WPS6590 Policy Research Working Paper 6590 Foreign Bank Behavior During Financial Crises Jonathon Adams-Kane Julian A. Caballero Jamus Jerome Lim The World Bank Development Economics Prospects Group September 2013 Policy Research Working Paper 6590 Abstract One of the persistent policy problems faced by developing countries, using a unique bank-level database governments contemplating financial liberalizations is of foreign ownership. In particular, the authors examine the question of whether to allow foreign banks entry whether the credit supply of majority foreign-owned into the domestic economy. This question has become financial institutions differ systematically conditional on ever more urgent in recent times, due to rapid financial a crisis event in their home economies. They show that globalization, coupled with the credit contractions foreign banks that were exposed to crises in their home experienced as a result of the 2007/08 financial crisis. countries exhibit changes in lending patterns that are This paper examines the question of whether opening the lower by between 13 and 42 percent than their non-crisis financial sector to foreign participation is a good idea for counterparts. This paper is a product of the Prospects Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at jonathonak@gmail.com, julianc@iadb.org, and jlim@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Foreign Bank Behavior During Financial Crises an A. Caballero Jonathon Adams-Kane, Juli´ & Jamus Jerome Lim∗ Keywords: Foreign bank ownership, �nancial crisis, bank lending JEL Classification: G21, G01, F34 Sector Board: EPOL ∗ The authors are with the Development Prospects Group at the World Bank, Humboldt State University, and the Inter-American Development Bank. Their respective emails are: jonathonak@gmail.com, jlim@worldbank.org, and julianc@iadb.org. Financial support from the KCPII Window 2 Grant TF095266 “Analyzing the Impact of the Financial Crisis on International Bank Lending to Developing Countries� was critical in supporting the data collection process. The paper has bene�ted from comments from Samuel Berlinski, Steve Kaplan, Andrew Powell, Hans Timmer, and Jasmine Xiao. The �ndings, interpretations, and conclusions expressed in this article are entirely those of the authors. They do not necessarily represent the views of the World Bank or the Inter-American Development Bank, their Executive Directors, or the countries they represent. A banker is a fellow who lends you his umbrella when the sun is shining, but wants it back the minute it begins to rain. Mark Twain American author and humorist (1835–1910) 1 Introduction ao VI set sail from Brazil to Portugal in an attempt On April 25, 1821, then prince regent Dom Jo˜ to deal with a revolution that was underway there, carrying with him a large part of the deposits of the Banco do Brasil, the colony’s major �nancial institution. The bank, which was already in ao’s crisis as a result of its close ties with the Portuguese Crown, was left bankrupt as a result of Jo˜ actions. Clearly, the concern that foreign banks may flee when their home countries experience difficult times is neither unwarranted nor unprecedented. Indeed, over the long course of history, governments have often weighed potential liquidity and growth bene�ts of foreign bank presence against fears that such banks may prove unreliable sources of capital in times of crisis. Policymakers in developing countries seeking to liberalize their �nancial sectors are routinely called on to decide whether foreign banks are to be allowed into their domestic �nancial markets, and if so, to what extent such banks have the freedom to operate a-vis domestic banks. vis-` This paper seeks to contribute to the literature on foreign bank presence in developing countries by asking whether foreign banks do indeed make different credit provision choices when their home economies are undergoing hard times. In particular, we examine whether the lending activity of majority foreign-owned �nancial institutions that experienced a crisis in their home countries differ systematically in their lending behavior relative to foreign-owned institutions that did not, within the general setting of the global �nancial crisis of 2007/08. Whether foreign-owned banks choose to scale back on their lending activity in such circum- stances is far from obvious. Foreign subsidiaries experiencing a crisis in their home country may choose to repatriate capital to an ailing parent bank, but it is just as plausible that parent banks reallocate asset portfolios toward markets relatively less affected by the crisis. The issue of foreign bank lending during �nancial crises is thus, ultimately, an empirical question. Our empirical exploration seeks to answer this question by relying on a quasi-experimental difference-in-difference (DiD) approach. Our baseline sample draws on a unique bank ownership dataset collected across countries and over time, and comprises 361 foreign-owned banks based in developing countries over the course of the recent 2007/08 global �nancial crisis and in the immediate pre- and post-crisis years (2006 and 2009). We de�ne our crisis “treatment� as a �nancial or banking crisis (Laeven and Valencia, 2012) experienced in the home country of the foreign-owned 2 bank. Crucial for our identi�cation strategy is the fact that, while �nancial crises experienced in the home economy may have been closely tied to the performance of banks based in the crisis economy, foreign subsidiaries of these banks are unlikely to have contributed to the crisis there, so that the crisis event was an “import� from high-income countries. From the perspective of these banks, then, the �nancial crisis was essentially an exogenous event, just as it was for other foreign banks situated within the host economy, with the crucial difference being that the former group could subsequently be subject to potential constraints resulting from the home-country crisis—such as the need to repatriate pro�ts to their parent banks—that foreign banks not facing similar shocks in their home economies would not experience. We exploit this exogenous variation to identify the effect of a home-country crisis on foreign bank lending behavior in our baseline difference-in-difference speci�cation. We further re�ne our baseline estimate by comparing pairs (or small groups) of particularly comparable foreign banks via a DiD design that matches them on a number of observables. In a host of robustness checks, we consider alternative strategies designed to isolate the causal effects of the crisis treatment, such as the inclusion of additional bank- and country-level controls, falsi�cation tests that consider whether alternative non-crisis mechanisms may be driving the results, and exploring various dimensions of heterogeneity in the crisis effect among the foreign banks. These complementary methodologies thus allow us to both identify the causal effect of a crisis in a foreign bank’s home country on the change in the bank’s lending activity before and after the crisis, as well as provide some sense of whether certain bank- or country-speci�c characteristics may have contributed to the estimated avaerage treatment effect on the treated. Our main result is that foreign banks owned by countries experiencing crises do in fact experience a post-crisis change in their lending that is relatively lower—by between 13 and 42 percentage points in our baseline—compared to non-crisis foreign banks. Thus, while foreign banks have, on average, been a force for �nancial stability in developing countries facing local �nancial crises (Clarke et al., ınez-Peria et al., 2005; Wu et al., 2011), this is not the 2003; de Haas and van Lelyveld, 2010; Mart´ case when the crisis originated from the foreign bank’s home country. Thus, rather than expanding lending in an attempt to diversify away from the shock experienced in their home countries, such banks probably repatriate capital to shore up the liquidity of their parents, or endure contractions in liquidity from their parents. When we explore the issue of heterogeneity among foreign banks further, we also �nd evidence suggesting that non-crisis foreign bank lending may have helped offset reductions in post-crisis lending by crisis-stricken foreign banks and domestic banks, and that the crisis that faced foreign banks in Eastern Europe was especially severe. The empirical literature on bank ownership and economic outcomes has grown dramatically over the past decade. However, in part due to data limitations, much of the literature tends to study a given country or region. Some of these studies have, like this one, been concerned with foreign bank 3 behavior during a crisis.1 For example, Peek and Rosengren (1997, 2000) document a reduction in lending by Japanese banks in the U.S. after the bust of the Japanese stock market in 1990; while Chava and Purnanandam (2011) and Schnabl (2012) present evidence of negative spillovers via foreign banks of the Russian crisis of 1998 to �rms in the U.S. and Peru, respectively. Similarly, de Haas and van Lelyveld (2006) study how foreign banks in Central and Eastern Europe responded to banking crises there. In the setting of the crisis of 2007/08, Galindo et al. (2010) document negative spillovers of foreign banks in Latin America, Popov and Udell (2012) and Ongena et al. (2012) in Eastern Europe, Aiyar (2012) and Rose and Wieladek (2011) in the UK, and Cetorelli and Goldberg (2012b) in the USA. In contrast to the relatively narrow geographic focus of these papers, our country coverage includes 51 developing economies, across all regions of the world. Relatively few papers have considered the speci�c issue of the influence of foreign bank ownership on credit across a wider range of countries (Cetorelli and Goldberg, 2011; Claessens et al., 2001; Clarke et al., 2006; Detragiache et al., 2008; Van Rijckeghem and Weder di Mauro, 2003). However, in these papers, foreign bank presence in an economy is typically measured at an aggregate level, rather than the bank- and home-country-speci�c level we employ in this paper (which permits us to map banks to home-speci�c shocks). To the extent that some papers have worked with bank-level data, their bases for compari- son have been different: de Haas and van Lelyveld (2010), for example, restrict their analysis to only subsidiaries of the 45 largest multinational banks, while Galindo et al. (2010) focus on Latin American host countries. Similarly, de Haas and van Lelyveld (2013) and Claessens and van Horen (2013) are concerned with benchmarking lending by foreign subsidiaries of multinational banks against that of domestic banks. Since we are interested in the effects of a crisis in the home country on lending activity by foreign banks, our study restricts itself to only the subset of foreign-owned banks operating in developing countries, since we believe that foreign banks from non-crisis home countries offer the purest control group for our treatment of interest. To our knowledge, the papers that are closest in approach to this paper are Wu et al. (2011), and three sets of papers by Peek and Rosengren (1997, 2000), de Haas and van Horen (2012a,b), and Giannetti and Laeven (2012a,b).2 Like this paper, Wu et al. (2011) consider foreign ownership at the bank level across emerging economies, and whether such banks respond differentially to exogenous shocks. However, the paper is primarily concerned with the effect of monetary shocks experienced 1 Other studies have explored the behavior of foreign banks in developing countries during normal times, or across the business cycle. For instance, Dages et al. (2000), Mart´ınez-Peria et al. (2005), Clarke et al. (2005), and Galindo et al. (2005) study the influence of foreign bank presence on lending patterns in Latin America, while Gormley (2010) does the same for India. Detragiache and Gupta (2006) and Khwaja and Mian (2008) examine episodes of liquidity shortages in Malaysia and Pakistan, respectively. 2 We should note that Micco et al. (2007) also approach the foreign ownership issue from a bank-level perspective. But the substantive focus of their paper is different, being concerned more about bank performance, rather than credit provision. 4 in the host economy, while our focus is instead on shocks experienced in home economies. The Peek and Rosengren (1997, 2000) papers are concerned with shocks experienced in foreign banks’ home countries, but the papers are essentially case studies of a pair of major developed economies (Japanese banks in the United States), as opposed to our more general developing country focus. Finally, while the concerns of the �nal four papers do overlap with ours, the analyses mainly rely on datasets with somewhat weaker coverage (for example, on cross-border syndicated lending), and, most importantly, are not focused on addressing causal concerns in a systematic fashion, which we regard as our central contribution. The paper is organized as follows. In the following section, we discuss the relevant theory underlying foreign bank lending during crises. This is followed by a description of our dataset and its main stylized features of banks during the �nancial crisis of 2007/08 (section 3). Section 4then outlines our econometric setup, along with a discussion of identi�cation issues. Our baseline results are reported in section 5, and robustness checks in section 6. In section 7 we explore heterogeneity among banks, for foreign relative to domestic banks, and between foreign banks. A �nal section concludes with policy implications and avenues for future research. 2 Foreign and domestic bank lending in a �nancial crisis In this section we discuss the main mechanisms by which foreign and domestic banks may differ in their lending behavior, and the channels by which a crisis may affect lending activity. There is no single, well-established theory of how foreign banks’ characteristics or lending deci- sions can be expected to differ systematically from those of domestically owned banks, nor of how a crisis in a foreign bank’s home country can be expected to influence its lending. However, some mechanisms by which shocks are transmitted through international banking have been discussed in the literature.3 One important consideration is that subsidiary banks—whether they are part of a multinational or domestic banking group—typically do not operate completely independently of their parent company. In the case of a multinational banking group, this has two main implications in regards to the transmission of a shock within the group, as emphasized by Morgan et al. (2004).4 On one hand, when a foreign-owned bank is hit by a crisis in its parent’s home country, the shock in the home country may be cushioned by repatriation of capital from the bank to its parent, or by reallocation to other subsidiary banks in the group that were relatively exposed to the shock (termed a “support effect� by de Haas and van Lelyveld (2010)). Similarly, when the parent encounters 3 Goldberg (2009) provides an overview of key international spillovers through banking, including some mechanisms of crisis transmission. 4 Although the model of Morgan et al. (2004) is applied to studying crises in host rather than home countries, the mechanisms in the model are applicable to the case of a crisis in the home country. de Haas and van Lelyveld (2010) and Cetorelli and Goldberg (2011) present similar ideas from the perspective of the balance sheet of a multinational bank. 5 liquidity problems, the parent may pass these on by supplying less liquidity to the subsidiary via the group’s internal capital market (Cetorelli and Goldberg, 2012a). This is analagous to a wealth or income effect. On the other hand, when the parent bank faces a less favorable risk-return tradeoff in its home country, there is a substitution effect as well. The parent has an incentive to reallocate its portfolio of assets toward countries less affected by the crisis; that is, to reallocate liquidity to its subsidiaries in relatively safe havens so that more loans can be made there, rebalancing the group’s loan portfolio in favor of these countries, and helping to shore up its overall balance sheet. Which of these two effects dominates determines the net effect of a crisis in the home country on a foreign bank’s access to liquidity and thus the net effect on its lending behavior, which is therefore ambiguous.5 The answer to this question is ultimately empirical. One strategy for addressing this question is to compare the lending behavior of foreign banks which face a crisis in their home countries to the lending behavior of domestic banks, and infer whether there are any systematic differences between post-crisis lending activity in the two groups. But this could be problematic, since foreign banks likely differ from domestic banks in systematic ways in terms of their pre-crisis characteristics, which could also make a difference in terms of how a crisis affects their lending. For example, some domestic banks are state-owned, and these may have a political mandate to cushion the economy from shocks by lending countercyclically (Bertay et al., 2012). In the context of asymmetric information, which is probably especially relevant in developing countries, foreign banks may also face greater costs of acquiring information about borrowers, potentially leading to “cherry picking� the most attractive, or largest, clients (Dell’Ariccia and Marquez, 2004; Detragiache et al., 2008). Domestic and foreign banks may also tend to differ in size (as shown in subsection 3.2), capital structure, sources of funding, pursuit of longer client relationships versus “transaction-by-transaction� lending, and degree of lending to foreign vs. domestic �rms. All these differences between foreign and domestic banks also suggest that the two groups likely face different demand schedules for loans in a given host economy. Thus, there is little reason to expect that lending by domestic and foreign banks in developing countries would have shared similar trends in lending between 2006 and 2009 had the crisis not occurred. This is the main reason why, in answering the question of how foreign banks’ lending is affected by a crisis in its home country, we consider it more appropriate to compare these crisis-stricken foreign banks to other foreign banks –those with non-crisis home countries of ownership– rather than domestic banks. This is the crux of the empirical strategy that we adopt in this paper. 5 Indeed, Cetorelli and Goldberg (2012c) present evidence that the operation of internal capital markets of U.S. banks in 2007 and 2008 was quite heterogeneous and depended on bank and host-country speci�c conditions. Some host markets operated as “funding� markets for some banks, seeing larger net flows to parent banks; while other host economies operated as “investment� markets, with increased net internal flows from parent to subsidiary. 6 3 Foreign Banks and the 2007/08 Financial Crisis 3.1 Data source and description The dataset used in this paper is based on an extensive data collection effort on banking sector evolution in developing countries for the period of 1995–2010. This dataset, in turn, builds on a previous dataset compiled by Claessens et al. (2008), which includes data for the decade between 1995–2005 for a smaller set of developing countries (about two-thirds of the current coverage).6 The coverage is for 4,496 commercial banks, saving banks, cooperative banks and bank holding companies in 131 developing countries.7 The information sources used to build the dataset include Bankscope (the primary source), supplemented by individual banks’ websites and annual reports, banking regulation agencies’ publications and announcements, parent companies’ reports, and news articles. A bank is de�ned as foreign-owned if 50 percent or more of its shares are directly owned by foreign entities. Majority ownership is assessed annually based on shareholder information at the end of the year, or as close to the end of the year when sufficient data are available. Nationality of ownership is based on direct ownership, except in certain cases when ultimate ownership is used.8 Ultimate ownership was used when the main country of direct foreign ownership was a tax haven (classi�ed by the OECD as a tax haven or as an OECD member country with a potentially harmful preferential tax regime), the owner(s) in the tax haven operated the �rm purely as a holding company, and total foreign ownership was 50 percent or more; or when the bank was majority owned by a single owner which functions purely as a holding company, and which was fully owned by a third entity; or when a bank was transferred from its parent to another of the parent’s subsidiaries for the purpose of being absorbed by that other subsidiary that year. If the majority of shares of a bank are held by foreigners but no single nationality accounts for a majority, then the foreign country with the highest share is considered the nationality of ownership. We now turn to our de�nition of a banking crisis, which relies on the banking crisis database of Laeven and Valencia (2012). In this database, a banking crisis is de�ned as a systemic banking crisis when two conditions are met:9 �rst, there are signi�cant signs of �nancial distress in the banking system (as indicated by signi�cant bank runs, losses in the banking system, and/or bank liquidations); and second, signi�cant banking policy intervention measures were undertaken in response to losses in the banking system. Importantly, this de�nition does not include isolated banks in distress. 6 The dataset has also been independently updated by Claessens and van Horen (2013), to the year 2009. We allude to some of the differences between this dataset and our own below, but the coverage is substantially similar. 7 The observations are at legal entity level, and cover both branches and subsidiaries. 8 Additional detail on the rules used to construct this variable is provided in the technical appendix. 9 A detailed description of the construction of this variable is provided in the technical appendix. 7 The starting year for a systemic banking crisis is identi�ed by the two conditions just mentioned, along with the ful�llment of at least three out of the following six policy interventions (Laeven and Valencia, 2012, p. 4): extensive liquidity support; large bank restructuring costs; signi�cant bank nationalizations; signi�cant asset purchases; signi�cant guarantees put in place; or deposit freezes and bank holidays.10 Because the quantitative thresholds used in this de�nition of systemic banking crises are ad hoc, events that almost meet these thresholds are classi�ed as “borderline.� With the methodology just described, Laeven and Valencia (2012) identify 147 crises in 115 countries for the period 1973–2009. Of these crises, thirteen events are classi�ed as borderline. We combine these ownership and banking crisis variables to construct our crisis treatment effect, which is our main independent variable of interest. We de�ne the crisis treatment as an indicator variable for every foreign-owned bank that takes on the value of unity when its main country of ownership experienced a banking crisis in the years 2007–2009, and zero otherwise. The baseline de�nition includes all systemic crises identi�ed by Laeven and Valencia (2012) for this period.11 To avoid confounding home and host country crises, we also exclude from the pool of host countries all crisis-hit countries. The baseline de�nition yields a total of 17 systemic banking crises. To enhance the quality of the data, we re�ne our working sample in several additional ways. First, we consider only host countries where there is at least one operating bank from a crisis- stricken country and at least one foreign bank from a non-crisis country, so that a comparison can be made between these two groups. Second, we drop from the sample all host countries that have less than �ve operating banks (after excluding the cases mentioned before), so that our results are not driven by unrepresentative outliers. Finally, we exclude foreign banks that change their main country of ownership between 2006 and 2009, so that the country effect is well de�ned. The resulting sample comprises 361 foreign banks from 66 home countries, operating in 51 host countries. Of these banks, 208 are treated banks (their main country of ownership is one of the 17 countries that faced a systemic banking crisis in 2007/08), and the remaining 153 banks are controls.12 We merge our crisis treatment variable with additional information drawn from the Bankscope database, which includes year-by-year balance sheet and performance information for each bank. Our main dependent variable is a bank’s total outstanding loans, net of reserves for impaired or nonperforming loans. The core set of bank-level controls includes bank size, solvency, the interest margin, and the income-to-loan ratio; these are measured in standard ways. Size, for example, 10 When a country has faced �nancial distress but less than three of these measures have been used, the event is nevertheless classi�ed as a crisis if either the country’s banking system exhibits signi�cant losses due to nonperforming loans or bank closures, or if �scal restructuring costs of the banking sector exceed �ve percent of GDP. 11 The baseline de�nition also includes the borderline crises of France, Portugal and Slovenia, because the banking systems of these Eurozone countries are highly integrated with those of other Eurozone countries that experienced nonborderline systemic crises, such as Austria, Germany, Italy, Spain and UK. In robustness checks, we show that our baseline results hold after including all borderline crises. 12 Tables A.2 and A.3 in the appendix provide additional information on the sample, organized by home and host countries. 8 is proxied with total assets, while the income-to-loan margin is the net income share of total loans. Additional bank-level variables—such as liquidity, reliance on wholesale funding, weakness as measured by loan loss provisions, and pro�tability—are treated as non-core bank covariates (either because they capture analogous concepts to the core variables, or suffer from weaker data availability). The core set of country-level variables consists of (lagged) real GDP growth, real GDP per capita, consumer price inflation, and the current account balance from the World Bank’s World Development Indicators (WDI); and a dummy for offshore �nancial center, as identi�ed by the Bank of International Settlements. Additional country-level covariates used in our robustness checks include trade openness and �nancial exports from the WDI; and the aggregate capital to assets ratio and ratio of banks’ nonperforming loans to total gross loans from the World Bank’s Financial Development and Structure database (Beck et al., 2000). Additional details on the de�nitions and sources of all variables, along with summary statistics for the main variables of interest, are provided in the appendix. 3.2 Stylized features of banks in the 2007/08 crisis To gain a better understanding of the research design, it is useful to consider several stylized facts a-vis domestic banks, and present in the data. These concern the lending patterns of foreign vis-` between foreign banks with home countries that experienced a crisis relative to those with non-crisis home countries of ownership. We begin by documenting the fact that domestic and foreign banks in developing countries exhibit systematically different lending behavior (Fact 1 ). For 2006, outstanding loans of an average domestic bank in developing countries amounted to $5 billion, while the average foreign bank had lent a third that amount, $1.7 billion (all measured in current U.S. dollars). The variation within each of the two groups is also substantial: the standard error of the mean for domestic banks in 2006 was $1.1 billion, compared to $295 million for foreign banks.13 These differences are statistically signi�cant: the test statistic for differences in the means, assuming unequal variances in the unpaired data, is t = 2.99 (p = 0.001), while the variance ratio test yields F = 27.50 (p = 0.000), for 2006, and statistically signi�cant differences hold for 2009 as well (we discuss theoretical reasons why this discrepancy between the two groups may exist in section 2).14 The second feature of our data is that foreign banks with home countries experiencing a crisis do differ, on average, in their amounts of outstanding loans as compared to non-crisis foreign banks 13 Equivalent �gures for the mean (standard error) for 2009 were $9.6 billion ($2.0 billion) and $2.4 billion ($380 million), respectively. 14 These statistical differences persist even in the log-transformed data that we use for our econometric analysis; in 2006, for instance, the analogous two-group t and F tests are t = 2.07 (p = 0.019) and F = 1.19 (p = 0.03). 9 (Fact 2 ); in 2006, the mean for loans from the former group were $2.3 billion, versus $540 million for the latter.15 Foreign banks exposed to the crisis in their home countries represent 61 percent of the sample of foreign banks—221 out of 361 banks—which provides some assurance that any estimated treatment effects are unlikely to be driven by outliers or small-sample problems. The third element in the data that deserves mention is the fact that lending by both groups of foreign banks essentially followed the same trend up through the eve of the crisis (Figure 1): the 3-year change in (log) average lending between 2004–06 is statistically indistinguishable between crisis and non-crisis banks. Just as important, the trend in lending among just the non -crisis banks remains unchanged following the crisis16 (Fact 3 ). Average of Gross Loans 600 Average of Gross loans (after normalizing 2000==100) 500 400 300 200 100 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 Crisis Years Foreign Banks Owned by No-Crisis Countries Foreign Banks Owned by Crisis Countries Figure 1: Trends in average gross loans, disaggregated by crisis treatment and nontreatment foreign banks, 2000–10. For comparison purposes, average loans for both groups are normalized to 100 for 2000. The crisis period is demarcated as 2007–08. Similar rising trends for both groups are evident until 2007, and the divergence in trends following 2008 is striking. Taken together, the second and third stylized facts argue strongly in favor of our difference-in- difference approach: the methodology allows for initial differences to exist between the two groups of interest (Fact 2 ), while the coincident pre-crisis trends (and stable pre- and post-crisis trend for the control group) captured by Fact 3 point to the fact that our two groups in question would 15 A matrix summarizing the means, dispersions, and differences for crisis-treated and nontreated foreign banks is provided in the appendix. 16 The 3-year change in (log) average lending between 2004–06 and 2008–10 are statistically indistinguishable for non-crisis banks. 10 likely have shared parallel trends in the absence of a crisis (the common support assumption). Finally, as a preview of the results to follow, it is useful to compare the extent to which lending for each group changed between 2006 and 2009. For foreign banks that experienced home country crises, post-crisis average lending increased by 41% ($972 million from $2.4 billion), whereas banks that did not experience crises increased their average post-crisis lending by 56% ($366 million from $651 million). Thus, the recovery in lending for noncrisis foreign banks was far sharper than that of crisis-stricken foreign banks; the difference in this change in lending patterns between the two groups is large and statistically signi�cant, and constitutes our stylized Fact 4.17 4 Econometric Methodology 4.1 A difference-in-difference design The point of departure in our empirical analysis is a straightforward difference-in-difference setup: lijk,t = α + γ0 crisisk + γ1 postt + δ (crisisk · postt ) + ijk,t where the dependent variable lijk,t is total lending of bank i = 1, . . . , I in host country j at time t = {2006, 2009}. Each foreign bank i also has as an attribute its home country of ownership k , crisisk is an indicator variable that takes on the value of 1 when country k experiences a systemic banking crisis in 2008 (the crisis treatment), and postt is an indicator variable that takes the value 1 if t = 2009 (the post-crisis period). is an idiosyncratic error term, which, depending on the speci�cation, may be clustered along host j and/or home k . In an application with only two periods, there is a well-known correspondence between the simple difference-in-difference estimator above and its differenced form, where the equation above is identical to the regression ˜ ∆lijk = β + δcrisisk + εijk , (1) ˜ = δ, where the operator ∆ denotes the change between two periods. The coefficient of interest, δ a-vis untreated banks. In captures the difference in the average change in lending for treated vis-` principle, if the treatment is randomly assigned, estimates of this coefficient will be identi�ed. Of course, this identi�cation of the treatment effect also hinges crucially on our common support assumption (Imbens and Wooldridge, 2009). As shown in subsection 3.2, however, this assumption is ful�lled in our sample. Given the wide variation in foreign bank types operating across different developing countries, ˜ may be reduced, and the �t of the model improved, by introducing additional controls to bias in δ (1). Country-level effects, for both the home and host, may be important in practice. For example, 17 These results are also tabulated formally in the appendix. 11 banks from Spain may adopt a different operational model for subsidiaries as compared to banks based in the United States, and as a consequence Spanish-owned banks may react differently to a crisis than U.S.-owned banks. By a similar token, banks operating in different countries need not react similarly after a crisis in a foreign country, as they face distinct economic environments (for example, different monetary or regulatory policy regimes). Accounting for these effects amounts to including bank- and country-level �xed effects in the basic difference-in-difference setup: lijk,t = α + γ0 crisisk + γ1 postt + δ (crisisk · postt ) + αi + αj + αk + γ2 (αj · postt ) + γ3 (αk · postt ) + ijk,t , where αi captures a bank-speci�c effect, and αj and αk represent country effects for the host and home countries, respectively. Note that we have allowed for time-varying country effects (γ2 and γ3 ), but have constrained the coefficient on bank effects to be constant across the two periods.18 The above speci�cation can be rewritten ˜ crisisk + α + α + ε . ∆lijk = β + δ (2) j k ijk Since the bank �xed effect αi is time-invariant, it drops out of the �rst-differenced speci�cation.19 To the extent that accounting for time-varying bank effects can further improve the efficiency of our estimate of δ (and possibly account for omitted sources of bank-speci�c trends that may lead to violations of the common support assumption), we may wish to explicitly introduce additional bank-speci�c controls into (2). More speci�cally, we can estimate ˜ crisisk + α + α + β Bi + ε , ∆lijk = β + δ (3) j k 1 ijk where Bi is a vector of bank-speci�c characteristics. Populating B with additional (observable) bank controls then allows us to capture potential time-varying idiosyncratic bank effects. Although including additional controls in (2) and (3) does mean that time-varying factors at the bank as well as the country level are accounted for, there are two problems with doing so. There is the possibility that introducing additional covariates may lead to a violation, rather than a strengthening, of our common support assumption.20 In addition, introducing time-varying coefficients for observable vectors of characteristics may also violate the exogeneity assumption ˜ and δ and hence result in biased estimates (Lechner, 2010). Consequently, δ ˜ may capture only a 18 In principle, a fully-saturated speci�cation would allow for an additional interaction term αi · postt . In practice, however, doing so would give rise to degrees-of-freedom issues that would inhibit estimation. 19 In other words, all time-invariant effects are implicitly accounted for in a �rst-differenced speci�cation. 20 This would result, for instance, if banks are relatively homogeneous within each group, so that adding additional covariates leads to a weakened likelihood that the groups maintain parallel trends. 12 relatively crude estimate of the average crisis treatment effect, whether conditioned on unobservable or observable controls. But if we are reasonably con�dent of the identi�cation of the treatment effect, comparing crisis- treated foreign banks against nontreated banks that share very similar observable characteristics—a matching difference-in-difference (matching DiD) speci�cation—can further improve the quality of our estimate of δ (Abadie and Imbens, 2006).21 Let  1 −i∈IM (i) ∆l−ijt if crisisk = 0,  ∆ˆcrisis lijt = M ∆l if crisisk = 1; ijt  ∆l if crisisk = 0, ijt ∆ˆnoncrisis lijt = 1 M −i∈IM (i) ∆l−ijt if crisisk = 1, where IM (i) is the set of M matching indices. These changes in lending outcomes correspond to, respectively, foreign banks exposed to the crisis treatment and those that were not. Then the matching difference-in-differences estimator of Abadie and Imbens (2006) generates our coefficient of interest given by I ˜= 1 δ ∆ˆcrisis lijt − ∆ˆnoncrisis lijt , (4) I i=1 which we implement with the nearest-neighbor (Mahalanobis) metric. Note that the identi�cation of the treatment effect in the matching DiD estimator depends on the assumption of unconfoundness, which requires, conditional on covariates, that there be no unobservables associated with both the treatment and with the potential outcomes (Imbens and Wooldridge, 2009). As we argue below, our treatment—a home-country crisis—is plausibly independent of the lending activities of these countries’ foreign subsidiaries in developing countries, thereby satisfying this assumption. 4.2 Identi�cation of the crisis treatment The estimation described in subsection 4.1 hinges on whether, conditional on our sample, the crisis treatment is well identi�ed. In this subsection, we discuss why we believe that this is the case. First, it is worth noting that only banks that were majority foreign-owned were considered in our setup. As discussed in subsection 3.2, this is because foreign banks with home countries that did not experience a crisis are the most appropriate comparison group for estimating the effect 21 Note that in contrast to propensity score matching difference-in-differences, the matching estimator of Abadie and Imbens (2006) is not effected to determine selection into the crisis treatment; rather, the matching algorithm ensures comparability of treated and untreated banks. We compute the Abadie and Imbens (2006) matching estimator following the implementation described in Abadie et al. (2004), which performs matching with replacement, and with the bias correction suggested in Abadie and Imbens (2011). 13 of the crisis treatment. Moreover, the difference-in-difference approach allows initial differences to exist between the two groups in question. To further establish identi�cation, it is necessary that the crisis treatment satis�es the exclusion restriction. We make this case in three steps. First, we assert that, by and large, developing country- based subsidiary banks of foreign multinationals are dwarfed by the size of their home country banking systems. Consequently, the likelihood that they influenced their respective home-country crises is extremely small. Second, only certain home—and, typically, high-income—countries underwent a �nancial crisis in 2008, and consequently, only a subset of the foreign-owned banking subsidiaries in our sample were exposed to the crisis treatment. It is this exogenous variation in home-country experiences that we exploit to identify the effect of a home-country crisis on foreign bank behavior in our baseline DiD speci�cation. The �nal issue with regard to identi�cation is with regard to relevance: that is, whether our banking and �nancial crisis treatment, as captured by our sample of home crisis countries, is capturing the effect of a crisis per se, or whether other country-level macro factors are responsible for the observed treatment effects. In our robustness checks, we test this condition by attempting to rule out the possibility that some other possible channels may be responsible for the observed treatment effect. One concern that may arise regarding our working sample is the possibility of survivorship bias. Although there is undoubtedly attrition in our sample between 2006 and 2009, we view this issue as mainly a red herring. There is little reason to believe that, after conditioning on �xed effects, there would be any systematic variation in bank failures between the two groups that are not directly attributable to the crisis. Indeed, if anything, the magnitude of our estimate of the crisis effect would be biased downward by such attrition. 5 Empirical Results 5.1 Baseline difference-in-differences Our baseline results are reported in Table 1. In the �rst column (B1) we report the baseline difference-in-difference speci�cation in (1) with no �xed effects. Below the estimated coefficient we report robust standard errors clustered either by home country, host country, or both.22 The 22 Unlike the inclusion of �xed effects, there is generally less consensus on the appropriate treatment of clustered errors, especially since such corrections in the presence of a small number of groups can lead to downward-biased standard errors in DiD settings (Donald and Lang, 2007). In our application, we cluster errors by host and/or home country (rather than treatment), so the number of groups is reasonably large, which mitigates this concern. Nevertheless, since the decision typically involves a tradeoff between robustness and efficiency, we report all three possible combinations of clustering in the baseline results. 14 coefficient on the crisis treatment is statistically signi�cant at the conventional levels, and negative; thus, the results indicate that foreign banks owned by entities in countries that experienced a �nancial crisis tended to reduce their lending more (or raise their lending less) than foreign banks owned by entities in non-crisis-stricken countries. As discussed in the introduction and section 2, while this result strikes us as reasonably intuitive, the alternative outcome (of increased lending) is an a priori theoretical possibility. One objection to this simple benchmark is that pooling all foreign banks, regardless of host country, may fail to adequately capture heterogeneity in changing host country conditions. For instance, foreign banks generally extend loans in domestic currency, and since our dependent vari- able is measured in nominal U.S. dollars, the conversion into a common numeraire currency may introduce distortions into our estimate of the treatment effect. In theory, there is no reason why this should be a major concern. Consider, for example, the case of a host country facing high rates of inflation. Even if exchange rates do not efficiently correct for the inflation differential, foreign banks experiencing the crisis treatment are just as likely as non-treatment foreign banks, ex ante, to locate in countries with high or low inflation, and thus would introduce no bias to the residual. Of course, systematic country-level differences need not be limited to inflation, but could in- clude all manner of idiosyncratic country-level shocks (resource-rich countries experiencing a posi- tive terms-of-trade shock from rising commodity shocks, say). Any host- or home-country-speci�c factors that give rise to initial differences between the lending of the treatment and non-treatment groups of banks are controlled for by the DiD strategy, as long as they do not also give rise to differences in changes in lending behavior over the crisis period. Any time-invariant bank-speci�c effects are also captured by even this simple speci�cation. In sum, as long as the identifying as- sumptions described in subsections 4.1 and 4.2 hold, we are assured that the estimates we obtain are unbiased (Imbens and Wooldridge, 2009). Nevertheless, if we believe that efficiency is enhanced by allowing for time-varying country effects, it is straightforward to introduce these into the baseline setup as in equation (2). One immediate concern is the need to introduce changes to demand-side effects for lending, which can be done by adding host country �xed effects. Accordingly, column (B2) reports results when we allow for these. Doing so substantially improves the �t of the model, and slightly raises the point estimate for the treatment effect. This is not the case when we control for only home �xed effects, where—as shown in column (B3)—the magnitude of the coefficient falls, although the lower R2 alongside the mostly smaller standard errors strongly suggest that this result is due to omitted variable bias. Column (B4) thus reports the fullest articulation of equation (2), where we include time-varying �xed effects for both home and host economies. The estimated crisis effect in this case is even stronger than in (B1), and is signi�cant at the 5 percent level or better, regardless of our choice of error clustering. 15 Table 1: Baseline difference-in-difference regressions for bank lending, 2006 and 2009† B1 B2 B3 B4 Crisis effect -0.316 -0.364 -0.127 -0.420 (0.13)∗∗ (0.12)∗∗∗ (0.00)∗∗∗ (0.16)∗∗∗ (0.14)∗∗ (0.16)∗∗ (0.39) (0.21)∗∗ (0.14)∗∗ (0.16)∗∗ (0.10) (0.17)∗∗ Fixed effects Home No No Yes Yes Host No Yes No Yes Adj. R2 0.021 0.307 0.245 0.490 Clusters (countries) 66, 51 66, 51 66, 51 66, 51 Estimation OLS OLS OLS OLS N (banks) 361 361 361 361 † The dependent variable is in log differenced form. Heteroskedastic- ity and intragroup correlation-robust standard errors are reported in parentheses; the rows correspond to standard errors: (1) clustered by home country; (2) clustered by host country; (3) with two-way cluster- ing. A constant term was included in the regressions, but not reported. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Fixed effects for home and host are time varying. Cluster sizes are reported for home and host, respectively. The magnitude of the coefficient is also economically signi�cant: using the �nal speci�cation (B4), foreign banks exposed to the 2007/08 �nancial crisis in their home countries pared back on their lending in their developing country hosts by an average of 42 percent (relative to foreign banks whose home countries did not experience a crisis). To place this �gure in some perspective, consider the change in lending the case that would obtain if a crisis-stricken foreign bank actually did not experience the crisis. If the 2006–09 change in lending for this bank was that of the average change in lending among crisis-stricken banks—of $935 million—the bank would hypothetically have lent $ [935/ (1 − 0.42)] = $1.6 billion instead in the absence of the home-country crisis. While this result accords well with the literature examining the influence of home crises on foreign bank behavior, it stands as a counterpoint to a larger literature that �nds that foreign banks tend to be stability-enhancing in countries undergoing crises (Clarke et al., 2003).23 Finally, it is worth considering what the time-varying country �xed effects mean for our es- 23 These disparate �ndings are easily reconciled: the larger literature has seldom examined the case where the �nancial crisis currently experienced in the developing country originates from abroad. de Haas and van Lelyveld (2013) conclude that multinational banks contribute to �nancial stability during local crisis episodes, but also increase the risk of importing instability from abroad. 16 timated coefficients. Columns (B2)–(B4) essentially allow home and host country �xed effects to take on different slopes in the post-crisis period, capturing distinct country-speci�c responses to the crisis. The higher coefficient in column (B4)—as compared to (B1)—thus suggests that the effect of the crisis on lending might well have been greater absent crisis mitigation policies such as the expansion of central bank balance sheets (since the smaller estimated effect in column (B1) would be due to not controlling for these heterogeneous policies).24 5.2 Baseline with additional bank covariates In Table 2 we consider the inclusion of a set of covariates at the bank level, along the lines of equation (3). As noted earlier, expanding the set of covariates is not necessary if the treatment is well identi�ed or if idiosyncratic bank effects are time-invariant, and there are reasonable objections to the indiscriminate inclusion of additional controls. The inclusion of covariates in our application is further complicated by the fact that such covariates may be correlated with the crisis treatment such that they violate the exogeneity assumption. Our resolution of this latter problem is to include our covariates as they are observed in the pre-crisis period (Bt = Bt+1 , t = 2006). Set against these potential disadvantages is the fact that including covariates allows us to capture the possibility that bank effects may vary over time, as they well could after a major shock such as a �nancial crisis. In Table 2, we incrementally introduce the four idiosyncratic bank controls included in the core set, along with two additional controls (wholesale funding and liquidity) (these are described in subsection 3.1). This core set is chosen to best capture important (observable) cross-bank heterogeneity that may potentially affect foreign bank lending behavior (further details of these measures are documented in the appendix). The main message from this set of results is that, compared to the bare-bones speci�cations in Table 1, the magnitude and signi�cance of the crisis effect generally holds, even after we allow for the possibility of time-varying bank-speci�c effects. While the coefficient in the �nal speci�cation, (C6), is not signi�cant at conventional levels, this is likely due to the compromised sample size when data availability demands are greater due to the relatively larger number of covariates included. Overall, the results here point to post-crisis lending by crisis-stricken banks that is 26 to 57 percent lower than that of their noncrisis counterparts. The point estimates here are also, on average, a hair larger than those in the baseline, accompa- nied by higher standard errors.25 There are two potential explanations for these larger coefficients. The �rst possibility is that controlling for time-varying effects of these covariates yields more precise 24 Admittedly, this interpretation would only be de�nitive if we are willing to make the ceteris paribus assumption of unchanged demand conditions in each country. The general point about the crisis effect being underestimated in the simple DiD speci�cation without �xed effects will continue to hold, however. 25 In the interests of space, the reported standard errors correspond to two-way clustering; analogous results are obtained when clustered by either home or host countries, and are available on request. 17 Table 2: Difference-in-difference regressions for bank lending, with core and additional bank-level covariates, 2006 and 2009† C1 C2 C3 C4 C5 C6 Crisis effect -0.256 -0.571 -0.548 -0.508 -0.397 -0.296 (0.14)∗ (0.26)∗∗ (0.24)∗∗ (0.27)∗ (0.22)∗ (0.25) Core bank-speci�c characteristics Size -0.110 0.028 0.029 0.017 0.008 0.009 (0.11) (0.09) (0.09) (0.08) (0.09) (0.08) Solvency 0.000 0.000 0.000 0.000 0.000 (0.00)∗ (0.00)∗ (0.00) (0.00) (0.00) Interest margin -0.000 -0.000 -0.000 -0.000 (0.00) (0.00) (0.00) (0.00) Income-to-loan -0.007 0.247 0.256 (0.01) (0.04)∗∗∗ (0.04)∗∗∗ Additional bank-speci�c characteristics Wholesale 0.001 0.002 (0.00) (0.00) Liquidity 0.003 (0.00) Fixed effects Home Yes Yes Yes Yes Yes Yes Host Yes Yes Yes Yes Yes Yes Adj. R2 0.502 0.548 0.550 0.558 0.660 0.668 Clusters (countries) 66, 51 66, 51 66, 51 66, 51 66, 51 66, 51 Estimation OLS OLS OLS OLS OLS OLS N (banks) 361 361 361 361 344 343 † The dependent variable is in log differenced form. Heteroskedasticity and intragroup correlation-robust standard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). Fixed effects for home and host are time varying. A constant term was included in the regressions, but not reported. Cluster sizes are reported for home and host, respectively. estimates of the average treatment effect, since including these additional observable controls may improve efficiency. The second possibility is that the result may be biased due to the introduction of time-varying coefficients on covariates into the DiD design Lechner (2010). This may be a problem for all the estimates in Table 2, although the very comparable coefficients (relative to Table 1) would suggest that any such bias is small. Nevertheless, a more powerful way to control for covariates is to follow a matching DiD strategy, which is the exercise we undertake in the following subsection. 18 5.3 Matching difference-in-differences Table 3 reports results for the crisis treatment effect of equation (4), estimated by DiD estimates matched on the set of home and host country-speci�c covariates (columns (M1)–(M3)),26 and with additional bank-speci�c covariates added (columns (M4)–(M6)). Since there is no agreement on an optimal number of matches that should be chosen (Imbens and Wooldridge, 2009), we present results for one, two, and four matches, for each of the two cases.27 Table 3: Matching difference-in-difference regressions for bank lending, with bank- and country-level controls, 2006 and 2009† M1 M2 M3 M4 M5 M6 Crisis effect -0.497 -0.367 -0.496 -0.071 -0.277 -0.381 (0.13)∗∗∗ (0.13)∗∗∗ (0.11)∗∗∗ (0.12) (0.11)∗∗∗ (0.11)∗∗∗ Core host covariates Yes Yes Yes Yes Yes Yes Core home covariates Yes Yes Yes Yes Yes Yes Core bank covariates Yes Yes Yes Yes Yes Yes Additional bank controls No No No Yes Yes Yes Estimation Matching Matching Matching Matching Matching Matching Matches 1 2 4 1 2 4 N (banks) 340 340 340 322 322 322 † The dependent variable is in log differenced form. Point estimates computed from matching with replacement based on the Mahalanobis metric and are Abadie and Imbens (2011) bias-corrected. Heteroskedasticity-robust standard errors reported in parentheses. ∗ indicates signi�cance at 10 per- cent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Covariates used for matching are the core country and bank controls listed in the appendix. Additional bank covariates are wholesale and liquidity. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). The qualitative �ndings remain largely unchanged. The matching DiD estimates are in the same ballpark as the simple DiD regressions and those obtained when bank covariates are included, and the (statistically signi�cant) crisis effect coefficients range from -0.28 to -0.50. In the fullest ar- ticulation of the baseline model with four matches—shown in column (M6)—foreign banks exposed to a �nancial crisis in their home countries have changes in lending that are 38 percent smaller, on average, than an otherwise comparable foreign bank whose home country did not experience a 26 In lieu of host and home �xed effects we use a set of core country variables (GDP per capita, GDP growth, inflation, and current account balance) as matching variables. As was the case for simple difference-in-differences, there is a case against over�tting of covariates. 27 The choice of one match is entirely reasonable—we wish to compare only banks existing in the data, rather than synthetic comparators—and the choice of four matches has been shown to perform well in terms of mean- squared error (Abadie and Imbens, 2011). We include two matches as an intermediate case, noting that the difference between coefficient estimates for three and four matches is small. We also considered higher numbers of matches. In general, these decreased the magnitude of the estimated coefficient, but even for the (extreme) case of 20 matches, the coefficient remained statistically signi�cant. These additional results are available on request. 19 crisis. Although the magnitudes of the point estimates are comparable to those reported in Tables 1 and 2, it is worth noting that matching DiD may in fact provide a superior estimate. In the simplest DiD implementation with no additional controls, all crisis and noncrisis banks are pooled together—even if it were the case that they differed along a signi�cant number of dimensions—and identi�cation of the average crisis treatment effect relies on a more-or-less random distribution of additional characteristics across the sample. Such pooling may fail to accurately gauge the true extent of the crisis effect, to the extent that bank characteristics are correlated both with the treatment and with the error term. In contrast, the matching estimator forces the comparison to occur either with an otherwise similar (at least along observable dimensions) bank, or against a synthetic equivalent. To the extent that such matching on observables is indeed appropriate (and does not introduce any selection bias), the estimate renders a better apples-to-apples comparison. 6 Robustness Checks 6.1 Additional controls and alternative measures In this subsection we consider a range of robustness checks that offer variations on our choice of controls in the baseline.28 These are reported in the six columns on the left panel of Table 4. We �rst supplement the core set of bank covariates with two alternative bank-level controls, bank weakness and pro�tability.29 These are given in columns (R1) and (R2) which build on, respectively, the DiD speci�cation with the core bank covariates—extending speci�cation (C4)— and the matching DiD equivalent, which extends speci�cation (M3). Next, we allow for the possibility of time-varying effects that operate at the country-pair level (as opposed to independently at the country level). More speci�cally, we replace the home and host country �xed effects αj and αk in (2) with a �xed effect αjk for each unique home-host dyad. This approach will absorb greater unobservable heterogeneity insofar as pairwise effects, such as those arising from economic closeness at the bilateral level (de Haas and van Horen, 2012b), are relevant to lending behavior. Since this approach to capturing �xed effects is fundamentally distinct from even our simple augmented model (2), we show the DiD results obtained both without and with core bank covariates in columns (R3) and (R4) (analogous to speci�cations (B4) and (C4), respectively). Third, we exploit the additional flexibility for including covariates offered by the matching DiD 28 Alternative permutations, such as using matching with core and additional bank controls, did not qualitatively alter our �ndings, and are available on request. 29 The reason why we choose not to include all four additional covariates together is twofold: some of these variables capture very similar concepts, and so including them simultaneously may introduce multicollinearity; moreover, doing so would seriously erode the size of our sample (since the coverage of these additional controls do not overlap perfectly). 20 Table 4: Robustness of DiD and matching DiD regressions for bank lending, with alternative and additional bank- and country- level controls, 2006 and 2009† R1 R2 R3 R4 R5 R6 R7 R8 R9 R10 t = 2006, t+1 = 2009 t = 2005–06, t+1 = 2009–10 Crisis effect -0.320 -0.413 -0.521 -0.375 -0.684 -0.478 -0.530 -1.014 -0.437 -0.990 (0.19)∗ (0.10)∗∗∗ (0.00)∗∗∗ (0.16)∗∗ (0.14)∗∗∗ (0.15)∗∗∗ (0.19)∗∗∗ (0.14)∗∗∗ (0.18)∗∗ (0.14)∗∗∗ Fixed effects/core covariates Home Yes Yes No No Yes Yes Yes Yes Yes Yes Host Yes Yes No No Yes Yes Yes Yes Yes Yes Pair No No Yes Yes No No No No No Yes Bank Yes Yes No Yes Yes Yes Yes Yes Yes Yes Noncore covariates? Additional country-speci�c No No No No Yes Yes Yes No No No 21 Additional bank-speci�c No No No No No No No No Yes Yes Alternative bank-speci�c Yes Yes No No No Yes No No No No Adj. R2 0.670 0.800 0.880 0.527 0.561 Clusters (countries) 66, 51 - 66, 51 66, 51 - - 66, 51 - 66, 51 - Estimation OLS Matching OLS OLS Matching Matching OLS Matching OLS Matching Matches - 4 - - 4 4 - 4 - 4 N (banks) 340 321 361 361 247 234 361 347 361 347 † The dependent variable is in log differenced form. Matching estimates computed with Mahalanobis metric and are Abadie and Imbens (2011) bias- corrected. Heteroskedasticity (all speci�cations) and intragroup correlation (OLS only)-robust standard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Fixed effects for home, host, and pair are time varying. Core bank and country covariates are listed in the appendix. Additional bank covariates are wholesale and liquidity, alternative bank covariates are weakness and pro�tability, additional country covariates are related to the banking system (bank capital, bank nonperforming loans) and economic openness (trade openness, �nancial services exports), and are listed in the appendix. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). Cluster sizes are reported for home and host, respectively. estimator by adding additional country- and bank-level covariates (although with the same caveat as before that doing so may lead to signi�cant sample size reductions, which justi�es our decision not to use them in the baseline estimates). These additional covariates relate to country-level characteristics (regarding the quality of the �nancial system and the openness of the economy) and the alternative bank-level controls (weakness, pro�tability) introduced in the �rst two columns. The results for each are reported, respectively, in columns (R5) and (R6). In the right panel we examine the robustness of our results to an alternative measure of our pre- and post-crisis periods: rather than utilizing data from two individual years (2006 and 2009), we average observations from 2005 and 2006 for the pre-crisis period, and 2009 and 2010 for the post- crisis period.30 Here, for reasons of space, we report only the DiD and matching DiD estimates when controlling for only country-level �xed effects/core covariates (columns (R7) and (R8), respectively), and with core and additional bank covariates (columns (R9) and (R10)). As evident from Table 4, our baseline results by and large survive this array of robustness checks. There is little variation in the magnitude of the estimated crisis treatment effect, except in the two matching DiD estimates when two-year averages are considered (namely, speci�cations (R8) and (R10)). The coefficient estimates are bound by [−0.32, −1.01], and all retain their sta- tistical signi�cance at the 10 percent level or lower. We suspect that the somewhat higher point estimates obtained in (R8) and (R10) may be because the matching algorithm is inflated by the rel- atively higher 2010 lending data among noncrisis countries, resulting in arti�cially strong synthetic comparators, and thereby possibly overestimating the crisis treatment effect. We make one �nal, brief remark regarding the estimates in Table 4: the stability of the coef- �cients across this broad array of speci�cations lends a fair amount of con�dence that the crisis treatment effect is not only real, but reliably estimated. This lends con�dence that even the most parsimonious DiD speci�cation, given by equation (1), is probably sufficient for our analysis. With this in mind, we turn away from estimating the crisis treatment effect, and toward possible falsi�- cation tests in order to build our case that these estimates are indeed valid. 6.2 Falsi�cation tests for alternative channels In this subsection we introduce a set of distinct placebo tests designed to rule out the possibility that the estimated effect of the crisis treatment may either be due to noncrisis-related trends in the two groups, or to other, distinct noncrisis shocks that occurred between 2006 and 2009 which were correlated with the crisis treatment. Our �rst test alters the pre- and post-crisis dates to an earlier period; we choose 2002 and 2005 30 We perform the period averaging to avoid serial correlation problems that may arise from difference-in-difference treatments that span multiple time periods (Bertrand et al., 2004). Data limitations prevent us from using longer averages. 22 Table 5: Falsi�cation tests for difference-in-difference regressions for bank lending, 2003, 2006, and 2009† F1 F2 F3 F4 F5 F6 t=2002, t+1=2005 treatment=trade treatment=�scal Treatment effect 0.077 -0.389 0.889 0.435 0.517 0.675 (0.32) (0.29) (0.33)∗∗∗ (0.74) (0.23)∗∗ (0.27)∗∗ Fixed effects/core covariates Home Yes Yes Yes Yes Yes Yes Host Yes Yes Yes Yes Yes Yes Bank No Yes No Yes No Yes Adj. R2 0.442 0.516 0.490 0.558 0.490 0.558 Clusters (countries) 49, 42 49, 42 66, 51 66, 51 66, 51 66, 51 Estimation OLS OLS OLS OLS OLS OLS N (banks) 265 264 361 361 316 316 † The dependent variable is in log differenced form. Heteroskedasticity and intragroup correlation-robust stan- dard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). Fixed effects for home and host are time varying. A constant term was included in the regressions, but not reported. Cluster sizes are reported for home and host, respectively. as alternative years.31 This falsi�cation test is designed to rule out the possibility that trends in lending behavior in the two groups may already have been diverging prior to 2006. Consequently, if coefficient estimates for the crisis variable are in signi�cant, we can more con�dently assert that our crisis effect is capturing a genuine shock experienced between 2006 and 2009. The �rst panel of Table 5 reports the results of this �rst set of placebo tests, for the augmented DiD speci�cations (2) and (3) (columns (F1) and (F2), corresponding to speci�cations (B4) and (C4), respectively). The insigni�cant estimated coefficients indicate that the baseline estimations of the treatment effect in the DiD model are indeed capturing an effect unique to the period between 2006 and 2009.32 Our second falsi�cation exercise considers the other major non-�nancial crisis-related event that occurred in the intervening period: the great trade collapse of 2008/09 (Baldwin, 2009).33 31 We pick these two years to maximize data coverage, as data availability for most bank-level controls is quite limited for years prior to 2002. 32 One concern that may arise is that the sample sizes in Table 5 are substantially smaller than those in our baseline estimates, and so what is being captured is due to changes in the sample, rather than a genuine insigni�cant effect. To allay this concern, we replicated the three speci�cations for two other subsamples: �rst, by repeating the exercise for this smaller subsample for 2006–09, and second, by re-estimating the 2002–05 placebo for the subsample resulting from the �rst step (due to incomplete data coverage, the �rst step shrinks the subsample even further). Our crisis effect actually remains signi�cant in the reduced subsample from the �rst step, and the placebo remains insigni�cant in the second step. Taken together, these additional tests indicate that the results are not due to sample variations. 33 It is reasonable to argue that the trade collapse occurred in 2008 as a direct consequence of the �nancial crisis, and 23 Of course, �nancial crises and other economic crises are likely to be correlated, so home countries we identify as having experienced a banking crisis may have also underwent trade-related changes around the same time, which could in turn have affected their banks’ subsidiaries’ lending abroad. For example, if Spain’s imports from Mexico collapse, and Spanish-owned banks in Mexico tend to lend more to �rms that export to Spain (perhaps because such exporters are also Spanish-owned) a-vis other banks in Mexico, then these banks will face a greater decrease in loan demand during vis-` the crisis than other banks. Thus, their lending will fall more than other banks’ lending in Mexico. If this is systematic across country pairs, it will show up as an effect of the home-country banking crisis, when the effect is actually that of a collapse in home-country import demand (and associated �nancing needs abroad). To rule out this channel, the falsi�cation test requires the construction of a new treatment variable that captures the effect of a trade collapse, replaces our crisis variable with this dummy, and relies on differences between the two sets of treatments to identify the trade contraction effect. The second panel of Table 5 shows the results of this second set of falsi�cation exercises. In columns (F3) and (F4), we again estimate the augmented DiD speci�cations both without and with core bank covariates, but replace the crisis treatment with a trade collapse treatment that takes on the value of unity when the contraction in the home country’s total trade falls below the median of all declines in trade (that is, the 50th percentile of all decreases in home-country trade flows; full details for the construction of this treatment are provided in the appendix). The coefficients in this case are positive, and in one of the two cases, statistically signi�cant. Not only does this indicate that our crisis treatment effects are not driven by trade contractions; if anything, foreign banks from economies that experienced trade collapses lent relatively more than those that did not have trade collapses in their home countries. Thus, to the extent that trade collapses affected lending by foreign banks from crisis economies at all, this effect operated in the opposite direction as the crisis treatment, and diminished its estimated effect. The �nal falsi�cation exercise that we implement turns to the possibility that �scal stimuli, introduced as a result of the �nancial crisis, may have supported growth in economies receiving such a positive real shock, which subsequently allowed bank parents in these countries to systematically expand the lending activities of their developing country subsidiaries (possibly crowding out lending by banks with home economies that did not undergo stimulus). In this case, we can no longer claim that the treatment is capturing a �nancial crisis effect, but rather that the treatment was due to a beggar-thy-neighbor spillover from �scal policy expansion. We thus code economies as having a �scal stimulus treatment as those who exceeded the median so cannot be treated as an entirely separate event. However, the main mechanisms involved are distinct: a real side shock affects other economies via exports, while a �nancial sector one via capital flows. Moreover, there is imperfect overlap between economies suffering trade contractions as opposed to �nancial crises. Both of these reasons suggest that a separate treatment of the issue is warranted. 24 expenditure among all countries that adopted �scal stimuli (variable construction details are in the appendix). The two columns in the third panel, (F5) and (F6), perform this falsi�cation test when we replace our crisis treatment with this �scal stimulus one. As in the case of the trade collapse, we �nd no evidence that the �scal stimulus explanation is driving our results obtained earlier, and the positive and signi�cant coefficients point to a bias against our estimated crisis effect operating via �scal policy expansions. 7 Heterogeneity in the Crisis Effect 7.1 Comparing foreign banks to domestic banks In subsection 3.2, we made the case for why non-crisis foreign banks are the most appropriate comparison group for estimating the effect of the crisis treatment. Introducing domestic banks into the working sample would be inappropriate from an econometric perspective when using the DiD estimator, since it pools all control banks in the comparison group, despite that domestic and foreign banks likely differ along important dimensions related to changes in lending outcomes during the crisis period (for example, facing different loan demand schedules). However, much of the empirical literature (as cited in the introduction) is interested in distinc- tions between domestic and foreign bank behavior, and it is useful to expand our consideration of banks to include domestic banks. While the use of the DiD estimator remains circumspect for the purposes of obtaining a causal estimate of the crisis effect with domestic banks included in the sample, the matching estimator offers an ideal solution to this problem: since matching uses the closest possible match(es) for each treated bank, we are assured that the control is, in fact, an appropriate counterfactual. A further advantage of this empirical strategy is that it allows us to expand to a larger pool of controls from which matches are chosen. Table 6 presents the results of the matching estimator using this expanded sample, analogous to Table 3, which now allows both domestic banks and non-crisis foreign banks to be used to construct controls. In these regressions, we match treated banks only with nontreated banks that operate in the same host economy; accordingly, the only covariates used for matching—other than constraining matches to the same host country—are bank-level covariates.34 The results are in line with our baseline matching DiD (and simple DiD) estimates, though the magnitudes are slightly lower. The smaller coefficients obtained when allowing for matching with domestic banks suggests that the post-crisis recovery in lending for foreign non-crisis banks, as 34 The algorithm performs exact matching, or as exact as possible, on the country of operation. Because for each crisis-stricken bank there may not exist M members among the noncrisis group, fewer than M matches may be used. To provide a sense of how many exact matches exist (as a rough gauge of the quality of the matches), we report the percentage of exact matches possible in the sample, which ranges from 96% when using one matching noncrisis bank to 82% when matching on four banks. 25 Table 6: Matching difference-in-difference regressions using exact matching for host country with expanded sample, 2006 and 2009† D1 D2 D3 D4 D5 D6 Crisis effect -0.229 -0.364 -0.334 -0.161 -0.160 -0.210 (0.08)∗∗∗ (0.07)∗∗∗ (0.07)∗∗∗ (0.08)∗∗ (0.07)∗∗ (0.07)∗∗∗ Core bank covariates Yes Yes Yes Yes Yes Yes Noncore covariates? Additional bank-speci�c No No No Yes Yes Yes Exact host matching Yes Yes Yes Yes Yes Yes Exact matches (%) 95.7 94.5 87.0 92.8 91.8 81.8 Estimation Matching Matching Matching Matching Matching Matching Matches 1 2 4 1 2 4 N (banks) 1,099 1,099 1,099 1,021 1,021 1,021 † The dependent variable is in log differenced form. Point estimates computed from matching with replacement based on the Mahalanobis metric and are Abadie and Imbens (2011) bias-corrected. Heteroskedasticity- robust standard errors reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Covariates used for matching are the core country and bank controls listed in the appendix. Additional bank covariates are wholesale and liquidity. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). compared with domestic banks, may have been larger. This also suggests that total lending in the host countries may well have been lower in the absence of these non-crisis stricken foreign banks, whose post-crisis lending likely exceeded that of domestic banks. This is important, as most studies so far compare domestic and foreign banks, with no distinction between crisis and noncrisis foreign banks (Claessens and van Horen, 2013; de Haas and van Lelyveld, 2013). Some of these studies are a-vis domestic banks; then led to conclude that foreign banks may have reduced their lending vis-` our �ndings, in contrast, point to important differences among foreign banks. To explore this issue further, we compare lending by foreign banks headquartered in noncrisis countries with that of domestic banks. We assign treatment to the �rst group, and run similar DiD and matching DiD regressions as before.35 The results in Table 7 verify that foreign banks from noncrisis countries increased their lending relatively more than domestic banks: the coefficients on the foreign non-crisis term are uniformly positive, and in a number of speci�cations, statistically signi�cant. This result also helps us reconcile our results with the claim in the literature that, on aggregate, foreign banks may increase their lending following a crisis (Clarke et al., 2003; de Haas ınez-Peria et al., 2005; Wu et al., 2011). and van Lelyveld, 2010; Mart´ 35 a-vis domestic banks apply The same caveats we remarked on before about the comparability of foreign banks vis-` to this exercise. We nevertheless present estimates with the simple DiD estimator for completeness, and to allow for comparisons with existing studies that run �xed effects regressions (which are similar in spirit to the our basic DiD model). 26 Table 7: DiD and matching DiD regressions comparing noncrisis foreign banks (treat = noncrisis) with domestic banks using exact matching for host country, 2006 and 2009† N1 N2 N3 N4 N5 N6 N7 N8 Crisis effect 0.251 0.166 0.140 0.184 2.158 1.299 3.383 2.997 (0.12)∗∗ (0.22) (0.20) (0.27) (0.09)∗∗∗ (0.08)∗∗∗ (0.14)∗∗∗ (0.08)∗∗∗ Fixed effects/core covariates Home No Yes Yes Yes Yes Yes Yes Yes Host No Yes Yes Yes Yes Yes Yes Yes Bank No No Yes Yes Yes Yes Yes Yes Noncore covariates? Additional bank-speci�c No No No Yes No No No Yes 27 Exact host matching - - - - Yes Yes Yes Yes Exact matches (%) - - - - 95.4 87.6 73.7 69.5 Adj. R2 0.012 0.395 0.405 0.452 - - - - Clusters (countries) 74, 51 74, 51 74, 51 74, 51 - - - - Estimation OLS OLS OLS OLS Matching Matching Matching Matching Matches - - - - 1 2 4 4 N (banks) 891 891 891 827 891 891 891 827 † The dependent variable is in log differenced form. Matching estimates computed with Mahalanobis metric and are Abadie and Imbens (2011) bias-corrected. Heteroskedasticity (all speci�cations) and intragroup correlation (OLS only)-robust standard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Cluster sizes are reported for home and host, respectively. Fixed effects for home, host, and pair are time varying. Core bank and country covariates are listed in the appendix. Additional bank covariates are wholesale and liquidity. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). 7.2 Comparing distinct features among foreign banks The heterogeneity among foreign banks that arose in the previous subsection hints at the potential value of investigating further how differences among foreign owned banks are related to changes in lending when a bank’s home country experiences a crisis. We disaggregate foreign banks along several dimensions, and in this subsection we report results for two dimensions considered: geo- graphical region, and ownership structure. Our empirical strategy in this regard is straightforward; we add an interaction term to our difference-in-difference setup: ˜1 (crisisk · statei ) + α + α + ε , ˜0 crisisk + σstatei + δ ∆lijk = β + δ (5) j k ijk where state distinguishes various dimensions along which a given bank i can differ from another. ˜1 is now our coefficient of interest. To avoid over�tting the model and assist in our iterpreta- δ ˜1 , our speci�cation builds on the relatively parsimonous augmented difference-in-difference tion of δ speci�cation (2) with time-varying country �xed effects. These results are reported in Table 8. The left panel reports the triple interaction on six developing-country geographical regions (as de�ned by the World Bank), while the right panel reports ownership in terms of whether the banks were publicly listed, and whether they were government-owned. ˜1 estimate applies The �rst observation we make about these results is that the only signi�cant δ to Eastern Europe, and this point estimate is extremely large. This is consistent with the �nding in the literature that the region was especially hard-hit by the crisis (Claessens et al., 2010), and suggests that the crisis, as experienced in Eastern Europe, was such that foreign banks there which faced home-country crises tended to contract their lending more than those in the rest of the developing world on average. Note that the insigni�cant coefficient on the uninteracted crisis term for Eastern Europe and Central Asia is no real cause for concern. The total effect of the crisis has to be inferred from the sum of both the uninteracted and the interaction term, and if we treat statistically insigni�cant coefficients as equal to zero, the total effect for all cases remains signi�cantly negative.36 Second, there also appears to be no signi�cant influence of ownership structure in terms of crisis treatment: both publicly listed and government-owned banks in crisis treatment economies had loan outcomes indistinguishable from privately-held banks. Thus ownership structure does not 36 Another possible interpretation of the insigni�cant independent crisis effect in the ECA speci�cation is that only ECA banks are responsible for our results; that is, the crisis effect would not be signi�cant if ECA banks, which were especially hard hit by the crisis, were not included in our sample. To rule out this possibility (as well as the possibility that selected regional subsamples may be giving rise to our overall crisis effect), we ran regressions using our baseline speci�cation that systematically excluded one region at a time from the sample. The results, which are reported in the appendix, generally hold up to this selective exclusion, indicating that the crisis effect is not due to the lending behavior of foreign banks from any one region. 28 Table 8: Difference-in-difference regressions for bank lending with additional interactions, 2006 and 2009† S1 S2 S3 S4 S5 S6 S7 S8 Regions Ownership Crisis effect -0.425 0.010 -0.447 -0.621 -0.467 -0.515 -0.544 -0.355 (0.08)∗∗∗ (0.37) (0.07)∗∗∗ (0.08)∗∗ (0.07)∗∗ (0.07)∗∗∗ (0.28)∗ (0.21)∗ Crisis × EAP -0.304 (0.50) Crisis × ECA -1.560 (0.61)∗∗ Crisis × LAC 0.551 (0.63) Crisis × MNA 0.549 (0.43) Crisis × SAS 0.301 (0.42) Crisis × SSA 0.562 29 (0.43) Crisis × Pub. List. 0.273 (0.38) Crisis × Govt. -0.552 (0.55) Fixed effects Home Yes Yes Yes Yes Yes Yes Yes Yes Host Yes Yes Yes Yes Yes Yes Yes Yes Adj. R2 0.491 0.505 0.493 0.492 0.491 0.492 0.491 0.4948 Clusters (countries) 66, 51 66, 51 66, 51 66, 51 66, 51 66, 51 66, 51 66, 51 Estimation OLS OLS OLS OLS OLS OLS OLS OLS N (banks) 361 361 361 361 361 361 361 1,021 † The dependent variable is in log differenced form. Heteroskedasticity and intragroup correlation-robust standard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Regions correspond to World Bank regions (EAP = East Asia and Paci�c; ECA = Eastern Europe and Central Asia; LAC = Latin America and Caribbean; MNA = Middle East and North Africa; SAS = South Asia; SSA = Sub-Saharan Africa). Ownership is either publicly-listed (pub. list.) or government-owned (govt.), the other group being privately-held banks. Fixed effects for home and host are time varying. A constant term was included in the regressions, but not reported. Cluster sizes are reported for home and host, respectively. appear to be a source of .37 8 Conclusion In this paper, we examined the question of whether foreign banks whose home countries were hit by the 2007/08 �nancial crisis altered their lending behavior as a result of the shock. We �nd strong and consistent evidence that they do indeed scale back on their lending: in our baseline, by between 13 and 42 percent relative to foreign banks that did not experience such a crisis in their home countries. This result holds up to a battery of robustness checks, which include a range of controls for covariates, and falsi�cation tests for alternative hypotheses. Consequently, we are con�dent that this effect is causal. To infer from our results that developing country policy makers should close their �nancial markets to foreign banks would be to carry our results too far; after all, foreign banks probably carry a host of additional bene�ts in terms of �nancial stability and enhanced competition (Clarke et al., 2003). Our results apply to a very speci�c situation—when foreign countries are experiencing crises of their own—and it is questionable that the unique circumstances of the 2007/08 crisis would necessarily outweigh the other bene�ts that generally accrue from maintaining a liberalized �nancial sector. Our results do, however, suggest that domestic monetary authorities should be aware of the potential for greater credit contraction by foreign banks, and support domestic liquidity formation during crises accordingly. 37 We would, however, caution against excessive inference in this regard. As is the case for certain regions (no- tably EAP and SAS), the number of banks that are publicly-listed or government-owned is fairly small—70 and 24, respectively—which would serve to limit statistical power. 30 References Abadie, A., Drukker, D. M., Herr, J. L. and Imbens, G. W. (2004). Implementing matching estimators for average treatment effects in stata. Stata Journal, 4 (3), 290–311. — and Imbens, G. W. (2006). Large sample properties of matching estimators for average treatment effects. Econo- metrica, 74 (1), 235–267. — and — (2011). Bias-corrected matching estimators for average treatment effects. Journal of Business & Economic Statistics, 29 (1), 1–11. Aiyar, S. (2012). From �nancial crisis to great recession: The role of globalized banks. American Economic Review, 102 (3), 225–230. Baldwin, R. E. (ed.) (2009). The Great Trade Collapse: Causes, Consequences and Prospects. London, England: Center for Economic Policy Research. Beck, T., Demirg¨ ¸ -Kunt, A. and Levine, R. (2000). A new database on the structure and development of the uc �nancial sector. World Bank Economic Review, 14 (3), 597–605. Bertay, A. C., Demirguc-Kunt, A. and Huizinga, H. (2012). Bank Ownership and Credit Over the Business Cycle: Is Lending by State Banks Less Procyclical? Policy Research Working Paper Series 6110, The World Bank, Washington, DC. Bertrand, M., Duflo, E. and Mullainathan, S. (2004). How much should we trust differences-in-differences estimates? Quarterly Journal of Economics, 119 (1), 249–275. Cetorelli, N. and Goldberg, L. S. (2011). Global banks and international shock transmission: Evidence from the crisis. IMF Economic Review, 59 (1), 41–76. — and — (2012a). Banking globalization and monetary transmission. Journal of Finance, 67 (5), 1811–1843. — and — (2012b). Follow the money: Quantifying domestic effects of foreign bank shocks in the great recession. American Economic Review, 102 (3), 213–218. — and — (2012c). Liquidity management of u.s. global banks: Internal capital markets in the great recession. Journal of International Economics, 88 (2), 299–311. Chava, S. and Purnanandam, A. K. (2011). The effect of banking crisis on bank-dependent borrowers. Journal of Financial Economics, 99 (1), 116–135. Claessens, S., Demirg¨ ¸ -Kunt, A. and Huizinga, H. P. (2001). How does foreign entry affect domestic banking uc markets? Journal of Banking and Finance, 25 (5), 891–911. —, van Horen, N., Gurcanlar, T. and Mercado, J. (2008). Foreign bank presence in developing countries 1995–2006: Data and trends, mimeograph. Claessens, S. A., Dell’Ariccia, G., Igan, D. and Laeven, L. A. (2010). Cross-country experiences and policy implications from the global �nancial crisis. Economic Policy, 25 (62), 267–293. — and van Horen, N. (2013). Foreign banks: Trends and impact. Journal of Money, Credit and Banking, (forth- coming). Clarke, G. R., Cull, R. J. and Mart´ ınez Per´ ıa, M. S. (2006). Foreign bank participation and access to credit across �rms in developing countries. Journal of Comparative Economics, 34 (4), 774–795. —, —, Mart´ ınez Per´ıa, M. S. and Sa ´ nchez, S. M. (2003). Foreign bank entry: Experience, implications for developing economies, and agenda for further research. World Bank Research Observer, 18 (1), 25–59. 31 —, —, — and Sa´ nchez, S. M. (2005). Bank lending to small businesses in latin america: Does bank origin matter? Journal of Money, Credit and Banking, 37 (1), 83–118. Dages, B. G., Goldberg, L. S. and Kinney, D. (2000). Foreign and domestic bank participation in emerging markets: Lessons from mexico and argentina. Economic Policy Review, 6 (3), 17–36. de Haas, R. and van Horen, N. (2012a). International shock transmission after the lehman brothers collapse: Evidence from syndicated lending. American Economic Review, 102 (3), 231–237. — and — (2012b). Running for the exit? international bank lending during a �nancial crisis. Review of Financial Studies, 26 (1), 244–285. — and van Lelyveld, I. (2006). Foreign banks and credit stability in central and eastern europe. a panel data analysis. Journal of Banking and Finance, 30 (7), 1927–1952. — and — (2010). Internal capital markets and lending by multinational bank subsidiaries. Journal of Financial Intermediation, 19 (1), 1–25. — and — (2013). Multinational banks and the global �nancial crisis: Weathering the perfect storm? Journal of Money, Credit and Banking, (forthcoming). Dell’Ariccia, G. and Marquez, R. (2004). Information and bank credit allocation. Journal of Financial Economics, 72 (1), 185–214. Detragiache, E. and Gupta, P. (2006). Foreign banks in emerging market crises: Evidence from malaysia. Journal of Financial Stability, 2 (3), 217–242. —, Tressel, T. and Gupta, P. (2008). Foreign banks in poor countries: Theory and evidence. Journal of Finance, 63 (5), 2123–2160. Donald, S. G. and Lang, K. (2007). Inference with difference-in-differences and other panel data. Review of Eco- nomics and Statistics, 89 (2), 221–233. ´ rez, L. (2010). Financial Integration and Foreign Banks in Latin Galindo, A. J., Izquierdo, A. and Rojas-Sua America. IDB Working Paper IDB-WP-116, Inter-American Development Bank, Washington, DC. —, Micco, A. and Powell, A. P. (2005). Loyal Lenders or Fickle Financiers: Foreign Banks in Latin America. on 529, Inter-American Development Bank, Washington, DC. Red de Centros de Investigaci´ Giannetti, M. and Laeven, L. A. (2012a). The flight home effect: Evidence from the syndicated loan market during �nancial crises. Journal of Financial Economics, 104 (1), 23–44. — and — (2012b). Flight home, flight abroad, and international credit cycles. American Economic Review, 102 (3), 219–224. Goldberg, L. S. (2009). Understanding banking sector globalization. IMF Staff Papers, 56 (1), 171–197. Gormley, T. A. (2010). The impact of foreign bank entry in emerging markets: Evidence from india. Journal of Financial Intermediation, 19 (1), 26–51. Grail Research (2009). Global Financial Crisis Bailout/Stimulus Tracker. Tech. rep., Grail Research, Cambridge, MA. Imbens, G. W. and Wooldridge, J. M. (2009). Recent developments in the econometrics of program evaluation. Journal of Economic Literature, 47 (1), 5–86. Khwaja, A. I. and Mian, A. R. (2008). Tracing the impact of bank liquidity shocks: Evidence from an emerging market. American Economic Review, 98 (4), 1413–1442. 32 Laeven, L. A. and Valencia, F. (2012). Systemic Banking Crises Database: An Update. Working Paper WP/12/163, International Monetary Fund, Washington, DC. Lechner, M. (2010). The estimation of causal effects by difference-in-difference methods. Foundations and Trends in Econometrics, 4 (3), 165–224. Mart´ınez-Peria, M., Powell, A. and Vladovka-Hollar, I. (2005). Banking on foreigners: The behavior of international bank claims on latin america, 1985–2000. IMF Staff Papers, 52 (3), 430–461. Micco, A., Panizza, U. and Yan ˜ ez, M. (2007). Bank ownership and performance: Does politics matter? Journal of Banking and Finance, 31 (1), 219–241. Morgan, D., Rime, B. and Strahan, P. E. (2004). Bank integration and state business cycles. Quarterly Journal of Economics, 119 (4), 1555–1584. Ongena, S., Peydro ´ Alcalde, J.-L. and Van Horen, N. (2012). Shocks abroad, pain at home? bank-�rm level evidence on �nancial contagion during the 2007–2009 crisis, tilburg University. Peek, J. and Rosengren, E. S. (1997). The international transmission of �nancial shocks: The case of japan. American Economic Review, 87 (4), 495–505. — and — (2000). Collateral damage: Effects of the japanese bank crisis on real activity in the united states. American Economic Review, 90 (1), 30–44. Popov, A. and Udell, G. F. (2012). Cross-border banking, credit access, and the �nancial crisis. Journal of International Economics, 87 (1), 147–161. Rose, A. K. and Wieladek, T. W. (2011). Financial Protectionism: the First Tests. NBER Working Paper 17073, National Bureau of Economic Research, Cambridge, MA. Schnabl, P. (2012). The international transmission of bank liquidity shocks: Evidence from an emerging market. Journal of Finance, 67 (3), 897–932. Van Rijckeghem, C. and Weder di Mauro, B. (2003). Spillovers through banking centers: A panel data analysis of bank flows. Journal of International Money and Finance, 22 (4), 483–509. Wu, J., Luca, A. C. and Jeon, B. N. (2011). Foreign bank penetration and the lending channel in emerging economies: Evidence from bank-level panel data. Journal of International Money and Finance, 30 (6), 1128–1156. 33 Technical Appendix A.1 Data sources and de�nitions A.1.1 Rules for ownership determination in the ownership database Ultimate ownership was used if: 1. The main country of foreign ownership is a tax haven, the owner(s) in the tax haven is a holding company runs the �rm as a holding company and not as an operational �rm,38 and total foreign ownership is 50% or more; • If the main foreign owner(s) is a holding company located in a country classi�ed by the OECD as a tax haven or classi�ed by the OECD as an OECD member country with a potentially harmful preferential tax regime; whenever the direct owner is a holding company resident in one of these countries, we assume that the arrangement exists for tax purposes and uses direct ownership, except in cases when there is evidence that the owner is not merely a holding company but an operational �rm in its own right; 2. Majority ownership by a holding company, which functions purely as a holding company, and which is fully owned by a third �rm; • When a Bank is majority owned by a holding company, and that holding company is not itself an operational bank and is deemed to exist purely for the purpose of ownership (according to the best judgment of the authors); and that holding company is fully owned by a parent �rm; then the nationality of the holding company’s parent is used.39 3. Transfer of a bank from its parent to another of the parent’s subsidiaries for the purpose of being absorbed by that other subsidiary; • Ownership was transferred from a parent company to another subsidiary for the purpose of absorption by that other subsidiary. The nationality of the parent is applied to that year (the �nal year of the bank’s existence), since the bank is in effect still directly owned by that parent at the time when it becomes part of another bank which is owned by that parent, and effectively loses its autonomy more or less at the time of transfer to the domestic sibling. • For example, Banca Italo Albanese (Albania) is owned by Intesa (Italy) for several years, until March 2008, when the bank is acquired by ABA, another Albanian subsidiary of Intesa, from the parent com- pany, for the purpose of absorbing it. Banco Italo Albanese is immediately absorbed by ABA, and ceases to exist as a bank. Despite that the last owner of the bank in 2008 was Albanian, 2008 ownership is recorded as Italian. A.1.2 Classi�cation of banking and �nancial crises Systemic banking crises are taken from Laeven and Valencia (2012). In this dataset, a banking crisis is de�ned as a systemic banking crisis when two conditions are met: 1. Signi�cant signs of �nancial distress in the banking system (as indicated by signi�cant bank runs, losses in the banking system, and bank liquidations); and 2. Signi�cant banking policy intervention measures in response to losses in the banking system. The de�nition does not include isolated banks in distress. The year in which a systemic banking crisis starts is identi�ed by the two conditions just mentioned and when at least three out of the following �ve policy interventions have been used (Laeven and Valencia, 2012, p. 4): • Extensive liquidity support (ratio of central bank claims on the �nancial sector to deposits and foreign liabilities exceeds 5% and more than doubles relative to its pre-crisis level); 38 In practice, it is occasionally difficult to de�nitively ascertain whether a given �rm operates as a pure holding company or not, and so holding company status was established with reference to relevant public documentation. 39 In most cases when this rule is applied, the ultimate owner is a large global bank with a familiar name (HSBC, Citibank/Citigroup, etc.) 34 • Large bank restructuring costs (at least 3% of GDP, excluding asset purchases and direct liquidity assistance from the treasury); • Bank nationalizations (treasury or central bank asset purchases exceeding 5% of GDP); • Signi�cant asset purchases; • Signi�cant guarantees put in place (excluding increases in the level of deposit insurance coverage); or • Deposit freezes and bank holidays. When a country has faced �nancial distress but fewer than three of these measures have been used, the event is classi�ed as a crisis if one of the following two conditions has been met: 1. Country’s banking system exhibits signi�cant losses resulting in a share of nonperforming loans above 20% or bank closures of at least 20% of banking system assets; 2. Fiscal restructuring costs of the banking sector exceed 5% of GDP. A.1.3 Construction of trade collapse and �scal stimulus treatments The treatment variable for trade collapse was constructed by �rst compiling total trade (the sum of imports and exports) for a given economy, and computing the percentage change in total trade flows between 2006 and 2009. Only the economies that experienced a net decline in trade flows between the two periods were then sorted, and the threshold for what constituted a trade collapse was then de�ned as contractions that fell below the median (the 50th percentile) of this group. This is equivalent to a percentage decrease of total trade of -3.9%. By this de�nition, this treatment includes 66 treated banks, with 295 nontreated banks. Comparable results to that reported in the text were obtained when more stringent (e.g. the 30th percentile, or a fall of 13.9%) or relaxed (e.g. the 70th percentile, which implies a fall of 3.9%) de�nitions of a trade collapse were employed (these are reported in the optional tables in the appendix). The �scal stimulus treatment is based on the dataset by Grail Research (2009), which compiles, inter alia, the total announced bailout amounts in U.S. dollar terms. These were then normalized by 2008 GDP from the World Development Indicators. Stimulus amounts ranged from 86 and 47 percent of GDP at the high end ($400 billion and $2.1 trillion, Saudi Arabia and China, respectively) to 0.07 and 0.04 percent of GDP at the low end ($15 billion and $ 200 billion, Jamaica and Romania, respectively). An economy is coded as having experienced a stimulus treatment if the stimulus amount exceeded 2.5% of GDP. By this de�nition, this treatment includes 226 treated banks, with 135 nontreated banks. Comparable results to that reported in the text were obtained when a more stringent de�nition of a �scal stimulus was employed (e.g. a stimulus of 5% of GDP), although the treatment in this latter case comprises 162 treated banks and 199 nontreated ones. 35 Table A.1: Sources and de�nitions for main variables of interest Variable De�nition Source Variable of interest † Loans Stock of gross loans less reserves for impaired loans/NPLs Bankscope Crisis 1 if the home country experienced a systemic banking crisis; 0 otherwise‡ Authors/Laeven and Valencia (2012) Core bank level covariates Size Stock of total earning assets Bankscope Solvency Ratio of equity to total assets (%) Bankscope Income to loan ratio Net current income/Total loans (%) Bankscope Interest margin Interest income on assets less expense paid on liabilities/Total assets (%) Bankscope Core country level covariates GDP growth Real GDP growth, lagged one year WDI* GDP per capita GDP per capita (constant 2000 USD) WDI 36 Inflation Inflation, consumer prices (annual %) WDI Current account balance Current account balance (% of GDP) WDI Offshore Dummy for home country classi�ed as offshore �nancial center BIS§ Additional and alternative bank, and additional country covariates Liquidity Liquid assets/Total Assets (%) Bankscope Wholesale Net loans as a percentage of customer funding (%) Bankscope Pro�tability Return on average equity (%) Bankscope Weakness Ratio of loan loss provisions to net interest revenue (%) Bankscope Trade openness Imports plus exports (% of GDP) WDI Financial exports Insurance and �nancial services (% of service exports, BoP) WDI Bank capital Bank capital to assets ratio (%) Beck et al. (2000) Nonperforming loans Ratio of banks’ nonperforming loans to total gross loans (%) Beck et al. (2000) † Gross loans include residential mortgage, other mortgage, other consumer/retail, corporate and commercial, and other loans. ‡ The construction of this variable is described in detail in the text. * WDI = World Development Indicators. § BIS = Bank of International Settlements. A.2 Data sample Table A.2: Baseline sample of home countries by crisis and noncrisis status, with corresponding number of banks Country Banks Country Banks Country Banks Crisis countries * (17 countries; 208 banks) Austria 10 Ireland 1 Portugal† 7 Belgium 3 Italy 6 Slovenia† 1 Denmark 1 Latvia 1 Spain 16 France† 28 Luxembourg 3 United Kingdom 46 Germany 13 Netherlands 18 United States 38 Greece 14 Nigeria 2 Noncrisis countries (49 countries; 153 banks) Argentina 4 Honduras 1 Panama‡ 6 Australia 2 Hong Kong‡ 2 Peru 2 Azerbaijan 1 Hungary 3 Russia 9 Bahrain‡ 6 India 9 Saudi Arabia 1 Botswana 2 Indonesia 1 Singapore‡ 6 Brazil 9 Israel 4 South Africa 9 Canada 8 Japan 10 Sweden 1 China 1 Jordan 1 Switzerland 4 Colombia 4 Kazakhstan 1 Thailand 1 Costa Rica 2 Kenya 4 Togo 5 Croatia 1 Korea, Rep. 2 Turkey 5 Dominican Rep. 2 Lebanon‡ 2 UAE 4 Ecuador 1 Libya 4 Uruguay 3 Egypt 1 Liechtenstein 1 Uzbekistan 1 Estonia 1 Malaysia 1 Venezuela 1 Finland 1 Mauritius‡ 1 Guatemala 1 Mexico 1 * As de�ned by Laeven and Valencia (2012). † Borderline banking crisis. ‡ Offshore �nancial center. 37 Table A.3: Baseline sample of host countries, and corresponding number of foreign and domestic banks Country Foreign Domestic Country Foreign Domestic Host Countries (51 countries; 361 foreign banks; 738 domestic banks) Algeria 5 3 Kenya 5 15 Angola 4 4 Lebanon 3 20 Argentina 15 41 Lithuania 5 3 Armenia 6 2 Macedonia 2 3 Belarus 4 4 Malaysia 11 22 Bolivia 4 6 Mauritius 6 3 Bosnia & Herz. 8 5 Mexico 14 19 Botswana 3 5 Moldova 2 7 Brazil 26 52 Nepal 2 10 Bulgaria 7 7 Pakistan 7 11 Cameroon 5 1 Panama 17 9 China 5 58 Paraguay 7 3 Colombia 5 6 Peru 6 5 Congo, Dem. Rep. 4 1 Romania 15 3 Costa Rica 5 34 Russia 23 168 Cˆote d’Ivoire 4 1 Senegal 5 1 Dominican Rep. 2 27 Sierra Leone 2 3 Ecuador 2 13 South Africa 7 19 Egypt 9 10 Tanzania 11 4 El Salvador 4 2 Tunisia 5 8 Georgia 4 2 Turkey 10 11 Guatemala 3 10 Uganda 9 1 Honduras 3 7 Uruguay 13 3 India 6 48 Venezuela 3 11 Indonesia 16 18 Zambia 6 1 Kazakhstan 6 8 38 A.3 Additional tables In Table A.4, we report a comparison of means (and accompanying standard errors) between crisis treatment and nontreatment foreign banks, along with the difference in the two groups, for the years 2006 and 2009. Table A.5 provides summary statistics for the main variables in the effective sample. Table A.7 reports baseline difference- in-difference regressions (corresponding to speci�cation (B4)) on subsamples that either exclude or include a given region.40 Table A.4: Student’s t-tests for bank lending, 2006 and 2009† 2006 2009 Difference Crisis treatment 5.83 6.36 0.52 (0.14) (0.18) (0.20)∗∗ Nontreatment 4.63 5.51 0.88 (0.17) (0.26) (0.27)∗∗∗ Difference 1.20 0.85 -0.36 (0.30)∗∗∗ (0.26)∗∗∗ (0.14)∗∗ † Means are for bank lending in log form. Standard er- rors are reported in parentheses and are estimated by linear regression with clustering by host country. Dif- ferences are calculated as that between 2009 (treat- ment) and 2006 (nontreatment). ∗ indicates signif- icance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. 40 Naturally, small sample sizes substantially reduce our ability to draw inferences from subsamples that only include banks within a region, but these are reported for completeness. 39 Table A.5: Summary statistics for main variables of interest† Variable N Mean Std. Dev. Min Max Crisis banks Total loans, 2006 221 2,344.02 7,007.38 0.97 53,633.43 Total loans, 2009 221 3,279.40 8,963.77 0.03 68,622.90 Size 221 6.78 1.76 0.75 11.17 Solvency 221 1,404.57 1,166.55 104.00 9,394.00 Income-to-loan 221 0.39 3.51 -0.24 44.05 Interest margin 221 580.95 354.66 -921.00 2,240.00 GDP per capita 221 26,494.67 8,865.31 2,166.32 53,628.23 Lagged GDP growth 221 2.56 1.13 0.68 10.60 Inflation 221 2.41 0.95 1.06 8.59 Current account balance 221 -2.08 6.53 -22.68 13.78 Noncrisis banks Total loans, 2006 143 539.55 1,358.54 0.10 7,736.05 Total loans, 2009 143 910.83 1,945.48 1.97 12,925.52 Size 143 5.57 1.63 -0.01 9.88 Solvency 143 1,895.05 1,609.80 237.00 8,031.00 Income-to-loan 143 0.07 0.37 -0.57 4.12 Interest margin 143 606.87 491.09 -208.00 4,085.00 GDP per capita 143 10,668.40 12,149.55 257.40 39,965.86 Lagged GDP growth 143 5.51 3.29 0.93 26.40 Inflation 139 4.95 3.76 0.24 14.45 Current account balance 142 2.99 10.40 -15.39 39.25 † Notes: All variables are for 2006, unless otherwise stated. 40 Table A.6: Falsi�cation tests for linear difference-in-difference re- gressions for bank lending with alternative de�nitions of the trade treatment, 2006 and 2009† A.F1 A.F2 A.F3 A.F4 treat = trade30p treat = trade70p Crisis 0.025 -0.071 0.889 0.435 (0.17) (0.27) (0.33)∗∗∗ (0.74) Fixed effects/core covariates Home Yes Yes Yes Yes Host Yes Yes Yes Yes Bank No Yes No Yes Adj. R2 0.490 0.558 0.490 0.558 Clusters (countries) 66, 51 49, 42 66, 51 66, 51 Estimation OLS OLS OLS OLS N (banks) 361 361 361 361 † The dependent variable is in log differenced form. Point estimates and heteroskedasticity and intragroup correlation-robust standard errors (re- ported in parentheses) with two-way clustering. All bank-level covariates enter with their values set in the pre-crisis period (t = 2006). A constant term was included in the regressions, but not reported. ∗ indicates signif- icance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. 41 Table A.7: Baseline difference-in-difference regressions for bank lending, 2006 and 2009, regional subsamples† A.S1 A.S2 A.S3 A.S4 A.S5 A.S6 Crisis effect -0.435 -0.386 -0.352 -0.468 -0.411 -0.574 (0.17)∗∗∗ (0.22)∗∗ (0.27) (0.20)∗∗ (0.16)∗∗∗ (0.20)∗∗∗ Fixed effects Home Yes Yes Yes Yes Yes Yes Host Yes Yes Yes Yes Yes Yes Subsample Excluding EAP ECA LAC MNA SAS SSA Adj. R2 0.499 0.516 0.531 0.493 0.531 0.410 Clusters (countries) 59, 48 51, 39 51, 35 64, 47 65, 48 59, 38 Estimation OLS OLS OLS OLS OLS OLS N (banks) 329 269 232 339 346 290 A.S7 A.S8 A.S9 A.S10 A.S11 A.S12 Crisis effect -1.482 -0.296 -0.630 -0.104 -2.735 -0.373 (1.76) (0.35) (0.32)∗∗ (0.00)∗∗∗ (0.00)∗∗∗ (0.45) Fixed effects Home Yes Yes Yes Yes Yes Yes Host Yes Yes Yes Yes Yes Yes Subsample Including EAP ECA LAC MNA SAS SSA Adj. R2 0.508 0.561 0.404 0.644 0.374 0.792 Clusters (countries) 16, 3 27, 12 27, 16 11, 4 9, 3 16, 13 Estimation OLS OLS OLS OLS OLS OLS N (banks) 32 92 129 22 15 71 † The dependent variable is in log differenced form. Heteroskedasticity and intragroup correlation-robust standard errors with two-way clustering reported in parentheses. ∗ indicates signi�cance at 10 percent level, ∗∗ indicates signi�cance at 5 percent level, and ∗∗∗ indicates signi�cance at 1 percent level. Regions correspond to World Bank regions (EAP = East Asia and Paci�c; ECA = Eastern Europe and Central Asia; LAC = Latin America and Caribbean; MNA = Middle East and North Africa; SAS = South Asia; SSA = Sub-Saharan Africa). Own- ership is either publicly-listed (pub. list.) or government-owned (govt.), the other group being privately-held banks. Fixed effects for home and host are time varying. A constant term was included in the regressions, but not reported. Cluster sizes are reported for home and host, respectively. Subsample are de�ned by the exclusion of all banks within a given region, or the inclusion of only banks within a given region. 42