45346 '44 - -i Volume 21 2007 Number 1 S-v É t1 - .'- ,,i OXFORID JOURNALS OXFORD UNIVERSI1-Y PRESS THE WORLD BANK ECONOMIC REVIEW EDITOR Jaime de Melo, University of Geneva ASSISTANT TO THE EDITOR Marja Kuiper, World Bank EDITORIAL BOARD Chong-En Bai, Tsinghua University, Jan Willem Gunning, Free University, China The Netherlands Jean-Marie Baland, University ofNamur, Hanan Jacoby, World Bank Belgium Graciela Kaminsky, George Washington Kaushik Basu, Cornell University, USA University, USA Alok Bhargava, Houston University, USA Peter Lanjouw, World Bank Frangois Bourguignon, World Bank Thierry Magnac, Universit de Toulouse l, Kenneth Chomitz, World Bank France Maureen Cropper, University ofMaryland, Jonathan Morduch, New York University, USA USA Jishnu Das, World Bank Juan-Pablo Nicolini, UniversidadTorcuato Klaus Deininger, World Bank di Tella, Argentina Ash Demirgiiq-Kunt, WorldBank Boris Pleskovic, WorldBank Stefan Dercon, University of Oxford, UK Martin Rama, World Bank Ishac Diwan, World Bank Ritva Reinikka, World Bank Augustin Kwasi Fosu, UNEconomic Elisabeth Sadoulet, University of Calfornia, Commission for Africa (ECA), Ethiopia Berkeley, USA Alan Harold Gelb, World Bank Joseph Stiglitz, Columbia University, USA Paul Gertler, World Bank Jonathan Temple, University of Bristol, UK Indermit Gill, World Bank L. Alan Winters, World Bank The World Bank Economic Review is a professional journal for the dissemination of World Bank-sponsored and other research that may inform policy analysis and choice. It is directed to an international readership among economists and social scientists in government, business, international agencies, universities, and development research institutions. The Review seeks to provide the most current and best research in the field of quantitative development policy analysis, emphasizing policy relevance and operational aspects of economics, rather than primarily theoretical and methodological issues. It is intended for readers familiar with economic theory and analysis but not necessarily proficient in advanced mathematical or econometric techniques. Articles illustrate how professional research can shed light on policy choices. Consistency with World Bank policy plays no role in the selection of articles. Articles are drawn from work conducted by World Bank staff and consultants and by outside researchers. Non-Bank contributors are encouraged to submit their work. Before being accepted for publication, articles are reviewed by three referees-one from the World Bank and two from outside the instittition. Articles must also be endorsed by two members of the Editorial Board before final acceptance. For more information, please visit the Web sites of the Economic Review at Oxford University Press at www.wber.oxfordjournals.org and at the World Bank at www.worldbank.org/research/journals. Instructions for authors wishing to submit articles are available online at www.wber.oxfordjournals.org. Please direct all editorial correspondence to the Editor at wber@worldbank.org. THE WORLD BANK ECONOMIC REVIEW Volume 21 - 2007 - Number 1 In Memoriam 111 Growth and Risk: Methodology and Micro Evidence 1 Chris Elbers, Jan Willem Gunning, and Bill Kinsey Dollars, Debt, and International Financial Institutions: Dedollarizing Multilateral Lending 21 Eduardo Levy Yeyati Business Cycle Synchronization and Regional Integration: A Case Study for Central America 49 Norbert Fiess Protecting the Vulnerable: the Tradeoff between Risk Reduction and 73 Public Insurance Shantayanan Devarajan and William Jack The Incidence of Public Spending on Healthcare: 93 Comparative Evidence from Asia Owen O'Donnell, Eddy van Doorslaer, Ravi P. Rannan-Eliya, Aparnaa Somanathan, Shiva Raf Adhikari, Deni Harbianto, Charu C. Garg, Piya Hanvoravongchai, Mohammed N. Huq, Anup Karan, Gabriel M. Leung, Chiu Wan Ng, Badri Raj Pande, Keith Tin, Kanjana Tisayaticom, Laksono Trisnantoro, Yubui Zhang, and Yuxin Zhao Did the Health Card Program Ensure Access to Medical Care for 125 the Poor during Indonesia's Economic Crisis? Menno Pradhan, Fadia Saadah, and Robert Sparrow New Development Data Bases A Short Note on Updating the Grilli and Yang Commodity 151 Price Index Stephan Pfaffenzeller, Paul Newbold and Anthony Rayner Trade, Production, and Protection Database, 1976-2004 165 Alessandro Nicita and Marcelo Olarreaga In Memoriam: Enzo Grilli 1943-2006 Enzo's friends and colleagues were stunned and deeply saddened to learn of his recent sudden and premature death. All those who worked with him will remember his great enthusiasm, his tireless energy, and his keen interest in research. But above all he will be remembered for his dedication to improving economic policy. At the World Bank, Enzo held several senior positions in the Development Economics Department which he directed prior to holding positions in his country's government. During a brief leave of absence from the Bank, as general secretary of economic planning under the minister of the Treasury in Italy, Enzo laid the foundations for rigorous project evaluation in Italian public investment. Enzo left the Bank a second time to serve as Executive Director for Italy at the International Monetary Fund (following a stint in the same position at the World Bank) before retiring from the Bank in 1998 to devote full time to his great love, teaching, at his alma mater, The Johns Hopkins School of Advanced International Studies. By coincidence, the Editorial Board of the World Bank Economic Review was looking forward to publishing in this issue an update of the seminal study on the evolution of the commodity terms of trade for developing countries by Enzo and Maw Chen Yang, which the Review had published in 1988, shortly after its launch. With great sadness for his untimely death, the Editorial Board dedicates this issue to the memory of Enzo Grilli. Growth and Risk: Methodology and Micro Evidence Chris Elbers, Jan Willem Gunning, and Bill Kinsey How exposure to risk affects economic growth is a key issue in development. This article quantifies both the ex ante and ex post effects of risk using long-running panel data for rural households in Zimbabwe. It proposes a simulation-based econometric methodology to estimate the structural form of a micro model of household invest- ment decisions under risk. The key finding is that risk substantially reduces growth in this particular setting: the mean capital stock in the sample is (in expectation) 46 percent lower than in the absence of risk. About two-thirds of the impact of risk is due to the ex ante effect (that is, the behavioral response to risk), which is usually not taken into account in policy design. These results suggest that policy interventions that reduce exposure to shocks or that help households manage risk could be much more effective than is commonly thought. JEL Codes: D10, D91, C51, 012 Growth and risk are central issues in development. While the two phenomena are usually studied in isolation, it is often suggested that they are closely linked. For example, Collier and Gunning (1999a) use microeconomic evidence to show that the responses of agents to risk are an important part of the explanation for Africa's poor growth performance. Risk management involves changes in the choice of activities: households may choose low-risk activities or portfolios of activities that are well diversified. Diversification is, of course, costly: the household forgoes the gains from specialization. In itself this is a level effect that does not necessarily affect growth. However, level effects easily translate into growth rate effects-for example, when there are indivisibilities in investment and imperfections in credit markets, typical for many rural areas in Africa. All three authors are with the Department of Economics, Free University, Amsterdam. Their email addresses are celbers@feweb.vu.nl, jgunning@feweb.vu.nl (corresponding author), and bkinsey@feweb. vu.nl. For very helpful comments we are grateful to the Editor and three anonymous referees of the Review and to Hans Binswanger, Stefan Dercon, Ravi Kanbur, Peter Lanjouw, Ethan Ligon, Martin Ravallion, Elisabeth Sadoulet, T.N. Srinivasan, Erik Thorbecke, Steve Younger, and seminar participants at Amsterdam, Berkeley, Clermont-Ferrand, Cornell, Harvard, Leuven, Oxford, Rotterdam, and the World Bank. We are grateful to Trudy Owens, who provided us with the aggregate crop income data which she constructed. A supplemental appendix to this article is available at www.wber.oxfordjournals.org. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 1-20 doi:10.1093/wber/1hl008 Advance Access Publication 4 January 2007 () The Author 2007. 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 1 2 THE WORLD BANK ECONOMIC REVIEW In addition to managing risk through activity choice, households use various risk-coping strategies. In the absence of well functioning credit and insurance markets, self-insurance (saving and lack of saving in response to income shocks to smooth consumption) and informal social security institutions are used instead. Consumption smoothing typically involves changes in food stores, live- stock, or both. These assets are also subject to substantial risk (theft, vermin, spoilage, livestock illnesses), making consumption smoothing less effective as a risk-coping strategy. There is growing evidence on the cost of such risk management and coping mechanisms: they lower income, perpetuate poverty, and cause the effects of shocks to persist over long periods (Rosenzweig and Binswanger 1993; Morduch 1999; Dercon 2003, 2004; Jalan and Ravallion 2005). The signifi- cance of growth-reducing responses to risk is now widely recognized, but neither the theoretical nor the empirical literature provides much guidance for quantifying the effect of risk on growth. Much of growth theory is, of course, deterministic, so the issue cannot be addressed. Interest in growth under uncertainty has recently been revived (for example, Binder and Pesaran 1999; de Hek 1999; after early contributions such as Levhari and Srinivasan 1969), but these contributions typically examine special cases. For example, risk has no ex post effect in the Levhari- Srinivasan model and no ex ante effect in the Binder-Pesaran model.' The empirical literature also shows a growing interest in the effect of risk on growth, at both the macro and the micro levels (Ramey and Ramey 1995; Guillaumont and Chauvet 2001; Jalan and Ravallion 2001). Most of these studies use a reduced-form specification for the income-generating process. There are a few examples of structural models. For example, Rosenzweig and Wolpin (1993), who analyzed optimal accumulation under risk in Indian vil- lages, concluded that introducing actuarially fair insurance in these villages would not raise welfare because farmers are already adequately protected through informal insurance. Formal insurance therefore offers no benefits and (in their model) does not lower the cost of risk coping either.2 As a result, Rosenzweig and Wolpin find no effect on investment when risk is reduced. In this article an intertemporal optimization model is estimated for a rural household using a unique long-running panel dataset for rural households in Zimbabwe that were initially part of a government land reform and resettle- ment program. These households make little use of financial assets and infor- mal insurance, and their investment takes the form largely of building up their own livestock herd. The analysis of their behavior focuses on consumption 1. Similarly, Lucas's (1987, 2003) famous back-of-envelope calculation of the welfare effect of eliminating business cycles implicitly assumes that risk has no ex ante effect. 2. Profits are net of the implicit premium paid for informal insurance, but this premium is not known. Hence, if farmers were to switch from informal to formal insurance, the constant term in the profits function is not adjusted: farmers would implicitly continue to pay for informal insurance (Rosenzweig and Wolpin 1993). Elbers, Gunning, and Kinsey 3 smoothing as the key risk-coping strategy. In this respect the model is similar to that of Deaton (1991). However, in Deaton's model households have no incentive to save in the absence of risk, and they have access to a safe asset.' The model used here to describe the Zimbabwean farmers differs in both respects. First, it exhibits conditional convergence in the absence of risk. Households start very poor, with asset holdings far below the steady-state level. As a result, they have an incentive to accumulate capital (livestock). This process of growths provides a benchmark for addressing the growth and risk question because accumulation under risk can be compared with this risk-free counterfactual. Second, households are exposed to shocks that affect both live- stock and the income from agriculture where livestock is used as an input. Hence, these farmers have no access to a safe asset. Since the costs and benefits of consumption smoothing are explicitly modeled, the impact of actuarially fair insurance or, equivalently, the effect of risk on growth can be assessed. Whether that effect is strong or weak and posi- tive or negative is an empirical matter: the model can in principle generate widely different results. In particular, whether risk increases or reduces house- holds' propensity to accumulate assets is not implied by the specification. It depends on the nature of risk (notably the relative importance of income and asset shocks) and on how risk-averse households are. These issues are not resolved a priori but are left to the estimation phase. Using the estimated micro growth model to assess the effect of risk on growth yields a very strong nega- tive effect. Under risk the expected value of the capital stock at the end of a 50-year simulation period is 46 percent lower than it would have been in the counterfactual risk-free case. Risk not only reduces capital accumulation and hence growth; it also has a negative effect on household welfare. The results suggest that policy measures that reduce the risk exposure of these households or that offer more efficient risk coping (for example, through insurance) would have powerful effects not only on growth but also on household welfare. I. EX ANTE AND EX POST EFFECTS OF RISK A household's economic decisions (for example, on savings) are affected by risk in two ways: through the household's experience of shocks and through its 3. Dercon (1996) extends Deaton's model by making agricultural income depend on the allocation of labor between two crops, one risky, the other risk-free. In this model removing risk leads to full specialization, but the effect on growth cannot be analyzed. As in Deaton's model, there is no incentive for accumulation in the absence of risk. 4. Dercon (2005) stresses that models of consumption smoothing (e.g., Deaton 1991) often assume that agents have access to a safe asset. This overstates the effectiveness of consumption smoothing as a risk-coping strategy. 5. Since the model has a steady state, growth should be understood as transitional dynamics. 4 THE WORLD BANK ECONOMIC REVIEW perception of the distribution of the shocks it faces. It is useful to formalize this distinction. Suppose household investment decisions can be summarized as: (1) kt+ = (p(kt, st; a) where k denotes the household's capital stock, s a shock (with expected value Es = 1), and o a parameter characterizing the distribution from which s is drawn. At this stage the definition of o- is irrelevant, except that an increase in a can be interpreted as an increase in risk and that o = 0 denotes the risk-free case. In general, risk affects k,+ (and thereby growth in the sense of transi- tional dynamics) not just through s but also through o. Changes in o- will in general change the household's behavior: the household will choose a different value of ktj for the same values of k, and the current shock s,. This is the ex ante effect of risk. An example is the effect on investment behavior of the possibility of civil war breaking out. The household has not yet been exposed to a shock, but it knows that peace is precarious, so its assessment of the likeli- hood of violent conflict will have a powerful effect on its investment decisions. Hence, the ex ante effect results from the household's view of the world: two households that differ in their perception of the risks they face but that are identical in all other respects will in general make different investment decisions. By contrast, the ex post effect measures the impact of the shocks themselves. The effect of risk can be decomposed into ex ante and ex post components. Applying equation (1) repeatedly leads to (2) kt+T = g(kt; st, st+1,...,stT-1; a) for some suitably defined function go. Taking expectations EtktT = Etg(kt; st, st+1,...,St+T-1; a) = g(kt; 1,1,...,1; o-) - [g(kt;1,1,...,1; o-) - Etg(kt;st,sti,...,st+T-1; o7)] = g(kt; 1,1,..., 1; o-) - [ex post effect] = g(kt; 1,1.. 1; 0) - [g(kt; 1,1,..1; 0) - g(kt; 1,1,..1; o-)] - [ex post effect] = g(kt; 1, 1, ... 1; 0) - [ex ante effect] - [ex post effect] hence (3) g(kt; 1, 1, ..., 1; 0) - Etkt±T = [ex ante effect] + [ex post effect]. Here g(kt; 1, 1, ..., 1; 0) is the value of kt+T that the household would attain in a risk-free world (with o= 0); g(kt;1, 1, ..., 1; oa) is the hypothetical value Elbers, Gunning, and Kinsey 5 reached if the household expects shocks drawn from a distribution with posi- tive o- but in fact experiences no shocks (s = 1 in all periods). Note from equation (3) that the two effects are defined in such a way that a positive value implies that growth is reduced. In section V equation (3) is applied in the analysis of risk in Zimbabwe. That shocks can make the path of k volatile is obvious, but that risk affects the expectation Ek is not. Indeed, there is no presumption in theory about the sign (let alone the size) of the ex ante and ex post effects. The effect of risk on growth is therefore an empirical issue. Usually in empirical research equation (1) is estimated by regressing k,+T (or some other proxy for growth) on various controls (country characteristics in macro growth regressions, household characteristics in micro studies), measures of s, and, possibly, measures of o-. Two cases can arise, depending on data availability. If o- does not vary in the sample (for example, because all house- holds face the same rainfall risk, o- = u), without further identifying restrictions the effect of changes in o, (the ex ante effect) can obviously not be identified, and only kt+1 = 1.1 or s, < 0.9) in any year is 71 percent. The rate of technical progress cannot be precisely estimated. An earlier estimate using a different methodology (Gunning and others 2000) was higher but within the 95 percent confidence interval of the point estimate here (slightly below 1 percent a year). The estimated value of 7T is very significant but remarkably low.25 In this model every household is essentially a single-agent economy. In par- ticular, households do not pool idiosyncratic risk. This is clearly a very strong assumption that might seem to cause the results to overstate the effect of growth on risk. This is not the case. Consider a proportional risk-pooling arrangement that provides partial insurance, presenting an agent with a shock es(0 < 6 < 1) when a shock s occurs. By ignoring the existence of such an arrangement, the estimation procedure here understates the extent of risk. It will in fact produce an estimate of o6 rather than of o: the partial insurance 25. This is a common empirical finding; see, for example, Rosenzweig and Binswanger (1993). 14 THE WORLD BANK ECONOMIC REVIEW Fi;URE 1. Growth and Risk: Capital Accumulation for a Selected Household 4 - No risk ----- No ex-post risk 3.5 - - - - Average under risk - - - Sample path under risk 0 .5 C 2 4-1 -, -I -- - - - - - In) 0.5 0 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 Years Note: The "No risk" path corresponds to g(ko,t,...,1; 0), the "No ex post risk" path corresponds to g(ko, 1, ..., 1; a), and the "Average under risk" path corresponds to Eok,. Source: Authors' analysis based on Zimbabwe survey data. institution is observationally equivalent to a reduction in risk. In this case the estimate of the effect of risk on growth would be unaffected.26 V. RESULTS The estimated model is used to determine the effect of risk on growth. Four paths of the real value of the capital stock (scaled by the household's labor force) are followed over a 50-year period, starting at t = 0 from the house- hold's actual starting position (figure 1).27 26. Going slightly further, partial insurance can be tested for by replacing shock s with insurance-modified shock A + Bs, where -A can be interpreted as the insurance premium and B as the degree of risk mitigation. Note that with log-normally distributed shocks s and A, B # 0, A + Bs is not log-normally distributed, so the model would be misspecified for this type of risk pooling. If risk sharing is important, a significantly improved fit could be obtained by allowing A and B to vary. The signs of A and B-I would depend on whether the sample households' average position is long or short. The unrestricted estimation results give a value of A that is not significantly different from zero and a value of B that is not significantly different from 1. Also, the likelihood shows very little improvement as a result of dropping the restrictions A = 0, B = 1. In this sample the weak test of "no risk pooling" is passed. This may reflect the particular nature of the sample: the resettlement farmers came from different parts of the country and therefore had no previous ties. 27. This household has values of total factor productivity and initial capital close to the median in the sample. Calculating means over households yields similar results. The capital stock is scaled by dividing it by the labor force times the factor (1 + r)' so that (in these units) the capital stock converges to a constant level in the deterministic case. Elbers, Gunning, and Kinsey 15 The sample path is one possible growth path, defined by a particular series of 50 randomly drawn shocks, one for each year. The point to note is how volatile this path is: the capital stock (livestock) frequently changes by as much as 50 percent in one or two years.28 Clearly, time-series data for part of this growth path would make it very difficult to say something about the underlying growth process. The model was used to generate 100,000 growth paths. For each year the expected value of the household's capital stock was then calculated as the mean over these paths. The time path of this mean is shown in figure 1 as EOkT (for T = 1, .. . , 50). The averaging procedure, of course, removes the volatility. The path shows how much livestock the household would expect to obtain at future dates, from the standpoint of t = 0. Since the household starts out very poor (with ko = 0.56), it initially grows very rapidly (in expectation), at some 9 percent a year in the first 10 years. Now consider the effect of risk on growth. The distribution of shocks is initially kept unchanged, but instead of drawing shocks s from this distribution the household is presented with s = Es = 1 in each period. Hence, the house- hold faces the same risk as before but, as it happens, never experiences a shock. This is the path of g(ko;1,...1;o), using the notation of section II. The vertical difference between this curve and path EOkT is the ex post effect. Next, the path of k is calculated in the absence of risk by taking the (co)variances of the shocks to 0 while, as before, presenting the household with s = Es = 1 in each period. Clearly, this implies that the household solves a nonstochastic optimization problem: it knows that it faces no risk and indeed experiences no shocks at any point in time. This yields the path of g(ko;1, ... 1; 0). By con- struction the vertical distance between this curve and the path g(ko; 1, ... 1; o) measures the ex ante effect. The effect of risk is massive: the household would have accumulated much more capital in the absence of risk. For this household the total effect of risk is dominated by the ex ante effect (figure 1). These results also apply to the sample households as a group. In the sample risk reduces the expected long-run value of the capital stock, Eok5o, 46 percent below the steady-state value, k*, in the deterministic risk-free case: Eok5o = 0.54k*. This is a striking result. Risk not only makes growth very volatile (illustrated vividly by the sample path), it also greatly lowers growth on average. Two-thirds of this reduction is accounted for by the ex ante effect, the rest by the ex post effect-also a remarkable result. Much of the empirical literature (e.g., Ravallion 1988; Dercon and Krishnan 2000; Dercon 2004) implicitly assumes that the actual shocks are an adequate measure of the effect of risk. But they are not in the case of Zimbabwean 28. These simulation results are confirmed by the data, which show large shocks in the k time-series. A simple regression of In k on its lagged value gives a residual standard error of 0.3. This implies that changes of 50 percent are indeed quite common. 16 THE WORLD BANK ECONOMIC REVIEW households: much of the expected impact is internalized as different invest- ment decisions. Chronic poverty is often diagnosed as the result of poor endowments, as opposed to transient poverty, which is seen as the result of risk. The calculations here show that risk has a very substantial effect on mean consumption as well. In that sense risk is a structural determinant of chronic poverty. The risk-free case can be interpreted as what would happen if actuarially fair insurance were introduced. The sum of the ex ante and ex post effects then measures the effect of such insurance on capital accumulation. Figure 1 would look very similar if welfare had been plotted; in particular, the ranking of the three cases is the same: risk causes a substantial welfare loss, and much of this loss is reflected in the ex ante effect. The implication is that policies designed to reduce the exposure of households to risk or to help households to cope with risk improve welfare. In particular, these households would gain substan- tially from actuarally fair insurance. VI. CONCLUSION Empirical work using micro datasets for rural households has uncovered much evidence of the impact of risk on income levels, investment, and portfolio decisions (for example, crop diversification). While the effect of risk on growth is recognized as a key issue in development, micro studies have seldom quanti- fied it. This quantification is the central objective here. This article makes three contributions. First, it proposes a framework for analyzing the effect of risk on growth, distinguishing between the ex ante and ex post effects of shocks. Second, it estimates a stochastic growth model in its structural form using simulation methods. If all households face the same risks (as assumed here), the effect of risk on growth cannot be identified from a reduced-form regression. Moreover, using simulation methods to estimate growth models eliminates the need for the simplifications usually adopted in applied work to make the estimation problem tractable (for example, lineariza- tion around the steady state). Third, turning from the methodology to the micro evidence, the application here shows that for a sample of rural households in Zimbabwe (observed for almost a generation) risk has a very substantial effect on capital accumulation (and hence on poverty). The average (across households) expected long-run capital stock is estimated to be 46 percent lower than in the absence of risk. This confirms the suggestion in the literature that self-insurance and other microeconomic responses to risk may substantially reduce growth. The magnitude of the impact of risk on economic growth in Zimbabwe suggests that policymakers may need to reconsider the balance between interven- tions that address structural determinants of poverty (for example, raising pro- ductivity through education or improvements in farm practices) and interventions that reduce exposure to shocks or help households manage risk. The results here Elbers, Gunning, and Kinsey 17 suggest that the welfare costs of risk can be very high. The potential benefits of policy interventions to reduce exposure to risk or promote insurance or credit may therefore be much greater than previously envisaged.29 Such policies are usually seen as reducing the volatility of household income around a given mean. That risk can massively reduce the mean implies that this perspective (common in the literature on household vulnerability) can be misleading: much of what is classified as structural poverty may in fact reflect households' exposure to risk. The design of the land reform program makes the Zimbabwe case special, for example, by severely limiting the scope for diversification and reducing the incentives for investment in education. Extending this work to other countries is therefore an important area for research: to what extent the Zimbabwe results can be generalized remains an unanswered question. Also, the model here involves some stark simplifications. These are intended to be relaxed in future work, notably by increasing the number of assets and allowing for infor- mal risk pooling. APPENDIX. SOLVING THE STOCHASTIC RAMSEY MODEL Consider the case where there is no technical progress, so that j is constant. Define wealth on hand, w, as w = slif(k) + S2(1 - 8)k and shocks as s = (sl, S2). If a solution exists, the model can be written in recursive form as the stationary Bellman equation: (A-1) V(w(k, s)) = max u(w(k, s) - k) + 6EV(w(k,s)) k with associated policy function (A-2) p(w(k, s)) = arg max u(w(k, s) - k) + EV(w(k, )) k where k denotes the capital stock at the beginning of the current period and k at the end, and s denotes current shocks and § future ones. Equation (A-1) applies to every period, so time subscripts can be suppressed. Note that the policy function, p, maps the current (k, s) into k, next period's k. Hence, ( can be seen as an investment function, giving k,+± as a function of wealth on hand, wt (itself a function of the capital stock, k, and the current shocks, st). A finite value function V, that satisfies the Bellman equation (A-1) for all (k, s) is a solution to the original maximization problem equation (4). V and (P satisfy 29. A third possibility is to introduce a fixed return (safe) asset, which would increase total accumulation (in the two assets taken together) while reducing accumulation of the risky asset because the return on the safe asset establishes a floor under the expected marginal return of the risky asset. This is similar to a deterministic model where the existence of the fixed return asset induces a switch from capital accumulation, subject to decreasing marginal productivity, to the fixed return asset. 18 THE WORLD BANK ECONOMIC REVIEW the first-order condition: (A-3) u'(w(k,s) - o(w(k,s))) = 6Ev(w(k,)) (k,) and the envelope condition (A-4) V'(w) = u'(w - (w)). Condition (A-3) equates the current marginal utility of consumption to the expected discounted value of a future extra unit of wealth on hand. Condition (A-4) states that the marginal value of wealth on hand, w, equals the marginal utility of the corresponding consumption, w - p(w). It is typically not possible to solve the two conditions analytically. An approximation is used that restricts and rounds variables to a fine but finite grid of (w, k, s) values. The key to the solution of the resulting discrete system is the observation that the program value, V(.), and the policy function, ((.), are func- tions of a single variable, w. With only finite sets of values for (k, s) and i1v rounded to the nearest grid value for wealth on hand, it is easy to calculate the probabilities pij = Prob[w(ki, s) = wi] so that the equation to solve becomes (A-5) V(wf) = maxu(w - ki) + 8 pijV(wj), for all . This equation can be solved by iteration, with arbitrary initial values for V(wf), = 1, 2, . Since 3 < 1, the iteration converges.30 Given the solution, V(wf), it is straightforward to derive the corresponding policy function, p(wf). The extension to the case with technical progress (T > 0) is also straightforward. This policy function is a proximate solution to the stochastic Ramsey problem. The approximation involves discretization and imposition of a finite end time. Obviously, this may affect the solution. The authors are currently experimenting with alternative solution methods that yield a policy function that is continuous in wealth on hand, unlike the method described here. REFERENCES Alderman, Harold, John Hoddinott, and Bill Kinsey. 2004. "Long Term Consequences of Early Childhood Malnutrition." Discussion Paper 168. International Food Policy Research Institute, Washington, D.C. Binder, Michael, and M. Hashem Pesaran. 1999. "Stochastic Growth Models and Their Econometric Implications." Journal of Economic Growth 4(2):139-83. Bliss, Christopher. 1999. "Galton's Fallacy and Economic Convergence." Oxford Economic Papers 51(1):4-14. 30. See Stokey and Lucas (1989). Elbers, Gunning, and Kinsey 19 Carroll, Christopher. 2001. "Death to the Log-Linearized Consumer Euler Equation! (And Very Poor Health to the Second-Order Approximation)." Advances in Macroeconomics 1(1). Collier, Paul, and Jan Willem Gunning. 1999a. "Explaining African Economic Performance." Journal of Economic Literature 37(1):64-111. -. 1999b. Trade Shocks in Developing Countries. Oxford, UK: Oxford University Press. Deaton, Angus. 1991. "Saving and Liquidity Constraints." Econometrica 59(5):1221-48. -. 1992. Understanding Consumption. Oxford, UK: Clarendon Press. de Hek, Paul. 1999. "On Endogenous Growth under Uncertainty." International Economic Review 40(3):727-44. Dehn, Jan. 2000. "Commodity Price Uncertainty and Shocks: Implications for Investment and Growth." D.Phil. thesis, University of Oxford, UK. Dercon, Stefan. 1996. "Risk, Crop Choice and Savings: Evidence from Tanzania." Economic Development and Cultural Change 44(3):485-514. -. 2003. "The Microeconomics of Poverty and Inequality." In Proceedings of the AFD-EUDN Conference 2003. Paris: Agence francaise de d&veloppement. -. 2004. "Growth and Shocks: Evidence from Rural Ethiopia." Journal of Development Economics 74(2):309-29. -. 2005. "Risk, Insurance and Poverty: a Review." InStefan Dercon ed., Insurance against Poverty. Oxford, UK: Oxford University Press. Dercon, Stefan, and Pramila Krishnan. 2000. "Vulnerability, Seasonality, and Poverty in Ethiopia." Journal of Development Studies 36(6):25-53. Elbers, Chris, and Jan Willem Gunning. 2003. "Estimating the Stochastic Ramsey Model from Household Panel Data." Department of Economics, Free University, Amsterdam. Epstein, Larry G., and Stanley E. Zin. 1991. "Substitution, Risk Aversion, and the Temporal Behavior of Consumption and Asset Returns: an Empirical Analysis." Journal of Political Economy 99(2):263-86. Gouri6roux, Christian, and Alain Monfort. 1996. Simulation-Based Econometric Methods. New York: Oxford University Press. Guillaumont, Patrick, and Lisa Chauvet. 2001. "Aid and Performance: A Reassessment." Journal of Development Studies 37(6):66-92. Gunning, Jan Willem, John Hoddinott, Bill Kinsey, and Trudy Owens. 2000. "Revisiting Forever Gained: Income Dynamics in the Resettlement Areas of Zimbabwe." Journal of Development Studies 36(6):131-54. Hoogeveen, Hans. 2001. "Risk and Insurance in Rural Zimbabwe." Ph.D. thesis. Free University, Amsterdam. Jalan, Jyotsna, and Martin Ravallion. 2001. "Behavioral Responses to Risk in Rural China." Journal of Development Economics 66(1):23-49. -. 2005. "Household Income Dynamics in Rural China." In Stefan Dercon, ed., Insurance against Poverty. Oxford, UK: Oxford University Press. Levhari, David, and T.N. Srinivasan. 1969. "Optimal Savings under Uncertainty." Review of Economic Studies 36(2):153-63. Lucas, Robert E. Jr. 1987. Models of the Business Cycle. New York: Basil Blackwell. -. 2003. "Macroeconomic Priorities." American Economic Review 93(1):1-14. Morduch, Jonathan. 1999. "Between the State and the Market: Can Informal Insurance Patch the Safety Net?" World Bank Research Observer 14(2):187-207. Murphy, Kevin M., and Robert H. Topel. 1985. "Estimation and Inference in Two-Step Econometric Models." Journal of Business and Economic Statistics 3(4):370-79. Newbery, David M.G., and Joseph E. Stiglitz. 1981. The Theory of Commodity Price Stabilization.. Oxford, UK: Clarendon Press. 20 THE WORLD BANK ECONOMIC REVIEW Ramey, Garey, and Valerie A. Ramey. 1995. "Cross-Country Evidence on the Link Between Volatility and Growth." American Economic Review 85(5):1138-51. Ravallion, Martin. 1988. "Expected Poverty under Risk-Induced Welfare Variability." Economic Journal 98(393):1171-82. Rosenzweig, Mark R., and Hans Binswanger. 1993. "Wealth, Weather Risk and the Composition and Profitability of Agricultural Investments." Economic Journal 103(416):56-78. Rosenzweig, Mark R., and Kenneth Wolpin. 1993. "Credit Market Constraints, Consumption Smoothing, and the Accumulation of Durable Production Assets in Low-Income Countries: Investment in Bullocks in India." Journal of Political Economy 101(2):223-44. Stokey, Nancy L., and Robert E. Lucas, Jr. 1989. Recursive Methods in Economic Dynamics. Cambridge, Mass.: Harvard University Press. Dollars, Debt, and International Financial Institutions: Dedollarizing Multilateral Lending Eduardo Levy Yeyati Financial dollarization is increasingly seen as a concern because of its tendency to con- tribute to financial crises and output volatility. As a result the debate on financial dol- larization has shifted in favor of a more proactive stance on dedollarization. While often neglected, lending from international financial institutions is an important source of financial dollarization in emerging economies and must be considered in any dedollarization strategy. This article revisits old and new arguments in favor of inter- national financial institution lending in the local currency and argues that any such initiative should rely, at least initially, on demand from residents seeking stable returns in units of the local consumption basket but who are reluctant to take on sovereign risk. Superior enforcement capacity enables international financial insti- tutions to intermediate these savings, currently invested in dollarized foreign assets, back into the local economy. The international financial institutions can offer invest- ment-grade local currency bonds and use the proceeds to dedollarize their own lending to noninvestment-grade countries, thereby reducing financial dollarization and fostering the development of local currency markets. JEL codes: F34, F41, H63, G11. Financial dollarization, defined as the holding by residents of foreign currency-denominated financial assets and liabilities, has moved to the fore- front of the policy debate in many emerging market economies due to concerns about the impact of the associated currency mismatch on output volatility and financial fragility.! As a result, the central issue of the financial dollarization debate has moved from a generally passive stance aimed at minimizing the Eduardo Levy Yeyati is the financial sector advisor for the Latin America and the Caribbean Region at the World Bank, and professor of economics (currently on sabbatical leave) at the Business School of Universidad Torcuato Di Tella, Buenos Aires, Argentina, where he also directs the Center for Financial Research; his email address is ely@utdt.edu. The author thanks Jaime de Melo, Esteban Molfino, three anonymous referees, participants at the International Monetary Fund/Bank of Spain Conference on "Dollars, Debts, and Deficits-60 Years after Bretton Woods" (Madrid, June 14-15, 2004), and, especially, Eduardo Fernandez Arias for many fruitful discussions and suggestions. The author also thanks Ramiro Blazquez and Daniel Chodos for their excellent research assistance. A supplemental appendix to this article is available at http://wber.oxfordjournals.org/. 1. Many of these concerns have been validated by recent empirical work. See, for example, Berganza and Garcia Herrero (2004); Calvo, Izquierdo, and Mejia (2004); De Nicol6, Honohan, and Ize (forthcoming); Frankel (forthcoming) and Levy Yeyati (2006). THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 21-47 doi:10.1093/wber/lhlOll Advance Access Publication 24 January 2007 () The Author 2007. 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 21 22 THE WORLD BANK ECONOMIC REVIEW negative implications to a more proactive one aimed at offsetting the incentives that favor dollarization while developing local currency substitutes.2 One of the most important sources of financial dollarization in emerging market economies and developing economies is lending by international financial insti- tutions, denominated in hard currency (or a basket of currencies). The convenience of dedollarizing international financial institution lending has been highlighted in recent proposals (see, for example, Eichengreen and Hausmann 2004), but for various reasons these ideas have been received with skepticism or indifference. This article discusses old and new theoretical and empirical arguments in favor of international financial institution lending in local currencies. In con- trast to proposals that stress potential demand from nonresidents seeking a currency-diversified portfolio, this article argues that international financial institution lending should rely, at least initially, on demand from residents in search of local currency assets to minimize the volatility of returns measured in the local consumption basket but reluctant to take on sovereign risk. Saving in the local currency faces two fundamental obstacles in developing countries: high nominal volatility (unpredictable inflation due to nominal shocks or attempts to dilute the real value of local currency liabilities through inflation) and high credit risk (a high probability of default, including through violation of the terms and conditions of both public and private contracts under local jurisdiction or the imposition of confiscatory taxes on savings). The first obstacle can be mitigated through indexation, typically to the local consumer price index (CPI). The second obstacle is more difficult to overcome. Country risk encompasses not only the possibility of default but also a number of sovereign actions that erode creditor rights, including alterations of the indexation methodology. This risk induces residents to relocate their savings to countries where property rights are better protected. Thus, even in the absence of nominal instability (or despite the mitigating presence of indexation), resi- dents may end up dollarizing their savings simply because of the lack of local currency assets free of country-specific credit risk. Under these conditions there is potential demand for investment-grade securities in local currency that cannot be satisfied by noninvestment-grade countries. It is easy to see how moving savings abroad can deepen financial dollariza- tion. As a large part of the domestic pool of savings moves abroad, the country is forced to rely on foreign, particularly multilateral, credit.3 International financial institutions can make up at least partially for the lack of domestic 2. In this context dedollarization is understood as a voluntary process, as opposed to a compulsory currency conversion. Following the standard dollarization literature, dollar and foreign currency, and peso and local currency are used here interchangeably. 3. De la Torre and Schmukler (2004) discuss offshoring as a mechanism for coping with country-specific risk. If offshoring could successfully protect against country risk, foreign borrowing could readily substitute for the decline in domestic funds. However, the extent to which offshore claims are less exposed than onshore assets to country-specific risk is not obvious, as witnessed in the recent Argentine default. Levy Yeyati 23 funds through their ability to enforce an implicit preferred creditor status that reduces their exposure to sovereign risk and enables them to collect funds and issue loans at close to risk-free rates to countries facing high country-risk premiums. Thus, international financial institutions are able to mitigate the agency problem underlying the high cost of capital in emerging market econo- mies, playing a risk-transformation role. This makes international financial institutions natural candidates to launch investment-grade local currency markets. By issuing debt in emerging market currencies to fund local currency loans, they could dedollarize an important portion of the country's external liabilities while keeping a balanced currency position, in the process providing the needed liquidity to start up these missing markets. Note that this strategy requires dedollarization of outstanding loans rather than additional international financial institution lending (which could ultimately weaken the capacity of international financial institutions to enforce their preferred creditor status).4 The main deterrent has been the untested conjecture that the represent- ative international investor would not be attracted by local currency assets. However, the minimum liquidity needed to launch markets for investment- grade local currency securities can be obtained from latent demand for these securities from the country's residents. The financial dollarization literature distingushes, both analytically and empirically, between resident and nonresident potential demand for local cur- rency assets. Analytically, local currency instruments will look relatively more attractive to risk-averse local savers and borrowers, because they mirror their stream of future consumption and income more closely.s This "home currency bias" is consistent with the evidence that shows that past debt dedollarization processes were driven largely by a deepening of the domestic markets based on local demand.6 4. IIt follows that the credit risk exposure of the international financial institution should remain constant or may even decline to the extent that the default probability depends on the currency mismatch. 5. This distinction was originally made by Thomas (1985) in a two-country model and more recently by Ize and Levy Yeyati (2003). 6. Bordo, Meissner, and Redish (2002, p. 18), when analyzing the evolution of debt denomination in four British Dominions (Australia, Canada, New Zealand, and South Africa), highlight that "the onset of World War I essentially closed the London capital market, and the response was similar in all four dominions. The gold convertibility of the domestic currency was suspended (and not resumed until 1925), and governments raised funds domestically, essentially creating a domestic bond market.... Foreign issues relative to domestic issues (at least for sovereign debt) would never regain their earlier dominance." Similarly, Claessens, Klingebiel, and Shmukler (2003) find that the dollarization ratio of (domestic plus external) government bonds is negatively related to the size of domestic financial markets. See also Martinez and Werner (2002) for the development of local currency markets in Mexico, Herrera and Valdez (forthcoming) for Chile, and Caballero, Cowan, and Kearns (2005) for Australia. Note that this evidence does not imply that there is necessarily any pent-up demand for peso assets; rather, it indicates that, if that demand exists at all, it will likely come from resident investors. 24 THE WORLD BANK ECONOMIC REVIEW The case of pension funds is illuminating. By acquiring a credit risk-free asset denominated in CPI units, fund managers would fulfill their role by ensuring a stable stream of retirement benefits while avoiding the risk of confis- cation. However, to diversify credit risk, pension funds typically invest a fraction of their portfolio in dollarized foreign assets. As a result, while the government of the emerging market economy borrows in dollars from international financial institutions, a share of residents' retirement savings is invested abroad in dollarized investment-grade paper (such as that issued by international financial institutions to fund their own lending). The international financial institutions may then intermediate these funds by selling to pension funds the bonds that finance the country loans. This intermediation could be done in CPI units to the benefit of both parties involved. The article makes the case for a similar type of arrangement as a natural first step to dedollarizing the external debt of developing countries. I. DEFINITIONS, IMPLICATIONS, AND MEASUREMENT There is still disagreement on the proper definition and measurement of financial dollarization (Eichengreen, Hausmann, and Panizza 2003; Goldstein and Turner 2003). While the literature has emphasized the country's foreign currency pos- ition relative to nonresidents (typically measured by its foreign currency- denominated external debt),7 the aggregation argument underlying this distinc- tion (that the currency exposure of resident creditors and debtors should cancel out) ignores important aggregate effects. Even if a financially dollarized economy is currency balanced as a whole, it will likely be imbalanced at a micro level, leading to capital flight, bank runs, and massive bankruptcies at the time of a real exchange rate adjustment, with important real consequences. Hence, the significant effects of domestic financial dollarization found in the literature. With that in mind, this article looks at both domestic and external sources of financial dollarization. Domestic dollarization is captured by domestic dollar deposits, which, given the standard prudential limits on banks' net currency position, provide a good proxy for the volume of domestic dollar loans. External dollarization is represented by private external loans and hold- ings of external bonded debt and by multilateral lending, broken down by International Monetary Fund (IMF) and non-IMF loans.8 Finally, liability dollar- ization is computed as the ratio of dollar liabilities to total liabilities, where domestic dollar loans are proxied by domestic dollar deposits, and domestic 7. This focus on external debt implicitly presumes a link between bondholders' residence and debt jurisdiction that is imperfect at best. 8. This distinction is important for two reasons. First, one key ingredient of the proposal discussed here is that to avoid a currency mismatch while dedollarizing lending, an international financial institution needs to fund itself in the same exotic currency in which it intends to lend. This option is unavailable to the IMF, which unlike multilateral banks, does not fund itself in the capital market. Second, in contrast to the long-term financing of investment projects by multilateral banks, the IMF's short-term liquidity assistance can hardly be seen as a partial substitute for domestic loanable funds. Levy Yeyati 25 loans are proxied by domestic deposits. These values are presented for a balanced sample of countries in table 1. Official financial dollarization (ratio of non-IMF international financial institution lending to GDP) levels are compar- able to those of external loans and higher than those associated with domestic dollarization and external bonded debt (the focus of much of the empirical financial dollarization literature). Based on median values, official financial dollarization accounts for more than a fourth of total external dollarization and a fifth of total dollarization. These numbers indicate that a strategy aimed at reducing financial dollarization cannot ignore the role of international financial institutions. II. OFFSHORING AND FINANCIAL DOLLARIZATION One aspect often overlooked in the financial dollarization literature is the inter- action between country-risk and the degree of dollarization for noninvestment- grade economies. If country-risk drives domestic savings abroad where no local currency assets are available, other things being equal, higher country risk would be associated with higher total dollarization ratios (inclusive of offshore deposits), leading to a smaller volume of local currency loanable funds. In turn, this deficit would be partially compensated for by a greater dependence on foreign dollar borrowing (to the extent that it insulates investors from country risk better than domestic assets do) and, ultimately, on international financial institution lending. This section presents a stylized analytical illus- tration of this intuition and tests its empirical implications in the data. An Analytical Example The links between country risk, offshoring, and dollarization can be illustrated by a simple extension of Ize and Levy Yeyati's (2003) portfolio model. Consider the following scenario. A continuum of measure A of risk-averse resi- dent investors endowed with a unit of cash can invest in four alternative assets: peso and dollar debt issued in a noninvestment-grade emerging market economy (the home economy) and peso and dollar debt issued in an investment-grade developed economy (the foreign economy). Denote the portfolio shares by xHP x , xFP, and xFD where the first superscript denotes the place of issuance (home or foreign country) and the second the denomination (pesos or dollars). Real returns as measured by the resident investor are given by: rHP = E(rP) - - rHD = E(rHD) s c _ rFD E(rFD) s1 rFP = E(rFp) - /, O TABLE 1. Sources of Financial Dollarization for a Balanced Sample of Nonindustrial Economical (Percent of GDP) International Domestic Dollar-bonded financial dollar External external debt institution lending, IMF lending Total external deposits (a) loans (b) (c) excluding IMF (d) (e) (b) + (c) + (d) + (e) Total 0 O 1996 Mean 0.0825 0.1427 0.0406 0.2238 0.0130 0.4201 0.5027 0 Median 0.0625 0.1297 0.0272 0.0872 0.0055 0.3323 0.4326 Minimum 0 0.0271 0.0014 0 0 0.09118 0.1185 Maximum 0.3390 0.5295 0.2937 2.4379 0.0591 2.6977 2.9492 Number of observations 30 30 30 30 30 30 30 2001 Mean 0.1197 0.1286 0.0908 0.1838 0.0183 0.4214 0.5411 Median 0.0823 0.1207 0.0583 0.0964 0.0028 0.3479 0.4422 Minimum 0.0003 0.0359 0.0019 0.0011 0 0.1783 0.1814 Maximum 0.5101 0.2484 0.3251 1.973382 0.0987 2.2831 2.7932 Number of observations 30 30 30 30 30 30 30 Note: Countries in the sample: Argentina, Bulgaria, Chile, Costa Rica, Czech Republic, Dominican Republic, Egypt, Estonia, Guatemala, Croatia, Hungary, Indonesia, Jamaica, Kazakhstan, Lithuania, Latvia, Moldova, Mexico, Malaysia, Nicaragua, Peru, Philippines, Poland, Romania, Slovak Republic, Thailand, Turkey, Uruguay, Venezuela, and South Africa. Source: Author's analysis based on data sources discussed in the text. C I OZ 'Llja-nqoj uo punjrjPjuoW Ituoiluuolul ju /B105 lu.TnopJojxo-joqm//:lt[ umoij poppolumo(I Levy Yeyati 27 where E(r') denotes the expected real return on the assets, and p-, IL, and Ac are zero-mean disturbances to the local inflation rate T, the real (peso-dollar) exchange rate s, and the home country's credit risk (or country risk) c, assumed to be distributed with variance-covariance matrix [Sxy], with Scs= SC, = 0. Assume further that investors maximize risk-adjusted real return: c (2) max U = E(r) - -Var(r) x1 2 subject to the no-short sales condition x' > 0, where r = r . Any solution to the portfolio problem can be characterized by the following dollarization and offshoring ratios:10 (3) A xHD +XFD 1 E(rHP _ HD cV and (4) *y xFD FP = 1 E(rHD _ D cSCC where (5) V Var(rH - rHD) = SJ7 + Sss + SIs and (S., + Sus) (6) AU = +V is the dollarization share in the absence of real return differentials, or the underlying dollarization ratio. In turn, the demand for funds is characterized by a continuum of measure L of local borrowers, assumed to be risk-neutral for simplicity (the assumption can be relaxed without altering the qualitative results, as in Ize and Levy Yeyati 2003), with the option to invest in a production technology with known 9. The qualitative results are not driven by these simplifying assumptions. The derivation of the analytical solutions (reported in the supplemental appendix) presents the solution for the case in which country and real exchange rate risk are not independent. Similarly, the results still carry through if inflation and country risk are partially correlated. A good reference is provided by the correlation between country risk (the sovereign bond spread over a risk-free bond of similar duration) and the rates of inflation and real devaluation. For the sample used in the tests and based on monthly data, these correlations are 0.15 and 0.03, respectively. 10. A derivation of the solution is included in the supplemental appendix. 28 THE WORLD BANK ECONOMIC REVIEW real returns. This investment can be financed from four sources: domestic peso and dollar debt and external peso and dollar debt (where the share of external debt in the liability portfolio is denoted by x*). External debt can be contracted either with private lenders or with the inter- national financial institutions. Under the assumption that country risk is miti- gated by foreign borrowing or by the international financial institutions' preferred creditor status, foreign funds would be cheaper than domestic funds, and in equilibrium there would be no domestic borrowing (and therefore no domestic intermediation of funds). More realistically, credit risk associated with foreign borrowing would increase with the outstanding stock of external debt, and this should be reflected in borrowing costs. For the current exercise it suffices to assume that foreign borrowing costs command a premium 4(X) on the international rate, with ('(X) > 0, where X = x*L. Then, assuming away intermediation costs, interest rate arbitrage implies that (7) E(rHP) = E(rHD) = E(rFP) + qP = E(rFD) + ( which, substituted into equations (3) and (4), implies that A = A,, and = I - (PCSC,. This solution characterizes a continuum of portfolios that maximize the investors' utility. While the availability of external peso assets has no impact on the dollarization ratio if y < Au, it does if y > A,, where external assets in excess of the desired stock of dollar assets should be denominated in pesos (that is, x > y - A). Therefore, if external peso assets were not available, dollarization would be driven entirely by the deposit offshoring ratio (the ratio of resident deposits held abroad to total resident deposits), because in this case the deposit offshoring ratio could never exceed the deposit dollarization ratio. More generally, it can be shown that the new dollarization ratio (identified by the lower bar) is given by:11 (8) (S + S", + SC) 1 V + Scc C V + SeC - where the difference A - A,, increasing in country risk, is due to the absence of investment-grade peso assets: capital flight translates directly into an increase in the dollarization ratio. This analysis can be readily extended to the case of CPI-indexed domestic assets. Perfect indexation could be expressed as /_, = 0. Therefore, if domestic peso assets were indexed to the local CPI, the dollarization ratio from equation 8 would equal S,c/S,+ S,,, and would be driven entirely by country risk. 11. Note that this is also the dollarization ratio that would be obtained should local dollar deposits be banned, leaving the placement of assests abroad as the only option to dollarize savings. Levy Yeyati 29 In other words country risk sets a floor to the extent to which a noninvestment- grade country can reduce financial dollarization through monetary policy (redu- cing inflation volatility and the exchange rate pass-through) or CPI indexation. How does the placement of domestic savings abroad affect the currency composition of resident liabilities? To answer this question, note that the domestic balance of funds requires that (9) (1 - x*)L = (1 - y)A and, for the peso market, (10) (1 - A B)L = (1 - A)A where AB denotes the borrower's dollarization ratio. If the offshoring ratio were not binding, an increase in country risk would not alter the degree of dollarization of domestic savings and, in turn, the amount of peso funds available, as the resulting decline in the domestic supply of pesos would be perfectly offset by an increase in the supply of pesos abroad. However, when the offshoring ratio is binding, the supply of peso funds is automatically determined by the offshoring ratio y, since equations (9) and (10) yield: (11) (1 - A B)L = (1 - y)A = (1 - x*)L or AB = x*. In this case, the borrower effectively has two options: domestic peso loans (limited by the domestic supply of funds) and dollarized foreign borrowing. As country risk mounts, the cost of domestic peso loans relative to dollarized foreign borrowing increases, raising the offshoring ratio (and, as a result, the liability dollarization ratio) to the point at which foreign and domes- tic borrowing costs are again equalized.12 It can be seen that the presence of country-risk-free offshore assets restores condition (10), as the borrower's financing needs can now be met by borrowing pesos abroad (at the peso risk-free rate plus the transaction cost, 4), decoupling the choice of currency and location. In particular, increases in x as a result of higher country risk need no longer have an impact on AB.13 In sum, high country risk is associated with a high offshoring ratio and, if the ratio is sufficiently high, with a smaller supply of peso-loanable funds. 12. It can be readily seen that AB = 1 - (1 - y) (A/L), so that Oy/OS,, > 0 implies that dAB/ aScc > 0. 13. The model implicitly assumes constant investment returns, to focus on the currency choice. Note, however, that under diminishing marginal returns an increase in country risk would raise borrowing costs and reduce aggregate investment without altering the qualitative results in terms of currency composition. 30 THE WORLD BANK ECONOMIC REVIEW To the extent that foreign borrowing is relatively immune to country risk, higher country risk would lead to a larger share of foreign borrowing and, in the absence of risk-free peso assets, higher liability dollarization ratios. Are Nonresidents Different? An argument made repeatedly in the literature is that hedging considerations indicate that resident investors are likely to exhibit smaller dollarization ratios than nonresident investors (see Thomas 1985 for an early reference). The pre- vious example helps illustrate the point. First, note that using A, p e - A, where /_ce denotes nominal exchange rate shocks, underlying dollarization simplifies to Sir (12) A,- S e See the coefficient of a simple regression of the inflation rate on the nominal exchange rate, that is, a crude measure of the exchange rate pass-through. Starting from equation (12), and exploiting the symmetry of this setup, it is easy to verify that in a stylized two-country world the degree of underlying "pesoification" (that is, the share of foreign currency assets over total assets) of foreign residents would be equal to (13) A= S7*e* Se*e* where e* denotes the dollar-peso exchange rate, and 7T* the rate of inflation in the foreign country. For any pair of countries with comparable pass-through coefficients it follows that any coefficient below 50 percent would imply that the demand for local currency assets from residents should exceed that from nonresidents-an asymmetry that deepens when inflation volatility (and the pass-through coeffi- cient) declines and, by extension, when the peso assets are indexed to the local inflation rate. The example oversimplifies the portfolio choice of the representative resident and nonresident investors. In particular, it ignores cross-border transaction costs and differences in the liquidity services provided by financial assets at home and abroad, which in practice introduce an important source of investor heterogeneity. This helps explain why in practice fully dollarized investment portfolios (where the condition y > A, binds) can coexist with a stock of domestic dollar deposits (owned by small investors that face proportionally large transaction costs, or by large investors for liquidity purposes). However, the exercise provides a valid intuition regarding two points that are critically important for assessing the role of international financial Levy Yeyati 31 institutions in the development of local currency markets: the impact that country risk may have (through the shifting of domestic savings abroad) on financial dollarization in noninvestment-grade countries, and the home cur- rency bias that makes local currency assets more appealing to resident investors than to foreigners, particularly in countries with stable inflation. Offshoring and Dollarization in the Data Thus, in the absence of risk-free instruments in emerging market economy (exotic) currencies, noninvestment-grade countries may see a substantial portion of their domestic savings dollarized simply as a result of capital flight to safer investments abroad. This capital flight, inasmuch as it reduces the volume of domestic loanable funds, increases both financing costs and the country's dependence on external finance, to the extent that external finance is perceived as less exposed to country-specific credit risk. In particular, if exter- nal finance offers only partial protection against country risk, one would expect limited access to domestic finance to make noninvestment-grade countries more dependent on international financial institution lending. This section explores whether these intuitions are consistent with the empirical evidence. This is done by distinguishing between the domestic dollarization (ratio of domestic dollar deposits to total domestic deposits) and deposit dollarization, A,, (the ratio of dollar deposits to total resident deposits at home and abroad). The second variable, while less frequently used in the literature, is a more accu- rate measure of the degree of dollarization of residents' portfolios as depicted in the analytical example and more clearly reflects both underlying dollariza- tion and the deposit offshoring ratio. The domestic supply of loanable funds is proxied by the ratio of domestic deposits to GDP.14 Two sources of external dollarization are examined, nonofficial lending (external loans plus external bonded debt) and official lending (which distinguishes between IMF and non-IMF lending). Liability dollarization is computed as the ratio of total foreign currency liabilities to total liabilities, where the currency composition of domestic loans is proxied by that of domestic deposits. Finally, country risk is measured as the stripped spread between sovereign debt and comparable U.S. Treasuries, as captured by J.P. Morgan's EMBI Global.'5 A correlation matrix of period averages for the links between country risk, location and currency composition of resident savings, and the sources of exter- nal dollarization shows an association of country risk with a smaller volume of domestic funds and a higher deposit offshoring ratio, on the one hand, and a high and positive correlation between the offshoring ratio and deposit 14. Domestic deposits are used instead of M2 or domestic credit to be consistent with the way the deposit offshoring ratio is computed. However, all three variables are highly correlated and yield virtually identical results. 15. For methodogical issues concerning the EMBI Global, see J.P. Morgan (1999). 32 THE WORLD BANK ECONOMIC REVIEW dollarization (and in turn between deposit dollarization and country risk), on the other (table 2). Both findings are consistent with the implications of the previous model. Regarding the sources of liability dollarization, there is no clear link between country risk and nonofficial external finance, suggesting that while external debt may provide some protection against country risk (hence, the weaker negative link between these two variables), it does not offset the decline in domestic funds. This decline is ultimately compensated for by a larger dependence on international financial institution lending, as reflected in a larger ratio to GDP as well as in a larger ratio of international financial insti- tution to total external credit. More detailed exploration shows that high-risk countries are associated with smaller domestic deposits (table 3, columns 1-3) and higher deposit offshoring ratios (columns 4-6). This remains true even after controlling for domestic dollarization and for restrictions on dollar deposits that may bias residents toward holding dollar assets abroad if capital flight were motivated by currency risk. (Interestingly, restrictions appear to be positively related to domestic deposits and negatively related to the deposit offshoring ratio.) Thus, country risk contributes to a weak demand for peso assets in noninvestment-grade countries in addition to the underlying dollarization channel identified by the portfolio approach. Indeed, country risk appears to affect deposit dollarization mainly through its effect on the offshoring ratio. As column 7 shows, even after controlling for underlying dollarization, the additional effect of country risk on the share of dollar deposits is similar to the effect on external deposits (which, measured at a country risk mean of 575 basis points yields a sizable increase of about 10 percent). This effect virtually vanishes, however, after controlling for the deposit offshoring ratio, which is reflected almost one to one in the deposit dollarization ratio (column 8).16 The link between the deposit dollari- zation ratio and the deposit offshoring ratio also holds for the larger sample obtained by dropping the country-risk index (column 9) and in a dynamic setting with country-fixed effects (columns 10 and 11).7 A partial regression plot from a regression of excess deposit dollarization (measured as the difference between the observed and the underlying dollariza- tion ratios) on the offshoring ratio (including per capita GDP as a proxy for development and institutional quality) provides another illustration of the 16. Underlying dollarization is computed here directly from equation (6), based on monthly inflation and real exchange rate data. Note that the fact that observed dollarization depends on both underlying dollarization and offshoring follows directly from the model, once differential cross-border transaction costs for individual savers are allowed for. See the supplemental appendix for a derivation. 17. To control for the possible incidence of common macroeconomic variables omitted in the specification, the cross-section regressions were rerun including per capita GDP as a proxy for time-invariant institutional and development factors and the panel regressions of columns 10 and 11 adding the inflation and the real growth rates. The results, which remain virtually unchanged, are reported in the supplemental appendix. TABLE 2. Correlation Matrix for Measures of Financial Dollarization in emerging Market Economies (Period Averages) Dollar external International International Deposit Deposit liabilities, excluding financial institution financial institution Domestic offshoring dollarization international financial lending, excluding lending / total deposits ratio ratio institutions IMF IMF lending external liabilities Deposit offshoring ratio -0.5429 (0.0023) Number of observations 29 Deposit dollarization ratio -0.4163 0.6803 (0.0540) (0.0005) Number of observations 22 22 Dollar external liabilities, 0.4209 -0.2224 -0.1390 excluding international (0.0455) (0.3076) (0.5588) financial institutions Number of observations 23 23 20 International financial -0.2383 0.2876 0.2052 -0.4280 institution lending, (0.2313) (0.1542) (0.3722) (0.0469) excluding IMF Number of observations 27 26 21 22 IMF lending -0.2714 0.1522 0.2676 -0.1778 0.2520 (0.1709) (0.4578) (0.2408) (0.4285) (0.2048) Number of observations 27 26 21 22 27 International financial -0.1246 0.1128 0.1667 -0.5470 0.7935 0.1307 institution lending / total (0.5357) (0.5832) (0.4703) (0.0084) (0.0000) external liabilities (0.5159) Number of observations 27 26 21 22 27 27 Country risk -0.5646 0.5005 0.5047 -0.1886 0.5633 0.5073 0.3652 (0.0012) (0.0057) (0.0166) (0.3774) (0.0022) (0.0069) (0.0610) Number of observations 30 29 22 24 27 27 27 Note: Numbers in parentheses are significance levels. Averages are computed based on observations for which the country risk index is available. Domestic deposits, dollar external liabilities, and international financial institution and IMF lending are computed as a ratio of GDP. Source: Author's analysis based on data sources discussed in the text. 0lOt 'LXjr-aqoj no p-njrjPjuoW Ituop~utuolul u /'Jo-sltunoFpJojxoxoqm//:dBlt umoij popolumo( TABLE 3. Country Risk, Deposit Offshoring Ratio and Deposit Dollarization in Emerging Market Economies Domestic deposits as share Deposit offshoring ratio Deposit dollarization ratio o of GDP Ordinary least Ordinary least squares (period squares (period averages) averages) Ordinary least squares Fixed effects (annual (period averages) data)a (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) O Country risk -0.030* -0.039* -0.031* 0.015* 0.016*** 0.013*** 0.018*** 0.005 0.004** o (0.007) (0.008) (0.007) (0.004) (0.008) (0.007) (0.009) (0.004) (0.002) Domestic deposit dollarization ratio 0.089 0.033 (0.164) (0.128) Restrictions 0.045 -0.031*** (0.047) (0.016) Underlying dollarization ratio 0.426* 0.390* 0.265* (0.089) (0.062) (0.050) Deposit offshoring ratio 0.815* 0.818* 0.852* 0.701* (0.158) (0.100) (0.118) (0.034) Constant 0.560* 0.588* 0.521* 0.216* 0.210* 0.254* 0.261* 0.093 0.179* 0.222* 0.290* (0.069) (0.092) (0.072) (0.029) (0.047) (0.046) (0.054) (0.056) (0.043) (0.036) (0.023) Number of observations 30 23 24 29 22 22 21 21 78 107 584 R-squared 0.32 0.39 0.35 0.25 0.25 0.31 0.68 0.86 0.62 0.99 0.98 *Significant at the 1 percent level; **significant at the 5 percent level; ** *significant at the 10 percent level. Note: Numbers in parentheses are robust standard errors. aRegressions include year dummy variables. Source: Author's analysis based on data sources discussed in the text. 00Zt'L,AIja-iqoj uo p-njrPjuoW Ituoiliuuu ju /'Jo-sltunoFpJojxoxnoqm//:dBlt umoij poppolumo( Levy Yeyati 35 FicURE 1. Deposit Offshoring Ratio and Excess Deposit Dollarization (partial regression plot) 8* X* t*. 0 -* * * 0** *0* -0.4 -0.2 -5.551e-17 0.2 0.4 e (offshorization I X) Coefficient = 0.64200696, (robust) se = 0.07970967, t= 8.05 Note: Excess dollarization is measured as the difference between actual and underlying deposit dollarization. Source: Author's analysis based on data sources discussed in the text. influence of the deposit offshoring ratio on the deposit dollarization ratio (figure 1). The correlation is highly significant and is not driven by a few outliers. The deposit offshoring ratio is also associated with greater dependence on international financial institution lending, both at different points in time and over time for IMF and non-IMF lending (table 4). The offshoring ratio is associated with a shortage of loanable funds rather than an excess supply of them (as would be the case in middle-income countries like Chile and most high-income countries where outflows are typically associated with a scarcity of investment options and low returns). In contrast, no significant link is found between the deposit offshoring ratio and other sources of external credit (table 5, columns 1 and 2) or total external long-term liabilities, used as a proxy for foreign currency external debt (columns 3 and 4)." This suggests that if a foreign jurisdiction provides better protection against credit risk as is sometimes speculated, this benign effect is not enough to make up for the 18. Data on total external long-term liabilities are available from the World Bank's Global Development Finance for a larger sample and a longer period. The implicit assumption that all external debt issued by nonindustrial countries is denominated in foreign currency seems to be a reasonable approximation of reality. o TABLE 4. Deposit Offshoring ratio and Lending by International financial Institutions, Emerging Market Economies Non-IMF lending as share of GDP IMF lending as share of GDP Ordinary least squares Ordinary least squares (averages) Fixed effects (annual data) (averages) Fixed effects (annual data) )aO (1), (2) (3) (4) (5)* (6) (7) (8) Deposit offshoring ratio 0.678* 1.221*** 0.376*** 0.235** 0.025 0.083*** 0.030* 0.086** (0.349) (0.306) (0.079) (0.094) (0.026) (0.023) (0.008) (0.041) Country risk 0.008*** 0.001 (0.002) (0.001) Constant -0.004 0.078 0.387*** 0.082* ** 0.009 -0.001 0.018*** -0.013 (0.096) (0.095) (0.033) (0.029) (0.008) (0.007) (0.003) (0.016) Number of observations 25 120 815 125 26 120 816 125 R-squared 0.19 0.09 0.97 0.98 0.02 0.09 0.93 0.82 *Significant at the 10 percent level; **significant at the 1 percent level; ***significant at the 5 percent level. Note: Numbers in parentheses are robust standard errors. alncludes only observations for which the country risk index is available. Source: Author's analysis based on data sources discussed in the text. C I OZ 'LXjr-mqoj no p-njrPjuoIt uopiuIulu u /'Jo-slunopojxo-jaqm//:dt[ umoij popolumo(I TABLE 5. Offshoring and Liability Dollarization in Emerging Market Economics Dollar external liabilities as share Total external liabilities as share of GDP, excluding international of GDP, excluding international Liability dollarization ratio, Liability dollarization financial institutions financial institutions excluding IMF ratio Ordinary least Ordinary least squares Fixed effects squares Fixed effects Ordinary least squares Ordinary least squares (averages) (annual) (averages) (annual) (averages) (averages) (1) (2) (3) (4) (5) (6)' (7) (8)a Deposit offshoring ratio -0.194 0.209 -0.034 0.080* 0.547* 0.564** 0.558** 0.561*** (0.184) (0.136) (0.075) (0.036) (0.213) (0.132) (0.213) (0.111) Constant 0.278** 0.109** 0.141*** 0.082** 0.462** 0.515** 0.465*** 0.521*** (0.064) (0.054) (0.015) (0.074) (0.054) (0.074) (0.046) Number of observations 23 301 120 815 38 88 38 88 R-squared 0.05 0.80 0.00 0.84 0.18 0.23 0.18 0.23 *Significant at the 5 percent level; **significant at the 1 percent level; ***significant at the 10 percent level. Note: Numbers in parentheses are robust standard errors. aUses total long-term external debt as a proxy for dollar long-term external debt. Source: Author's analysis based on data sources discussed in the text. Z C I OZ 'LXjr-mqoj no p-njrPoo X~uIt uop11uulu 1u /'Jo-sltunopJojxoxoqm//:dBqt uioij popolumo( 38 THE WORLD BANK ECONOMIC REVIEW deficit in domestic funds when sovereign risk is driving capital abroad. Finally, liability dollarization is positively correlated with deposit offshoring (columns 5-8), a reflection of the higher dependence on international financial insti- tution lending revealed by the previous results. In sum, the evidence is consistent with the hypothesis that, in the absence of risk-free instruments in exotic currencies, noninvestment-grade countries may see a substantial portion of their domestic savings dollarized simply as a result of the flight of capital to safer investments abroad. In turn, this capital flight, inas- much as it reduces the volume of domestic loanable funds, increases the coun- try's dependence on dollarized international financial institution lending (better equipped to cope with country risk than international investors), shifting the currency liability composition towards the foreign currency as a consequence. III. FINANCIAL DEDOLLARIZATION AND THE INTERNATIONAL FINANCIAL INSTITUTIONS As the previous discussion highlights, in noninvestment-grade countries inter- national financial institutions tend to substitute for domestic sources of finance. Crucially, the role of international financial institutions in this context does not necessarily entail, as sometimes argued, a significant subsidy to emerging market economies. Indeed, a key characteristic of international financial insti- tutions is that, unlike private investors, they experience a surprisingly good repayment record.19 Thus, international financial institutions can be seen as contributing to a "sovereign risk transformation," matching the supply of private funds in search of investment-grade securities and the demand for funds by noninvestment-grade economies, exploiting their superior enforce- ment capabilities to channel these funds into virtually risk-free lending. It is only natural, then, to use this advantage to foster the supply of local currency funds that are lost due to sovereign risk considerations. As noted, this would not require the extension of additional lending by the international financial institutions, but rather the issuance of investment-grade paper, to meet the demand for risk-free local currency securities, and use of the proceeds to convert part of the outstanding stock of international financial institution loans to keep a balanced currency position. This section reviews the facilities offered by international financial insti- tutions to hedge their clients' currency exposure and proposes a scheme to dedollarize international financial institution lending to meet the demand for investment-grade local currency securities by residents. It addresses the main criticisms leveled against both old and new initiatives of this type. 19. Recent work has revealed that the subsidy component in IMF nonconcessional lending to emerging economies is virtually nill (Jeanne and Zettelmeyer 2001)-a result that would also apply to other international financial institutions with preferred creditor status. Levy Yeyati 39 What's on the Menu? The supply of hedging instruments available from international financial insti- tutions is still rather limited. The World Bank, for example, offers the option of converting outstanding loan obligations (or of requesting a swap of its foreign currency obligations) into local currency. Since World Bank loans are funded in foreign currency, the transaction requires that the World Bank arrange a local- foreign currency swap with another financial institution to transfer the currency exposure.20 These local currency products are not without benefits, particularly for noninvestment-grade clients that would otherwise be unable to access cur- rency swaps directly in capital markets. However, they are typically limited in volume, shorter than the loan they are intended to hedge, and granted on a case-by-case basis subject to the existence of a liquid swap market. Their greatest shortcoming, however, is that rather than expanding the pool of local currency funds, these local currency products tap into existing swap markets. Thus, while they may benefit local borrowers through reduced trans- action costs, they are also likely to crowd out the available supply of hedging instruments. At any rate, possibly because of their limited benefits, these rela- tively new products have not been in high demand.22 The World Bank has also launched a few issues in investment-grade exotic currencies.23 The modality is not uniform. For example, in February 2000, three-year euronote in Mexican pesos was issued abroad and was placed largely among U.S. investors on the back of strong external demand shortly after rating agencies announced that they were considering an upgrade of Mexico's debt to investment grade. By contrast, in May 2000, Chilean CPI-indexed peso five-year euronote was distributed mainly among domestic institutional investors, who purchased about 75 percent of the issue. While these issues may have positive spillovers for the development of local currency markets, they contribute little to a dedollarization agenda, as the risk- transformation role is less valuable for economies that already enjoy investment- grade status. The previous analysis suggests that the best use of the international financial institutions' advantage entails external issues (to minimize the crowd- ing out of available domestic funds) in noninvestment-grade countries that are unable to attract domestic investors in search of low-risk assets. 20. For details, see the brochure on local currency financial products posted on www.worldbank. org/fps/hedging.htm. Currency swaps are also offered by the IDB. 21. The emerging market economies that satisfied this condition by end-2003 according to the World Bank were Brazil, Chile, Colombia, the Czech Republic, Hungary, India, Indonesia, Republic of Korea, Mexico, Malaysia, Philippines, Poland, the Slovak Republic, South Africa, and Thailand. Of these, only Colombia, India, Indonesia, and Philippines are noninvestment-grade countries. 22. As of April 2004, only three countries had signed the Master Derivatives Agreement required by the World Bank to request a currency swap. 23. A list of recent World Bank issues can be found in www.worldbank.org/debtsecurities/ recent issues.htm. 40 THE WORLD BANK ECONOMIC REVIEW A move in this direction was the March 2004 Colombian CPI-indexed bond, issued and placed domestically by the World Bank with domestic institutional investors. While it still has the potential to crowd out existing (captive) demand for peso assets, the bond was nonetheless welcomed by the govern- ment as a way to satisfy demand from the growing private pension system for long-term risk-free assets in the local currency instead of abroad. Closer to the scheme proposed here was the May 11, 2004, eurobond in Brazilian reais issued by the Inter-American Development Bank (IDB) that included restric- tions on domestic sales to avoid crowding out domestic resources. While these issues reflect a welcome shift in the funding strategies of some international financial institutions, they have been motivated by the search for lower funding costs. Indeed, the proceeds have been immediately swapped into dollars rather than being used to convert outstanding loans into the same exotic currencies. Thus, despite the merits, their effective impact in terms of dedollarizing the liabilities of emerging economies has been virtually nill. What Has Been Proposed? Most of the discussion about what type of peso instrument to substitute for the current dollar assets has centered on CPI indexation, an avenue that proved suc- cessful in containing and undoing financial dollarization in Chile and Israel. While the local CPI is the most obvious candidate index for domestic residents, several alternatives have been proposed when the aim is to attract foreign investors. Of these, two stand out: a GDP index (see Borensztein and Mauro 2004) and a com- modity index (see Caballero and Panageas 2003 on copper-indexed debt in Chile). While the existence of derivative markets makes commodities easier to price and hedge, their use is limited to commodity exporters and by the correlation of commodity exports with the country's income. Moreover, as with the cur- rency swap discussed above, it is not clear how indexation improves on a short hedge purchased directly by the issuer in the derivatives markets. And while GDP indexation may be more suitable for smoothing out countercyclical varia- tions in debt-to-GDP ratios and borrowing costs, it is difficult to see how GDP risk can be stripped and hedged by potential investors, particularly in the absence of a market for GDP indexes.24 Similar caveats apply in principle to CPI-indexation as a way of luring non- resident investors. Eichengreen and Hausmann (2004) stress the attractiveness of a basket of CPI-indexed exotic currencies for nonresidents and propose that international financial institutions issue debt in these currencies to fund their own lending to emerging market economies and provide the needed startup liquidity. This requires matching not only demand and supply in each currency but also across currencies to allow for the needed diversification strategy, a not insubstantial coordination effort. Furthermore, while speculative nonresident 24. This may explain why the GDP-indexed clause attached to the bond offered by Argentina in the 2005 debt exchange was significantly underpriced by the market. Levy Yeyati 41 demand for specific currencies perceived as undervalued is not unlikely (as the IDB issue in Brazilian reais attests), interest from long-run international investors seeking a diversified portfolio with stable returns is more difficult to envisage. International Financial Institutions and Intermediation of Resident Savings Once the focus shifts from foreign investors to demand from residents, there are important advantages to using CPI indexation. It can be measured fre- quently (improving the accuracy of the indexation) by an autonomous agency (ensuring that the index is free from government manipulation). More impor- tantly, unlike other indexes, there is a natural demand for the CPI arising from the hedging properties highlighted above. That makes it the obvious choice to jumpstart the dedollarization process with the help of an investment-grade issuer (international financial institutions) that decouples sovereign and cur- rency risk to attract domestic investors seeking stable real returns at a reason- able level of credit risk.25 Resorting directly to the domestic market, however, may have economic (and political) drawbacks, as it crowds out other local currency funds by indu- cing a shift from high-risk government and corporate debt to investment-grade international financial institution paper. While the new issue may extend the domestic market for local currency securities by bringing in investors pre- viously reluctant to assume country risk, it is likely to increase the cost of funds domestically, inducing the government to borrow abroad, resulting in only minor changes in the overall composition of government liabilities. Thus, to maximize the beneficial composition effect, the new debt should be issued in international markets. Both the literature and recent experience point to institutional investors as the natural target of these issues. Consider, for example, the case of pension funds. By acquiring a credit-risk-free asset denominated in units of the consump- tion basket, they fulfill their role as guarantors of a stable stream of real income after retirement, while avoiding country-specific credit risk. Pension funds are typically allowed to invest a share of their portfolio in investment-grade foreign assets, a share that has been growing since the Argentine debacle sounded the alarm on excessive exposure to sovereign risk. Thus, while the government borrows from international financial institutions, a share of residents' retirement savings is being invested abroad in triple A paper such as that issued by inter- national financial institutions to fund their loans. The international financial institutions can readily channel these funds back into the domestic economy by selling CPI-indexed bonds to the pension funds to finance loans denominated in the same index. This demand for long-dated investment-grade local currency 25. Risk decoupling is at the heart of the Eichengreen and Hausmann (2004) proposal as well. However, currency risk is tolerated to the extent that it can be diversified away in a basket of exotic currencies. In the version proposed here CPI indexation eliminates currency risk from the resident's standpoint, so that no currency diversification is required. 42 THE WORLD BANK ECONOMIC REVIEW paper from institutional investors could reach high levels as pension fund stocks accumulate (see table S2 in the supplemetal appendix). An alternative approach is the use of international financial institution guar- antees of local currency debt to reduce the credit risk of noninvestment-grade issues in exotic currencies. Halfway between a risk-free international financial institution bond and a risky emerging market economy paper, this combination (if guarantees are capped in dollars) would entail similar risks for the inter- national financial institutions as those associated with existing guarantees of dollar bonds.26 As noted, a few successful international financial institution issues in exotic currencies have already revealed the existence of demand for these securities. That makes the redollarization of their proceeds particularly puzzling and at odds with the concerns about financial dollarization repeatedly expressed by international financial institution officials and publications (see, for example, IDB 2006). Considering the high exposure of these institutions with many of their clients, it is easy to see how instead of the swap with a third financial institution that followed bond issuance, they could have dealt directly with the client, par- tially dedollarizing outstanding obligations. While the cash flows of the bond and loan would typically differ, swap markets provide sufficient flexibility to match both schedules with little, if any, additional transaction costs. The settlement cur- rency of both streams of cash flows would be immaterial in this case. Even if the currency of the original loan is preserved, this obligation would be indexed to the local currency (or the local CPI), eliminating any currency exposure-an argu- ment also valid for new lending. And for the same reason there is no obvious rationale for limiting the currency conversion to the local expenditure component of the loan (as is currently the case for existing local currency products). An appropriate hedging strategy would need to match the currency composition of liabilities with that of future earnings (as opposed to past expenses). In October 2005 the Asian Development Bank launched a Philippine bullet peso bond in the Philippine capital market. The issue was oversubscribed and broadly placed among resident institutional investors including banks, insur- ance companies, fund management companies, and trust departments. The proceeds of the bond issue will be used for a peso-denominated loan to Balikatan Housing Inc. (BHI), a special-purpose vehicle jointly owned by the National Home Mortgage Finance Corporation of the Philippines and Deutsche Bank. In sum, there seems to be no obvious obstacle to onlending the funds obtained from local currency issues to emerging market economy clients. Addressing the Skeptics Besides the mixed reviews from market participants, the proposals to dedol- larize international financial institution lending have faced internal criticism. 26. In this vein the IDB has recently approved a policy change to enhance its partial credit guarantees for Latin American and Caribbean debt issuers. Levy Yeyati 43 This article briefly discusses two of them, as summarized by Rajan (2004).27 First, he points out that a portfolio approach to financial dollarization should take into account the correlation between financial returns and nonfinancial income. More precisely, to the extent that economic activity (and, as a result, real nonfinancial income) is negatively correlated with the nominal exchange rate, local savers would demand lower returns on dollar assets that are used as a hedge against economic downturns. In principle, however, this effect would be offset by the mirror impact that exchange rate procyclicality has on the borrower: because real earnings decline just when the burden of dollar debt increases in real terms, local debtors would be willing to pay the higher returns demanded on peso assets to hedge their own income stream. Ultimately, the net effect of introducing nonfinancial income into the picture is not straightforward. The second argument is more relevant to the discussion here. Under myopic behavior, emerging market economy borrowers would be expected to exploit the lower dollar borrowing costs in good times, disregarding the contingent cost of the associated exposure-likely to be borne by others.28 Note that while the peso interest rate charged by international financial institutions would be below that demanded by private lenders (because of the lower credit risk), the conversion of outstanding international financial institution loans to the local currency would not save debtors the currency risk premium that induced dollarization in the first place. In other words if financial dollarization were the result of asymmetric risk pricing, rather than lack of investment-grade local currency assets as argued here, opportunistic debtors would turn down the offer to insure against future balance sheet effects at a fair price. If so, the proposed dedollarization strategy, rather than suffering from lack of investor interest, might be condemned by the indifference of the very debtors that it is intended to relieve. This moral hazard argument looks rather less persuasive in light of recent dedollarization efforts in emerging market economies.29 Moreover, the 27. A third, more general argument concerns the fear of de-indexation (that is, the forcible conversion of indexed assets to nominal ones) that may inhibit the development of the market for CPI-indexed assets. Although CPI indexation is far from a necessary condition for dedollarization, it has played an important role in most recent attempts to issue local currency debt at reasonably long maturities. However, unilateral de-indexation would affect the international financial institution loan, but not the bond (which would still be indexed). Indeed, de-indexation would amount to defaulting on an international financial institution-an unlikely outcome. 28. Variations on this argument have been examined in the literature in relation to market imperfections such as implicit guarantees (Burnside, Eichenbaum, and Rebelo 2001) and currency-blind regulation (Broda and Levy Yeyati 2006) that are conducive to excessive dollarization. 29. Some examples include the gradual dedollarization of public debt in post-tequila crisis Mexico and, more recently, in Brazil; the introduction of local currency assets in Peru and CPI-indexed assets in Uruguay, along with a revision of the prudential framework on dollar intermediation; and the imposition of quantitative restrictions on the on-lending of onshore dollar deposits in Argentina after the demise of the currency board. 44 THE WORLD BANK ECONOMIC REVIEW successful international placement of local currency sovereign bonds by noninvestment-grade countries at long maturities (such as the recent "Brazil 2016") appears to confirm the view that the inability to issue local currency debt in international markets (the so-called "original sin") may have reflected the sovereign's unwillingness to pay the needed currency premium-an attitude that has been fading as the premium declines and awareness of the perils of dollarization increases. Nonetheless, taking the moral hazard argument at face value, one can only conclude that, if the international financial institutions have correctly interna- lized the welfare of the country, it would be in their best interest to undo this imperfection by changing the terms on which they offer dollar lending (more generally, by including dedollarization in their standard set of conditionalities) rather than by supplying misleadingly cheap dollar loans that perpetuate this perverse cycle. Ultimately, agency problems provide yet another reason for international financial institutions to adopt a more proactive stance. The IMF's new focus on currency mismatches as prudential indicators represents a move in this direction. IV. CONCLUDING REMARKS This article has tried to convey a simple message: to the extent that country risk induces financial dollarization through the placement of domestic savings abroad, international financial institutions can exploit their superior enforce- ment ability to intermediate these savings back into the domestic economy, undoing financial dollarization. For international financial institutions this would not require expanding credit, transferring resources, or incurring cur- rency risk. Rather, it would involve issuing local currency bonds and using the proceeds to gradually convert current loans into (or refinance maturing loans in) the local currency. This initiative represents a feasible starting point for the much needed development of local currency markets, not a final solution to the dollarization problem. While successful issues of international financial insti- tution debt in exotic currencies are an encouraging first step, a coordinated effort is still needed to convince governments and the international financial institutions of the benefits of using the proceeds to dedollarize multilateral lending. The scheme described above is not a sufficient condition for reducing financial dollarization in emerging market economies. Needless to say, the demand for local currency assets (and, more generally, the achievement of financial stability) would be contingent on the consistent implementation of responsible economic policies. However, while following good policies is by definition good advice, doing so is not always sufficient for collecting the full reward. It is here that the international financial institutions can make a contribution. Levy Yeyati 45 APPENDIX Variable sources and definitions * Onshore dollar (peso) deposits: Foreign (local) currency deposits with dom- estic banks. Source: Levy Yeyati (2006). * Onshore deposits: Onshore dollar deposits + onshore peso deposits. * Offshore deposits: Cross-border deposits by residents with banks domiciled in Bank for International Settlements reporting countries. Source: Bank for International Settlements. * Total deposits: Offshore deposits + onshore deposits. * Deposit offshoring ratio: Offshore deposits/total deposits. * Deposit dollarization ratio: (Onshore dollar deposits + offshore deposits)/ total deposits. * Onshore deposit dollarization ratio: Onshore dollar deposits/onshore deposits. * External loans: Cross-border loans to residents from banks domiciled in Bank for International Settlements reporting countries. Source: Bank for International Settlements. * Dollar-bonded external debt: Private and public external bonds denomi- nated in foreign currency; stocks outstanding. Source: Bank for International Settlements. * Peso-bonded external debt: Private and public external bonds denominated in local currency; stocks outstanding. Source: Bank for International Settlements. * International financial institution lending: Long-term debt with official creditors. Public and publicly guaranteed debt from official creditors includes loans from international organizations (multilateral loans) and loans from governments (bilateral loans). Unit: millions of U.S. dollars. Source: World Bank (2003). * IMF lending: Use of IMF credit. Denotes repurchase obligations to the IMF with respect to all uses of IMF resources, excluding those resulting from drawings in the reserve tranche. Source: World Bank (2003). * Dollar external liabilities: External loans + dollar-bonded external debt + International financial institution lending. * Total external debt: Includes public and publicly guaranteed long-term debt, private nonguaranteed long-term debt, use of IMF credit, and esti- mated short-term debt outstanding. Source: World Bank (2003). * Short-term external debt: Debt with an original maturity of one year or less. Source: World Bank (2003). 46 THE WORLD BANK ECONOMIC REVIEW * Total long-term external debt: Total external debt - short-term external debt. Used in some tests as an alternative measure of dollar external liabil- ities. Source: World Bank (2003). * Liability dollarization ratio: (Dollar external liabilities + dollar domestic deposits)/(dollar external liabilities + peso-bonded debt + domestic depos- its). The currency composition of deposits is used to proxy the currency composition of domestic loans. * Country risk: J.P. Morgan Bond EMBI Global index. Included in the EMBI Global are U.S. dollar-denominated Brady bonds, eurobonds, traded loans, and local market debt instruments issued by sovereign and quasi- sovereign entities. Source: J.P. Morgan. * Restrictions: Index of restrictiveness of rules on resident holdings of foreign currency deposits onshore as of beginning of 2001. Source: Levy Yeyati (2006) based on IMF (2001), following the methodology proposed by De Nicol6, Honohan, and Ize (forthcoming). * Underlying dollarization ratio: [Var(7T) - Cov (7, s)] / [Var (7T) + Var(s) - 2Cov (iw, s)], where 7T and s are the monthly inflation and real devaluation rates. Source: IMF, various years, International Financial Statistics. REFERENCES Berganza, J. C., and A. Garci Herrero. 2004. "What Makes Balance Sheet Effects Detrimental for the Country Risk Premium?" Bank of Spain, Madrid. Bordo, M., C. Meissner, and A. Redish. 2002. "How "Original Sin" Was Overcome: The Evolution of External Debt Denominated in Domestic Currencies in the United States the British Dominions 1800-2000." NBER Working Paper 9841. Cambridge, Mass.: National Bureau of Economic Research. Borensztein, E., and P. Mauro. 2004. "The Case for GDP-Indexed Bonds." Economic Policy 19(38): 165-216. Broda, C., and E. Levy Yeyati. 2006. "Endogenous Deposit Dollarization." Journal of Money, Credit, and Banking 38(4):963-88. Burnside, C., M. Eichenbaum, and S. Rebelo. 2001. "Hedging and Financial Fragility in Fixed Exchange Rate Regimes." European Economic Review 45(7):1151-93. Caballero, R., and S. Panageas. 2003. "Hedging Sudden Stops and Precautionary Recessions: A Quantitative Framework." NBER Working Paper 9778. Cambridge, Mass.: National Bureau of Economic Research. Caballero, R., K. Cowan, and J. Kearns. 2005. "Fear of Sudden Stops: Lessons from Australia and Chile." The Journal of Policy Reform 8(4):313-54. Calvo, G., A. Izquierdo, and L. F. Mejia. 2004. "On the Empirics of Sudden Stops: The Relevance of Balance-Sheet Effects." Inter-American Development Bank, Washington, D.C. Claessens, S., D. Klingebiel, and S. Schmukler. 2003. "Government Bonds in Domestic and Foreign Currency: The Role of Macroeconomics and Institutional Factors." World Bank, Washington, D.C. De la Torre, A., and S. Schmukler. 2004. "Coping with Risk through Mismatches: Domestic and International Financial Contracts for Emerging Economies." International Finance 7(3):349-90. Levy Yeyati 47 De Nicol6, G., P. Honohan, and A. Ize. (Forthcoming). "Dollarization of the Banking System: Good or Bad?" Journal of Banking and Finance. Eichengreen, B., and R. Hausmann. 2004. "Original Sin: The Road to Redemption." In B. Eichengreen, and R. Hausmann eds., Other People's Money: Debt Denomination and Financial Instability in Emerging-Market Economies. Chicago: University of Chicago Press. Eichengreen, B., R. Hausmann, and U. Panizza. 2003. "Currency Mismatches, Debt Intolerance and Original Sin: Why They Are Not the Same and Why It Matters." NBER Working Paper 10036. Cambridge, Mass.: National Bureau of Economic Research. Frankel, J. (Forthcoming). "Contractionary Currency Crashes in Developing Countries." IMF Staff Papers. Goldstein, M., and P. Turner. 2003. Controlling for Currency Mismatches in Emerging Economies. Washington, D.C.: Peterson Institute for International Economics. Herrera, L. 0., and R. Valdez. (Forthcoming). "Dedollarization, Indexation and Nominalization: The Chilean Experience." Journal of Policy Reform. IDB (Inter-American Development Bank). 2006. Living with Debt: How to Limit the Risk of Sovereign Finance. Cambridge, Mass.: Harvard University Press. IMF (International Monetary Fund). 2001. Annual Report on Exchange Arrangements and Exchange Restrictions. Washington, D.C. Various years. International Financial Statistics. Washington, D.C. Ize, A., and E. Levy Yeyati. 2003. "Financial Dollarization." Journal of International Economics 59(2):323-47. Jeanne, 0., and J. Zettelmeyer. 2001. "International Bailouts, Moral Hazard, and Conditionality." Economic Policy 16(33):409-32. Levy Yeyati, E. 2006. "Financial Dollarization: Evaluating the Consequences." Economic Policy 21(45):61-118. Morgan, J. P. 1999. "Introducing the J. P. Morgan Emerging Markets Bond Index Global (EMBI Global)." August 3, 1999, J. P. Morgan Securities Inc. Martinez, L., and A. Werner. 2002. "Capital Markets in Mexico: Recent Developments and Future Challenges." Central Bank of Mexico, Federal District of Mexico, Mexico City. Rajan, R. 2004. "How Useful Are Clever Solutions?" Finance & Development 41(1 March):56-57. Thomas, L. R. 1985. "Portfolio Theory and Currency Substitution." Journal of Money, Credit, and Banking 17(3):347-57. World Bank. 2003. Global Development Finance 2003. Washington, D.C. Business Cycle Synchronization and Regional Integration: A Case Study for Central America Norbert Fiess Deeper trade integration between Central America and the United States, as envisaged under the Central American Free Trade Agreement, is likely to lead to closer links between Central American and U.S. business cycles. This article assesses the degree of business cycle synchronization between Central America and the United States-relevant not only for a better understanding of the influence of important trading partners on the business cycle fluctuations in the domestic economy but for evaluating the costs and benefits of macroeconomic coordination. JEL codes: F15, F42. In early January 2003 the United States and Costa Rica, El Salvador, Guatemala, Honduras, and Nicaragua launched official negotiations for the Central American Free Trade Agreement (CAFTA, renamed DR-CAFTA after the Dominican Republic joined the negotiations in 2004).1 Once ratified by all members, DR-CAFTA will expand trade barrier reductions similar to those in the North American Free Trade Agreement (NAFTA) to Central America. DR-CAFTA is part of a bigger project to promote regional integration through- out the Americas, with the ultimate aim of establishing a Free Trade Area of the Americas. An open question for any trade integration initiative is the macroeconomic consequences and so the implications for macroeconomic policies. Like NAFTA, DR-CAFTA contains no explicit provisions on macroeconomic policy. But just as NAFTA has affected Mexico's macroeconomic dynamics (Lederman, Maloney, and Serven 2005), DR-CAFTA has the potential to change the macroeconomic dynamics between Central America and the United States, which could in turn alter the desirability to coordinate fiscal and monet- ary policies between these countries. Norbert Fiess is the deputy director of the Centre for Development Studies in the Department of Economics at the University of Glasgow; his email address is n.fiess@socsci.gla.ac.uk. A supplemental appendix to this article is available at http://wber.oxfordjournals.org. 1. This article's focus is limited to the initial CAFTA countries. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 49-72 doi:10.1093/wber/1hl014 Advance Access Publication 24 January 2007 () The Author 2007. 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 49 50 THE WORLD BANK ECONOMIC REVIEW Trade is often perceived as an important-if not the most important- transmission channel for business cycles from one country to another. Theoretically, the impact of trade integration on business cycle synchronization is unclear. Increased trade can lead business cycles to diverge or converge. If trade integration leads to increased interindustry trade as a part of a specialization process, business cycles are likely to diverge because shocks specific to particular industries will become responsible for shaping business cycles. But if trade inte- gration leads to a higher share of intraindustry trade, business cycles will converge because industry-specific shocks will affect trading partners in a similar way. If business cycles are similar and shocks are common, coordination of macroe- conomic policies can become desirable, with a common currency as the ultimate form of policy coordination. But if shocks are predominately country-specific, resulting in little business cycle synchronization, independent monetary and fiscal policies are generally seen as important in helping an economy adjust to a new equilibrium. Clearly, if business cycles are affected predominately by country- specific shocks and are likely to continue to be so, intensified macroeconomic coordination as part of regional integration might do more harm than good. Frankel and Rockett (1988) show that if macroeconomic coordination is based on the wrong economic model, it can make countries worse off than under noncooperation. The aim of this article is to come closer to the "true" economic model by providing information about the current trade structure and the degree of business cycle synchronization between Central America and the United States. Because both business cycle synchronization and trade structure are expected to change with trade integration, knowledge of the status quo will provide crucial information for future policy analysis. This article does three things. First, it uses state-of-the-art econometric tech- niques to measure the degree of business cycle synchronization between Central America and the United States-its main trading partner. Second, it calculates inter and intraregional trade for Central America to quantify the relationship among trade intensity, trade structure, and business cycle synchronization; this is followed by a discussion of how trade integration within DR-CAFTA is likely to shape future business cycle patterns in the region. Third, it offers policy advice on the appropriateness of macroeconomic coordination for Central America con- ditional on its trade structure. Given El Salvador's unilateral dollarization in 2000, it seems highly relevant to inform the debate on this front. Restricted data availability for Central America seriously limits the scope for econometric analysis. To maximize inference about the level of business cycle synchronization and the link between trade structure and business cycle synchro- nization in Central America, two sets of data are analyzed: annual data on GDP from 1965 to 2005 and monthly data on economic activity from 1995 to 2005. The annual data span a longer time period and allow an analysis of changes in business cycle synchronization over time. Because business cycles are usually defined as 6-32 quarters, the higher frequency of the monthly data should provide additional insight into business cycle synchronization over the more Fiess 51 recent period. Most Central American countries went from extreme instability marked by hyperinflation and civil war to a period of peace and economic reform in the 1990s. Thus, the later, more tranquil period is likely to be more useful for predicting future developments in business cycle synchronization. Most of Central America's trade structure is interindustry, and current business cycle synchronization with the United States is low. Thus, to date neither the trade structure nor the degree of business cycle synchronization of Central America appears to make a compelling case for macroeconomic coordination within Central America or between Central America and the United States. I. BUSINESS CYCLE SYNCHRONIZATION IN CENTRAL AMERICA The degree of business cycle synchronization is important because it shows the necessity of independent fiscal and monetary policy. If business cycles are similar and shocks are common, coordination of macro policies can become desirable, with a common currency as the ultimate form of policy coordination. But if shocks are predominately country-specific, independent monetary and fiscal policies are usually seen as important in helping an economy adjust to a new equilibrium. Data and Methodology Because shocks are not directly observed, empirical studies rely on econometric methods for their identification. Bayoumi and Eichengreen (1993) and Helg and others (1995) adopt a structural vector autoregression approach, whereas Artis and Zhang (1995) develop an identification scheme based on cyclical com- ponents. Rubin and Thygesen (1996), Beine and Hecq (1997), and Beine, Candelon, and Hecq (2000) use a codependence framework. Filardo and Gordon (1994), Beine, Candelon, and Sekkat (1999), and Krolzig (2001) use a Markov switching vector autoregression model. This empirical work shows that it is important to distinguish between short- and long-run effects. Bayoumi and Eichengreen (1993), Helg and others (1995), and Rubin and Thygesen (1996) use differenced variables in the vector autoregression representation. However, such a specification does not allow for a long-run relationship among the vari- ables. Beine, Candelon, and Hecq (2000) overcome this by investigating common trends and common cycles simultaneously, where evidence of a common European cycle is taken as evidence of perfect synchronization of shocks. Breitung and Candelon (2001) use a frequency domain common cycle test to analyze synchronization at different business cycle frequencies. The analysis here uses annual data on real GDP and trade figures for 1965- 2002 and monthly data on industrial production and economic activity for 1995-2002. Data on GDP are from the International Monetary Fund's International Financial Statistics database, data on industrial production are from each country's central bank statistics, and data on trade are from the World Bank's World Integrated Trade Solution database and the International Monetary Fund's Direction of Trade Statistics database. 52 THE WORLD BANK ECONOMIC REVIEW The key variable is the degree of business cycle synchronization between countries i and j. Frankel and Rose's (1998) approach is used to measure this variable; the correlation between the cyclical component of the output in countries i and j is computed, with a higher correlation implying a higher degree of business cycle synchronization. The cyclical component of output is obtained using different de-trending methods. Given the lack of consensus on the optimal procedure and the sensitivity of the cycle to the de-trending method, this approach should provide a robustness check of the results. For de-trending, first-differencing and band-pass filtering (Baxter and King 1999) are used for the annual data and spectral analysis for the monthly data. Two aspects of business cycle synchronization are analyzed here. First is the degree of business cycle synchronization, which is measured using simple contemporaneous correlations between the cyclical components of economic activity (at monthly and annual frequencies) across countries. Regression analy- sis is then used to study whether the sensitivity of business cycles to develop- ments in major trading partners has changed over time. Second is the link between trade integration on business cycle synchronization, which is assessed using measures of bilateral trade intensity and trade structure in combination with the measure of the degree of business cycle synchronization. Measuring the Degree of Business Cycle Synchronization ANNUAL DATA: 1965-2005. Band-pass filtered data, the preferred method for business cycle extraction in this section, show that in Central America business cycle synchronization is highest among Costa Rica, El Salvador, Guatemala, and Honduras (table 1). Nicaragua and Panama appear to follow a different cycle, as correlation across business cycles is in most cases negative, though not statistically significant.2 Correlation with the U.S. business cycle is also high. In Costa Rica, El Salvador, and Honduras business cycle synchronization with the United States appears even higher than with regional neighbors, indicating that bilateral relationships with the United States through trade and remittances are more important than regional effects. Somewhat surprising, business cycle synchroni- zation between the United States and Panama, which adopted full dollarization in 1904, appears to be much lower than synchronization between the United States and the rest of Central America, except Nicaragua. On the basis of the business cycle synchronization, the rest of Central America would be better can- didates for a currency union with the United States than Panama would be. In fact, business cycle synchronization between the United States and Costa Rica, El Salvador, Guatemala, and Honduras is higher than the EU average (table 2). 2. Results based on first differences, reported in supplemental appendixes S-1 and S-2, broadly confirm the band-pass filtered results. The supplemental appendixes are available at http://wber. oxfordjournals.org. Fiess 53 TABLE 1. Business Cycle Synchronization, Band-Pass Filter, Central America Country Costa Rica El Salvador Guatemala Honduras Nicaragua Panama Costa Rica 1.000 El Salvador 0.604a 1.000 Guatemala 0.632a 0.238 1.000 Honduras 0.524a 0.442a 0.590a 1.000 Nicaragua -0.214 0.015 -0.142 -0.157 1.000 Panama -0.007 -0.062 -0.087 -0.011 0.088 1.000 Argentina 0.354a 0.111 0.187 0.043 -0.086 0.148 Brazil 0.350a 0.028 0.407a 0.174 -0.162 -0.001 Mexico 0.151 -0.335 0.395a 0.168 -0.255 0.323 Canada 0.621a 0.276 0.492a 0.359a -0.214 -0.336 United States 0.687a 0.506a 0.463a 0.679a -0.163 -0.148 France 0.239 0.113 0.394 0.152 -0.170 -0.138 Germany 0.167 0.107 0.308 0.107 -0.138 0.280 Portugal 0.124 -0.088 0.504a 0.423a -0.127 -0.085 Spain 0.175 0.136 0.389a 0.057 0.167 -0.218 United Kingdom 0.402a 0.479a 0.241 0.459' -0.268 -0.323 Note: Displays bilateral correlations of the cyclical components of band-pass filtered annual GDP data. aSignificant at the 5 percent level. Source: Author's calculations based on data described in the text. Business cycle synchronization between Argentina and Brazil, two Common Market of the South (Mercosur) countries, is lower than among Costa Rica, El Salvador, and Guatemala. While business cycle synchronization is also substantial between Canada and the United States, it is surprisingly low between Mexico and the United States. Appendix table A-1 shows business cycle synchronization among Central American countries after controlling for the common impact of the U.S. business cycle using a two-step procedure. First, the cyclical component of GDP in Central America is regressed on a constant and then on the cyclical component of the United States. Second, the regression residuals are correlated to assess the degree of business cycle synchronization in Central America, which is independent from the U.S. business cycle.3 Once the common impact of the U.S. business cycle is removed, only synchronizations between Costa Rica and El Salvador, Costa Rica and Guatemala, and Guatemala and Honduras are affected by common factors other than the U.S. business cycle. Because these countries also account for the largest share of intraregional trade, this finding supports the often postulated positive relationship between trade intensity and business cycle symmetry. 3. The regression was yj= at + yus + si, where yj is the cyclical component of the band-pass filtered GDP in Central American country i, yus is the component in the United States, and ej is the ordinary least squares regression residual (which is orthogonal to the U.S. cyclical component). The regressions correct for serial correlation and heteroscedasticity. TABLE 2. Business Cycle Synchronization, Other Free Trade Agreements Mercosur NAFTA European Union Country Argentina Brazil Mexico Canada United States France Germany Portugal Spain United Kingdom Costa Rica 0.354' 0.350' 0.151 0.621a 0.687a 0.239 0.167 0.124 0.175 0.402a El Salvador 0.111 0.028 -0.335 0.276 0.506a 0.113 0.107 -0.088 0.136 0.479a 2 O Guatemala 0.187 0.407a 0.395a 0.492a 0.463a 0.394a 0.308 0.540a 0.389a 0.241 Honduras 0.043 0.174 0.168 0.359a 0.679a 0.152 0.107 0.423a 0.957 0.459a F Nicaragua -0.086 -0.162 -0.255 -0.214 -0.163 -0.170 -0.138 -0.127 0.167 -0.268 Panama 0.148 -0.001 0.323 -0.336 -0.148 -0.138 0.280 -0.085 -0.218 -0.323 Argentina 1.000 0.202 0.093 -0.095 -0.033 -0.212 0.273 -0.091 -0.067 -0.100 Brazil 1.000 0.122 0.514' 0.286 0.080 0.070 0.209 0.223 0.320 Mexico 1.000 0.161 0.086 -0.007 0.156 0.159 0.013 -0.290 Canada 1.000 0.771a 0.338a -0.088 0.170 0.370a 0.607a United States 1.000 0.337a 0.104 0.292 0.329 0.727a France 1.000 0.372a 0.656a 0.711a 0.482a Germany 1.000 0.328a 0.348a -0.044 Portugal 1.000 0.559a 0.431a Spain 1.000 0.429a United Kingdom 1.000 Note: Displays bilateral correlations of the cyclical components of band-pass-filtered annual GDP data. aSignificant at the 5 percent level. Source: Author's calculations based on data described in the text. 00ZO~'LXA.rruqoj no punjrjPjuoW Ituop1uulu ju /'Jo-sltunopojxoqm//:d t[ uioij popolumo( Fiess 55 MONTHLY DATA: 1995-2004. The business cycle is usually defined in the range of 6-32 quarters, and thus the low frequency of annual data might be insufficient to fully assess business cycle synchronization. This section complements the analysis of the previous section by using monthly data, where output is proxied by seasonally adjusted monthly indices of industrial production and economic activity. Because the data cover a relatively short time span of less than 10 years, at most two to three business cycles are likely to be captured. Unfortunately, neither monthly nor quarterly data exist on a consistent basis prior to 1995. To make the most of the short time span, spectral analysis is used to estimate the correlation at different frequencies, and the average coherence at business cycle frequency (6-32 quarters) of year-over-year changes in monthly economic activity is used as a summary measure of business cycle synchronization (Anderson, Kwark, and Vahid 1999; Garnier 2004). The advantage of using cross-spectral densities over simple correlations in the analysis of business cycle synchronization is twofold. First, spectral analysis avoids possible business cycle distortions due to filtering; it is well known that the cycles change with the de-trending method (Canova 1998). Second, contemporaneous correlation is unable to take lagged co-movement into account. Because coherence measures the correlation between two series in the frequency domain and provides infor- mation on the phase lead or lag, spectral analysis provides a richer analysis of business cycle dynamics. While the coherence shows to what extent two business cycles are dominated by the same frequency, the phase lag shows to what extent elements with the same frequency lag each other. In sum, a high degree of business cycle synchronization implies a high coherence and a low phase lag. Coherence and phase lag are calculated from the spectral density function.4 To calculate the average coherence at business cycle frequency, the spectral coherence pxy((o) is calculated between series x and y at each frequency, assum- ing that frequencies are independent across and between series: (1) , (0) x (to) PXY - F__(o)Fy (o) Frequencies outside the business cycle ranges are omitted, and the average coherence is then calculated as the average over frequencies within the business cycle band. A high average coherence consequently implies that two series are dominated by the same frequencies within the business cycle frequency bands. 4. The cross-spectral density Fya(w) between series x and y is the Fourier transformation of the cross-covariance function Cxy(T), where -oo < T< 00 is the lag. The cross-spectral density Fxy(() is defined as Exy(s) = 27r e-TCxy(r)dr. 56 THE WORLD BANK ECONOMIC REVIEW The average coherence at business cycle frequency between year-over-year growth rates of economic activity between 1995 and 2004 broadly confirms the findings of the previous section (table 3). Within Central America, business cycle synchronization is found to be the highest between Costa Rica and El Salvador, El Salvador and Guatemala, El Salvador and Nicaragua, and Honduras and Nicaragua. Business cycle synchronization with the United States is the highest for Costa Rica, El Salvador, and Honduras but lower than for NAFTA and Mercosur member countries.s The higher level of business cycle synchronization between Mexico and the United States, as well as between Argentina and Brazil, is explained partly by the long time period (1965-2005) under consideration. Business cycle synchro- nization between Mexico and the United States has increased substantially since the mid-1990s, which Cuevas, Messmacher, and Werner (2002) attribute to increasing integration due to NAFTA.6 Changes in Business Cycle Synchronization over Time Except Costa Rica, Central American countries suffered deep crises prior to the 1990s: from hyperinflation in Nicaragua to civil war in El Salvador. These episodes, which led to low growth and high volatility, clearly marked business cycles, and are likely to have affected the degree of business cycles synchroniza- tion with the United States. A basic regression analysis is used to assess changes in business cycle synchro- nization. Annual growth rates of GDP in Central American countries were regressed against their lagged values and that of U.S. GDP growth. The regressions take the following general form, with coefficient estimates reported in table 4: Ayi = ao + dum90 + j8jAyj,t-j +1 24yi,t-1 x dum90 n + S: 81,kAYIJ,t-k k=0 n (2) + 5 82,kAYIUS,t-k x dum90 k=0 5. The phase lag is not reported here because it is very poorly estimated if the coherence is small, which is the case for most country pairings in table 3. 6. That business cycle synchronization can increase significantly with structural reform has been documented in the case of Mexico and the United States. Cuevas, Messmacher, and Werner (2002) attribute higher business cycle synchronization between Mexico and the United States during the 1990s to increasing integration due to NAFTA. TABLE 3. Average Coherence at Business Cycle Frequency Country Costa Rica El Salvador Guatemala Honduras Nicaragua Argentina Mexico Canada France Costa Rica 0.381 El Salvador 0.524' Guatemala 0.381 0.534' Honduras 0.456 0.340 0.381 Nicaragua 0.393 0.510a 0.421 0.544a Mexico 0.332 0.453 0.242 0.366 0.288 0.537a 0.361 0.345 United States 0.454 0.427 0.336 0.421 0.322 0.486 0.468 0.554a 0.429 Brazil 0.318 0.322 0.382 0.319 0.272 0.500a 0.608a 0.467 0.319 Germany 0.510a 0.536a 0.604a 0.529a 0.248 0.355 0.447 0.584a 0.601 Note: Displays the bilateral average coherences of monthly data of economic activity. aSignificant at the 5 percent level. Source: Author's calculations based on data described in the text. 0lOZ 'L,AIja-qoj uo punjrjPjuoW tuoiluuiolul ju /'Jo-slunopojxo-jqm//:dBt[ uoij poppolumo( 58 THE WORLD BANK ECONOMIC REVIEW TABLE 4. Changes in Business Cycle Synchronization with the United States over Time GDP growth volatility (coefficient of variation) Country 8 82 81+ 82 R2 1965-89 1990-2005 Costa Rica 0.630a 0.808a 1.438 0.36 0.82 0.57 El Salvador -0.153a 0.552a 0.400 0.58 3.47 0.55 Guatemala 0.426' 0.308' 0.734 0.53 0.86 0.32 Honduras 0.790' -0.339 0.451 0.26 0.81 0.76 Nicaragua -0.295 0.842' 0.547 0.56 23.88 0.71 Panama -0.091 1.488' 1.397 0.46 1.50 0.60 United States 0.73 0.49 Mexico 0.432a 1.106a 1.538 0.62 0.83 1.04 [1990-95] 0.65 [1996-05] Note: Coefficients 81 and 82 are from regression 2. Coefficient b1 measures the sensitivity of economic developments in Central American countries to developments in the United States. Relationship (81 + 82) indicates how this sensitivity has changed over time. The last two columns compare volatility, based on the coefficient of variation over two periods: 1965-89 and 1990- 2005. A low coefficient indicates a more tranquil period. Comparative statistics for Mexico and the United States are included. To account for Mexico's Tequila Crisis of 1994-95, the coefficient of variation is calculated for two additional subperiods: 1990-95 and 1996-2005. The regressions for Mexico also include a dummy variable for the Tequila Crisis. aSignificant at the 10 percent level or higher. Source: Author's calculations based on data described in the text. where Ayj is the annual GDP growth rate of Central American country i, Ayus,t is the annual U.S. GDP growth rate, dum90 is a bivariate dummy variable that takes the value of one from 1990 onwards. Several lag structures were explored, but additional lags of the dependent and independent variables proved generally insignificant even though contemporaneous lags are usually highly significant. This simple regression allows two issues to be assessed. First is how sensitive the dependent variable is to developments in the United States (as given by 81). Second is how this sensitivity has changed over time (as given by 81 + 82) and whether the changes are statistically significant. The sensitivity to developments in the United States has generally increased over time and the negative co-movement with the United States for Honduras and Nicaragua (see table 4) appears to vanish in the 1990s, indicating that the end of the civil war most likely lessened the impact of country-specific shocks.7 For Costa Rica, Mexico, and Panama the sensitivity coefficient becomes larger than one, which indicates that GDP in these countries may respond more than proportionally to changes in U.S. output.8 7. There also appears to be a positive link between the size of the sensitivity coefficient and the macroeconomic volatility. The sensitivity coefficient is generally higher during more tranquil periods. 8. Lederman, Maloney, and Serven (2005) report a similar finding for Mexico. Fiess 59 Table 4 also explains an apparent contradiction between the correlation results of the longer, annual sample and the shorter, monthly correlation exer- cise reported in previous sections. Because business cycle synchronization between Mexico and the United States increased significantly during the 1990s, it is not surprising to find substantially higher correlation in the more recent sample. Cuevas, Messmacher, and Werner (2002) attribute higher business cycle synchronization between Mexico and the United States during the 1990s to increased integration due to NAFTA. II. TRADE STRUCTURE AND BUSINESS CYCLE SYNCHRONIZATION The impact of trade liberalization on business cycle synchronization is theoreti- cally ambiguous. Standard trade theory (Heckscher-Ohlin) predicts that removing trade barriers leads to an increasing specialization in production, which leads to interindustry trade patterns. As industry-specific specialization increases, industry-specific shocks-for example, a shock to commodity prices-will make business cycles more dissimilar and hence decrease business cycle synchronization. Experience from developed countries, however, shows a trend toward intraindustry rather than interindustry trade. If intraindustry trade is vertical- for example, particular countries specialize in different production stages of the same good-industry-specific shocks will make business cycles more similar. The same results occur if intraindustry trade is horizontal-for example, countries trade and compete with the same products. In that case industry- specific shocks are also expected to increase business cycle synchronization.9 To summarize, intraindustry trade, vertical or horizontal, is expected to increase business cycle synchronization. Central America's Trade Structure Appendix tables A-2 and A-3 provide information about Central America's trade structure. Trade patterns of NAFTA countries and some EU and Mercosur countries are again provided for comparison. Trade (measured as the ratio of bilateral exports to total exports) in Central America is not predomi- nantly intraregional as it is for EU, Mercosur, and NAFTA members. Even within the so-called Northern Triangle (El Salvador, Guatemala, and Honduras) and between El Salvador and Nicaragua, bilateral exports as a ratio of total exports barely exceed 10 percent. The United States is by far Central America's most important trading partner, although trade with the European Union is somewhat significant. As exports to the United States appear to be under-reported, U.S. imports from Central America are provided as an 9. The authors thank an anonymous referee for noting that horizontal intraindustry trade leaves the field open for asymmetric "taste" shocks to occur, such as shifts away from Fiat cars to BMWs. In this context business cycles would become less similar. 60 THE WORLD BANK ECONOMIC REVIEW alternative measure. On the basis of this measure, exports to the United States account for more than 60 percent of total exports in Costa Rica, El Salvador, and Guatemala. Appendix table A-4 provides information on the importance of intraindustry trade in Central America based on the adjusted Grubel-Lloyd intraindustry trade index, AIIT: n n E(Xi + MA) - E|X - Mil (3) A IIT = n n i i where X is exports of industry i and M is imports. The AIIT, which adjusts for trade imbalances, can take values between zero (no intraindustry trade) and one (all trade is intraindustry). Intraindustry trade appears to be some- what important within Central America. But except Costa Rica (0.3), there is virtually no evidence of intraindustry trade with the United States. For El Salvador and Guatemala intraindustry trade appears to be quite high with Brazil and Mexico. Business Cycle Synchronization and Trade Empirical evidence on trade integration and business cycle synchronization is somewhat mixed. Frankel and Rose (1998), Choe (2001), Calderon, Chong, and Stein (2002), and Calderon (2003), to name a few, find that a higher trade intensity tends to increase business cycle synchronization. Shin and Wang (2003) find that increasing trade itself does not necessarily lead to more syn- chronized business cycles, with evidence for East Asia suggesting that only the expansion of intraindustry trade had such an effect. But Garnier (2004) finds only weak or no relations between intraindustry trade and business cycle synchronization for 16 developed countries and concludes that intraindustry trade at most only partially explains business cycle transmission. The low correlations reported by Calderon, Chong, and Stein (2002) suggest a similar interpretation for trade intensity and business cycle synchronization. A cross-plot of bilateral export to GDP ratios and average coherence at business cycle frequency fails to show a positive relationship between trade intensity and business cycle synchronization (figure 1), a finding in line with other research.10 The slope of the regression line, however, is quite flat because most countries fall into a relatively narrow range of business cycle synchroniza- tion, independent of their level of trade intensity. For example, despite a big difference in trade intensity, France and Mexico have a similar degree of business cycle synchronization with the United States. This seems to support 10. The results are similar if bilateral exports as a share of total exports are used as a measure of trade intensity. Fiess 61 FIGURE 1. Trade Intensity and Business Cycle Synchronization 0.7 0.6- 0.5 0.4 ~0.3 c0.2-* 0.1 y= 0.387 + 0.123x 1?=o0.095 (0.017) (0.062) 0 1 0 10 20 30 40 50 60 70 80 Bilateral exports as a share of GDP (percent) Note: Regression is based on country pairings based on countries listed in appendix table A-4. Numbers in parentheses are standard errors. Source: Author's calculations based on data described in the text. Kenen's (2000) argument that business cycle symmetry is only partly explained by trade intensity. In other words, for El Salvador to reach Mexico's level of business cycle synchronization with the United States-which is only slightly higher in GDP terms-El Salvador would have to more than double its exports to the United States. Figure 2 shows a similar regression for trade intensity and intensity of intraindustry trade. As explained by Shin and Wang (2003) and Garier (2004), the link between intraindustry trade and business synchroniza- tion is found to be stronger and more significant. FIGURE 2. Intraindustry Trade Intensity and Business Cycle Synchronization 0.7 - 0.6 - 0O.5 - . CO0.4 -_ _ _ _ _ _ _ _ _ _ _ _ _ _ 2~ 0,3-~ 0.2- y 0.3741 + 0.185x (0.22) (0.065) 0.1 - R = 0.161 0.0 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Intraindustry trade Note: Regression is based on country pairings based on countries listed in appendix table A-4. Numbers in parentheses are standard errors. Source: Author's calculations based on data described in the text. 62 THE WORLD BANK ECONOMIC REVIEW Some studies argue that that macroeconomic coordination, and in particular exchange rate stability, per se can lead to higher trade and as a consequence more synchronized business cycles. Frankel and Rose (1998, 2002) show that larger trade flows are associated with greater business cycle correlation and argue that increased trade flows can be the result of monetary and economic integration. Fontagn6 and Freudenberg (1999) establish a negative relation between intraindustry trade and exchange rate volatility and draw attention to the fact that monetary integration, by suppressing exchange rate uncertainty, has promoted intraindustry trade in Europe. If trade structure is a good proxy for output structure, business cycles should become more synchronized because cycles will be increasingly affected by the same shocks. Panama dollarized in 1990 and Argentina adopted a currency board that anchored the currency to the U.S. dollar between 1991 and 2001. Both countries fully eliminated exchange rate volatility with respect to the dollar during these periods. Given this high level of monetary integration, Frankel and Rose (1998) and Fontagn6 and Freudenberg (1999) predict an increase in bilateral trade and intraindustry trade with the United States for both countries. But trade as a percentage of GDP as well as intraindustry trade with the United States declined, providing little empirical support that exchange rate stability alone promotes trade (figures 3 and 4). This exercise is far from being conclusive. Nevertheless, this finding might not be too surprising, given that Kenen (2000) and Hughes-Hallet and Piscitelli (1999) question a causal link FIGURE 3. Trade between the United States and Argentina and the United States and Panama, 1997-2001 20 16 14 12 10 -a -Argentina 2 ~ --*-Panama 4 2 0 1997 1998 1999 2000 2001 Source: Bilateral trade data from the U.S. Census Bureau. Fiess 63 FiGURE 4. Grubel-Lloyd Index for Argentina and Panama, 1997-2001 0.30 0.25 0.20 0.15- 0.10 - Pantama -Argentina 0.05- 0.00- 1997 1998 1999 2000 2001 Note: Calculation of the Grubel-Lloyd index uses data at the three-digit level of the Standard International Trade Classification. Source: Intraindustry trade data from the Hamburg Institute of International Economics. between business cycles and trade, when countries are not similar enough. Hughes-Hallet and Piscitelli (1999) demonstrate that a currency union increases business cycle synchronization only after sufficient symmetry exists in institutional structures and market responses across countries. This is likely to be the case for most Latin American countries and the United States (Lederman, Maloney, and Serven 2005). Business Cycle Synchronization and Remittances While trade is often perceived as the most important channel of business cycle synchronization, financial integration is increasingly being recognized as another. Worker remittances provide a growing financial link between Central America and the United States, and they are likely to increase even further in the context of CAFTA; in particular, if provisions are made for temporary or permanent migration of labor. This section assesses the impact of worker remittances on business cycle synchronization. There is little theoretical guidance on how worker remittances are expected to affect business cycle synchronization. Nevertheless, it seems plausible that under certain conditions worker remittances can contribute to synchronization between recipient and sending countries, with the adjustment taking place in the recipient country. For this to be the case, remittances would need to be countercyclical to economic activity in the recipient country, with remittances 64 THE WORLD BANK ECONOMIC REVIEW TABLE 5. Correlation between Remittances and GDP Correlation with Correlation with Remittances as share of Country own GDP U.S. GDP GDP in 2004 (percent) Argentina -0.270 (0.42) 0.460 (0.15) 0.2 Brazil -0.376 (0.08) 0.103 (0.65) 0.4 Mexico -0.650 (0.02)b 0.341 (0.11) 2.5 Costa Rica' 1.6 El Salvador -0.147 (0.47) 0.284 (0.17) 16.1 Guatemala -0.656 (0.01)b -0.178 (0.52) 9.3 Honduras -0.230 (0.54) 0.210 (0.44) 15.4 Nicaragua -0.503 (0.12) 0.376 (0.25) 11.4 Panama -0.412 (0.05)b 0.112 (0.62) 0.9 Note: Numbers in parentheses are p-values. aCorrelations are not reported because too few observations are available. bSignificant at the 5 percent level. Source: Author's calculations based on data described in the text. increasing during times of crisis, and pro-cyclical with economic activity in the sending country, with a growing economy providing a larger outflow of remit- tances. This would create forces to pull the business cycle of the receiving country toward the business cycle of the sending country. But if remittances are uncorrelated with the economic activity in the sending country and pro- cyclical with economic activity of the recipient country, business cycles could become more dissimilar. Simple correlations between changes in remittance flows and GDP growth provide weak evidence that remittances are countercyclical (table 5). Correlations are significant only in Guatemala, Mexico, and Panama. There is also some indication that remittance flows are positively correlated with econ- omic growth in the United States but not at a statistically significant level. The observed correlation patterns suggests that except Guatemala, Mexico, and Panama, worker remittances do not appear to significantly smooth asymmetric shocks. A general lack of significant correlation with economic activity in the United States further suggests that remittances do not contribute in a major way to the synchronization of business cycles between Central America and the United States. III. SUMMARY AND CONCLUDING REMARKS This article offers the following findings: * Business cycle synchronization within Central America is quite low com- pared with synchronization in NAFTA and the European Union, but not compared with synchronization in Mercosur. Fiess 65 * Business cycle synchronization in Central America is highest between Costa Rica and El Salvador, El Salvador and Guatemala, El Salvador and Nicaragua, and Honduras and Nicaragua. * Costa Rica and Honduras have a higher degree of business cycle synchroni- zation with the United States than with any other Central American country. However, business cycle synchronization with the United States is still below the levels of business cycle synchronization among NAFTA and Mercosur members. * Central American countries have become more sensitive to developments in the U.S. economy. * Unlike trade in NAFTA, the European Union, and Mercosur, trade in Central America is not predominantly intraregional. The United States is by far Central America's most important trading partner. * Except for Costa Rica, there is virtually no evidence of intraindustry trade between Central America and the United States. The level of intraindustry trade within Central America is comparable to that of Mercosur, but less than in NAFTA (Canada and the United States) and the European Union (France and Germany). * The degree of business cycle synchronization seems only weakly related to trade intensity and trade structure (intraindustry trade), although the relationship between intraindustry trade and business cycle synchronization is slightly stronger. As such, the gain in business cycle synchronization through trade expansion seems quite low. * Macroeconomic coordination per se is unlikely to promote business cycle synchronization or trade because institutional structures and market responses in Central America and the United States lack sufficient symmetry. * Remittances provide a growing financial link between Central America and the United States. But there is little evidence that remittances lessen the impact of asymmetric shocks or contribute in a major way to the synchronization of business cycles between Central America and the United States. Neither Central America's trade structure nor its degree of business cycle synchronization makes a compelling case for macroeconomic coordination within Central America or between Central America and the United States. Central America's trade structure is predominately interindustry, and the current level of business cycle synchronization with the United States is not that high, despite an increase since the mid-1990s. Clearly, trade integration is a dynamic process, and as trade intensities and compositions of trade flows change so will business cycle patterns. To fully assess the consequences of closer trade integration for the conduct of 66 THE WORLD BANK ECONOMIC REVIEW macroeconomic policies, information about the future evolution of trade struc- tures in DR-CAFTA are needed. If trade becomes more intraindustry (vertical or horizontal), business cycles are expected to become more similar, and inde- pendence of macro policy will be less of a concern. However, if trade inte- gration takes the form of higher interindustry trade, business cycles are likely to diverge from current levels, and the ability to conduct independent macro policies will grow more important. While information about the future developments of trade patterns within DR-CAFTA is not available, Mexico's experience in NAFTA might provide some guidance. Trade between Mexico and the United States has grown exponentially since the signing of NAFTA-from $89.5 billion in 1993 to $275.3 billion in 2004. The United States has become not only Mexico's top trading partner but also its main investor. Since 1994 the United States has accounted for 62 percent of all foreign direct investment in Mexico. But the two economies are increasingly linked not only through trade and invest- ment but also through worker remittances. In 2005 worker remittances from the United States accounted for 3 percent of Mexico's GDP. Closer econ- omic integration through NAFTA has had a clear impact on business cycle synchronization. Cafias, Coronado, and Gilmer (2006) find that based on the coincidence indexes for economic activity for both countries the degree of business cycle synchronization since 1993 is about a third higher than in 1980-93. Since the signing of NAFTA there has also been a consistent upward trend in intraindustry trade between Mexico and the United States. According to Bruehlhart and Thorpe (2001), between 1980 and 1998 the unadjusted Grubel-Lloyd index for manufacturing products between Mexico and the United States grew from 0.36 to 0.61.11 Mexico's dramatic shift in intraindus- try trade with the United States is explained mostly by increased vertical intraindustry trade in textiles and apparel and in auto industries (Burfisher, Robinson, and Thierfelder 2001). The increase in vertical intraindustry trade has been accompanied by higher business cycle synchronization. Cuveas, Messmacher, and Werner (2002) claim that macroeconomic synchronization between Mexico and the United States has increased substantially due to NAFTA. Despite the higher level of business cycle synchronization between Mexico and the United States, Cuevas, Messmacher, and Werner (2002) and Lederman, Maloney, and Serven (2005) do not advocate adopting common stabilization policies in NAFTA. Most of their arguments transfer directly to DR-CAFTA. Despite increased sensitivity to the U.S. economy, idiosyncratic shocks continue to be important for Mexico, and idiosyncratic volatility remains higher in Mexico than in the United States. Lederman, Maloney, 11. For products at the three-digit level of the Standard International Trade Classification. At the same time, intraindustry trade with Canada remained at a relatively constant low level of 0.17. Fiess 67 and Serven (2005) argue that nominal price and wage flexibility are lacking in Mexico, and NAFTA does not provide unrestricted labor mobility or mechanisms of fiscal redistribution to facilitate Mexico's adjustment to shocks in the absence of independent stabilization policies. A similar case can be made for Central America because idiosyncratic volatility is also higher and DR-CAFTA, like NAFTA, does not come with any built-in shock absorbers. Further, that the Mexican economy responds more than proportionally to shocks in the United States indicates that Mexico would require a higher dosage for the treatment of the same shock. A common policy response would not be able to effectively counteract output and employment fluctuations in Mexico. The picture is even more complex for Central America, where the same shock would require a larger policy response for Costa Rica and Panama, but a smaller dosage for the remaining countries. Finally, policy transmission channels are different and require the ability to apply stabilization policies in different quantities. Lederman, Maloney, and Serven (2005) argue that Mexico's lower level of financial development and domestic credit to the private sector implies that interest and credit channels are less developed relative to United States, while exchange rate channels are more important for Mexico because trade accounts for a larger share of GDP. Central America appears an even greater mismatch in this respect; it lags far behind Mexico in terms of financial sector development but leads Mexico in terms of openness. APPENDIX TABLE A-1. Business Cycle Synchronization, Orthogonal to U.S. Business Cycle Country Costa Rica El Salvador Guatemala Honduras Nicaragua Costa Rica 1.000 El Salvador 0.409' 1.000 Guatemala 0.488' 0.006 1.000 Honduras 0.104 0.157 0.421' 1.000 Nicaragua -0.141 0.115 -0.076 -0.063 1.000 Panama 0.134 0.014 -0.021 0.118 0.065 Note: Displays bilateral correlations of the cyclical components of band-pass filtered annual GDP data orthogonal to the U.S. business cycle. aSignificant at the 5 percent level. Source: Author's calculations based on data described in the text. 00 0 TABLE A-2. Central America's Trade Structure: Bilateral Exports as a Share of Total Exports, 1995-2001 (percent) Country Costa Rica El Salvador Guatemala Honduras Nicaragua Argentina Mexico Canada France Costa Rica 4.4 3.5 1.1 4.8 0.1 0.2 0.0 0.0 El Salvador 2.3 9.9 3.1 11.1 0.1 0.2 0.0 0.0 Guatemala 3.2 12.4 2.5 2.8 0.1 0.4 0.0 0.0 a Honduras 1.7 6.8 2.0 5.3 0.0 0.1 0.0 0.0 Nicaragua 2.9 3.8 3.1 2.2 0.0 0.1 0.0 0.0 Mexico 1.1 0.7 2.3 0.3 2.8 1.2 0.5 0.4 Brazil 0.1 0.0 0.0 0.0 0.0 26.9 0.5 0.4 0.7 United States 21.3 11.1 50.7 61.1 38.0 9.4 87.1 85.3 7.3 Germany 3.6 6.1 3.3 3.8 9.9 2.3 0.9 0.9 15.7 European Union 16.0 10.7 10.4 12.2 23.1 18.5 3.6 4.9 61.6 Free trade zone 39.1 54.5 U.S. reported imports c.i.f. 62.4 68.1 66.3 Note: Data are averages for 1995-2001. The table should be read column-wise, where each row represents the share in total column-countries exports. As an example, the top-left figure indicates that exports from Costa Rica to El Salvador represent 2.3 per cent of Costa Rica's total exports. Source: International Monetary Fund's Direction of Trade Statistics. 0lOZ 'LXA.r-aqoj uo p-njrjPjuoW Ituoiluujolul ju /'Jo-slm.inoFpJojxo-jaqm//:dBlt umoij popolumo( TABLE A-3. Central America's Trade Structure: Bilateral Exports as a Share of GDP, 1995-2001 (percent) Country Costa Rica El Salvador Guatemala Honduras Nicaragua Argentina Mexico Canada France Costa Rica 0.8 0.7 0.6 1.2 0.01 0.05 0.01 0.01 El Salvador 0.8 1.8 1.5 2.9 0.01 0.05 0.00 0.01 Guatemala 1.1 2.3 1.2 0.7 0.01 0.11 0.01 0.00 Honduras 0.6 1.3 0.4 1.4 0.00 0.03 0.00 0.00 Nicaragua 1.0 0.7 0.6 1.1 0.00 0.01 0.00 0.00 Mexico 0.4 0.1 0.4 0.2 0.7 0.1 0.2 0.1 Brazil 0.0 0.0 0.0 0.0 0.0 2.4 0.1 0.13 0.1 United States 7.1 2.1 9.5 30.1 9.8 0.8 24.1 30.3 1.6 Germany 1.2 1.1 0.6 1.9 2.6 0.2 0.3 0.3 3.3 European Union 5.3 2.0 1.9 6.0 5.9 1.6 1.0 1.7 13.2 Free trade zone 13.0 10.1 U.S. reported imports c.i.f. 19.4 11.8 11.7 Note: Data are averages for 1995-2001. Interpretation of this table is as follows. The table should be read column-wise, where each row represents the share of bilateral exports in the column-countries GDP. As an example, the top-left figure indicates that exports from Costa Rica to El Salvador represent 0.8 per cent of Costa Rica's GDP. Source: International Monetary Fund's Direction of Trade Statistics and International Financial Statistics database. 0lOZ 'LXjr-aqoj uo punjrjPjuoW ItuoiluIuju ju /'Jo-sltunoFpJojxo-joqm//:dBlt umoij poppolumo( o TABLE A-4. Intraindustry Trade, 2001 Country Costa Rica El Salvador Guatemala Honduras Nicaragua Argentina Mexico Canada France El Salvador 0.36 Guatemala 0.38 0.45 Honduras 0.40 0.27 0.33 Nicaragua 0.34 0.15 0.21 0.15 Mexico 0.18 0.43 0.42 0.11 0.02 0.26 0.49 0.57 Brazil 0.30 0.05 0.05 0.06 0.02 0.10 0.46 0.66 0.56 United States 0.08 0.43 0.51 0.03 0.28 0.39 0.51 0.17 0.11 Germany 0.06 0.02 0.01 0.13 0.79 0.33 0.70 Note: A five-digit level of disaggregation is used. Source: Author's calculations based on trade data from the UN Commodity Trade Statistics (Comtrade) database. 0lOZ 'LXjr-aqoj uo punjrjejouoIt uoil1uuu ju /'Jo-slunopojxo-jqm//:lth uioij popolumo(I Fiess 71 REFERENCES Artis, M., and W. Zhang. 1995. "International Business Cycles and the ERM: Is There a European Business Cycle." Oxford Economic Papers 51(1):120-32. Anderson, H.M., N. Kwark, and F. Vahid. 1999. "Does International Trade Synchronize Business Cycles?" Working Paper 8/99. Caulfield East, Austrialia: Monash University, Faculty of Business and Economics, Department of Econometrics and Business Statistics. Bayoumi, T., and B. Eichengreen. 1993. "Shocking Aspects of European Monetary Unification." In F. Torres and G. Giavazzi eds., Adjustment and Growth in the European Monetary Union. London: Cambridge University Press. Beine, M., and A. Hecq. 1997. "Asymmetric Shocks inside Future EMU." Journal of Economic Integration 12(2):131-40. Breitung, J., and B. Candelon. 2001. "Is There a Common European Business Cycle? New Insights from a Frequency Domain Analysis." Vierteljahrshaefte zur Wirtschaftsforschung 70(3):331-38. Bruehlhart, M., and M. Thorpe. 2001. "Export Growth of NAFTA Members, Intra-Industry Trade and Adjustment." Global Business and Economics Review 3(1):94-110. Baxter, M., and R.G. King. 1999. "Measuring Business Cycles: Approximate Band-Pass Filters for Economic Time Series." Review of Economic and Statistics 81(14):575-93. Beine, M., B. Candelon, and A. Hecq. 2000. "Assessing a Perfect European Optimal Currency Area: A Common Cycle Approach." Empirica 27:115-32. Beine, M., B. Candelon, and K. Sekkat. 1999. "Stabilization Policy and the Business Cycle Phases in Europe: A Markov Switching VAR Analysis." SFB 373 Working Paper 91/1999. Berlin: Humboldt-Universitit. Burfisher, M.E., S. Robinson, and K. Thierfelder. 2001. "The Impact of NAFTA on the United States." Journal of Economic Perspectives 15(1):125-44. Calderon, C. 2003. "Do Free Trade Agreements Enhance the Transmission of Shocks Across Countries?" Working Paper 213. Central Bank of Chile, Santiago. Canova, F. 1998. "Detrending and Business Cycle Facts." Journal of Monetary Economics 41(3):475- 512. Choe, J. 2001. "An Impact of Economic Integration through Trade: On Business cycles for 10 East Asian Countries." Journal of Asian Economics 12(4):569-86. Calderon, C., A. Chong, and E. Stein. 2002. "Trade Intensity and Business Cycle Synchronization: Are Developing Countries Any Different?" Working Paper 195. Central Bank of Chile, Santiago. Cafias, J., R. Coronado, and R.W. Gilmer. 2006. "U.S., Mexico Deepen Economic Ties." Southwest Economy 1. Federal Reserve Bank of Dallas, Texas. Cuevas, A., M. Messmacher, and A. Werner. 2002. "Macroeconomic Synchronization between Mexico and Its NAFTA partners." World Bank, Washington, D.C. Filardo, A.J., and S.F. Gordon. 1994. "International Co-Movements of Business Cycles." Working Paper 94-11. Federal Reserve Bank of Kansas, MO. Fontagne, L., and M. Freudenberg. 1999. "Endogenous Symmetry of Shocks in a Monetary Union." Open Economics Review 10(3):263-87. Frankel, J.A., and K.E. Rockett. 1988. "International Policy Coordination When Policy Makers Do Not Agree on the True Model." American Economic Review 78(3):318-40. Frankel, J.A., and A. Rose. 1998. "The Endogeneity of the Optimum Currency Area Criteria." Economic Journal 108(449):1009-25. -. 2002. "An Estimate of the Effect of Currency Unions on Trade and Growth." Quarterly Journal of Economics 117(2):437-66. Garnier, J. 2004. "To What Extent Does Intraindustry Trade Matter in Business Cycle Comovements? Distinguishing Common and Transmitted Cycles." European University Institute, Florence, Italy. 72 THE WORLD BANK ECONOMIC REVIEW Hughes-Hallet, A., and L. Piscitelli. 1999. "EMU in Reality: The Effect of a Common Monetary Policy on Economies with Different Transmission Mechanisms." CEPR Discussion Paper 2068. Centre for Economic Policy Research, London. Helg, R., P. Manassa, T. Monacelli, and R. Rovelli. 1995. "How Much (A)symmetry in Europe? Evidence from Industrial Sectors." European Economic Review 39(5):1017-41. Kenen, P. 2000. "Currency Areas, Policy Domains and the Institutionalization of fixed exchange rates." CEP Discussion Paper 467. London School of Economics, Center for Economic Performance, London. Krolzig, H. M. 2001. "Business Cycle Measurement in the Presence of Structural Change: International Evidence." International Journal of Forecasting 17(3):349-68. Lederman, D, W. Maloney, and L. Serven. 2005. Lessons from NAFTA for Latin America and the Caribbean. Palo Alto, Calif.: Stanford University Press. Rubin, J., and N. Thygesen. 1996. "Monetary Union and the Outsiders: A Cointegration-Codependence Analysis of Business Cycles in Europe." Economie AppliquWe 49(3):123-71. Shin, K., and Y. Wang. 2003. "Trade Integration and Business Cycle Synchronization in East Asia." Asian Economic Papers 2(3):1-20. Protecting the Vulnerable: the Tradeoff between Risk Reduction and Public Insurance Shantayanan Devarajan and William Jack In a risky world should governments provide public goods that reduce risk or compen- sate the victims of bad outcomes through social insurance? This article examines a basic question in designing social protection policies: how should a government allo- cate a fixed budget between these two activities? In the presence of income and risk heterogeneities a simple public insurance scheme that pays a fixed benefit to all house- holds that suffer a negative shock is an effective redistributional instrument of public policy. This is true even when a well functioning private insurance market exists, and so the role of public insurance is not to correct a market failure. In fact, the existence of a private insurance market means that the public system has desirable targeting properties-all but the poor and high-risk take up private insurance. The provision of public goods that reduce risk for all should therefore be complemented with public insurance that (automatically) benefits those who are especially vulnerable. JEL Codes: H41, H42, 138. When unanticipated disasters hit individuals, businesses, and communities, gov- ernments are often expected to respond. Governments provided much of the relief after the devastating Indian Ocean tsunami of 2004 and the 2005 earth- quake in Pakistan, and they continue to do so. Similarly, the U.S. government was the main source of compensation for victims of the terrorist attacks of September 11, 2001. Governments are likewise called upon to distribute food aid in the event of drought, and they are expected to provide emergency medical care in response to disease outbreaks. All these post-event compensatory actions can be thought of as publicly provided insurance-public transfers to individuals in the event of bad luck-that spread risk across the population. But governments can also affect the chances that individuals suffer direct nega- tive shocks. For example, early warning systems for tsunamis and drought can reduce the negative shock associated with bad events. Similarly, dams prevent and control flooding, and mosquito spraying can lower the risk of malaria. Shantayanan Devarajan is the chief economist of the World Bank's South Asia Region and editor of World Bank Research Observer; his email address is sdevarajan@worldbank.org. William Jack (corresponding author) is an associate professor in the Department of Economics at Georgetown University; his email address is wgj@georgetown.edu. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 73-91 doi:10.1093/wber/1hl007 Advance Access Publication 24 January 2007 () The Authors 2007. 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 73 74 THE WORLD BANK ECONOMIC REVIEW How should spending on these two types of risk-management activities- public insurance and risk-reducing expenditures-be balanced? What is the tra- deoff between a dam and public flood insurancel or between mosquito eradica- tion and antimalarial drugs?2 These policy choices are important in the context of reducing poverty. There is increasing recognition that poverty is not only a situation of low income, but also one of vulnerability to severe income shocks, such as loss of work, ill-health, and the like. For example, in Indonesia (before the East Asian crisis) about twice as many people were vulnerable to being poor as were poor (defined as having an income in the 20th percentile).3' 4 Although vulnerability has been widely studied and documented, much less has been written about what governments should do about it.s For this reason this article accounts for individual heterogeneity in both risk and income and for the distributive effects of government spending. This article develops a model in which the government provides a public good that reduces the probability of a negative shock. There is no private provision of public goods, so no crowding-out of precautionary actions occurs. But publicly provided insurance that pays a fixed amount to all victims is allowed to affect the risk-sharing behavior of individuals and communities, as it is assumed that an efficient private insurance market exists. Individuals then have the option of availing themselves of the free but possibly incomplete coverage under the public scheme or opting out and purchasing private insurance. To be sure, governments sometimes provide insurance because private insurance markets are inefficient.6 1. Bangladesh has considered spending large amounts of money to build or reinforce dams along three rivers that frequently flood. An alternative use of these funds would be to provide flood insurance to compensate flood victims. 2. This kind of choice is also evident at the aggregate, or macroeconomic, level. Sound macroeconomic policies, with the associated political costs of fiscal restraint, can smooth fluctuations in incomes, while countercyclical transfers protect individuals who suffer during downturns. 3. In particular, 30-50 percent of the population had a 50 percent chance of having their income fall below that of the 20th percentile in the following year (Pritchett, Sumarto, and Suryahadi 2000). 4. The definition of vulnerability is itself the subject of debate. World Bank (2000) introduces dynamic and longer term concepts such as reslience, and Ligon and Schechter (2003) and Kamanou and Morduch (2003) develop more welfare economic concepts. This article does not contribute to this debate: vulnerability is measured here purely by a probability, that is, the chance of suffering a negative shock. 5. Albarran and Attanasio (2003), Dercon and Krishnan (2003), and Barrett, Holden, and Clay (2003) examine the effects of particular policy interventions on risk exposure. Morduch's (1999) review of safety net programs briefly compares public policies that reduce risk with means-tested transfers, unemployment benefits, health insurance, and social security. He argues, however, that risk-reducing expenditures are "policy areas that are on the table for other reasons, and are best judged by other criteria". This suggests that risk-reducing expenditures are best dealt with outside the design of public insurance schemes. By contrast, this article explicitly examines the tradeoff between these two policy instruments. 6. It is well understood that if insurance markets are missing, public safety nets can enhance efficiency, not just because of the reduced uncertainty of income, but also because they can increase average incomes, due to changes in production techniques, for example, Jalan and Ravallion (2003). Banerjee (2003) argues that lack of insurance leads to similar dynamic inefficiencies and poverty traps. Devarajan and Jack 75 Allowing for efficient private insurance, it can be demonstrated that there is a redistributive case for public insurance beyond correcting insurance market failure: to improve the welfare of individuals who are high risk, poor, or both. Income and risk heterogeneity create the possibility of a redistributive role for government, although they themselves do not support public provision of insurance. Two superior instruments are a lump-sum transfer from low-risk households to high-risk households and a progressive income tax. It is assumed here that both instruments are unavailable. Taxes and transfers based on ex ante risk characteristics are very difficult to implement in even the most capacity-rich countries,7 and sophisticated and well functioning income tax systems are particularly rare in developing countries (Thirsk 1998). In the absence of these instruments, publicly provided insurance can serve a redistri- butive role. Such a motivation underlies much of the increased interest in devel- oping so-called "social protection" policies as part of broad poverty reduction strategies in developing countries.8 How can it be assumed that the government can operate a public insurance system but not a redistributive tax-and-transfer system? First, suppose there is an optimal nonlinear income tax, based on earned income. Because of standard asymmetric information problems (Mirrlees 1971), this tax system will not fully redistribute income across the population; there will be some (typically much) residual heterogeneity in after-tax incomes due to the distortionary effects of taxes on taxable income. The public insurance system described here is an additional redistributive instrument that can be grafted onto the tax system to improve welfare. The exogenous distribution of income assumed here can thus be interpreted as the distribution of after-tax income in a more com- plete model. In this model public insurance pays a fixed amount to anyone in the system who suffers a negative shock. This amount complements transfers inherent in the income tax system because the flat insurance benefit is a function solely of the realized state of nature rather than of an individual's realized income. This design feature is the second reason that governments can operate a public insurance system even if their tax systems are rudimentary: it is much easier to verify that an individual has suffered a shock than to estimate the value of the loss.9 This feature in turn has two effects that are absent from an income tax: transfers are made to unlucky individuals (independent of their actual income), and this kind of insurance is more valuable to individuals who expect to have especially low incomes in the event of a negative shock or who expect to suffer such a shock more often. Thus the public insurance scheme represents a useful 7. See, for example, the literature on risk adjustment in health insurance: Glazer and McGuire 2000, 2002; Newhouse 2002. 8. See, for example, the chapter on social protection in World Bank (2002), available at www. worldbank.org/poverty/strategies/chapters/socprot/socprot.htm. 9. After the Indian Ocean tsunami the Sri Lankan government compensated individuals who had lost their house, but the value of their house was not accurately assessed in calculating this payment. 76 THE WORLD BANK ECONOMIC REVIEW additional targeting instrument: resources are directed to individuals with either low (expected) incomes, high risk, or both. An important feature of this model is that individuals can choose to partici- pate in the public insurance system or to purchase insurance on the private market, but not both. This rules out individuals using the public system but taking out complementary insurance to pay for uncovered losses. With this feature the public insurance scheme naturally targets the poor, who choose to opt into the system. Without it, everyone would use the (free) public system, and its targeting properties would be reduced. It would, however, still possess some targeting features, directing resources (on average) to individuals with a higher probability of suffering negative shocks, for example.10 This self-targeting property is why governments might find it desirable to prohibit complementary insurance. Private insurance companies may, conver- sely, want to insure only individuals who are not covered by the public system for moral hazard reasons. Although, for simplicity, it is assumed that the private market works under conditions of complete information, in practice most insurance policies provide incomplete coverage so as to maintain individ- uals' incentives to take precautions against negative shocks (engaging in active job search, exercising and eating well, and so on). If public insurance increases an individual's coverage, these incentives will be weakened. Similarly, moral hazard provides an additional reason that the government may prohibit comp- lementary private coverage.n It is necessary to be explicit regarding the government's institutional capacity to implement an insurance system with the targeting properties described here. First, the government must be able to determine whether an individual has experienced a negative shock; second, it must be able to exclude privately insured individuals from using the public system. Both of these may require a degree of administrative capacity beyond what already exists, especially in poor countries. In practice governments would be expected to base insurance benefits on more aggregate measures of shocks, including, say, local rainfall levels (as is currently under consideration in Ethiopia) and areas of disease outbreak, which are easily measured. Regarding the ability to exclude privately insured individ- uals, the incentives identified above for both public and private providers are strong enough for them to do so in practice, if only imperfectly. Finally, the gov- ernment may have other instruments for redistributing income, including an income tax and public spending targeted to the poor. These policy instruments are not excluded, but the possibility is noted that public insurance could be an additional instrument to enhance social welfare. 10. It would not, by contrast, direct resources to individuals with low incomes in the good state and even lower incomes in the bad state. 11. In some developed countries complementary insurance (also referred to as "gap" insurance) is permitted-for example, assurance complementaire in France-but in most it is prohibited. Devarajan and Jack 77 The tradeoff between public risk reduction and public insurance depends of course on the effectiveness of the public good in reducing risk.12 But the optimal allocation of public spending between the two depends on flows into and out of the public insurance scheme as budget allocations are altered. Increasing the generosity of public insurance directly benefits the poor and vulnerable, but also causes more individuals to opt into the scheme, thereby putting upward pressure on insurance expenditures. Investment in risk reduction has the opposite effect on the pool of publicly insured: it lowers the cost of private insurance and therefore induces relatively low-risk and relatively high-income individuals to leave the public system. Thus public good provision affects the targeting properties of the public insurance system. The analysis here focuses on the interaction of these effects, holding constant the technical efficiency of the public good in reducing risk. This article illustrates the complexity of the public policy tradeoff between prevention and cure. As a result of this complexity, robust but nontrivial con- clusions about optimal public spending allocations are not easily forthcoming. Simulation techniques are used to gain further insight into an issue that is relevant to the debate on the multidimensional nature of poverty. In particular, if being income poor is associated with exposure to greater risk, should public spending be focused more toward one instrument than the other? The simulation results show however that the optimal allocation of the budget between the public good and insurance is nearly independent of the correlation between income and risk. Risk makes the poor poorer, but it does not significantly affect the government's appropriate antipoverty program, at least in the framework considered here.13 I. SETUP OF THE MODEL There is a continuum of individuals in the economy, each of whom is endowed with certain income and risk characteristics. An individual's "income type," y E [Ymin, Ymax], is exogenous, meaning that there is no labor supply decision. There are two states of the world: good and bad. In the good state an individual earns income y; in the bad state he or she earns ay, where a E (0, 1) is fixed and the same for all individuals, meaning that they cannot influence the size of a loss in the bad state (there is no hidden information moral hazard). In the absence of public good provision the good state occurs with an underlying probability p E [0, 1], which is exogenous to each individual, meaning that he or she cannot affect this probability (there is no hidden action moral hazard), 12. Earlier literature (for example, Schlesinger and Venezian 1986) examined the incentive for a profit-maximizing monopoly insurer to invest in risk reduction. The straightforward assumptions are that monopoly profit is a concave function of the probability that an insured individual suffers a loss and that a private insurer might want to alter the probability to maximize profits, net of the costs of manipulating individuals' exposure to risk. Even in a model with no public insurance the optimal level of public expenditure to reduce risks would likely be positive, for similar reasons. 13. It may of course mean that the overall budget should be increased. 78 THE WORLD BANK ECONOMIC REVIEW but varies across individuals. The strong assumption is made that shocks are idiosyncratic (discussed in more detail in section III when specifying the govern- ment's budget constraint). Thus each individual in the economy is indexed by a pair (y, p). Individuals are distributed over the set f C [ynin, Ymax] X [0, 1] with suitably differentiable density function 4(.,-). All individuals have the same von Neumann-Morgenstern utility index u(.), defined over income. The focus is the allocation of a fixed government budget R, which can be spent on a pure public good G and on state-contingent transfers (insurance). G decreases the probability of the bad state occurring (it increases p). Thus let ir (G, p) be the probability that a p-individual faces the good state given G, where 7G > 0, 7p > 0, and 7T E [0, 1]. The effect of G on 7r is assumed to be independent of an individual's income.14 An alternative public expenditure is state-contingent transfers or services at a uniform rate m per capita. For example, if the risk is health-related, m could be the level of medical care available to an individual contingent on the person being sick. Alternatively, m could be a flat dollar amount paid to workers who become unemployed or otherwise lose their livelihood. Individuals can purchase insurance at actuarially fair prices in a private market. It is assumed that there are no administrative costs associated with insurance. Thus, individuals with higher incomes (and hence higher losses in the bad state) tend to purchase more insurance in the private market than those with lower incomes, and individuals with less risk pay lower premiums. For reasons outlined in the introduction, an individual with private insurance is not permitted to use the public system-that is, by purchasing private cover- age individuals effectively opt out of the public system."s The decision to do so is of course endogenous and depends on the government's choice of policy instruments m and G. II. PARTICIPATION IN THE PUBLIC SYSTEM Let S(y,7; m) be defined as the net surplus an individual earns from purchasing private insurance instead of enrolling in the public insurance system when income in the good state is y and the probability of being therein is 7. Clearly Sm(y,7, m) < 0: the more generous the public scheme, the greater an individ- ual's expected utility of enrolling. An individual whose income in the good 14. The sign of the cross derivative, iGp, is not specified at this stage, but as p -+ 1, it is necessary that iTG - 0. There is little effect of the public good on individuals who already have virtually no chance of being in the bad state. Thus the public good naturally favors those who are more vulnerable (have lower p) but public insurance does as well. 15. In addition to the reasons mentioned in the introduction, some kinds of public insurance are likely to be provided in kind rather than in cash, making individuals not want to double dip. For example, it may be difficult to use both public hospital services and private medical care for the treatment of a given condition. Devarajan and Jack 79 state is less than m/(1 - a) will definitely choose the public system. In this case his or her income in the bad state is ay + m, which is higher than in the good state, providing expected utility greater (albeit with some risk) than could be obtained with full private insurance, that is, S < 0. Public insurance overinsures the very poor. Conversely, the very rich will definitely purchase private insur- ance. To see this, consider a very large y (and hence ay): public insurance yields a pair of incomes, (y, ay + m), which is close to (y, ay) and so delivers virtually no improvement in expected utility. Private insurance by contrast yields a first-order increase in expected utility, so S > 0. On the basis of these limiting properties, the net surplus earned from private insurance is assumed to be increasing in y over the whole range of incomes. That is, SY(Y,T, M) > 0.16 When 7r= 1, the net surplus from private insurance is of course zero; neither public nor private insurance increases the individual's expected utility. But when 7 = 0, the net surplus is unambiguously negative; private insurance does nothing for the individual, but the public system guarantees a transfer of m. S is either always negative (except at 7T = 1, when S = 0), in which case all individuals choose the public system, or first negative, then positive, and then zero at 7r = 1. In this case, there is a value r (y; m) such that S i 0 as 7T r. Because it is assumed that S > 0, & is decreasing in y, whereas it is increasing in m (figure 1). This behavior of &- with respect to y and m allows the decomposition of the population into those who opt into the public system and those who opt out (figure 2).18 Assuming a given level of G, and hence a given mapping from p into 7T, the set of individuals who join the public system is denoted by P, its complement by P', and the set of those who are indifferent between the public and private systems by 0P. It is convenient in the next section to describe the boundary of the participation set, OP, as a function )(p; m, G). An increase in the generosity of the public system (an increase in m) shifts the boundary OP to the right, increasing the share of the population publicly insured. But an increase in public good spending G, holding m fixed, shifts OP 16. A sufficient, but not necessary, condition for this is that u'(y) < 0. 17. The net surplus function can be written S(y, ; m) = u([I + a(1 - I)]y) - [-u(y) + (1 - i)u(ay + m)] V(Y,ir) - -(y,IT; M) Note that &(-) s linear in iT, whereas v(.) is concave. Also, v(y, 0) < w(y, 0; m), and v(y, 1) = &(y, 1; m), so that either S(y, w; m) < 0 for all iT and all individuals opt into the public system or only those for whom iT is high enough do. 18. Figure 2 is drawn assuming that individuals with very low risk (p near 1) but low incomes (y near ymin) opt into the public system and that those with high incomes (up to ymax) but high risk (p near 0) also opt in. This need not be the case-that is, the line OP may intersect one or both of the axes-but it has no substantive impact on the analysis. 80 THE WORLD BANK ECONOMIC REVIEW FIGURE 1. Net Surplus Earned from Purchasing Private Insurance Instead of Participating in the Public System S S S(y, ir m) S(y, r m) Y U y um1up (i) (ii) Note: y is an individual's income in the good state, ir is the probability of an individual being in the good state, and m is the benefit paid by the government insurance scheme in the bad state. FIGURE 2. Individuals' Decision Whether to Opt in to the Public Insurance System y Ymax Ymin 1 p Note: P is the set of individuals who join the public insurance system, P' is the set of individuals who join the private insurance system, and OP are those individuals who are indifferent between the public and private systems. Devarajan and Jack 81 to the left, reducing participation in the public system. Public good spending reduces the value of public insurance relative to private insurance, even holding m constant, and hence allows the public system to be better targeted to individ- uals with low incomes and high risk. The reason low-income and high-risk individuals opt into the public system while others opt out is that the value of public insurance differs correspond- ingly, with the payment in the bad state independent of income. Such a payment is worth more to individuals with lower incomes (due to declining marginal utility of income) and to individuals with higher risk, that is, those who expect to receive the payment more often. This kind of insurance is worth less to individuals with a low marginal utility of income and to individuals who do not expect to need it very often. Thus, undifferentiated public insur- ance is necessarily valued differently by different individuals and is thereby self- targeted to the poor and vulnerable.19 Of course, more closely targeted transfer systems are usually considered better because they allow more to be spent on each recipient. However, the improvement in targeting associated with an increase in G does not obviously increase welfare because m is unlikely to increase or remain constant. Expenditure on the public good must be financed by reductions in public insurance payments, and although there are fewer recipients, those who quit the system have low risk and receive the transfer infrequently (compared with those who remain). By contrast, unlike administrative expenditures (for example, on outreach and monitoring) designed to improve the targeting of certain transfer programs, the public good embodies direct benefits of its own, both to individuals who opt out of the public insurance system and to those who stay in (unless m is so large that the transfer-inclusive income of publicly insured individuals is higher in the bad state than in the good state). The optimal balance between public insurance and risk reduction accounts for both these targeting and risk reduction effects of the public good. III. OPTIMAL PUBLIC EXPENDITURE Since shocks are independently (but not identically) distributed across the population, there is no aggregate uncertainty about public (or private) insur- ance payments. Given the policy variables m and G, the cost of running the public system is (1) M(m, G) = m (1 - -g(G, p))P(y, p)dy dp. 19. A similar targeting mechanism is used in Besley and Coate's (1991) article on in-kind transfers. 82 THE WORLD BANK ECONOMIC REVIEW The costs of public insurance plus the public good must be no greater than the revenue available, that iS20 (2) R > M(m, G) + G. Since increasing G induces more individuals to opt out of public insurance, the impact on the level of per capita spending, m, is ambiguous. On the one hand, in the case of medical care, for example, fewer public patients means higher quality (m) for those remaining. On the other hand, the smaller insur- ance budget (equal to R - G) means lower quality per capita. Using the definitions of w(y,7; m) and v(y,7) in footnote 17 and assuming a utilitarian welfare function,21 the government's optimization problem is max W(G, m) = w(T, y, m)4(y, P)dp dy G,m JP (3) + J v(T, y) (y, p)dp dy subject to i = 7T(G, p) and R > M(m, G) + G. This optimization problem is potentially nonconvex. To begin however, the first-order conditions are assumed to be sufficient for a maximum. A heuristic illustration of a possible nonconvexity and its implication is then provided. First-Order Approach To derive the first-order conditions for the government's problem, the Lagrangian is defined as (4) L(G, m; A) = W(G, m) + A[R - (G + M(m, G))] 20. In a world with covariate risk, aggregate public (and indeed private) insurance spending would be uncertain, and how unusually high (or low) expenditures would be financed would need to be carefully specified. If governments and insurers have access to reinsurance markets, equation (2) holds in expected value and individuals can be shielded from the aggregate fluctuations. They can be partially shielded if governments and insurers can borrow on capital markets to cover unusually large costs. Without access to such markets, governments will have to save and dissave as events necessitate, but the mix of public spending is unlikely to be significantly affected. 21. Some readers might prefer to adopt a welfare function that exhibits a degree of inequality aversion, such as the type suggested by Atkinson (1970). Within the standard expected utility framework, risk aversion induces declining marginal utility of income, so that even a utilitarian welfare function would lead to redistributive policies. Devarajan and Jack 83 where A is the multiplier on the constraint. The first-order condition for G is dq) + jp Vdq (5)J OG O - A 1 -M m ' d) - 0r( - 77) G de = 0 -[1 UP OG fo where 9G(P; m, G) is the increase in the income of individuals with probability p who are indifferent between public and private insurance, given the policy variables m and G, and dq) is shorthand for 0(y, p)dy dp. The following expressions for the partial derivatives can then be substituted into this first- order condition: (6) G = [u(y) - u(ay + m)] G Ov (7) = u'(y)(1 - a)y7G. OG Condition (4) can be usefully interpreted as balancing the marginal benefits and costs of expanding the public good. S0d(D + OVdD p OG p, OG Marginal benefit to insiders Marginal benefit to outsiders (8) = A 1 -m dq) - j(1 w9 d(D 11 jP OG OP Marginal cost of public good Savings on insiders Fall in public enrollment The first term represents the benefits to users of the public system (insiders) and the second term the benefits to users of the private system (outsiders). The mar- ginal cost of expanding the public good, priced at the shadow cost of public funds (A), has three elements. The financial cost of an extra unit of G is simply one dollar; as a result of the expansion, the probability of the bad state falls, so expenditures on users of the public insurance system fall; finally, the reduction in risk for all individuals induces some of them (those on the boundary OP) to opt out of the public system (note YG < 0), yielding a per capita cost saving of m with probability (1 - 7) to the public budget. 84 THE WORLD BANK ECONOMIC REVIEW The first-order condition for m is (9) [d - A (1 - T(G, P))d) + m (1 - 7(G, P))Ymd1 = 0 where (10) _ = (1 - -(G, p))u'(ay + m). Om The first-order condition thus simplifies to (11 (1 - 7)u'(ay + m)d(D A jp(1 - -r)D + m J,p(1 - )mdeD Marginal benefit -Intensive marginal cost Extensive marginal cost The term on the left side is the marginal social benefit of expanded quality of public insurance, comprising the expected marginal utility of additional income in the bad state for users of the public system. The term on the right side is the marginal social cost, again valued in terms of public revenue, com- prising the cost of paying an extra dollar to public insurance beneficiaries in the bad state and the cost of paying the full benefit m in the bad state to indi- viduals who join the public system as a result of the increased benefits. Heuristic Approach Equations (8) and (11) show the policy tradeoffs at the optimum, assuming that the second-order conditions are satisfied. However, even if simple func- tional forms are assumed for utility and the effect of the public good, they prove too complex to solve analytically. This section presents a more heuristic analysis of the tradeoff between public insurance and risk reduction. An increase in the public insurance budget, M, would be effected through an increase in m and would be matched by a reduction in public good spending, G. For individuals who participate in the public system there is a direct benefit: payments in the bad state increase, even with the increased par- ticipation. The social marginal benefit (the sum of the marginal benefit across participants) may initially increase with M, as participation increases dominate. At some point it is assumed that the marginal benefit to insiders begins to fall as M increases, while remaining positive. The marginal benefit per dollar of extra spending is the ratio of the left side of equation (11) to the square-bracketed term on the right side. An increase in M is costly to the extent that it must be matched by a reduction in G. This cutback in public good spending has direct implications Devarajan and Jack 85 for individuals in the public insurance system and individuals who opt out. For insiders, the marginal cost is assumed to be initially positive and increasing, but as long as the total budget, R, is large enough, it must become negative (and hence decrease) at some (possibly large) value of M. This is because when M, and hence m, is large, at least some publicly insured individuals receive higher income in the bad state than in the good, and a fall in G (which increases the likelihood of the bad state) increases their expected utility. The marginal cost imposed on insiders is shown in figure 3 as the dotted line MCin. This corresponds precisely to the ratio of the first term on the left side of equation (8) to the square-bracketed term on the right (with suitable change of sign). For individuals who do not participate in the public scheme, the increase in insurance budget reduces welfare, and further increases in M initially prove costly to those outsiders. Of course, as M increases, participation in the public system becomes more attractive (both because m is higher and because G is lower), so the increase in total costs imposed by the shift in spending on outsi- ders as a group is less than if participation was fixed. Indeed, if M increases enough, the whole population might join the public system, and the marginal cost imposed on outsiders (of whom there are now none) would be zero. This is shown as the dashed curve MCout in figure 3 and corresponds to the ratio of the second term on the left side of equation (8) to the square-bracketed term FIGURE 3. Marginal Costs Associated with Decreased Public Good Spending and an Increase in the Public Insurance Budget / out4 Note: MC is the total marginal cost, MCout is the marginal cost to outsiders (users of the private insurance system), and MCin is the marginal cost to insiders (users of the public insurance system). 86 TIE WORLD BANK ECONOMIC REVIEW on the right (again, with suitable change of sign). Total marginal costs of increasing M are denoted MC. Figure 4 combines the marginal cost and marginal benefit curves. Point A is a local maximum, at which the budget devoted to the public insurance system is M*. Point B is a local minimum, and welfare is increased by either spending more or less on the system (as indicated by the arrows). Clearly, the primary determinants of the optimal level of spending on public insurance (and hence also on the public good) are the levels of the two curves MB and MC. In particular, because MC reflects the marginal benefits of public good spending, the position of this curve will depend crucially on how effective such spending is at reducing risk. If it is ineffective, the MC curve will be lower and more spending should be allocated to public insurance (point A shifts right). It is even possible that the public good is so ineffective that MC lies below MB everywhere, in which case the whole budget should be spent on the insurance scheme. Increasing the available overall budget, R, means that for a given insurance budget, M, there is more spending on the public good, making the bad state less likely. This shifts the MB and MC curves down in figure 4, with an ambig- uous effect on optimal insurance spending. However, if the available budget increases above a certain threshold, R*, optimal spending on public insurance abruptly jumps from M* to R and the whole budget should be spent on transfers in the bad state. This is simply because with a large budget, transfer-inclusive income in the bad state can be FicURE 4. Marginal Cost and Marginal Benefit of an Increase in the Public Insurance Benefit MC A C B MB Devarajan and Jack 87 larger than income in the good state, so the bad state is preferred.22 However, this possibility should be viewed only as a technical curiosity, since social protection budgets are extremely limited in most poor countries (likely leaving governments constrained at point A) and since if the budget was so large, the government would surely search for alternative ways to distribute it to the population, instead of just doing so in the bad state. IV. SIMULATING THE EFFECTS OF MULTIDIMENSIONAL POVERTY Several comparative static exercises can be contemplated within this frame- work, most of which require simulation methods.23 The issue of most policy relevance-and pertinent to the discussion of vulnerability and multidimen- sional poverty-is the effect of correlation between income and risk. If income- poor people tend to face greater risk, how should policy respond in terms of the allocation of the budget to insurance and public goods? Because analytic answers to this kind of question are hard to come by, a simulation exercise is used below to develop some intuition. Specification Individuals are assumed to be distributed on fl = [ymin, ymax] x [0, 1] accord- ing to the bivariate log-normal distribution with mean parameters Ay, /ap, dis- persion parameters o7y, o-p, and correlation coefficient p (with the external probability distributed proportionately across the domain). The effect of the public good on risk is parameterized by assuming that the 7 function takes the form (12) -r(G, p) = p + 8(1 - p)(1 - e-kG) for some 6 E (0, 1). This has the properties that 7T(O, p) = p, 7Tp > 0, 7TG > 0, and -gGp < 0. All individuals have the same von Neumann-Morgenstern utility functions, specified by the constant relative risk aversion form 1-or (13) u(y) = where o- is the coefficient of relative risk aversion. With these parameterizations, it is straightforward to show that for each underlying probability p, there is a cutoff income level 9(p; m, G) such that 22. The authors thank an anonymous referee for providing the intuition for this result. 23. These include, for example, variations in risk aversion, the effects of including the administrative costs of running public insurance systems, and changes in the within-state productivity of the public good (so far it has been assumed that G affects only the probability of different states occurring, but not the realized income in those states). 88 THE WORLD BANK ECONOMIC REVIEW individuals with probability of the good state equal to p choose private insur- ance whenever y > 9(p; m, G). The expression for 9(p; m, G) is m (14) (p; m, G) = - z(7(G, p)) - a where (15) Z(7T) (,7T± a(l -T))(1't) - T In the simulation a simple grid search is performed over (G, m) pairs. Because of the government's budget constraint, there is only one degree of freedom, so G is simply iterated over. For each G, m is iterated over using a basic Newton's method until m(G) is found such that the budget constraint is satisfied. Welfare is calculated at each G to find the maximum. Income-Risk Correlation and Public Policy The coefficient of relative risk aversion is fixed at o-= 1.5, and the correlation between risks and income is varied. Recall that p is the probability of the good state, so a positive correlation indicates an environment in which individuals with low incomes on average face a greater chance of being in the bad state. In changing the distribution of individuals in fl, as occurs when the correlation is varied, aggregate income in the economy is naturally altered. Holding public sector revenue constant in such a comparative statics exercise may not be appropriate. Therefore the budget is fixed as 20 percent of GDP (figure 5). The share of public expenditure devoted to the public good is higher at the extremes-correlations near + 1 and - 1-but nearly constant within this range. Participation in the public system has a similar (but inverted) shape- lower rates at correlations near + 1 and - 1 but nearly constant for a wide range of subunitary correlations. These simulation results suggest that the mix of public spending between risk reduction and insurance is not very sensitive to the correlation between risk and income heterogeneity. Of course, facing more risk (to the extent this is so) makes the poor poorer. The simulation suggests, however, that the impact this has on policy may be relatively small. The heuristic approach of section III is useful in providing intuition for the apparently small and ambiguous impact of changes in p on optimal public good spending. The initial impact of an increase in p is to increase the resources devoted to public insurance, due to its beneficial targeting properties. But the concomitant reduction in public good spending induces more individuals to take up public insurance, thereby lower- ing insurance benefits per beneficiary, which in turn mitigates the social benefit of the initial increase in the public insurance budget. Devarajan and Jack 89 FiCURE 5. Relationship between Optimal Expenditure Policy and the Correlation of Incomes and Risks 100% - 55% 90% - 50% 80% - - 45% O z 70% - GIR t 40% 60% -participation in public system 35% 50% 40%30% -1 -0.5 0 0.5 1 P Of course, the simulation does not definitively support this conclusion; no simulation can. The simulation does indicate that empirical observations about the correlation between income and risk do not automatically support a shift toward either public good spending or public insurance. V. CONCLUSIONS This article presents a model of the allocation of budgetary resources in a risky environment. In particular, it examines the division of public expenditures between those that reduce underlying uncertainty and those that provide expli- cit insurance. In addition to deriving conditions for the optimal allocation of public resources, this formulation permits an evaluation of alternative incre- mental changes to each kind of expenditure when public expenditures are not necessarily optimal, in the spirit of cost-benefit analysis. An important feature of this article is that it assumes an efficient private insurance market. Some implications of inefficient insurance markets are dis- cussed below, but first it is noted that if insurance markets are efficient, it might be expected that there should be no role for public insurance and that all public spending should be directed toward risk reduction. This is correct in the absence of distributional concerns, but when individuals are heterogeneous 90 THE WORLD BANK ECONOMIC REVIEW with regard to either income levels (y) or risk exposure (p), the kind of public insurance described here performs a redistributive function. In particular, its self-targeting properties-individuals with high incomes or low risk tend to opt out of the public system-make it a useful tool of social protection in the broad sense of the term. Although this redistributive role underpins much of the support expressed for public insurance systems in the context of social protection, the somewhat complicated analytics of self-targeting are underappreciated in the literature. The exact nature of the targeting inherent in the system depends on the division of spending between the public good and the insurance program through their impact on the participation decision. Public good spending makes public insur- ance less valuable, thus focusing participation on individuals who are relatively poor and relatively high risk. But such spending must be financed with reductions in insurance benefits: so the better targeted public insurance scheme may provide less generous benefits for each person enrolled. The nonlinearities induced by changes in the participation decisions mean that characterization of the optimal spending allocation is nontrivial and that optimal allocations do not vary monotonically with underlying parameters (as shown, for example, in figures 4 and 5). What changes should be expected with a more realistic view of the potential inefficiencies of the private insurance market? For instance, it has been assumed that risk heterogeneity does not lead to adverse selection and that insurance can be purchased at actuarially fair prices. Although the potential for adverse selection is clear, the qualitative features of the model would be expected to hold if introduced explicitly, for two reasons. First, while adverse selection leads to individuals with low risk opting out of private insurance markets, much practical experience (for example, in Chile) suggests that indi- viduals with high risk tend to end up in the public system. This is exactly the pattern of participation the model here predicts. Second, the model ignores issues of both hidden action and hidden infor- mation moral hazard. The social returns to public insurance might be expected to fall in the presence of moral hazard, so that public good provision might be more favored-a seemingly valid argument for the case of hidden information moral hazard. If hidden action moral hazard is thought to be important, then just as incentives for precautionary actions can be reduced by insurance, so too can public good provision crowd out private precautions, and so the net social productivity of G may fall as well. The net impact on the division between G and M would then be ambiguous. Notwithstanding these shortcomings, the analysis in this article points to a role for publicly provided insurance that is distinct from its usual role as a cor- rection for failures in the private market. In countries where governments have limited instruments for redistributing income, especially between low- and high-risk individuals, publicly provided insurance can go a long way toward achieving this welfare-enhancing redistribution. Devarajan and Jack 91 REFERENCES Albarran, Pedro, and Orazio P. Attanasio. 2003. "Do Public Transfers Crowd Out Private Transfers? Evidence from a Randomized Experiment in Mexico." InStefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Atkinson, Anthony. 1970. "On the Measurement of Inequality" Journal of Economic Theory 2(3):244-63. Banerjee, Abhijit. 2003. "The Two Poverties." InStefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Barrett, Christopher B., Stein Holden, and Daniel C. Clay. 2003. "Can Food-for-Work Programmes Reduce Vulnerability?" InStefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Besley, Timothy, and Stephen Coate. 1991. "Public Provision of Private Goods and the Redistribution of Income" American Economic Review 81(4):979-84. Dercon, Stefan ed. 2003. Insurance against Poverty. Oxford, U.K.: Oxford University Press. Dercon, Stefan, and Pramila Krishnan. 2003. "Food Aid and Informal Insurance." InStefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Glazer, Jacob, and Thomas McGuire. 2000. "Optimal Risk Adjustment of Health Insurance Premiums: An Application to Managed Care" American Economic Review 90(4):1055-71. Glazer, Jacob, and Thomas McGuire. 2002. "Setting Health Plan Premiums to Ensure Efficient Quality in Health Care: Minimum Variance Optimal Risk Adjustment" Journal of Public Economics 84(2):153-75. Jalan, Jyotsna, and Martin Ravallion. 2003. "Household Income Dynamics in Rural China." In Stefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Kamanou, Gisele, and Jonathan Morduch. 2003. "Measuring Vulnerability to Poverty." In Stefan Dercon ed., Insurance against Poverty. Oxford, U.K.: Oxford University Press. Ligon, Ethan, and L. Schechter. 2003. "Measuring Vulnerability" Economic Journal 113(486):C95- C102. Mirrlees, James. 1971. "An Exploration in the Theory of Optimum Income Taxation" Review of Economic Studies 38(2):175-208. Morduch, Jonathan. 1999. "Between the State and the Market: Can Informal Insurance Patch the Safety Net?" World Bank Research Observer 14(2):187-207. Newhouse, Joseph. 2002. Pricing the Priceless: A Health Care Conundrum. Cambridge, Mass.: MIT Press. Pritchett, Lant, Sudarno Sumarto, and Asep Suryahadi. 2000. "Quantifying Vulnerability to Poverty: A Proposed Measure, Applied to Indonesia" Research Working Paper 2437. World Bank, Washington, D.C. Schlesinger, Harris, and Emilio Venezian. 1986. "Insurance Markets with Loss-Prevention Activity: Profits, Market Structure, and Consumer Welfare" Rand Journal of Economics 17(2):227-38. Thirsk Wayne, ed. 1998. Recent Experience with Tax Reform in Developing Countries. Washington, D.C.: World Bank. World Bank. 2002. "Social Protection" Poverty Reduction Strategy Sourcebook. vol. 2. Washington, D.C. The Incidence of Public Spending on Healthcare: Comparative Evidence from Asia Owen O'Donnell, Eddy van Doorslaer, Ravi P. Rannan-Eliya, Aparnaa Somanathan, Shiva Raj Adhikari, Deni Harbianto, Charu C. Garg, Piya Hanvoravongchai, Mohammed N. Huq, Anup Karan, Gabriel M. Leung, Chiu Wan Ng, Badri Raj Pande, Keith Tin, Kanjana Tisayaticom, Laksono Trisnantoro, Yuhui Zhang, and Yuxin Zhao The article compares the incidence of public healthcare across 11 Asian countries and provinces, testing the dominance of healthcare concentration curves against an equal distribution and Lorenz curves and across countries. The analysis reveals that the distribution of public healthcare is prorich in most developing countries. That distri- bution is avoidable, but a propoor incidence is easier to realize at higher national incomes. The experiences of Malaysia, Sri Lanka, and Thailand suggest that increas- ing the incidence of propoor healthcare requires limiting the use of user fees, or protecting the poor effectively from them, and building a wide network of health facilities. Economic growth may not only relax the government budget constraint on propoor policies but also increase propoor incidence indirectly by raising richer individuals' demand for private sector alternatives. JEL Codes: H22, H42, H51. Owen O'Donnell (corresponding author) is an assistant professor of quantitative methods at the University of Macedonia, Greece; his email address is ood@uom.gr. Eddy van Doorslaer is a professor of health economics at Erasmus University, the Netherlands; his email address is vandoorslaer@few.eur. nl. Ravi P. Rannan-Eliya is director of the Institute for Health Policy in Sri Lanka; his email address is ravi@ihp.1k. Aparnaa Somanathan is a fellow at the Institute for Health Policy, Sri Lanka; her email address is aparanaa@ihp.1k. Badri Raj Pande is the director of the Nepal Health Economics Association; his email address is neil@info.com.np. Shiva Raj Adhikari is a researcher at the Nepal Health Economics Association; his email address is sssadhikari@yahoo.com. Laksono Trisnantoro is a professor of health policy at Gadjah Mada University, Indonesia; his email address is trisnantoro@ yahoo.com. Deni Harbianto is a researcher at Gadjah Mada University, Indonesia; his email address is d_harbianto@yahoo.com. Charu C. Garg is a health economist at the World Health Organization; her email address is gargc@who.int. Piya Hanvoravongchai is a researcher at the International Health Policy Programme, Thailand; his email address is piyaorn@ihpp.thaigov.net. Kanjana Tisayaticom is a researcher at the International Health Policy Programme, Thailand; her email address is kanjana@ihpp. thaigov.net. Mohammed N. Huq is a lecturer at Jahangirnagar University, Bangladesh; his email address is m_nazmul@proshikanet.com. Anup Karan is a Takemi Fellow at the Harvard School of Public Health; his email address is akaran@hsph.harvard.edu. Gabriel M. Leung is a professor of THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 93-123 doi:10.1093/wber/1hl009 Advance Access Publication 24 January 2007 () The Author 2007. 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 93 94 THE WORLD BANK ECONOMIC REVIEW Propoor public spending on healthcare and other services is a stated objective of national governments and international agencies. It is central to the mission of the World Bank and is a key component of the Heavily Indebted Poor Countries Initiative and the International Monetary Fund's Poverty Reduction and Growth Facility. Motivations include redressing inequity in the distribution of healthcare, reducing health inequality, and raising the human capital of the poor and thereby the growth potential of the economy. In low-income countries, where administrative constraints on redistribution through cash transfers are particularly binding, a subsidiary justification for public spending on healthcare may be the alleviation of poverty and the reduction of inequality (Besley and Coate 1991). The validity of these arguments for public spending on healthcare rests on the empirical question of whether the spending is in fact targeted to the poor. Benefit incidence analysis identifies the recipients of public spending in relation to their position in the income distribution. Benefit incidence studies, many conducted by the World Bank, generally find that public spending on healthcare in developing countries is not concentrated on the poor (van de Walle 1995; Castro-Leal and others 2000; Mahal and others 2000; Sahn and Younger 2000; Filmer 2003). Most of these studies have been conducted on an ad hoc basis, with relatively little attention to consistency in methods. Limitations in the comparability of the evidence make it difficult to draw lessons about the economic, political, and health system characteristics that explain greater and lesser success in targeting health spending to the poor. This article presents comparable evidence on the incidence of public health spending using consistent methods across eight Asian countries (Bangladesh, India, Indonesia, Malaysia, Nepal, Sri Lanka, Thailand, and Vietnam) and three Chinese provinces or regions (Gansu, Heilongjiang, and Hong Kong Special Administrative Region). Dominance tests are used to determine whether the distribution of public healthcare deviates significantly from perfect equality. Many indicators show that poorer individuals are generally less healthy (Gwatkin and others 2003) and, one may presume, in greater need of healthcare. From an egalitarian perspective an equitable distribution of health- care demands that resources be concentrated on the poor. Evidence that the translational public health at the University of Hong Kong; his email address is gmleung@hku.hk. Keith Tin is a researcher at the University of Hong Kong; his email address is tinyiukei@hkusua.hku.hk. Chiu Wan Ng is a lecturer at the University of Malaya, Malaysia; her email address is chiuwan.ng@ummc. edu.my. Yuxin Zhao is a professor of health economics at the National Health Economics Institute, China; her email address is yuxin.zhao@cnhei.edu.cn. Yuhui Zhang is a researcher at the National Health Economics Institute, China; his email address is zyh@nhei.cn. The authors thank three anonymous referees and the editor for valuable comments. The European Commission International Research Cooperation with Developing Countries (INCO-DEV) program (ICA4-CT-2001-10015) funded the Equity in Asia-Pacific Health Systems (Equitap) project from which this article derives. The Health, Welfare, and Food Bureau of the government of the Hong Kong Special Administrative Region funded the analysis for Hong Kong. A supplemental appendix to this article is available at http://wber. oxfordjournals.org/. O'Donnell and others 95 poor do not receive their population share of health spending would be suffi- cient to reject equity in the allocation of public healthcare. While the main jus- tification for public provision of healthcare is likely to be its impact on the level and distribution of population health, redistribution of living standards may be a further motivation in largely informal economies that are constrained in the execution of tax and cash transfer policies.' To assess the redistributive impact of public health spending, its distribution is compared with the Lorenz curve of household income. One limitation of many previous benefit incidence studies is the crudeness of the unit cost data used to value services (van de Walle 1998; Sahn and Younger 2000). This study derives costs from detailed health accounts, avail- able for most of the countries and provinces, which document public expendi- tures across health services, facilities, and regions. This allows examination of whether conclusions about the incidence of public healthcare are sensitive to analysis of use or expenditure data. Data and methods are described in the next section and results are presented and discussed in section II. The findings are summarized in section III. I. DATA AND METHODS The objective is to estimate and assess the distribution of public healthcare in relation to economic status. For each country data are from recent health or socioeconomic surveys that provide information on both use of public health- care and a suitable measure of living standards (see table S-1 in the supplemen- tal appendix, available at http://wber.oxfordjournals.org/). All are nationally representative except for the surveys of Chinese provinces. The preferred proxy for living standards is household (per adult equivalent) consumption, which includes the value of goods produced by the household for its own consump- tion and a use-value of housing and durable goods.2 Household expenditure, rather than consumption, is used for Hong Kong SAR, where household production is much less significant. For Malaysia the only available measure of living standards included in the health survey is household income, which is likely to understate the living standards of rural households. It is, however, the measure that has been used in previous incidence studies of Malaysia (Meerman 1979; Hammer, Nabi, and Cercone 1995). Distributions of three categories of public healthcare-hospital inpatient care, hospital outpatient care, and nonhospital care-are examined. 1. In Latin America cash transfers are increasingly used to affect the distribution of income, as well as that of health and education services, but this is less so in the low-income economies of Asia, where in-kind transfers, such as healthcare, continue to predominate. 2. The equivalence scale used is eh = (A, + 0.5Kh)0.75, where Ah is the number of adults in household h, and Kh is the number of children 0-14 years old. Parameter values were set on the basis of estimates summarized in Deaton (1997, pp. 241-70). 96 THE WORLD BANK ECONOMIC REVIEW Nonhospital care is an aggregate of visits to doctors, polyclinics, health centers, and antenatal care (table S-2). For inpatient care the recall period is 12 months, except in Bangladesh (3 months) and Sri Lanka (2 weeks). For all other care the recall period is generally 2 weeks to 1 month, except in Bangladesh where it is 3 months. Use data do not capture variations in the quality of services received across facilities and geographic locations. This is a potentially important deficiency given evidence of marked quality differences favoring richer neighborhoods even within a single city, such as Delhi, India (Das and Hammer 2005). The service-specific non-negative public subsidy received by an individual can be defined as: (1) Ski = max(0, qkickj - fki) where qki is the quantity of service k used by individual i, ckj is the unit cost of providing k in region j where i resides, and fki is the amount paid for k by i. Where possible, variations in costs by facility (local, district, teaching hospital) and service (inpatient/outpatient) are taken into account. Unit costs are com- puted as: TREk1 (2) Ckj kj k' E qkiWi iej where TREkj is total recurrent public expenditure and wi is an expansion factor that inflates sample use to population use. The total public subsidy received by an individual is computed as Si = Yk akSki, where the ak terms are scaling factors that standardize use recall periods across services. National health accounts, available for Bangladesh, the Chinese provinces, Hong Kong SAR, Sri Lanka, and Thailand, are used to disaggregate expenditure figures by facility, service, and region. Full accounts are not available for India, Indonesia, Malaysia, Nepal, and Vietnam. For India unit subsidies computed for another benefit incidence study are used (Mahal and others 2000). These are specific to 960 subgroups (three facilities, 16 major states, urban-rural residence, gender, and five income quintiles). For Indonesia public health expen- diture review figures allow expenditures to be disaggregated for each of 30 provinces. For Malaysia expenditure data were disaggregated to five levels of public hospital care, but geographic disaggregation was not undertaken since the use data could not be analyzed by this dimension. Incomplete health accounts for Nepal allow disaggregation by hospital and nonhospital care by region. For Vietnam public accounts and hospital costing estimates were used to compute unit costs by service and facility but not by region (World Bank 2001). Subtraction of the user payment from equation (1) to get the net benefit of the service is appropriate provided that quality is not responsive to the O'Donnell and others 97 payment. This is an untestable assumption with the available data. For China, India, Indonesia, Malaysia, Nepal, and Sri Lanka either the survey data do not contain information on payments made by individuals for public health ser- vices or the data are not considered sufficiently reliable, for example, because payments for public and other care are likely to be confused. For these countries it is assumed that all users in a particular region pay the same charge for a given service. Waiting and travel time also reduce the net benefit from care and should, in principle, be valued and subtracted in computing the subsidy. The survey data do not permit this, however. As a consequence, benefits to the rural poor, in particular, may be overstated to the extent that they travel long distances to access better quality care. By contrast, the cost of waiting time will be less for the poor if time is valued according to wage rates. The incidence of public healthcare is described by its concentration curve, which plots the cumulative proportion of healthcare use and subsidy against the cumulative proportion of the population ranked by household consumption per adult equivalent. To establish whether the subsidy is propoor, in the sense that lower income individuals receive more of the subsidy than the better-off, a test is conducted of whether the concentration curve dominates (lies above) the 450 line. Whether the poorest 20 percent of individuals consume more than 20 percent of healthcare is also tested. Dominance of the concentration curve over the Lorenz curve of household consumption is tested to establish whether spending on public healthcare reduces inequality. For the dominance tests standard errors of the ordinates of curves and of differences in ordinates are computed, allowing for dependence between curves where appropriate (Bishop, Chow, and Formby 1994; Davidson and Duclos 1997)." A multiple comparison approach to testing is adopted (Beach and Richmond 1985; Bishop, Formby, and Thistle 1992), with the null defined as curves being indistinguishable. This is tested against both dominance and cross- ing of curves (Dardanoni and Forcina 1999). The null is rejected in favor of dominance if there is at least one significant difference between the ordinates of two curves in one direction and no significant difference in the other direc- tion across 19 evenly spaced quantile points from 0.05 to 0.95. The null is rejected in favor of crossing if there is at least one significant difference in each direction. The 5 percent level of significance is used with critical values from the studentized maximum modulus distribution to allow for the joint nature of the test (Beach and Richmond 1985).4 An alternative dominance test consistent with the intersection-union prin- ciple (Kaur, Rao, and Singh 1994; Howes 1996), which has been used in the 3. The computation is carried out in Stata. 4. Dardanoni and Forcina (1999) show that the probability that this test will falsely reject the null in favor of dominance does not exceed the significance level and report Monte Carlo evidence suggesting that the actual significance level is well below its nominal value. 98 THE WORLD BANK ECONOMIC REVIEW benefit incidence literature (Sahn and Younger 2000; Sahn, Younger, and Simler 2000), takes nondominance as the null and tests this against the alterna- tive of strict dominance. This is a conservative test that requires statistically significant differences in ordinates at all points of comparison for the null to be rejected. Dardanoni and Forcina (1999) present Monte Carlo evidence showing that while this test reduces the probability of falsely rejecting nondominance to a negligible value, compared with the multiple comparison approach it has greatly reduced power of detecting dominance when true. Given these results, most weight in the discussion below is given to the results from the multiple comparison tests, but discrepancies with the more conservative intersection- union test are pointed out. II. RESULTS In Hong Kong SAR, Malaysia, and Thailand the concentration curve of the total public health subsidy dominates both the Lorenz curve and the 450 line of equality (table 1, final column), indicating that the subsidy is both inequality-reducing and propoor. With the exception of the comparison with the 450 line in the case of Thailand, these dominance results are robust to use of the stricter test. In Sri Lanka an equal distribution of the total subsidy is not rejected. In relative terms this shifts the distribution of living standards toward the poor, as the concentration curve dominates the Lorenz curve. In the remaining countries and provinces the concentration curve of the total subsidy is dominated by the 45' line but, with the exceptions of India and Nepal, dom- inates the Lorenz curve. That is, the subsidy is prorich but inequality reducing. For Bangladesh and the two Chinese provinces nondominance relative to both the Lorenz curve and the 450 line cannot be rejected when the more conserva- tive intersection-union test is employed.5 The degree to which the public health subsidy is targeted to the poor can be seen more explicitly by examining the share of the subsidy received by the poorest 20 percent of individuals (table 2). Public healthcare is clearly most propoor in Hong Kong SAR, with the poorest fifth of the population receiving almost two-fifths of the total subsidy (table 2, final column). In Malaysia the poorest quintile also receives significantly more than 20 percent of the total subsidy, but the propoor bias is much less than it is in Hong Kong SAR. In Sri Lanka and Thailand the poorest quintile's share of the total subsidy does not differ significantly from 20 percent. In the remainder of countries and pro- vinces, with the exception of Bangladesh, the poorest 20 percent of individuals receive significantly less than 20 percent of the public health subsidy. The share going to the poorest 20 percent of individuals is lowest in Nepal, at less than 7 percent, followed by the two Chinese provinces, at 8-10 percent. In these 5. Concentration and Kakwani indices, which provide summary measures of the magnitude by which the concentration curve deviates from the 450 line and the Lorenz curve, are given in table S-3. TABLE 1. Tests of Dominance of Concentration Curves for Public Health Service Use and Subsidy against the Lorenz Curve and the 45 Degree Line of Equality Use Subsidy Hospital Hospital Hospital Hospital inpatient outpatient Nonhospital inpatient outpatient Nonhospital Total Country, province, or region Lorenz 45 Lorenz 45 Lorenz 45 Lorenz 45 Lorenz 45 Lorenz 45 Lorenz 45 Bangladesh - + - + + - + + - Gansu, China + +* - n.a. n.a. + - + - n.a. n.a. + - Heilongjiang, China +* - n.a. n.a. + - + - n.a. n.a. + - Hong Kong SAR +* +* +* +* +* +* +* +* +* +* +* +* +* +* India -* +* -* +* + -* +* +* -* Indonesia - x -* +* + -* * + +* Malaysia +* + +* +* + +* + +*-1 +* +* +* Nepal' + n.a. n.a. +* - -* n.a. n.a. x -* x Sri Lanka +* +* + +* +*+ n.a. n.a. +* Thailand +* +* x +* - +* +* + -P +* +* + Vietnam + -* +* + +* - + -* +* + +* - Blank cell indicates failure to reject the null hypothesis that curves are indistinguishable using the multiple comparison test (Bishop, Formby, and Thistle 1992) at the 5 percent significance level. 0 x indicates rejection of the null hypothesis that curves are indistinguishable in favor of curves crossing using the same test. + /- indicates rejection of the same null hypothesis in favor of dominance using the same test. A + indicates that healthcare is more concentrated on the poor than is household consumption per adult (Lorenz) or equal per capita distribution (45), while a - indicates that it is less concentrated. *indicates rejection of the null hypothesis of nondominance in favor of an alternative of strict dominance using the intersection-union test (Howes 1996) and a 5 percent significance level. Dominance is in the direction indicated by the + or -, as above. n.a. means that data were not available to conduct the test. aThe results in the hospital inpatient columns refer to both inpatient and outpatient. Source: Authors' calculations based on survey data documented in table S.1 (see supplemental appendix available at http://wber.oxfordjournals.org/). 0lOZ 'LXjr-qoj no punjruP1uoIt uoiluuu ju /'Jo-slunopojxo-jqm//:dt[ uoij popolumo( TABLE 2. Share of Total Household Consumption and Public Healthcare Subsidy Received by Poorest Quintile of Individuals (percent) Hospital care Country, province, Household consumption or region per adult equivalent Inpatient Outpatient Nonhospital care Total subsidy z Bangladesh 7.25* (0.0437) 15.20 (6.3732) 11.60* (1.8853) 24.42 (5.5695) 16.78 (3.4916) p Gansu, Chinaa 5.24* (0.0695) 7.27* (1.5331) 9.57* (1.6473) n.a. 8.17* (1.2265) 2 O Heilongjiang, China' 5.98* (0.0759) 6.57* (1.8184) 12.32* (2.5677) n.a. 10.47* (1.8729) g Hong Kong SAR 6.82* (0.0377) 38.77* (3.2580) 38.68* (2.2048) 38.19* (1.7718) 38.73* (2.7463) n India 10.50* (0.0083) 10.70* (1.1086) 18.59 (1.6219) 26.23* (1.5471) 12.49* (0.9553) * Indonesia 9.77* (0.0078) 3.80* (0.3762) 5.77* (0.4857) 19.73 (0.3199) 13.46* (0.2582) 1 Malaysia 7.20* (0.0370) 21.19 (0.8807) 18.72 (1.1208) 32.25* (1.3422) 22.95* (0.6921) < Nepalb 8.05* (0.0534) 3.52* (1.4851) 3.52* (1.4851) 9.04* (1.7220) 6.64* (1.1780) Sri Lankac 8.31* (0.0725) 20.76 (2.6013) 21.11 (1.9418) n.a. 20.88 (1.8367) Thailand 6.94* (0.0589) 21.26 (1.4144) 17.70* (1.0278) 31.16* (1.9137) 20.06 (0.8963) Vietnam 8.78* (0.0429) 13.64* (1.9209) 11.55* (1.7049) 19.73 (1.7346) 14.79* (1.5416) *Significantly different from 20 percent at the 5 percent significance level. Bold indicates that the subsidy share is significantly different from the household consumption share. n.a. means that data were not available to conduct the test. Note: Numbers in parentheses are standard errors. aThere are no data on nonhospital care, but low-level hospitals, equivalent to polyclinics and health centers, are included. bIt is not possible to distinguish between hospital inpatient and outpatient visits. 'The subsidy specific to nonhospital care cannot be computed. Source: Authors' calculations based on data documented in table S.1 (see supplemental appendix available at http://wber.oxfordjournals.org/). El OZ'L A.iamqoj no punj r,uoW wuoiluuni.Iau /'Jo-sltunop.ojxonqm//:dBt[umo,ijpppolumou O'Donnell and others 101 cases, and in Bangladesh, India, and Indonesia, the richest quintile receives more than 30 percent of the total subsidy (not shown in table). In all cases but Nepal the share of the subsidy going to the poorest quintile is significantly greater than its share of total household consumption. Differences in Incidence across Healtb Services Only in Hong Kong SAR does the concentration curve dominate the 45' line for both hospital inpatient and outpatient care and for nonhospital care (see table 1), with the poorest quintile receiving about 39 percent of the subsidy to all three services (see table 2). In Malaysia the concentration curves for inpati- ent and nonhospital care lie above the 45' line, but the outpatient care curve does not deviate significantly from the line of equality (see table 1). In Thailand it is inpatient care that is equally distributed, while the concentration curves for the other types of care dominate the diagonal, at least using the less stringent test criteria. However, in both Malaysia and Thailand the poorest quintile receives significantly more than 20 percent of the subsidy only for non- hospital care (see table 2). In Sri Lanka there is equality in the distributions of all services except for a propoor distribution of outpatient care as measured by use (see table 1). In the remainder of countries and provinces, concentration curves for hospital care tend to lie below the diagonal-meaning that the better-off consume more-while the curves for nonhospital care lie above it. The poorest quintile fairly consistently receives less than 20 percent of the subsidy for hospital care and significantly more than 20 percent of the subsidy for nonhospital care only in India (see table 2). For most countries and provinces the distribution of nonhospital care domi- nates that of hospital inpatient and outpatient care (table 3), confirming that nonhospital care is generally more targeted to the poor than is hospital care. Comparison of Use and Subsidy Distributions Estimating the incidence of the public healthcare subsidy requires much more information than that of raw use. Unit costs must be estimated at the facility and regional levels and, where appropriate and possible, fees paid by individ- uals must be identified. The effort involved to obtain this extra information is worthwhile only if there is significant variation in unit costs or fees with the indicator of household living standards and if this covariance is sufficiently large relative to that for use. The dominance tests reported in table 1 display a considerable consistency across the use and subsidy measures. Only in 10 of 58 pairwise comparisons do the conclusions of the test differ depending on whether the distribution of use or the subsidy is examined. This is not an insubstantial degree of disagreement, but it suggests that the results of domi- nance tests are generally robust to the measure over which incidence is exami- ned and that variation in use, not unit subsidies, is the main driver of the public subsidy distribution. This increases the confidence that can be placed in studies that look only at use. It is consistent with the findings of Sahn and TABLE 3. Tests of Dominance between Concentration Curves for Different Public Health Services and between Use and Subsidy Distributions Use Use and subsidy Country, province, Hospital inpatient Hospital inpatient Hospital outpatient Hospital Hospital or region versus outpatient versus nonhospital versus nonhospital inpatient outpatient Nonhospital Bangladesh op>ip Gansu, China op>ip* n.a. n.a. use>subsidy n.a. Heilongjiang, China op>ip n.a. n.a. n.a. Hong Kong SAR use>subsidy India op>ip* non-h>ip* non-h>op* subsidy>use subsidy>use Indonesia op>ip non-h>ip* non-h>op use>subsidy* use>subsidy* use>subsidy Malaysia non-h>ip non-h>op use>subsidy* use>subsidy* n.a. Nepal non-h> (ip + op)a non-h> (ip + op)' use> subsidy Sri Lanka op>non-h use>subsidy use>subsidy n.a. Thailand non-h>ip non-h>op* subsidy>use Vietnam non-h>ip* non-h>op* subsidy>use* ip is inpatient, op is outpatient, non-h is non hospital. Blank cell indicates failure to reject the null hypothesis that curves are indistinguishable using the multiple comparison test at the 5 percent significance level. > indicates that the null hypothesis is rejected in favor of dominance, for example, op > ip indicates that outpatient care is more propoor than inpaticare and use > subsidy indicates that the use distribution is more propoor than the subsidy distribution. *indicates rejection of the null hypothesis of nondominance in favor of an alternative strict dominance in the direction indicated by >, as above, using the intersection-union test and a 5 percent significance level. aTest is between all hospital care (inpatient and outpatient) and all nonhospital care. Source: Authors' calculations based on survey data documented in table S.1 (see supplemental appendix S.1 available at http://wber.oxfordjournals. org/). 0lOZ 'LXjr-aqoj uo punjruP1uoW Ituoputuolul ju /'Jo-slunopojxoqm//:dBt[ uoij popolumo( O'Donnell and others 103 Younger (2000) but somewhat stronger, since the current study allows for more sources of heterogeneity in unit subsidies. Notwithstanding this result, there are significant differences between the distributions of use and subsidy. In Indonesia, Malaysia, and Sri Lanka the use distributions dominate-they are more propoor than the subsidy distributions-for all services, and in Gansu, Hong Kong SAR, and Nepal this is true for some services (see table 3). Dominance is not always found using the more conservative test, however. Urban-rural and regional differences in the quality of care are the most likely reason that the subsidy is less propoor than use. Only in India, Thailand, and Vietnam does the subsidy distribution domi- nate the use distribution for certain services, indicating that the subsidy per unit of care falls as household consumption rises. This is likely due to user payments rising with household consumption, whether because of exemptions granted to the poor or because richer households are paying for higher quality care that is not reflected in the unit cost figures. Cross-Country Comparisons As would be expected from the results already presented, the subsidy concen- tration curve of Hong Kong SAR dominates that of all other countries and provinces (table 4).6 The incidence of public care is so skewed toward the poor that the distribution of total healthcare (public and private) in Hong Kong SAR is propoor (Leung, Tin, and O'Donnell 2005).7 While this is in striking contrast with the distribution of healthcare in the low- and middle-income countries examined in this article, it is consistent with the distribution that pre- vails in most high-income economies (Van Doorslaer, Masseria, and Koolman 2006). There are no significant differences between the concentration curves of Malaysia, Sri Lanka, and Thailand, where the subsidies range from slightly propoor to evenly distributed. On the less strict test the Vietnamese distri- bution is dominated by that of Hong Kong SAR, Malaysia, and Thailand and it is indistinguishable from that of Sri Lanka. It dominates the subsidy distri- butions of all the remaining countries and provinces using the less stringent test.8 For most pairwise comparisons the subsidy concentration curves of Bangladesh, Gansu, Heilongjing, India, Indonesia, and Nepal are indistinguish- able. Exceptions are that India and Indonesia dominate Gansu and Nepal using the less strict test. In all these countries and provinces the public health subsidy is significantly and substantially prorich (see tables 1 and 2). This is 6. See table S-4 for cross-country dominance tests for each type of health service subsidy. 7. Some 43.5 percent of total expenditure on health in Hong Kong SAR is funded from private sources (Hong Kong Domestic Health Accounts 1999-2000). 8. This is not due simply to the fact that unit subsidies are negatively correlated with household consumption in Vietnam, unlike in most other countries and provinces. Only one cross-country dominance result for Vietnam becomes insignificant when use of each service rather than the subsidy to each service is examined. TABLE 4. Cross-Country Dominance of Public Health Subsidy Concentration Curves Malaysia Thailand Sri Lanka Vietnam Bangladesh Indonesia India Gansu Heilongjiang Nepal Hong Kong SAR D* D D* D D* D* D* D* D* D* Malaysia n.s. n.s. D D D* D* D* D* D* Thailand n.s. D D D* D* D D* D* Sri Lanka n.s. ns D D D D* D* n Vietnam D D* D D* D D* Bangladesh ns ns ns ns ns Indonesia ns D ns D India D ns D Gansu, China ns ns Heilongjiang, China ns n.s. indicates failure to reject the null hypothesis that the curves are indistinguishable using the multiple comparison test at the 5 percent significance level. D indicates rejection of the null in favor of dominance (more propoor) of the row country over the column country by the same test. *indicates that the intersection-union test rejects the null of nondominance against the alternative of strict dominance at the 5 percent significance level. Source: Authors' calculations based on survey data documented in table S.1 (see supplemental appendix available at http://wber.oxfordjournals.org/). 0lOt 'LXjr-Eqoj uo p-njrUP1uoW Ituoutuolul ju /'Jo-sunopojxo-jqm//:dt[ uioij poppolumo( O'Donnell and others 105 consistent with the findings of the majority of benefit incidence studies con- ducted in developing countries (van de Walle 1995; Castro-Leal and others 2000; Mahal and others 2000; Sahn and Younger 2000; Filmer 2003). But Malaysia, Thailand, Sri Lanka, and to a lesser extent Vietnam stand out as exceptions to this norm of prorich bias. Why is it that public healthcare is more propoor in these four countries than it is in other developing countries of Asia and elsewhere? National income is an obvious candidate to explain cross-country variation in the targeting of public health spending. Public healthcare is strongly targeted to the poor in Hong Kong SAR in large part because Hong Kong is rich enough to afford a dual system of universal public healthcare funded from general taxation and a private healthcare system used predominantly by the better-off to bypass the bottlenecks and inconveniences of the public system. It is surely no coincidence that Malaysia and Thailand are the only other two countries where public health spending is significantly propoor. While they are not nearly as rich as Hong Kong SAR, they are considerably better off than the other countries included in this study (see table S-5). Economic development is not the sole explanation for cross-country differ- ences in the incidence of public healthcare. It does not explain why Sri Lanka, despite a lower GDP per capita than Indonesia, achieves a distribution of health resources that is much more favorable to the poor. Levels of public spending on health and health system characteristics might be expected to explain part of the residual cross-country variation in targeting of the poor. In per capita terms Sri Lanka spends 2.5 times as much as Indonesia on public healthcare (table S-5). The scale of public spending may influence its incidence by affording a wider geographic distribution of public health facilities and so bring services closer to poor, rural populations. There may also be a trickle-down effect. At low levels of spending the politi- cally powerful, higher income urban elite may be more successful than the rural poor in capturing spending for programs that meet their own needs. As spending levels rise and more of the health needs of higher income groups are satisfied, additional programs can be better targeted to the needs of the poor (Lanjouw and Ravallion 1999). Countering this tendency, the pressure from higher income groups for prioritization of tertiary-level city hospitals may be maintained by the attraction of continuing advances in medical technology (Victora and others 2000). The extent to which higher income groups claim the benefits from public healthcare will depend on whether an attractive private sector alternative exists. Income-elastic demand for healthcare quality, in particular amenities and convenience of service, will lead to greater substitution of private for public care by an expanding middle-class as the economy grows. Hammer, Nabi, and Cercone (1995) argue that this mechanism was largely responsible for the increased propoor incidence of public health spending in Malaysia between the mid-1970s and the mid-1980s. The private sector continues to 106 THE WORLD BANK ECONOMIC REVIEW grow in Malaysia, driven in part by dissatisfaction with the responsiveness of the public system (Shepard, Savedoff, and Phua 2002). In Thailand, which has also achieved impressive economic growth in recent years, the private sector is also expanding rapidly (Towse, Mills, and Tangcharoensathien 2004). The combination of (near) universal public provision, a private sector offering an attractive alternative, and incomes that make demand for this alternative effective leads to redistribution through public provision in the way that theory predicts (Besley and Coate 1991). This mechanism implies a possibly uncomfortable tradeoff between the quality of public healthcare and the extent to which it is targeted to the poor. In lower income countries, such as Bangladesh, India, and Indonesia, separation of low- and high- income groups into the public and private sectors is constrained not only by the limited purchasing power of the middle class but also by marked intra- sectoral quality differentials. There is evidence of pronounced income gradi- ents in the quality of private sector care used in India (Das and Hammer 2005). There, as in Bangladesh, the poor make extensive use of unqualified private providers. This discussion suggests that economic development, the scale of public health spending, and the availability and quality of private sector alternatives may each help explain cross-country variation in the incidence of public health spending. Regression analysis is used to examine whether this is the case across the study countries and provinces and others for which benefit incidence results are available from other studies (Filmer 2003). Only 24 observations are avail- able for this analysis, and so the results (table 5) should be treated with due caution. It is an exploratory exercise and not an empirical test of hypotheses. The dependent variable is the (log) percentage of the total public subsidy received by the poorest quintile. This share increases significantly with GDP per capita, with an elasticity of about 0.3. At a lower level of significance (10 percent), the poorest quintile's share is also increasing with public health spending as a percentage of GDP, with an elasticity of about 0.5. So, for a given GDP there is some evidence that the share of the subsidy going to the poor is increasing with the scale of public health spending. To examine whether, for a given level of public expenditure, the share of the subsidy going to the poor increases with use of private sector alternatives, public spending as a percentage of total expenditure on health is included in the regression. Consistent with the hypothesis, the coefficient is negative but does not reach conventional levels of significance. The regression residuals are largest, in absolute value, for the two Chinese provinces. Public health spend- ing in these provinces is much less targeted on the poor than would be expected given GDP and the scale of public spending and its share of total health financing. This is most likely due to the extensive imposition of user charges with no income-related exemptions. Excluding these two provinces increases the magnitude and significance of the coefficients. In particular, the O'Donnell and others 107 TABLE 5. Cross-Country Regression Analysis of Targeting of the Public Health Subsidy (Dependent variable: log of percentage of public health subsidy received by poorest quintile) Excluding Gansu and Full sample Heilongjiang Robust standard Robust standard Coefficient errora Coefficient errora Log of gdp per capitab 0.3214*** 0.1002 0.3426*** 0.0889 Public health expenditure 0.2337* 0.1190 0.2971*** 0.0884 as percent of gdp Public health expenditure -0.0080 0.0049 -0.0110** 0.0043 as percent of total health expenditure Eastern Europe and -0.3308 0.2091 -0.4895* * 0.1889 Central Asia Latin America and -0.2478 0.3535 -0.4338 0.2990 Carribean Sub-Saharan Africa -0.8630*** 0.3004 -1.0750*** 0.2093 Constant 0.0691 0.7465 0.0294 0.7118 Sample size 24 22 R2 0.5712 p-value 0.7421 p-value RESET (F3,n-k-3) 0.76 0.5371 0.71 0.5671 *Significant at the 10 percent level; * *significant at the 5 percent level; " **significant at the 1 percent level. Note: Observations are the 11 countries and provinces for the years of this study plus those from Filmer (2003): Armenia (1999) Bangladesh (1995), Bulgaria (1995), Costa Rice (1992), Cote d'Ivoire (1995), Ecuador (1998), Georgia (2000), Ghana (1994), Guinea (1994), Honduras (1995), Nicaragua (1996), South Africa (1994), and Vietnam (1993). aRobust to heteroscedasticity of general form. bGross domestic product per capita in purchasing power parity dollars at constant 2000 prices. source: Dependent variable, authors' calculations based on data in table S. 1 (see supplemental appendix available at http://wber.oxfordjournals.org/) and that reported in Filmer (2003). GDP, World Bank, various years, World Development Indicators. Health expenditure, WHO, various years, National Health Accounts and World Health Report Statistical Annexes. negative coefficient on the public health financing share becomes significant at 5 percent.9 Although this study has found that the public health subsidy is not targeted on the poor in the majority of the 11 Asian countries and provinces examined, the distribution appears to be even more skewed toward the better- off in Eastern Europe and Central Asia and in Sub-Saharan Africa.10 9. The results are similar if the weight given to observations with large absolute residuals is reduced, but not set to zero, using robust regression. The results are also robust to the exclusion of Hong Kong SAR, where the subsidy is much more propoor and GDP is much higher than in the other countries and provinces. 10. Other potential explanatory factors, including the Gini coefficient, the urbanization rate, and the doctor supply rate, were not found to be significant. 108 THE WORLD BANK ECONOMIC REVIEW These regression results tell only of associations in a fairly small sample of countries and should not be interpreted as causal effects. GDP may be acting as a proxy for a number of primary determinants of incidence, such as the quality of governance and preferences for redistribution. Through human capital acquisition, assuming that the marginal product of investments in health is higher for poorer (and sicker) individuals, GDP may itself be responsive to the targeting of healthcare to the poor. Polices are of course endogenous. The positive correlation between the scale and the propoor incidence of public spending may derive from the degree of political commitment to reaching the poor. Reducing racial conflict in post-independence Malaysia was a major motivation for the expansion in access to healthcare and the channeling of public resources to the rural Malay population (Hammer, Nabi, and Cercone 1995). The early adoption of democracy and female suffrage in Sri Lanka contributed to the high priority given to healthcare and the wide geographical distribution of health resources in response to the lobbying of local politicians (McNay, Keith, and Penrose 2004). In fact, a 1928 commission proposed the full enfranchisement of women at the same time as men as a means of securing a political lobby for the prioritization of healthcare (Rannan-Eliya 2001). High rates of female literacy and a relatively high degree of female autonomy have raised awareness of maternal and child health problems, leading to high rates of use of modern health facilities and medicines (Caldwell 1986). Political and economic circumstances determine the motivation and resources for the pursuit of propoor public healthcare, but realization of the objective depends on the specific health sector policies adopted. One policy has been to minimize charges for poor patients in accessing care. There are virtually no fees for public health services in Sri Lanka, and fees are minimal in both Hong Kong SAR and Malaysia (table S-6). In all three cases fees are not retained by facilities or even by the health sector, but accrue to general revenues, thus undermining providers' incentives for generating fee revenue. The near avoidance of user fees in resource-poor Sri Lanka has been feasible only by driving down unit costs (Rannan-Eliya 2001). Nonmonetary incentives, such as professional development and opportunities to work simultaneously in the private sector, help maintain high levels of staff productivity. In Thailand fees have been much higher. Prior to the introduction of universal coverage in 2001, public hospitals received 20-50 percent of their revenue from user fees (Towse, Mills, and Tangcharoensathien 2004). But the disincentive effect on use by the poor was limited through a fairly effective healthcard scheme that covered about two-thirds of the poor. Crucially, this scheme compensated providers for fee exemptions from a designated budget. A geographically dispersed network of health facilities close to the rural population also appears to contribute to the propoor targeting of health spend- ing. In Malaysia half the population lives within 10 kilometers of a public O'Donnell and others 109 hospital and within 4.6 kilometers of a public clinic." In Sri Lanka most of the population has lived within 5 kilometers of a healthcare facility since the early 1970s, and most of the rural population is within 5-10 kilometers of a peripheral facility (Hsiao 2000). In Thailand, although beds and doctors are highly concentrated in Bangkok, an extensive rural infrastructure has been developed over decades. There are primary care health centers in all subdis- tricts and community hospitals in all districts (Towse, Mills, and Tangcharoensathien 2004). The introduction of universal coverage has initiated a major shift of resources from urban hospitals to primary care. Vietnam also has a relatively high level of provision in rural areas through a comprehensive network of commune health centers. But the contribution of primary care to propoor public health spending should not be exaggerated. Public health spending is better targeted on the poor in Hong Kong SAR, Malaysia, Thailand, Sri Lanka, and Vietnam because the distribution of hospital care is more favorable to the poor and not because more resources are devoted to nonhospital care (see table S-3). Of course, hos- pitals differ. In Malaysia and Sri Lanka many hospitals are small in scale and not particularly well equipped. But their wide geographic distribution makes them accessible to the rural poor. In many other low-income countries, such as Bangladesh, resources are more concentrated in large, well-equipped hospitals in urban centers that are inaccessible to the poor. III. CONCLUSION The analysis reveals substantial variation across Asia in the incidence of public subsidies for healthcare. Public spending is strongly propoor in high-income Hong Kong SAR. The total public health subsidy is more moderately propoor in low- to middle-income Malaysia and Thailand and it is evenly distributed in low-income Sri Lanka. At a still lower level of national income the subsidy is mildly prorich in Vietnam. In the remainder of the low-income countries and provinces examined, which account for the far greater share of the Asian popu- lation, the better-off receive substantially more of the subsidy than do the poor. In most cases there is prorich bias in the distribution of hospital care, while nonhospital care is propoor. A greater share of the healthcare subsidy goes to hospital care, and so this dominates the overall distribution. While public health subsidies are typically not propoor, they are inequality reducing in all cases except India and Nepal. Most within- and between-country dominance tests are robust to whether the distribution of healthcare use or the value of the subsidy is examined. This is a reassuring result since the health accounts data required for analysis of subsidy incidence are often unavailable and raw use data must be relied on. There are, however, significant differences between the distribution of 11. Authors' calculations from the 1996 National Health and Morbidity Survey. 110 THE WORLD BANK ECONOMIC REVIEW healthcare use and healthcare subsidies, with use often more propoor. Where this occurs, the likely explanation is urban-rural and interregional differences in the nature and funding of facilities. The analysis shows that the prorich distribution of public healthcare subsi- dies that is pervasive in most developing countries is avoidable but that effec- tive targeting is easier to realize at higher levels of national incomes. The experiences of Malaysia, Sri Lanka, Thailand, and Vietnam suggest that achiev- ing a more propoor incidence of public health spending requires limiting the use of user fees, or at least effectively protecting the poor from them; building a wide geographic network of health facilities; and ensuring that hospital care, which absorbs most spending, is sufficiently targeted at the poor. APPENDIX: THE INCIDENCE OF PUBLIC SPENDING ON HEALTHCARE: COMPARATIVE EVIDENCE FROM ASIA TABLE S1. Description of sample surveys Institution Survey Survey conducting Survey Sampling Response Sample size Country year name survey coverage Survey design unit rate individuals Bangladesh 1999- Health and Bangladesh Bureau National Stratified Household and 99% 56,010 2000 Demographic of Statistics Individual Survey (HDS) (BBS) 2000 Gansu 2003 National Health Ministry of Health Gansu province Stratified, cluster Household 100% 15,535 (China) Household (poor in sample. Self- Interview west China) weighting Surveys Heilongjiang 2003 Heilongfiang Health bureau of Heilongjiang Stratified, cluster Household 100% 11,572 (China) Health Heilongjiang province sample. Self- Household province (north-east weighting Interview Survey China) Hong Kong April- Thematic Census and National Stratified. Household 78.4% 31,672 SAR June Household Statistics Sample (noninstitutional; (noninstitutional); 2002 Survey in the Department, weights individual 97.2% second quarter Government of applied (institutional) (institutional) of 2002 Hong Kong SAR C India 1995-96 National Sample National Sample National Stratified, cluster Household 100% 629,024 Survey 52nd Survey sample. round Organisation Weights applied Indonesia 2001 Socioeconomic National Board of National Stratified, cluster Household 98% 889,413 Survey Statistics sampling. Self- (SUSENAS) weighted (Continued) El00Z'1jrqo AI nUo punjkXjPjouoW rtuoputuonq 11u /'Jo-slunopjojxo-jaqm//:I1lt[ mto.ij pppoIlumo( TABLE Si. Continued Institution 7 Survey Survey conducting Survey Sampling Response Sample size Country year name survey coverage Survey design unit rate individuals O Malaysia 1996 National Health Public Health National Stratified, cluster Household 86.90% 59,903 and Morbidity Institute, sample. Survey II Ministry of Weights Health applied Nepal 1995/96 Nepal Living Central Bureau of National Stratified, cluster Household 96.60% 18,855 Standards Statistics (CBS) sample. Survey Weights applied Sri Lanka 1996/97 Consumer Finance Central Bank Excluded Stratified Household 98% 399,28 Survey Northern Province due to civil war. Thailand Jan-June Socioeconomic National Statistical National Stratified Household 80% 17,489 2002 Survey Office Vietnam 1998 Living Standards General Statistical National Stratified, cluster Household 70% 28,623 Survey Office sample. Weights applied 0l0Z 'AXAr-qoj uo punjkLiPjouow ItuoiPlulu ju /B'JoslUu.nopJojxo-JoqM//:dflq[ mo.ij ppapEoluI TABLE S2. Measures of healthcare utilisation Hospital care Nonhospital care Polyclinic/health Inpatients outpatients Doctor visits center Antenatal care Comments Bangladesh Reference period last episode in previous 3 months last episode in previous 3 months 3months Care at satellite and community Measurement unit Number of days Number of Number of clinics also included but not visits visits child immunisation Gansu and Heilongjiang (China) Reference period 12months 2 weeks n.a. 2 weeks n.a. Data on hospital care only. Five Measurement unit Number of days Number of Number of levels of hospital are visits visits distinguished, the lowest of which are equivalent to polyclinics. Hong Kong SAR Reference period 12 months 30 days 30 days n.a. n.a. Hospital outpatient includes Measurement unit Number of days Number of Number of visits to specialist and A&E. visits visits Doctor visits is general outpatient visits. India Reference period 12 months 2 weeks 2 weeks 2 weeks 2 weeks Measurement unit Number of days any visits any treatment period any visits Indonesia Reference period 12 months 1 month n.a. 1 month 1 month Puskesmas (inpatients and a Measurement unit Number of days Number of n.a. Number of visits, Number of outpatients) and visits Number of days for visits supplementary Puskesmas inpatient (outpatients) included in health centre/polyclinic. Polindes and Posyandu in antenatal care. (Continued) 0lOZ 'LXjr-sqoj uo p-njrjPjouoW Ituopiuuu u /'Jo-sltunopJojxoxoqm//:dBqt umoij poppolumo(j TABLE S2. Continued Hospital care Nonhospital care Polyclinic/health Inpatients outpatients Doctor visits center Antenatal care Comments O Malaysia Reference period 12 months 2 weeks n.a. 2 weeks n.a. Measurement unit Number of Number of Number of visits admissions visits Nepal Reference period 30 days n.a. 30 days n.a. Data does not allow distinction 0 Measurement unit Number of n.a. Number of visits n.a. between hospital IP and OP visits Sri Lanka Reference period 2 weeks 2 weeks 2 weeks 2 weeks n.a. Measurement unit Any admission Any visit Any visit Any visit Thailand Reference period 12 months 1 month n.a. 1 month n.a. A distinction is made between Measurement unit Number of Number of Number of visits public and private care only admissions visits for the last 2 IP admissions and the last episode of other care. Assumed all care received in same sector Vietnam Reference period 12 months 4 weeks n.a. 4 weeks n.a. No distinction between public Measurement unit Number of days Number of Number of visits and private sector for IP care. visits Since vast majority of hospitals were public, assumed all IP is public IP inpatient. OP outpatient. n.a. not applicable. CJ OZ'LXjr-aqoj no p-njrUP1uoW Ituopiuulu ju /'Jo-slunopjojxoqm//:dt[ tuoij popolumo( TABLE S3. Summary indices of incidence of incidence of the public healthcare subsidy Hospital care Inpatient Outpatient Non-hospital care Total public subsidy Bangladesh Concentration index 0.2325 (0.1154) 0.1356 (0.0360) 0.0474 (0.0838) 0.1588 (0.0609) Kakwani index -0.1338 (0.0909) -0.2388 (0.0372) -0.3358 (0.0692) -0.2244 (0.0499) Subsidy share 47.99% 25.33% 26.69% 100% Gansu (China) Concentration index 0.2442 (0.0509) 0.1199 (0.0373) 0.1199 (0.0373) 0.1970 (0.0365) Kakwani index -0.2286 (0.0439) -0.3529 (0.0360) -0.3529 (0.0360) -0.2758 (0.0332) Subsidy share 65.42% 34.58% 34.58% 100% Heilongjiang (China) Concentration index 0.03232 (0.0605) 0.2192 (0.0474) 0.2192 (0.0474) 0.2527 (0.0385) Kakwani index -0.1242 (0.0652) -0.2281 (0.0510) -0.2281 (0.0510) -0.1946 (0.0424) Subsidy share 60.09% 39.91% 39.91% 100% Hong Kong SAR Concentration index -0.3193 (0.0355) -0.2762 (0.0264) -0.2444 (0.0232) -0.3104 (0.300) Kakwani index -0.6919 (0.0356) -0.6491 (0.0265) -0.6173 (0.0232) -0.6831 (0.0301) Subsidy share 82.47% 13.36% 4.17% 100% India Concentration index 0.2630 (0.0193) 0.00296 (0.0211) -0.1325 (0.0328) 0.2117 (0.0164) Kakwani index 0.0122 (0.01928) -0.2476 (0.02113) -0.3830 (0.03281) -0.0390 (0.0165) O Subsidy share 83.68% 9.62% 6.65% 100% O Indonesia Concentration index 0.4896 (0.0254) 0.3891 (0.0186) -0.0078 (0.0045) 0.1822 (0.0081) Kakwani index 0.1752 (0.0248) 0.0880 (0.0187) -0.3142 (0.0047 -0.1245 (0.0080) Subsidy share 26.54% 14.86% 58.59% 100% Malaysia Concentration index -0.0416 (0.0124) -0.0165 (0.0231) -0.2410 (0.0181) -0.0807 (0.0116) Kakwani index -0.4100 (0.0131) -0.3863 (0.0235) -0.3863 (0.0235) -0.4493 (0.0123) (Continued) Cj0Z~'L,1jv-aqoj uo punjrjPjuoW tuop~utuoutl u /'Jo-slunopjojxo-jaqm//:dBt[ uoij popolumo( TABLE S3. Continued a Hospital care Inpatient Outpatient Non-hospital care Total public subsidy Subsidy share 37.02% 38.53% 24.45% 100% Nepal Concentration index 0.3422 (0.0709) 0.3422 (0.0709) 0.1865 (0.0411) 0.2541 (0.0398) Kakwani index 0.1268 (0.0605) 0.1268 (0.0605) -0.0677 (0.0487) 0.0384 (0.405) Subsidy share 54.58% 54.58% 45.24% 100% Sri Lanka Concentration index 0.0220 (0.0377) -0.0486 (0.0304) -0.0486 (0.0304) -0.0020 (0.0269) Kakwani index -0.3313 (0.0252) -0.4042 (0.0172) -0.4042 (0.0172) -0.3561 (0.0284) Subsidy share 68.00% 32.00% 32.00% 100% Thailand Concentration index -0.0242 (0.0308) -0.0392 (0.0227) -0.2506 (0.0325) -0.0404 (0.0195) Kakwani index -0.4199 (0.0317) -0.4348 (0.0242) -0.6463 (0.0335) -0.4361 (0.0210) Subsidy share 50.74% 45.16% 4.18% 100% Vietnam Concentration index 0.0354 (0.0359) 0.1672 (0.0349) -0.1065 (0.0272) 0.0114 (0.0283) Kakwani index -0.1495 (0.0471) -0.0599 (0.0667) -0.4623 -0.2573 (0.0458) Subsidy share 86.88% 2.13% 10.98% 100% Robust standard errors in parentheses. Source: Authors' calculations from data documented in table S-1. 00ZO~'L,1jv-qoj uo punjrjPjuoIt uoiluuolul ju /'Jo-slunopjojxo-jaqm//:dt[ uoij popolumo( TABLE S4. Cross-country Dominance of Public Health Subsidy Concentration Curves Total subsidy Malaysia Thailand Sri Lanka Vietnam Bangladesh Indonesia Gansu India Heilongjiang Nepal Hong Kong SAR D* D* D* D D* D* D* D* D* D* Malaysia n.s. n.s. D D D* D* D* D* D* Thailand n.s. D D D* D D* D* D* Sri Lanka n.s. n.s. D D D* D* D* Vietnam D D* D* D D D* Bangladesh n.s. n.s. n.s. n.s. n.s. Indonesia D n.s. n.s. D Gansu (China) n.s. n.s. India D n.s. D Heilongjiang (China) n.s. Hospital inpatient subsidy Malaysia Thailand Sri Lanka Vietnam Bangladesh Gansu India Heilongjiang Nepal Indonesia Hong Kong SAR D* D* D* D D D* D* D* D* D* Malaysia n.s. n.s. D n.s. D* D* D* D* D* Thailand n.s. D n.s. D D* D D* D* Sri Lanka n.s. n.s. D D* D D* D* Vietnam n.s. D D D D* D Bangladesh n.s. n.s. n.s. n.s. n.s. Gansu (China) n.s. n.s. n.s. n.s. India n.s. D D* Heilongjiang (China) n.s. n.s. Nepala Indonesia D Hospital outpatient subsidy Sri Lanka Thailand Malaysia India Gansu Bangladesh Vietnam Heilongjiang Indonesia Hong Kong SAR D* D* D D* D* D D* D* D* Sri Lanka n.s. n.s. n.s. D D D* D D* Thailand n.s. n.s. D D D D D* (Continued) -A 0lOZ 'L,1j,a-iqoj uo punjkjPjuoW Ituoilsuujolul ju /'Jo-slunopjojxo-jqm//:dt[ uioij poppolumo( 00 TABLE S4. Continued Total subsidy Malaysia Thailand Sri Lanka Vietnam Bangladesh Indonesia Gansu India Heilongjiang Nepal Malaysia n.s. D D D* D D* India D D D* D D* Gansu (China) n.s. n.s. n.s. D Bangladesh n.s. n.s. D Vietnam n.s. D o Heilongjiang (China) D Non-hospital subsidy Hong Kong Malaysia India Vietnam Indonesia Bangladesh Nepal Thailand D D D D D* D D* Hong Kong SAR n.s. D D D* D* D* Malaysia D D D D D* India n.s. D* D D* Vietnam D D D* Indonesia n.s. D* Bangladesh n.s. Note: Countries/provinces are ranked from most to least propoor according to values of concentration indices. Tests follow the multiple comparison approach with the null hypothesis defined as curves being indistinguishable. n.s. indicates failure to reject the null at 5% significance. D indicates that the subsidy concentration curve of the row country/province dominates (is more pro-poor) than that of the column country/province. There are no cases of crossing concentration curves. *indicates that the intersection union principle test rejects the (different) null of nondominance against the alternative of strict dominance at 5%. If no *appears, then this test does not reject its null. acomparison with Nepal are for the aggregate of inpatient and outpatient subsidies. 0lOZ 'L,1j,a-iqoj uo punjrjP1uoW Ituop~uuojul ju /.Jostiopojojq/:t uoij popolumo( TABLE S5. National Income and Government Expenditure on Health General government General government General government expenditure GDP per capita, expenditure on expenditure on health on health as % total Territory Yeara ppp $b health as % GDPC per capital, PPP $ expenditure on health Bangladesh 1999 1495 0.98 15 27 China 2002 4568 2.26 103 42 Gansu (China) 2002 2661 2.38 63 42 Heilongjiang 2002 5434 1.48 80 36 (China) Hong Kong 2001/02 26049 3.26 849 57 SAR India 1996 1994 0.81 16 16 Indonesia 2001 3146 0.57 18 36 Malaysia 1996 8254 1.34 111 58 Nepal 1995/96 1179 1.20 14 24 Sri Lanka 1996/97 2951 1.63 48 50 Thailand 2000 6740 2.04 138 61 Vietnam 1998 1854 1.44 27 33 aYear of survey used for distributional analysis. bGDP per capita in international $ using purchasing power parity (PPP) exchange rates. Constant year 2000 prices. cGeneral government expenditure on health including social insurance. Source: GDP per capita-World Development Indicators, World Bank. Health expenditures-National health accounts estimates, except: India, Malaysia and Vietnam from World Health Report, Stastistical Annexes, WHO, and Nepal from (HMG/Nepal 2000 and Hotchkiss, Rous and others 1998). 00ZO~'LXjr-mqoj uo p-njrPjuoIt uoiluuolul ju /'Jo-slunopojxo-jqm//:dBt[ uoij poppolumo(I TABLE S6. Charges and exemptions for public healthcare Nonpoor groups exempt from Charged services Free Services Income/poverty related fee waivers charges Bangladesh Secondary services (nominal Most primary care (or local Poor exempt or pay lower charge Civil servants (selected services) registration fee for inpatient/ services); medicines within outpatient); Inpatient care in facility; immunization; some major hospitals reproductive healthcare China Inpatient (including etc Family planning None Old Red Army soldiers and medicines); Outpatient Retirees (including medicines); Immunisation Hong Kong Inpatient (including medicines); Accident and emergency (until Welfare recipients exempt Civil servants and dependents SAR outpatient (including December 2002) (reduced rate for Inpatients); medicines); dental hospital staff and dependents India Inpatient bed charge; Hospital consultation and certain None formally. Indirect relation to Civil sevants outpatient registration medicines. Primary care/health income through price charge; certain medicines; center/polyclinic consultation differentiation in inpatient care. tests/x-rays; dental and medicines. Family planning. Informally, "poor" can be Vaccinations and immunizations exempted partially or fully from charges Indonesia All medical care and medicines None Poor exempt from all charges. Charges determined at local Indirect relation of inpatient government level. Some better charges to income through price off local govts. Provide free discrimination health centre care Malaysia Hospital inpatient and Family planning and vaccinations/ Hospital directors have discretion Infants less than 1 year outpatient. Primary care. immunizations. Outpatient ante to waive fees for destitute. (outpatient). State rulers, Dental care. Diagnostics and and postnatal care. Treatment of Upper limit on charges for third Governors and families. Civil x-rays infectious diseases on third class class ward patients servants (including retired) wards, Dental care for pregnant and dependents. Local women and pre school children authority employees and dependents 00ZO~ 'LXjrruqoj uo punjr/PuocInuoiuoulaiun /BJoslwnopJojxoxnqm//:dBt[ uoij popolumo Nepal All medical care and medicines. Emergency services; selected Poor either exempt or pay reduced None Nominal charge for vaccines, immunization and charge but not fully outpatient varying with reproductive health services. implemented. facility. 60% subsidy for medicines at Health Posts and Primary Care centres. Sri Lanka Family planning services. All medical and medicines except No official exemptions, but limited None Patients occasionally asked family planning. survey evidence suggests that to buy medicines/supplies facility staff tend to avoid asking from private retailers when the poorest patients to self- out of stock at facility. purchase medicines and supplies, or ration available stocks to them. Thailand All medical care and medicines. Nonpersonal healthcare; EPI Poor exempted from user fees and children <12; elderly >60; After Oct 2001, fixed fee (30 vaccination co-payments. Informally, those public health volunteers; Baht) UC scheme means very "unable to pay" are exempted. monks. minimal co-payment. Vietnam Fees for most services Outpatient services at commune Fee exemptions for individuals Families of health personnel, introduced in 1989. health centres. who have certification of certain classes of patients (like Medicines rarely provided indigency from neighbourhood handicapped, TB), orphans. free of charge. or village People's Committee. 00Z~'LXA.r-nqoj uo punjrjPjuoW Ituoil1uuu u /'Jo-sltunopojxoxoqm//:dl[ uioij poppolumo( 122 THE WORLD BANK ECONOMIC REVIEW REFERENCES Beach, Charles M., and James Richmond. 1985. "Joint Confidence Intervals for Income Shares and Lorenz Curves." International Economic Review 26:439-50. Besley, Timothy, and Stephen Coate. 1991. "Public Provision of Private Goods and the Redistribution of Income." American Economic Review 81(4):979-84. Bishop, John A., K. Victor Chow, and John P. Formby. 1994. "Testing for Marginal Changes in Income Distributions with Lorenz and Concentration Curves." International Economic Review 35(2):479-88. Bishop, John A., John P. Formby, and Paul D. Thistle. 1992. "Convergence of the South and Non-South Income Distributions, 1969-1979." American Economic Review 82(1):262-72. Caldwell, John C. 1986. "Routes to Low Mortality in Poor Countries." Population and Development Review 12(2):171-220. Castro-Leal, Florencia, Julia Dayton, Lionel Demery, and Kalpana Mehra. 2000. "Public Spending on Health Care in Africa: Do the Poor Benefit?" Bulletin of the World Health Organisation 78(1):66-74. Dardanoni, Valentino, and Antonio Forcina. 1999. "Inference for Lorenz Curve Orderings." Econometrics Journal 2(1):49-75. Das, Jishnu, and Jeffrey Hammer. 2005. "Which Doctor? Combining Vignettes and Item Response to Measure Clinical Competence." Journal of Development Economics 78(2):348-83. Davidson, Russell, and Jean-Yves Duclos. 1997. "Statistical Inference for the Measurement of the Incidence of Taxes and Transfers." Econometrica 65(6):1453-65. Deaton, Angus. 1997. The Analysis of Household Surveys: A Microeconometric Approach to Development Policy. Baltimore, MD: Johns Hopkins University Press. Filmer, Deon. 2003. "The Incidence of Public Expenditures on Health and Education." Background note for the World Development Report 2004: Making Services Work for the Poor. Washington, D.C.: World Bank. Gwatkin, Davidson R., Shea Rustein, Kiersten Johnson, Roshni Pande, and Adam Wagstaff. 2003. Initial Country-Level Information about Socio-Economic Differentials in Health, Nutrition and Population, vols.I and II. Washington, D.C.: World Bank, Health, Population, and Nutrition Division. Hammer, Jeffrey S., Ijaz Nabi, and James A. Cercone. 1995. "Distributional Effects of Social Sector Expenditures in Malaysia, 1974-89." In Walle Domenique van de, and Nead Kimberly eds., Public Spending and the Poor. Baltimore, MD: Johns Hopkins University Press. Howes, Stephen. 1996. "A New Test for Inferring Dominance from Sample Data." World Bank, Washington, D.C. Hsiao, William. 2000. "A Preliminary Assessment of Sri Lanka's Health Sector and Steps Forward." Cambridge, Mass.: Harvard School of Public Health. Kaur, Aamrjot, B. L. S. Prakasa Rao, and Harshinder Singh. 1994. "Testing for Second-Order Stochastic Dominance of Two Distributions." Econometric Theory 10(5):849-66. Lanjouw, Peter, and Martin Ravallion. 1999. "Benefit Incidence, Public Spending Reforms and the Timing of Program Capture." World Bank Economic Review 13(2):257-73. Leung, Gabriel M., Keith Tin, and Owen O'Donnell. 2005. "Redistribution or Horizontal Equity in Hong Kong's Mixed Public-Private Health System: A Policy Conundrum." EQUITAP Working Paper 22, Erasmus, Rotterdam, Netherlands, and Institute for Policy Studies, Colombo. Mahal, Ajay, Janmejaya Sing, Farzana Afridi, Vikram Lamba, Anil Gumber, and V Selvaraju. 2000. "Who Benefits from Public Health Spending in India?" New Delhi: National Council for Applied Economic Research. McNay, Kirsty, Regina Keith, and Angela Penrose. 2004. Bucking the Trend: How Sri Lanka Has Achieved Good Health at Low Cost-Challenges and Policy Lessons for the 21st Century. London: Save the Children. O'Donnell and others 123 Meerman, Jacob. 1979. Public Expenditure in Malaysia: Who Benefits and Why? New York: Oxford University Press. Rannan-Eliya, Ravindra P. 2001. "Strategies for Improving the Health of the Poor-the Sri Lankan Experience." Health Policy Programme, Institute of Policy Studies, Colombo. Sahn, David E., and Stephen D. Younger. 2000. "Expenditure Incidence in Africa: Microeconomic Evidence." Fiscal Studies 21(3):321-48. Sahn, David E., Stephen D. Younger, and Kenneth R. Simler. 2000. "Dominance Testing of Transfers in Romania." Review of Income and Wealth 46(3):309-27. Shepard, Donald S., William Savedoff, and Kai Hong Phua. 2002. Health Care Reform Initiatives in Malaysia: Report of a Consultation with the Planning and Development Division, Ministry of Health, Malaysia. Waltham, Mass.: Brandeis University, Shneider Insitute for Health Policy. Towse, Adrian, Anne Mills, and Viroj Tangcharoensathien. 2004. "Learning from Thailand's Health Reforms." British Medical journal 328:103-05. van de Walle, Domenique. 1995. "The Distribution of Subsidies through Public Health Services in Indonesia, 1978-87." In Dominique van de Walle, and Nead Kimberly eds., Public Spending and the Poor. Baltimore, MD: Johns Hopkins University Press. van de Walle, Domenique. 1998. "Assessing the Welfare Impacts of Public Spending." World Development 26(3):365-79. Van Doorslaer, Eddy, and Cristina Masseria, Xander Koolman and the OECD Health Group Equity Research. 2006. "Inequalities in Access to Medical Care by Income in Developed Countries." Canadian Medical Association Journal 174(2):177-83. Victora, C. G., J. P. Vaughan, F. C. Barros, A. C. Silva, and E Tomasi. 2000. "Explaining Trends in Inequalities: Evidence from Brazilian Child Health Studies." The Lancet 356(9235):1093-98. World Bank. 2001. Vietnam: Growing Healthy-A Review of Vietnam's Health Sector. Hanoi. World Bank. Various years. World Development Indicators. Washington, D.C. World Health Organization. Various years. World Health Report: Statistical Annexes. Geneva. Did the Health Card Program Ensure Access to Medical Care for the Poor during Indonesia's Economic Crisis? Menno Pradhan, Fadia Saadah, and Robert Sparrow The Indonesian Social Safety Net health card program was implemented in response to the economic crisis that hit Indonesia in 1997, to preserve access to health care ser- vices for the poor. Health cards were allocated to poor households, entitling them to subsidized care from public health care providers. The providers received budgetary support to compensate for the extra demand. This article focuses on the effect of the program on primary outpatient health care use, disentangling the direct effect of allo- cating health cards from the indirect effect of government transfers to health care facilities. For poor health card owners the program resulted in a net increase in use of outpatient care, while for nonpoor health card owners the program resulted mainly in a substitution from private to public health care. The largest effect of the program seems to have come from a general increase in the supply of public services resulting from the budgetary support to public providers. These benefits seem to have been cap- tured mainly by the nonpoor. As a result, most of the benefits of the health card program went to the nonpoor, even though distribution of the health cards was propoor. The results suggest that had the program, in addition to targeting the poor, established a closer link between provision of services to the target groups and funding, the overall results would have been more propoor. JEL codes: H51, Il, 138. In the current debate on the provision of health care services in developing countries, many researchers have found high inequalities in the use of public health care and hence in the benefit incidence of public spending. Menno Pradhan is a senior poverty specialist at the World Bank in Jakarta; his email address is mpradhan@worldbank.org. Fadia Saadah is a sector manager for health, nutrition, and population in the East Asian and Pacific Region at the World bank; her email address is fsaadah@worldbank.org. Robert Sparrow is a lecturer in development economics at the Institute of Social Studies, The Hague; his email address is sparrow@iss.nl. The work for this article was done while Robert Sparrow was at the Vrije Universiteit Amsterdam and the Tinbergen Institute. Support from the Netherlands Foundation for the Advancement of Tropical Research (WOTRO) is gratefully acknowledged. The authors thank Jan Willem Gunning, Maarten Lindeboom, Aparnaa Somanathan, Dominique van de Walle, three anonymous referees, and seminar participants at the World Bank, Tinbergen Institute, Institute of Social Studies, and 2001 International Health Economics Association and Global Development Network conferences for comments on earlier drafts. A supplemental appendix to this article is available at http:/ wber.oxfordiournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 125-150 doi:10.1093/wber/1hl010 Advance Access Publication 12 January 2007 () The Author 2007. 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 125 126 THE WORLD BANK ECONOMIC REVIEW Supply-driven public policy typically lacks incentives for health care providers to serve the poor. Ensuring that the poor benefit from health care and receive a basic package is a widely shared policy objective (World Bank 2004). Targeted price subsidies for medical care are often advocated to increase access to medical care for the poor. However, there is little empirical evidence about the efficacy of such systems. This case study looks at a particular kind of health care intervention that was implemented in Indonesia, which included both a targeted price subsidy and a public spending component. This combined program was part of a larger Social Safety Net (SSN) program, initiated in 1998 to protect the poor from the effects of the Southeast Asian economic crisis.' Households that were thought to be most vulnerable to economic shocks were allocated health cards, which entitled household members to the price subsidy. Health care facilities that provided the subsidized care received extrabudgetary support to compen- sate for the increased demand. There are some distinct features to the SSN health program. First, the price subsidy applied only to public service providers. Second, the program followed a decentralized design, with both geographic and community-based targeting. Third, there was a weak link between use of the health card and compensation of health care providers. Compensation was allocated to districts based on the estimated number of households eligible for the health card program and not on actual use of health cards. This article focuses on the effect of the Indonesian health card program on demand for primary outpatient health care. The design of the program makes it possible to investigate several interesting questions. First, it provides the opportunity for an ex post evaluation of a targeted price subsidy for health care that was implemented on a national scale. There are relatively few empiri- cal studies that evaluate actual pricing policies in health care. Of those that do, only a handful take account of the endogenous nature of public interventions in their estimation strategy.2 Most studies that discuss the effectiveness of health policy draw on health care demand models that make ex ante predic- tions of possible policy scenarios. The drawback of these simulations is that the underlying estimates are often based on cross-sectional data that typically show little (spatial) variation. In effect, these simulations concern forecasted 1. The program also included an education program, a labor creation program, and food assistance. See Daly and Fane (2002) for an overview of the programs. 2. Using data from a randomized health insurance and cost sharing experiment in the United States, Manning and others (1987) estimate the demand for outpatient care. Gertler and Molyneaux (1997) use panel data to evaluate an experiment with a user fee increase for outpatient services in Indonesia. The Medicaid program in the United States is probably the most studied targeted health care subsidy program. Currie and Gruber (1996) and Currie and Thomas (1995) exploit variation in legislature across states to control for endogeneity of the program. In an analysis of a school-based health insurance scheme in Egypt, Yip and Berman (2001) treat participation as selection on observed characteristics. Pradhan, Saadah, and Sparrow 127 interventions that lie outside the range of the observed price data (Gertler and Hammer 1997). Second, since the health card entitles users to free services only at public providers, substitution effects between private and public providers can be investigated directly. This is difficult in health care demand studies since infor- mation on the price menu offered by alternative health care providers is often unavailable. As an alternative to exogenous price data many models estimate the demand for medical care based on proxy variables derived from (endogen- ous) household expenditure data (Gertler, Locay, and Sanderson 1987; Lavy and Quigley 1993; Ching 1995; Mocan, Tekin, and Zax 2000) or variations in indirect cost measures, such as opportunity costs due to loss of work or travel time to the nearest provider (Gertler and van der Gaag 1990; Dow 1999). However, opportunity costs do not vary by public or private provider. Studies that do manage to identify price variation across provider types generally find substantial substitution effects between public and private providers as a result of public price policy (Mwabu, Ainsworth, and Nyamete 1993; Sahn, Younger, and Genicot 2003). The third contribution of this article is that it evaluates the impact of the public spending component of the program and compares its magnitude to that of the targeted price subsidy. Empirical evidence is inconclusive on the causal- ity between spending and health outcomes (World Bank 2004; Filmer and Pritchett 1999). There are empirical studies using provider or community data that show a positive effect of supply and quality of care (especially drug avail- ability) on use (Lavy and Quigley 1993; Mwabu, Ainsworth, and Nyamete 1993; Lavy and Germain 1994; Akin, Guilkey, and Denton 1995; Akin and others 1998; Sahn, Younger, and Genicot 2003). The problem with these quality and supply variables is that they are often endogenous due to govern- ment policy, and the measured effects are likely to capture both supply and demand effects. Although some studies manage to control for the endogeneity problem, it is much harder to control for the second effect. This article identifies effects of both the price subsidy and of the budgetary support on health care use and shows that the largest share of the program's effect is due to increased public spending. The effects of the price subsidy and the supply impulse differ by income group. For low-income groups there is both a substitution from private to public care and an increase in total use because of the health card, but little effect from increased spending. For the more wealthy groups the substitution effect dominates, and the supply-induced effect of the budget increase is larger, possibly since the rich typically face fewer barriers to access to medical care than the poor do. Overall, the nonpoor captured most of the benefit, despite the propoor targeting of the health cards, because of the weak link between financing and use of health cards. The next section gives an overview of the data. Section II describes the health card program in more detail. Section III focuses on the evaluation problem and the strategy for estimating the impact of the health card on use of 128 THE WORLD BANK ECONOMIC REVIEW health care services. The results are discussed in section IV, and section V high- lights some caveats and examines the sensitivity of the results to the main assumptions of the study. I. THE DATA The study is based on data from Indonesia's nationwide socioeconomic house- hold survey (susenas) conducted by the Indonesian Bureau of Statistics. The 1999 survey round contained a module on the use of SSN interventions, includ- ing the health card program. The health card program started in September 1998, and the survey was fielded in February 1999. The survey data therefore reflect the experience of the first months of operation. For this reason, and other data limitations, the analysis here is limited to the impact of the program on access to medical care (in terms of use), and no effort is made to estimate the effect on health. The survey sampled 205,747 households and collected a wide range of socioeconomic indicators along with a measure of consumption. In the area of health the survey collected information on self-reported illness, use of health care services, user fees, and ownership and use of the health card. Data from the 1998 household survey were used to provide pre-intervention data for the analysis. This round, also fielded in February, includes 207,645 households and covers the same questionnaire and variables as the 1999 survey, except for the SSN module. A 1996 village-level census (podes) provides pre-intervention information on accessibility and supply of health services and on various other community characteristics. The 1996 podes includes 66,486 villages and urban precincts and can be merged with the national household survey. Administrative data concerning the 1998/99 budget for the SSN program were also used. These data include the budget allocated to 293 districts to implement the health card program and to compensate the public health clinics and village midwives for the expected extra demand for health services result- ing from the health card program. The largest share of this budget was trans- ferred directly to public health care providers. The transfers were made in two to four phases, depending on the province, starting in the last quarter of 1998. By the time of the survey SSN budgets had arrived at the health centers. 11. USE OF HEALTH CARE SERVICES AND THE HEALTH CARD PROGRAM The economic crisis hit Indonesia in the fall of 1997, exacerbated by social and political unrest in 1998. Real GDP dropped roughly 15 percent in 1998 causing poverty to rise sharply. Suryahadi, Sumarto, and Pritchett (2003) esti- mate an increase in the poverty headcount ratio from 15 percent in May 1997 to 33 percent at the end of 1998. The consumer price index rose 78 percent in 1998. The price of food doubled, with rice and staple foods experiencing the Pradhan, Saadah, and Sparrow 129 most severe increase. There is little evidence of rising overall unemployment during the crisis. Instead, real wages dropped about 40 percent in the formal sector during the first year of the crisis, whereas agriculture seems to have absorbed part of the displaced labor from other sectors (Cameron 1999; Smith and others 2002; Frankenberg, Smith, and Thomas 2003). The severity of the crisis undoubtedly affected households' health care use and expenditures. Frankenberg, Smith, and Thomas (2003) find that household con- sumption declined by 20 percent in 1998, with investment in human capital (health care and education) decreasing 37 percent. Data from household surveys on use of modern health care in February of each year for several years before and during the crisis indicate a sharp decrease in the use of modern health care from 1997 to 1998, due largely to declining use of public sector providers (table 1). Waters, Saadah, and Pradhan (2003) attribute the decline to a worsening in the quality of public sector providers. The deterioration was due mainly to the growing shortage of drugs and supplies at public facilities during the crisis, especially in rural areas (Frankenberg, Thomas, and Beegle 1999; Knowles, Pernia, and Racelis 1999). From 1998 to 1999 total use of modern health care providers remained the same, but the share of the public sector increased. One possible explanation for the change is the introduction of the health card program.4 Under the SSN health program the allocation of health cards and funds is delegated to lower administrative levels. The amount of subsidy for public health care providers to be distributed across districts, along with the number of health cards to be issued, is determined by a pre-intervention poverty esti- mate. This poverty measure is constructed by the national family planning board [Badan Koordinasi Keluarga Berencana Nasional (BKKBN)] and counts the number of poor households per district based on "prosperity status." Under this definition a household is deemed poor when it has insufficient funds for any one of the following: to worship according to its faith, to eat basic food twice a day, to have different clothing for school or work and home, to have an improved floor (not made of earth), or to have access to modern medical care for children or access to modern contraceptives. The BKKBN reg- ularly collects this information on a census basis.5 3. Modern health care is defined as public health care providers-hospitals, health clinics (puskesmas), village maternity posts (polindes), and integrated health posts (posyandu)-and private providers-hospitals, doctors, clinics, and paramedical services. Traditional health care is not included. 4. Another explanation for the dip in 1998 would be that households postponed preventive care, in anticipation of introduction of the health card. But this is unlikely because the program had not been announced at the time of the 1998 survey. 5. Use of the BKKBN prosperity measure results in a higher poverty headcount (42 percent of households in December 1997) than use of consumption-based measures [24 percent of population in February 1998, according to estimates by Suryahadi, Sumarto, and Pritchett (2003) based on the susenas consumption module for 1998]. The BKKBN measure has been criticized for being an unsuitable allocation criterion for the SSN, since its components are fairly inflexible and inappropriate for measuring economic shocks or the impact of a crisis. However, at the time of implementation it was the only up-to-date welfare measure at hand. 130 THE WORLD BANK ECONOMIC REVIEW TABLE 1. Changes in Outpatient Contact Rates for Public and Private Care with SSN Program, 1995-99 (percentage of population that visited provider at least once in previous month) 1999 without Provider 1995 1997 1998 1999 SSN program Public 7.00 (0.083) 6.65 (0.085) 5.03 (0.062) 5.34 (0.071) 4.87 Private 6.48 (0.073) 6.71 (0.079) 6.11 (0.070) 5.80 (0.078) 5.67 Overall outpatient 12.83 (0.111) 12.83 (0.118) 10.48 (0.098) 10.53 (0.110) 9.98 care (public or private)a Number of 873,647 887,266 880,040 864,580 observations Note: Numbers in parentheses are standard errors. aThe contact rate for all modern care is smaller than the sum of the contact rates for public and private care since individuals who sought both public and private care are counted only once in the aggregate. Source: Authors' analysis based on data from Indonesia's annual susenas household survey; see description in text. At the district level committees were formed to deal with the allocation of funds to the health clinics and village midwives. The allocation was based on the BKKBN estimate of poor households eligible for a health card in the village or subdistrict served by each public provider rather than on actual services pro- vided to health card owners. The district committee allocated health cards to villages, again based on the BKKBN measure, for distribution by village com- mittees headed by local leaders. Along with the health cards the village com- mittees received guidelines on which criterion to use when selecting households for the health card program. Besides households that were classified as poor by the BKKBN, the village committees were to consider households that were severely affected by the crisis. Local leaders maintained considerable discretion to distribute health cards according to their own insights, however. Health cards were usually distributed through local health centers and village midwives. The health card entitled the owner and family members to free services at public health care providers consisting of outpatient and inpatient care, con- traceptives for women of child-bearing age, prenatal care, and assistance at birth. A health card was not transferable to other households. The analysis here looks only at the impact of the health card program on outpatient health care use. By February 1999 the health card program was already of substantial size, with 10.6 percent of Indonesians reporting that their household had been allo- cated a health card. The share rose to 18.5 percent among individuals from the Pradhan, Saadah, and Sparrow 131 TABLE 2. Distribution of Health Card Allocation and Use, by per Capita Consumption Quintile, Gender, and Urban or Rural Residence (percent) Characteristic Coverage Share in allocation Share in use Quintile 1 (poor) 18.5 33.7 31.3 Quintile 2 13.7 25.7 24.4 Quintile 3 10.6 20.1 20.4 Quintile 4 7.1 13.4 14.9 Quintile 5 (rich) 3.7 7.1 9.0 Male 10.5 49.3 43.8 Female 10.8 50.7 56.2 Urban 7.2 26.8 29.5 Rural 12.8 73.2 70.5 Indonesia 10.6 100.0 100.0 Note: Number of observations is 822,607. Source: Authors' analysis based on data from Indonesia's annual susenas household survey; see description in text. poorest 20 percent of the population (table 2) and 13.7 percent among those in the second poorest quintile (about half of whom were estimated to live below the poverty line). There was considerable leakage to more wealthy households, however. Whereas the poorest 20 percent of the population own 34 percent of the health cards, households from the wealthiest 60 percent of the population possess about 40 percent of the health cards. Use of health cards is also propoor but slightly less so. Those who received benefits were on average weal- thier than those who received the card. Use of outpatient care is higher among households that own a health card, especially, use of public services (table 3). Overall, 12 percent of health card owners visited an outpatient provider in the month before the survey compared with 10 percent of those without a health card. TABLE 3. Use of Health Card (percentage that sought care at least once in previous month) Head of household reports Head of household having received a reports not having received Use characteristic health card a health card Received outpatient care 12.4 10.4 Went to public provider 8.2 5.0 Went to private provider 5.0 5.9 Number of observations 81,126 741,481 Source: Authors' analysis based on data from Indonesia's annual susenas household survey; see description in text. 132 THE WORLD BANK ECONOMIC REVIEW III. IMPACT OF HEALTH CARD PROGRAM ON USE OF HEALTH CARE SERVICES What would the use of outpatient health care services have been if the health card program had not existed? This question incorporates two effects: the effect of the health card price subsidy and the effect of the additional budget- ary resources made available to public sector services through the SSN program. Because of the weak link between these two components of the program, the two effects are treated as separate interventions. Disentangling Two Interventions The assumption is that the first intervention-the distribution of health cards- benefits only those who own a health card, whereas the second intervention can potentially benefit the whole population, depending on the size of the grant to the health care provider. This assumption rules out external or general equilibrium effects. Because only short-term impacts are considered, health-related general equilibrium effects are assumed not to be substantial since they are likely to take longer to materialize. However, externalities through congestion or crowding out induced by the program may also compro- mise the independence assumption. Sensitivity to these effects is examined in section V. Under the independence assumption the combined average impact of the two interventions can be written as the sum of the two impacts separately. Let Yj (hi, qj) denote the outcome for individual i living in district j as a function of the two interventions, with hi = 1 if a person lives in a household that has received a health card and 0 if not. The amount of SSN budgetary support to public health care providers in the area where the person lives (indicated by SSNj) is reflected by qj. The analysis seeks to establish to what extent the observed development in use from 1998 to 1999 is due to these two interventions. The overall impact of the program can be expressed as a weighted mean of the impact on people with a health card (h; = 1, qj = SSNi) and on people who did not receive a health card but who benefited only from the budget increase (hi = 0, qj = SSNj). Under the independence assumption, the overall impact can be written as p{E[Yi(1,SSN)| hi= 1,qj= SSNj] - E[Yi(0, 0)|hj= 1,qj= SSNj]} + (1 - p)f{E[Yi(0, SSNj)|hj = 0, qj = SSNj] - E[Yi(0,0)|hj= 0,q,= SSNj]} (1) where p = Pr (hi = 1). The observed average outcome for people with a health card is E [Yj (1, SSNj) | hi = 1, qj = SSNj], whereas E [Yi (0, SSNj) | hi = 0, Pradhan, Saadah, and Sparrow 133 q= SSN] reflects the observed average outcome for people who did not receive a health card. The other two terms reflect the expected counterfactual outcomes for the two groups: what would have happened if the programs had not been implemented. Equation 1 can be rewritten by adding and subtracting pE [Y; (0, SSNj) | hi = 1, q, = SSN], as p {E[Yi(1, SSNj)|hi = 1, qj = SSNJ] - E[Yi(0, SSNi)|hi = 1, qj = SSNi]} + E[Yi(0, SSNj)|qj = SSN] - E[Yi(0, 0)|qj = SSNj]. (2) Here the first two terms (weighted by p) give the impact of the pure health card program, conditional on the budget increase, for those who own a health card. This is referred to as the direct effect of the program. The last two terms reflect the effect of the budget increase for the whole population, referred to as the indirect effect of the SSN program. Estimation Strategy Both the direct health card effect and the overall effect are estimated. The indirect effect of the program cannot be identified directly. Instead, the impact of the general increase in funding to public services is derived by subtracting the direct effect estimate from the total effect estimate. For estimating the direct effect of the health card intervention, a control group is formed from the population that did not receive a health card. Since both those with and those without health cards benefited from the transfer of funds to health care providers, this measures the differential effect of owning a health card conditional on the transfer program. Since selection was not random, a direct comparison of those with and those without health cards after introduction of the program does not yield a valid impact estimate. The health card was distributed to poor households, and even without a health card use of health care services by these households would have been different from that of wealthier households without health cards. It is also possible that health cards were allocated based on need. In that case health card recipients would, on average, use more health care, even without the health card. Propensity score matching is used to correct for nonrandom placement of the program, relying on matching on observables and the assumption of con- ditional independence6 (that is, conditional on a set of observed characteristics 6. We experimented with instrumental variables but abandoned this approach because we were not convinced that we could construct adequate instruments. We used variables that measure the perception of fairness of the distribution of health cards in the district. But the results were very sensitive to specification and choice of instrument. We also experimented with 1997 district BKKBN estimates. However, when using 1998 data we found that these variables appear to be correlated with the level of use (but not with changes). 134 THE WORLD BANK ECONOMIC REVIEW selection into the program can be treated as random'). The unit of analysis is the household, as health cards were distributed at this level. Households in the treatment group are then matched to households in the potential control group. The extent to which propensity score matching will reduce the bias depends on the specification of the propensity score model and the quality of the control variables (Heckman, Ichimura, and Todd 1997). It is therefore crucial to understand the program design and to include sufficient information about the selection procedure (at all allocation levels) in the model. There are two main sources of bias. The first is the endogenous placement of health cards with households. The second relates to systematic differences in regional program intensity between the control and the treatment groups. District-fixed effects are included to control for these regional differences. They capture any between-district variation in the allocation of health cards and SSN funding. BKKBN poverty estimates for subdistricts control for the allocation of subsidies within districts and the number of health cards issued in the areas covered by the public health facilities. Thus, matched households live in areas that enjoy similar program intensity in terms of health card coverage and SSN budget. Endogenous program placement is caused by purposive targeting at different stages in the decentralized allocation process. To control for endogenous program placement at the village level, variables from the village-level census are included that reflect pre-program access to health care: number of public health clinics, auxiliary health clinics, and maternity facilities in the village; dummy variables indicating whether the majority of village traffic is by land; and a dummy variable reflecting village leaders' opinions about the accessibility of health clinics. Because health cards are distributed by local facility staff, the number of doctors and village midwives living in the village (per 1,000 inhabitants) is included as a proxy for informal networks within the village. Finally, the level of education of the village leader is included, as well as dummy variables indicating eligibility for Inpres Desa Tertinggal (IDT), an antipoverty program for economically less developed villages, and whether the village is located in a rural area. For households the five criteria of the BKKBN prosperity status are included as dummy variables. Other household welfare variables include housing charac- teristics (type of house occupied; type of roof, walls, and floor; sewage, sani- tation, and drinking water facilities; and source of light), sector of main source of household income, and employment status of the head of household. Other controls include household composition (gender and age), household size, and characteristics of the head of household (gender and education level). Per capita consumption is endogenous (a health card reduces health care expenses) and is 7. Following Rosenbaum and Rubin (1983). Smith and Todd (2005) provide an insightful discussion on the application and pitfalls of propensity score matching in the recent literature. 8. As a result, the sample sizes of the treatment and matched control group differ because households vary in number of people. Pradhan, Saadah, and Sparrow 135 therefore omitted. A household with a health card would, on average, report a lower consumption level than it would if it had not received a health card.9 If household expenditure were added to the propensity score function, the control group would be less wealthy than the intervention group. Consequently, the health card effect would be overestimated. Health status is the one important unobserved variable that is missing from this specification. Soelaksono and others (1999) provide some evidence that health cards were allocated based on illness. This is reinforced by the fact that self-reported illness is higher among health card owners. This would suggest that the positive bias due to health status outweighs the negative bias due to propoor targeting. The susenas survey records self-reported illness, but this is prone to reporting bias and may be endogenous to health card allocation.10 Health status is therefore omitted from the propensity score function. Many of the individual characteristics included will reduce the health bias (for example, age, housing and sanitary conditions, BKKBN criteria), but some may remain. However, it is shown later that the potential bias from omitting health status is small and that the estimation results are robust and within reasonable bounds. The propensity score function was estimated as a logit, separately for each of the five main regions in Indonesia." This restricted the match to households in the same region. A household with a health card living in Java, for instance, will not be matched with a household without a health card living in Sumatra. The pseudo R2 for the regional models ranged from 0.12 to 0.26. Nearest neighbor matching, the simplest matching procedure, was applied to house- holds within the range of common support.12 The matched households are very similar in the individual observed charac- teristics that entered into the matching function (table 4). From table 4 it appears that, before the match, households that owned a health card perform worse on the BKKBN criteria, are slightly larger, and work more often in agri- culture compared with households that do not own a health card. Heads of households with a health card have less education on average and are more likely to be female. After the match the control and intervention groups are well balanced across the observed characteristics. The second panel of table 4 shows variables that were not included in the matching function. Both program intensity variables are balanced for the 9. See van de Walle (2003) for a discussion on assumptions about behavioral responses regarding the effect of public policy on household consumption. 10. Using an experiment with increases in user fees in Indonesia, Dow and others (2000) provide an illustration of the problem of reporting bias and measurement error in self-reported health status. Whereas objective measures of health show that increasing user fees leads to a deterioration in health status, self-reported measures suggest an improvement in health. Dow and others (2000) argue that this reporting bias is correlated with exposure to the health system, which is affected by the user fee increase. See also Strauss and Thomas (1998) for a more general discussion. 11. The five regions are defined as Java and Bali, Sumatra, Sulawesi, Kalimantan, and Other Islands. 12. The estimation results for the propensity score function and details on the matching procedure are reported in the supplemental appendix, available at http://wber.oxfordiournals.org/. TABLE 4. Descriptive Statistics for Households with and without a Health Card and for Matched Pairs All households Matched pairs O No health Health No health Health Difference Standard a Variable card card card' card in means error Included in matching Propensity score 0.0823 0.2488 0.2433 0.2433 -0.0000 0.0018 Female head of household 0.1268 0.1608 0.1618 0.1601 -0.0017 0.0038 Education head of household No education completed 0.3641 0.5087 0.5090 0.5073 -0.0017 0.0052 Primary 0.2985 0.3324 0.3289 0.3327 0.0038 0.0049 Junior secondary 0.1220 0.0814 0.0818 0.0818 0.0000 0.0028 Senior secondary 0.1689 0.0667 0.0693 0.0674 -0.0019 0.0026 Higher 0.0465 0.0107 0.0111 0.0108 -0.0003 0.0011 Head of household unemployed 0.0079 0.0074 0.0075 0.0074 -0.0001 0.0009 Household size 4.2043 4.2576 4.2211 4.2449 0.0238 0.0189 BKKBN household prosperity criteria Worship 0.9343 0.8894 0.8911 0.8902 -0.0010 0.0032 Food 0.9835 0.9778 0.9785 0.9790 0.0004 0.0015 Clothing 0.9645 0.9473 0.9487 0.9487 0.0000 0.0023 Floor 0.8193 0.5935 0.5962 0.5954 -0.0008 0.0051 Health 0.8899 0.9061 0.9056 0.9057 0.0001 0.0030 Main source of household income Agriculture, farming 0.4551 0.5568 0.5526 0.5546 0.0020 0.0051 Mining, quarrying 0.0097 0.0089 0.0089 0.0089 -0.0001 0.0010 0lOZ 'LXA.r-aqoj uo punjrjPjuoW Ituopiuuu ju /'Jo-slunopJojxo-jaqm//:dBt uioij popolumo( Processing industry 0.0687 0.0685 0.0655 0.0682 0.0027 0.0026 Electricity, gas, water 0.0022 0.0007 0.0009 0.0007 -0.0002 0.0003 Construction 0.0400 0.0494 0.0507 0.0496 -0.0011 0.0023 Trade 0.1482 0.1180 0.1206 0.1193 -0.0013 0.0034 Transport, storage, communications 0.0510 0.0519 0.0522 0.0521 -0.0002 0.0023 Finance, insurance, real estate 0.0091 0.0031 0.0026 0.0031 0.0005 0.0006 Services 0.1462 0.0931 0.0957 0.0936 -0.0021 0.0030 Other 0.0028 0.0037 0.0033 0.0036 0.0004 0.0006 Income recipient 0.0672 0.0459 0.0470 0.0464 -0.0006 0.0022 Rural area 0.6792 0.7880 0.7856 0.7862 0.0006 0.0042 IDT villageb 0.2822 0.3495 0.3476 0.3444 -0.0032 0.0049 BKKBN rate per subdistrict 0.3088 0.4407 0.4417 0.4390 -0.0028 0.0026 Not included in matching Program intensity at district level SSN budget per capita 1.6164 1.8178 1.8154 1.8147 -0.0007 0.0099 Health card coverage 0.0886 0.1885 0.1865 0.1870 0.0004 0.0012 Weight for age Z-score, children under -1.2116 -1.2943 -1.2924 -1.2987 -0.0063 0.0244 fivec Member of household ill 0.3110 0.3620 0.3293 0.3605 0.0312 0.0049 Number of observations 173,366 18,993 18,727 18, 727 alncludes 406 households that are matched more than once. bVillages in the IDT antipoverty program. cSusenas nutrition module, 1999. Total sample is 7,902 children with a health card and 64,946 children without a health card. Matched sample is 7,502 children with a health card and 6,891 children without a health card. Source: Authors' analysis based on data described in text. 00ZO~'LXjr-mqoj uo punjrjPjuoIt iuoiluuiu ju /'Jo-slunopJojxo-jaqm//:dl[ uioij popolumo( 138 THE WORLD BANK ECONOMIC REVIEW matched households, while they differ strongly for the nonmatched households. This confirms that the district dummy variables in the matching function managed to control for variation in the size of the grants and health card cov- erage in the district. The match has also balanced the weight for age Z score of children under age five.13 Weight is indicative of the health of children over a period of time, which, in this case, will reflect mostly the period before the launch of the health card. In the absence of panel data it is thus the best proxy for balance in pre-intervention outcome variables.14 Section V further investi- gates the robustness of the impact estimate to health status. Comparing means of the matched treatment and control group yields the average direct effect of the health card intervention on the use of outpatient services by health card owners. This is obtained by estimating the regression: Yi = 8 + 6 HCj + si (3) on the matched sample, applying sample weights. The term is an unbiased estimate of the treatment effect for those who are selected into the program: 1 = E[Yi(1, SSNj) - Yi(0, SSNj)|hj = 1, qj = SSNj]. (4) Weighting this by the probability of selection into the program, = Pr(h; = 1), gives the average direct health card effect, fi , defined in equation (2). The overall impact of the program, as defined in equation (1), is obtained by exploiting regional variation in the financial compensation of public health care providers for the health card program and the fact that the allocation to districts was based on pre-intervention poverty estimates. Pre-health card use rates-based on the 1998 susenas survey-are compared with health care use rates right after introduction of the health card program. The impact estimate is a result of the two interventions acting simultaneously. The robustness of this approach is evaluated later in the article. Administrative data on the 1998/99 budget allocated for transfers to public health facilities were used to measure the variation in SSN compensation. The variation was substantial. For example, Sulawesi's allocation was (weighted by district population) 29 percent higher than Sumatra's and 34 percent higher than Java and Bali's, but about half that allocated to the smaller islands. The total SSN budget allocated in block grants to public providers amounted to 159 billion rupiah, or 1,432 rupiah per capita. The effect of the general increase in funding is modeled as a linear function of the budget allocation. For district j in time period t, use of health services is 13. The weight for age Z score is based on a 1999 susenas nutrition module covering 72,848 children under age five. 14. Waters, Saadah, and Pradhan (2003) find that the crisis had no observable effect on the weight for age Z score. Effects of the health card program on the score are unlikely as it did not cover nutritional programs. Pradhan, Saadah, and Sparrow 139 written as s SSN Ylt = a, + Oodt + Ordrdt + Ny + O it + sit (5) where SSNj is the amount of compensation for public health clinics allocated to district j, and N denotes the district population size. The time subscript, t, refers to either the time period before the intervention (1998) or the time period after the intervention (1999). The time dummy variable, dt, takes a value of zero if the period is 1998 or one if it is 1999 and is interacted with five region-specific fixed effects, dr, to allow for some flexibility in capturing the time effect.15 In the pre-intervention year SSNj equals zero for all districts. A set of regional welfare and demographic characteristics, Wit, are also added to the model. These include the poverty rate, Po, and poverty gap, P1, for the districts, the average age and household size, the district population size, and the fraction of the population living in a rural area. Frankenberg, Smith, and Thomas (2003) show evidence of changes in household size and migration between urban and rural areas as households restructured their composition in response to the crisis. Although the average household size increased in (lower cost) rural areas, the number of working age family members increased in urban households. The nonrandom allocation of the SSN budget is accounted for by a district- fixed effect, aj. This removes any bias due to unobserved time-invariant factors that affect geographic allocation and are also correlated with health care use. The fact that the SSN budget allocation was determined by static pre-program poverty estimates, and not on the basis of dynamic changes in poverty, legiti- mizes the fixed-effects approach. Taking differences across districts over time gives 5 SSNj99 AYlt = 0o + Ordr + NY - + SAWit + Aet. (6) Estimating equation (6) by ordinary least squares (OLS) yields unbiased estimates under the assumption that the allocation of SSN funds is not correlated with time-variant unobservables. If the geographic allocation is correlated with important district-level trends that are not captured by the time dummies or AWj,, then OLS estimates may still be biased. This is not very likely, given that the BKKBN indices are badly suited for capturing changes in welfare. Further reassurance is given by the fact that no corre- lation is found between SSN allocation (per capita) and pre-program changes in use from 1997 to 1998. 15. Java and Bali (region 1) are used as the reference group. 140 THE WORLD BANK ECONOMIC REVIEW The overall impact of the program is then obtained by taking a population- weighted average of the effects for the districts J SSNj Nj NJN = SSN (7) Nj where SSN is the average financial compensation for the health card program per person across the country, and J is the number of districts. The estimated impact of the supply impulse on the use of outpatient services (the indirect effect) is given by the difference between the estimate of the average total effect and the average direct health card effect. Inserting equation (7) and the estimate of /3 into equations (3) and (2) yields an expression for the impact of the general budget increase for public service providers E[Yi(0, SSN)|qj = SSNj] - E[Yi(0, 0)|qj = SSN_]N = j'SSN - (8) IV. RESULTS The estimation results of the direct health card effect on outpatient use for health card owners, /3, and the average direct effect, , are summarized in table 5. The estimate of f is simply the fraction of individuals (for a specific population group) living in a household that owns a health card. The table also shows the percentage change relative to the counterfactual. Health card ownership resulted in a 1 percentage point increase in the use of outpatient services, a 9.1 percent increase relative to the base counterfactual. This increase was due to an increase in use among the poorest quintiles. Only a substitution effect is observed among the richest quintile, from private to public health care providers. For all income groups health card ownership resulted in an increase in the use of public sector services and a decrease in the use of private sector services. For the richest quintile the two effects cancelled out each other. There was a small but statistically insignificant increase in overall use. The shift from private to public care seems to have occurred in both urban and rural areas. The health card program affected use more among women than among men, possibly because maternity services were covered under the program. Both the overall increase in outpatient visits and the substi- tution effect from private to public services were larger for women. Table 5 also presents the estimates of y from equation (6), and estimates of the overall effect of the program (5 SSN), defined in equation (7). The results indicate an absolute increase in the use of outpatient services of 0.5 percentage point, which stems mainly from an increase in the use of public services, as the TABLE 5. Impact of Health Card Ownership and Overall Effect of SSN Interventions on Use of Outpatient Services (one month reference period) Direct effect of health card' Overall effect of SSNb Indirect effectc Intervention Control Difference Change Direct effect Coefficient Overall effect (percentage share group group (0) (%) (Ip) (9) ( SSN) of overall effect) p All outpatient visits Quintile 1 (poor) 0.0993 0.0869 0.0123* 14.2 0.0023* 0.0039 0.0056 (58.93) Quintile 5 (rich) 0.1510 0.1451 0.0059 4.0 0.0002 0.0075** 0.0108** 98.15 Male 0.1158 0.1069 0.0089* 8.3 0.0009* 0.0037** 0.0053*** 83.02 Female 0.1270 0.1157 0.0113* 9.8 0.0012* 0.0040*** 0.0058*** 79.31 Urban 0.1392 0.1281 0.0110* 8.6 0.0008* 0.0045 0.0064 (87.50) Rural 0.1149 0.1061 0.0088* 8.3 0.0011* 0.0060** 0.0086** 87.21 All 0.1215 0.1113 0.0101* 9.1 0.0011* 0.0039*** 0.0055*** 80.00 Outpatient public Quintile 1 (poor) 0.0729 0.0542 0.0187* 34.6 0.0035* 0.0035 0.0049 (28.57) Quintile 5 (rich) 0.0841 0.0590 0.0251* 42.5 0.0009* 0.0094* 0.0134* 93.28 Male 0.0734 0.0537 0.0197* 36.8 0.0021* 0.0026*** 0.0037*** 43.24 Female 0.0871 0.0622 0.0249* 40.1 0.0027* 0.0040** 0.0057** 52.63 Urban 0.0869 0.0628 0.0241* 38.3 0.0017* 0.0016 0.0023 (26.09) Rural 0.0779 0.0565 0.0214* 38.0 0.0027* 0.0053* 0.0076*** 64.47 All 0.0804 0.0580 0.0224* 38.6 0.0024* 0.0033* 0.0047** 48.94 Outpatient private Quintile 1 (poor) 0.0305 0.0371 -0.0066* -17.7 -0.0012* 0.0020 0.0028 Quintile 5 (rich) 0.0803 0.0983 -0.0179* -18.2 -0.0007* -0.0016 -0.0023 Male 0.0501 0.0601 -0.0100* -16.6 -0.0010* 0.0012 0.0017 (Continued) 00ZO~'L,1j,a-qoj uo punjrjPjuoW Ituoil1uuu u /'Jo-slunopojxo-jqm//:dt[ uioij poppolumo( TABLE 5. Continued Direct effect of health carda Overall effect of SSNb Indirect effectc o Intervention Control Difference Change Direct effect Coefficient Overall effect (percentage share a group group ( ) (%) (jip) (j) (i SSN) of overall effect) P Female 0.0477 0.0606 -0.0129* -21.3 -0.0014* 0.0005 0.0008 Urban 0.0613 0.0726 -0.0113* -15.5 -0.0008* 0.0031 0.0045 Rural 0.0442 0.0565 -0.0123* -21.7 -0.0016* 0.0013 0.0019 All 0.0489 0.0604 -0.0115* -19.0 -0.0012* 0.0009 0.0013 *Significant at the 1 percent level; **significant at the 5 percent level; * *significant at the 10 percent level. Note: Detailed estimation results are available in the supplemental appendix (tables S.7, S.9-S.11). aNumber of observations: 76,903 individuals in treatment group, 73,986 in control group. Bootstrapped standard errors with 500 replications. bNumber of observations: 293 districts. cNumbers in parentheses are indirect effect estimates that are based on imprecise estimates of the total effect. Source: Authors' analysis based on data described in text. C I OZ 'LXjr-mqoj no punjrjPjuoIt uopiuIulu ju /.Jostiopojojq/ t uioij popolumo( Pradhan, Saadah, and Sparrow 143 program does not seem to affect the private sector. The effect is larger for wealthier households. For poor households the estimates are smaller and impre- cise. As with the direct health card effect, the overall effect of the program on public services is larger for females than for males. The program had the largest impact on the use of public care in rural areas; in urban areas the esti- mates are imprecise. Since private care seems unaffected, the results are similar for the overall effect on use. The indirect effect that could be attributed to an overall supply or quality impulse as a result of the extra budget support in the public sector seems to have been a main contributor to the increase in the use of public health care services. Combining the estimates of the direct health card effect with the overall effect of the SSN permits investigation of what share of the increase in the use of public sector services is due to the indirect effect (as defined in equation (8)). The share of the indirect effect in the total effect is given by 1 - [(fi /3) / (' SSN)]. The indirect effect accounts for about 80 percent of the overall increase in use. In the public sector about half of the total increase can be attributed to the indirect effect of the budget increase. The results also suggest that the indirect benefits of the program increase with income. For the richest quintile only 7 percent of the increased use of public care can be attrib- uted to the health card itself whereas for the poor there is less clear evidence of an indirect effect. The indirect effect for the poor is smaller, but based on an imprecise estimate. Finally, the supply impulse had an above average effect in rural areas, emphasizing the shortage of resources of rural public health care providers. The overall effects for the private sector are not significant, so the indirect effects are not calculated. So, can the revival in use of public sector health services be attributed to the SSN program? The answer appears to be yes. The results reported in table 5 can be used to estimate use had the health card program not existed. Table 1, which reported trends in health care use, included the counterfactual for public and private health care use in the absence of the health card program. From 1998 to 1999 the contact rate for public health services increased from 5.0 percent to 5.3 percent, whereas the contact rate for all modern health care providers remained stable at 10.5 percent. The estimates suggest that without the health card program public health care use would have dropped further to 4.9 percent and the overall contact rate would have dropped to 10.0 percent. V. CAVEATS AND SENSITIVITY ANALYSIS This section discusses some caveats to the empirical analysis and examines the robustness of the results with respect to specification and the main assumptions. 144 THE WORLD BANK ECONOMIC REVIEW Crowding Out, Congestion, and Interaction Effects The main assumption underlying the study is that use of health care services by households with health cards is independent of that by households without health cards. This implies that the number of health card recipients (program intensity) in the region does not affect use of care for nonrecipients and that both groups enjoy similar benefits from the SSN budget. However, if health care supply were inelastic, then distributing health cards could lead to conges- tion and crowding out. For example, if services were delivered to health card owners according to set standards, resources would be redistributed from non- recipients to health card recipients. In this case the estimated direct effect of the health card would be biased upward. The difference in use would consist of the "true" health card effect and the crowding out effect. Alternatively, external- ities can manifest themselves if the direct benefits of the health cards do not follow set standards but are contingent on available resources. The quality of care provided to health card owners will then increase with the SSN budget. One might argue that the external effects of health card allocation are likely to be small. Since health card coverage is 11 percent and concentrated among the poor, whose health care demand is typically low, it is unlikely that the program would strain the capacity of health care facilities. For example, dou- bling the use of public health services for health card owners would result in 16 percent more outpatient visits for a typical public health care facility. The district dummy variables included in the matching functions do capture program intensity and the supply shock induced by the SSN program (Table 4). Moreover, the estimation method allows for effect heterogeneity due to regional variation in program intensity, since the estimated impact for all the households with a health card is simply averaged. Nevertheless, it is possible to test for the presence of externalities by control- ling for program intensity when estimating the direct effect, and including interaction effects of health card ownership with the average number of health cards distributed in the district and the per capita amount of the SSN subsidy. Crowding out or congestion would imply that the interaction effects for program intensity are statistically significant. If crowding out or congestion effects are important, they would be expected to be stronger in areas where the program is underfunded. These are areas where a large number of health cards are distributed compared with the budget that is received. This can be tested by including the amount of the SSN subsidy per allocated health card as regressor and interacting this with the health card dummy variable. Statistically signifi- cant interaction effects would indicate the presence of general equilibrium and external effects. The top panel of table 6 shows estimates of the direct effect given different specifications. The results suggest that the estimated direct effect is not biased due to externalities. Specification 1 gives the initial estimates. Specification 2 controls for the fraction of the population with a health card and the SSN TABLE 6. Sensitivity of Impact Estimates to Different Specifications and Assumptions Overall outpatient care Public Private Direct effect of health card 1. Original estimate table 5 0.0101* (0.0020) 0.0224* (0.0015) -0.0115* (0.0015) 2. Program intensity control variablesa 0.0109* (0.0020) 0.0235* (0.0015) -0.0113* (0.0015) 3. Interaction effects' a. SSN per capita and health card allocation per capita 0.0106** (0.0052) 0.0272* (0.0038) -0.0140* (0.0038) b. SSN per health card in district 0.0114* (0.0021) 0.0239* (0.0016) -0.0109* (0.0016) 4. Selection on needsb 0.0081* (0.0020) 0.0201* (0.0016) -0.0117* (0.0015) Total effect of SSN program 5. Original estimate table 5 0.0039*** (0.0022) 0.0033** (0.0015) 0.0009 (0.0013) 6. Health card coverage in districtc 0.0042 (0.0028) 0.0037 (0.0019) 0.0007 (0.0017) 7. Interaction effectsd 0.0044 (0.0032) 0.0039 (0.0022) 0.0007 (0.0020) *Significant at the 1 percent level; **significant at the 5 percent level; **significant at the 10 percent level. Note: The coefficients of the interaction terms and other covariates are omitted for convenience. Detailed estimation results are available in the supplemental appendix (tables S.12-S.16). aProbit marginal effects. Specification 2 includes SSN budget per capita and health card coverage in districts. Specification 3 adds interaction terms of these two variables with the treatment dummy variable. Both specifications include age, gender, characteristics of head of household (gender, education), household size, BKKBN prosperity status, main source of income (agriculture/no agriculture), village status (rural, IDT antipoverty program), availability of health providers in the village or township, subdistrict BKKBN index, and district poverty profile (Po, PI). The sample concerns the same set of individuals from matched households as in table 5. Numbers in parentheses are robust standard errors. bThe propensity score function includes a dummy variable that indicates whether a health complaint has disrupted work, school, or the daily activities of a household member. N treated = 76,956, N control = 74,263. Numbers in parentheses are bootstrapped standard errors (with 500 replications).z cSimilar to specification 5, with health card coverage in districts added. dSimilar to specification 5, with interaction term (health card coverage in districts) x (SSN per capita in district) added. Source: Authors' analysis based on data described in text. CI0Z~'LXr-mqoj uo pu~njrjPjouo Ituoutuomuil ju /B1Jo-slnu1notpiojxofoqA//:d1lt umoij popolumo(I 146 THE WORLD BANK ECONOMIC REVIEW budget per capita allocated to the districts, the subdistrict BKKBN index, and district poverty indicators P0 and P1. It further includes a set of individual and household characteristics, IDT village and rural area dummies, and the avail- ability of health facilities in the village. Specification 3 includes the interaction terms. The interaction effects are not statistically significant for public and overall outpatient care. For private care, there is a small positive and weakly significant effect only for the SSN subsidy interaction term. This is an interest- ing finding, because doctors working at public facilities in Indonesia often maintain private practices. This could suggest that in districts with relative SSN budget abundance, some doctors have used the SSN subsidy to treat health card recipients in their private practice. The impact estimate is robust to differ- ent specifications. The point estimate for the direct health card effect on overall use is slightly larger, but still within one standard deviation, whereas the substi- tution effect between public and private is also slightly larger. Selection on Health Status A potentially more serious problem is the failure to take into account the possi- bility that households may have been selected based on health status. Those with poor health may have received a health card because of their higher antici- pated need while otherwise similar individuals did not receive one. Officially, health cards should have been distributed based on BKKBN criteria, but health status could well have played a role in actual distribution. If so, failing to include a measure of health status in the matching function will result in an intervention group with a worse health status than the control group. Poor health will, other things being equal, increase the demand for health care. The resulting impact estimate will be larger or equal to the true effect. The only measure of health status that the susenas survey collects is self- reported illness. However, including self-reported illness in the matching function would likely have resulted in an underestimate of the true health card effect. Evidence indicates that self-reported illness depends on the affordability of care. The rich report illness more often than the poor, which is surely not a result of the rich having a worse health status than the poor. If self-reported illness depends on the affordability of health care, and health care is more affordable for those who own a health card, then matching on self-reported illness will result in a control group with worse health status than the intervention group. Better health will, other things being equal, decrease the demand for primary health care. Thus the impact estimate would have been an underestimate. Two impact estimates, one obtained without and one with self-reported illness included in the matching function, can provide some notion of the extent of the bias. The health card effect should lie between the estimate that controls for self-reported illness (lower bound) and the one that does not (upper bound). The results suggest that the estimates presented earlier are not sensitive to systematic differences in health status, since the estimated bounds lie close to each other. A dummy variable was included in the matching Pradhan, Saadah, and Sparrow 147 function that indicated whether a health complaint had disrupted work, school, or daily activities for any member of the household during the last month. Specification 4 in table 6 gives the results for a one-month reference period. The impact estimate for all outpatient care decreases slightly, from 0.0101 to 0.0081. The point estimates are within one standard deviation. This leads to an upper and a lower bound for the direct effect of 0.11 to 0.09 per- centage point. The difference comes from the change in demand for public care. The estimated effect for private care remains unchanged. Total Effect Is the combined effect of the SSN funding and the allocation of health cards, as defined in equation (1), identified if general equilibrium effects compromise the independence assumption? It could be, for example, that the indirect effect of the subsidy decreases if health card allocation is relatively high. Alternatively, there could be districts with a high SSN allocation but with a delay in health card distribution at the time of the survey. Does the variation in SSN budget then adequately capture the total effect, and does this allow clear interpretation of the indirect effect? To investigate, health card coverage was added to the model, as well as an interaction term with the SSN variable. If the budget allocation does not identify the total effect, the results would be expected to be sensitive to the new variables. Note that health card allocation data are likely to be endogenous. Unlike the SSN budget, these data are not driven by pre-program welfare indicators. The data reflect the actual allocation of health cards, which depend on district- specific infrastructure, organization, and welfare characteristics, and are likely to be correlated with the heterogeneous effects of the crisis. Therefore, the BKKBN indices from December 1997 are used as instruments for health card allocation.16 The results are given in the bottom panel of table 6 and suggest that the original estimates are fairly robust and capture the combined effect of the program. When health card coverage or the interaction term is added to the regression, the coefficients for the SSN grants are slightly larger and a little less precise. VI. CONCLUSION This article presented an impact evaluation of the health card program as it operated under the SSN program in its first months. It found that in many ways the program was a success. It is also the case that the program may have worked in ways that were not the objective at the outset. The health card 16. The indices for the two poorest BKKBN classifications are used (pre-prosperous and KS1). Households ranked in one of these groups are eligible for a health card. The instruments are not correlated with the pre-crisis trend. An overidentifying restrictions test further validates the instruments. Detailed results are available in the supplemental appendix (tables S.16 and S.17). 148 THE WORLD BANK ECONOMIC REVIEW program has a weak link between the delivery of services to health card owners and the financial compensation of health care providers. Service providers are reimbursed using a lump sum transfer based on the number of health cards dis- tributed to their area of influence. As a result, serving a health card owner did not result in a direct financial reward to the service provider. This makes the health card program a rather particular case of a targeted price subsidy scheme. There is clear evidence that the health card program was propoor. The poor had a higher probability of receiving a health card, and those with a health card had increased use of health services, presumably making them healthier. However, there was considerable leakage of benefits to the richer quintiles, and use of services is less propoor than is ownership. Conditional on ownership, the rich have a higher propensity to use their health card. Returning to the questions posed initially, for all households health card ownership was found to result in a large substitution effect away from private providers to public providers, with a net increase in the overall use of outpati- ent medical services. A dynamic analysis further indicates that the combined SSN program resulted in an increase in the outpatient contact rate at modern health care providers of 0.55 percentage point. Without the program use of outpatient facilities would have fallen further in 1999. However, the direct health card effect contributed only about 20 percent to the increased use. A considerable proportion of the impact of the program seems to have been through the budgetary support for public health services. If this is true, the revival of public health services can be attributed in large part to the supply impulse induced by the increased spending under the SSN health program. However, the effects of both the health card and the supply impulse show a strong heterogeneous pattern across subgroups of the population. While the targeting and impact of the health cards were propoor, the total effect was not. The poor responded to a price subsidy but not to the supply impulse. The health card increased use and led to a substitution effect from private to subsi- dized public care. For the nonpoor, however, use seems to be mainly supply driven, as the health card affected only their choice of health care provider without increasing use. These results suggest that in the absence of clear incentive mechanisms for health care providers, general increases in public spending are relatively ineffec- tive in reaching the poor. A stronger link between provision of services to health card owners and budget support would likely have improved targeting to the poor. Health card distribution was propoor, and use of modern care among health card owners increased. A stronger link could have been estab- lished, for example, by tying compensation for providers to services delivered to health card owners, on a fee for service or capitation basis. An alternative would be to establish contracts with providers, with budgets dependent on monitorable target indicators in the communities they serve. The empirical evi- dence worldwide on alternative mechanisms to stimulate demand for health services among the poor is still scarce and merits further research. Pradhan, Saadah, and Sparrow 149 REFERENCES Akin, John S., David K. Guilkey, and Hazel Denton. 1995. "Quality of Services and Demand for Health Care in Nigeria: A Multinominal Probit Estimation." Social Science and Medicine 40(11):1527-37. Akin, John S., David K. Guilkey, Paul L. Hutchinson, and Michael T. McIntosh. 1998. "Price Elasticities of Demand for Curative Health Care with Control for Sample Selectivity on Endogenous Illness: An Analysis for Sri Lanka." Health Economics 7(6):509-31. Cameron, Lisa A. 1999. "Survey of Recent Developments." Bulletin of Indonesian Economic Studies 35(1):2-40. Ching, Panfila. 1995. "User Fees, Demand for Children's Health Care and Access across Income Groups: The Philippine Case." Social Science and Medicine 41(1):37-46. Currie, Janet, and Jonathan Gruber. 1996. "Health Insurance Eligibility, Utilization of Medical Care, and Child Health." Quarterly Journal of Economics 111(2):431-66. Currie, Janet, and Duncan Thomas. 1995. "Medical Care for Children: Public Insurance, Private Insurance, and Racial Differences in Utilization." Journal of Human Resources 30(1):135-62. Daly, Anne, and George Fane. 2002. "Anti-Poverty Programs in Indonesia." Bulletin of Indonesian Economic Studies 38(3):309-29. Dow, William H. 1999. "Flexible Discrete Choice Demand Models Consistent with Utility Maximization: An Application to Health Care Demand." American Journal of Agricultural Economics 81(3):680-85. Dow, William H., Paul Gertler, Robert F. Schoeni, John Strauss, and Duncan Thomas. 2000. "Health Care Prices, Health and Labor Outcomes: Experimental Evidence." Working Paper. Santa Monica, Calif: RAND Corporation. Filmer, Deon, and Lant Pritchett. 1999. "The Impact of Public Spending on Health: Does Money Matter?" Social Science and Medicine 49(10):1309-23. Frankenberg, Elizabeth, James P. Smith, and Duncan Thomas. 2003. "Economic Shocks, Wealth, and Welfare." Journal of Human Resources 38(2):280-321. Frankenberg, Elizabeth, Duncan Thomas, and Kathleen Beegle. 1999. "The Real Costs of Indonesia's Economic Crisis: Preliminary Findings from the Indonesian Family Life Survey." Labor and Population Program Working Paper 99-04. Santa Monica, Calif: RAND Corporation. Gertler, Paul, and Jeffrey S. Hammer. 1997. "Strategies for Pricing Publicly Provided Health Services." Policy Research Working Paper 1762. World Bank, Washington, D.C. Gertler, Paul, Luis Locay, and Warren Sanderson. 1987. "Are User Fees Regressive? The Welfare Implications of Health Care Financing Proposals in Peru." Journal of Econometrics 36(1-2):67-88. Gertler, Paul, and Jack Molyneaux. 1997. "Experimental Evidence on the Effects of Raising User Fees for Publicly Delivered Health Care Services: Utilization, Health Outcomes, and Private Provider Response." Santa Monica, Calif: RAND Corporation. Gertler, Paul, and Jacques van der Gaag. 1990. The Willingness to Pay for Medical Care: Evidence from Two Developing Countries. Baltimore, Md.: Johns Hopkins University Press. Heckman, James J., Hidehiko Ichimura, and Petra E. Todd. 1997. "Matching as an Econometric Evaluation Estimator: Evidence from Evaluating a Job Training Programme." Review of Economic Studies 64(4):605-54. Knowles, James C., Ernesto M. Pernia, and Mary Racelis. 1999. "Social Consequences of the Financial Crisis in Asia." Economic Staff Paper 60. Manila: Asian Development Bank. Lavy, Victor, and Jean-Marc Germain. 1994. "Quality and Costs in Health Care Choice in Developing Countries." Living Standards Measurement Study Working Paper 105. World Bank, Washington, D.C. Lavy, Victor, and John M. Quigley. 1993. "Willingness to Pay for the Quality and Intensity of Medical Care: Low-Income Households in Ghana." Living Standards Measurement Study Working Paper 94. World Bank, Washington, D.C. 150 THE WORLD BANK ECONOMIC REVIEW Manning, Willard G., Joseph P. Newhouse, Naihua Duan, Emmett B. Keeler, and Arleen Leibowitz. 1987. "Health Insurance and the Demand for Medical Care: Evidence from a Randomized Experiment." American Economic Review 77(3):251-77. Mocan, H. Naci, Erdal Tekin, and Jeffrey S. Zax. 2000. The Demand for Medical Care in Urban China. NBER Working Paper 7673. Cambridge, Mass.: National Bureau of Economic Research. Mwabu, Germano, Martha Ainsworth, and Andrew Nyamete. 1993. "Quality of Medical Care and Choice of Medical Treatment in Kenya." Journal of Human Resources 28(4):838-62. Rosenbaum, Paul R., and Donald B. Rubin. 1983. "The Central Role of the Propensity Score in Observational Studies for Causal Effects." Biometrika 70(1):41-55. Sahn, David E., Stephan D. Younger, and Garance Genicot. 2003. "The Demand for Health Care Services in Rural Tanzania." Oxford Bulletin of Economics and Statistics 65(2):241-59. Smith, James P., Duncan Thomas, Elizabeth Frankenberg, Kathleen Beegle, and Graciela Teruel. 2002. "Wages, Employment and Economic Shocks: Evidence from Indonesia." Journal of Population Economics 15(1):161-93. Smith, Jeffrey A., and Petra E. Todd. 2005. "Does Matching Overcome LaLonde's Critique of Nonexperimental Estimators?." Journal of Econometrics 125(1-2):305-53. Soelaksono, Bambang, Sri Budiyati, Hastuti, Musriyadi Nabiu, Akhmadi, Pamadi Wibowo, Sri Kusumastuti Rahayu, and John Maxwell. 1999. "The Impact of the Crisis on the People's Behavior and Perceptions of the Use and Effectiveness of Puskesmas, Posyandu, and the Role of Midwives." Jakarta: SMERU Research Institute. Strauss, John, and Duncan Thomas. 1998. "Health, Nutrition, and Economic Development." Journal of Economic Literature 36(2):766-817. Suryahadi, Asep, Sudarno Sumarto, and Lant Pritchett. 2003. "The Evolution of Poverty during the Crisis in Indonesia." Asian Economic Journal 17(3):221-41. van de Walle, Dominique. 2003. "Behavioral Incidence Analysis of Public Spending and Social Programs." In Frangois Bourguignon, and Luiz A. Pereira da Silva eds. The Impact of Economic Policies and Poverty and Income Distribution: Evaluation Techniques and Tools. Washington, D.C. and New York: World Bank and Oxford University Press. Waters, Hugh, Fadia Saadah, and Menno Pradhan. 2003. "The Impact of the 1997-98 East Asian Economic Crisis on Health and Health Care in Indonesia." Health Policy and Planning 18(2):172-81. World Bank. 2004. World Development Report 2004: Making Services Work for Poor People. Washington, D.C.: Oxford University Press and the World Bank. Yip, Winnie, and Peter Berman. 2001. "Targeted Health Insurance in a Low Income Country and its Impact on Access and Equity in Access: Egypt's School Health Insurance." Health Economics 10(3):207-20. A Short Note on Updating the Grilli and Yang Commodity Price Index Stephan Pfaffenzeller, Paul Newbold, and Anthony Rayner The Grilli and Yang commodity price index is one of the most widely used commod- ity price series in the applied economics literature. This note provides some practical advice on updating this data series by listing the base period index values, identifying relevant data sources, and describing a method for computing subindex weights. JEL codes: 013, Fl. In 1988, Enzo Grilli and Maw Cheng Yang published their seminal article on the long-run development of an index of 24 primary commodity prices (GYCPI) deflated by an index of manufactured goods' unit values. The sample of average annual primary commodity prices covers about 54 percent of the primary commodity trade in the index reference period, 1977-79 (Grilli and Yang 1988, p. 3, n. 2). The deflators considered for manufacturing prices were the U.S. manufacturing price index (USMPI) and the manufacturing unit value index (MUV).1 Widely used and discussed, the Grilli and Yang data set has been extended by a number of researchers (for example, Lutz 1999; Le6n and Soto 1997; Cashin and McDermott 2002). The data have been employed in a variety of contexts in later studies (for example, Bleaney and Greenaway 2001; Kim et al. 2003). Thus, the GYCPI data continue to enjoy wide popularity in their own right and as a benchmark for new approaches to empirical studies. However, Stephan Pfaffenzeller (corresponding author) is a lecturer in economics at the University of Liverpool; his email address is s.pfaffenzeller@liverpool.ac.uk. Paul Newbold is a professor of econometrics at the University of Nottingham; his email address is paul.newbold@nottingham.ac.uk. Anthony Rayner is emeritus professor of economics at the University of Nottingham; his email address is anthony.rayner@nottingham.ac.uk. The authors thank the late Enzo Grilli for providing background information on data sources and Betty Dow for providing data from the primary commodity price database. They are also indebted to Prof. David Sapsford and Dr Paul Cashin as well as three anonymous referees for helpful comments. Stephan Pfaffenzeller gratefully acknowledges the financial support provided by a UK Ministry of Agriculture, Fisheries, and Food (MAFF) studentship. Supplemental appendixes to this article are available at http://wber.oxfordjournals.org/. 1. The MUV had to be interpolated for 1914-20 and 1939-47. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 151-163 doi:10.1093/wber/1hl013 Advance Access Publication 31 January 2007 () The Author 2007. 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 151 152 THE WORLD BANK ECONOMIC REVIEW because the data sources consulted for updates have differed, it may not always be clear when differences in results arise from differences in the data used and when they arise from differences in the econometric methodology employed. The obvious need for occasional updates of the series is in marked contrast with the absence of an accessible central reference for suitable data sources and the appropriate weights to be applied to individual commodity prices over the various subindices. This note aims to identify suitable data sources and compo- site index and subindex weights. These have not all been directly and publicly available from an accessible source.2 The data sources identified are a compro- mise between continuity with the original Grilli and Yang data and accessibility. The intention is to allow individual researchers a realistic opportunity to obtain identical updates of the Grilli and Yang index from clearly identified sources. I. DATA SOURCES FOR UPDATES Data for updating the commodity price data come directly from the World Bank Development Prospects Group's primary commodity price database, the International Monetary Fund (IMF) commodity price tables, and the Organization for Economic Cooperation and Development (OECD) inter- national trade by commodities statistics. Most of the World Bank's primary commodity prices and the IMF's commodity price tables are available online. Online access to the OECD trade statistics requires a subscription. MUV updates were obtained from the Global Economic Prospects team of the World Bank's Development Prospects Group.3 A possible cause of confusion is the frequent revisions of the reported data and the occasional lack of continuity in the available data series. Often, more than one series is available for one commodity, and the researcher will have to use dis- cretion in selecting the most appropriate one. Obviously, a close correspondence to the historic series is desirable, but a perfect match may not always be possible. The following list of the 24 commodities in the Grilli and Yang data set ident- ifies the data series used to update them from 1987 onward. The IMF commod- ity price table data are identified by their series descriptors or series code. OECD trade data are identified by their four-digit Standard International Trade Classification, Revision 2 (SITC Rev. 2) code. Data obtained directly from the primary commodity price database of the World Bank's Development Prospects Group are also identified.4 Any decision to deviate from the data series used in the original GYCPI data set is identified in the commodity description.5 2. At least some of this information was reportedly published in a working paper preceding Grilli and Yang (1988). However, we have been unable to obtain a copy despite repeated efforts. 3. The data are quoted online in the commodity price appendix to World Development Indicators. The 2005 edition is available at http://devdata.worldbank.org/wdi2005/Table6_4.htm. 4. Updates are available online from the World Bank's "Pink Sheets" (www.worldbank.org/ prospects). 5. Information on the original data sources was obtained from a list kindly provided by Enzo Grilli detailing data sources and definitions for Grilli and Yang's work. Pfaffenzeller, Newbold, and Rayner 153 Aluminium: London Metal Exchange (LME), unalloyed primary ingots, high grade, minimum 99.7 percent purity, from the primary commodity price database. Bananas: Central and South American, U.S. import price, free on truck (f.o.t.) gulf ports, from the primary commodity price database. Beef: IMF commodity price tables series PBEEF, beef, Australian and New Zealand 85 percent lean fores. These data deviate from the Argentinean export unit values used in the original study, which were obtained from Argentinean national statistics ("Comercio Exterior" Argentina, Instituto Nacional de Estadistica y Censos). The national statistics are not readily available online or in print at the required level of detail. Cocoa: International Cocoa Organization daily price, average of the first three positions on the terminal markets of New York and London, nearest three future trading months, from the primary commodity price database. Coffee: International Coffee Organization, other mild Arabica, from the primary commodity price database. Copper: LME grade A minimum 99.9935 percent purity, cathodes and wire bar shapes, settlement price, from the primary commodity price database. Cotton: Cotton Outlook A Index, middling 1 (3/32) inch staple, Europe cost, insurance, and freight (c.i.f.), from the primary commodity price database. Hides: IMF commodity price tables series PHIDE, hides, heavy native steers, over 53 pounds. Jute: Raw white D, free on board (f.o.b.) Chittagong. This series, obtained directly from the World Bank and quoted on the Pink Sheets, was discontin- ued after 2004. More recent jute prices are quoted by the Food and Agriculture Organization (FAO). Lamb: New Zealand, frozen whole carcasses, wholesale price; London, from the primary commodity price database. Lead: LME refined, 99.97 percent purity, settlement price, from the primary commodity price database. Maize: U.S. no. 2 yellow, f.o.b. gulf port, from the primary commodity price database. Palm oil: 5 percent bulk, Malaysian, c.i.f. NW Europe, from the primary com- modity price database. Rice: Thai 5 percent, milled, indicative price based on weekly surveys of export transactions, government standard, f.o.b. Bangkok, from the primary com- modity price database. The original Grilli and Yang data set used the Board of Trade-posted price series, which was phased out after 1991. In the inter- est of continuity, the series from the primary commodity price database is used for the updated series from 1987 onward. 6. These are available online from the Food and Agriculture Organization Commodities and Trade home page (www.fao.org/es/esc/en/index.html). At the time of writing, prices could be obtained from an interactive databank by following the "Prices" link under "Publications." 154 THE WORLD BANK ECONOMIC REVIEW Rubber: RSS no.1 Rubber Traders Association spot New York, from the primary commodity price database. Silver: Handy & Harman 99.9 percent New York, from the primary commod- ity price database. Sugar: International Sugar Agreement daily price, raw, f.o.b. and stowed at greater Caribbean ports, from the primary commodity price database. Tea: Three-auction average (Kolkata, Colombo, Mombasa), from the primary commodity price database. The original Grilli and Yang data set used the London auctions series, which was phased out in 1998. In the interest of continuity, the three-auction average listed in the World Bank's Pink Sheets is used in the updated series from 1987 onward. Timber: OECD international trade by commodities statistics, through ESDS International, UK import unit values, SITC Rev.2 series 2482 (sawn wood, coniferous species). Tin: LME 99.85 percent purity, settlement price, from the primary commodity price database. Tobacco: U.S. import unit values, unmanufactured leaves. Data were obtained directly from the primary commodity price database of the World Bank's Development Prospects Group.7 Wheat: No.1 Canadian western red spring, in store, St Lawrence, export price, from the primary commodity price database. Wool: IMF commodity price tables series PWOOLC, wool, coarse, 23 micron, Australian Wool Exchange spot quote. Zinc: LME, special high grade, minimum 99.995 percent purity, weekly average bid/asked price, official morning session; prior to April 1990, high grade, minimum 99.95 percent purity, settlement price, from the primary commodity price database. II. FURTHER DETAILS ON COMMODITY PRICES All commodity price series have been indexed to their 1977-79 average in con- structing the GYCPI and its subindices. Both the new data and the original Grilli and Yang component series were first indexed to their 1980 values, and the updated component of this series was subsequently indexed to the 1977-79 average of the combined index series. Table 1 shows the 1977-79 index values for each commodity, as well as the 1990 and 2000 index values, to facilitate future extensions of the series. The individual commodity price series can be used to update the GYCPI and its various subindices. 7. The data series is not listed in the Pink Sheets but is available online from the commodity price data appendix of World Development Indicators. The 2005 edition is available at http://devdata. worldbank.org/wdi2005/Table6_4.htm. Pfaffenzeller, Newbold, and Rayner 155 TABLE 1. Commodity Price Index Data for Selected Years and Index Weights (1980 = 100) Weights (% share) Commodity 1977 1978 1979 1990 2000 GYCPI Subindices Food Bananas 72.479 75.803 85.909 142.718 111.873 0.9 1.64 Beef 47.098 50.394 84.283 92.872 70.125 5.1 9.27 Cocoa 145.689 130.935 126.619 48.654 34.791 2.7 4.91 Coffee 154.493 106.410 110.897 56.901 55.386 10.3 18.73 Lamb 57.191 69.049 87.550 92.044 90.722 0.9 1.64 Maize 76.060 80.370 92.185 87.231 70.658 6.8 12.36 Palm oil 89.146 102.340 111.137 49.666 53.171 8.3 15.09 Rice 62.732 84.693 76.354 65.942 49.275 3.0 5.45 Sugar 28.322 27.206 33.693 43.805 28.544 7.3 13.27 Tea 120.454 98.122 96.640 124.000 113.076 1.6 2.91 Wheat 60.691 70.649 90.356 81.855 77.113 8.1 14.73 Nonfood primary commodities Cotton 85.951 77.863 88.414 88.862 63.613 4.3 15.81 Hides 80.559 102.785 159.132 200.862 174.718 2.3 8.46 Jute 91.034 106.476 107.462 132.565 90.077 0.2 0.74 Rubber 56.410 68.171 87.568 62.836 51.217 2.8 10.29 Timber 66.483 65.982 82.204 114.992 94.658 12.0 44.12 Tobacco 80.101 93.367 97.632 149.051 130.773 2.9 10.66 Wool 80.750 83.702 96.032 79.887 52.185 2.7 9.93 Metals Aluminium 73.917 76.511 85.587 112.569 106.397 5.1 28.65 Copper 64.282 63.988 90.114 121.975 83.110 5.9 33.15 Lead 72.358 79.311 124.092 89.514 50.104 1.3 7.30 Silver 22.406 26.172 53.749 23.688 24.226 1.7 9.55 Tin 63.191 74.420 84.186 36.277 32.404 2.2 12.36 Zinc 91.958 82.810 99.707 198.817 148.244 1.6 8.99 III. RECONSTRUCTING THE GRILLI AND YANG COMMODITY PRICE INDEX This section provides further information on how to reconstruct the GYCPI commodity price index and the various subindices mentioned in Grilli and Yang (1988) from individual price series. The Index and Subindices The basic GYCPI is a trade-weighted average of all 24 of the commodity price series shown in table 1. In addition, Grilli and Yang (1988) constructed subindices for agricultural food commodities (GYCPIF), nonfood agricultural commodities (GYCPINF), and metals (GYCPIM). The weights are based on each commodity's average export share during the 1977-79 base period and are quoted for the GYCPI as percentage weights in Cuddington (1992) and in table 1. 156 THE WORLD BANK ECONOMIC REVIEW Fi;URE 1. GYCPI/MUV Index and Update 2 1.8 - 1.6 - 1 - 0.8 0.4 0.2 - 1900 1916 1932 1948 1964 1980 1996 - Original series - - Update The composite index is then computed simply as a weighted average of the commodity prices in question as (1) CPI, = aiPi,t i=1 where CPI is the commodity price index in question, n = 24 for the overall GYCPI, a; is the appropriate commodity weight, and Pi,, is commodity i's price in period t indexed to its 1977-79 average. The GYCPI relative to the MUV index and the update undertaken here are shown in figure 1. Weights for the subindices are easily reconstructed from the percentage shares for the overall index.8 The arithmetically weighted index described in equation (1) has been used most frequently in the literature. However, Cuddington and Wei (1992) argued that a geometric aggregation is more appropriate. Such a geometric index would be computed as (2) GPI, = Pa. i= 1 The properties of this alternative index are discussed in depth by Cuddington and Wei (1992). This note reports geometric index alternatives alongside the conventional arithmetic aggregations in the Appendix and in figure 2. 8. With one commodity (commodity 1) used as the numeraire, the ith commodity's weight in any subindex is given by Si S where the i subscript refers to the ith commodity in the relevant subindex and s, is the ith commodity's share in the overall GYCPI. Pfaffenzeller, Newbold, and Rayner 157 FicURE 2. GYCPI/MUV Arithmetic and Geometric Indices 2- 1.8- 1.6- 0.6 - 0.4- 0.2 1900 1916 1932 1948 1964 1980 1996 ---- Geometric weights - Arithmetic weights The GYCPIF, GYCPINF, and GYCPIM subindices are listed together with the GYCPI and the MUV in Grilli and Yang (1988) for the period 1900-86. A comparison of the indices reconstructed on the basis of the weights shown in table 1 with those in Grilli and Yang (1988) shows a close overall correspon- dence for the 1900-86 period. Table 1 lists the percentage shares for each commodity in the overall GYCPI index and the weights for the food, nonfood, and metals indices (last column). The Manufacturing Unit Value index The deflator used alongside the GYCPI index is the manufacturing unit value index, currently implemented as the MUV-G5 index.9 It is a trade-weighted index of the five major developed countries' (France, Germany, Japan, United Kingdom, and United States) exports of manufactured commodities to develop- ing countries. The most frequently used deflator in the literature, the MUV, is also used by the World Bank. As a measure of developing country imports, it is far from perfect. Its use in the present context is based on the rather strong assumption that G-5 manufacturing exports are generally representative of developing country imports. However, the MUV is the only readily available trade-based manufacturing price measure available over a suitably long time horizon, which explains its continued use. Updates of the MUV index were obtained from the World Bank Development Prospects Group, Global Economic Prospects team.1o At the time of writing, the MUV is typically indexed to a 1990 base, whereas Grilli and 9. This index is referred to as either the MUV or the MJV-G5. The MUV-G5 is more specific, since it takes the current definition of the MUV index as an explicit point of reference. Grilli and Yang (1988) refer to the index as the MUVUN. 10. The MUV series from Cashin and McDermott (2002), kindly supplied by Dr Cashin, was used for the 1987-98 period. 158 THE WORLD BANK ECONOMIC REVIEW Yang consistently use the 1977-79 average as their base period. The 1977-79 average for the MUV with a base year of 1990 is 60.008. This figure can be used to reindex the series. IV. CONCLUSION This note has explained how to update the Grilli and Yang index and how to obtain index weights for the various subindices of the GYCPI. This method can also be used to compile new subindices from subsets of the individual data series. In the future, it would seem highly desirable for the World Bank or the IMF to publish updates to the GYCPI and its component indices. Meanwhile, this note should enable interested researchers to extend the Grilli and Yang index series further in the absence of a published updated version. APPENDIX. COMMODITY PRICE INDICES This appendix lists the various commodity price indices11 and the MUV as well as an update of the data published in Grilli and Yang (1988). All series are indexed to their 1977-79 averages. The price indices listed are as follows: GYCPI: the Grilli and Yang commodity price index; MUV: the manufacturing unit value index; GYCPIM: the metals index (aluminium, copper, lead, silver, tin, and zinc); GYCPINF: the index of agricultural nonfood commodities (cotton, hides, jute, rubber, timber, tobacco, and wool); GYCPIF: the index of agricultural food commodities (bananas, beef, cocoa, coffee, lamb, maize, palm oil, rice, sugar, tea, and wheat). Alternative geometric aggregations of the composite indices are identified by a CW suffix in the following table. 11. A spreadsheet with the individual commodity price indices and the composite index series is available in appendix S.1. at http://wber.oxfordjournals.org/. Year GYCPI MUV GYCPIM GYCPINF GYCPIF GYCPI-CW GYCPIM-CW GYCPINF-CW GYCPIF-CW 1900 19.309 14.607 27.778 21.310 15.587 12.866 20.064 11.049 12.014 1901 18.236 13.858 27.522 19.292 14.716 12.008 18.485 10.366 11.232 1902 18.145 13.483 25.518 19.268 15.209 11.878 17.268 10.433 11.220 1903 19.006 13.483 26.668 22.860 14.634 12.173 18.672 11.424 10.938 1904 20.586 13.858 27.526 24.450 16.444 13.070 18.835 11.337 12.459 1905 21.621 13.858 29.150 26.226 16.924 13.624 20.898 11.693 12.793 1906 21.610 14.607 31.726 27.547 15.422 13.759 23.904 12.519 12.057 1907 22.757 15.356 36.699 25.967 16.672 14.089 25.435 12.420 12.386 1908 20.427 14.232 24.245 22.291 18.276 13.526 17.797 11.501 13.410 1909 21.554 14.232 20.822 28.973 18.143 13.787 16.623 12.837 13.443 1910 22.630 14.232 21.026 32.924 18.088 14.224 16.781 13.504 13.834 1911 21.909 14.232 19.923 28.122 19.498 14.773 16.731 12.863 15.195 1912 22.640 14.607 23.176 28.166 19.739 15.611 19.731 13.338 15.642 1913 20.461 14.607 23.134 25.440 17.149 14.592 19.151 13.488 13.893 1914 20.210 13.858 19.291 22.239 19.509 14.642 16.272 13.231 14.878 1915 24.468 14.232 31.321 24.388 22.292 17.723 23.260 15.839 17.158 1916 31.933 17.603 50.327 30.897 26.497 22.237 34.085 21.671 19.614 1917 39.396 20.974 45.271 40.257 37.074 27.255 34.109 30.256 24.070 1918 42.028 25.468 35.121 42.841 43.861 30.731 30.419 35.787 28.595 1919 39.208 26.966 30.853 43.292 39.902 31.150 25.861 34.477 31.464 1920 41.951 28.839 29.684 39.641 47.052 29.631 24.672 32.652 29.968 1921 21.356 24.345 20.219 21.605 21.602 16.904 16.433 18.896 16.145 1922 21.910 21.723 19.919 24.771 21.147 17.176 17.097 19.401 16.195 1923 26.407 21.723 24.587 29.989 25.234 19.615 20.316 22.481 18.128 1924 26.521 21.723 25.066 28.365 26.086 20.319 20.534 20.843 19.996 1925 29.381 22.097 26.315 36.778 26.637 22.112 22.075 24.005 21.243 1926 25.758 20.974 25.962 28.691 24.250 20.117 21.822 20.045 19.628 1927 25.143 19.850 24.028 26.823 24.677 19.759 20.124 19.904 19.570 1928 24.423 19.850 23.585 25.393 24.217 19.970 20.015 19.890 19.995 (Continued) 1 ElOW L Å.Ia-lqåj uo p-un TIUo~~. Iituom~jI ju /2ostnýjjo.åm/d uloij poppEol"Mo Continued Year GYCPI MUV GYCPIM GYCPINF GYCPIF GYCPI-CW GYCPIM-CW GYCPINF-CW GYCPIF-CW 1929 23.266 19.101 25.210 22.332 23.098 19.114 21.143 18.161 18.975 1930 18.277 18.727 21.655 16.949 17.838 15.085 16.727 15.769 14.272 4 1931 13.610 15.356 18.479 12.308 12.675 11.002 12.894 10.130 10.888 1932 10.797 12.734 16.883 8.958 9.734 8.787 10.734 7.722 8.779 1933 12.591 14.232 18.388 12.357 10.833 10.243 13.292 10.291 9.393 1934 15.763 16.854 18.591 16.427 14.522 12.825 14.813 13.623 11.880 1935 17.294 16.479 18.383 16.229 17.465 13.654 15.178 13.270 13.382 1936 18.418 16.479 18.677 18.348 18.369 14.543 15.299 14.632 14.263 1937 21.361 16.854 20.931 20.366 21.988 17.127 18.064 16.840 16.976 1938 16.552 17.603 18.474 16.198 16.105 13.481 15.153 14.173 12.663 1939 16.019 16.105 19.188 17.499 14.267 13.062 16.027 14.749 11.513 1940 17.237 17.603 18.932 20.547 15.063 14.122 16.023 17.897 12.058 1941 20.093 18.727 18.452 24.844 18.288 17.032 16.131 22.107 15.237 1942 23.073 21.723 18.039 27.716 22.419 19.593 16.107 24.761 18.594 4 1943 24.283 24.345 18.132 29.094 23.905 20.605 16.350 26.568 19.583 1944 25.243 27.715 18.132 30.786 24.816 21.278 16.350 28.627 20.010 1945 25.832 28.464 18.232 30.112 26.186 21.504 16.584 27.149 20.843 1946 31.232 28.839 19.485 32.688 34.314 25.501 18.361 29.720 26.293 1947 40.389 34.831 24.709 37.349 46.952 33.426 23.146 33.635 37.532 1948 38.722 35.581 27.980 40.934 41.107 33.168 26.122 37.349 33.790 1949 35.845 33.333 26.479 35.727 38.930 30.331 24.901 30.596 32.192 1950 39.263 30.337 27.767 45.060 40.130 32.653 26.036 36.674 33.175 1951 48.093 35.955 32.466 58.702 47.929 39.647 30.557 48.526 39.031 1952 40.508 36.704 31.825 45.983 40.623 35.287 29.973 41.726 34.241 1953 37.897 35.206 32.214 40.839 38.289 33.477 29.985 36.948 33.041 1954 38.565 34.457 33.066 39.797 39.738 34.216 30.543 35.714 34.752 1955 38.233 34.831 38.267 42.537 36.107 34.213 34.695 38.528 32.116 1956 39.895 36.330 40.977 41.517 38.747 36.664 36.931 38.421 35.741 1957 40.108 36.704 35.376 42.372 40.525 36.585 32.541 39.059 36.790 1958 36.231 36.330 32.546 38.647 36.235 33.525 30.069 36.057 33.498 1959 37.113 36.330 35.379 40.667 35.926 34.707 32.675 37.072 34.256 £I OZ'L ÅIEnJqý4 uo pund piuo ltuo],UOJýul ju /.w ostiopox-iqýk/di uioij papl" 1960 37.327 37.079 36.781 41.799 35.305 35.045 33.763 39.222 33.548 1961 36.466 37.453 35.242 40.424 34.917 34.053 32.738 38.153 32.604 1962 36.486 37.453 34.734 39.893 35.377 33.719 32.650 37.110 32.495 1963 41.419 37.453 34.747 39.084 44.723 36.693 33.158 36.075 38.236 1964 41.046 38.202 37.620 39.782 42.774 38.251 36.349 37.174 39.441 1965 38.119 38.951 40.499 39.990 36.429 35.758 39.095 38.326 33.569 1966 37.935 39.700 40.568 37.445 37.325 35.297 38.953 35.370 34.154 1967 36.846 39.700 41.509 33.813 36.830 34.425 39.874 32.214 33.921 1968 37.431 39.326 43.914 34.620 36.718 35.211 42.213 33.233 34.167 1969 39.761 40.449 47.712 37.459 38.322 37.805 45.324 36.109 36.468 1970 42.201 42.697 53.500 36.438 41.381 40.194 49.763 35.596 39.833 1971 42.324 45.318 50.293 37.638 42.051 40.034 46.982 37.038 39.504 1972 46.625 48.689 49.613 43.823 47.037 43.607 46.895 42.538 43.119 1973 69.472 58.801 55.720 69.054 74.123 63.951 53.118 66.967 66.380 1974 102.410 71.161 79.813 74.718 123.330 84.803 77.415 73.730 93.600 1975 85.156 79.026 76.090 65.807 97.598 73.494 74.452 65.020 77.757 1976 83.110 78.652 81.408 78.946 85.707 80.105 79.964 77.760 81.336 1977 93.125 86.517 87.752 90.681 96.064 92.193 87.316 90.251 94.823 1978 93.627 98.876 91.149 94.173 94.159 93.426 90.943 94.141 93.890 1979 113.250 114.610 121.100 115.150 109.780 112.388 120.204 114.786 108.827 1980 138.830 125.470 144.720 126.490 142.990 128.818 138.845 125.138 127.547 1981 117.940 119.100 124.210 108.870 120.380 113.227 123.156 108.409 112.583 1982 96.784 115.730 110.540 96.727 92.364 94.597 108.241 95.845 89.976 1983 102.780 110.490 118.370 103.150 97.566 100.094 114.464 102.299 94.814 1984 103.540 108.610 112.810 105.290 99.686 100.297 108.995 104.256 95.783 1985 91.268 109.590 105.590 90.490 87.022 88.034 100.879 89.379 83.608 1986 88.358 130.300 105.340 86.026 84.013 84.122 97.014 84.284 80.253 1987 95.215 142.900 108.047 118.203 79.694 90.567 103.931 117.070 76.295 1988 116.574 153.300 155.777 124.230 100.101 109.537 142.006 121.927 95.511 1989 118.705 152.925 151.529 129.335 102.826 108.972 140.342 127.158 93.027 1990 113.918 166.647 135.879 139.309 94.255 102.306 124.454 135.078 83.691 1991 103.689 165.558 111.752 130.249 87.945 94.312 102.724 125.237 79.734 (Continued) ElOW L Åja-nqåj uo p-unläE,uoW~ Ituoil~jul ic /2ostnýjjo.åm/dl uioij p~~opMo" Continued Year GYCPI MUV GYCPIM GYCPINF GYCPIF GYCPI-CW GYCPIM-CW GYCPINF-CW GYCPIF-CW 1992 101.897 171.841 111.151 122.767 88.580 91.564 102.345 117.187 78.179 8 1993 99.068 170.123 95.373 115.683 92.048 89.730 88.804 111.072 81.015 1994 114.839 170.123 115.390 132.639 105.858 109.502 107.210 130.351 101.148 3 1995 128.768 171.841 138.508 154.315 112.983 121.763 126.433 151.673 107.908 1996 123.471 168.075 118.752 146.121 113.797 115.837 112.093 142.603 105.637 1997 120.882 168.634 122.247 142.559 109.720 115.790 112.656 138.579 106.891 1998 106.333 167.617 99.364 125.907 98.909 101.750 94.276 119.959 96.139 1999 93.311 165.445 97.679 115.649 80.850 87.322 92.105 108.425 77.115 2000 92.753 161.945 107.338 112.766 78.136 84.939 99.488 107.261 71.907 2001 88.680 157.179 95.077 106.410 77.842 79.425 88.170 100.669 68.292 2002 92.114 155.212 90.655 112.585 82.463 83.803 84.233 107.987 73.805 2003 98.879 166.853 99.672 127.848 84.297 90.456 93.482 125.158 76.220 £I OZ'L ÆIEflJqý4 uo pund piuo ltuo],UOJv.J ju /.W ostiopox-iqýk/di uioij papl" Pfaffenzeller, Newbold, and Rayner 163 REFERENCES Bleaney, Michael, and David Greenaway. 2001. "The Impact of Terms of Trade and Real Exchange Rate Volatility on Investment and Growth in sub-Saharan Africa." Journal of Development Economics 65(2):491-500. Cuddington, John. 1992. "Long-run Trends in 26 Primary Commodity Prices." Journal of Development Economics 39(2):207-27 Cashin, Paul, and John McDermott. 2002. "The Long-Run Behavior of Commodity Prices: Small Trends and Big Variability." IMF Staff Papers 49(2):175-99. Cuddington, John, and Hong Wei. 1992. "An Empirical Analysis of Real Commodity Price Trends: Aggregation, Model Selection, and Implications." Estudios Econ6micos 7(2):159-179. Economic and Social Data Service (ESDS) International (subscription only). www.esds.ac.uk/inter national/. Grilli, Enzo, and Maw Cheng Yang. 1988. "Primary Commodity Prices, Manufactured Goods Prices, and the Terms of Trade of Developing Countries: What the Long Run Shows." The World Bank Economic Review 2(1):1-47 IMF Primary Commodity Price Tables. www.imf.org/external/np/res/commod/index.asp. Kim, Thae-Hwan, Stephan Pfaffenzeller, Anthony Rayner, and Paul Newbold. 2003. "Testing for Linear Trend with Application to Relative Primary Commodity Prices." Journal of Time Series Analysis 24(5):539-51. Lutz, Matthias. 1999. "A General Test of the Prebisch-Singer Hypothesis." Review of Development Economics 3i(1):44-57. Le6n, Javier, and Raimundo Soto. 1997. "Structural Breaks and Long-Run Trends in Commodity Prices." Journal of International Development 9(3):347-66. World Bank Commodity Price Data (Pink Sheet). www.worldbank.org/prospects. World Bank World Development Indicators. www.worldbank.org/data/. Trade, Production, and Protection Database, 1976-2004 Alessandro Nicita and Marcelo Olarreaga The database described in this article provides researchers with a broad set of data on trade, production, and protection for 28 manufacturing sectors at the three-digit level of the International Standard Industrial Classification, Revision 2. The database covers up to 100 developing and developed countries over the period 1976-2004, but data availability varies by country and year. The trade, production, and protection database is available online and can be freely accessed through the World Bank trade website. JEL code: C8. The trade, production, and protection database includes annual data on trade flows (exports and imports), domestic production (output, value added, employment), and trade protection (tariffs and nontariff barriers) for up to 100 countries over the period 1976-2004. The main contribution of this database is that it merges data from different sources in a common industry classifi- cation. The data are disaggregated into 28 manufacturing sectors, correspond- ing to the three-digit level of the International Standard Industrial Classification, Revision 2 (ISIC Rev. 2). The database is available for download free of charge on the World Bank trade website (www.worldbank.org/trade; click on the 'Data & Statistics' tab). The database updates the earlier release made available in Nicita and Olarreaga (2001). Besides the longer time coverage, the database has been improved in a number of ways. The coverage has increased to 100 countries. The concordance table between Standard International Trade Classification, Revision 2 (SITC Rev. 2) and ISIC Rev. 2 has been updated, and more Alessandro Nicita (corresponding author) is a consultant in the Development Economics Research Group (Trade) at the World Bank; his email address is anicita@worldbank.org. Marcelo Olarreaga is a senior economist in the Office of the Chief Economist for Latin America at the World Bank; his email address is molarreaga@worldbank.org. The authors are grateful to three anonymous referees, the editor of the World Bank Economic Review, and many users of the earlier release of the database for their comments and suggestions. They also thank the World Bank's Research Support Budget for funding. A supplemental appendix to this article is available at http://wber.oxfordiournals.org/. THE WORLD BANK ECONOMIC REVIEW, VOL. 21, NO. 1, pp. 165-171 doi:10.1093/wber/1hl012 Advance Access Publication 31 January 2007 () The Author 2007. 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 165 166 THE WORLD BANK ECONOMIC REVIEW variables have been added (nontariff barriers and their ad valorem equivalents, elasticities of import demand, import price indices, import-weighted tariffs and their standard deviation). The database has also been made more compact by not including the ISIC four-digit classification. A supplemental appendix to this article (available at http://wber.oxfordjournals.org) describes the different dimensions of the database, variable definitions, and data availability in more detail. Researchers may wish to use this database jointly with others available on the web (see online resources in the reference list for details). These include the trade-related data provided by Jon Haveman and Raymond Robertson; data provided by the Centre d'etudes prospectives et d'informations internationales (CEPII), in particular the extension by Mayer and Zignago (2005) of the earlier release of the trade and production database; and another wide collec- tion of trade and protection data provided by the University of California at Davis, Center for International Data. I. PRODUCTION The source of domestic production-related data is the United Nations Industrial Development Organization (UNIDO), which collects annual data from member countries. These data are published annually in the International Yearbook of Industrial Statistics using the three-digit level of ISIC Rev. 2. The production-related data of this database include information on output, value added, gross fixed capital formation, index of real output, wage bill, number of establishments, number of employees, and number of female employees for each of the 28 manufacturing sectors. Both country and time coverage of the production-related data are limited. Some countries, especially in the developing world, report data only sporadi- cally and do not report on all the variables mentioned above. UNIDO makes a great effort to standardize the data so that they are comparable across countries and years, but some problems persist. Yamada (2005) provides detailed information on the UNIDO data and issues related to its use in research projects. II. TRADE The source of trade data is the commodity trade statistics database (Comtrade) kept by the United Nations Statistic Division. Comtrade provides trade data using the SITC Rev. 2. These data are converted into ISIC Rev. 2 using a con- cordance table. Trade data are reported at both the aggregate and the bilateral level for exports and imports. Country and time coverage are both very complete. However, there are some missing observations, especially among developing countries. To fill missing observations, researchers often resort to mirrored data (see, for example, Feenstra and others 2005). For example, the Nicita and Olarreaga 167 missing export data of country C are calculated using imports from country C reported by its trading partners. The database provides mirrored data for both imports and exports. Individual researchers can determine whether and how to use mirrored data. The trade data include information on the value of shipments (in thousands of U.S. dollars) and physical quantities (in kilograms). The database also con- tains unit values, measured in dollars per kilogram, calculated as the ratio of the value of shipments and physical quantities at the three-digit level of ISIC Rev. 2.1 They are provided for exports and imports at the aggregate and bilat- eral level. Important cautionary notes regarding the use of mirrored data and unit values are provided in section V. III. PROTECTION The protection data in this database include data on tariffs and nontariff bar- riers. The main source of protection data is the United Nations Conference on Trade and Development (UNCTAD) Trade Analysis and Information System (TRAINS). The authors also collected additional data from national statistical documents and websites. Protection data availability starts with 1988. These data are received at the Harmonized System (HS) six-digit level and converted to the three-digit level of ISIC Rev. 2 using a concordance table. Both country and time coverage of the protection data are far from complete, especially for the data on nontariff barriers. The tariff data contain simple and import-weighted average tariffs for the 28 manufacturing sectors. Standard deviation and maximum and minimum values at the six-digit level of the HS are also reported within each ISIC code for applied tariffs and most favored nation tariffs. Applied tariffs take into con- sideration the available data on preferential schemes and are therefore calcu- lated at the aggregate level using bilateral tariff and import data.2 Most favored nation tariffs are the rates granted to all World Trade Organization members to which no preferential access is granted. Data on nontariff barriers are reported as a single category core nontariff barrier (Core NTB), which includes price-control measures, finance-control measures, and quantity-control measures. Nontariff barrier data are reported using coverage ratios (the percentage of imports subjected to nontariff barriers) 1. For about 5 percent of products at the SITC five-digit level, quantities are not reported at all or are reported in number of units rather than metric weight. The aggregate quantities at the ISIC level do not take these products into account. 2. Bilateral protection data cover most preferential trade agreements, but not all. Researchers interested in bilateral tariffs should refer directly to the UNCTAD TRAINS database, available through World Integrated Trade Solutions (WITS) (wits.worldbank.org). An alternative source of bilateral tariff data that includes a careful calculation of ad valorem equivalents of specific tariffs is Bouet and others (2004). 168 THE WORLD BANK ECONOMIC REVIEW and frequency ratios (the percentage of tariff lines subjected to nontariff barriers). This database also provides simple and import-weighted averages of nontariff barriers at the three-digit level of the ISIC. The methodology to obtain the averages of nontariff barriers is described in Kee, Nicita, and Olarreaga (2006). IV. OTHER DATA Trade, production, and protection data are often used in gravity models. To facilitate work with these types of models, this database includes a number of gravity-type variables: geodesic distance between national capitals, language, GDP, GDP per capita (adjusted for purchasing power parity), and dummy variables for common language, shared border, being landlocked, and being an island. Some of these data come from the World Bank's World Development Indicators, and some have been collected and constructed by the authors. Researchers working with gravity models may find additional data at CEPII's website: http://www.cepii.fr/anglaisgraph/bdd/distances.htm and at Andrew Rose's website: http://faculty.haas.berkeley.edu/arose/RecRes. htm#Software. The trade, production, and protection database also includes import demand elasticities for each country and each of the ISIC codes at one point in time. These data are particularly useful for simulation exercises. The methodo- logy used to estimate the import demand elasticities is described in detail in Kee, Nicita, and Olarreaga (2004). Import demand elasticities are provided in a separate file. The database also contains information on input-output tables from the Global Trade Analysis Project (GTAP) database version 4, which is based on data from the early 1990s.3 Because the GTAP 4 database aggregates some countries by region, countries in the same GTAP region will have the same input-output table. To give more flexibility to users of these tables, the data are split into two tables.4 One table reports the share of each manufacturing sector output that is sold as an input to the production of each sector. Another table reports the amount of all of the other sectors' output necessary to produce one unit of final output in each sector. Because the GTAP industry dis- aggregation does not exactly match the ISIC Rev. 2 three-digit level industry disaggregation, the input-output data are provided at a higher level of aggregation. 3. The GTAP database is now in its sixth release. The use of newer releases of the GTAP database permits building updated input-output tables. For information on how to access the GTAP databases, visit www.gtap.org. 4. The supplemental appendix provides more detail on the construction of these tables. Nicita and Olarreaga 169 V. SPECIAL CONSIDERATIONS The data in the trade, production and protection database are organized to facilitate its use for many different purposes. The objective of the database is not to produce quick answers for researchers, but rather to ease the lengthy and cumbersome exercise of collecting and organizing data into a common classification. To use the database meaningfully, it is important that researchers be aware of its limitations. The database is an unbalanced panel, containing many missing observations. Thus comparisons must be made with care. Industry or country averages may not be very meaningful, for example, if they correspond to different time periods or contain different countries in different years. While missing data may be interpolated, the decision on whether and how to do so is left to the researcher. A second issue relates to the use of mirrored data. In theory, export data are recorded as free on board (f.o.b.), while import data are recorded as cost, insurance, and freight (c.i.f.). That makes it appealing to use the difference between the bilateral import value and its corresponding export value to impute trade costs. In practice, however, the differences between a trade flow (whether imports or exports) and its mirrored counterpart should not be con- sidered a good measure of trade costs, especially for countries with weak customs capacity. In many cases the discrepancies between the two values are attributable to many reasons besides trade costs, such as customs corruption, underinvoicing, weak accounting methods, existence of entrep6ts, and different product classifications. And for trade aggregates discrepancies may be due to missing reporting partners in the mirrored data (not all trade partners report trade to Comtrade). Consider the existence of entrep6ts (countries that are neither the origin nor the final destination of trade but through which transits of trade occur), that may create accounting discrepancies between reported data and mirrored data (Hanson and Feenstra, 2001). In some cases the country of origin mistakenly reports the entrep6t as the destination. Meanwhile, the entrep6t country does not report the import, and the final importer reports the original exporter as the country of origin. This creates discrepancies when bilateral imports and exports are compared. The researcher should keep this in mind, especially when analyzing bilateral trade flows that may involve such entrep6ts as Hong Kong, China; Macao, China; Singapore; and the Netherlands. In some cases, however, mirrored data may be considered of better quality than reported data, such as when the partner has much better customs adminis- tration than the reporter. See Yeats (1995) for a discussion. Another important consideration relates to a few cases (about 1 percent of observations) in which the value of exports is larger than the sum of output plus imports. This could arise for several reasons. First, there could be discre- pancies between the year of production and trade flows if goods produced one 170 THE WORLD BANK ECONOMIC REVIEW year are exported the next year. Second, production data may be misallocated across ISIC categories. Third, for some countries reported production data may exclude a significant portion of industrial activity because coverage of small-scale establishments is incomplete or because the data refer only to a certain area of the country or only to part of the manufacturing sector (exclud- ing the informal sector, for example). Thus, researchers should be attentive to the possibility of measurement error. Protection data also raise some issues. First, while the applied tariff data take into account the preferential access schemes of developed countries, some of the smaller agreements between developing countries may not be included. This issue is usually more relevant for data for early years. Second, there is no systematic information on preference scheme utilization rates. The data assume that the schemes are fully utilized, but some are not, often because of an inability to meet origin requirements. Third, tariff data include only the ad valorem component of tariff schedules, with no ad valorem equivalents for specific duties. This is not a major omission, however, because the database focuses on manufacturing, where specific duties are rare. Fourth, the database provides simple and import-weighted average tariffs. Neither of these has a sound theoretical basis as a measure of trade restrictiveness. For example, in calculations of import-weighed average tariffs, goods subject to prohibitively high tariffs have zero weight, underestimating trade restrictiveness. Similarly, very low tariffs on economically meaningless goods downwardly bias simple average tariffs as a measure of trade restrictiveness. For a detailed discussion and some solutions to these problems, see Kee, Nicita, and Olarreaga (2006). There is also an important caveat on the use of unit values. Unit values are useful for analyzing many international trade issues, in particular price compe- titiveness. However, they should be used with caution as they are a noisy proxy for prices.5 This is particularly the case for large product aggregates, such as those in this database. The main reason is that changes in product quality or product mix can affect average unit values. There are no straightforward sol- utions to these problems but rather a need for awareness and caution when using the data. In econometric work, one way out of some of these issues is to instrument unit prices to get rid of measurement error. VI. TECHNICAL INFORMATION Most of the data are stored as ASCII files and can be read with any text editor or statistical software. A supplemental appendix to this article describes the different dimensions of the database, variable definitions, and data availability in more detail. It can be found at http://wber.oxfordjournals.org or on the World Bank Trade website (www.worldbank.org/trade). 5. In practice, sudden jumps in the time series of unit values, when not substantiated by other data, are usually an indication of a change in the accounting or recording methods. Nicita and Olarreaga 171 REFERENCES Bouet, Antoine, Yvan Decreux, Lionel Fontagn6, S6bastian Jean, and David Laborde. 2004. "A Consistent, Ad-valorem Equivalent Measure of Applied Protection across the World: The MacMap-HS6 Database." CEPII Working Paper 2004-22. Paris: Centre d'&tudes prospectives et d'informations internationals. Centre d'tudes prospectives et d'informations international. Additional data for gravity models. http:// www.cepii.fr/anglaisgraph/bdd/distances.htm. Centre d'etudes prospectives et d'informations internationals. Extension by Mayer and Zignago (2005) of the earlier release of the trade and production database. www.cepii.fr/anglaisgraph/bdd/ TradeProd.htm. Feenstra, Robert, Robert E. Lipsey, Haiyan Deng, Alyson C. Ma, and Hengyong Mo. 2005. "World Trade Flows: 1962-2000". NBER Working Paper 11040. Cambridge, Mass.: National Bureau of Economic Research. Hanson, Gordon H., and Robert C. Feenstra. 2001. "Intermediaries in Entrep6ts Trade: Hong Kong Re-exports of Chinese Goods". NBER Working Paper 8088. Cambridge, Mass.: National Bureau of Economic Research. Haveman, Jon, and Raymond Robertson. Additional trade-related data. www.macalester.edu/research/ economics/PAGE/HAVEMAN/Trade.Resources/TradeData.html. Kee, Hiau Looi, Alessandro Nicita, and Marcelo Olarreaga. 2004. "Import Demand Elasticities and Trade Distortions." Policy Research Working Paper 3452. Washington, D.C: World Bank. -. 2006. "Estimating Trade Restrictiveness Indices." Policy Research Working Paper 3840. Washington, D.C: World Bank. Mayer, Thierry, and Soledad Zignago. 2005. "Market Access in Global and Regional Trade." CEPII Working Paper 2005-02. Paris: Centre d'6tudes prospectives et d'informations internationals. Rose, Andrew. Additional data for gravity models. http://faculty.haas.berkeley.edu/arose/RecRes. htm#Software. Nicita, Alessandro, and Marcelo Olarreaga. 2001. "Trade and Production: 1976-1999." Policy Research Working Paper 2701. Washington, D.C: World Bank. United Nations Conference on Trade and Development (UNCTAD). TRAINS database. Data on bilat- eral tariffs. www.worldbank.org. University of California at Davis, Center for International Data. Wide collection of trade and protection data. cid.econ.ucdavis.edu. World Bank. Various years. World Development Indicators. Washington, D.C. Yamada, Tetsuo. 2005. "Relevance and Applicability of the UNIDO Industrial Statistics Database for Research Purposes." UNIDO ESA/STAT/AC.105/21. Vienna. Yeats, Alexander. 1995. "Are Partner Country Statistics Useful for Estimating 'Missing' Trade Data?" Policy Research Working Paper 1501. Washington, D.C: World Bank. Forthcoming papers in THE WORLD BANK ECONOMIC REVIEW * Land Tenure, Investment Incentives and the Choice of Techniques: Evidence from Nicaragua. Oriana Bandiera * Child Labor, School Attendance and Intra-household Gender Bias in Brazil Patrick M. Emerson and Andre Portela Souza * Is Land Titling in Sub-Saharan Africa Cost-Effective? Evidence from Madagascar Hanan G. Jacoby and Bart Minten * The Anarchy of Numbers: Aid, Development, and Cross-country Empirics David Roodman * Tracking Poverty Over Time in the Absence of Comparable Consumption Data David Stifel and Luc Christiaensen * Are Remittances Insurance? Evidence from Rainfall Shocks in the Philippines Dean Yang and Hwajung Choi * Incremental Reform and Distortions in China's Product and Factor Markets Xiaobo Zhang and Kong-Yam Tan 個―一－ &- 、｀ 乞L 必Z 尸～ j 鬥日 騙 & && 斗 柑丰》L 日辟革 仕沫任 之rn二邢 卜Zr侃二」日 &，才〞鬥亡 『k&O一鬥一 必州巴j，〕亡〞 尸黜、‘戶｀騙戶 一～j尸r→→ 卜＇，、州－& &&＇、j‘州了 一于＇卜＇f拙 必弋目＇他 化日日荊上 才｀細C二》 江士計丰叩 :＿斤一日匹蔆 鬱弄喜審賽亡， ，■■一■■一■一■■一－■