International aid and financial crises in donor countries

European Journal of Political Economy 32 (2013) 232–250

Contents lists available at ScienceDirect

European Journal of Political Economy journal homepage: www.elsevier.com/locate/ejpe

International aid and financial crises in donor countries Hai-Anh Dang, Stephen Knack ⁎, F. Halsey Rogers The World Bank, 1818 H St. NW, Washington, DC 20433, United States

a r t i c l e

i n f o

Article history: Received 7 February 2012 Received in revised form 19 August 2013 Accepted 23 August 2013 Available online 30 August 2013 JEL classification: F35 H87 O19 Keywords: Foreign aid Financial crises

a b s t r a c t The recent global financial crisis placed new economic and fiscal pressures on donor countries that may have long-term effects on their ability and willingness to provide aid. Not only did donor-country incomes fall, but the cause of the drop — the banking and financial-sector crisis — may exacerbate the long-term effect on aid flows. This paper estimates how donor-country banking crises have affected aid flows in the past, using panel data from 24 donor countries between 1977 and 2010. We find that banking crises in donor countries are associated with a substantial additional fall in aid flows, beyond any income-related effects, at least in part because of the high fiscal costs of crisis and the debt hangover in the post-crisis periods. Aid flows from crisis-affected countries are estimated to fall by 28% or more (relative to the counterfactual) and to bottom out only about a decade after the banking crisis hits. In addition, our results confirm that donor-country incomes are robustly related to per-capita aid flows, with an elasticity of about 3. Findings are robust to estimation using either static or dynamic panel data methods to account for possible biases. Because many donor countries, which together provide two-thirds of aid, were hit hard by the global recession, this historical evidence indicates that aggregate aid could fall by a significant amount (again, relative to counterfactual) in the coming years. We also explore how crises affect different types of aid, such as social-sector and humanitarian aid, as well as whether strategic interaction among donors is likely to deepen or mitigate the fall in aid. © 2013 Elsevier B.V. All rights reserved.

1. Introduction: financial crises, economic downturns, and aid outflows Aid's impact on growth in recipient countries has been the topic of a large empirical literature, with highly disparate findings (e.g. Doucouliagos and Paldam, 2008). Clemens et al. (2012) argue that many of the less favorable findings on aid's effects can be attributed to a failure to correct adequately for serious endogeneity issues. Donors tend to give more aid to countries with poor growth prospects, and even when researchers attempt to correct for this problem, their 2SLS estimates are biased towards OLS estimates due to weak instruments (Bazzi and Clemens, 2010; Roodman, 2009). Several recent studies using very different identification strategies have produced evidence that aid has small but positive and significant effects on growth (Clemens et al., 2012; Bruckner, 2011; Frot and Perrotta, 2012). Moreover, aid may matter for other important outcomes. For example, Nielsen et al. (2011) show that, controlling for other factors, a negative aid shock (defined as a sudden reduction in aid of at least 0.5% of GDP) more than doubles the risk of violent conflict in the average country, from 2.1% to 5%. Young and Findley (2011) find that aid targeted to education, health, alleviation of conflict, governance, and civil society significantly reduces the incidence of terrorist events. For example, they estimate that a one standard deviation increase in education aid reduces the number of terrorist attacks by 71% on average, controlling for other variables and correcting for endogeneity. Cuts in aid associated with recent economic crises and growth slowdowns in donor countries could thus have potentially serious consequences. This paper investigates how financial crises and economic conditions in donor countries have affected flows of international aid in the past, with the aim of understanding how the recent global financial crisis could affect future flows. We do this by ⁎ Corresponding author. Tel.: +1 202 458 9712. E-mail addresses: [email protected] (H.-A. Dang), [email protected] (S. Knack), [email protected] (F.H. Rogers). 0176-2680/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.ejpoleco.2013.08.003

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

233

analyzing a 24-nation panel dataset covering 1977–2010 to estimate the effects of banking crises and other economic and political variables on aid.1 These effects turn out to be large, significant, and surprisingly long-lived, with aid from crisis-stricken countries falling relative to the counterfactual by more than 30%, even more than a decade after the banking crisis hits. Moreover, because they are estimated in regressions that control for donor-country per-capita income, these banking-crisis effects are additional to those of the economic downturns that coincide with the banking crises. As a result, the recent global financial crisis could substantially reduce aid in the years to come, given that the countries hit by banking-sector crises provided two-thirds of aid before the crisis. 1.1. Aid and economic slowdowns in donor countries From the perspective of aid policy, the recent global financial crisis was distinctive. A key difference from the emerging-market East Asian and Tequila crises of the 1990s, when multilateral and bilateral donors provided substantial financial assistance to help spark recovery in the crisis countries, is that both of those crises took place at a time of sustained growth in donor countries. This time, it was the donor countries that were hit first by the financial crisis, with GDP falling by 3.2% in 2009 for rich countries as a group (International Monetary Fund, 2010). With the economic slowdown and fiscal strains that resulted from the crisis, donor-country policymakers have found themselves pressured to redirect aid funds to domestic needs such as unemployment benefits and emergency infrastructure programs. Aid from several bilateral donors, such as Ireland and Spain, fell sharply in the initial years after the crisis,2 and although overall ODA from DAC donors continued to rise initially, in 2011 it too declined, by 2.7% in real terms (OECD, 2012). What do we know empirically about the sensitivity of aid flows to donor-country economic and financial conditions? While there is an established literature that examines how recipient-country characteristics — such as income level, population, and political system — affect aid inflows (see, for example, Alesina and Dollar, 2000; Dollar and Levin, 2006), much less is known about these donor-side determinants of aid. Several recent papers have begun to fill this gap, mostly using panel data with donor fixed effects. Aid is positively associated with income per capita in most of these studies (Round and Odedokun, 2004; Boschini and Olofsgard, 2007; Chong and Gradstein, 2008), although one finds no significant relation (Faini, 2006). Faini finds that aid is lower when budget deficits and the stock of public debt are higher, but others find no statistically significant relationship between deficits and aid provision (Round and Odedokun, 2004; Boschini and Olofsgard, 2007). Analyses of survey data find that higher levels of financial insecurity in donor countries are associated with weaker voter support for foreign aid (Paxton and Knack, 2012). None of these studies explore the impact of financial crises in donor countries on aid provision, although two recent studies (discussed below) do. 1.2. The possible exacerbating effects of donor-country banking crises Not all economic slowdowns are created equal, and it is possible that the nature and causes of any particular slowdown may influence its effects on aid flows. Roodman (2008) shows graphically that several OECD donor countries — Finland, Japan, Norway, and Sweden — reduced their aid flows substantially after they suffered domestic banking crises in the 1990s. Although Roodman does not provide further analysis, there are reasons to believe that banking crises could have direct effects on aid beyond their effects on incomes.3 Bank rescues and recapitalizations place massive new fiscal demands on the public sector; even if the government is eventually able to recoup many of the costs of these rescues through asset sales, the short-term effect is to worsen sharply the government's cash flow. In their comprehensive analysis of systemic banking crises, Laeven and Valencia (2008) estimate that fiscal costs of banking crisis management, even net of recoveries from asset sales, average more than 13% of GDP. For the advanced economies, the fiscal cost has typically been less than that — perhaps 3 to 5% of GDP (in gross terms) — but Finland is estimated to have spent 12.8% of GDP to resolve its banking crisis in the early 1990s. With these additional costs added to the usual cyclical revenue shortfalls from recession, donors may find it more difficult to continue giving aid during and after those crises than they would in a normal downturn of the same magnitude. This direct banking-crisis effect may be particularly relevant for aggregate aid flows in the wake of the recent crisis, for two reasons. First, two of the largest donors (in absolute terms), the United States and United Kingdom, were the major economies whose banking sectors were hit first, and most severely, by systemic crises beginning in 2007. Second, the crisis spread to banking sectors in other major donor countries in 2008, particularly those with the greatest exposure to US financial institutions (Laeven and Valencia, 2010). Two other recent papers attempt to estimate the impact of donor-country financial-sector troubles on aid provision. Mendoza et al. (2009) use 1967–2007 data for the US only and show that stock market volatility, which they call “a proxy for financial volatility and economic uncertainty,” is associated with reduced ODA. The other study is more directly parallel to our paper: Frot (2009) uses data for a panel of donors for the 1986–96 period to estimate how financial crises affect aid flows. He finds that a banking crisis in a donor country decreases aid by 13% (when estimated as a level effect) or that aid falls by 5% per year after the 1 Throughout this paper, we use “aid” as shorthand for official development assistance (ODA) provided by bilateral donors that have reported aid through the OECD Development Assistance Committee (DAC) for an extended period (a group that includes donor countries that are not members of the DAC). 2 For details, see discussion and sources in Section 5 below. 3 Devarajan (2008) questions this analysis, pointing out that since 1960 the only sustained decline in aid came in the 1990s, and not necessarily for economic reasons. He cites results from Paxton and Knack (2012) in arguing that “aid is motivated largely by non-economic factors.”

234

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

onset of a crisis (when estimated in trends). Frot also estimates the impact of GDP, unemployment, and budget deficits on aid, using a vector autoregression (VAR) model. 1.3. Contributions Our paper extends the literature in several important respects. First, our panel data analysis exploits a larger set of crises over a longer period of time (1977–2010), taking advantage of a recent update by Laeven and Valencia (2010) of their banking-crisis database. This rich database allows us to differentiate crises that are still ongoing at a given point in time from those that have been resolved (using crisis end dates provided in Laeven and Valencia, 2010), and also to differentiate second crises from first crises, in the handful of donors with multiple banking crises. Second, we analyze the effects of crises not only on total aid volumes, but also on aid broken down by component (namely social sector and humanitarian aid) and channels of contribution (namely bilateral aid or contributions to multilateral agencies). We believe that it is useful to do so, since donors may protect certain types of aid — such as aid that is perceived to respond to immediate needs, or aid through favored channels — and that crises may therefore have differing effects on these components. Third, we investigate donors' strategic behavior by differentiating between donors' responses to “isolated” crises (such as Japan's in the late 1990s) and their responses to “correlated” crises (such as in the Nordic countries in the early 1990s, and in the recent crisis beginning in 2007). Such differentiation is important, because this is the most highly correlated set of crises that the donor countries have suffered, so strategic interactions may cause their post-crisis aid patterns to deviate from the typical historical pattern. And finally, we subject our main results to a battery of sensitivity tests for a range of issues, such as treatment of “borderline” crises and alternative start dates for Japan's crisis. We also test robustness of results to dropping potentially influential observations (donors or years), and to inclusion of a range of economic and political variables that may affect aid levels. Results from tests that include debt and other fiscal variables suggest that banking crises affect aid in part through fiscal channels. 1.4. Structure of the paper The methodology and data are described in the following section, and results for overall ODA volumes are presented in Section 3. In Section 4, we test for effects of crises on the composition of aid, including aid for the social sectors and aid through different channels. We provide more detailed interpretation of the results and discussion of the implications for aid of banking crises in Section 5. The final section considers the implications of these findings for aid flows in the coming years, and for our understanding of how donors interact strategically in their voluntary provision of development aid as a global public good. 2. Methodology and data 2.1. Methodology Our general approach is to estimate some version of the following equation: Aidit ¼ αbanking crisisit þ βXit þ γZit þ μ i þ ηt þ εit

ð1Þ

where Aidit is overall aid from the country and i indexes the donor country and t the year. As our primary dependent variable, we use the log of net ODA disbursements in current dollars.4 banking crisisit, the primary variable of interest, captures whether the country has recently been hit by a banking crisis (with crises coded as described in the next section) in any given year. Xit is a vector of time-varying control variables, with per-capita GDP and population the most important. These two variables are included in virtually all studies of aid determinants, either individually or combined as total GDP, and so we include them in all of our specifications. Given the heterogeneity among countries and the likely importance of unobservables correlated with the error term in determining aid flows, we use donor fixed effects (μi) to capture time-invariant country-specific influences on aid provision. In all specifications we include year dummies (ηt) to account for common shocks to aid volumes in any given year. Coefficient estimates for the variables indexed with i and t are therefore informed only by within-country variation over time. The vector Zit represents a set of other plausible time-varying control variables that appear in the literature less frequently than income and population do. These variables have typically not been shown to be robustly significant in other studies, but we use them to test the robustness of our key variables and to explore informally the channels through which banking crises may reduce aid flows.5 Following Mosley (1985) and Round and Odedokun (2004), we control in most of our tests for the total volume of aid provided by the other 23 donors (lagged a year); in their tests, they interpret the positive and significant coefficient on this variable as supporting a “peer pressure” effect. In robustness tests, we also include several economic control variables from the literature: inflation (Mendoza et al., 2009), trade integration (Boschini and Olofsgard, 2007), unemployment (Boschini and

4 Boschini and Olofsgard (2007) and a few other studies similarly use aid in dollars. Faini (2006) uses aid's share of GDP, while Frot (2009) uses (log of) aid per capita, without including population as a regressor. Our results change very little if we substitute aid/GDP or aid per capita as the dependent variable. Detailed results are available on request. 5 Summary statistics for our main sample are shown in Appendix Table 3.

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

235

Olofsgard, 2007; Frot, 2009; Mendoza et al., 2009), and fiscal budget surplus and stock of public debt (Round and Odedokun, 2004; Faini, 2006; Boschini and Olofsgard, 2007; Frot, 2009). We add to this list government expenditures and real exchange rates. The political variables for which we control include corruption (Chong and Gradstein, 2008) and political orientation of the governing party (Round and Odedokun, 2004; Boschini and Olofsgard, 2007; Chong and Gradstein, 2008; Mendoza et al., 2009). 2.2. Aid data Data on official development assistance are taken from the OECD-DAC databases, which are the standard source used in the literature, and cover the period from 1977 through 2010. In the empirical work reported below, we follow most other studies in using data on net aid disbursements from the DAC Tables rather than aid commitments. There is typically a wide gap between commitments and disbursements, and ultimately it is the disbursements of aid — and not the higher committed amounts — that matters to the recipient countries. When analyzing how crises affect social-sector and humanitarian aid, we use commitment data (from the DAC's Creditor Reporting System), because far fewer observations are available on disbursements at the sector level. We also test the robustness of our findings to using two other measures of disbursements, in addition to net disbursements. The first alternative is gross disbursements, which measures only total disbursements and does not net out repayments of loan principal. The second alternative is Net Aid Transfers (NAT), using data made available by the Center for Global Development and described in Roodman (2006, 2010). NAT subtracts out not only repayments of principal but also interest payments and cancelation of non-ODA loans (aka “debt relief”). Thus NAT more closely approximates the current budgetary outlays associated with ODA. Our sample covers 24 donor countries from 1977 to 2010. (For a list, see Appendix Table 1.) For 19 of these countries, aid data are available for the entire 34-year period. For the other five, which became donors after 1977, the period of coverage is between 19 and 30 years. We omitted several new donors that began reporting aid data to the DAC only in the late 1990s.6 Most of these transition economies are also dropped by Demirguc-Kunt and Detragiache from their recent banking-crisis analysis (2005: 71) because “[banking] problems in these countries were of a special nature.” 2.3. Data on banking crises Our primary source on banking crises is the comprehensive new database assembled by Laeven and Valencia (2010), which covers 1970 to 2009. This dataset includes only systemic crises, which Laeven and Valencia define as those meeting two conditions: • Significant signs of financial distress in the banking system (as indicated by significant bank runs, losses in the banking system, and bank liquidations); • Significant banking policy intervention measures in response to significant losses in the banking system. Policy intervention in turn is deemed to be significant if at least three of the following six measures have been used: (1) extensive liquidity support, (2) bank restructuring that costs at least 3% of GDP, (3) significant bank nationalizations, (4) significant guarantees in liabilities put in place, (5) significant asset purchases (at least 5% of GDP), and (6) deposit freezes and bank holidays. In most of the 10 donor countries in our dataset that Laeven and Valencia classify as having entered a crisis in either 2007 or 2008, the first four of these conditions were met; in the US and United Kingdom, the fifth condition was also met. Laeven and Valencia classify five additional donor countries in 2008 as “borderline” cases that “almost” met their criteria for a systemic crisis, because they had undertaken two (in most cases, the first and fourth on the list above) of the six policy measures. A sixth donor country banking crisis classified as “borderline” by Laeven and Valencia is the US savings and loan crisis, beginning in 1988. In most of our tests, we include these six borderline cases as crises. Including them means we have more crises and therefore more variation in our key independent variable. Of course, if these borderline crises are much less severe than the others, they may have smaller impacts on aid, biasing downward the coefficients on banking crisis variables in the analysis. However, the crises that Laeven and Valencia classify as borderline based on the extent of policy interventions do not appear to be any less severe in terms of outcomes such as estimated output losses (measured as deviations from trend GDP) or fiscal costs incurred by governments. For example, the fiscal costs of the borderline US 1988 crisis are estimated at 3.7% of GDP, compared with 3.6% for Sweden and 2.7% for Norway in their crises of the early 1990s (Laeven and Valencia, 2008). Estimated output losses exceed 30% of GDP for 5 of the 10 donor countries fully meeting the criteria for a crisis in 2007 or 2008, but also for 3 of the 5 donor countries with borderline crises (Laeven and Valencia, 2010). Sweden's output losses are estimated at 33% for its 1991 crisis and 31% for its borderline 2008 crisis (Laeven and Valencia, 2010). We therefore include the borderline cases as crises in most of our tests, but show how results change if the borderline cases are excluded. Appendix Table 1 lists beginning and ending dates of the crises occurring in our 24 donor country sample between 1977 and 2010, with borderline cases shown in italics. The end year of the crisis is defined as the last year of a crisis preceding two consecutive years of real GDP growth and real credit growth. Based on this criterion, all of the crises beginning in 2007 or 2008 were still ongoing as of the end of 2010, because GDP declined in all crisis countries in 2009. 6 These recent donors reporting only low levels of aid include Czech Republic, Greece, Hungary, Poland, and Slovak Republic. In 2008, total net disbursements from these countries were only 1.2% of the total for the 24 countries in our sample. The DAC tables also record aid from “Arab countries” and “other donors” that sums to about 5% of the total for countries in our sample in 2008.

236

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

2.4. Data on income, population, and other control variables For per-capita income, population, and the other control variables in Zit, we use data from a variety of sources, including the OECD (OECD, 2011a, 2011b), World Development Indicators database (World Bank, 2011), World Economic Outlook (IMF, 2010), International Country Risk Guide (PRS Group, 2010), and Database of Political Institutions (Beck et al., 2001, updated 2010). Appendix Table 2 provides a detailed list of variables and sources for the data we use in this paper.

3. Results: donor-country financial, economic, and political characteristics and ODA

150 100 50

disbursed aid index with t0=100

Do donors give less aid after suffering banking crises? Fig. 1 gives a straightforward first-cut answer to that question, by comparing aid flows for crisis and non-crisis countries before and after a crisis hits.7 Crisis countries are defined as those that suffered any systemic banking crisis during 1977–2010, according to the Laeven and Valencia (2010) definition; non-crisis countries are those that did not. For the crisis group, the figure shows the indexed unweighted mean of the indices of aid volumes (in current dollars) before and after the onset of the crisis, with the first year of each crisis synchronized to t = 1. For example, year 1 refers to 1991 in Finland, 1997 in Japan, and 1988 in the United States. For non-crisis countries, we take the mean of aid levels from t = −10 to t = 13 for each of the different crisis years t = 1, then take the mean of those means to come up with the non-crisis trend line shown in the figure. Aid levels are indexed to 100 for year 0 for both crisis and non-crisis countries. The crisis countries include only mature donors for which we have at least 10 years of pre-crisis and 13 years of post-crisis aid data, to ensure a reliable comparison of long-term trends. (Fig. 2 shows the corresponding graph for the recent global financial crisis, for which few post-crisis years are available, of course.) Five countries meet those criteria and are included as the crisis countries in the figure: Finland, Japan, Norway, Sweden, and the US.8 Fig. 1 shows a clear correlation between banking crises and subsequent declines in (relative) aid flows. For years − 10 through year 0, aid from crisis countries follows roughly the same upward trends as that of non-crisis countries. After the crisis hits, in

-10

-5

0

t

5

10

mean net disbursed aid for crisis countries, excl. S. Korea mean net disbursed aid for non-crisis countries

Fig. 1. Net disbursed aid from crisis vs. non-crisis donor countries. Note: Crisis countries are defined as those mature donor countries that suffered any systemic banking crisis during 1977–2010 for which there are sufficient years of aid data before and after the crisis (10 years and 13 years, respectively). For these crisis countries — Finland, Japan, Norway, Sweden, and the US — the figure shows the unweighted mean of the indices of aid volumes (in current dollars, normalized to 100 in the last pre-crisis year) before and after the onset of the crisis, with year 1 of all crises synchronized to t = 1. The non-crisis countries' trends are calculated by taking the mean indices for all non-crisis countries for each t = 0, then averaging them. Source: Authors' calculations from OECD-DAC aid data, using definitions of crisis-onset dates from Laeven and Valencia (2010)

7

The figure thus generalizes and averages the individual country crisis–aid relationships shown by Roodman (2008) in his original post on this issue. Spain and Turkey are excluded because their crises began in 1977 and 1982, respectively, but they began reporting aid data to the DAC only in 1980 (Spain) and 1991 (Turkey). Korea, which was a relatively new donor at the time of its crisis in 1997, is also excluded. Its aid started from a low base and grew rapidly for secular reasons through much of this period as Korea sought a stronger role on the global stage, whereas the mature donors that went through banking crises starting in 2007–08 are generally much more like the other mature donors depicted in Fig. 1. When Korea is included in the graph, the differential pre- and postcrisis behavior of crisis and non-crisis countries is still apparent, but it is less dramatic. Aid from crisis countries grows much more rapidly than aid from non-crisis countries in the years leading up to the crisis, while the two move roughly together in the years after the crisis. The inclusion of country fixed effects in the regressions reported below ensures reliable comparisons without dropping these countries. 8

60

80

100

120

237

40

disbursed aid index with t0=100

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

-5

-4

-3

-2

-1

0

1

2

3

t mean net disbursed aid for crisis countries mean net disbursed aid for non-crisis countries

Fig. 2. Net disbursed aid from crisis and non-crisis donor countries before and after the global financial crisis of 2007–08. Note: Crisis countries are defined as those donor countries that are coded by using definitions of crisis-onset dates from Laeven and Valencia (2010) as having been hit by a banking crisis in 2007 or 2008. For these 15 crisis countries, the figure shows the unweighted mean of the indices of real aid volumes (in current dollars, normalized to 100 in the last pre-crisis year) before and after the onset of the crisis, with year 1 of all crises synchronized to t = 1. The non-crisis countries' trends are calculated by taking the mean indices for all non-crisis countries for each t = 0, then averaging them. Source: Authors' calculations using data from Laeven and Valencia (2010), as described in the text.

years 1 through 11, aid from the crisis countries stagnates, bottoming out in relative terms only after a decade of declines that cumulatively drive aid 30 to 40% below the levels in non-crisis countries. By contrast, aid from non-crisis countries on average continues on nearly the same upward trend as before the crisis.

Table 1 Aid and banking crises 1977–2010. Equation

1.1

1.2

1.3

1.4

1.5

1.6

Variations

LSDV (static model)

GMM

LSDVC (bias correction)

2nd crisis variables

2nd crisis observations dropped

Borderline cases not coded as crises

−0.048⁎⁎ (0.023) 0.0019⁎⁎⁎ (0.0006) 2.971⁎⁎⁎ (0.503) 0.203 (1.295) −3.692⁎⁎⁎ (1.352)

0.770⁎⁎⁎ (0.027) −0.015⁎⁎ (0.006) 0.0007⁎⁎⁎ (0.0002) 0.803⁎⁎⁎ (0.141) 0.219 (0.298) 0.554 (0.482)

0.831⁎⁎⁎ (0.026) −0.013⁎⁎⁎ (0.005) 0.0006⁎⁎⁎ (0.0002) 0.469⁎⁎⁎ (0.127) 0.233 (0.259) 0.397 (0.334)

0.831⁎⁎⁎ (0.024) −0.014⁎⁎⁎ (0.005) 0.0006⁎⁎ (0.0002) 0.500⁎⁎⁎ (0.112) 0.190 (0.221) 0.399 (0.416)

0.830⁎⁎⁎ (0.026) −0.012⁎⁎ (0.005) 0.0006⁎⁎⁎ (0.0002) 0.457⁎⁎⁎ (0.126) 0.245 (0.261) 0.217 (0.323)

b0.0001 – 24 763 0.01

b0.0001 0.45 24 731 0.005

b0.0001 0.94 24 731 0.04

0.826⁎⁎⁎ (0.022) −0.011⁎⁎ (0.005) 0.0004⁎⁎ (0.0002) 0.533⁎⁎⁎ (0.126) 0.168 (0.248) 0.384 (0.394) 0.035 (0.031) −0.001 (0.003) b0.0001 0.95 24 731 0.04

b0.0001 0.99 24 694 0.03

b0.0001 0.95 24 731 0.004

Lagged dependent variable Number of years from start of banking crisis Number of years from crisis start squared Log per capita income Log population Lagged aid from other donors Number of years from start of 2nd crisis Number of years from start of 2nd crisis squared F or Wald test p value Sargan test p value No. of donors No. of observations Joint significance of crisis variables

Dependent variable is log of net ODA disbursements. All regressions control for country and year effects, not shown for space reasons. Banking crises dates are from Laeven and Valencia (2010). Equation 1.1 reports results from a LSDV static panel data model, with standard errors adjusted for clustering at the country level. Equation 1.2 results are from one-step system GMM with robust standard errors. The Arellano–Bond test rejects the null hypothesis of zero autocorrelation in the first-differenced errors at order 1 and order 2. Equations 1.3–1.6 report LSDVC models (fixed effects with bias correction) at level 2 of accuracy. Standard errors are bootstrapped with 200 replications. ⁎ p b 0.1. ⁎⁎ p b 0.05. ⁎⁎⁎ p b 0.01.

238

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

The same pattern holds for the recent global financial crisis. Fig. 2 presents another normalized graph of aid from crisis and non-crisis countries, but includes only the crises of 2007–08. This figure includes fifteen crisis countries — many more than in Fig. 1 — though with only three to four years of post-crisis-onset aid data for each country. To avoid major compositional changes in year 4, we restrict the data graphed in Fig. 2 to the first three years of the crisis for each country. Nevertheless, the early returns strongly suggest an echo of earlier patterns: aid from crisis countries fell substantially relative to aid from non-crisis countries in the first three years of the crisis, despite having increased somewhat more rapidly in the years before 2007. Of course, these drops in aid from crisis countries could result from other factors — most obviously, the fall in per-capita income that tends to accompany crises. The rest of this section therefore explores this result econometrically, to add income and other control variables and to account for country fixed effects and common shocks. 3.1. Banking crises: effects on aid during and after the crisis Table 1 shows our core results on banking crises and aid for the full 1977–2010 period and the full set of donor countries. We first discuss the banking-crisis variables, before returning below to the effects of donor-country incomes and populations. We capture the crisis effects using a counter variable that records the number of years since a banking crisis hit, with the first year of the crisis taking a value of 1. To allow the effect to diminish over time, we include the counter in both linear and square terms; both turn out to be highly significant, and the magnitudes are large. Crises potentially affect aid for many years after they have ended, so we do not truncate these variables at the crisis end dates, or at any other dates. Banking crises emerge as a strong predictor of lower aid, based on tests using a variety of estimation methods and specifications (see Table 1). In equation 1.1, we use a simple fixed-effects (LSDV) model, with standard errors clustered at the country level, to correct for heteroskedasticity and within-panel serial correlation.9 The banking crisis coefficient estimates imply that five years after a crisis hits, aid is about 18% lower than it would have been in the absence of a crisis (other variables held equal). Perhaps surprisingly, at that point aid is still declining relative to the no-crisis counterfactual. It does not bottom out until 12 years after the crisis hits, at which point aid is down about 28%. The quadratic specification in equation 1.1 approximates a relationship that can be estimated in a less economical way, with an exhaustive set of dummy variables for the number of years (1, 2, 3, and so on) that have elapsed since onset of a crisis. Rather than attempting to report the full results from this regression for all 33 dummies in Table 1, we depict their coefficients graphically in Fig. 3. The post-crisis dummies show the largest (negative) coefficients for the 8th through 14th years following crisis onset. The dummies for those 7 years are all significant at the 0.10 level (and all but two at the 0.05 level). The pattern depicted by these dummy coefficients is similar enough to the effect traced out by the statistically significant quadratic relationship estimated in equation 1.1 that we use the latter specification in the remainder of the paper. 3.2. Dynamic panel models Aid levels tend to be persistent, because they reflect not only economic conditions but also foreign policy objectives and altruistic concerns of citizens in donor countries that do not change much from year to year. Even when governments make deliberate decisions to reduce or increase aid, they may prefer to phase in changes gradually. Arguably, we should therefore control for lagged aid levels. The model with lagged aid is defined as follows Aidit ¼ αbanking crisisit þ δAidit−1 þ βXit þ γZit þ μ i þ ηt þ εit

ð2Þ

where all the variables are defined as before, except that the lagged dependent variable is now included as one of the explanatory variables, creating a dynamic panel bias. We cannot use the simple LSDV model with country fixed effects to estimation Eq. (2), because the within-group transformation of the variables does not eliminate the dynamic panel bias (Bond, 2002). To address this bias, we use two dynamic panel data models to estimate Eq. (2). One is the system GMM model (Blundell and Bond, 1998), and the other a country fixed-effect model with bias correction (Bruno, 2005a, 2005b). For the former model, Blundell and Bond (1998), building on the previous work by Arellano and Bond (1991), suggest that Eq. (2) can first be differenced to eliminate the country fixed effects, and then both the lagged dependent variables and lagged differences of the dependent variable can be used as instruments respectively for the differenced equation and level equation.10 For the latter model, Bruno (2005a, 2005b) shows that the bias due to the lagged variables can be addressed if a bias correction term is subtracted from the estimated coefficients obtained from the usual fixed-effects model.11 Between these two dynamic panel data models, we prefer the estimation results from the fixed-effects with bias correction (LSDVC) model to those from the system GMM as, strictly speaking, the system GMM model is more appropriate for settings in which there are “small T, large N” panels. In our data, this condition does not hold: the number of time periods (T) is roughly equal to or even larger than the number of countries (N). However, it is useful to show the estimation results from the system GMM model for comparison. Results are provided in Table 1. Coefficient magnitudes from the GMM model (equation 1.2) and the 9 Using conventional standard errors could be equally valid given the small number of cross-sectional units (countries) in our panel (Wooldridge, 2002, 275– 76; Wooldridge, 2003). However, we use the corrected standard errors to be more conservative. 10 For more details on the GMM model, see the cited studies and Blundell et al. (2000). We estimate this model in Stata using xtabond. 11 We estimate this model in Stata using xtlsdvc, written by Bruno (2005a, 2005b).

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

239

Estimated banking-crisis coefficient

0.1 Years after onset of crisis 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22

-0.1

-0.2

-0.3

-0.4

-0.5

-0.6 Fig. 3. Estimated impact of banking crises on net disbursed aid provided by crisis-affected donors, 1977–2010. Note: The figure shows the coefficient on the individual-year counter dummy for years after crisis, as estimated in LSDV regressions controlling for donor-country per-capita income and population, as well as country fixed effects and year dummies across all donors. Year t = 1 corresponds to the first year of the crisis. Source: Authors' calculations using data from Laeven and Valencia (2010), as described in the text.

LSDVC model (equation 1.3) are very similar, particularly for the crisis-related variables.12 Overall, banking crises are found to have a negative impact on aid in both of these models, with reductions in aid predicted to accumulate until 11 (LSDVC) or 12 (GMM) years from the crisis onset. The decline in aid at that point relative to the counterfactual can be calculated at 39 and 44 percentage points, respectively, for the GMM and LSDVC models. This reduction is somewhat larger than the cumulative 28-percentage-point decline estimated above for the LSDV model that does not include lagged aid.13 Caution is required, however, in interpreting coefficients from the dynamic models. These models measure the short-run impacts only, which may be different from the long-run impacts if aid and other variables interact with each other. Generating longer-run estimates of effects on aid (i.e. 5 or 10 years following the onset of the crisis) from these models would require making further assumptions about the dynamic feedback mechanism between aid and other variables. To summarize, banking crises reduce predicted aid by 28% or more below trend, depending on the estimation method, with the largest effect coming only about 10 to 13 years after crisis onset. Moreover, it is likely that the relationship is a causal one, flowing from financial-sector troubles to reduced aid. It is implausible, of course, that the reverse would be true — that future reductions in aid outflows trigger banking crises in the donor countries. A more plausible endogeneity concern is that an omitted third variable could both weaken the banking system and reduce aid several years after the crisis.14 While we cannot completely rule out this possibility, the robustness checks and alternative specifications suggest that it is unlikely. Banking-crisis effects on aid remain large and statistically significant even after control variables representing the major channels for an economic shock (such as trade and GDP) are included in regressions, and even when the regressions are estimated using these dynamic panel estimation methods. 3.3. Robustness to changes in sample, definition of crises or specification Our main banking crisis variables are defined only for the first crisis in each country, but four countries — the US, Turkey, Spain, and Sweden — are hit by a second crisis as well. Including the second-crisis years for these countries in our sample, but not coding for the second crisis, can be viewed as a source of measurement error that could affect results for our main crisis variables. In equation 1.4 we add a second set of crisis variables, based on the start date of second crises, to our base specification from equation 1.3. Controlling for the first-crisis variables (which are left unchanged), the second-crisis indicators may capture additional effects on aid levels. Coefficients on the second-crisis counter variables for these four countries are opposite in sign to the main crisis variables in equation 1.4, but they are not significant either individually or jointly. Controlling for these second-crisis variables, coefficients on the first-crisis variables are slightly smaller (in absolute value) but remain significant at the 0.01 level. Another way of dealing with the possibly confounding influence of second crises is to drop them from the analysis. These dropped observations include Turkey from 2000 to 2010, the US from 2007 to 2010, and Spain and Sweden from 2008 to 2010. 12 We use all lags of the dependent variables as instruments in equation 1.2. To take into account the problem of too many instruments (Roodman, 2009), we also tried restricting the number of lags of the dependent variables to ten or alternatively to five. The crisis variables remain highly significant, with coefficient magnitudes larger in absolute value than in equation 1.2. Estimation results are available upon request. 13 For these dynamic panel models, we use the long-run multiplier coefficient to measure the cumulative impacts of banking crises. 14 An example is the suggestion from one referee that poor executive governance could be one such omitted third variable. However, this explanation is unlikely: before the global financial crisis hit the US, for example, the Bush administration had increased aid substantially even as the governance-related conditions for the crisis were gathering.

240

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

With second-crisis observations dropped in equation 1.5, results for the banking crises variables are very similar to those in the base specification of equation 1.3. Equation 1.6 differs from equation 1.3 in not treating as crises the six cases that Laeven and Valencia (2010) classify as “borderline.” The estimated reduction in aid associated with banking crises remains highly significant, but it is slightly smaller than in equation 1.3, where borderline cases are counted as crises. This finding is consistent with the evidence described above that the economic consequences of borderline crises tend to be as severe as for other banking crises, and supports our decision to treat them together with other crises in the remainder of our empirical tests. Table 2 reports results from LSDVC regressions that depart in one way or another from the base specification of equation 1.3. To conserve space, only the banking crisis results are shown, although we still control for income, population, aid from other donors, and donor and year fixed effects. The banking-crisis effect does not vary much using different measures of aid, as shown in equations 2.1 and 2.2. For both gross disbursements and net aid transfers, the point estimates of the effects are slightly larger than for net disbursements from equation 1.3. Gross disbursements are estimated to decline by 34% after 12 years, and net aid transfers by 33% after 11 years. Equation 2.3 differs from 1.3 in re-coding the start date of Japan's banking crisis from 1997 to 1992, the date assigned by Caprio et al. (2005) and used by Roodman (2008) and Frot (2009) in their analyses of aid. The 1992 date is based on the bursting of asset bubbles in Japan, but the Laeven–Valencia criteria for a crisis were not met until 1997. The choice of start dates does not matter for our results, however. The coefficients on banking-crisis variables in equation 2.3 are nearly identical to those in equation 1.3. Equations 2.4 to 2.6 omit particular countries that are potentially influential in our main results. Finland is dropped in equation 2.4, because Roodman (2008) shows that Finland had the deepest post-crisis trough in aid among the four cases he examined. It turns out not to be an influential case in our analysis; however, the banking crisis variables remain highly significant. Equation 2.5 drops Iceland. Because observations from each of the 24 countries are weighted equally in the regressions, results may be sensitive even to very small donors that have little effect on total aid volumes. Iceland is the smallest, with a population and aid budget equal to less than 1% of the mean values in the sample. However, dropping it has very little effect on estimates for the crisis variables. Korea and Turkey are dropped in equation 2.6. Both countries are new donors, with large fluctuations around an overall upward trend in aid, and their aid levels as a share of GDP or government spending are lower than those of the other 22 donors in the sample. Dropping them also turns out to have negligible impact on our results. Equation 2.7 drops for all countries the four most recent years of data, which cover the years of the global financial crisis. As of this writing, these crises were still fairly new (with the “years since crisis onset” variable equal to either 3 or 4 for 2010) and ongoing. Results for the crisis variables change very little when the analysis is limited to the years 1977–2006. Equation 2.8 controls for an exchange-rate deflator. Aid flows measured in current US dollars are potentially sensitive to exchange-rate changes. However, controlling for exchange rates has little impact on the banking crisis results.

Table 2 Aid and banking crises: robustness to changes in sample and definition of aid. Specification/sample change

No. of years from crisis start

No. of years squared

No. of donors

No. of obs.

2.1) Gross disbursements

−0.016⁎⁎⁎ (0.006) −0.014⁎⁎ (0.006) −0.012⁎⁎ (0.005) −0.013⁎⁎ (0.006) −0.013⁎⁎ (0.005) −0.012⁎⁎ (0.005) −0.016⁎⁎ (0.007) −0.012⁎⁎ (0.005) −0.015⁎⁎ (0.005)

0.0007⁎⁎⁎ (0.0002) 0.0006⁎⁎⁎ (0.0002) 0.0005⁎⁎⁎ (0.0002) 0.0006⁎⁎⁎ (0.0002) 0.0006⁎⁎⁎ (0.0002) 0.0005⁎⁎ (0.0002) 0.0008⁎⁎⁎ (0.0003) 0.0005⁎⁎⁎ (0.0002) 0.0006⁎⁎⁎ (0.0002)

24

731

24

718

24

731

23

698

23

717

22

691

24

635

24

731

24

731

2.2) Net aid transfers 2.3) Japan crisis start date of 1992 2.4) Finland dropped (largest aid decline) 2.5) Iceland dropped (smallest donor) 2.6) Korea, Turkey dropped (lowest aid/GNI) 2.7) 1977–2006 only 2.8) Control for exchange rate deflator 2.9) Time trend and Cold War dummy added

Dependent variable is log of gross ODA disbursements in 2.1, log of net aid transfers in 2.2, and log of net ODA disbursements in 2.3–2.9. All regressions control for country and year effects, per capita income, population, and lagged total aid from other donors. Estimates are from LSDVC models (fixed effects with bias correction) at level 2 of accuracy, with standard errors bootstrapped with 200 replications. Banking crises dates are from Laeven and Valencia (2010). ⁎ p b 0.1. ⁎⁎ p b 0.05. ⁎⁎⁎ p b 0.01.

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

241

3.4. Time trends For space reasons, results for the many year dummies are not shown in tables. Including a full set of year dummies, rather than a single linear time trend variable, makes it possible to control for common shocks or other events that affect all donor countries in particular years. These may include oil price shocks, for example, or new donor commitments like aid to the transition countries in the 1990s. If we add a linear time trend to the baseline regression, its coefficient is negative and significant at the 0.05 level. Thus, controlling for income and other variables in the model, there is some evidence of a downward secular trend in aid over the full period. If we also control for a Cold War dummy (defined as equaling 1 through 1990 and 0 afterward), its coefficient turns out to be positive but insignificant, while the time trend remains negative and significant. There are two opposing arguments about how the end of the Cold War likely affected aid. On the one hand, new recipients in East Europe and Central Asia may have stimulated aid flows, but on the other hand, the reduced geopolitical importance of many traditional aid recipients following the end of the Cold War may have depressed aid flows. Our result suggests the two effects roughly cancel each other out in the aggregate aid data. Inclusion of the time trend and Cold War dummies does not weaken the impact of the crisis variables, as shown in equation 2.9 of Table 2. 3.5. Effects of donor-country income levels As expected, the level of per-capita income (in logs) of a donor country is a strong, robust, and very significant predictor of the amount of ODA that the country gives each year. In the Table 1 LSDV regression predicting disbursements (1.1), the coefficient on log GDP per capita is about 3.0. In the GMM (1.2) and LDSVC regressions (1.3), the implied elasticity is about 3.4 and 2.8 respectively.15 Even when other control variables are added (in Tables 3 and 4 below), which in some cases reduces the sample size by about one third, the estimated income effect varies only slightly. These estimates are substantially higher than coefficients on income in pooled OLS regressions that do not control for country fixed effects (results not shown), probably because of omitted-variable bias in OLS. Most notably, the United States, one of the highest-income donors, has historically given less aid relative to its income than most other OECD countries, for reasons that clearly have little to do with the slope of the aid–income relationship. To some extent, control variables such as the trade/GDP and fiscal deficit/GDP ratios may capture these unobserved variables, and we include those in the later regressions. But because heterogeneity in unobservables is likely to be important as well, the fixed-effects results are likely to be the most illuminating. Because the dependent variable is also in logs, this coefficient on income can be interpreted as an elasticity, implying that a 1% increase in GDP per capita is associated with an increase of about 3% in aid outflows. To gauge the reasonableness of this finding, consider what it implies for aid shares in GDP. With an elasticity of 3, a decade of steady per-capita income growth of 1% per year (with no population growth) would raise a donor country's aid/GDP ratio from 0.35% to 0.43%.16 Given that the incomes of donor countries have grown consistently over recent decades, why do the data not show the steady increase in aid/GDP ratios predicted by these results? From an econometric perspective, a major reason is that over much of the 1977–2010 period, the coefficients on the year dummies were consistently shrinking, indicating a secular decline in average propensity for aid-giving that partially offset the tendency toward higher aid-giving associated with higher donor incomes. The post-1990 decline has been attributed in part to the end of the Cold War, which reduced the geopolitical motivation for aid-giving. Since the late 1990s, however, the year dummies suggest that this secular decline has leveled off. The result that donor-country income is a significant and quantitatively important predictor of aid is consistent with findings of some other recent studies. Our elasticity estimate of about 3 is roughly twice as large as that implied by the income coefficients reported by Boschini and Olofsgard for the somewhat earlier 1970–97 period. Note that our results on income are not particularly sensitive to the indicator of aid that we use. 3.6. Effects of donor-country size While we focus primarily on the banking-crisis and income variables, we also include population (in logs) as a control variable in the core regressions in Table 1. Perhaps surprisingly, population is not a significant predictor of aid disbursements in most of our tests. This finding implies that over time for a given donor, years with higher population are not associated with significantly higher aid disbursements.17 The inclusion of donor fixed effects means that we are not merely showing that aid is low relative to GDP in the most populous donors, the US and Japan. In fact, in random-effects and pooled OLS regressions (not shown here), where estimates are informed by cross-country variation in the data, population is a stronger and highly significant predictor of aid outflows as expected, with an elasticity of close to 1. Overall our results indicate that changes in donor-country population have far less effect on aid levels than do changes in per-capita income. Thus, even though aggregate GDP is a good predictor of aid levels (in regressions that are not shown here), the 15

For example, in the case of equation 1.3, the long-run multiplier coefficient for income is calculated as 0.469 / (1 − 0.831) = 2.78. We also re-ran the regressions with aid/GDP as the dependent variable. In those cases, the coefficient on log GDP per capita was typically on the order of 0.004. From an initial average aid/GDP share of 0.0035, say, this translates into an increase of a little more than 1% in the aid share with 1% income growth. This coefficient magnitude is very similar to those reported by Round and Odedokun (2004). 17 This result is consistent with Boschini and Olofsgard (2007). Log of population is negative and significant in Round and Odedokun (2004), who use aid/GDP as their dependent variable. These studies, like ours, include donor fixed effects. If we use aid per capita as the dependent variable in our tests, the effect of population is significantly negative (results available on request). 16

242

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

Table 3 Aid and banking crises: robustness to economic variables. Equation

3.1

3.2

3.3

3.4

Lagged aid

0.822⁎⁎⁎ (0.025) −0.012⁎⁎ (0.005) 0.0005⁎⁎⁎ (0.0002) 0.454⁎⁎⁎ (0.117) 0.179 (0.212) 0.275 (0.394) −0.001 (0.001) 0.001 (0.001)

0.827⁎⁎⁎ (0.027) −0.011⁎ (0.007) 0.0005⁎⁎ (0.0002) 0.434⁎⁎⁎ (0.136) 0.163 (0.290) 0.163 (0.467) −0.001 (0.001) 0.001 (0.001) −0.002 (0.005)

0.791⁎⁎⁎ (0.032) −0.005 (0.007) 0.0004 (0.0002) 0.462⁎⁎⁎ (0.171) −0.472 (0.428) 0.342 (0.443) 0.0001 (0.005) 0.001 (0.001)

0.831⁎⁎⁎ (0.026) −0.011⁎⁎ (0.005) 0.0005⁎⁎⁎ (0.0002) 0.485⁎⁎⁎ (0.124) 0.178 (0.211) 0.321 (0.391) −0.0003 (0.001) 0.0004 (0.001)

Number of years from crisis start Number of years from crisis start squared Log per capita income Log population Lagged aid from other donors Inflation rate Trade share of GDP Unemployment rate Government expenditures (% of GDP)

0.007 (0.005) 0.003 (0.006) −0.002⁎⁎ (0.001)

Budget surplus/deficit (% of GDP) Net debt (% of GDP) Growth in aggregate GDP (% change from preceding year) Wald test p value Sargan test p value No. of donors No. of observations Joint significance of crisis variables

b0.0001 0.95 24 729 0.02

b0.0001 0.95 24 661 0.06

b0.0001 0.99 24 541 0.13

0.013⁎⁎⁎ (0.004) b0.0001 0.97 24 729 0.03

Dependent variable is log of net ODA disbursements. All regressions control for country and year effects, not shown for space reasons. Estimates are from LSDVC models (fixed effects with bias correction) at level 2 of accuracy, with standard errors bootstrapped with 200 replications. Banking crises dates are from Laeven and Valencia (2010). ⁎ p b 0.1. ⁎⁎ p b 0.05. ⁎⁎⁎ p b 0.01.

elasticity of aid with respect to GDP is not constant. Its magnitude will depend on whether aggregate income has increased as a result of population growth or growth in per-capita income.18 Olson and Zeckhauser (1966) argued that smaller countries (as measured by aggregate GDP) will tend to free-ride on larger countries' contributions to international public goods, and they cited development aid as an example. Whether our data support their argument depends on the estimation method used. Implied elasticities from LSDV, LSDVC, and random-effects models are all greater than 1, while elasticities from the pooled OLS model are slightly below 1. Elasticities above 1 can be interpreted as supporting Olson and Zeckhauser, while estimates below 1 reflect the more commonly observed fact that aid as a share of national income in the US, Japan, and other large donors is relatively small.

3.7. Correlations with other economic variables In addition to income levels, several other possible economic determinants of aid have been examined in the literature: inflation rate, trade exposure, unemployment rate, size of government, budget surplus or deficit, and government debt. In Table 3, we enter each of these in turn. Equation 3.1 adds inflation and trade openness to the base specification of equation 1.3. Inflation has a negative but insignificant coefficient in equation 3.1.19 Contrary to the evidence from bilateral trade relationships (Boschini and Olofsgard, 2007), greater openness as measured by trade share/GDP is positively but not significantly related to aid flows. Unemployment is added in equation 3.2; its inclusion reduces the sample size by about one tenth as unemployment data are not available for most

18 A corollary is that the recent downturn in OECD economies may have larger negative effects on aid than would a downturn of equivalent size caused by a fall in population growth (to suggest an unrealistic thought experiment). 19 Mendoza et al. (2009) show that US aid over time is inversely related to a “misery index” equal to the inflation rate plus the unemployment rate. They do not test these two variables separately, nor do they offer a theoretical justification for their assumption that a percentage point increase in unemployment has the same impact on aid as a percentage point increase in the inflation rate.

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

243

Table 4 Aid and banking crises: robustness to political variables. Equation

4.1

4.2

4.3

Lagged aid

0.812⁎⁎⁎ (0.028) −0.010⁎⁎ (0.005) 0.0004⁎⁎ (0.0002) 0.625⁎⁎⁎ (0.121) 0.191 (0.221) 0.336 (0.391) 0.067⁎⁎ (0.034) 0.066⁎⁎ (0.034)

0.779⁎⁎⁎ (0.027) −0.014⁎⁎ (0.006) 0.0006⁎⁎⁎ (0.0002) 0.603⁎⁎⁎ (0.111) 0.465⁎ (0.279) 0.077 (0.401)

0.833⁎⁎⁎ (0.027)

Number of years from crisis start Number of years from crisis start squared Log per capita income Log population Lag of total aid from other donors Right-wing party Left-wing party Corruption index (ICRG, 0–6 scale)

0.019 (0.018)

No. of years from crisis start (isolated) No. of years from crisis start squared (isolated) No. of years from crisis start (correlated) No. of years from crisis start squared (correlated) Wald test p value Sargan test p value No. of donors No. of observations Joint significance of crisis variables Joint significance of crisis variables (isolated) Joint significance of crisis variables (correlated)

0.465⁎⁎⁎ (0.130) 0.256 (0.264) 0.483 (0.355)

b0.0001 0.93 23 689 0.09 – –

b0.0001 0.63 24 619 0.004 – –

−0.016⁎⁎ (0.007) 0.0007⁎⁎⁎ (0.0002) −0.009 (0.010) 0.0004 (0.0006) b0.0001 0.94 24 731 – 0.01 0.61

Dependent variable is log of net ODA disbursements. All regressions control for country and year effects, not shown for space reasons. Estimates are from LSDVC models (fixed effects with bias correction) at level 2 of accuracy, with standard errors bootstrapped with 200 replications. Banking crises dates are from Laeven and Valencia (2010). Switzerland is missing data on party ideology (equation 4.1). ⁎ p b 0.1. ⁎⁎ p b 0.05. ⁎⁎⁎ p b 0.01.

countries prior to 1980. As with inflation, the coefficient for unemployment has the predicted sign but is not significant. In both equations 3.1 and 3.2, results for the banking crisis variables are little affected by the addition of these economic variables. Equation 3.3 adds three fiscal variables — using OECD data — to the equation 3.1 model (leaving out unemployment, to maximize the sample size). In doing so, we lose about one third of our observations in each test due to missing data, mostly for earlier years in the sample period. Perhaps surprisingly, the coefficient for government expenditures as a share of GDP is positive but not significant. The budget surplus is positively (but not significantly) associated with aid disbursements. The debt stock as a share of GDP is negatively associated with aid disbursements as expected, and is significant at the 0.05 level. This finding is consistent with results in Faini (2006). If banking crises affect aid through fiscal channels, then controlling for debt levels and other fiscal variables should substantially weaken the relationship between crises and aid in our regressions. This pattern is in fact what we observe in equation 3.3, where neither crisis variable is significant at conventional levels. While they retain their hypothesized signs, the coefficient magnitudes are smaller in absolute value than in equation 3.1.20 Equation 3.4 adds the growth in the donor-country's aggregate GDP (in constant dollars) to the equation 3.1 specification. For a given income level, higher growth rates may be associated with reduced budgetary pressures. As expected, the coefficient on growth in equation 3.4 is positive and significant. Results for the banking crisis (and income) variables change very little when we control for growth.

20 These differences are partly due to the change in sample, but even holding the sample constant, the coefficient on years from crisis start declines in absolute value from −0.009 to −0.005 with the addition of the fiscal variables. The coefficient on years from crisis start squared declines only slightly, from 0.00042 to 0.00038.

244

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

3.8. Correlations with political variables Table 4 shows the effects of including other variables that are political in nature or thought to work primarily through the political process. Results for banking crisis variables are robust to their inclusion. The first political variable is ideology (left or right) of the party in government, as coded by Beck et al. (2001, updated 2010).21 In equation 4.1, we test dummies for both rightand left-wing governments (as coded by Beck et al., 2001), with centrist and non-ideological governments as the omitted category. Both dummies have positive and significant coefficients that are very similar in magnitude, indicating that both leftist and rightist governments tend to provide higher aid levels than centrist or non-ideological governments.22 Corruption ratings, from the International Country Risk Guide (with higher values on the 0 to 6 scale indicating less corruption), are not significantly related to aid in equation 4.2. Chong and Gradstein (2008) argue that the corruption index proxies for voter attitudes regarding their government's competence to spend resources effectively. In their study, corruption ratings produced coefficients that were not only positive, as in our regressions, but also statistically significant. Again, the most important result from these analyses is that including these variables does not challenge the main conclusions from earlier tables. The banking-crisis indicators remain significant in Table 4, with their usual signs, although the estimated decline in aid is somewhat smaller in equations 4.1 and 4.2 than in our base specification of equation 1.3.23

4. Do donor-country economic conditions affect the composition of aid? In Table 5 we analyze components of ODA that might be affected differently by banking crises in donor countries. When donors cut aid in response to crises, they might disproportionately reduce contributions to multilateral agencies and protect bilateral aid, which may be linked more closely to high-priority strategic and commercial objectives. On the other hand, contributions to multilaterals may be harder to cut in the short run, as they tend to be negotiated in advance with agreements covering several years. In equation 5.1, we net out donor countries' aid that is channeled through UN or other multilateral agencies, including the EC, World Bank, and regional development banks.24 The estimated effects of banking crises on bilateral aid in equation 6.1 are very similar to those in our base specification of equation 1.1. The elasticity of bilateral aid with respect to income is 2.6, which is not far below the elasticities for total aid implied by the Table 1 regressions. When analyzing only aid funneled through multilaterals in equation 5.2, coefficients for the crisis variables increase somewhat in absolute value. The income elasticity (only 1.9) is somewhat lower, however, than for bilateral aid.25 Economic conditions in donor countries may also affect the sectoral composition of aid. Our hypothesis is that in more difficult economic conditions, donor-country voters may be less supportive of some types of aid than of others. When their own country suffers from high rates of unemployment, voters may feel that their first priority is to stimulate the domestic economy, rather than to try to spur growth, development, and job creation in poorer countries. These sentiments could lead to reduced aid for infrastructure, economic management, or general budget support. By contrast, emergency aid and disaster relief may command support even in difficult times. By the same token, support for projects and programs that promote basic education and health may remain higher, if voters feel that the beneficiaries — children and the sick — cannot be expected to fend for themselves. In a simple test of this hypothesis, we vary the dependent variable from our basic specification of equation 1.3. Instead of using overall ODA disbursements, in equation 5.3 we use “social sector aid,” defined as aid commitments to the social sectors (health and education) and for humanitarian purposes (disaster relief and emergency aid). For comparison, equation 5.4 reports a similar regression using the residual category of “non-social sector aid,” defined as all aid commitments net of “social sector aid”. As noted above, not all aid is classified into categories, so “non-social sector aid” includes not only aid identified as going to other specific sectors, but aid commitments designated as “multi-sector” or not classified sectorally.26 This residual category of aid commitments is about six times as large on average as “social sector aid.” As shown in equation 5.3, social-sector aid appears to be unaffected by banking crises. The coefficients on the banking-crisis variables are much smaller than in equation 1.3, and fall far short of statistical significance. In contrast, the crisis variables are highly significant predictors of non-social-sector aid, and their coefficient estimates imply that aid eventually bottoms out at 38%

21 These ruling-party ideology dummies are missing data for Switzerland, which had a four-party ruling coalition over the whole period that spanned the left– right spectrum. Other observations dropped due to missing data on these variables include France (1977–78), Portugal (1980–82), and Turkey (2003–08). 22 Round and Odedokun (2004) and Boschini and Olofsgard (2007) find no evidence that government ideology matters, but they test only left v. right governments — for which we find little difference — without a third centrist category. Chong and Gradstein (2008) similarly test leftist governments against all others, finding no significant result in fixed effects regressions similar to ours, but a positive effect in dynamic panel data regressions. 23 Equation 4.3 is presented in Section 6 below, as evidence on strategic interactions among donors. 24 Multilateral contributions average somewhat less than 30% of all aid in recent years from these 24 donors. For EU members, it averages about 35%, with contributions to EC aid accounting for about one half of this amount. Contributions to the EC do not appear to “crowd out” contributions to other multilateral agencies, as these account for a little more than one sixth of aid on average for both EU members and non EU countries. 25 Results for multilateral contributions change very little when we exclude contributions to the EC, in equation 5.3. Aid from the EC is often viewed as more similar to bilateral than to multilateral aid (Martens et al., 2002). 26 Also note that sectoral composition of aid is not reported by Iceland and Turkey during 1977–2010, and is missing for certain years for some other donors. We therefore do not control in 5.3 and 5.4 for the lagged sum of sectoral aid from other donors.

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

245

Table 5 Components of aid. Equation

5.1

Aid component Lagged aid component Number of years from start of banking crisis Number of years from crisis start squared Log per capita income Log population Lagged aid from other donors (total bilateral or multilateral) Wald test p value Sargan test p value No. of donors No. of observations Joint significance of crisis variables

5.2

5.3

5.4

Bilateral

Multilateral contributions

Social sectors

Non-social sector

0.843⁎⁎⁎ (0.023) −0.014⁎⁎ (0.007) 0.0007⁎⁎ (0.0003) 0.406⁎⁎ (0.189) 0.244 (0.389) 0.017 (0.434) b0.0001 0.14 24 731 0.04

0.582⁎⁎⁎ (0.025) −0.030⁎⁎⁎ (0.011) 0.0008⁎⁎ (0.0004) 0.788⁎⁎⁎ (0.195) −1.185⁎⁎ (0.541) 1.905⁎⁎⁎ (0.654) b0.0001 0.01 24 731 0.01

0.554⁎⁎⁎ (0.041) 0.001 (0.015) −0.0002 (0.0006) 1.179⁎⁎⁎ (0.434) 1.424⁎⁎ (0.653)

0.738⁎⁎⁎ (0.040) −0.016⁎⁎ (0.008) 0.0006⁎⁎⁎ (0.0003) 0.661⁎⁎⁎ (0.228) −0.089 (0.317)

b0.0001 0.03 22 597 0.94

b0.0001 0.97 22 597 0.11

Dependent variable is log of net ODA disbursements (equations 5.1–5.2) or log of ODA commitments (equations 5.3–5.4) for the specified aid component. All regressions control for country and year effects, not shown for space reasons. Banking crises dates are from Laeven and Valencia (2010). Iceland and Turkey are missing data in equations 5.3 and 5.4. ⁎ p b 0.1. ⁎⁎ p b 0.05. ⁎⁎⁎ p b 0.01.

below the counterfactual after 13 years. These results are consistent with the hypothesis that under financial strain, donors may be more responsive to what they see as immediate needs of vulnerable groups than to appeals for aid to productive sectors (such as infrastructure or finance). There is no certainty that these shifts in sectoral commitments translate into changes in the sectoral composition of disbursements, of course. For example, if education commitments are seen as a signal of aid effort, donor countries may make greater use of that signal in tough fiscal times — even though it may turn out later that they will not be able to follow through with higher disbursements, as suggested by our earlier results on the effects of crisis on disbursements. Thus we cannot be sure what the results imply for actual aid flows. Moreover, the null hypothesis that the overidentifying restrictions are valid is rejected in equations 5.2 and 5.3. Still, these preliminary findings suggest the value of examining the economic determinants of sectoral as well as aggregate aid flows.

5. Interpreting the crisis results What drives these large banking crisis effects on aid? As noted earlier, one likely channel is the major fiscal costs of crises. Aid is forced to compete with other priorities, most of them with larger domestic constituencies, as aid to the financial sector and other expenditures necessitated by the crisis crowd out pre-crisis spending. Articles on aid spending in crisis countries show this effect in action. In the years after the Savings and Loan bailout in the United States, Lancaster (2006) notes that “…new purposes … were not enough to protect foreign aid — especially development aid — from being slashed during the efforts of the 1990s to cut the federal budget deficit and the size of government.” Similar pressures restrained Japan's aid in the early 2000s, as suggested by this quote from 2003: “Japan has relied heavily on aid to project influence overseas. But it has been forced to scale back on this diplomatic clout by a fiscal crisis that will see public debt soar to ¥686 trillion at the end of this fiscal year” (Watts, 2003). In the wake of the recent global crisis, conflicting needs quickly put aid budgets under pressure. Aid is being forced to compete with domestic programs for those in need; for example, as Ireland cut its aid budget substantially in 2008 and 2009, newspapers noted that a leading development campaigner “refused to condemn the cuts”, citing “people in this country who are seriously hurt by this current malaise” (Cunningham, 2009). Such pressures may help explain the recent aid declines in aid from a number of donors. OECD statistics show that despite the overall growth in development aid in 2009, Austria and Italy each cut their aid budget by 31%, Ireland by 19%, Portugal by 16%, and Germany and Greece each by 12% (OECD, 2010). In the cases of Greece and Ireland, the OECD specifically cited fiscal or budgetary pressure as the reason for the fall in aid. In 2010, Ireland, Spain, and Sweden reduced aid by 5 to 7% each, even as overall ODA rose (OECD, 2011a, 2011b). Again, the OECD explained the declines for Ireland and Spain as resulting from fiscal pressures. More recently, overall ODA from DAC donors declined nearly 3% in 2011 (according to preliminary data), and the Secretary General of the OECD referred to an environment in which DAC donors were having to implement “tough fiscal consolidation plans” (OECD, 2012). Twelve of fifteen EU DAC members reduced ODA in real terms, with Greece and Spain cutting aid by 39 and 33%,

246

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

respectively, because of their fiscal and financial crises. Even the UK and US, which had continued to increase aid in the immediate wake of the 2008 crisis, saw slight declines in ODA in 2011. A natural follow-up question is why banking crises should have their most severe effects on aid only in the medium to long term. The negative effects on aid do appear to start soon after the crisis hits, as suggested by both the raw data in Fig. 1 and the regression results, as well as the example from Ireland given above. But according to our analysis, the effects continue to deepen (relative to counterfactual) for up to 10 to 13 years, which seems remarkably long. Recent analysis of the effects of banking crises by Reinhart and Rogoff (2008) suggest a possible explanation. That paper finds that major banking crisis have effects on growth and fiscal positions that are especially large and long-lasting. In particular, they cause a major shock to government revenue, which flips from rising at an average rate of 3% per year in the years immediately before the crisis to declining at about 3% in the first three years of the crisis. At the same time, governments must pay not only for the banking-sector bailouts — the full cost of which are not typically known for years — but also other expenditures needed to address the economic effects of the crisis. As a result, public debt increases by a daunting 86% on average in the first 3 years of crisis. Reinhart and Rogoff conclude that “[a]rguably, the true legacy of banking crises is higher public indebtedness — far over and beyond the direct headline costs of big bailout packages.” Given the size of the debt overhang, it is quite plausible that debt pressures lead governments to cut back or at least restrain growth in aid for years afterwards, as suggested by Faini's (2006) finding that a larger public sector debt predicts lower aid levels. And in fact, the US and Japanese examples cited above both refer to aid cutbacks that took place five years or more after banking crises hit, during a period of attempts to deal with the debt legacy. In the case of Sweden, an OECD development co-operation review from 1996 — five years after that country's banking crisis broke out — makes this link explicitly as it explains the recent fall in ODA: “Because of budget deficits Sweden has embarked on a stringent budget cutting exercise to definitively halt the large deficits of recent years and reverse the serious buildup in public debt”, leading the OECD to expect “a further large fall in [aid] disbursements” (OECD, 1996). An investigation of the impacts of banking crises on fiscal variables is beyond the scope of the present study. However, we show in Table 3 (equation 3.3) that debt is associated with lower aid, and that controlling for debt, coefficients on the banking crisis variables decline in magnitude and are no longer significant. These findings are entirely consistent with the hypothesis that the primary channel by which crises affect aid flows is increasing fiscal pressure. 6. Implications for aid over the next decade, with strategic interactions What do these results on income and banking-crisis effects imply for the likely path of aid disbursements over the next decade? Past is not necessarily prologue, particularly given the unprecedented magnitude of the recent crisis. And without a fully worked-out general equilibrium model, we cannot credibly estimate the joint effects of the numerous variables that we have identified as significant predictors of aid flows. Nevertheless, there are reasons to believe that aid is already well below the counterfactual and could continue to decline substantially in relative terms, given the large and consistently significant effects of the income and banking-crisis variables. First, incomes have taken a big hit. According to recent IMF estimates, GDP in donor countries remained considerably below potential even in 2011, with output gaps of 2 to 3% in several major OECD countries and over 5% in Japan and the US (International Monetary Fund, 2011). But the income costs of the crisis are even greater than these estimates suggest, because the crisis also reduced potential GDP: “much of the loss relative to trends is not expected to be recovered over the medium term”, in the IMF's words (p. 14). The Fund estimates that GDP therefore lags the pre-crisis trend by more than 12% in the US and the UK and by 6– 12% in several other large donor economies (including Canada and Italy, which Laeven and Valencia do not classify as having suffered banking crises). Even in 2016, the Fund projects that real GDP for the advanced countries as a group will still be 10% below the pre-crisis trend. The depth and length of these income gaps are far larger than we would expect in a “normal” recession, underlining the distinctive nature of a downturn caused by banking crises. If the aid–income elasticities of 2.8 to 3.4 operate cyclically the way they do in the long-term panel regressions, these gaps would imply that aid levels are already well below the counterfactual. On top of these income effects come the banking-crisis effects: seven years after the crisis hit, our estimates suggest that the banking-crisis effect will have grown rather than attenuated. The point estimates from the quadratic specification suggest a drop in aid of at least 23% after 7 years, and the individual year dummies graphed in Fig. 3 suggest an even larger drop. Thus after seven years, the combined effects of the financial crisis through these two channels could be to reduce aid by at least 40 to 50% or more relative to the counterfactual.27 These estimated aid effects are very rough, and they do not take into account the effects of other determinants that will also be shifting and that will offset each other. In the short term, for example, the crisis has already led to higher budget deficits and lower inflation relative to trend, both of which are associated with higher aid; but it has also led to higher unemployment, higher inequality, and a lower trade share, all of which are associated with lower aid. Since we do not jointly estimate these

27

Even under the Table 3 specifications that include lagged aid as a regressor, the implied fall in aid is significant, at around 10%.

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

247

effects, these casual simulations should not be interpreted too literally. The point is that, especially in the donor countries that have suffered financial crises — including the US, UK, and thirteen other countries that together provide two-thirds of ODA — aid is likely to fall well below the counterfactual and remain lower for the next decade. And in some countries, the effects on aid could be larger than those suggested above, because the crisis effects are more severe. In the case of Ireland, for example, income per capita fell by about 15% from 2007 to 2010. The evidence from Fig. 2 and the regressions suggest that this fall in aid has already begun. One caution is that this analysis does not take into account any possible strategic interactions among donors: in determining their own aid levels, donors might react to other donors' aid volumes. The sample of banking crises in our data includes a mixture of “isolated” crises that did not spread from one donor to another (Spain in 1977, Turkey in 1982, the US in 1988, Japan in 1997, Korea in 1997) and other crises that are “correlated” either regionally (Finland, Norway and Sweden in 1991) or globally (beginning with the US in 2007). Synchronization of crises could either exacerbate or dampen the effects on aid, depending on the strategic interactions among donor countries. Imagine that the standard approach of the aid community in the past has been for donors to compensate for each other in the case of economic shocks — that is, for one donor to increase its aid when another is forced to cut back in a recession. In that standard case, the effect on aggregate aid of a downturn in one country would be less than the effect on that country's own aid. But in a synchronized global recession like the one starting in 2007, there may be no donor-of-last-resort willing to step in and compensate for shortfalls elsewhere, so aid could eventually drop farther than predicted based on past crises. In fact, strategic behavior could deepen the drop in aid: a European Commission report warned that “governments might be more willing to reduce aid if they believe that their peers are doing the same, thus expecting only a collective blame rather than becoming exposed individually” (European Commission, 2009). We might expect this type of behavior if, as argued by Hillman (2010) and Potrafke and Ursprung (2013), development aid is a form of “expressive giving”. No individual donor will reduce aid unilaterally, Hillman suggests, because doing so would express that donor's identity as a country that does not “care about poor people in poor countries” (Hillman, 2010); if this is the case, the crisis might provide a coordinating mechanism or focal point allowing cutbacks by many donors. Alternatively, if donor countries become concerned about the dangers of simultaneous cutbacks in aid across many countries, they may be able to coordinate to maintain aid above the levels that would otherwise emerge from their unilateral decision making. In that case, aggregate aid would not drop by as much as empirical evidence from the past would lead us to expect. We cannot fully answer which of these models of strategic interaction applies in practice, but the data can provide some suggestive answers. A first approach to this issue is to examine how donors in general respond to changes in aid volumes of other donors. Two previous studies found that a donor's aid volumes responded positively to changes in aid by other donors. Mosley (1985) interpreted this finding as indicating that higher disbursements by other countries “encourage, or shame, donors to keep up with the Jones's [sic] within the DAC”. Round and Odedokun (2004) similarly interpret their positive coefficient on other donors' aid as “strong evidence to support the postulated ‘peer pressure’ effect”. If indeed strategic behavior operates in the way suggested by those results, then an exogenous decline in aid from banking-crisis countries could cause non-crisis countries to cut their aid too, unless donors manage to coordinate to avert this. However, in tests that include many more recent years of data, we find that the relationship between aid and other donors' (lagged) aid volumes is positive but insignificant. This evidence on how donors react in general to other donors' aid decisions — while suggestive — does not necessarily indicate how they may respond in the specific case of banking crises, including crises that simultaneously affect several donors. Are donors likely to coordinate to avert a steep drop in aid in the wake of the recent crises? One approach to answer this question would be to compare the effects of “isolated” and “correlated” banking crises. The only crisis in our data that is synchronized over a large number of donors is very recent and ongoing, so it is impossible to reliably test the differential effects of “isolated” and “correlated” crises. In Table 4 (equation 3), however, we provide a rough first test, classifying the Nordic crises starting in 1991 and the recent global crises as “correlated” crises and all others as “isolated.” For “isolated” crises, the quadratic specification for “years since crisis onset” produces coefficient estimates similar to those from our base specification in equation 1.3. Coefficients for the “correlated” crises variables are signed similarly, but are smaller in magnitude and insignificant. This result is particularly notable because the “correlated” crises include that of Finland — and as noted above, Roodman (2008) showed Finland to be the most dramatic instance of a drop in aid following onset of a banking crisis. These results are consistent with the conjecture that donors coordinate in an attempt to maintain aggregate aid levels when crises are synchronized. Findings from this test should be interpreted cautiously, however; results for the “correlated” crises depend heavily on the Nordic crises, because the number of years from crisis onset does not exceed 4 for any country during the later global financial crisis. Although it is likely that there was some strategic interaction on aid provision among the Nordic countries (where aid is more of a politically salient issue than in most donors), these three donors collectively accounted for only about 8% of total DAC aid in 1991. By contrast, the donors affected by the crises beginning in 2007 and 2008 accounted for more than two-thirds of total aid in those years. 7. Conclusions Using panel data from 24 donor countries between 1977 and 2010, results presented above show that banking crises in donor countries are associated with substantial declines in aid flows. Even after controlling for income and other factors, net ODA

248

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

disbursements from crisis-affected countries fall by an average of 28% or more (relative to the counterfactual) and bottom out only about 10 to 13 years after the onset of the banking crisis. One possible channel for this effect is governments' attempts to deal with the large fiscal debt overhangs that typically result from banking crises. Our estimates suggest a larger effect on non-social-sector aid than on aid for health, education, and humanitarian purposes, perhaps because of greater voter willingness to support these more immediate needs in times of fiscal stringency. Our results also confirm that donor-country per capita income is strongly related to aid flows, with an elasticity of about 3. This income effect could reduce aid further, even if donor countries continue to recover in the medium term, because potential GDP is estimated to have dropped as well as a result of the crisis. What is the relevance of these findings to the current situation? If aid flows follow historical patterns, the global financial crisis could continue to reduce aid significantly over the next decade (at least relative to counterfactual). All donor countries suffered from the recent global recession, which drove their incomes far below pre-crisis trends. And numerous donor countries, which together contribute around two-thirds of DAC aid, suffered full or borderline banking crises starting in 2007 or 2008. Historical patterns suggest that aid from these countries could end up a third or more below the pre-crisis counterfactual in the coming years, as they wrestle with the long-term fiscal and other consequences of banking-sector stresses. Whether this scenario plays out will depend in part on how donors interact strategically: do they respond to each others' fiscal troubles by taking the opportunity to reduce their own aid, by increasing aid to compensate for others' cuts, or by coordinating to sustain aid levels? The historical evidence from this and other studies suggests that either the first or third of these alternatives is most likely. The pattern this time around should become clear over the next five years, the period that our results indicate is when aid may be most likely to drop sharply in the absence of a coordinated response. Our findings also contribute to understanding how donors interact strategically in their voluntary provision of development aid as a global public good. Some aid is motivated by “private” interests of particular donor countries, but aid is a collective good for donors “to the extent that they all value the development of the less developed areas” (Olson and Zeckhauser, 1966). Results from our analysis and related studies as reported in Section 6 suggest that aid flows are generally consistent with a model of informal burden sharing or conditional cooperation. Donors reduce their aid when confronted by idiosyncratic shocks (in the form of a country-specific banking crisis) but not when faced by common shocks (in the form of multi-country crises). More generally, aid-giving appears to be positively correlated with other donors' giving, controlling for income and time trends. In contrast, there is little evidence in favor of the Olson–Zeckhauser hypothesis that larger donor countries will provide a disproportionately large amount of foreign aid due to greater incentives for free riding on the part of smaller donor countries.

Appendix Appendix Table 1 Donor countries studied, with years of banking crises. Source: Laeven and Valencia (2010). Donor Australia Austria Belgium Canada Denmark Finland France Germany Iceland Ireland Italy Japan Korea, Rep. of Luxembourg Netherlands New Zealand Norway Portugal Spain Sweden Switzerland Turkey United Kingdom United States Borderline crises are in italics.

Start-end dates 2008– 2008– 2008– 1991–1995 2008– 2008– 2008– 2008– 1997–2001 1997–1998 2008– 2008– 1991–1993 2008– 1977–1981, 2008– 1991–1995, 2008– 2008– 1982–1984, 2000–2001 2007– 1988–1988, 2007–

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

249

Appendix Table 2 Data sources. Data

Sources

Net and gross aid disbursements Social sector and other sector commitments Banking crises beginning, ending dates GDP per capita, PPP adj. (current USD) Total population Government expenditure as % of GDP Trade as % of GDP Budget deficit as % of GDP General government net debt as % of GDP Unemployment rate Inflation rate Real exchange rate Corruption index (0–6) Right-wing/left-wing party

DAC Table 2A; OECD DAC Creditor Reporting System (CRS); OECD Laeven and Valencia (2010) OECD World Development Indicators; World Bank (2011) OECD World Development Indicators; World Bank (2011) OECD OECD World Economic Outlook, IMF (2010) World Development Indicators; World Bank (2011) OECD International Country Risk Guide (PRS Group, 2010) Database of Political Institutions (Beck et al., 2001; updated 2010)

Appendix Table 3 Summary statistics.

Net disbursements (log, current mil. USD) Gross disbursements Net aid transfers Bilateral net disbursements Contributions to multilateral donors Social sector commitments Other (non-social sector) commitments No. years from start of banking crisis No. years from start of crisis squared GDP per capita (log, PPP adjusted) Population (log, millions) Log of total aid from other donors Inflation rate (%) Real exchange rate Trade intensity (% of GDP) Unemployment rate (% of labor force) Government spending (% of GDP) Surplus/deficit (% of GDP) Net debt (% of GDP) Leftist ruling party (0–1 dummy) Rightist ruling party (0–1 dummy) Corruption (0 worst, 6 best)

Mean

Std. dev.

Min.

Max.

6.69 6.75 6.72 6.19 5.53 5.14 6.90 2.58 43.23 9.83 16.52 10.75 5.63 99.14 74.26 6.80 44.962 −1.93 47.52 0.37 0.48 4.9

1.77 1.81 1.65 2.01 1.87 1.56 1.42 6.05 134.70 0.55 1.54 0.55 10.09 25.53 46.13 3.83 9.38 4.60 28.72 0.48 0.50 1.1

1.39 1.39 0.63 −4.61 −4.61 −1.51 3.21 0 0 8.28 12.46 9.09 −4.32 53.26 16.01 0.19 17.56 −16.01 0.82 0 0 2

10.27 10.30 10.26 10.15 8.63 8.85 10.01 33 1089 11.401 19.54 11.71 137.98 233.02 324.33 24.12 71.72 19.09 183.53 1 1 6

References Alesina, A., Dollar, D., 2000. Who gives foreign aid to whom and why? J. Econ. Growth 5 (1), 33–63. Arellano, M., Bond, S., 1991. Some tests of specification for panel data: Monte Carlo evidence and an application to employment equations. Rev. Econ. Stud. 58 (2), 277–297. Bazzi, S., Clemens, M., 2010. Blunt instruments: a cautionary note on establishing the causes of economic growth. Working Paper 171. Center for Global Development. Beck, T., Clarke, G., Groff, A., Keefer, P., Walsh, P., 2001 (updated 2001). New tools in comparative political economy: the database of political institutions. World Bank Econ. Rev. 15 (1), 165–176. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. J. Econometrics 87 (1), 115–143. Blundell, R., Bond, S., Windmeijer, F., Baltagi, B.H., 2000. Estimation in dynamic panel data models: improving on the performance of the standard GMM estimator. Nonstationary panels, panel cointegration, and dynamic panels. Advances in Econometrics, vol. 15. Elsevier Science, JAI, Amsterdam; New York and Tokyo, pp. 53–91. Bond, S., 2002. Dynamic panel data models: a guide to micro data methods and practice. Port. Econ. J. 1, 141–162. Boschini, A., Olofsgard, A., 2007. Foreign aid: an instrument for fighting communism? J. Dev. Stud. 43 (4), 622–648. Bruckner, M., 2011. On the simultaneity problem in the aid and growth debate. J. Appl. Econ. 28 (1), 126–150. Bruno, G.S.F., 2005a. Approximating the bias of the LSDV estimator for dynamic unbalanced panel data models. Econ. Lett. 87 (3), 361–366. Bruno, G.S.F., 2005b. Estimation and inference in dynamic unbalanced panel-data models with a small number of individuals. Stata J. 5 (4), 473–500. Caprio, G., Klingebiel, D., Laeven, L., Noguera, G., 2005. Banking Crisis Database. Systemic Financial Crises: Containment and Resolution. Cambridge University Press, Cambridge and New York 307–340. Chong, A., Gradstein, M., 2008. What determines foreign aid? The donors' perspective. J. Dev. Econ. 87 (1), 1–13. Clemens, M., Radelet, S., Bhavnani, R., Bazzi, S., 2012. Counting chickens when they hatch: the short-term effect of aid on growth. Econ. J. 122 (561), 590–617. Cunningham, G., 2009. Geldof understands cuts to aid budget. Ir. Indep. (Oct. 24 2009). http://www.independent.ie/irish-news/geldof-understands-cuts-to-aidbudget-26575958.html. Demirguc-Kunt, A., Detragiache, E., 2005. Cross-country empirical studies of systemic bank distress: a survey. Natl. Inst. Econ. Rev. 192, 68–83.

250

H.-A. Dang et al. / European Journal of Political Economy 32 (2013) 232–250

Devarajan, S., 2008. Will the financial crisis reduce foreign aid? Africa Can… End Poverty (blog) October 22. Dollar, D., Levin, V., 2006. The increasing selectivity of foreign aid, 1984–2003. World Dev. 34 (12), 2034–2046. Doucouliagos, H., Paldam, M., 2008. Aid effectiveness on growth: a meta study. Eur. J. Polit. Econ. 24 (1), 1–24. European Commission, 2009. Economic cycles and development aid: what is the evidence from the past? ECFIN Economic Brief. Faini, R., 2006. Foreign aid and fiscal policy. CEPR Discussion Papers, 5721. Frot, E., 2009. The Consequences of Financial Crises on Aid. Stockholm School of Economics, Stockholm. Frot, E., Perrotta, M., 2012. Aid effectiveness: new instrument, new results? SITE Working Paper 11. Stockholm Institute of Transition Economics, Stockholm School of Economics. Hillman, A.L., 2010. Expressive behavior in economics and politics. Eur. J. Polit. Econ. 26 (4), 403–418. International Monetary Fund, 2010. World Economic Outlook: Recovery, Risk, and Rebalancing. International Monetary Fund, Washington, DC. International Monetary Fund, 2011. World Economic Outlook: Slowing Growth, Rising Risks. International Monetary Fund, Washington, DC. Laeven, L., Valencia, F., 2008. Systemic banking crises: a new database. IMF Working Paper WP/08/224. Laeven, L., Valencia, F., 2010. Resolution of banking crises: the good, the bad, and the ugly. IMF Working Papers, 10/146. International Monetary Fund (35 pages). Lancaster, C., 2006. Foreign Aid: Development, Diplomacy, Domestic Politics. University of Chicago Press, Chicago. Martens, B., Mummert, U., Murrell, P., Seabright, P., 2002. The Institutional Economics of Foreign Aid. With a Foreword by Elinor Ostrom. Cambridge University Press, Cambridge; New York and Melbourne. Mendoza, R., Jones, R., Vergara, G., 2009. Will the global financial crisis lead to lower foreign aid? A first look at United States ODA. Fordham Economics Discussion Paper Series. Fordham University, Department of Economics. Mosley, P., 1985. The political economy of foreign aid: a model of the market for a public good. Econ. Dev. Cult. Chang. 33 (2), 373–393. Nielsen, R.A., Findley, M.G., Davis, Z.S., Nielson, D.L., 2011. Foreign aid shocks as a cause of violent armed conflict. Am. J. Polit. Sci. 55 (2), 219–232. OECD, 1996. Development Co-operation Review Sweden. OECD, 2010. Development Aid Rose in 2009 and Most Donors Will Meet 2010 Aid Targets. from http://www.oecd.org/document/51/0,3343,en_2649_34447_ 44981579_1_1_1_1,00.html. OECD, 2011a. Development Aid Reaches an Historic High in 2010. from http://www.oecd.org/document/35/0,3746,en_2649_34447_47515235_1_1_1_1,00.html. OECD, 2011b. OECD StatExtract. stats.oecd.org/Index.aspx. OECD, 2012. Development: Aid to Developing Countries Falls Because of Global Recession. from http://www.oecd.org/newsroom/developmentaidto developingcountriesfallsbecauseofglobalrecession.htm. Olson, M., Zeckhauser, R., 1966. An economic theory of alliances. Rev. Econ. Stat. 48 (3), 266–279. Paxton, P., Knack, S., 2012. Individual and country-level factors affecting support for foreign aid. Int. Polit. Sci. Rev. 33, 171–192. Potrafke, N., Ursprung, H., 2013. Development Aid as Expressive Giving. University of Munich and ifo Institute. PRS Group, 2010. International Country Risk Guide. www.prsgroup.com. Reinhart, C.M., Rogoff, K.S., 2008. Banking crises: an equal opportunity menace. NBER Working Paper 14587. Roodman, D., 2006 and 2010. An index of donor performance. Center for Global Development Working Paper Number 67 (Data set updated 2010). Roodman, D., 2008. History Says Financial Crisis Will Suppress Aid (October 13, Views from the Center (Blog)). Roodman, D., 2009. A note on the theme of too many instruments. Oxf. Bull. Econ. Stat. 71 (1), 135–158. Round, J.I., Odedokun, M., 2004. Aid effort and its determinants. Int. Rev. Econ. Financ. 13 (3), 293–309. Watts, J., 2003. Cash-strapped Japan cuts international aid budget. Lancet 361 (9355), 406. Wooldridge, J.M., 2002. Econometric Analysis of Cross Section and Panel Data. MIT Press, Cambridge, MA. Wooldridge, J.M., 2003. Cluster-sample methods in applied econometrics. Am. Econ. Rev. 93 (2), 133–138. World Bank, 2011. World Development Indicators. data.worldbank.org/indicator. Young, J.K., Findley, M., 2011. Can peace be purchased? A sectoral-level analysis of aid's influence on transnational terrorism. Public Choice 149, 365–381.