 16: 563– 76. Flug, K., A. Spilimbergo, and E. Wachtenheim. 1998. “Investment in Education: Do Economic Volatility and Credit Constraints Matter?�? Journal of Development Economics 55: 465– 81. Foster, D.P., and E.I. George. 1994. “The Risk Inﬂation Criterion for Multiple Regression.�? Annals of Statistics 22: 1947– 75. Hanushek, E.A., J.F. Kain, and S.G. Rivkin. 2004. “Disruption versus Tiebout Improvement: The Costs and Bene�?ts of Switching Schools.�? Journal of Public Economics 88: 1721–46. Heston, A., R. Summers, and B. Aten. 2006. Penn World Table Version 6.2. University of Pennsylvania, Center for International Comparisons of Production, Income, and Prices, Philadelphia, Pa. Available at http://pwt.econ.upenn.edu. 302 THE WORLD BANK ECONOMIC REVIEW Gertler, P., D.I. Levine, and M. Ames. 2004. Schooling and Parental Death. Review of Economics and Statistics 86: 211– 25. Guha-Sapir, D., and R. Below. 2002. “The Quality and Accuracy of Disaster Data: A Comparative Analysis of 3 Global Data Sets.�? Working paper prepared for the Disaster Management Facility. World Bank and Collaborating Centre for Research on the Epidemiology of Disasters, Brussels. Kass, R.E., and A.E. Raftery. 1995. “Bayes Factors.�? Journal of the American Statistical Association 90, 773 –795. Kass, R.E., and L. Wasserman. 1995. “A Reference Bayesian Test for Nested Hypotheses and Its Relationship to the Schwarz Criterion.�? Journal of the American Statistical Association 90: 928– 34. Kim, N. 2008. “Impact of Extreme Climate Events on Educational Attainment: Evidence from Cross Section Data and Welfare Projection.�? UNDP/ODS Working Paper. United Nations Development Programme, New York. Ley, E., and M.F. Steel. 2009. “On the Effect of Prior Assumptions in Bayesian Model Averaging with Applications to Growth Regression.�? Journal of Applied Econometrics 24: 651–74. Masanjala, W.H., and C. Papageorgiou. 2008. “Rough and Lonely Road to Prosperity: A Reexamination of the Sources of Growth in Africa Using Bayesian Model Averaging.�? Journal of Applied Econometrics 23: 671 –82. Marshall, M.G., and K. Jaggers. 1995. Polity IV Project. Political Regime Characteristics and Transition, 1800–2004, database. Version 2005. www.systemicpeace.org/polity/polity4.htm. Noy, I. 2009. “The Macroeconomic Consequences of Disasters.�? Journal of Development Economics 88 (2): 221 –31. Okuyama, Y. 2009. “Critical Review of Methodologies on Disaster Impact Estimation.�? Background paper for the joint World Bank– United Nations Assessment on the Economics of Disaster Risk Reduction. The Global Facility for Disaster Reduction and Recovery, Washington, DC. Raftery, A. E. 1995. “Bayesian Model Selection for Social Research.�? Sociological Methodology 25: 111 –63. Rasmussen, T. N. 2004. “Macroeconomic Implications of Natural Disasters in the Caribbean.�? IMF Working Paper WP/04/224. International Monetary Fund, Washington, DC. Sacerdote, B. 2008. “When the Saints Come Marching in: Effects of Hurricanes Katrina and Rita on Student Evacuees.�? Dartmouth College, Department of Economics, Hanover, New Hampshire. Sala-i-Martin, X., G. Doppelhofer, and R. Miller. 2004. “Determinants of Long-Term Growth: A Bayesian Averaging of Classical Estimates (BACE) Approach.�? American Economic Review 94: 813 –35. Skidmore, M. 2001. Risk, Natural Disasters, and Household Saving in a Life Cycle Model. Japan and the World Economy 13: 15 –34. Skidmore, M., and H. Toya. 2002. Do Natural Disasters Promote Long-run Growth?, Economic Inquiry 40: 664– 87. ———. 2007. “Economic Development and the Impacts of Natural Disasters.�? Economic Letters 94: 20 –25. Stijns, J-P. 2006. “Natural Resource Abundance and Human Capital Accumulation.�? World Development 34: 1060–83. Tol, R., and F. Leek. 1999. “Economic Analysis of Natural Disasters.�? in T.E. Downing, A.J. OlsthoornR.S.T. Tol eds., Climate, Change and Risk. London: Routledge. Bank, World. 2006. World Development Indicators 2006. Washington, DC: World Bank. Zellner, A. 1986. “On Assessing Prior Distributions and Bayesian Regression Analysis with g-prior Distributions.�? In P.K. Goel, and A. Zellner eds., Bayesian Inference and Decision Techniques: Essays in Honour of Bruno de Finetti. Amsterdam: North Holland. Managing for Results in Primary Education in Madagascar: Evaluating the Impact of Selected Workﬂow Interventions ´ rard Lassibille, Jee-Peng Tan, Cornelia Jesse, Ge and Trang Van Nguyen The impact of speci�?c actions designed to streamline and tighten the workﬂow pro- cesses of key actors in Madagascar’s primary education sector are evaluated. To inform the strategy for scaling up, a randomized experiment was carried out over two school years. The results show that interventions at the school level, reinforced by interventions at the subdistrict and district levels, succeeded in changing the behavior of the actors toward better management of key pedagogical functions. In terms of education outcomes, the interventions improved school attendance, reduced grade rep- etition, and raised test scores ( particularly in Malagasy and mathematics), although the gains in learning at the end of the evaluation period were not always statistically signi�?cant. Interventions limited to the subdistrict and district levels proved largely ineffective. JEL codes: I21, I28, J24 Ge´ rard Lassibille (corresponding author; gerard.lassibille@u-bourgogne.fr) is a research director at the Centre National de la Recherche Scienti�?que in the Institut de Recherche sur l’Economie de l’Education, Dijon, France. Jee-Peng Tan ( jtan@worldbank.org) is advisor in the Education Department of the Human Development Network at the World Bank. Cornelia Jesse (cjesse@worldbank.org) is operations of�?cer in the Africa Human Development Department at the World Bank. Trang Van Nguyen (tnguyen16@worldbank.org) is an economist in the Poverty Reduction and Economic Management Sector Department in the East Asia and Paci�?c Region at the World Bank. The authors are deeply indebted to Esther Duﬂo for invaluable help and guidance, particularly in setting up the experimental design for this impact evaluation. They thank their counterparts at the Madagascar Ministry of Education, led at the time of this study by Tahinarinoro Raza�?ndramary and Paul Randrianirina, and the staff of Aide et Action for help in conceptualizing the workﬂow tools, implementing the experiment, and collecting the data. Among close collaborators on this work, the authors especially appreciate Pierre-Emmanuel Couralet and Erika Strand for on-the-ground supervision during the experiment and Mathieu Laroche and Muriel Nicot-Guillorel for technical support to the Malagasy counterparts. The authors also thank colleagues and friends who provided input, advice, and comment to guide the design of the experiment, among them Sajitha Bashir, Benu Bidani, Robert Blake, Deon Filmer, Elizabeth King, Arianna Legovini, Robert Prouty, Lina Rajonhson, Patrick Ramanantoanina, and Venkatesh Sundararaman. The authors thank three anonymous referees and the editor of the journal for insightful comments. Finally, they acknowledge the support of the World Bank and the governments of France, Ireland, Madagascar, and Norway, as well as that of the donor partners of the Education for All Fast Track Initiative through the Education Program Development Fund. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 303 –329 doi:10.1093/wber/lhq009 Advance Access Publication August 6, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 303 304 THE WORLD BANK ECONOMIC REVIEW Over the last 10 years, low-income countries in Africa have made striking pro- gress in expanding coverage of primary education. However, in many of these countries the education system continues to deliver worse results than expected, putting at risk the goal of universal primary school completion. Weak adminis- tration, inadequate focus on results, and poor governance structures are thought to be some of the reasons for the meager results. Better management of workﬂow processes at each point along the service delivery chain might therefore improve productivity and provide a useful tool for raising the per- formance and ef�?ciency of education systems. This article reports on the �?rst known attempt at a randomized impact evalu- ation of interventions to improve management of the teaching process in an African setting. It adds to the growing literature on randomized impact evalu- ation of education programs in developing countries. Kremer and Holla (2009) group such evaluations into �?ve broad categories, with the interventions typically aiming to broaden access or improve learning outcomes: lowering private costs, increasing subsidies, or providing students with health care (examples include Bobonis and Finan 2009; Evans, Kremer, and Ngatia 2008; King and Orazem 1999; Kremer, Moulin, and Namunyu 2003; Miguel and Kremer 2004; Tan, Lane, and Lassibille 1999; and Todd and Wolpin 2006); enhancing teacher inputs or other resources such as textbooks or ﬂipcharts (Barnejee and others 2006; Duﬂo, Dupas, and Kremer 2007; Glewwe and others 2004; Glewwe, Kremer, and Moulin 2009); reforming pedagogy through radio- or computer- assisted instruction (Barnejee and others 2007; Jamison and others 1981); redu- cing teacher absenteeism through better incentives (Barnejee and Duﬂo 2006; Glewwe, Ilias, and Kremer 2008); and informing and involving local commu- nities and transferring to them control of school management in areas such as teacher hiring and �?ring (Barnejee and others 2006; Duﬂo, Dupas, and Kremer 2007; Gertler, Patrinos, and Rubio-Codina 2008). The experiment reported in this article targets Madagascar’s public primary school system, which enrolls 80 percent of students at this level. Thanks to increases in public spending since the late 1990s, the country has been making good progress in expanding coverage in primary education. At the same time, the government put in place important reforms to strengthen administration of the education system. The technical staff of the ministry developed tools to streamline and tighten the workﬂow processes of all actors along the service delivery chain, focusing on measures to make explicit the functional responsi- bilities of teachers, school directors, and district and subdistrict administrative staff through a coherent and detailed manual of operations and to increase information ﬂows and accountability through report cards at the school and higher administrative levels. These tools are discussed in detail in section II. To evaluate the impact of the workﬂow-enhancing interventions on the be- havior of service providers and on school outcomes, a randomized experiment was carried out over two school years (2005/06–2006/07) in a sample of school districts. The experiment was designed to answer three main questions: Lassibille, Tan, Jesse, and Van Nguyen 305 How much do the workﬂow-enhancing interventions alter the behaviors of service providers and schooling outcomes? How do the interventions differ in impact when aimed at the school, subdistrict, and district levels? Do the inter- ventions have spillover effects? The results after two years of experimentation suggest that only when com- bined as a package of school-level interventions reinforced by interventions at the subdistrict and district levels do the workﬂow-enhancing tools improve the management practices of service providers. School attendance rose and grade repetition fell in the intervention schools by statistically signi�?cant amounts; test scores also rose, though the gain was not always statistically signi�?cant. Interventions limited to the subdistrict and district levels were largely ineffec- tive, probably due to weak mechanisms for monitoring and control and to the lack of a true leadership culture among these actors. These �?ndings are sugges- tive and potentially useful for other Sub-Saharan African countries facing the same management issues in education. They can help inform similar work to clarify the options for addressing these issues. The remainder of the article proceeds as follow. Section I describes the edu- cation sector in Madagascar. Section II explains the interventions. Section III focuses on the design of the experiment and the implementation of the impact evaluation. Section IV presents the empirical results. And section V concludes. I . MA D A G A SCA R ’ S E D U CAT I O N S E C TO R IN CONTEXT Madagascar has not always prioritized education in the allocation of public spending, but in recent years the government has begun to channel more resources to the sector as part of its commitment under the Heavily Indebted Poor Countries Debt Initiative and in its poverty reduction strategy. As a result, public recurrent spending on education rose from 2.2 percent of GDP in 2000/ 01 to 3.3 percent in 2006/07. In tandem with the increase of funding, the gov- ernment introduced several important reforms (Government of Madagascar 2004b). Among the keys reforms are eliminating school fees for primary edu- ´ coles) and using school cation, launching a system of capitation grants (caisse e grants to incentivize performance, providing kits scolaires and textbooks to primary school students, creating school boards in all primary schools, provid- ing public subsidies to supplement the pay of non–civil service teachers in public schools who in the past have been hired and paid entirely by parent associations, restructuring the primary and secondary cycles of schooling, and introducing new pedagogical approaches.1 The most visible signs of progress are the large increase in coverage in primary education in recent years. In 2006/07, the education system enrolled some 3.8 million students in both public and private schools—more than twice 1. The new government that came into power in early 2009 is reviewing these reforms to determine their future direction. 306 THE WORLD BANK ECONOMIC REVIEW the enrollment in 1996. As a result of this increase, the gross enrollment ratio in primary education rose from 83 percent in 1996 to 123 percent in 2006, and thus the number of primary school teachers in the public sector also increased dramatically, from some 28,000 to 59,000. More than 4,000 new primary public schools have been created during the last 10 years. While the progress in coverage has been impressive, enormous challenges and problems remain for system performance. Entry rates to grade 1 are high, but less than half of each cohort reaches the end of the �?ve-year primary cycle. Despite government interventions, grade repetition has not improved as rapidly as expected. In fact, repetition rates are still uniformly high throughout the primary cycle, averaging about 18 percent. With regard to student learning, the 2005 Programme d’Analyse des Syste ` mes Educatifs de la Confemen survey indi- cates that Malagasy students performed better than their peers in other low- income African countries (for example, Benin, Cameroon, Chad, and Mauritania). But their scores were still low in absolute terms: about 30 percent on the test in French and about 50 percent on the tests in Malagasy and math- ematics (PASEC 2007). They also performed worse than the 1998 cohort of Malagasy students in grade 5, who scored 48 percent in French and 59 percent in mathematics (World Bank 2002). Several factors explain the poor performance of the primary education system. Many demand-side conditions are impervious to policy interventions, particularly in the short run, such as household poverty, community character- istics, the opportunity cost of children’s time, and parental perceptions of the value of education (see for example, Haveman and Wolfe 1995). Supply-side factors under the control of policymakers and managers in the education sector (see for example, Hanushek 1997) include the allocation of teachers across schools and the management of the pedagogical process. In Madagascar the degree of randomness in the allocation of teachers across primary school has diminished to a large extent over the last 10 years (Government of Madagascar 2008). Assigning teachers more consistently across schools clearly signals better administrative management of the system. Within schools, however, many aspects of the pedagogical processes are poorly managed, and tasks essential for student learning are neglected. A survey of workﬂow processes conducted as part of the impact evaluation reported in this article offers particularly telling revelations.2 The data show that student absenteeism is poorly monitored by teachers, with attendance taken an average of 13 days a month, and poorly supervised by school direc- tors, with 10 percent neglecting the task entirely and only a third signing off on the attendance records kept by teachers. Essential pedagogical tasks are 2. The workﬂow details presented in the following paragraphs are based on data codi�?ed from about 850 workﬂow artifacts collected from about 150 teachers in 40 schools that provide a record of their work over a full school year (see section III); the results discussed here pertain to teachers in schools that did not receive any of the interventions associated with the experiment reported in this article. Lassibille, Tan, Jesse, and Van Nguyen 307 often neglected: 20 percent of teachers do not prepare daily lessons plans, only 15 percent consistently prepare daily and biweekly lessons plans, and a third of school directors never discuss daily lesson plans with teachers. Student academic progress is also poorly monitored. The results of tests and quizzes are rarely recorded, if at all, and 25 percent of teachers do not prepare individual student report cards. Communication from teachers to parents on student learning is often perfunctory, and student absences are rarely communi- cated to parents. School directors seldom follow up on student performance: 75 percent do not discuss learning outcomes with their teachers, and only 20 percent sign off on test results and student report cards. The same goes for teacher absences, which average nearly 10 percent3—hardly a negligible �?gure. Only 8 percent of school directors monitor teacher absences by taking daily attendance or tracking and posting a monthly summary of absences, and more than 80 percent fail to report teacher absences to subdistrict and district admin- istrators. The general impression that emerges from these observations is one of obvious lack of organization, control, and accountability—all of which can compromise the performance of the system and the chances of success of the many ongoing reforms. II. DESCRIPTION OF THE INTERVENTIONS To put the evaluated interventions in context, this section discusses a few note- worthy features of the administrative arrangements in Madagascar’s primary education system. The system consists of a large number of public schools, funded and managed by centrally appointed civil servants, and a small number of mostly urban private schools, funded largely by student fees and managed by religious and other entities. In 2006/07 the network of some 18,000 public primary schools enrolled about 3.1 million students and employed 18,000 school directors and 59,000 teachers (table 1). These schools were supervised by 1,500 subdistrict adminis- trators (chefs ZAP), who provide the �?rst line of administrative and pedagogi- cal support to the schools under their care and who in turn report to one of 111 district administrators (chefs CISCO), who in turn report to one of 22 regional administrators (DREN)4 and the ministry of education. The subdis- trict and district administrators channel resources to schools, supervise teaching and learning practices and the collection of school statistics, and administer the annual national examination at the end of grade 5. Their work also includes distributing paychecks to teachers; organizing school building and maintenance projects; overseeing the distribution of books and small grants to schools; 3. This estimate compares reasonably well with others studies conducted on the topic in Madagascar (see World Bank 2008). 4. The regional administrators are new. They were introduced when the impact evaluation was already well under way, but were not (and still are not) yet operational. 308 T A B L E 1 . Administrative Structure of Madagascar’s Public Primary School System, 2006–07 School Subdistrict District Regional Ministry of Actors Students Teachers directors administrators administrators administrators Education Total number of actors 3,103,000 59,000 17,600 1,544 111 22 1 Number under each actor: THE WORLD BANK ECONOMIC REVIEW Students 53 177 2,010 27,900 141,000 3,103,000 Teachers 3 38 530 2690 59,000 School directors 11 160 800 17,600 Subdistrict administrators 14 70 1,544 District administrators 5 111 Regional administrators 22 Source: Authors’ analysis of the 2006/07 Fiches d’Enque ˆ tes Rapides. Lassibille, Tan, Jesse, and Van Nguyen 309 designing and implementing in-service training for teachers, school directors, and other staff; providing in-service support; and ensuring timely returns on the school census questionnaires. The key actors in the service delivery chain are responsible for only a few actors at the level below them. Each district administrator manages an average of 14 subdistrict administrators, and each subdistrict administrator is responsible for about 10 school directors. Each school director manages three teachers and 177 students on average. The interventions being considered by the Madagascar Ministry of Education seek to tighten the management of pedagogical workﬂow processes at each point along the service delivery chain described above and to increase the focus on results by making explicit to the actors their responsibilities and supporting them with tools and procedures to accomplish their tasks, inserting supervision and follow-up at key points in the administrative hierarchy, and facilitating school-community interactions and promoting accountability for results around school report cards. Each set of actors performs many tasks, but a few activities at each point along the service delivery chain are central to the mission of managing for results and must therefore receive routine attention by the relevant actor. After intensive consultation and discussion, the ministry of education identi�?ed the main responsibilities of the various actors. Out of this exercise emerged the core job description for each service delivery agent in the education system. Based on this job description, ministry of�?cials developed the corresponding operational tools and processes, which focus on six broad domains of activity: pedagogy, student learning and follow-up, management of instructional time, administration, school statistics, and partnership with the local community.5 Typically, the set of products for each actor is a rationalization and simpli�?- cation of existing templates that have become unwieldy, incoherent, and out- dated over time. More than 30 tools were developed for tasks considered essential for a well functioning system, among them 7 designed for use by tea- chers, 8 by school directors, 8 by subdistrict administrators, and 9 by district administrators. Each operational tool encourages superiors to pay closer atten- tion to the work of their direct reports (for example, requiring school directors to review information supplied by a teacher and suggest speci�?c follow-up actions in case of need) and is accompanied by a simple and clear guidebook that explains how and when each tool should be used. Training modules have been speci�?cally customized and offered to all the relevant actors (see section III). Depending on the actor and task, the templates and accompanying work- ﬂow processes are used daily, weekly, monthly, or at a speci�?c time in the year. As noted above, they replace tools that had become increasingly incoherent and idiosyncratic for lack of upkeep and updating. Teachers and administrators 5. For details on the tools, their actual presentation, their purpose, and how they are to be used by the various actors in the education system, see Government of Madagascar (2004a) and www.education. gov.mg. 310 THE WORLD BANK ECONOMIC REVIEW use the existing tools only sporadically, if at all, and the tools themselves have all but lost their power to reinforce the reporting relationships among teachers, school directors, and subdistrict and district administrators. While it is important to rationalize and tighten workﬂow processes through the operational tools described above, these processes may still be ineffective without an explicit focus on results. Report cards with customized information for each school, subdistrict, or district, one way to provide better accountability and address this concern,6 are part of the interventions created by this impact evaluation. The report cards are prepared using data from the Ministry of Education’s annual school census. To make them easy to understand for illiter- ate parents and members of poor rural communities, the content is kept simple. At the school level they contain a small set of performance indicators, information on enrollments and resources endowments, and selected compara- tive data that show a school’s performance and resource endowment rank rela- tive to other schools. Schools fall into one of four performance categories, each marked by a relevant icon: excellent, satisfactory, in dif�?culty, and disappoint- ing.7 At the subdistrict and district levels the report cards contain aggregate information on the same indictors as the school report cards, indicate the per- formance category to which the subdistrict or district in question belongs, and lists the schools or subdistricts in each performance category. At the district level report cards can serve as an instrument for self- evaluation and as a basis for designing and implementing an action plan for improvement. At the school and subdistrict levels report cards can likewise help focus attention on results and encourage action toward better outcomes. At the school level recent policy changes have been, until the coup in early 2009, strengthening the prospects in this regard. In the second phase of the Programme National pour l’Ame ´ lioration de l’Education (Government of Madagascar 1997), the government decided to encourage greater participation in school-based management by various actors through the nationwide contrats-programmes initiative, which set out explicit responsibilities for parents, teachers, and school directors. It essentially sought to foster commit- ment by the various parties to shared targets for increases in enrollment and examination pass rates and decreases in teacher absenteeism and to explicit contributions of materials and labor for classroom renovation and construc- tion. School report cards can provide key information to clarify and motivate 6. School report cards have been used extensively in other countries; for an analysis of their impact on student performance, see, for example, Banerjee and others (2008) and Hanushek and Raymond (2004). 7. A school is “excellent�? when it achieves above average results despite receiving a smaller endowment of resources than average, “satisfactory�? when it achieves the expected above average result with an above average endowment of resources, “in dif�?culty�? when it achieves below average results with a below average endowment of resources, and “disappointing�? when it achieves below average results despite being endowed with more than the average amount of resources. A sample school report card is shown in �?gure A1. Lassibille, Tan, Jesse, and Van Nguyen 311 decisionmaking by the relevant actors. For this reason, one of the interventions in impact evaluation reported below included structured school meetings with the staff of the school, parents, and community members in order to engage them in purposeful discussion about the school improvement plan for their school in light of the information in its report card, thus promoting parental monitoring of the school and accountability. To summarize, the interventions under consideration by the Madagascar Ministry of Education to tighten management of the primary education system consisted of providing workﬂow templates or tools, report cards, and related instruction guidebooks; facilitating meetings between school staff and the com- munity to develop and agree on a school improvement plan; and structuring training sessions to follow a set agenda that informs and motivates teachers, school directors, and subdistrict and administrators toward better performance (the Ame ´ lioration de la Gestion de l’Education a` Madagascar [AGEMAD] interventions). III. DESIGN AND I M P L E M E N TAT I O N OF THE I M P A C T E VA L U A T I O N Because of the diversity of actors and the large number of workﬂow processes and operational templates involved, many possibilities exist for implementing the interventions presented above. Because the operational tools by themselves might not be suf�?cient to improve productivity unless combined with account- ability measures to focus attention on schooling outcomes, the evaluation team, in discussion with counterparts in the Madagascar Ministry of Education, packaged the interventions and targeted them to two main sets of actors: midlevel bureaucrats (subdistrict and district administrators) and front- line service providers at the school level (teachers and school directors). Figure 1 provides an overview of the intervention design of the experiment,8 which addresses three important questions. How much do the AGEMAD inter- ventions affect the way schools function and their performance in terms of key outcome indicators? Would subdistrict- or district-level interventions produce the desired impact on service provider behaviors and on schooling outcomes, or would such interventions be ineffective unless reinforced with direct follow-up and training at the school level? Would all schools in a subdistrict require direct follow-up and training to achieve the desired impact, or would it be suf�?cient to target some schools to serve as demonstration projects for repli- cation in the remaining schools? These questions are addressed with an eye to obtaining results that inform the strategy for scaling up the interventions by implementing a nested random- ized experiment (see, for example, Boruch 1997; Dennis and Boruch 1989; and 8. For a more complete picture of the experimental design, see �?gure A2, which details the precise nature and type of interventions in each treatment group and the population and sample sizes involved. 312 THE WORLD BANK ECONOMIC REVIEW F I G U R E 1. Stylized Description of the Experimental Design Source: Authors’ construction. Duﬂo, Glennester, and Kremer 2007). First, 30 relatively accessible and similar districts were selected as the universe for the experiment,9 with 15 districts ran- domly chosen to receive the interventions and the remaining 15 chosen to be managed as usual (as control districts). The 30 districts contain 573 subdis- tricts and 6,488 schools. For practical reasons, the most remote and inaccess- ible subdistricts and schools were removed (107 subdistricts and 2,714 schools), leaving 466 subdistricts and 3,774 schools, with 259 subdistricts and 2,053 schools in the intervention districts and 207 subdistricts and 1,721 schools in the control districts. From the 3,774 schools, a sample of 1,212 schools in 179 subdistricts was randomly selected to evaluate the impact of the interventions. With 303 schools in each intervention group, the sample size is large enough to detect the impact of the interventions on behaviors and indi- cators of school performance. The relevant administrators in the intervention districts received oper- ational tools, corresponding guidebooks explaining the use of the tools, and training relevant to their tasks, along with a district report card. Within the intervention districts, some subdistricts were randomly selected to receive 9. Power calculations show that this sample size is suf�?cient to detect a 0.25 standard deviation effect in school-level average test score, assuming a 0.1 correlation across schools within a district. To save space, baseline data on the pro�?les of the experimental and nonexperimental schools are not shown here but are available on request from the authors. To summarize, the data indicate that the 30 experimental districts are comparable to the 81 nonexperimental districts in number of students and teachers, percentage of repeaters, and number of schools managed by each subdistrict administrator. However, on average, the district administrators in the experiment are responsible for more subdistrict administrators and therefore for more schools than the district administrators in the nonexperimental districts. Lassibille, Tan, Jesse, and Van Nguyen 313 operational tools relevant to subdistrict-level functions, corresponding guide- books, speci�?c training, and a subdistrict report card. The other subdistricts received only training and a report card, schools in these subdistricts received only a report card. A sample of schools in the intervention subdis- tricts was randomly selected to receive the operational tools designed for use by school directors and teachers, corresponding guidebooks, training, and a school report card to be discussed in meetings with the community. For the other schools, subdistrict administrators received school report cards for onward distribution and templates of the school-level tools for replication. These tasks were left to the initiative and discretion of these administrators.10 To summarize, the nested design of the experiment creates four types of schools: schools in the control districts, which receive no intervention at any level in the system (control group), schools in intervention districts and in nonintervention subdistricts with no school-level interventions (group 1), schools in intervention districts and in intervention subdistricts with no school-level interventions (group 2), and schools in intervention districts and in intervention subdistricts that directly received school-level interventions (group 3). The diffusion of interventions to schools in groups 1 and 2 is left to the initiative of the subdistrict or district administrator, following the training-of-trainers model extensively used in Madagascar ( formation en cascade). By contrast, schools in group 3 receive the interventions directly, and these school-level processes are reinforced at the subdistrict and district levels.11 Since intervention assignment was random, comparing these four groups of schools gives a consistent estimate of the impact and makes it possible to answer key questions about the interventions. Comparing group 1 and the control group reveals the effect of a training-of-trainers or “cascade�? model that involves only the district-level administrators to implement the interven- tions. Comparing group 2 and the control group reveals the impact of a more intensive form of the cascade model that involves both subdistrict- and district-level interventions. Comparing group 3 with the control group reveals the maximum impact of the full set of interventions at the school, subdistrict, and district levels. Comparing groups 2 and 3 reveals the marginal impact of the most intensive package of interventions that directly target teachers and school directors along with subdistrict and district administrators over the 10. Schools and subdistricts in intervention districts were randomly assigned to one of the interventions after the sample was strati�?ed by school size and repetition rate. Within the largest districts, 40 intervention schools were selected from each of about 10 subdistricts, which were themselves randomly selected; and within the smaller intervention districts, 13 intervention schools were selected at random from each of about 4 subdistricts, which were themselves randomly selected. 11. The interventions received by schools in group 3 are inherently more costly than those received by schools in groups 1 and 2 A central issue in the impact evaluation is to determine whether they achieve greater impact than the less costly packages. 314 THE WORLD BANK ECONOMIC REVIEW second-most intensive package that directly targets only subdistrict and district administrators.12 Because schooling processes take time to evolve and produce results, the experiment took place over two school years. In the �?rst year, changes, if any, were expected mostly in the behaviors of the various actors, and in the second year, these changes would have had suf�?cient time to translate into changes in at least some of the schooling outcomes. The experiment started in September 2005 and ended in June 2007. A team comprising central-level ministry staff was formally constituted to oversee the experiment’s implementation. It was reinforced by technical experts and facili- tators (the latter supervised by Aide et Action, a nongovernmental organization that assisted in implementing the school-level activities). In total, more than 50 people were involved in the impact evaluation. Some 200,000 operational tools and 11,000 guidebooks were distributed to the actors, and some 10,000 report cards were produced for school directors and subdistrict and district adminis- trators by the Ministry of Education using school census data. Training modules, role-playing games, and supporting material for each set of actors were provided to nearly 4,000 participants. In the �?rst year, central-level minis- try staff involved in the experimentation organized and delivered a four-day course for the subdistrict and district administrators, who in turn delivered a two-day course for the teachers and school directors under their supervision. In the second year, training for the various actors was shortened to two days for the subdistrict and district administrators and to just one day for teachers and school directors. And over the two years, 1,500 meetings between schools and their communities were organized with the support of specially recruited facili- tators to discuss and prepare the school’s improvement plan. Data on the behavior of school personnel were collected from December to May during each of the two school years using a school questionnaire adminis- tered following random unannounced visits to the schools in the experiment. The questionnaire yielded detailed information on the pedagogical and admin- istrative organization of the school, on the personal characteristics of the tea- chers and their quali�?cations, and on teacher and student absenteeism. It also gathered data on various aspects of teacher behavior and how teachers do their jobs. The dataset contains information on some 4,000 teachers in the 1,200 schools involved in the impact evaluation.13 Besides the school-level data, information was also collected from subdistrict and district administrators during the second year. And at the end of the exper- iment, some 850 administrative and pedagogical tools that were used by tea- chers and school directors in 2006/07 were collected from 40 schools 12. The comparison thus measures the impact of the most intensive package of interventions after netting out the direct and indirect effects of a package centered on the administrators in which a cascade of bene�?ts ﬂowing through the relevant administrators to the school-level personnel under their supervision. 13. Five teachers were randomly selected in each school. Lassibille, Tan, Jesse, and Van Nguyen 315 randomly selected from those in the control group and in group 3. Because this information is codi�?ed directly from the artifacts actually used by teachers and school directors in the course of their work, it provides an independent exter- nal check on the data reported by school personnel.14 As indicated below, there is a high degree of consistency between the two sources of information, a feature that strengthens the analytical results based on the reported data. In terms of schooling outcomes, the indicators used to evaluate the impact of the interventions are based on school census data (Fiches de Fin d’Anne ´ e and Fiches d’Enque ˆ tes Rapides) on the number of repeaters and dropouts and on pass rates on the primary school leaving exam (Certi�?cat d’Etudes Primaires Ele´ mentaires, CEPE) for grade 5 students. Some schools have no candidates sitting for the CEPE exams; so to compare the schools across the full spectrum of outcomes, achievement tests were administered in both years of the exper- iment. The baseline test was administered in February 2006 to about 25,000 stu- dents in grade 3; the post-intervention test was administered in May 2007 to about 22,000 students in the same cohort (in grade 4). The test instrument is based on test items from the 2005 Programme d’Analyse des Syste ` mes Educatif de la Confemen survey; in both years, a maximum of 25 students per school were tested in French, Malagasy, and mathematics. To minimize selection bias in the data, assiduous effort was made to administer the tests to all students in the baseline sample. The survey enumerators were thus instructed to identify stu- dents who were absent from school on the day of the test, seek out these students in their homes, and administer the tests at home whenever feasible.15 I V. E M P I R I C A L R E S U L T S The empirical results are presented here in two parts. The �?rst pertains to the direct impact of the interventions on the behavior of school personnel. The second pertains to the indirect effect of the interventions on students and their schooling outcomes. Because the experiment is randomized, the impact of a 14. Detailed data on artifacts, such as those collected in this experiment, are extremely rare— unsurprising, given that teachers may be unwilling to part with a source of information that provides such a complete a record of their work. In The Gambia, for example, Adekanmbi, Blimpo, and Evans (2009) report on an ongoing impact evaluation in which the enumerators collected data on only one artifact—the lesson plan—and then only by asking teachers to show it to them during the visit. The study found that 35 percent of the teachers surveyed declined to show their lesson plan while claiming to have it, 17 percent did not have a lesson plan at all, and 48 percent said they had a lesson plan and were able to show it. 15. The baseline characteristics of schools in the intervention and control groups are summarized in table A1. The data show that schools in the various groups are comparable in terms of size (about 250 students), number of teachers (four), repetition rate (about 18 percent), CEPE pass rate (about 61 percent), percentage of correct answers on the test administered in year one (about 61 percent), and loss of students from the sample between the two school years spanning the AGEMAD experiment for whom test scores were collected (about 11 percent). An important feature in the resulting database is that differences in baseline test scores between students who dropped out of the sample and those who remained were small and comparable across intervention and control groups. 316 THE WORLD BANK ECONOMIC REVIEW particular intervention on the behavior of school personnel and on student out- comes ( y) is estimated using the following regression in year 2: X 3 ð1Þ y¼aþ bi gi þ e i¼1 where gi is an indicator for whether the school is in an intervention group i (with i ¼ 1, 2, or 3), with the control group being the omitted category; a measures the average value of y in the control group; and bi estimates the impact of each intervention on y. A similar formulation is used to evaluate the pairwise impacts of the three interventions relative to each other. Direct Effects on Service Providers’ Behaviors Alternative methods exist for evaluating and presenting the direct effects of the interventions because many actors are involved, each with multiple responsibil- ities. This article focuses on the school personnel closest to the students’ daily learning activities, the teachers and school directors. It further focuses on the seven tasks for each group that Malagasy educators deem essential to the mission of managing for results (table 2). For teachers these tasks include taking daily roll call, preparing the lesson of the day, monitoring student learning, helping lagging students, and the like. For school directors the tasks include keeping a register of enrollments, analyzing student absences on a regular basis, following up lesson planning with teachers, reviewing student performance, and so on. To keep the analysis tractable, a “good�? teacher (or more accurately, a minimally conscientious teacher) and a “good�? school director are de�?ned here as one who performs all seven workﬂow tasks that Malagasy educators consider essential to the role. In the same vein, a well managed school is one where the school director and all the teachers perform all their essential tasks.16 With regard to teacher absenteeism, the interventions had no signi�?cant impact.17 Absenteeism was about 9 percent in the control group, compared with 9–10 percent in the three intervention groups (table 3). Teacher absentee- ism is much lower in Madagascar than in other developing countries such as Bangladesh, India, and Uganda (Chaudhury and others 2006). It is linked mostly to the fact that many teachers must travel to a central location, often 16. An alternative approach is to follow the methodology of Kling, Katz, and Liebman (2007), who analyze program effects using a seemingly unrelated regression for each task—either separately or in clusters of tasks. This approach was considered for this article, but ultimately the judgment of Malagasy educators that the individual tasks are an integrated package of closely connected actions required for managing the teaching and learning process was accepted. 17. Data on teacher absenteeism are based on information supplied by the school directors in the sample schools to the enumerators, who make unannounced visits following a random schedule. Teachers are considered absent if they were absent, for whatever reason, for at least one day during the week preceding the visit. Lassibille, Tan, Jesse, and Van Nguyen 317 T A B L E 2 . Tasks Considered by Malagasy Educators to be Essential for Teachers and School Directors Teachers School directors Takes daily roll call Keeps a register of enrollments Prepares daily lesson plan Signs off on daily roll call Prepared bimonthly lesson plans Analyzes student absences on a monthly or bimonthly basis Monitors student learning Reviews student test results Has tested students during the past two Takes stock of teacher absences months Helps lagging students Informs subdistrict or district administrator of teacher absences Discusses student learning issues with Follows up with teachers on lesson planning school director Source: Authors’ construction based on Government of Madagascar 2004a. far from their place of work, to collect their salaries. In this context, absentee- ism cannot be tackled simply by tightening supervision to ensure that teachers report to work. The problem may require action outside the education sector— for example, to replace the current method of paying teachers with a more con- venient and secure system. With regard to task execution, 42 percent of teachers in the control group performed all the tasks deemed essential for good classroom management, and 24 percent of schools saw all teachers perform all the essential tasks (see table 3). The corresponding values rise to 63 percent and 43 percent, respect- ively, for schools in group 3, and the differences are statistically signi�?cant. In groups 1 and 2 about 53 percent of teachers performed all the tasks, and 30– 36 percent of schools had all teachers perform all tasks, but the differences between these values and those for the control group are not statistically signi�?- cant. These results show that the interventions changed behavior only when targeting the entire chain of service delivery, with schools bene�?ting directly from the interventions and indirectly through the interventions at the subdis- trict and district levels. In schools where the interventions cascade down to schools indirectly through subdistrict and district administrators, the impact on the extent to which teachers perform their essential tasks is limited. The results for well managed schools are similar. The share of well managed schools in group 3 exceeds the corresponding share in the control group by 22 percentage points, and the difference is statistically signi�?cant. The correspond- ing difference for groups 1 and 2 is also positive but not statistically signi�?cant. Even for schools in group 3, however, there is no room for complacency. After bene�?ting from the interventions for two years, the share of well managed schools in this group is still only 37 percent, perhaps reﬂecting the fact that it takes time to change behavior. Comparisons across the three intervention groups can be used to evaluate the existence of spillover effects that might arise from the diffusion of practices T A B L E 3 . Impacts of the Interventions on Providers’ Behaviors and Schooling Outcomes (percent) 318 Differences between intervention Diffusion of the programa Differences with respect to control groupb groupsc Direct to Direct to teachers and Via subdistrict teachers and Via subdistrict school and district Via district school and district Via district Group 2 Group 3 Group 3d Control directors administrators administrators directors administrators administrators versus versus versus Indicator groupa (group 3d) (group 2) (group 1) (group 3d) (group 2) (group 1) group 1 group 1 group 2 Impact on service provider behaviors Teacher absenteeism 9.2 8.7 9.9 10.1 –0.5 0.7 1.0 ratee –0.2 – 1.4 – 1.2 (1.3) (1.4) (1.4) (1.8) (1.7) (1.5) Teachers performing 42.4 63.0 53.6 53.4 20.6** 11.2 11.1 0.2 all tasksf THE WORLD BANK ECONOMIC REVIEW 9.6*** 9.4*** (9.7) (9.6) (9.5) (4.1) (4.2) (3.3) Schools with performing 23.9 42.8 36.4 30.0 18.9** 12.5 6.1 6.4 all all tasksg teachers 12.9*** 6.4 (9.0) (7.7) (7.5) (5.3) (5.3) (6.4) Well 14.6 36.7 23.5 22.0 22.2*** 9.0 7.5 1.5 14.7*** 13.2*** managed (8.5) (8.0) (7.2) (4.4) (5.0) (4.7) schoolsg Impact on students’ schooling outcomes and learning Attendance 86.6 90.7 88.1 89.6 4.1** 1.5 3.0 –1.5 1.1 2.6 rateh (1.9) (2.3) (2.1) (1.1) (0.9) (1.4) Repetition 22.6 17.5 20.0 18.1 – 5.1** –2.6 –4.5** 1.9 – 0.6 – 2.5 ratei (1.9) (1.9) (2.0) (1.2) (1.3) (1.5) Dropout 6.1 5.5 5.5 4.3 – 0.6 –0.6 –1.8 1.2 1.2 0.0 ratej (1.6) (1.5) (1.5) (1.3) (1.4) (1.4) Pass rate at 69.1 73.0 76.3 76.7 3.9 7.2 7.6 –0.4 – 3.7 – 3.3 CEPEk (5.9) (5.9) (5.8) (2.9) (3.2) (2.4) Year two test score l French 29.9 30.5 29.1 29.7 0.6 –0.7 –0.7 0.0 1.3 1.4 (2.5) (2.5) (2.5) (1.1) (1.2) (0.8) Mathematics 50.0 51.7 49.2 48.9 1.6 –0.8 –1.1 0.3 2.7** 2.5** (2.9) (2.9) (2.9) (1.5) (1.5) (1.1) Malagasy 49.8 52.1 50.1 49.8 2.3 0.3 –0.3 0.3 2.3 2.0** (2.4) (2.3) (2.4) (1.3) (1.3) (0.9) All three 42.7 44.3 42.3 42.2 1.5 –0.4 –0.6 0.2 2.1 1.9** subjects (2.5) (2.4) (2.5) (1.2) (1.2) (0.9) *** Signi�?cant at the 1 percent level; ** signi�?cant at the 5 percent level. Note: Numbers in parentheses are standard errors. a. Average; observations are weighted by the probability of selection of the school. b. Standard errors are clustered at the district level. c. Standard errors are clustered at the subdistrict level. d. Interventions at the school level are combined with those at the subdistrict and district levels. e. Absenteeism during the week before the school visit; information provided by the school director; the unit of observation is the school. f. The unit of observation is the teacher. g. Schools with up to �?ve teachers only; the unit of observation is the school. h. Attendance during the month before the school visit; information provided by the school director; the unit of observation is the school. i. Percentage of repeaters in the total enrollment; the unit of observation is the school. j. Rate of net exit; impacts on this rate are evaluated at the end of the �?rst year of the experiment; the unit of observation is the school. k. Percentage of students who pass the Certi�?cat d’Etudes Primaires Ele ´ mentaires (CEPE); impacts on this rate are evaluated at the end of the �?rst year of the experiment; the unit of observation is the school. l. Raw percentage of correct responses; the unit of observation is the student. Source: Authors’ analysis of the 2006/07 Ame ´ lioration de la Gestion de l’Education a ` Madagascar (AGEMAD) school survey, the 2007 AGEMAD posttest, the 2005/06 Fiches de Fin d’Anne ´ e, and the 2006/07 Fiches d’Enque ˆ tes Rapides. Lassibille, Tan, Jesse, and Van Nguyen 319 320 THE WORLD BANK ECONOMIC REVIEW and knowledge by administrators with responsibility for multiple schools in different locations. Schools in group 2 were not better managed than those in group 1, however, suggesting that the interventions that bene�?ted subdistrict administrators did not reinforce the impact of the interventions that bene�?ted the district administrators. Comparing groups 2 and 3 also shows no evidence of diffusion of the package of practices in schools that bene�?ted directly from the interventions to other schools in the same subdistrict and districts under the supervision of the same administrators. These results imply that targeting some schools to serve as demonstration project for replication by the subdistrict and district administrators in the remaining schools under their care is unlikely to be effective.18 Among the possible reasons: lack of interest and motivation by subdistrict and district administrators, which may be linked to their heavy work load and also to their low level of technical competence; weak mechan- isms for monitoring and supervising schools by the subdistrict administrators, who are most directly responsible for this part of the workﬂow in the service delivery chain; and absence of a true leadership culture among both the subdis- trict and district administrators. In subdistricts and districts where only the administrators received the interventions, practically none of the schools received the workﬂow templates designed to tighten management of the peda- gogical processes or training or support,19 and in 75 percent of the schools report cards were never discussed with the communities. A legitimate concern about the conclusions thus far is that the underlying information is based on data reported by teachers and school directors during unannounced school visits by survey enumerators and may be biased since the enumerators did not directly examine relevant workﬂow artifacts (such as teacher attendance records or lesson plans). The latter approach was proposed but rejected by government counterparts as being impractical for two reasons: it would have put the (relatively junior) enumerators in the same position as school inspectors, and it would have required a overly costly commitment of time to capture all the details of the workﬂow processes managed by each teacher and school director in the sample. Instead, an external check on the data collected by the survey enumerators—by gathering and codifying the workﬂow artifacts (cov- ering an entire school year) from a sample of randomly selected schools, 20 in the control group and 20 in intervention group 3—was used (see section III). This unique database reveals how teachers and school directors in each group performed their duties throughout the school year, in particular regard- ing following actions: monitoring and following up on student absenteeism, preparing lesson plans, and tracking progress in student learning (table A2). The data unequivocally indicate that after two years of bene�?tting from the 18. The �?nding that interventions at the subdistrict and district levels have no signi�?cant impact on teachers’ behavior does not rule out the possibility of an impact beyond the two-year time frame of the experiment reported in this article. 19. Only 3 percent of the schools in group 1 and 2 received a copy of the tools from the subdistrict administrators. Lassibille, Tan, Jesse, and Van Nguyen 321 interventions, teachers in group 3 were more conscientious in executing their duties than teachers in the control group, a �?nding that con�?rms the conclusion based on self-reported information. To illustrate, compared with teachers in the control group, those in group 3 monitor student absences for an average of 12 more days during the year; prepare lessons for a larger portion of the school year, covering an average of one more bimester; and record their work in log- books more than 1.5 times as often and for twice as long during the school year. Teachers in group 3 appear to have a better grasp of their students’ edu- cation progress and learning dif�?culties, and the information they communicate to parents via student report cards is more comprehensive and exhaustive. School directors in group 3 are also more conscientious in exercising their supervisory and monitoring duties. Indirect Effects on Students’ Schooling Outcomes and Learning The �?rst of the schooling outcomes pertains to student attendance at the school level, which is based on the number of students present at school during the month before the survey. This information was provided by the school directors during the unannounced school visits. Other measures of schooling outcomes include dropout and repetition rates and pass rates on the CEPE exam taken at the end of the primary cycle—all of which are computed from administrative data maintained by the Ministry of Education.20 Finally, data on learning are from the achievement test scores administered at the end of the experiment. Interventions at the school level reinforced by those at the subdistrict and district levels signi�?cantly improved student attendance compared with the control group, boosting the rate by about 4 percentage points (see table 3). The interventions also had a positive and signi�?cant impact on grade repetition, reducing it by 5.1 percentage points. By contrast, interventions left to the initiative of subdistrict and district administrators had no signi�?cant impact on either student attendance or grade repetition. While the package of interventions targeting school personnel as well as the subdistrict and district administrators that supervise them helped reduce student absenteeism and repetition, it did not reduce dropout rates or improve pass rates on the CEPE exam. Likewise, the interventions targeting only administrators failed to alter any schooling outcomes apart from lowering the repetition rate in intervention group 1. As noted earlier, however, the time frame for evaluation was relatively short, particularly in view of the processes involved in altering beha- viors; the results observed so far may still be partial or incomplete at this time. As for student learning, students in the schools receiving interventions throughout the whole service delivery chain had better scores in all three tested subjects than their peers in the control schools did, but the gains were not 20. For practical reasons, the data on dropout and CEPE pass rate are for 2005/06. 322 THE WORLD BANK ECONOMIC REVIEW statistically signi�?cant.21 The comparisons between group 3 and groups 1 and 2 are also interesting: students in group 3 achieved higher scores than those in group 1 (students in schools that received the interventions through district administrators only), with the gain in mathematics being statistically signi�?- cant. The students in group 3 also outperformed those in group 2 (students in schools that received the interventions through subdistrict and district adminis- trators), with statistically signi�?cant gains in Malagasy and mathematics but not in French. The absence of a pattern of unambiguous gains in test scores across all three subjects is not surprising given the short duration of the evalu- ation period. And the absence of gains in French in any of the comparisons is also consistent with the fact that most Malagasy primary school teachers have a poor grasp of the language themselves, an impediment that better workﬂow processes can hardly be expected to address. In sum, the results on schooling outcomes are consistent with the earlier �?ndings. In its most direct and intensive form, the interventions changed the behavior of all actors toward better management. These changes translated immediately into increases in student attendance and sizable reductions in dropout rates. However, changing service providers’ behavior takes time and effort, and a two-year time frame was probably too short to produce clear-cut impacts on student test scores. V. C O N C L U S I O N Inadequate funding does not appear to be the only reason for the poor per- formance of Madagascar’s primary schools. A detailed analysis of how edu- cation is delivered in schools reveals that many aspects of the pedagogical process are poorly managed and that far too many school personnel and administrators neglect tasks deemed essential for student learning. As men- tioned earlier, 20 percent of teachers do not prepare daily lessons plans, school directors rarely follow up with their teaching staff on student performance, and in only 15 percent of the sample schools do all the teachers and school direc- tors consistently perform the package of seven tasks considered essential by Malagasy educators. There is thus substantial scope to improve the manage- ment of the pedagogical process as part of the country’s effort to raise the per- formance and ef�?ciency of public primary schools. The randomized experiment to evaluate selected interventions to streamline and tighten the workﬂow processes of keys actors yields some interesting results. A package of intensive and direct interventions involving school-level personnel and subdistrict and district administrators changed the behavior of teachers and school directors toward more conscientious execution of the tasks for which they are responsible. It also improved school attendance, reduced grade rep- etition, and raised test scores, although the gain in scores was not statistically 21. Larger sample sizes might have detected statistically signi�?cant gains. Lassibille, Tan, Jesse, and Van Nguyen 323 signi�?cant. A laissez-faire (and less costly) version of the interventions, targeting only the subdistrict and district administrators—in the hopes that they would in turn disseminate and supervise implementation of the interventions to improve workﬂow processes at the school level—proved largely ineffective. Because the experiment did not test the impact of simply providing resources without the managerial changes, the relative roles of these factors cannot be separated out. Beyond their intrinsic interest, the results also offer a basis for exploring the implications for policy development in Madagascar. Scaling up the most effective interventions among those evaluated—that is, interventions directly targeted to teachers and school directors and reinforced by support from subdistrict and dis- trict administrators—by simply replicating the arrangements used during the experiment would almost certainly cost too much. The reason is that during the experiment, the option of relying on existing institutions to carry out the training of teachers and to facilitate community-school meetings was unavailable. Thus, a separate team had to be hired and trained to implement the AGEMAD interven- tions; this approach was also the best way to adhere to the strict requirements of the experiment for implementation and data collection. In the scaling up phase, a potentially more sustainable strategy would be to integrate the AGEMAD inter- ventions into existing structures and programs. For example, ongoing training programs for teachers, school directors, and subdistrict and district administrators could include short modules to familiarize participants with the AGEMAD tem- plates and processes for workﬂow management and train them in using the tools. Similarly, the procurement and distribution of the workﬂow templates and the facilitation of school meetings could be integrated into ongoing initiatives to decentralize funding and decisionmaking to schools, districts, and regions through the system of localized school improvement grants. More broadly, scaling up the AGEMAD interventions effectively means motivating a large number of actors in the system to change their behavior at work. The task is daunting at best and will almost certainly require sustained effort to foster a culture of results-oriented management and leadership throughout the system. Finally, beyond the implications for Madagascar, the impact evaluation reported in this article also enriches discussion of promising approaches for improving educational outcomes in low-income countries. The results suggest that packaging an increase in resources to schools in the form of support for improving workﬂow processes and accountability measures (that is, teachers’ personal engagement in their work, proper monitoring and supervision of their work, involvement of parents and the community, and the like) is worth con- sideration. Such changes are probably especially relevant in dysfunctional schools—where teachers neglect their essential pedagogical duties, where school directors routinely fail to support and supervise the teachers, and so on. Future research could therefore be directed at enhancing knowledge about the nature of the problems and promising interventions in this regard. Such knowl- edge would inform the design of policies and programs for education advance- ment in low-income countries. 324 THE WORLD BANK ECONOMIC REVIEW APPENDIX F I G U R E A1. School Report Card Source: Authors’ construction. Lassibille, Tan, Jesse, and Van Nguyen 325 F I G U R E A2. Experimental Design and Description of the Interventions Note: Size of the follow-up sample in parentheses. Source: Authors’ construction. T A B L E A 1 . Baseline School Characteristics and Attrition in End Line Test Scores Differences with respect to control group Direct to Via subdistrict and teachers and district Via district Control school directors administrators administrators Characteristic group (group 3a) (group 2) (group 1) Baseline school characteristics Number of students 239.5 – 4.8 – 25.3 – 39.3 (28.2) (24.9) (23.3) Number of sections 5.7 – 0.0 – 0.4 – 0.4 (0.3) (0.3) (0.2) Number of teachers 4.4 0.0 – 0.4 – 0.2 (0.4) (0.3) (0.3) Repetition rate (percent) 17.7 3.7 2.2 2.9 (2.0) (2.0) (2.0) Pass rate at CEPEb 60.7 1.7 1.4 1.7 (percent) (3.0) (3.0) (3.0) (Continued ) 326 THE WORLD BANK ECONOMIC REVIEW TABLE A1. Continued Differences with respect to control group Direct to Via subdistrict and teachers and district Via district Control school directors administrators administrators Characteristic group (group 3a) (group 2) (group 1) Year one test scoresc 60.9 0.7 – 1.0 – 1.0 (percent) (1.9) (1.8) (1.9) Loss of sample for year two test Loss rated (percent) 11.1 3.1 1.4 3.4 (2.2) (1.5) (2.2) Difference in year one test – 0.68 2 2.7 2 0.2 2 2.0 scores between quitters (1.3) (1.6) (1.6) and stayers in year 2e (percentage points) Note: Numbers in parentheses are standard errors clustered at the district level. a. Interventions at the school level are combined with those at the subdistrict and district levels. b. Percentage of students who pass the Certi�?cat d’Etudes Primaires Ele ´ mentaires (CEPE). c. Raw percentage of correct responses in French, Malagasy, and mathematics. d. Percentage of students who did not take the year two test scores. e. Differences in the raw percentage of correct responses in French, Malagasy, mathematics in year one test scores between students according to their status in year 2 as quitters or stayers. Source: Authors’ analysis of the 2005/06 Fiches d’Enque ˆ tes Rapides, the 2004/05 Fiches de Fin d’Anne´ e, and the Ame´ lioration de la Gestion de l’Education a ` Madagascar pretest. T A B L E A 2 . Task Execution Rates by Teachers and School Directors in a Subsample of Schools (percent, unless otherwise indicated) Control Direct interventions to teachers Task schools and school directors (group 3a) Roll call (average over the school year) Number of days task was performed per month 13.0 14.5 Monthly sheets with a recapitulation of rate of 21.0 76.0 student absenteeism Monthly sheets not signed by the school 67.0 46.2 director Weekly lesson plans Teachers not using the tool 13.6 1.5 Teachers using the tool at least once in the year 86.4 98.5 Frequency of use (average number of 2.8 4.2 bimestrial segments) Teachers using weekly lessons plans 21.1 55.1 throughout the year (Continued ) Lassibille, Tan, Jesse, and Van Nguyen 327 TABLE A2. Continued Control Direct interventions to teachers Task schools and school directors (group 3a) Teachers not reporting the subtitle of the 74.1 33.3 lessons Teachers not annotating their observations on 60.5 30.3 the lesson plans Weekly lesson plans not signed by the school 49.0 35.2 director Daily lesson plan Teachers not using the tool 52.4 21.7 Teachers using the tool at least once in the year 47.6 78.3 Average number of days used in the year 34.1 62.7 (maximum 150) Weekly sheets with annotated observations 11.5 30.8 Teachers who have followed the weekly 38.8 51.3 lesson plans Weekly sheets not signed by the school 71.9 60.2 director Record of tests Evaluation reports with no annotation on the 62.7 52.1 number of students missing their tests Evaluation reports with no annotation on the 74.4 14.9 number of students performing above the class average Evaluation reports not signed by the school 80.0 59.0 director Student report cards With missing information on student’s class 11.7 5.0 rank With no information on the average score of 94.4 33.9 students in the class With no observations on student’s attendance 46.9 28.1 Not signed by the teacher 16.7 13.2 Not signed by the school director 80.2 62.8 Teacher’s travel pass for authorized trips b Teachers not using the tool 92.0 74.0 Teachers using the tool at least once in the year 8.0 26.0 Pass is signed on teacher’s arrival at 13.8 68.8 destination by the relevant authority Pass is signed on teacher’s leaving the locality 20.7 81.3 by the relevant authority a. Interventions at the school level are combined with those at the subdistrict and district levels. b. Teachers are routinely granted permission for travel to collect their salaries at designated localities. These trips may result in excessive absences from the classroom unless adequately supervised. Source: Authors’ analysis of the 2006/07 Ame ´ lioration de la Gestion de l’Education a ` Madagascar tools survey. 328 THE WORLD BANK ECONOMIC REVIEW REFERENCES Adekanmbi, A., M. Blimpo, and D. Evans. 2009. The State of The Gambia Lower Basic Education. Banjul, The Gambia: Department of State for Basic and Secondary Education. Banerjee, A., R. Banerji, E. Duﬂo, R. Glennerster, and S. Khemani. 2006. Can Informational Campaigns Spark Local Participation and Improve Outcomes? A Study of Primary Education in Uttar Pradesh, India. Policy Research Working Paper 3967. Washington, D.C.: World Bank. ———. 2008. Pitfalls of Participatory Programs: Evidence from a Randomized Evaluation in Education in India. Policy Research Working Paper 4584. Impact Evaluation Series 21. Washington, D.C.: World Bank. Banerjee, A., S. Cole, E. Duﬂo, and L. Linden. 2007. “Remedying Education: Evidence from Two Randomized Experiments in India.�? Quarterly Journal of Economics 122(3): 1235–64. Banerjee, A., and E. Duﬂo, 2006. “Addressing Absence.�? Journal of Economic Perspectives 20(1): 117 –32. Banerjee, A., S. Jacob, M. Kremer, J. Lanjouw, and P. Lanjouw. 2005. “Promoting School Participation in Rural Rajasthan: Results from Some Prospective Trials.�? Working Paper. Massachusetts Institute of Technology, Cambridge, Mass. Bobonis, G., and F. Finan. 2009. “Neighborhood Peer Effects in Secondary School Enrollment Decisions.�? Review of Economics and Statistics 91(4): 695– 716. Boruch, R. 1997. Randomized Experiment for Planning and Evaluation: A Practical Guide. London: Sage. Chaudhury, N., J. Hammer, M. Kremer, K. Muralidharan, and H. Rogers. 2006. “Missing in Action: Teacher and Health Worker Absence in Developing Countries.�? Journal of Economic Perspectives 20(1): 91 –116. Dennis, M., and R. Boruch. 1989. “Randomized Experiments for Planning and Testing Projects in Developing Countries: Threshold Conditions.�? Evaluation Review 13: 292– 309. Duﬂo, E., P. Dupas, and M. Kremer. 2007. “Peer Effects, Pupil-Teacher Ratios, and Teacher Incentives.�? Working Paper. Massachusetts Institute of Technology, Cambridge, Mass. Duﬂo, E., R. Glennester, and M. Kremer. 2007. Using Randomization in Development Economics Research: A Toolkit. Discussion Paper 6059. London: Centre for Economic Policy Research. Evans, D., M. Kremer, and M. Ngatia. 2008. “The Impact of Distributing School Uniforms on Children’s Education in Kenya.�? Working Paper. Harvard University, Cambridge, Mass. Gertler, P., H. Patrinos, and M. Rubio-Codina. 2008. Empowering Parents to Improve Education: Evidence from Rural Mexico. World Bank Policy Research Working Paper 3935. Washington, D.C.: World Bank. Glewwe, P., N. Ilias, and M. Kremer. 2008. “Teacher Incentives.�? Working Paper. Harvard University, Cambridge, Mass. Glewwe, P., M. Kremer, and S. Moulin. 2009. “Many Children Left Behind? Textbooks and Test Scores in Kenya.�? American Economic Journal: Applied Economics 1(1): 112– 35. Glewwe, P., M. Kremer, S. Moulin, and E. Zitzewitz. 2004. “Retrospective vs. Prospective Analyses of School Inputs: the Case of Flipcharts in Kenya.�? Journal of Development Economics. 74(1): 251–68. Hanushek, E. 1997. “Assessing the Effects of School Resources on Student Performance: An Update.�? Educational Evaluation and Policy Analysis 19(2): 141– 164. Hanushek, E., and M.A. Raymond. 2004. Does School Accountability Lead to Improved Student Performance? NBER Working Paper Series 10591. Cambridge, Mass.: National Bureau of Economic Research. Haveman, R., and B. Wolfe. 1995. “The Determinants of Children’s Attainment: A Review of Methods and Findings.�? Journal of Economic Literature 33(4): 1829– 78. Lassibille, Tan, Jesse, and Van Nguyen 329 Jamison, D., B. Searle, K. Galda, and S. Heyneman. 1981. “Improving Elementary Mathematics Education in Nicaragua: An Experimental Study of the Impact of Textbooks and Radio on Achievement.�? Journal of Educational Psychology 73(4): 556– 67. King, E.M., and P.F. Orazem. 1999. “Evaluating Education Reforms: Four Cases in Developing Countries.�? World Bank Economic Review 13(3): 409– 13. Kling, J.R., J.B. Liebman, and L.F. Katz. 2007. “Experimental Analysis of Neighborhood Effects.�? Econometrica 75(1): 83– 119. Kremer, M., and A. Holla. 2009. “Improving Education in the Developing World: What Have We Learned from Randomized Evaluations?�? Annual Review of Economics 1: 513– 42. Kremer, M., S. Moulin, and R. Namunyu. 2003. “Decentralization: A Cautionary Tale.�? Working Paper. Harvard University, Cambridge, Mass. ´ lioration de l’Enseignement – Madagascar, Government of. 1997. Programme National pour l’Ame PNAE2 du MinESEB. Antananarivo: Government of Madagascar. ———. 2004a. Ame ´ lioration de la Gestion de l’e ` Madagascar : approche strate ´ ducation a ´ gique et plan d’actions. Antananarivo: Government of Madagascar. ———. 2004b. Rapport sur le suivi des objectifs du mille ´ naire pour le de ` Madagascar. ´ veloppement a Antananarivo: Government of Madagascar. ———. 2008. Ele ` me e ´ ments de diagnostic du syste ´ ducative ´ ducatif malagasy. Le besoin d’une politique e ´ naire et la re nouvelle pour l’atteinte des objectifs du mille ´ . Antananarivo: ´ duction de la pauvrete Government of Madagascar. Miguel, E., and M. Kremer. 2004. “Worms: Identifying Impacts on Education and Health in the Presence of Treatment Externalities.�? Econometrica 72(1): 159– 217. PASEC (Programme d’Analyse des Syste ` mes Educatifs de la Confemen). 2007. Quelques pistes de ´ ﬂexion pour une e re ´ ducation primaire de qualite ´ pour tous. Rapport Madagascar. Dakar: Confe ¸ ais en Partage. ´ rence des Ministres de l’Education des Pays ayant le Franc Tan, J.P., J. Lane, and G. Lassibille. 1999. “Student Outcomes in Philippine Elementary Schools: An Evaluation of Four Experiments.�? World Bank Economic Review 13(3): 493– 508. Todd, P., and K.I. Wolpin. 2006. “Assessing the Impact of a School Subsidy Program in Mexico: Using a Social Experiment to Validate a Dynamic Behavioral Model of Child Schooling.�? American Economic Review 96(5): 1384–1417. World Bank. 2002. Education and Training in Madagascar: Toward a Policy Agenda for Economic Growth and Poverty Reduction. Washington, D.C.: World Bank. ———. 2008. Service Delivery in the Education and Health Sectors in Madagascar. Washington, D.C.: World Bank. Potential Implications of a Special Safeguard Mechanism in the World Trade Organization: the Case of Wheat Thomas W. Hertel, Will Martin, and Amanda M. Leister The special safeguard mechanism—both quantity- and price-based—was key in the July 2008 failure to reach agreement in the World Trade Organization negotiations under the Doha Development Agenda. A stochastic simulation model of the world wheat market is used to investigate the effects of the special safeguard mechanism. As expected, the quantity-based safeguard is found to reduce imports, raise domestic prices, and boost mean domestic production in the countries that implement it. However, rather than insulating developing countries in those regions from price vola- tility, the quantity-based safeguard increases domestic price volatility, largely by restricting imports when domestic output is low and prices are high. The quantity- based safeguard shrinks average wheat imports nearly 50 percent in some regions, and global wheat trade falls by 4.7 percent. The price-based safeguard discriminates against lower price exporters and contributes to producer price instability. G33, Doha Development Agenda, Gaussian quadrature, safeguard, special safeguard mechanism, SSM, wheat, World Trade Organization, WTO, trade and development. JEL codes: F1, F13, F51, O24, Q1, Q17. The special safeguard mechanism was key in the July 2008 failure to reach agreement in the World Trade Organization (WTO) negotiations under the Doha Development Agenda. The draft agreement would allow members to impose speci�?ed additional duties when the volume of imports of an agricul- tural product exceeds a speci�?ed level or when import prices from a particular supplier fall below a speci�?ed price (WTO 2008a). Given the substantial gains available under the Doha Development Agenda (Martin and Mattoo 2008), the fact that the negotiations were unable to proceed for lack of consensus on this issue highlights its importance to many WTO members. Wolfe (2009) Thomas W. Hertel (corresponding author; hertel@purdue.edu) is Distinguished Professor and Executive Director of the Center for Global Trade Analysis at Purdue University. Will Martin (wmartin1@worldbank.org) is a Research Manager in the Development Research Group of the World Bank. Amanda M. Leister (aleister@purdue.edu) is a graduate research assistant at the Center for Global Trade Analysis at Purdue University. The authors wish to acknowledge valuable comments provided by the editor of the journal and three anonymous reviewers. THE WORLD BANK ECONOMIC REVIEW, VOL. 24, NO. 2, pp. 330– 359 doi:10.1093/wber/lhq010 Advance Access Publication August 18, 2010 # The Author 2010. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oxfordjournals.org 330 Hertel, Martin, and Leister 331 attributes the breakdown of negotiations primarily to inadequate analysis of the safeguard operation and implications. Analysis based speci�?cally on the proposed measure was limited because the proposed safeguard mechanism is one of the most technically complex aspects of the entire modalities, because it attempts to deal with variations in prices rather than—as with tariffs—merely their level, and because it was presented to ministers only a few days before the meeting. Agricultural producers in developing countries are vulnerable to both dom- estic shocks—particularly from weather-related shocks to output—and shocks to international markets. However, developing country consumers are also par- ticularly vulnerable to shocks to food prices, given that the poorest people spend as much as 75 percent of their incomes on food. Policies that raise food prices by imposing an import duty may help farmers whose incomes have fallen due to a harvest shortfall but do so at the expense of net buyers of food—including many farmers. Farmers isolated from world markets by poor infrastructure and communications receive little or no bene�?t from protection that raises the cost of food to poor consumers linked to world markets. This highlights the need for careful analysis of the impact of the special safeguard mechanism that takes into account the differences between imported and dom- estic goods. It is important to consider the implications of the special safeguard mechan- ism for global markets because they would apply to all developing countries, which account for about two-thirds of the value of world agricultural pro- duction. Thus, this article assesses the global implications of the proposed price- and quantity-based safeguards for a key agricultural staple, wheat, taking into account not just their direct impacts on import prices but also the resulting impacts on world prices when many developing countries use them at the same time. The analysis also traces the resulting impacts on key variables such as the volume of imports, domestic producer prices, and the returns to land on which the incomes of many farm households depend. Both the average impact and the volatility of these variables are considered, since part of the motivation for the safeguard is to reduce volatility by offsetting shocks from international markets. Countries can use either a quantity-based or price-based safeguard in a given year. While it would be interesting to consider a situation where countries could choose between the two types at each point in time, it is unclear which option policymakers would choose when both are available. Therefore, the article focuses on the important prior objective of assessing each type separately. The article is organized as follows. Section I examines some of the key prior contributions to the literature on the use of special safeguards. Section II con- siders the nature of the speci�?c proposals under consideration. Section III intro- duces a diagrammatic assessment of the qualitative effects of such interventions, including an analysis of the extent to which it might be used. 332 THE WORLD BANK ECONOMIC REVIEW Section IV discusses how the special safeguard mechanism can be used. Section V explains the empirical model used to estimate the potential implications of the safeguard for domestic and global markets. Section VI applies the model to the quantity-based safeguard, and section VII applies it to the price-based safe- guard. Section VIII compares the two sets of results. Section IX offers sugges- tions for future research. I . WH AT THE L I T E R AT U R E S AY S A B O U T THE SPECIAL SAFEGUARD MECHANISM While much technical work was available at the time of the Doha ministerial, many key questions either had not been asked or had not been satisfactorily resolved. Montemayor (2007, 2008) and Valde ´ s and Foster (2005) focus on the broad impacts of different duty rates on imports into individual countries without taking into account the impacts on world markets. Not surprising, much of this initial work examines the frequency with which the special safe- guard measure could be used rather than on whether it would achieve its under- lying economic goals, such as moderating the impact of commodity market volatility on the incomes of farmers and the living costs of poor consumers. Valde ´ s and Foster (2005) rule out the quantity-based safeguard a priori, arguing that increased import quantities are likely to be due to declines in har- vests, making it dif�?cult to justify import restrictions. They also express concern about developing countries’ dif�?culties in maintaining data on imports, and the inevitable lags between increases in imports and the implementation of safeguards. Finger (2009) raises several other important questions about the proposed special safeguard mechanism. For instance, would the mechanical trigger rules allow import duties to be imposed when import prices are constant or rising? What objectives of the safeguard would be consistent with such mechanical rules? And would use of a quantity-based trigger reduce—or actually increase—the variability of domestic prices by raising duties during periods of short domestic supply? He also raises questions about the shipment-by-shipment nature of the duty calculation of the price-based safe- guard. After calculating duties under the safeguard by comparing the price of each shipment with the average price of all shipments, he �?nds that countries that export lower price products—typically developing countries—would likely face considerably above average safeguard tariffs. De Gorter, Kliauga, and Nassar (2009) suggest that under both the quantity- and price- based triggers most invocations of the safeguard in China, India, Indonesia, and the Republic of Korea would be against exports from develop- ing countries. Grant and Meilke (2009) made an important step forward in the analysis of the proposed special safeguard mechanism by taking into account the potential impact of the safeguard as proposed in the July 2008 modalities on international Hertel, Martin, and Leister 333 and domestic prices. They �?nd that the safeguard increases the volatility of world prices. Although they believe that the impacts on world markets overall would be fairly modest, several developing countries—most notably in the Middle East and North Africa—would experience large increases in the vola- tility of domestic prices. Because Grant and Meilke use a net trade model (where imports are not linked to particular exporters), they were unable to deal with the issue raised by Finger (2009) regarding the discriminatory nature of a price-based safeguard toward lower price developing country exports. The framework outlined in section II enriches the analysis of Grant and Meilke (2009) by incorporating the important features of key agricultural pro- ducts such as wheat, which show evidence of strong differentiation by country of origin (Uri and Beach 1997). This differentiation, due partly to differences in physical qualities of wheat from different countries and partly to less tangi- ble factors such as differences in the terms and conditions of sale, results in price differences that inﬂuence the extent to which the price-based safeguard is invoked. Therefore, this article examines the price- and quantity-based safe- guards within a modeling framework, as proposed in the draft WTO modal- ities of December 2008, that allows for differences in relative prices of exports from different suppliers, thereby permitting Finger’s (2009) hypothesis of dis- crimination against developing country exporters to be tested. I I . F E AT U R E S OF THE PROPOSED SPECIAL SAFEGUARD MECHANISM The impacts of the special safeguard mechanism likely depend substantially on its design. The one under discussion is based broadly on the current special agricultural safeguard, which includes two triggers—one based on the price of imports and one on the quantity of imports (GATT 1994). In contrast with standard WTO safeguards under Article XIX of the General Agreement on Tariffs and Trade, there is no requirement to demonstrate that imports have caused injury to domestic producers. The price-based safeguard uses a reference price based on a three-year moving average of import prices from all sources (WTO 2008a). When the price of an individual shipment falls below 85 percent of the reference price, a duty can be used to remove 85 percent of the shortfall. One important feature of this shipment-by-shipment trigger is that it imposes higher duties on imports from lower price exporters. Finger (2009) and de Gorter, Kliauga, and Nassar (2009) argue that a price-based safeguard generally imposes higher duties on exports from developing countries. The quantity-based safeguard can be used when imports in a year exceed base imports—a three-year moving average of imports.1 The duty that can be 1. Since imports in any one year are compared with a three-year moving average of past imports, steady annual import growth of 5 percent compounds to a “surge�? in imports of more than 10 percent, against which a safeguard can be imposed. 334 THE WORLD BANK ECONOMIC REVIEW applied increases as imports exceed this base. Imports of 110–115 percent of the base allow an additional duty of 25 percent of the current binding or 25 percentage points, imports of 115–135 percent of the base allow an additional duty of 40 percent of the binding or 40 percentage points, and imports of more than 135 percent of the base allow an additional duty of 50 percent of the binding or 50 percentage points. A quantity-based safeguard can be imposed for only two years, and if used four years in a row, cannot be used for another two years. If a safeguard duty is imposed and imports fall below the level in the period before imposition, the trigger level is not reduced—thus preventing the duty itself from causing the trigger level to decline. The draft modalities do not, in general, permit total applied duties to exceed the pre-Doha limit. A major focus of debate has been on exceptions to this limit for the quantity-based safeguard, and two speci�?c proposals have been advanced. The “Lamy compromise�? would permit duties to exceed the bind- ings by 15 percentage points on 2.5 percent of tariff lines when imports exceed the base by 40 percent (ICTSD 2008). The compromise proposed by the G-33 (2008) and its negotiating partners would permit tariffs up to 30 percent (or percentage points) above the pre-Doha bindings on 7 percent of tariff lines when imports exceed 110 percent of base levels. The draft modalities consider increases of 12 percent and 15 percent above the bound rate (WHO 2008a, para. 145). The next section examines the qualitative implications of using the quantity and price-based safeguards as a guide to understanding the model-based results in subsequent sections. I I I . Q U A L I TAT I V E I M PAC T S O F US I N G T H E P R I C E - AND QUANTITY-BASED SAFEGUARDS The impacts of a price-based safeguard in a small trading economy are straight- forward. First, consider the market for a single imported food crop, as shown in �?gure 1. The domestic supply of the good is shown by the curve S, while the demand is represented by curve D. The world price falls from an initial level of p0 to p1. Introducing a duty of t completely offsets the decline in the domestic price.2 A partially offsetting duty that diminished the size of the reduction in domestic prices by 85 percent would reduce the variance of domestic prices in response to this type of shock to 2 percent of its original level. Imports would, of course, decline relative to their level without the safe- guard. Had domestic prices fallen from p0 to p1, imports would have increased from (q0 – d0) to (q1 – d1). For a small economy in which producer output is distributed independently of world output, average farm income would rise and the variability of farm income would decline. The average cost of food to consumers would rise because of the safeguard tariff, but the variability in the 2. Complete stabilization would require a full set of taxes and subsidies on imports and exports. Hertel, Martin, and Leister 335 F I G U R E 1. Impacts of a Decline in World Prices in a Single Country Source: Authors’ design. cost of food would decline. Consumers eat less food because of its higher price, which generates an economic cost measured by area def in �?gure 1. Another cost—measured by area bcg—arises because lower cost imports are replaced by higher cost domestic production. If a safeguard is introduced in a large group of countries, the world market price for the commodity is no longer constant. In this case, it is useful to con- sider import demand from the group of countries using the safeguard (ED) together with the export supply (ES) from the rest of the world, as illustrated in �?gure 2. An increase in supply—perhaps from a large harvest—that shifts the excess supply curve from the rest of the world from ES to ES0 would, F I G U R E 2. Implications of a Price-Based Safeguard for the World Market Source: Authors’ design. 336 THE WORLD BANK ECONOMIC REVIEW without a safeguard, cause the world price to decline from p0 to p1. The decline in prices in importing countries would cause their imports to increase from m0 to m1. If a safeguard reduces the decline in import prices in importing markets, the decline in world prices must be larger, because more of the price adjustment is forced onto the exporting countries. If 85 percent of the decline in world prices is offset by a safeguard, the increase in imports for a given reduction in world prices is reduced to 15 percent of its level in the absence of a safeguard. Thus, world prices would decline further, as illustrated by p2 in �?gure 2. For the importing countries, the reduction in the world price to p2 resulting from the safeguard requires a second-round increase in the safeguard duty on top of that shown in �?gure 1. For each country, the decline in the world price is not just the initial reduction from p0 to p1 but that from p0 to p2 shown in �?gure 2. Average world prices decline, since the safeguard sometimes increases—and never reduces—duties. Another key impact of the widespread use of a safeguard in importing countries is an increase in the volatility of world prices (see Tyers and Anderson 1992). An important implication of the analysis in �?gure 2 is that, when analyzing the impacts of introducing a safeguard that covers all developing countries, it may not be enough to simply consider experience in the absence of a safeguard. Once the safeguard is introduced in several important markets, the volatility of world prices is likely to be greater than would otherwise be the case. If this effect is large, it will increase the probability that the safeguard will be trig- gered in any period. As noted in Fraser and Martin (2008) and Valde ´ s and Foster (2005), the implications of a quantity-based safeguard depend heavily on the cause of the shock. If the cause is a decline in world prices of the type shown in �?gure 1, for instance, imports rise from (q0 – d0) to (q1 – d1). If this decline is large enough to trigger the quantity-based safeguard, a quantity-based safeguard could be an alternative to a price-based safeguard. If the same additional duty were generated by both safeguards, there is no effective difference. Because the link between the size of the price decline and the tariff imposed under the quantity-based safeguard is weak, this safeguard may permit a larger response than the price-based safeguard and may even cause the domestic price to rise when the import price falls. If the world price does not decline but imports increase, the quantity-based safeguard can be triggered even though the price-based safeguard is not. In this situation, it is important to examine the cause of the increase in imports. In agriculture, such an increase is likely due either to a shift in the domestic supply curve—such as a decline in the harvest associated with poor weather conditions—or to an increase in demand, as considered by Sen (1981). The South Centre (2009) concludes that more than 85 percent of import surges are not accompanied by declines in import prices, suggesting that most import surges are driven by domestic shocks, such as declines in domestic production. Hertel, Martin, and Leister 337 F I G U R E 3. Potential Effects of a Quantity-Based Safeguard Source: Authors’ design. Figure 3 focuses on a reduction in domestic supply. Domestic supply is initially shown by the supply curve S, which shows domestic production of q0 at price p0. Domestic demand is represented by curve D, and demand at price p0 by d0. Imports are initially given by (q0 – d0). Without a quantity-based safeguard, a decline in domestic supply from S to S0 does not affect the dom- estic price. Imports increase to make up the larger gap between domestic demand and supply, allowing the domestic price to remain stable. If a quantity- based safeguard is used, the effect is to apply an additional duty—and thus to raise the domestic price. The effect of a quantity-based safeguard in this situation is to destabilize the domestic price. For consumers, the adverse impact on prices can be avoided by importing to make up the shortfall. For producers, prices are destabilized, but revenues and net returns may be stabilized or destabilized.3 If the tariff imposed is slightly larger than the decline in the quantity of output, producer gross revenues will be stabilized.4 However, the effect on producer net returns may differ, depending on the nature of the shift in the supply curve (Martin and Alston 1994, 1997). Imposing a quantity-based safeguard would reduce imports below the level that would have prevailed in the absence of such a measure. In �?gure 3, the initial level of imports is given by (d0 2 q0). Without a safeguard, imports would rise to (d0 2 q00 ). A safeguard would reduce imports below this level, 3. Consumers and producers are typically not distinct groups—particularly in poor countries. Many farmers are net buyers of staple foods, and some households classi�?ed as urban for survey purposes are net sellers of these products (Ivanic and Martin 2008). 4. For a small change in output, the proportional effect on producer revenues is given by @ p/p þ @ q/ q where @ p is the change in price and @ q the changes in quantity. For larger changes, the interaction effect between the change in price and quantity becomes signi�?cant. 338 THE WORLD BANK ECONOMIC REVIEW perhaps to a level similar to (d1 2 q10 ). Whether this is greater than or less than the initial level of imports is unclear. With the quantity-based safeguard, there is a link between the extent of import penetration before the safeguard and the size of the duty that can be imposed during the following 12 months. A 35 percent increase in imports would allow an additional duty of 50 percentage points to be imposed. A relatively high elasticity of demand for imports at the tariff-line level, as is usually assumed, may be enough to reduce imports sub- stantially. However, the short-term nature of the measure makes signi�?cant supply response unlikely, lowering the probability that imports would be reduced below their initial level. I V. H O W THE SPECIAL SAFEGUARD MECHANISM MIGHT BE USED Because the special safeguard mechanism provides an option, but not an obli- gation, to protect, it is dif�?cult to be sure how frequently it would be used. One view is that most developing countries have considerable binding over- hang, with their bound tariffs considerably above their applied rates, so it is unlikely that a safeguard would make a signi�?cant difference to the protection allowed under WTO rules. Another is that decisions about the duties applied under a safeguard are likely to be taken in a different forum from those regard- ing ordinary tariffs, which may have real implications for choices about border protection. In many countries, applied tariff levels are decided by a tariff com- mittee, which includes representatives from different parts of government and which frequently takes a broad view about the desirability of low tariffs for export competitiveness and the overall ef�?ciency of the economy. A body with a narrower focus may be more willing to provide protection that bene�?ts pro- ducers in a particular sector. One promising approach to assessing the likely use of the special safeguard mechanism is to examine the frequency with which the special agricultural safeguard provided under the Uruguay Round has been used. Morrison and Sharma (2005) conclude that the ratio of times that the special agricultural safeguard was invoked to times that it could have been invoked was about 1 percent. They suggested three reasons for nonapplication: the complexity of the formulas; high tariff bindings in many developing countries, which make it feasible to raise applied tariffs or apply additional duties without exceeding bound rates; and a judgment that the costs of introducing such measures exceed the bene�?ts. Several other reasons for the limited use of the special agricultural safeguard have been offered. Finger (2009) notes that many major users of the special agricultural safeguard posted minimum prices when the Uruguay Round agree- ment came into effect. Under these circumstances, exporters knew that pricing at a lower level would result in a duty and thus had a strong incentive to price at the minimum posted price so that the safeguard would not be invoked. Hallaert (2005) suggests that many members have ignored the requirement to Hertel, Martin, and Leister 339 report use of the special agricultural safeguard to the WTO’s Committee on Agriculture within 10 days of implementation, though he concludes that the use of the safeguard has increased as WTO members become more familiar with its provisions. Another possible reason is that the special agricultural safeguard was often politically unattractive because of the weak relationship between its mechanical formulas and policymaker goals. While the quantity-based special agricultural safeguard might permit the use of safeguards following a crop failure, policy- makers may not have wanted to use the safeguard under those circumstances because of pressure from consumers concerned about high food prices. The next section turns to the empirical framework, which permits a more thorough assessment and comparison of the implications of the proposed quantity- and price-based safeguards. V. E M P I R I C A L F R A M E W O R K AND SCENARIO DESIGN The analysis builds on Valenzuela and others (2007), who use a stochastic simulation approach to validating computable general equilibrium models, with a focus on the world wheat market. This study employs a more recent version of the Global Trade Analysis Project model that has been speci�?cally tailored to agricultural applications (Keeney and Hertel 2005). It incorporates segmented factor markets to mimic short-run rigidities in supply response and more detailed information about supply and demand elasticities pertinent to agricultural production and food consumption.5 The Armington import demand speci�?cation with econometrically estimated elasticities of substitution between varieties of wheat in the model is used to allow for differentiation between wheat produced in different countries (Hertel and others 2007). As seen below, product differentiation by origin plays an important role in the price-based safeguard. Since demand for wheat is relatively stable and most shocks to the wheat market come from weather-induced shocks to production, supply-side shocks are introduced into the model. Speci�?cally, total factor productivity in wheat in each model region is shocked by sampling from historical distributions of supply deviations from trend in all world regions.6 The approach used in this stochastic simulation ensures that each time the impacts of a new policy regime 5. This model is �?rst validated based on historical variation in production and prices, following the approach proposed by Valenzuela and others (2007). For more details, see the working paper version of this article (Hertel, Martin, and Leister 2010). 6. Standard stochastic simulation techniques such as Monte Carlo procedures are cumbersome at best, given the large number of variables in the model so Valenzuela and others (2007) are followed in approximating the distribution of supply shocks using Gaussian quadrature. This has been shown to be an ef�?cient means of assessing the consequences of stochastic variation in parameters of the shocks to computable general equilibrium models (DeVuyst and Preckel 1997) and its implementation has been automated in the GEMPACK software used to solve the model (Arndt 1996; Pearson and Arndt 2000). 340 THE WORLD BANK ECONOMIC REVIEW are simulated, an identical set of stochastic shocks is administered. This elimin- ates the possibility that differences in the sample of supply-side shocks contrib- ute to differences in outcomes across policy regimes. Three sets of stochastic simulations are performed. The �?rst set establishes the baseline (no safeguard). This case assumes that tariffs remain �?xed at the level of scheduled applied tariff rates for 2001, except when countries made inter- national commitments to lower their WTO bound tariff rates—as in the case of China’s accession to the WTO—or to lower tariffs on a preferential basis. The second set of stochastic simulations permits developing countries to invoke the quantity-based safeguard, as detailed in the next section. The analy- sis focuses on the differences between the expected mean and standard devi- ation of key variables, which are computed as the outcome under the quantity-based safeguard minus the outcome under the baseline. The third set of stochastic simulations allows developing countries to implement the price- based safeguard, as detailed below. Again, the focus is on differences in the expected percentage change in the mean and standard deviation, computed as the price-based safeguard value minus the baseline value. The percentage change in the mean and standard deviation of model variables under any indi- vidual policy regime are in appendix tables A1–A5. The special safeguard mechanism duties considered in the second and third sets are distinct from—and additional to—initial applied tariff rates, in the same way that antidumping duties and Article XIX safeguard duties are in addition to scheduled applied tariff rates. Many developing countries can raise applied rates relative to bound rates, with China a notable exception (see below). All other regions are modeled according to the draft modalities, and applied tariffs plus the endogenously determined safeguard remain below the bound rates in all cases. VI. IMPLEMENTING THE QUANTITY-BASED SPECIAL SAFEGUARD MECHANISM The quantity-based trigger permits developing countries to apply a tariff on imports whenever trade volumes reach 110 percent of a three-year moving average. The resulting tariff can be as high as 25 percent of the bound tariff or 25 percentage points, whichever is higher (tier 1). If imports exceed 115 percent of the baseline, the additional duty can be as high as 40 percent of the bound tariff or 40 percentage points (tier 2). And if imports reach 135 percent of the baseline, the additional duty can be as high as 50 percent of the bound tariff or 50 percentage points (tier 3). For China, where binding overhang has largely been eliminated, a duty of up to 30 percentage points is allowed for, as proposed by the G-33,when the combination of applied tariffs and the safe- guard duty exceeds the bound tariff. This quantity-based safeguard is modeled as a nonlinear complementarity problem. More speci�?cally, letting Ti be the safeguard tariff and QRi be the Hertel, Martin, and Leister 341 ratio of observed imports to the trigger level of imports for the safeguard tier i ¼ 1, 2, 3 yields the complementary slackness condition Ti ! 0 ? (1 2 QRi) ! 0, which implies that either Ti ! 0,(1 2 QRi) ¼ 0 (the safeguard is binding) or Ti ¼ 0,(1 2 QRi) ! 0 (the safeguard is nonbinding). The benchmark year for the baseline level of imports is 2001. With the quantity-based safeguard, it is assumed that when imports reach but do not exceed a trigger level, the duty is adjusted to keep imports at that trigger level.7 The full duty permitted at a given trigger level is imposed only when imports exceed the speci�?ed trigger level. Attention then focuses on whether the next higher trigger is reached and the next higher duty imposed. Table 1 reports on the power of the safeguard tariff (that is, 1 þ the ad valorem tariff rate) for both the quantity- and price-based safeguards. The quantity-based safeguard columns relate to the tier 1 and tier 2 tariffs applied to imports from all sources, while the price-based safeguard columns report the bilateral changes in the power of the safeguard tariff. This section focuses on the quantity-based safeguard. For example, the mean power of the tier 1 safe- guard tariff in China is 9.7 percent higher than its baseline value (1.0). When cost, insurance, and freight (c.i.f.) prices are held constant, a 1 percen- tage point change in the power of the safeguard tariff translates directly into a one percentage point change in the domestic price of imported wheat. In the absence of the safeguard, this tariff—and hence the power of the tariff—is unchanged. However, when the safeguard is present, all regions except Other East Asia (where the safeguard is always nonbinding) show a positive mean change in the power of the tier 1 safeguard tariff, ranging from 2.9 percent in the Middle East and North Africa to 10.7 percent in Brazil, where domestic production is extremely volatile. Only Brazil and China invoke the tier 2 safe- guard tariff; the tier 3 tariff is not used in the simulations.8 Table 2 reports the changes in the mean and standard deviations of key vari- ables in developing countries that are permitted to apply the safeguard. It is assumed that they do so when imports reach 110 percent of baseline levels and that they apply an additional tariff when imports reach 115 percent of the baseline. By frequently invoking the safeguard tariff, developing countries raise the mean tariff-inclusive price of imported wheat over the course of the stochastic simu- lations. When the quantity-based safeguard is imposed, the mean import price in China rises by 10.2 percent (see table 2). By restricting imports when domestic 7. An alternative, and potentially much more trade-restrictive, scenario involves imposing the full duty permitted whenever imports reached the trigger in the past 12 months, even if doing so results in imports falling below the trigger. 8. It is also of interest how the safeguard tariff would change if only a single region used the safeguard. Separate simulations that permitted only one region to impose the tariffs were undertaken (but are not reported here). Not surprising, this results in lower mean tariffs in the country invoking the safeguard—that is, the effect of all developing countries using the safeguard is to increase the frequency and intensity of single-region safeguard tariffs. 342 T A B L E 1 . Means and Standard Deviations for quantity-based and bilateral price-based Safeguards: percent change in power of the tariff Percentage Changes in Means Quantity-based SSM Price-based SSM Wheat Exporters Importing Regions Tier 1 duty Tier 2 duty AUS CHN JPN OEASIA STHASIA CAN USA MEX ARG BRZ RLAmer EU15 OEUR RUS MENA SSA CHN 9.7 1.2 0.44 9.96 0.07 0.37 5.11 0.03 0 0.63 4.47 2.83 3.26 0.63 7.68 13.12 1.47 0.01 OEASIA 0 0 0.38 8.15 0.01 0.07 4.07 0.02 0 0.47 5.17 1.36 3.08 0.4 7.67 13.12 0.87 0 STHASIA 4.2 0 0.38 8.99 0.01 0.04 3.87 0.02 0 0.47 4.49 0.75 3.1 0.4 7.95 12.19 0.87 0 MEX 4.3 0 0.46 9.96 0.07 0.21 5.11 0.03 0 0.63 4.81 2.83 3.26 0.7 8.13 13.12 1.15 0.01 ARG 5.7 0 0.12 0 0 0 0.03 0 0 0 3.13 0 0 0 4.5 8.95 0 0 BRZ 10.7 3.9 0.14 0.1 0 0 0.08 0 0 0 3.01 0 0 0 4.92 9.49 0 0 RLAmer 3.9 0 0.35 7.32 0 0 3.65 0 0 0.27 4.49 0.71 3.07 0.06 8.18 12.39 0.48 0 MENA 2.9 0 0.33 6.52 0 0 2.91 0 0 0.24 4.51 0.01 3.1 0.01 7.7 12.16 0.28 0 THE WORLD BANK ECONOMIC REVIEW SSA 3.7 0 0.3 5.02 0 0 2.46 0 0 0.22 4.52 0 2.75 0 7.8 13.07 0.15 0 Percentage Changes in Standard Deviation Quantity-based SSM Price-based SSM Wheat Exporters Importing Regions Tier 1 duty Tier 2 duty AUS CHN JPN OEASIA STHASIA CAN USA MEX ARG BRZ RLAmer EU15 OEUR RUS MENA SSA CHN 11.6 1.8 1.68 12.79 0.39 0.88 6.66 0.19 0 1.31 6.2 5.15 4.09 1.2 9.38 15.92 2.72 0.05 OEASIA 0 0 1.55 10.93 0.07 0.24 5.65 0.1 0 1.1 6.9 2.83 3.89 0.81 9.37 15.92 2.02 0 STHASIA 6.6 0 1.55 11.78 0.05 0.16 5.43 0.1 0 1.1 6.23 1.93 3.92 0.82 9.7 14.75 2.02 0 MEX 6.1 0 1.69 12.79 0.39 0.54 6.66 0.19 0 1.31 6.66 5.15 4.09 1.31 9.92 15.92 2.36 0.05 ARG 8 0 0.57 0 0 0 0.13 0 0 0 4.72 0 0 0 6.58 12.24 0 0 BRZ 11.8 5.5 0.61 0.28 0 0 0.33 0 0 0 4.53 0 0.02 0 7.02 12.74 0 0 RLAmer 5.9 0 1.42 10.04 0 0 5.27 0.02 0 0.76 6.23 1.86 3.89 0.24 9.98 15.01 1.41 0 MENA 4.7 0 1.36 9.18 0 0 4.48 0 0 0.7 6.24 0.05 3.91 0.05 9.4 14.72 1.03 0 SSA 6 0 1.24 7.57 0 0 3.98 0 0 0.63 6.26 0 3.63 0 9.52 15.86 0.72 0 Source: Authors’ simulations T A B L E 2 . Difference in Percentage Changes of Mean and Standard Deviation between Baseline and Quantity- and Price-Based Safeguards for Key Variables in Developing Country Wheat Markets (percentage points) Import Pricea Import Quantity Producer Price Land Rents Output Country Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based or Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean ARG 5.8 0.5 2 46.9 2 7.3 2 1.9 2 1.5 2 7.6 2 6.7 2 2.5 2 2.6 BRZ 14.3 0.7 2 23.1 2 1.3 3.5 0.1 20.4 0.6 4.6 0.3 CHN 10.2 0.5 2 44.4 2 1.6 4.7 0.1 13.9 0.3 3.4 0.0 MEX 3.5 0.3 2 5.8 2 0.5 1.7 0.1 6.2 0.5 2.3 0.2 MENA 2.1 0.8 2 4.7 2 1.6 0.7 0.2 3.4 1.2 1.2 0.5 OEASIA 2 0.8 0.7 0.2 0.0 2 0.5 0.4 2 2.8 2.4 2 0.9 0.7 RLAmer 3.0 0.6 2 6.4 2 1.2 1.0 0.2 3.9 0.8 1.3 0.3 STHASIA 3.3 0.6 2 5.5 2 1.0 1.0 0.1 2.7 0.3 0.8 0.1 SSA 3.0 0.6 27 2 1.2 0.7 0.1 7.5 1.5 3.1 0.7 Standard deviation ARG 2.6 1.4 2 48.3 51.7 2 0.2 5.6 2 4.8 10 2 1.1 5.1 BRZ 13.0 2.1 2 24.7 10.2 4.3 34.3 17.8 2 19.1 2 5.5 11.5 CHN 10.0 2 3.8 2 54.2 51.4 5.8 18.8 15.8 2 28.9 2 4.0 2 15.8 MEX 2.1 0.3 2 7.7 2 0.1 2.3 2.0 2 6.3 1.7 2 3.3 2.5 MENA 1.4 2 11.5 2 6.9 2 49.4 1.1 2 23.1 2 2.8 2 5.7 22 2 6.6 OEASIA 0.2 0.3 0.0 2 1.1 0.1 1.1 2 0.2 2 4.0 0.0 1.3 RLAmer 2.1 2 4.4 2 7.6 2 57.5 1.9 2 37.8 2 5.1 2.2 2 3.2 2 21.9 STHASIA 2.8 2 0.1 2 7.5 19.6 1.3 7.5 3.4 2 23.5 2 1.0 2 8.2 SSA 2.1 0.2 2 9.9 5.1 1.0 1.6 2 6.4 7.3 2 4.6 5.1 a. Including the duty. Source: Authors’ calculations based on data described in the text. Hertel, Martin, and Leister 343 344 THE WORLD BANK ECONOMIC REVIEW production is low and prices are high, the expected domestic price of imports rises signi�?cantly across all regions except Other East Asia. This is expected to adversely affect the urban poor in particular because they tend to spend a larger share of their income on staple foods than do wealthier households. The expected quantity of imports into China is reduced under the quantity- based safeguard. Without the safeguard, the expected value of imports in China is 41.1 percent above the baseline (see appendix table A1). This large positive mean value arises because, when domestic production is low, the demand for imports is very strong; hence there is a large percentage increase from the base level. However, when domestic production is high, gross imports cannot fall below zero. So, the expected value of imports in a stochastic environment is higher than in the baseline. When the quantity-triggered safe- guard regime is overlaid on this same stochastic production environment, the mean change in imports is 3.3 percent below the baseline import value for China (see appendix table A1). So the difference is 44.4 percentage points. Other regions with large reductions in mean imports due to the quantity-based safeguard are Argentina (46.9 percentage points) and Brazil (23.1 percentage points). All developing countries and regions except Other East Asia show lower mean import quantities under the quantity-based safeguard. Higher prices for imports translate into higher mean prices for domestic pro- ducts (although the two are imperfectly linked due to the Armington product differentiation assumption) and higher mean returns to wheat producers under the special safeguard mechanism. For example, in China, mean wheat prices rise from 3.7 percent to 8.4 percent (see appendix table A1), a difference of 4.7 percentage points (see table 2). This in turn boosts mean land rents under the quantity-based safeguard in China’s wheat sector by 13.9 percent, which then boosts expected output 3.4 percent. Land rents rise in all developing countries and regions except Argentina and Other East Asia (see above), where they fall 7.6 percent because Argentina is a net exporter of wheat and because producers are hurt by the safeguard in other countries. The largest increase in land rents between the two policy regimes is for Brazilian wheat producers (20.4 percent rise in mean land rents), but other gains are also substantial. The quantity-based safeguard also results in higher mean wheat output in these regions, with the largest deviation from the nonsafeguard mean change in Brazil (4.6 percent higher under the safeguard; see table 2). Table 3 reports changes in the mean and standard deviation for key vari- ables in developed country wheat markets. These changes are the opposite of the developing country results. Mean output prices and mean land rents and output are lower in all developed countries and regions, and mean import quantities are higher except in Australia and Canada. On average, producers in these countries are adversely affected by the protection imposed in developing countries. For Canadian wheat producers, for example, wheat land rents fall 0.7 percent rather than rising an average of 8.1 percent (see appendix table A1), a difference of 8.8 percentage points. Australian wheat producers show T A B L E 3 . Difference in Percentage Changes of Mean and Standard Deviation between Baseline and Quantity- and Price-Based Safeguards for Key Variables in Developed Country Wheat Markets (percentage points) Import Pricea Import Quantity Producer Price Land Rents Output Country or Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based Quantity-based Price-based Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean Australia 2 0.1 2 0.3 2 5.0 4.5 2 0.9 0.2 2 6.7 0.6 2 3.0 0.3 Canada 2 0.6 0.1 2 2.0 0.4 2 1.0 0.3 2 8.8 2.0 2 4.5 1.0 European 2 0.3 0.0 0.3 0.1 2 0.2 0.0 2 2.2 0.4 2 1.3 0.2 Union Japan 2 0.8 0.2 0.2 2 0.1 2 0.5 0.1 2 4.2 1.0 2 1.4 0.3 Other Europe 2 0.5 2 0.5 1.8 1.3 2 0.1 2 0.2 2 0.6 2 1.0 2 0.3 2 0.6 Russian 2 0.2 2 0.2 1.0 0.2 2 0.1 2 0.2 2 0.3 2 0.8 2 0.2 2 0.5 Federation United States 2 1.0 0.2 0.7 0.3 2 0.7 0.2 2 4.1 1.3 2 1.8 0.6 Standard deviation Australia 0.2 0.3 2 0.3 4.9 0.2 0.2 2 2.5 2 1.9 2 0.8 2 0.7 Canada 0.1 0.0 2 0.2 0.0 0.0 0.0 2 1.9 0.2 2 0.8 0.0 European 0.0 0.0 0.1 0.0 0.0 0.0 2 0.2 2 0.1 2 0.3 0.0 Union Japan 0.1 0.0 0.1 0.0 0.1 0.0 2 0.4 0.3 0.0 0.0 Other Europe 0.0 0.1 1.0 0.7 2 0.1 0.1 0.0 0.2 2 0.1 2 0.5 Russian 2 0.1 0.1 1.2 0.3 2 0.1 0.1 0.0 0.7 0.0 2 0.6 Federation United States 0.0 0.0 0.4 0.0 0.2 0.1 2 1.5 0.0 2 0.7 2 0.1 a. Including the duty. Source: Authors’ calculations based on data described in the text. Hertel, Martin, and Leister 345 346 THE WORLD BANK ECONOMIC REVIEW T A B L E 4 . Changes in Mean and Standard Deviation between Baseline and Quantity- and Price-Based Safeguards in World Wheat Markets (percentage points) Trade Quantity-based Price-based Measure Safeguard Safeguard Mean Quantity 2 4.7 2 0.5 Price 2 0.8 0.0 Standard deviation Quantity 2 2.1 0.1 Price 0.1 0.1 Source: Authors’ calculations based on data described in the text. nearly as large a change in mean land rents. Consequently, expected output in developed countries and regions is also lower. Globally, mean wheat trade quantity falls sharply, from 7.3 percent to 2.6 percent (see appendix table A5), a difference of 4.7 percentage points (table 4).9 The expected decline in global wheat prices due to the quantity-based safeguard is 0.8 percentage points. The bottom panels of tables 2-4 show the volatility of key variables in the global wheat market, as measured by the changes in standard deviations. For example, the percentage changes in the standard deviation of the power of safe- guard tariffs on wheat imports into China are 11.6 percent for the tier 1 tariff and 1.8 percent for the tier 2 tariff (see table 1, bottom panel). Volatility in the power of the safeguard tariff, translates directly into volatility in import prices (inclusive of the tariff). The percentage change in the standard deviation of the domestic price of imports in China is 14.1 percent with the safe- guard and 4.1 percent without (see appendix table A1), a difference of 10 per- centage points (see table 2, bottom panel). Import quantities are inherently volatile in many of these countries, with the change in standard deviation suggesting that all countries and regions (except Other East Asia) will regularly exceed the 110 percent tier 1 threshold under the quantity-based safeguard. Without the safeguard, the greatest import volatility is in China, where the expected change in standard deviation of import volume equals 110.3 percent of baseline imports (see appendix table A1). The safeguard substantially reduces the volatility of imports into China, cutting the change in standard deviation to roughly half the nonsafeguard value (to 56.1 percent; see appendix table A1). Argentina shows a signi�?cant drop in wheat import volatility, with a change in standard deviation that is 48.3 percentage points lower with the safeguard, as does Brazil (79.2 percent to 54.5 percent, a difference of 24.7 percentage points; 9. The impact on global wheat trade of the special safeguard mechanism in only one country or region was also considered, as has been the case in most previous studies with single-country analyses. In this case, the change in world wheat trade is very similar to the nonsafeguard case. Hertel, Martin, and Leister 347 see table 2, bottom panel). All the other countries and regions except Other East Asia cut their import volume volatility by nearly half, translating into drops in expected change in standard deviation ranging from 6.9 percentage points to 9.9 percentage points. When duties are imposed on import surges, producer prices become more volatile (see �?gure 3). In China, the standard deviation of domestic prices rises from 25 percent to 30.8 percent (see appendix table A1), a difference of 5.8 percentage points, and in Brazil it rises from 46.2 percent to 50.5 percent, a difference of 4.3 percentage points (see table 2). The impact on land rents is more complex, with volatility increasing sharply in China and Brazil, but falling in Argentina, Mexico, Middle East and North Africa, Other East Asia, Rest of Latin America, and Sub-Saharan Africa. Finally, domestic output may be more stable under the safeguard because in a bad year, when production is low and there is a strong incentive for imports to surge, this competing source of supply is frustrated by rising tariffs, thereby lending extra incentive for pro- ducers to offset the weather-induced decline in output. The bottom panel of table 3 reports changes in the standard deviation of key market variables in developed countries, which are little affected by the safeguards in developing countries. Prices are slightly more volatile in the wheat-exporting regions of Australia, Canada, and the United States as well as in Japan, and output slightly more stable with the quantity-based safeguard than with no safeguard, but the differences are small. This reﬂects the predomi- nance of developed countries in global wheat trade. Globally, the volatility of wheat trade volume is slightly lower with the quantity-based safeguard, while price volatility is slightly higher (see table 4). VII. IMPLEMENTING THE PRICE-BASED SPECIAL SAFEGUARD MECHANISM Under the price-based safeguard, countries can implement a safeguard tariff when the import price on a shipment falls below 85 percent of the baseline level (three-year moving average). Retaining the previous notation of T for the safeguard tariff and introducing PR as the ratio of observed price per shipment to the price trigger yields the complementarity problem T ! 0 ? (PR 2 1) ! 0, which implies that either T ! 0, (PR 2 1) ¼ 0 (the safeguard is binding) or T ¼ 0,(PR 2 1) ! 0 (the safeguard is nonbinding). Unlike the quantity-based system, there is only one tier in the price-based safeguard. In addition, the safeguard tariff imposed can amount to only 85 percent of the difference between the shipment price and the baseline price. There are two key differences between the quantity- and price-based safe- guards. The �?rst has to do with bilateral price differences for wheat, and the second has to do with the price-based safeguard’s focus on shipments instead of average annual imports. Both features are important to the �?ndings and thus deserve discussion at this point. Turning �?rst to the bilateral price issue, 348 THE WORLD BANK ECONOMIC REVIEW because the price of each shipment of wheat is compared to a most favored nation average price to evaluate whether the safeguard has been triggered, it is important to account for bilateral differences in commodity prices. To better understand these bilateral price differences, average unit values for wheat exports from each region in the model over 2000–04 are computed as the ratio of each region’s export unit value to the global average export unit value (table 5, �?rst column). Developing countries show a general tendency for lower prices and developed countries for higher ones, as shown by Schott (2004) for exports in general. But this is not always the case, with some high- income regions specializing in lower price varieties of wheat and some poorer countries having higher unit values. The countries and regions with below average wheat export prices are Argentina, China, Rest of Latin America, South Asia, the European Union, Rest of Europe, and the Russian Federation. Countries and regions with above average unit values are Brazil, Mexico, Sub-Saharan Africa (largely South Africa), Australia, Canada, Japan, and the United States. The remainder of table 5 uses these unit values and the bilateral trade pattern from 2001 to compute the ratio of a given bilateral exporter price to the average import price in each importing market. Some exporters show sig- ni�?cant variation in the price of their product, relative to other suppliers, across destination markets. For example, Australian bilateral relative prices range from 0.98 in China to 1.19 in Argentina. Canadian export price ratios range from 1.02 to 1.24. In some cases the ratio falls below the 0.85 trigger point speci�?ed in the special safeguard mechanism. Therefore, these values were truncated at 85 percent of the average import price for use in the empiri- cal model, since values below 0.85 are not permitted. For such exporter- importer pairs, any further decline in price will immediately trigger the special safeguard mechanism. In the case of high unit value exporters (such as Canada), export prices will have to fall by more than 15 percent to trigger the special safeguard mechanism. The second key difference between the two safeguards has to do with the application of the price trigger on a shipment-by-shipment basis. This contrasts with the year-to-year price volatility reproduced by the model. The price of grain varies considerably both within a given year and across suppliers, but much of this variability is averaged out in the annual statistics used in the model. Thus, without any adjustments, the model would not invoke the bilat- eral, shipment-based safeguards with suf�?cient frequency. To remedy this problem while retaining the same basic model structure and the capability to compare results between the quantity- and price-based safe- guards, a multiplicative factor was introduced, kr ¼ abr, that operates on the bilateral c.i.f. prices in the model to compute the appropriate price trigger, ptriggerrs ¼ kr pcifrs. Setting the parameter a ¼ 1.15 bridges the gap between annual price volatility and the monthly price variations used as a proxy for the shipment-by-shipment volatility data that were unavailable. This factor was T A B L E 5 . Relative Global Export Price Ratio and Bilateral Import Price Ratios Developing Country Exporters Bilateral Import Price Ratios (Developing Country Wheat Importers) Global Export Middle East and Other Rest of Latin South Sub-Saharan Exporter Price Ratio Argentina Brazil China Mexico North Africa East Asia America Asia Africa Argentina 0.89 0.97 0.96 0.85 0.85 0.85 0.85 0.85 0.85 0.85 Australia 1.10 1.19 1.18 0.98 0.99 1.03 1.00 1.02 1.00 1.05 Brazil 1.05 1.14 1.13 0.94 0.94 0.99 0.96 0.97 0.96 1.00 Canada 1.14 1.24 1.23 1.02 1.02 1.07 1.04 1.06 1.04 1.09 China 0.97 1.05 1.04 0.87 0.87 0.91 0.89 0.90 0.88 0.93 European Union 0.98 1.06 1.05 0.88 0.88 0.92 0.89 0.91 0.89 0.94 Japan 1.03 1.12 1.11 0.92 0.92 0.97 0.94 0.95 0.94 0.99 Mexico 1.15 1.25 1.24 1.03 1.03 1.08 1.05 1.07 1.05 1.10 Middle East and 1.00 1.09 1.08 0.89 0.90 0.94 0.91 0.93 0.91 0.96 North Africa Other East Asia 1.00 1.09 1.08 0.89 0.90 0.94 0.91 0.93 0.91 0.96 Other Europe 0.85 0.92 0.91 0.85 0.85 0.85 0.85 0.85 0.85 0.85 Rest of Latin 0.9 0.98 0.97 0.85 0.85 0.85 0.85 0.85 0.85 0.86 America South Asia 0.96 1.04 1.03 0.86 0.86 0.90 0.88 0.89 0.88 0.92 Sub-Saharan Africa 1.15 1.25 1.24 1.03 1.03 1.08 1.05 1.07 1.05 1.10 Russian Federation 0.85 0.92 0.91 0.85 0.85 0.85 0.85 0.85 0.85 0.85 United States 1.10 1.19 1.18 0.98 0.99 1.03 1.00 1.02 1.00 1.05 Note: Italicized numbers have been truncated to 0.85, the trigger for the price-based safeguard, so that they can be incorporated into the model. Source: Authors’ calculations based on data from the United Nations Statistics Division’s Commodity Trade Statistics Database and the Global Trade Analysis Project’s GTAP6 Data Base. Hertel, Martin, and Leister 349 350 THE WORLD BANK ECONOMIC REVIEW estimated using monthly price data for Canadian wheat for January 1983–June 2008 as a proxy for the prices of individual shipments.10 The second adjustment factor, br, is indexed by exporting region and brings bilateral annual prices in line with those observed over the historical period. Together these ensure that the frequency with which the bilateral price trigger is activated more accurately represents the reality of this bilateral, shipment-based measure. Table 1 also reports the changes in the mean and the standard deviation of the power of the bilateral safeguard tariff for eight developing countries and regions. The safeguard tariff now varies, not only by importer (rows) but also by the source country or region (columns). The highest mean tariffs are imposed on the low unit value exporters: Argentina, China, Other Europe, Russian Federation, and South Asia. The volatility ranking for the safeguard tariffs is similar, as shown in the bottom panel of table 1, which reports the standard deviation in the percentage change in the power of the price-based safeguard tariff on each bilateral ﬂow. The price-based safeguard columns in tables 2-4 report the changes in means and standard deviations of key variables in developing and developed country markets. The price-based safeguard has a much more uniform impact on import prices than the quantity-based safeguard does—slightly raising mean prices in nearly all regions. This is because the bilateral safeguard duty levied against an individual exporter is less likely to vary across importers. With free on board (f.o.b.) prices to all destinations changing at the same rate, the only differences in these price-based safeguard duties arise due to differential trade and transport costs as well as differences in the weights determining the average import price for each region. Whereas the quantity-based safeguard is driven largely by domestic supply shocks, the price-based safeguard is driven primarily by supply volatility in the exporting countries. Since the composite import price is a blend of products from different exporters, there is much less variation in the mean import price changes under the price-based safeguard regime. The rise in mean import prices is also smaller because the price-based safeguard is imposed on only a subset of the exporting regions, and most importers are diversi�?ed in their export sourcing of wheat. This stands in sharp contrast to the quantity-based safeguard, which applies to all import sources. With marginally higher mean (tariff-inclusive) import prices, mean import volumes are lower than in the non-safeguard case, and mean domestic prices are higher in each developing country region except Argentina, which relies heavily on exports that are now facing safeguard tariffs in other developing countries. Higher domestic prices boost land rents, which translate into slightly higher mean output in all developing country regions except Argentina (see 10. The variability of prices across shipments is largely captured by the variability across suppliers and the intertemporal variability across months included in the analysis. However, other elements, such as variation across wheat varieties, can make the variance across prices of shipments even greater than is captured in the analysis. Given this, the analysis is expected to provide a lower bound estimate of the frequency with which the price-based safeguard is invoked. Hertel, Martin, and Leister 351 table 2, top panel). The expected change in global wheat exports falls from 7.3 percent to 6.8 percent (see appendix table A5), a difference of 0.5 percentage point (see table 4), and there is no difference in the mean global export price. The bottom panels of tables 2 –4 report the changes in standard deviations associated with the percentage changes in market variables in the developing and developed country markets, as well as for global trade, as a result of the price-based safeguard. Import quantities are more volatile in �?ve of the nine developing country regions, while domestic prices are more volatile in seven developing country regions. Global wheat export price volatility rises slightly— from 4.1 percent to 4.2 percent (see appendix table A5), for a difference of 0.1 percentage point (see table 4). Once this is taken into account, this measure appears to actually increase the volatility of domestic prices in most developing countries. This result highlights the pitfalls of approaches such as that used by Valde´ s and Foster (2005) that ignore the impacts of such a measure on world prices. VIII. COMPARISON OF PRICE- AND QUANTITY-BASED SAFEGUARDS Having analyzed the quantity- and price-based safeguards separately, the dis- cussion now turns to comparing the two types of safeguards using the results in tables 2 –4. The quantity-based safeguard tends to boost tariff-laden import prices by much more than the price-based safeguard in developing countries, and the impacts vary more across importing countries. In China, the quantity-based safeguard boosts mean duty-laden import prices 10.2 percent while the price- based safeguard raises them less than 1 percent. Higher mean prices for imports in the domestic market translate into lower mean import quantities. The quantity-based safeguard also boosts domestic prices, land rents, and output by a larger amount in all but two developing country regions (Argentina and Other East Asia). These larger changes are mirrored by larger output reductions in developed countries because the price-based safeguard tends to discriminate against low unit value exporters, which tend to be devel- oping countries (as well as the European Union). While the quantity-based safeguard boosts import price variability in all developing country cases (see table 2, bottom panel), the price-based safeguard has a mixed effect on the standard deviation of tariff-laden import prices. The standard deviation of import prices is lower in four of the nine developing country regions but in none of the developed country regions. Import volatility decreases sharply with the quantity-based safeguard for all developing country regions but increases under the price-based safeguard in �?ve of them. Domestic price volatility rises in seven developing country regions because of the increased volatility of export prices resulting from the price-based safeguard. The quantity-based safeguard increases domestic price volatility for all develop- ing country regions except Argentina. 352 THE WORLD BANK ECONOMIC REVIEW Finally, with the quantity-based safeguard, the expected volume of world trade is substantially reduced, whereas the price-based safeguard appears to be less damaging to global trade levels. Both types of safeguards boost world price volatility (see table 4). IX. CONCLUSIONS The special safeguard mechanism has been a controversial feature of the recent WTO negotiations under the Doha Development Agenda. Some advocates argue that it is needed to protect low-income domestic producers from the vag- aries of world markets. However, economic principles suggest that widespread use could destabilize world prices and deny domestic consumers access to affordable imports in the case of domestic shortages. This article investigates the key components of the special safeguard mechanism proposed in the draft WTO modalities of December 2008. It includes provisions for both quantity- and price-based safeguard measures and shows that these safeguards operate in very different ways. The empirical analysis is conducted by stochastically simulating a model of the world wheat market. The �?ndings suggest that the quantity-based safeguard is an order of magnitude more damaging to world trade than the price-based safeguard is. The quantity-based safeguard reduces imports, raises domestic prices, and boosts mean domestic production in the countries and regions that use it. Rather than insulating countries that use it from price volatility, the measure could increase price volatility in developing countries by restricting imports when they would otherwise alleviate the adverse impacts of harvest shortfalls. Based on the speci�?ed triggers and duties, the quantity-based safe- guard would shrink the expected value of wheat imports by nearly 50 percent in some regions, with overall world wheat trade falling by 4.7 percent. A more restrictive scenario under which the full permitted duty is used whenever imports have reached the trigger in the past 12 months could result in even larger reductions in imports and greater volatility. The price-based safeguard is less damaging to world trade because it is applied on a bilateral basis and since countries import wheat from a variety of sources, the impact of a safeguard tariff on some suppliers is diluted. The same is true of the impacts of the price-based safeguard on prices and production. The results suggest that the price-based safeguard would actually increase the volatility of producer prices in seven of the nine developing country regions considered, with trading partners potentially applying the safeguard when the country has a good harvest and increases its exports. Part of the rationale for the special safeguard mechanism is a concern that shocks from world markets could have adverse impacts on vulnerable produ- cers and consumers in developing countries. However, by imposing the duties permitted under the safeguard, developing countries are likely to increase, rather than decrease, the volatility of prices in domestic markets. If the Hertel, Martin, and Leister 353 ﬂexibility it provides to raise protection on agricultural products is to be used, it is important to consider the actual impacts of such duties on domestic out- comes, rather than to mechanically implement the duties provided for under the special safeguard mechanism proposal. Unfortunately, developing countries that opt not to use the safeguard may still see the volatility of their producer prices increase as a result of greater world price instability induced by countries that do employ the safeguard. This is particularly troublesome if increased greenhouse gas concentrations in the atmosphere give rise to greater climate volatility and hence greater volatility in the production of staple food products (Ahmed, Diffenbaugh, and Hertel 2009). In closing, many of the main arguments in favor of the special safeguard mechanism focus on the well-being of vulnerable agricultural producers. Yet many rural residents of poor countries are net purchasers of food, and in many countries urban poverty is growing ever more signi�?cant. In this context, the potential for policies based on the safeguard rules to lessen poverty vulner- ability seems very questionable. Future work should take into account the poverty dimension of the special safeguard mechanism as well as the broad dynamics considered in this article. APPENDIX T A B L E A 1 . Percentage Change in Mean and Standard Deviation for Key Variables in Developing Country Wheat Markets 354 with Quantity-Based Safeguard Import pricea Import quantity Producer price Land rents Output No No No No No Country or Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean Argentina 2 1.7 4.1 41.3 2 5.6 4.2 2.3 10.9 3.3 3.9 1.4 Brazil 2.3 16.6 19.4 2 3.7 16.1 19.6 2.7 23.1 0.7 5.3 China 0.5 10.7 41.1 2 3.3 3.7 8.4 2 3.7 10.2 2 2.1 1.3 Mexico 0.6 4.1 1.8 2 4.0 1.0 2.7 1.5 7.7 0.7 3.0 Middle East and 2 0.2 1.9 2.5 2 2.2 0.5 1.2 2 1.0 2.4 2 0.3 0.9 North Africa Other East Asia 0.3 2 0.5 0.0 0.2 0.7 0.2 2.8 0.0 0.4 2 0.5 Rest of Latin 0.3 3.3 3.3 2 3.1 1.2 2.2 2.2 6.1 0.9 2.2 THE WORLD BANK ECONOMIC REVIEW America South Asia 0.1 3.4 3.3 2 2.2 1.3 2.3 2 0.1 2.6 2 0.3 0.5 Sub-Saharan Africa 2 0.1 2.9 3.1 2 3.9 0.8 1.5 2 2.0 5.5 2 0.7 2.4 Standard deviation Argentina 6.0 8.6 69.5 21.2 11.9 11.7 29.6 24.8 23.7 22.6 Brazil 8.7 21.7 79.2 54.5 46.2 50.5 11.2 29.0 35.2 29.7 China 4.1 14.1 110.3 56.1 25.0 30.8 11.6 27.4 12.8 8.8 Mexico 4.7 6.8 22.1 14.4 8.1 10.4 17.6 11.3 16.4 13.1 Middle East and 3.9 5.3 20.0 13.1 8.0 9.1 8.6 5.8 11.1 9.1 North Africa Other East Asia 4.6 4.8 0.9 0.9 5.5 5.6 29.8 29.6 16.9 16.9 Rest of Latin 4.2 6.3 21.6 14.0 8.3 10.2 13.8 8.7 13.4 10.2 America South Asia 4.4 7.2 20.4 12.9 12.9 14.2 6.1 9.5 8.7 7.7 Sub-Saharan Africa 4.0 6.1 25.3 15.4 9.6 10.6 15.8 9.4 16.1 11.5 a. Including the duty. Source: Authors’ calculations based on data described in the text. T A B L E A 2 . Percentage Change in Mean and Standard Deviation for Key Variables in Developed Country Wheat Markets with Quantity-Based Safeguard Import Pricea Import Quantity Producer Price Land Rents Output Country or No No No No No Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean Australia 2 7.4 2 7.5 57.7 52.7 1.0 0.1 6.1 2 0.6 2.5 2 0.5 Canada 2 1.0 2 1.6 10.7 8.7 1.0 0.0 8.1 2 0.7 3.8 2 0.7 European Union 2 0.2 2 0.5 1.7 2.0 0.6 0.4 2.0 2 0.2 1.2 2 0.1 Japan 0.4 2 0.4 0.0 0.2 0.6 0.1 4.6 0.4 0.6 2 0.8 Other Europe 2 2.0 2 2.5 28.7 30.5 3.2 3.1 0.2 2 0.4 0.0 2 0.3 Russian 2 3.1 2 3.3 42.7 43.7 7.3 7.2 0.5 0.2 2 1.5 2 1.7 Federation United States 0.5 2 0.5 4.1 4.8 1.0 0.3 3.3 2 0.8 1.5 2 0.3 Standard deviation Australia 8.0 8.2 58.2 57.9 6.4 6.6 41.0 38.5 28.5 27.7 Canada 5.1 5.2 20.8 20.6 4.8 4.8 30.2 28.3 20.7 19.9 European Union 4.6 4.6 3.8 3.9 6.0 6.0 11.5 11.3 11.3 11.0 Japan 4.5 4.6 2.0 2.1 4.5 4.6 34.6 34.2 15.6 15.6 Other Europe 4.6 4.6 68.0 69.0 19.1 19.0 5.6 5.6 16.8 16.7 Russian 15.5 15.4 69.3 70.5 31.1 31.0 14.2 14.2 17.7 17.7 Federation United States 4.4 4.4 22.2 22.6 6.1 6.3 16.0 14.5 13.9 13.2 a. Including the duty. Source: Authors’ calculations based on data described in the text. Hertel, Martin, and Leister 355 356 T A B L E A 3 . Percentage Change in Mean and Standard Deviation for Key Variables in Developing Country Wheat Markets with Price-Based Safeguard Import Pricea Import Quantity Producer Price Land Rents Output No No No No No Country or Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean Argentina 2 1.7 2 1.2 41.3 34 4.2 2.7 10.9 4.2 3.9 1.3 Brazil 2.3 3.0 19.4 18.1 16 16.1 2.7 3.3 0.7 1.0 China 0.5 1.0 41.1 39.5 3.7 3.8 2 3.7 2 3.4 2 2.1 2 2.1 Mexico 0.6 0.9 1.8 1.3 1.0 1.1 1.5 2.0 0.7 0.9 Middle East and North 2 0.2 0.6 2.5 0.9 0.5 0.7 2 1.0 0.2 2 0.3 0.2 Africa Other East Asia 0.3 1.0 0.0 0.0 0.7 1.1 2.8 5.2 0.4 1.1 THE WORLD BANK ECONOMIC REVIEW Rest of Latin America 0.3 0.9 3.3 2.1 1.2 1.4 2.2 3.0 0.9 1.2 South Asia 0.1 0.7 3.3 2.3 1.3 1.4 2 0.1 0.2 2 0.3 2 0.2 Sub-Saharan Africa 2 0.1 0.5 3.1 1.9 0.8 0.9 2 2.0 2 0.5 2 0.7 0.0 Standard deviation Argentina 4.7 6.1 22.1 73.8 8.1 13.7 17.6 27.6 16.4 21.5 Brazil 6.0 8.1 69.5 79.7 11.9 46.2 29.6 10.5 23.7 35.2 China 8.0 4.2 58.2 109.6 6.4 25.2 41.0 12.1 28.5 12.7 Mexico 4.4 4.7 22.2 22.1 6.1 8.1 16.0 17.7 13.9 16.4 Middle East and North 15.5 4.0 69.3 19.9 31.1 8.0 14.2 8.5 17.7 11.1 Africa Other East Asia 4.5 4.8 2.0 0.9 4.5 5.6 34.6 30.6 15.6 16.9 Rest of Latin America 8.7 4.3 79.2 21.7 46.2 8.4 11.2 13.4 35.2 13.3 South Asia 4.6 4.5 0.9 20.5 5.5 13.0 29.8 6.3 16.9 8.7 Sub-Saharan Africa 3.9 4.1 20.0 25.1 8.0 9.6 8.6 15.9 11.1 16.2 a. Including the duty. Source: Authors’ calculations based on data described in the text. T A B L E A 4 . Percentage Change in Mean and Standard Deviation for Key Variables in Developed Country Wheat Markets with Price-Based Safeguard Import Pricea Import Quantity Producer Price Land Rents Output Country or No No No No No Region Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Safeguard Mean Australia 2 7.4 2 7.7 57.7 62.2 1.0 1.2 6.1 6.7 2.5 2.8 Canada 2 1.0 2 0.9 10.7 11.1 1.0 1.3 8.1 10.1 3.8 4.8 European Union 2 0.2 2 0.2 1.7 1.8 0.6 0.6 2.0 2.4 1.2 1.4 Japan 0.4 0.6 0.0 2 0.1 0.6 0.7 4.6 5.6 0.6 0.9 Other Europe 2 2.0 2 2.5 28.7 30.0 3.2 3.0 0.2 2 0.8 0.0 2 0.6 Russian 2 3.1 2 3.3 42.7 42.9 7.3 7.1 0.5 2 0.3 2 1.5 2 2.0 Federation United States 0.5 0.7 4.1 4.4 1.0 1.2 3.3 4.6 1.5 2.1 Standard deviation Australia 8.0 8.3 58.2 63.1 6.4 6.6 41.0 39.1 28.5 27.8 Canada 5.1 5.1 20.8 20.8 4.8 4.8 30.2 30.4 20.7 20.7 European Union 4.6 4.6 3.8 3.8 6.0 6.0 11.5 11.4 11.3 11.3 Japan 4.5 4.5 2.0 2.0 4.5 4.5 34.6 34.9 15.6 15.6 Other Europe 4.6 4.7 68.0 68.7 19.1 19.2 5.6 5.8 16.8 16.3 Russian 15.5 15.6 69.3 69.6 31.1 31.2 14.2 14.9 17.7 17.1 Federation United States 4.4 4.4 22.2 22.2 6.1 6.2 16.0 16.0 13.9 13.8 a. Including the duty. Source: Authors’ calculations based on data described in the text. Hertel, Martin, and Leister 357 358 THE WORLD BANK ECONOMIC REVIEW T A B L E A 5 . Impacts on World Wheat Trade of Quantity- and Price-Based Safeguards Trade Measure Baseline Quantity-based Safeguard Price-based Safeguard Mean Quantity 7.3 2.6 6.8 Price 0.1 2 0.7 0.1 Standard deviation Quantity 7.8 5.7 7.9 Price 4.1 4.2 4.2 Source: Authors’ calculations based on data described in the text. REFERENCES Ahmed, A., N. Diffenbaugh, and T. Hertel. 2009. “Climate Volatility Deepens Poverty Vulnerability in Developing Countries.�? Environmental Research Letters 4(3). Arndt, C. 1996. “An Introduction to Systematic Sensitivity Analysis Via Gaussian Quadrature.�? Technical Paper 2. Global Trade Analysis Project, West Lafayette, Ind. DeVuyst, E.A., and P.V. Preckel. 1997. “Sensitivity Analysis Revisited: A Quadrature-Based Approach.�? Journal of Policy Modeling 19(2):175– 85. Finger, J.M. 2009. “A Special Safeguard Mechanism for Agricultural Imports and the Management of Reform.�? World Bank, Washington, D.C. Fraser, R., and W. Martin 2008. “Price and Quantity Triggers in WTO Agricultural Safeguards.�? World Bank, Washington, D.C. G-33 (Group of 33). 2008. “Statement of G-33, African Group, ACP and SVEs on Special Products and Special Safeguard Mechanism.�? World Trade Organization, Geneva. GATT (General Agreement on Tariffs and Trade). 1994. “The Results of the Uruguay Round of Multilateral Trade Negotiations.�? GATT Secretariat, Geneva. Grant, J., and K. Meilke. 2009. “Triggers, Remedies, and Tariff Cuts: Assessing the Impact of a Special Safeguard Mechanism for Developing Countries.�? Estey Center Journal for International Law and Trade Policy 10(1):223– 46. de Gorter, H., E. Kliauga, and A. Nassar. 2009. “How Current Proposals on the SSM in the Doha Impasse Matter for Developing Country Exporters.�? Instituto de Estudos do Comercio e ˜ o Paolo, Brazil. Negociacoes Internacionais, Sa Hallaert, J. 2005. “Special Agricultural Safeguards: Virtual Bene�?ts and Real Costs—Lessons for the Doha Round.�? Working Paper WP/05/131. International Monetary Fund, Washington, D.C. Hertel, T.W., D. Hummels, M. Ivanic, M., and R. Keeney. 2007. “How Con�?dent Can We Be of CGE-based Assessments of Free Trade Agreements.�? Economic Modelling 24(4):611– 35. Hertel, T.W., W. Martin, and A.M. Leister. 2010. “Potential Implications of a Special Safeguard Mechanism in the WTO : The Case of Wheat.�? Policy Research Working Paper 5334. World Bank, Washington, D.C. ICTSD (International Centre for Trade and Sustainable Development). 2008. “Members Give Mixed Reactions to Lamy Compromise, Take a Good Step Forward on Services.�? Bridges Daily Update 6, 26 July. Ivanic, M., and W. Martin. 2008. Implications of Higher Global Food Prices for Poverty in Low-income Countries. Agricultural Economics 39(1):405–16 Keeney, R., and T.W. Hertel. 2005. “GTAP-AGR: A Framework for Assessing the Implications of Multilateral Changes in Agricultural Policies.�? Technical Paper 24. Global Trade Analysis Project, West Lafayette, Ind. Hertel, Martin, and Leister 359 Martin, W., and J. Alston. 1994. “A Dual Approach to Evaluating Research Bene�?ts in the Presence of Trade Distortions.�? American Journal of Agricultural Economics 76(1):26–35. ———. 1997. “Producer Surplus without Apology? Evaluating Investments in R&D.�? Economic Record 73(221):146 –58. Martin, W., and A. Mattoo. 2008. “The Doha Development Agenda: What’s on the Table?�? Policy Research Working Paper 4672. World Bank, Washington, D.C. Montemayor, R. 2007. “Implications of Proposed Modalities for the Special Safeguard Mechanism A Simulation Exercise.�? Issue Paper 10. International Center for Trade and Sustainable Development, Geneva. ———. 2008. “How Will the May 2008 ‘Modalities’ Text Affect Access to the Special Safeguard Mechanism, and the Effectiveness of Additional Safeguard Duties?�? Issue Paper 15. International Center for Trade and Sustainable Development, Geneva. Morrison, J., and R. Sharma. 2005. “Modalities for an SSM: Product Eligibility and Alternative Triggers.�? Presentation to the Informal ICTSD Dialog on Special Products and the SSM, November 25, Geneva. Pearson, K., and C. Arndt. 2000. “Implementing Systematic Sensitivity Analysis Using GEMPACK.�? Technical Paper 3. Global Trade Analysis Project, West Lafayette, Ind. Schott, P. 2004. “Across-Product versus Within-Product Specialization in International Trade.�? Quarterly Journal of Economics 119(2):647 –78. Sen, A. 1981. Poverty and Famines: An Essay on Entitlement and Deprivation. Oxford, U.K.: Oxford University Press. South Centre. 2009. “The Volume-Based Special Safeguard Mechanism: Analysis of the Conditionalities in the December 2008 WTO Agriculture Chair’s Texts.�? Analytical Note SC/TDP/AN/AG/9. South Centre, Geneva. Tyers, R., and K. Anderson. 1992. Disarray in World Food Markets: A Quantitative Assessment Cambridge, U.K.: Cambridge University Press. Uri, N., and E.D. Beach. 1997. “A Note on Quality Differences and United States/Canadian Wheat Trade.�? Food Policy 22(4):359 –67. Valde´ s, A., and W. Foster. 2005. “The New SSM: A Price Floor Mechanism for Developing Countries.�? International Centre for Trade and Sustainable Development, Geneva. Valenzuela, E., T.W. Hertel, R. Keeney, and J.J. Reimer. 2007. “Assessing Global CGE Model Validity Using Agricultural Price Volatility.�? American Journal of Agricultural Economics 89(2):382–97. Wolfe, R. 2009. “The Special Safeguard Fiasco in the WTO: The Perils of Inadequate Analysis and Negotiation.�? Working Paper. Groupe d’Economie Mondiale, Paris. WTO (World Trade Organization). 2008a. “Revised Draft Modalities for Agriculture.�? TN/AG/W/4/ Rev.4 and TN/AG/W/7. Committee on Agriculture, Special Session, World Trade Organization, Geneva. ———. 2008b. “An Unof�?cial Guide to Agricultural Safeguards: GATT, Old Agricultural (SSG) and New Mechanism (SSM).�? [www.wto.org/english/tratop_e/agric_e/guide_agric_safeg_e.htm].