American Economic Journal: Macroeconomics 2014, 6(4): 170–208 100191 http://dx.doi.org/10.1257/mac.6.4.170 Government Spending Multipliers in Developing Countries: † Evidence from Lending by Official Creditors  By Aart Kraay * I use a novel loan-level dataset covering lending by official creditors to developing country governments to construct an instrument for government spending. Loans from official creditors typically finance multiyear public spending projects, with disbursements linked to the stages of project implementation. The identification strategy exploits the long lags between approval and eventual disbursement of these loans to isolate a predetermined component of public spending asso- ciated with past loan approval decisions taken before the realiza- tion of contemporaneous shocks. In a large sample of 102 developing countries over the period 1970–2010, the one-year spending multi- plier is reasonably-precisely estimated to be around 0.4. (JEL E62, F34, F35, H50, O23) E mpirically estimating government spending multipliers requires the isolation of a source of variation in government spending that is likely to be uncorrelated with contemporaneous macroeconomic shocks. In this paper, I construct an instru- ment for fluctuations in government spending, drawing on a near-comprehensive dataset of about 60,000 individual loans from official creditors (primarily multi- lateral development banks and bilateral aid agencies) to developing country gov- ernments over the period 1970–2010, as recorded in the Debtor Reporting System database of the World Bank. My identification strategy exploits two key features of this data. First, for many developing country governments, loans from official creditors are a major source of financing for public spending. In my largest sample of 102 countries, disbursements on these loans account for 11 percent of govern- ment spending for the median country-year observation, while the seventy-fifth and ninetieth percentiles correspond to 19 and 28 percent, respectively. Second, rather than simply financing the difference between government expenditure and revenue in a given year, these loans typically are tied to specific multiyear spending projects, *  1818 H Street NW, Washington, DC 20433 (e-mail: akraay@worldbank.org). I am grateful to Alexandra Jarotschkin for outstanding research assistance, to Luis Serven, Carlos Végh, and seminar participants at the University of Maryland, the International Monetary Fund, and the International Growth Center at the London School of Economics, and to several anonymous referees for helpful comments, and to Ibrahim Levent, Nanasamudd Chhim, Evis Rucaj, and Shelley Lai Fu for their guidance with the Debtor Reporting System database. Financial support from the Knowledge for Change Program of the World Bank is gratefully acknowledged. The views expressed here are the author’s, and do not reflect those of the World Bank, its Executive Directors, or the countries they represent. †  Go to http://dx.doi.org/10.1257/mac.6.4.170 to visit the article page for additional materials and author disclosure statement(s) or to comment in the online discussion forum. 170 Vol. 6 No. 4 kraay: government spending multipliers in developing countries 171 and accordingly disburse over a period of several years following the original com- mitment of the loan, with disbursements linked to stages of project implementation. My core identifying assumption is that the decision to approve a loan in a given year, and to embark on the associated spending plans, is uncorrelated with shocks to growth occurring in subsequent years when the spending plans are implemented and the actual loan disbursements take place. If this identifying assumption holds, and if loan disbursements follow a schedule specified at the time of loan approval, then disbursements occurring in the years following loan approval will also be uncorre- lated with contemporaneous macroeconomic shocks. Moreover, the long disburse- ment profiles observed on these loans imply that disbursements occurring in the years following loan approval are substantial: for the average loan in my dataset, only 22 percent of the original commitment is disbursed in the year that the loan is initially approved, and only a further 18 and 13 percent are disbursed in the first and second years following the approval year. The remaining nearly 50 percent of the loan is disbursed three or more years after loan approval. This in turn implies that the bulk of disbursements on loans from official creditors in a given country-year reflects loan approval decisions made in many previous years, and—crucially— before contemporaneous macroeconomic shocks are known. An immediate concern with this strategy is that, even though loan approvals are by definition made prior to the realization of macroeconomic shocks that occur dur- ing the subsequent disbursement period, the size and timing of loan disbursements may be tailored in response to current events. For example, lenders might choose to accelerate disbursements on previously-approved loans to a country experiencing a natural disaster, as a way of rapidly delivering resources to the affected country. Or alternatively, lenders might suspend disbursements on existing loans in response to an adverse political event, such as the outbreak of civil conflict, that disrupts the implementation of the associated spending plans. In either case, this would under- mine my identification strategy by creating a correlation between contemporaneous macroeconomic events and actual disbursements on previously-approved loans. To circumvent this problem, I construct an artificial predicted disbursements series for each loan, based on the observed average disbursement rates for other loans from the same creditor approved in the same decade, to borrowers in the same geographical region. Conditional on my identifying assumption that loan approvals are indepen- dent of future macroeconomic shocks, these artificial loan-level predicted disburse- ments in the years following loan approval, and their aggregation to the country-year level, are by construction independent of contemporaneous c ­ ountry-specific macro- economic shocks. I am interested in estimating overall government spending multipliers, and not simply the short-run effects of public spending projects financed by official credi- tors. This distinction is important given potential concerns about the fungibility of the latter. Specifically, it is possible that increases in government spending associ- ated with official creditor-financed projects might very well lead to changes in the level and composition of other forms of government spending. To address this con- cern, I use fluctuations in predicted disbursements on previously-approved loans as an instrument for changes in total government spending. As a result, any responses of other forms of government spending to official creditor-finance spending will be 172 American Economic Journal: Macroeconomics october 2014 subsumed into the fi ­ rst-stage relationship between my instrument and changes in total government spending. I apply this methodology in a sample of 102 developing countries where loans from official creditors are an important source of financing for government ­spending. I find baseline estimates of the one-year government spending multiplier of around 0.4, i.e., a dollar of additional government spending raises GDP in the same year by about 40 cents. I subject these basic results to a battery of robustness checks designed to address potential concerns with data quality, as well as possible objec- tions to the identifying assumption. While the estimates of the one-year multiplier vary somewhat across these checks, they typically remain in a range from around 0.3 to 0.5. The large cross-sectional dimension of my dataset also makes it feasible to investigate the empirical relevance of a number of potential sources of heteroge- neity in multipliers. I find some suggestive evidence that multipliers are larger in recessions, in countries that are relatively less exposed to international trade, and in countries with flexible exchange rate regimes. This paper builds on my previous work in Kraay (2012), which exploited lags between the approval of and subsequent disbursements on individual World Bank-financed development projects to isolate a predetermined component of World ­ Bank-financed public spending that could be used as an instrument to estimate gov- ernment spending multipliers. Out of necessity, that paper focused on a small set of 29 mostly very poor countries where World Bank-financed spending is large relative to the size of the recipient economy, and over the period 1985–2009. In contrast, the loan-level Debtor Reporting System data used in this paper covers lending by virtu- ally all multilateral and bilateral official creditors to all developing countries, and extends back to 1970. The combined disbursements on loans from all these creditors account for a much larger fraction of public spending than World Bank financing alone. This substantially strengthens identification compared to the previous paper, and moreover permits extending the analysis to a much larger set of 102 developing countries where lending from official creditors is an important source of financing for public spending.1 This greater sample size in turn expands the relevance of the findings to a broader set of countries, and in addition makes it feasible to assess a variety of possible sources of heterogeneity in estimated multipliers, as is done in this paper. My strategy of exploiting delays between loan approval decisions and the ultimate spending that they finance is also related to Leduc and Wilson (2012), who study the dynamic effects of federal highway spending in the United States. The nature of the projects in question, and the institutional environment in which they are financed, also give rise to long delays between the authorization of f ­ederally-financed high- way spending, and the actual state-level spending itself. Their use of this lag struc- ture is different from mine, however, in that they estimate responses to “surprises” 1  In fact, restricting attention to the same set of 29 countries covered in Kraay (2012), the F-statistic from the first-stage regression of changes in government spending on changes in predicted disbursements increases from 15 in the previous paper, to 28 in this paper. Nevertheless, the point estimates of the multiplier I find using this much larger dataset are remarkably similar to those based on World Bank-financed spending alone. Using only World Bank project-level data to construct the instrument, I found a multiplier of 0.48 in the benchmark specification, while I obtain a much-more-precisely estimated multiplier of 0.61 using the DRS loan-level data in this paper. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 173 in spending, measured as deviations in spending from what would have been pre- dicted from past financing approvals. This strategy is appropriate in their context, given their emphasis on isolating responses to unanticipated spending shocks. In contrast, in my context I am concerned that such deviations of spending from pre- dictions based on past approvals are likely to be endogenous responses to mac- roeconomic shocks in the borrowing country. For example, as discussed above, an official creditor might suspend disbursements on previously-approved loans in response to negative shocks in the borrowing country that disrupt the implementa- tion of the projects being financed by the loans. In this case, unexpectedly low dis- bursements and associated government spending (relative to initial plans) would be an endogenous response to the contemporaneous negative shocks. Instead, I rely on predicted disbursements on previously-approved loans as a strategy for excluding this potentially-endogenous component of fluctuations in actual disbursements and government spending. This paper contributes to a rapidly-expanding empirical literature on identifying the short-run output effects of government spending, nearly all of which is focused on developed countries, most notably the United States. One strand of this literature has followed the seminal contribution of Barro (1981), who observed that fluctua- tions in defense spending are an important source of fluctuations in total government spending in the United States, and are driven primarily by geopolitical factors rather than domestic macroeconomic shocks. As a result, they can be viewed as a plausibly exogenous source of variation in government spending that can be used to estimate spending multipliers.2 Another strand of this literature has followed Blanchard and Perotti (2002) in assuming that discretionary fiscal policy changes take sufficiently long to implement that they cannot respond to macroeconomic shocks during the same quarter.3 This assumption permits the identification of VAR-based estimates of spending multipliers in those countries where high-frequency macroeconomic and fiscal data are available.4 2  Other papers extending this basic insight include Ramey and Shapiro (1998); Hall (2009); Fisher and Peters (2010); Ramey (2011b); and Barro and Redlick (2011). A shared drawback of these military spending-based stud- ies is that it is difficult to control for the macroeconomic effects of other key features of wartime economies, such as price controls or mandatory military service. Moreover, this approach to identification is applicable only to the United States, where the conflicts associated with the spending increases occurred outside the United States, so that there were no direct effects of wartime destruction on the US economy. Nakamura and Steinsson (2014) also focus on military spending in the United States, but exploit cross-state variation in the intensity of defense spending. This approach is shared with Giavazzi and McMahon (2012) who investigate heterogeneity across households in the response to these defense spending shocks. These papers are based on a weaker identifying assumption that military spending buildups are unrelated to differences in macroeconomic conditions across US states. 3  Notable recent contributions along these lines include Auerbach and Gorodnichenko (2012a, 2012b) and Ilzetzki, Mendoza, and Végh (2013). These studies also examine heterogeneity in multipliers, with the former emphasizing the state of the business cycle, and the latter a range of factors such as the exchange rate regime and trade openness, as I do in this paper as well. 4  A key practical difficulty with implementing this approach in developing countries is the relative scar- city of high-frequency data in these countries. In a notable effort to fill this gap, Ilzetzki, Mendoza, and Végh (2013) assemble quarterly data for a sample of 20 developed and 24 emerging markets, and use this to implement ­Blanchard-Perotti (2002)-style estimates of spending multipliers. There are however only 11 countries in common between their emerging-market sample and the sample of 102 low- and middle-income countries used in this paper. The countries included in their 24 country sample but not mine are primarily richer emerging-market economies that rely little on borrowing from official creditors. Conversely, my paper covers 91 developing countries not cov- ered in Ilzetzki, Mendoza, and Végh (2010), where this source of financing of government spending is important. 174 American Economic Journal: Macroeconomics october 2014 A third strand of the literature has proposed a variety of creative instruments to isolate a plausibly exogenous component of changes in government spend- primarily focusing on subnational government spending in the United States. ing, ­ Papers such as Cohen, Covall, and Malloy (2010) and Fishback and Kachanovskaya (2010) have exploited political determinants of federal transfers to states, while Chodorow-Reich et al. (2012); Serrato and Wingender (2011); and Wilson (2012) ­ emphasize particular institutional features driving federal-state transfers that are likely to be orthogonal to state-level economic activity. Clemens and Miran (2012) and Shoag (2010) study fluctuations in state-level spending driven by variations in the stringency of balanced-budget rules, and pension fund windfalls, respectively. Finally, three important caveats about these results are worth noting. The first is that the empirical spending multipliers I estimate are by no means deep structural parameters. As is well-known from a large body of theoretical work, the short-run effects of government spending on output depend on a host of factors including technology, preferences, the nature of spending, the associated burden of current and future taxes, the stance of monetary policy, and a range of other country- and episode-specific characteristics. In light of this, the multipliers that I estimate are best understood as a description of the short-run average empirical relationship between changes in a plausibly predetermined component of government spending and changes in output. This caveat is of course shared with the bulk of the existing empirical literature on the short-run effects of government spending. The second caveat is that, in contrast with much of the empirical literature on esti- mating government spending multipliers in advanced economies, this paper does not seek to estimate the response of output to unanticipated increases in government spending. To the contrary, I use variation in government spending that is driven by past loan approvals, which are public information, as are the associated likely disburse- ment profiles. While this variation is arguably predetermined with respect to contem- poraneous shocks to output, my estimated multipliers will not pick up any response of output to the announcements of subsequent spending plans that occur when loans from official creditors are approved. I return to this issue in Section IV of the paper. The final caveat is that, while my identification strategy relies on lending by offi- cial creditors, which frequently is a vehicle for foreign aid delivery, the resulting evidence should not be interpreted as contributing to the long-standing empirical debate on the growth effects of aid. In contrast with the aid-growth literature, which has primarily been concerned with the medium- to long-run effects of aid on growth, my interest in this paper is in the short-run cyclical effects of increased government spending on output. These short-run responses of output to government spending are potentially consistent with a variety of longer-run responses of output to aid. The rest of this paper proceeds as follows. Section I sets out the empirical meth- odology and identification strategy, and Section II describes the construction of the instrument based on loan-level data from the Debtor Reporting System database. Section III contains my core results, and Section IV subjects them to a variety of robustness checks designed to explore the plausibility of the identifying assumption. Section V provides some evidence on the longer-run effects of government spending on output, while Section VI investigates a number of potential sources of heteroge- neity in multipliers. Section VII concludes. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 175 I.  Empirical Strategy I estimate variants on this simple empirical specification to assess the short-run effects of government spending on output: y​  − ​ yi​t−1​ ​ g​  − ​ gi​t−1​ (1) ​ _ =  β ​ _ ​ it​ it​ ​ yi​t−1 ​  ​     yi​t−1 ​  ​ ​    + ​ μ​   + ​ i​   + ​ λt​​ ​​  . εit Here, ​ y​ gi​t​denote GDP and total government spending in country i and year t, it​and ​ both measured in constant local currency units, and the composite error term ​ μ​ + ​ i​ λ​t   ​ + ε  ​ ​ it​ denotes all other sources of GDP fluctuations, such as other fiscal or mone- tary policy changes, terms of trade shocks, changes in productivity, natural disasters, and many other shocks. For terminological convenience, I refer to the aggregate of these as “macroeconomic shocks.” I sweep out the country-specific and ­year-specific components of the error term, μ ​​ ​ t​​  , by including a full set of country and year i​and λ effects in all specifications. The key parameter of interest is β, which captures the short-run government spending multiplier, i.e., the contemporaneous change in out- put due to a change in government spending. As noted in the introduction, β cannot be interpreted as a deep structural parameter. Rather, it should simply be thought of as a reduced-form empirical summary of the contemporaneous relationship between annual fluctuations in government spending and output. The standard difficulty in statistically identifying β is that changes in government spending are likely to be correlated with other contemporaneous shocks to output captured in the error term, so that OLS estimation of equation (1) will be incon- sistent. For example, if government spending increases endogenously in response to an economic downturn, perhaps due to the role of automatic stabilizers, then OLS estimates of the multiplier would be biased downwards. On the other hand, if government spending is procyclical and falls with the realization of negative mac- roeconomic shocks, perhaps due to an inability of governments to borrow during bad times, then OLS estimates of the multiplier would be biased upwards. A further possibility is that, in aid-dependent countries such as many of those studied here, any procyclical tendencies in domestically-financed government spending are offset by countercyclical tendencies in aid-financed government spending, so that total spending could be either procyclical or countercyclical. I address this endogeneity problem by constructing an instrument based on the lags between commitments and eventual disbursements on loans by official credi- tors to developing country governments. Some institutional background is helpful in order to better understand this identification strategy. My dataset, described in more detail in Section II, covers loans from multilateral and bilateral official credi- tors to developing country sovereign borrowers. Table 1 provides some summary statistics on the lending activities of official creditors included in my dataset. I first report total disbursements, disaggregated by major multilateral and bilateral credi- tors, for the decades of the 1970s, 1980s, 1990s, and 2000s. Disbursements on these loans are substantial, totaling nearly $1.8 trillion in constant 2005 prices over the past 40 years. Over time, the importance of multilateral creditors relative to bilateral creditors has increased substantially, with the share of the former increasing from 176 American Economic Journal: Macroeconomics october 2014 Table 1—Summary Statistics on Lending by Official Creditors 1970–1979 1980–1989 1990–1999 2000–2010 Disbursements by multilateral creditors (billions constant 2005 $US) African Development Bank 0.9 4.9 12.6 8.9 African Development Fund 0.3 3.5 6.4 7.4 Asian Development Bank 2.4 8.7 26.0 37.0 Asian Development Fund 1.2 5.4 11.9 13.5 European Investment Bank 1.5 4.2 11.6 21.7 International Bank for Reconstruction and Development 37.9 112.0 135.0 131.0 International Development Association 19.2 41.2 59.3 68.0 Interamerican Development Bank 6.8 20.6 41.9 62.7 Interamerican Development Fund 2.9 4.7 3.9 3.8 Other multilaterals 5.5 25.9 49.6 65.1 Total multilateral disbursements 78.4 231.1 358.2 419.1 Disbursements by bilateral creditors (billions constant 2005 $US) Canada 10.2 11.0 6.2 1.9 China 4.8 2.8 3.3 26.6 France 7.7 14.5 15.5 8.2 Germany 18.5 23.0 32.5 9.6 Italy 1.7 4.8 5.6 1.6 Japan 20.5 53.1 105.0 72.7 Kuwait 5.8 6.8 4.1 4.5 Russia 0.0 0.2 2.5 5.6 Saudi Arabia 7.2 9.8 1.6 1.8 United Kingdom 6.2 2.7 1.3 1.3 United States 45.9 48.6 23.9 4.0 USSR 8.0 18.8 1.1 0.0 Other DAC bilaterals 8.9 17.0 16.8 10.0 Other non-DAC bilaterals 13.8 16.8 6.5 9.6 Total bilateral disbursements 159.2 230.0 225.9 157.3 Total disbursements all creditors 237.6 461.0 584.2 576.4 Loan-weighted average terms of new commitments Interest spread over 20-year US Treasury Bill rate (percent) −2.6 −4.7 −2.0 −2.1 Grace period (years) 6.9 5.9 6.0 6.4 Maturity (years) 25.2 22.9 21.4 22.6 Notes: This table provides summary statistics on disbursements on loans from official creditors to developing coun- tries captured in the Debtor Reporting Statistics (DRS) database of the World Bank. Disbursements are measured in billions constant 2005 $US. “Other multilaterals” includes all other multilateral creditors with loans recorded in DRS. “Other DAC bilaterals” and “Other non-DAC bilaterals” include all other bilateral creditors with loans recorded in DRS. DAC refers to membership in the OECD’s Development Assistance Committee as of 2012. Data on terms in the last three lines are calculated on a loan-weighted basis, weighting by initial loan commitment. The interest spread refers to the nominal interest rate on the loan less the 20-year US T-Bill rate, except for the period 1987–1992 when no 20-year T-Bills were issued: during this period the 30-year rate is used instead. about one-third in the 1970s to nearly three-quarters in the 2000s. This reflects a steady shift on the part of most bilateral creditors over the past 40 years to providing aid in the form of grants (which are not reflected in the Debtor Reporting System database), rather than loans (which are).5 5  For my purposes in this paper, the main consequence of this shift to grants by bilateral creditors is that it weak- ens the first-stage relationship between changes in total government spending and changes in predicted disburse- ments on loans from official creditors, relative to what would have been the case if all aid were provided in the form of loans. However, since in most specifications below, the first-stage relationship is sufficiently strong as to permit meaningful inferences, this is not a serious concern. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 177 These loans are a traditional vehicle for aid donors to provide financial assis- tance to developing country governments. Consistent with this objective, the loans in this dataset are highly concessional on average, typically with long grace and repayment periods, as well as below-market interest rates. The bottom panel of Table 1 shows that, on a loan-weighted average basis, these loans have a matu- rity between 20 and 25 years, and an initial grace period (during which no pay- ment is required) of approximately six years. The interest rates on these loans are also highly ­concessional, with nominal spreads over 20-year US Treasury Bill rates between approximately −2 percent and −5 percent. These simple spreads probably understate the concessional value of the loans to many recipient countries, given that the market rates they would otherwise face on international borrowing from private creditors are likely to be much higher, if not prohibitive. A key feature of these loans is that they typically are tied to specific public spend- ing projects identified by the donor and the recipient government.6 These projects might consist of infrastructure construction, health and education initiatives, public sector reform efforts, or any other of a wide variety of development projects sup- ported by aid donors. Crucially for my purposes, such projects often take several years to implement, and the loan disbursements typically are tied to various stages in the implementation of the project that they are intended to finance. As a result, disbursements on the original loan commitment usually are spread out over multiple years following loan approval, rather than the loan disbursing in full at the time of loan approval. These long disbursement profiles in turn imply that, in any given country-year, aggregate disbursements on loans from official creditors consist pri- marily of disbursements on loans approved in several previous years, rather than the current year. To construct my instrument, for each country-year I first isolate disbursements on loans approved in previous years, but not the current year. In order for this to be a valid instrument for fluctuations in total government spending, it must be the case that (i) loan approval decisions do not anticipate future macroeconomic shocks, and (ii) disbursements on previously-approved loans also do not respond to contempora- neous macroeconomic shocks. While (i) is plausible given the timing of events, with project and loan approvals occurring before the realization of future macroeconomic shocks, (ii) is much less plausible because the decision to disburse a portion of a loan is made in real time, and may very well respond to contemporaneous shocks. For example, creditors may suspend disbursements on previously-committed loans to a country falling into a civil conflict that disrupts the implementation of the asso- ciated projects. Conversely, creditors might choose to accelerate disbursements on previously-committed loans as a way of quickly delivering additional resources to a country experiencing an adverse shock. Either of these possibilities would lead to 6  For example, World Bank loan agreements typically contain several pages of text describing the specific proj- ect the loan is intended to finance, conditions for monitoring the implementation of the project, and guidelines for procurement and disbursement. These loan agreements also contain a standard clause specifically committing the borrower to the project, along the lines of “The Borrower declares its commitment to the objectives of the Project and the Program. To this end, the Borrower shall carry out the Project  …  in accordance with the provisions of Article V of the General Conditions.” This language is found in Article III of standard World Bank loan agreements. 178 American Economic Journal: Macroeconomics october 2014 a correlation between actual disbursements on previously-approved loans and con- temporaneous macroeconomic shocks. In order to circumvent this problem, I replace actual disbursements on previously- approved loans with predicted disbursements, based on typical disbursement pro- files for similar loans. In particular, I construct loan-level predicted disbursements by applying to each initial loan commitment the average disbursement profile across all other loans issued by the same creditor in the same decade to all countries in the same geographical region as the actual borrower. I then construct my instrument by aggregating these predicted loan-level disbursements on previously-approved loans to the country-year level. By construction, aggregate predicted disbursements reflect only the combination of country-specific loan approval decisions from previous years with typical disbursement profiles, based on averages taken across many loans to many countries. My identifying assumption is that these loan approval decisions do not anticipate future shocks to growth, and under this assumption, changes in aggregate predicted disbursements will be uncorrelated with the error term in equa- tion (1). I can therefore use changes in predicted disbursements as an instrument for changes in total government spending when estimating the government spending multiplier based on equation (1). II. Data I work with loan-level data drawn from the Debtor Reporting System (DRS) database maintained by the World Bank. The DRS database contains information on loan commitments, terms, disbursements, and repayments, for all external loans contracted or guaranteed by the government in the borrowing country, beginning in 1970. The DRS data are, in principle, comprehensive in their coverage of all individ- ual external public and publicly-guaranteed debt obligations, from all creditors, and for all countries that borrow from the World Bank. This is because annual reporting to DRS is mandatory for World Bank clients: a country must be in good standing with respect to these reporting requirements in order for new projects for that coun- try to be considered by the Board of Directors of the World Bank.7 Countries are required to report basic information on the amount, terms, and purpose of new com- mitments, drawings and repayments on existing loans, and details of loan restruc- turings when applicable.8 Loan-level transactions reported in DRS are confidential. However, the aggregation of this loan-level data to the country-year level provides the basis for the country-level debt data published by the World Bank in its annual World Development Indicators and Global Development Finance publications.9 7  See the World Bank’s Operational Manual, BP14.10, available at: http://web.worldbank.org/WBSITE/ EXTERNAL/PROJECTS/EXTPOLICIES/EXTOPMANUAL/0,,contentMDK:20064540~menuPK:4564187~ pagePK:64709096~piPK:64709108~theSitePK:502184,00.html. 8  Details on reporting requirements can be found in the World Bank Debtor Reporting System Manual, available at http://siteresources.worldbank.org/DATASTATISTICS/Resources/drs_manual.doc. These loan-level transactions are typically provided as paper records or in spreadsheet format, and staff in the Development Data Group of the World Bank enter it manually into DRS. 9  To my knowledge, the loan-level data in DRS has been used only in a handful of previous scholarly papers, all of which are focused in one way or another on alternative characterizations of the financial value of concessional loans (Chang, Fernandez-Arias, and Serven 2002; Dikhanov 2004; and Dias, Richmond, and Wright 2011). Vol. 6 No. 4 kraay: government spending multipliers in developing countries 179 I rely on commitment and disbursement transactions on loans extended by offi- cial creditors to developing countries, as recorded in the DRS database. Official creditors include a range of multilateral lenders such as the World Bank, the African Development Bank, the Asian Development Bank, and the European Investment Bank. My dataset also includes loans issued by bilateral official creditors. The majority of these are large OECD aid donors such as Japan, Germany, and the United States, but the dataset also includes a number of non-OECD creditors such as Kuwait, Saudi Arabia, Russia, and the former Soviet Union (in the earlier half of my sample).10 The dataset, retrieved from DRS in January 2012, contains 60,192 loans issued by 188 distinct creditor countries and organizations. A large number of credi- tors represented in DRS account for only a handful of loans each, and usually for very small amounts. I discard a total of 768 loans issued by 113 creditors who have fewer than 50 loans each in the DRS data (and on average fewer than seven loans each), leaving a total of 59,424 loans issued by 75 distinct major official creditors. Loan commitments are reported in the currency of origination of the loan, and sub- sequent disbursements are recorded in DRS in current US dollars. I discard a further 36 loans for which data on the exchange rate used to convert the disbursements denominated in the currency of origination into US dollars could not be retrieved from DRS. This reduces the sample further to 59,388 loans. The first step in the development of my instrument is to construct a disburse- ment profile for each loan, i.e., the fraction of the original loan commitment that is disbursed in the commitment year and each subsequent year. I calculate these by converting the US dollar disbursements on each loan back to the currency of denomination of the original commitment, using the corresponding disbursement- year exchange rates, and then express this as a fraction of the original commitment. Roughly 10 percent of loans have accumulated disbursements greater than initial commitments. This typically reflects increases in the loan amount that occur at some point during the disbursement period, but that are not recorded in the original com- mitment. Because these revisions in loan size are potentially endogenous responses to contemporaneous shocks, I use only the original commitment and not the sub- sequent increases in the denominator of the disbursement rate. For a small num- ber of loans, reported total disbursements exceed initial commitments by several multiples. To avoid data entry errors that may be responsible for such implausibly high disbursements relative to original commitments, I drop a further 194 loans for which accumulated disbursements are more than five times the initial commitment amount.11 As a final step, I discard ten loans that are implausibly large relative to 10  Lending by official creditors to governments is formally defined as follows “Public and publicly guaranteed debt from official creditors includes loans from international organizations (multilateral loans) and loans from gov- ernments (bilateral loans). Loans from international organization include loans and credits from the World Bank, regional development banks, and other multilateral and intergovernmental agencies. Excluded are loans from funds administered by an international organization on behalf of a single donor government; these are classified as loans from governments. Government loans include loans from governments and their agencies (including central banks), loans from autonomous bodies, and direct loans from official export credit agencies” (data.worldbank.org). From this total, I exclude IMF credits, as these typically take the form of budget support as opposed to financing specific projects, and typically also are approved for strongly cyclical reasons, i.e., in response to macroeconomic crises in the borrowing countries. 11  Anecdotally, a potential explanation for this pattern is that occasionally loans recorded in DRS take the form of revolving credit lines that can be drawn upon and paid down multiple times. In this case the maximum loan 180 American Economic Journal: Macroeconomics october 2014 recipient country GDP but have very low disbursement rates, again because these possibly reflect data entry errors. This results in a sample of 59,184 loans on which my results are based.12 As noted earlier, a key feature of these loans from official creditors is that disburse- ments are typically spread out over several years following the loan ­ commitment. This is apparent from Figure 1, which reports typical disbursement profiles, i.e., the fraction of the initial loan commitment that is disbursed in year zero (i.e., the year the loan was approved) and the ten subsequent years. The top panel reports the simple average disbursement profile, averaging across all loans in my dataset, while the bottom panel reports disbursement profiles separately for loans issued by bilateral and multilateral official creditors. Taking all loans together, on average only 22 percent of the initial loan commitment is disbursed in the year the loan is approved, and only another 18 percent in the next year, with the remaining 60 percent spread out over subsequent years. Disbursement profiles are even more strongly backloaded for multilateral creditors than for bilaterals. The average mul- tilateral loan disburses only 13 percent of the original commitment in the approval year, while for the average bilateral loan the figure is 29 percent. These long dis- bursement profiles in turn imply that actual aggregate disbursements on loans from official creditors in a given country-year are largely associated with past loan com- mitments, and not with loans approved in the current year. In the median country- year observation in my full sample of 102 countries over the period 1970–2010, 89 percent of disbursements are associated with loans approved in previous years. For the twenty-fifth and s ­ eventy-fifth percentiles, the corresponding figures are 72 percent and 99 percent. Nevertheless, as discussed in the previous section, even these disbursements on previously-approved loans might still be endogenous responses to contem- poraneous shocks, to the extent that the disbursement decision in the current year reflects current shocks rather than being predetermined at the time of loan approval. To address this difficulty, I rely on predicted, rather actual, disburse- ments. Predicted disbursements are based on the interaction of actual loan com- mitments with typical disbursement profiles such as those shown in Figure 1, but for more fi ­ nely-disaggregated groups of loans. Specifically, I begin by assigning loans to a set of creditor/decade/region-specific bins. The creditor bins are based on the major creditors listed in Table 1, as well as the residual categories of other multilateral and bilateral creditors. I then separate loans issued by each of these creditor groups into decades by approval year, and further divide them into six geographical regions in which the borrower is located.13 This procedure results in amount is recorded as the loan commitment, and subsequent multiple drawings are recorded as disbursements and can easily exceed the recorded commitment amount. Unfortunately, however, the DRS database does not systemati- cally identify such credit lines. 12  This strategy is conservative in the sense that if I exclude valid loans using these criteria, this will only weaken my first-stage relationship between changes in predicted disbursements and changes in government spending. 13  The geographical regions are sub-Saharan Africa, Middle East and North Africa, South Asia, East Asia and the Pacific, Europe and Central Asia, and Latin America and the Caribbean. For the regional development banks listed in Table 1, I omit the geographical disaggregation. Also, four major creditors (Saudi Arabia, United Kingdom, the USSR, and Russia) have only a fairly small number of loans in each region/decade bin. To avoid over-fitting my predicted disbursements measure for these creditors, I also omit the geographical disaggregation and compute typical disbursement profiles only by decade for these creditors. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 181 All creditors 0.25 All creditors commitment disbursed 0.2 Fraction of original 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 Years since loan commitment Bilateral versus multilateral creditors 0.35 0.3 Multilaterals commitment disbursed Bilaterals 0.25 Fraction of original 0.2 0.15 0.1 0.05 0 0 1 2 3 4 5 6 7 8 9 10 Years since loan commitment Figure 1. Average Disbursement Profiles Notes: This figure reports average disbursement profiles on loans from official creditors to developing country government, defined as the fraction of the initial loan commitment dis- bursed in the year of loan approval (year zero) and subsequent years (years one through ten), averaging across loans. The top panel reports averages across all 59,184 loans included in my loan-level dataset. The bottom panel reports average disbursement profiles separately for loans from multilateral and bilateral creditors. 443 creditor/decade/region bins, with a median of 65 loans in each bin. For each loan, I compute the average ­disbursement profile across all other loans within the same creditor/decade/region bin, i.e., excluding the loan in question. I then apply this average disbursement profile to the original commitment to obtain a series of predicted loan-level disbursements. Finally, I aggregate predicted disbursements across all loans to the country-year level, but excluding loans committed in the ­ orrower-specific information in this measure same year. By construction, the only b of aggregate predicted disbursements consists of the original loan commitment 182 American Economic Journal: Macroeconomics october 2014 0.06 Current year approvals Previous years’ approvals 0.04 Predicted 0.02 0 1970 1980 1990 2000 2010 Year Figure 2. Disbursements on Loans from Official Creditors—Kenya Notes: This graph reports annual disbursements on loans from official creditors to Kenya. The overall height of the bars shows total disbursements, and the light (dark) shaded por- tions separate this into disbursements on loans approved in the current year (past years). The solid line reports predicted disbursements on loans approved in previous years, as described in Section II. decisions made in p ­ revious years, which I assume are uncorrelated with contempo- raneous macroeconomic shocks. Figure 2 helps to visualize the steps in the construction of my instrument based on predicted disbursements on previously-approved loans, using data for Kenya. Kenya is a fairly typical country in my sample, in that financing from official credi- tors accounts for 3 percent of GDP and 15 percent of government spending on aver- age over the period 1970–2010. The overall height of the bars in the graph display disbursements on loans from official creditors as a fraction of GDP. These vary considerably from year to year, ranging from lows around 1 percent of GDP to highs around 6 percent of GDP. The dark-shaded lower portion of each bar corresponds to the portion of disbursements in each year that is associated with loans approved in previous years, but not the current year. In most years, these disbursements on previously-approved loans account for a sizeable majority of total disbursements. However, during a few years there are substantial disbursements on loans approved in the same year. Finally, the solid line shows predicted disbursements on previ- ously-approved loans, which reflects the combination of country-specific loan approval decisions from previous years with “typical” disbursement rates. I will use changes in this predicted disbursement series as an instrument for changes in total government spending. The second major data requirement for this paper is data on government spending itself. My primary source for this is the total government expenditure series reported in the IMF’s World Economic Outlook (WEO) database. While this is by far the most comprehensive single data source for government spending, its ­ country-year coverage is nevertheless limited, particularly for low-income ­ countries. Data on Vol. 6 No. 4 kraay: government spending multipliers in developing countries 183 ­ fficial creditor lending from DRS is available for 2,804 c o ­ ountry-years over the period 1970–2010 in my full regression sample. However, WEO data on ­government spending cover only 1,732, or 62 percent, of these observations, and mostly cover the 1990s and 2000s. To fill this gap, I substantially expand coverage of the govern- ment spending data by piecing together additional information from a variety of other published sources for earlier years, including current and previous editions of the IMF’s Government Finance Statistics, the African Development Indicators of the World Bank, and data on total government spending available in the dataset on health and education spending compiled by Clements, Gupta, and Nozaki (2011).14 The success of my identification strategy requires a strong correlation between fluctuations in government spending and fluctuations in predicted disbursements on loans from official creditors. This is unlikely to be the case in countries that do not rely significantly on official creditors as a source of financing for public spending. Accordingly, I restrict attention to those countries where disbursements on loans from official creditors are on average equal to at least 1 percent of GDP, averaging over the entire period 1970–2010. In addition, in order to have meaning- ful w­ ithin-country time series variation for each country, I further restrict the sample to those countries that have at least 15 years of data on government spending.15 This results in a core regression sample of 2,804 country-year observations covering 102 countries listed in Table 2, and averaging 28 annual observations per country. Averaging across countries, disbursements on loans from official creditors account for 13.4 percent of government spending, and range from a low of 3.1 percent in Latvia to a high of 37.6 percent in The Gambia. In the empirical work that follows, I will also consider two subsamples, cor- responding to (i) countries that are more reliant on official creditor financing, and (ii) countries that are poorer. I define the former as the set of 70 countries for which disbursements on loans from official creditors exceed 10 percent of gov- ernment spending (as opposed to 3 percent for the full sample), and the latter as a set of 60 countries that are currently eligible for concessional lending from the World Bank-administered International Development Association (IDA, indicated with asterisks in Table 2).16 In these two subsamples, disbursements on loans from 14  For each country, I begin with the available government spending data series in the WEO. I then augment this with data from one or more of the three supplementary sources noted in the main text, in order to obtain the longest possible time series for each country. When there is a difference in the level of two series in the years where they overlap of more than 1 percent, I shift the earlier series up or down to match with the WEO series in the first year where the WEO data are available. The implication of this is that the merged series preserves the information in the annual changes in the earlier series. This is appropriate for my purposes in this paper where the estimation also relies only on annual changes, and not levels, of the government spending series. Appendix Table A1 details the sources of government spending for each country in my sample. While the resulting merged dataset on govern- ment spending based on these various sources remains highly imperfect, it is important to keep in mind that the inevitable measurement error in government spending will not bias my estimates of the spending multiplier as long as it (plausibly) is uncorrelated with fluctuations in my instrument based on past loan approval decisions. If this is the case, the only consequence of measurement error in government spending is to reduce the strength of my first- stage relationship between changes in spending and changes in predicted disbursements, and accordingly also the precision of my 2SLS estimates of the multiplier. 15  In addition, to prevent a relatively small number of extreme changes in output, government spending, and the predicted disbursement instrument from unduly influencing my estimates, I trim the sample at the first and 99th per- centiles of the distributions of these three variables. 16  IDA eligibility depends on a country’s GDP per capita falling below a given threshold, equal to US$1,175 at market exchange rates as of 2012. A further eight countries with higher per capita GDP are nevertheless ­IDA-eligible 184 American Economic Journal: Macroeconomics october 2014 Table 2—Disbursements from Official Creditors as Percentage of Total Government Spending Country Disb./Gov. Country Disb./Gov. Country Disb./Gov. Gambia, The* 37.6% Belize 15.6% Peru 9.8% Mauritania* 29.4% Swaziland 15.4% Dominica 9.5% Guinea-Bissau* 28.1% Kenya* 15.3% Congo, Rep.* 9.2% Madagascar* 26.7% Haiti* 15.0% Mauritius 9.0% Mali* 24.0% Costa Rica 14.7% Maldives* 8.9% Ghana* 23.6% Cape Verde 14.1% Colombia 8.8% Guyana* 22.9% El Salvador 14.1% St. Lucia 8.7% Mozambique* 22.8% Jamaica 13.9% Egypt 7.9% Lao P.d.r.* 22.5% Jordan 13.8% Albania 7.2% Guinea* 22.4% Paraguay 13.7% Moldova* 7.1% Burkina Faso* 22.3% Togo* 13.7% Papua New Guinea 7.1% Malawi* 21.7% Tunisia 13.7% Macedonia 6.8% Honduras* 21.6% Philippines 13.5% Yemen* 6.7% Nicaragua* 21.6% Lesotho* 13.4% Panama 6.2% Nepal* 20.5% St. Vincent and Gren. 12.9% Thailand 6.1% Niger* 20.1% Sri Lanka* 12.8% Botswana 5.8% Chad* 19.8% Liberia* 12.8% Sudan* 5.8% Senegal* 19.7% Cameroon* 12.7% Vanuatu 5.4% Tajikistan* 19.3% Indonesia 12.0% Fiji 5.3% Tanzania* 19.0% Morocco 12.0% Uzbekistan* 5.2% Uganda* 18.9% Comoros* 11.9% Algeria 5.1% Kyrgyz Republic* 18.8% Mongolia* 11.8% Seychelles 5.1% Sierra Leone* 18.8% Grenada 11.3% Uruguay 5.0% Burundi* 18.7% Georgia* 11.3% Gabon 4.7% Bangladesh* 18.6% Guatemala 11.2% Romania 4.3% Pakistan* 18.4% Sao Tome and Principe* 11.2% Bulgaria 4.1% Benin* 17.6% St. Kitts and Nevis 11.2% Syria 3.9% Armenia* 17.5% Djibouti* 11.1% Turkey 3.5% Ethiopia* 17.4% Ecuador 11.1% Solomon Islands* 3.4% Zambia* 17.2% Vietnam* 10.9% Angola* 3.3% Bolivia* 17.1% Eritrea* 10.7% Malaysia 3.3% Bhutan* 16.3% o ​te d’Ivoire* C​   10.4% Latvia 3.1% Rwanda* 16.1% Central Afr. Rep.* 10.3% Dominican Republic 15.8% Congo, Dem. Rep.* 10.2% Cambodia* 15.6% Tonga 10.0% Averages   Full sample 13.4%  Ida sample 16.2%  High Disb. sample 16.7% Notes: This table lists the 102 countries that make up the full sample, together with the over-time average of dis- bursements on loans from official creditors as a fraction of total government spending, for each country. These countries satisfy the criteria that (i) disbursements on loans from official creditors average at least 1 percent of GDP over the period 1970–2010, and (ii) at least 15 annual observations on government spending are available. IDA-eligible countries that make up the IDA subsample are indicated with asterisks. The high disbursements sub- ­ sample consists of countries in the first two columns, in which disbursements on loans from official creditors aver- age more than 10 percent of government spending. o creditors are substantially higher than in the full sample, averaging 16.7 and ­ fficial ­ 16.2 percent of government spending, respectively. Not surprisingly, the ­ first-stage relationship between fluctuations in predicted disbursements and government spend- ing will be stronger in these subsamples. under the “small island economies exception” (Kiribati, Cape Verde, Tonga, Vanuatu, Dominica, Grenada, Saint Lucia, and Saint Vincent). I exclude these countries from the IDA subsample used in this paper. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 185 Table 3—Summary Statistics Total Predicted Government Observations SD disbursements disbursements GDP spending Full sample   Total disbursements 2,804 2.0% 1.00   Predicted disbursements 2,804 0.6% 0.15 1.00  GDP 2,804 4.0% 0.00 0.03 1.00   Government spending 2,804 3.3% 0.12 0.11 0.25 1.00 IDA sample   Total disbursements 1,508 2.3% 1.00   Predicted disbursements 1,508 0.7% 0.19 1.00  GDP 1,508 4.0% 0.00 0.06 1.00   Government spending 1,508 3.5% 0.11 0.17 0.23 1.00 High disbursements sample   Total disbursements 1,950 2.1% 1.00   Predicted disbursements 1,950 0.7% 0.19 1.00  GDP 1,950 3.8% 0.00 0.05 1.00   Government spending 1,950 3.1% 0.11 0.16 0.23 1.00 Notes: This table reports summary statistics on the indicated variables. All variables are expressed as constant local- currency price changes scaled by lagged GDP, as in equation (1). In addition, all variables are in terms of deviations from country- and year-averages, consistent with the inclusion of country and year fixed effects in equation (1). Table 3 reports summary statistics on fluctuations in real GDP, government spending, and actual and predicted disbursements, in the three samples of countries. All variables are expressed as constant price annual changes, scaled by lagged GDP (as defined in equation (1)). In addition, I remove country- and year-specific means before calculating summary statistics, in order to be consistent with the empirical specifications that follow, all of which will also include a full set of country and year dummies. Real GDP growth and changes in government spending are quite volatile, with standard deviations of 4.0 and 3.3 percent, respectively, in the full sample, and of similar magnitudes in the two subsamples. Actual disbursements on loans from official creditors are also quite volatile, with standard deviations around 2 percent in the three samples. Naturally, my instrument based on predicted disbursements is less volatile than actual disbursements, with standard deviations of around 0.6 to 0.7 percent of GDP, but it nevertheless exhibits substantial variation. Fluctuations in predicted disbursements are correlated with fluctuations in government spending, and much more strongly so in the two subsamples of countries. The strength of this first-stage relationship will of course be crucial to the success of my identification strategy. III.  Benchmark Estimates of the Government Spending Multiplier Table 4 reports benchmark estimates of the government spending multiplier based on equation (1). The three panels of the table report the ordinary least squares (OLS), two-stage least squares (2SLS), and first-stage regressions, while the three columns refer to the three country samples discussed in the previous section. In addition, Figure 3 displays the scatterplots corresponding to the first-stage and second-stage regressions, partialling out the country and year fixed effects. The OLS ­ 186 American Economic Journal: Macroeconomics october 2014 Table 4—Benchmark Estimates of the Government Spending Multiplier Full IDA Disb/G>10% Sample of countries (1) (2) (3) Panel A. OLS estimates (Dependent variable is change in real GDP) Change in total government spending 0.306*** 0.259*** 0.277*** (0.0377) (0.0501) (0.0431) Panel B. 2SLS estimates (Dependent variable is change in real GDP) Change in total government spending 0.375 0.408** 0.417** (0.248) (0.197) (0.204) Weak instrument consistent 95% confidence interval [−0.058, 0.827] [0.071, 0.774] [ 0.082, 0.776] Panel C. First-stage regressions (Dependent variable is change in total government spending) Change in predicted disbursements 0.531*** 0.796*** 0.699*** (0.150) (0.150) (0.149) First-stage F-statistic on excluded instrument 12.62 28.08 22.18 Observations 2,804 1,508 1,950 Number of countries 102 60 70 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data and include a full set of country and year fixed effects. Changes in GDP, govern- ment spending, and predicted disbursements are all scaled by lagged GDP. Weak instrument consistent confidence intervals computed using the Moreira (2003) conditional likelihood ratio statistic. Panels A and B report OLS and 2SLS estimates of equation (1). Panel C reports OLS estimates of the corresponding first-stage regression. The three columns correspond to the full sample of countries, the sample of IDA-eligible countries, and the sample of countries where disbursements on loans from official creditors average at least 10 percent of government spending over the sample period. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. estimates of the multiplier are quite similar across samples, ranging from 0.26 to 0.31, and are very precisely estimated, with standard errors ranging from 0.04 to 0.05. As discussed above, however, these OLS estimates are likely to be biased to the extent that fluctuations in government spending are correlated with other shocks to GDP growth that are reflected in the error term. The 2SLS estimates in panel B, which are designed to correct for such biases, are somewhat larger than the OLS estimates, ranging from 0.38 to 0.42. They are also fairly precisely estimated, with standard errors between 0.20 and 0.25.17,18,19 17  These estimated standard errors for spending multipliers are respectable when compared with other papers in the literature. For example, Barro and Redlick (2011) use US data over the past century to estimate defense spending multipliers, and obtain standard errors ranging from 0.06 to 0.27 (their Table 2, first row). Similarly, the confidence bands around VAR-based impulse responses reported in Figure 5 of Blanchard and Perotti (2002) imply a standard error for the impact multiplier of 0.35. The predicted disbursements measure is a generated instrument (consisting of actual loan commitments mul- 18  tiplied by estimated average disbursement rates). However, this does not matter for the asymptotic distribution of the 2SLS estimator as long as actual loan approvals in year t are not correlated with macroeconomic shocks in year t + 1 and higher, as per my core identifying assumption. See Wooldridge (2001, chapter 6.1.2). 19  One notable assumption underlying these estimates and standard errors is that cross-sectional dependence in the error term is adequately captured by year fixed effects. This embodies the simple but unappealing assumption that common shocks have the same effect on all countries, and can be swept out using year dummies, as I do in the default specification. As a robustness check, I relax this assumption by implementing an estimator based on the cross-sectional averages of moment conditions which is asymptotically valid as T → ∞ under very general cross-sectional dependence, as proposed by Driscoll and Kraay (1998). This delivers slightly smaller estimates of ­ the multiplier ranging from 0.22 to 0.29, with standard errors ranging from 0.17 to 0.29. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 187 First stage 0.2 Change in government spending 0.1 0 −0.1 −0.2 −0.02 −0.01 0 0.01 0.02 0.03 Change in predicted disbursements OLS (thin) and IV (thick) estimates of multiplier 0.2 0.1 Change in real GDP 0 −0.1 −0.2 −0.2 −0.1 0 0.1 0.2 Change in government spending Figure 3. Scatterplots for Benchmark Results Notes: This graph shows the first-stage relationship between government spending and pre- dicted disbursements (top panel) and the second-stage relationship between GDP and govern- ment spending (bottom panel). All three variables are scaled by lagged GDP and expressed as deviations from country- and year-specific means. While the estimated multipliers are significantly greater than zero in the IDA and high-disbursement samples, in all three specifications I can reject the null hypothesis that the multiplier is equal to one at the 5 percent level. A further noteworthy feature of these benchmark results is that in all cases the 2SLS estimates of the multiplier are larger than the OLS estimates, suggesting that the latter are biased downwards. This may reflect a combination of (i) attenuation bias in the OLS estimates due to 188 American Economic Journal: Macroeconomics october 2014 measurement error in the government spending, as well as (ii) a countercyclical response of overall government spending to macroeconomic shocks.20 However, these differences should not be over-interpreted as they are far from statistically significant: the p-values for the null hypothesis that the OLS and 2SLS estimates are equal are 0.76, 0.75, and 0.63 respectively in the three samples of countries. In panel C of Table 4, I report the corresponding first-stage regressions for the three country samples. The first-stage relationship between changes in government spending and changes in predicted disbursements is quite precisely estimated, with first-stage F-statistics greater than the Staiger and Stock (1997) rule of thumb of ten in all three samples. Not surprisingly, the first-stage relationship is also much stronger in the second and third columns, in which the first-stage F-statistics are 28.1 and 22.2, respectively. This reflects the fact that lending from official credi- tors is a relatively more important source of financing for government spending in these more aid-dependent countries, and so the fluctuations in the predetermined component of this spending captured by my instrument have greater explanatory power for fluctuations in overall government spending. As would be expected given the strength of the instrument, the weak-instrument consistent 95 percent confidence intervals reported in panel B are quite similar to ones based on the usual asymptotic normal approximation.21 Figure 4 provides further information on the timing of the relationship between my predicted disbursement instrument, government spending, and output, that speaks to two potential concerns with the identification strategy. The first concern is that fluctuations in predicted disbursements in year t (which reflect only fluctua- tions in loan approvals in years in year t − 1 and earlier) may be associated with increases in government spending not just in year t, but in subsequent or previous years as well. In this case, the estimated multipliers reported in Table 4 would reflect not only contemporaneous, but also past or future increases in government spending. To assess this possibility, in the left-hand column, I report the estimated coefficients and 95 percent confidence intervals from a regression of changes in government spending on contemporaneous changes in predicted disbursements, as well as five leads and lags of changes in predicted disbursements, and a full set of country and year dummies. I do this separately for the three samples of countries noted above. These graphs in the left-hand column show a strong and significant contemporaneous correlation between changes in predicted d ­ isbursements and 20  Absent direct information on the signal-to-noise ratio in fluctuations in government spending, it is not pos- sible to distinguish between these two possible explanations. Since the OLS estimates are roughly 75 percent of the IV estimates, standard textbook calculations suggest that the signal-to-noise ratio in government spending would have to be about three in order to account for the gap between the two estimates. If measurement error is less (more) extreme than this benchmark, then government spending would also have to be countercyclical (procyclical) in order to explain the differences between the OLS and IV estimates. 21  It is tempting to interpret the slope of the first-stage regression as an estimate of the extent to which a dollar of aid in the form of loans from official creditors is reflected in recipient government spending. This interpreta- tion would be misleading however because the first-stage regression captures the relationship between changes in government spending and changes in predicted disbursements (i.e., changes in disbursements that would have occurred had individual loans disbursed according to their typical disbursement profiles), and not changes in actual disbursements. The first-stage slope could either over- or underestimate the relationship between aid and gov- ernment spending, depending on the correlation between changes in government spending and deviations from predicted disbursements. While this is an interesting issue, it is not central to the analysis here and is postponed to future research. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 189 Government spending response Output response Panel A. Full sample 1 1 Government spending 0.75 0.75 0.5 0.5 Output 0.25 0.25 0 0 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 −0.25 −0.25 −0.5 −0.5 Panel B. IDA sample 1 1 Government spending 0.75 0.75 0.5 0.5 0.25 Output 0.25 0 0 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 −0.25 −0.25 −0.5 −0.5 Panel C. High-disbursement sample 1 1 Government spending 0.75 0.75 0.5 0.5 Output 0.25 0.25 0 0 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 t=−5 t=−4 t=−3 t=−2 t=−1 t=0 t=1 t=2 t=3 t=4 t=5 −0.25 −0.25 −0.5 −0.5 Figure 4. Responses of Government Spending and Output to the Predicted Disbursement Instrument Notes: The graphs in the left (right) column report the estimated coefficients and 95 percent confidence intervals from regressions of changes in government spending (output) on contemporaneous changes in predicted disburse- ments, as well as five leads and lags of changes in predicted disbursements, and country and year effects. The three panels correspond to the full sample of countries, the IDA sample, and the high-disbursement sample. changes in government spending, consistent with the strong first-stage regressions documented in Table 4. In contrast, there is little evidence of a significant cor- relation between changes in predicted disbursements and changes in government spending at any other horizons. This suggests that the predicted disbursement instrument, based on the interaction of loan approvals in year t − 1 and earlier with typical loan disbursement profiles, is doing a good job of isolating fluctua- tions in government spending at year t alone. The second concern has to do with the plausibility of the identifying assump- tion that loan approval decisions do not anticipate future shocks to growth. While this identifying assumption is of course fundamentally untestable, it is neverthe- less helpful to examine the relationship between changes in output and changes 190 American Economic Journal: Macroeconomics october 2014 in predicted disbursements at different leads and lags. I do this in the right-hand column of Figure 4, in the same way as for changes in government spending. The ­ contemporaneous correlation between changes in predicted disbursements and changes in output is small, positive, and statistically significantly different from zero only in the IDA and high-disbursement subsamples of countries. This is consistent with the pattern of estimated multipliers in Table 4. Moreover, there is little evidence that changes in output are correlated with lagged changes in predicted disburse- ments. To see why this is important, note that changes in predicted disbursements in year t − 1, for example, reflect only loan approval decisions made in years t − 2 and earlier. The key identifying assumption is that these past loan approval decisions are not correlated with subsequent shocks to growth in years t − 1 or later. If this assumption were violated, one might expect to see changes in output that are cor- related with lagged changes in predicted disbursements. While this is by no means a test of the identifying assumption, it is at least suggestive that the identification strategy is reasonable. Table 5 reports a series of 2SLS estimates of the multiplier for several alternative versions of the instrument, in order to better understand the source of identification of my results. The three columns again refer to the three country samples, which are fixed across these alternatives, and so the corresponding OLS regressions are the same as those reported in panel A of Table 4, and are not repeated here. The first variant corresponds to constructing the instrument by aggregating predicted dis- bursements on loans extended by multilateral creditors only, while the second vari- ant corresponds to a version of the instrument based on loans extended by bilateral creditors only. The difference between the two sets of results is stark. The strength of identification, as measured by the first-stage F-statistics, is much higher in the results based on multilateral predicted disbursements than for bilateral predicted disbursements. The first-stage F-statistics in panel A of Table 5 are similar to those in Table 4 and range from 11.7 to 23.6. In contrast, the first-stage F-statistics are below ten in all three samples when the instrument is based on predicted disburse- ments on loans from bilateral creditors. The reasons for these differences are straightforward. First, as shown in Table 1, loans from multilateral creditors account for the majority of total disbursements in my sample, averaging 59 percent over the whole sample, and substantially more in recent years. Second, as is apparent from Figure 1, disbursements on loans from multilateral creditors are on average substantially more backloaded than disbursement on loans from bilateral creditors. The first observation implies that fluctuations in disbursements on loans from official creditors are a relatively more important source of variation in government spending in my sample of devel- oping countries. The second observation implies that predicted disbursements on loans approved in previous years are larger in the case of loans from official creditors. Together, these two factors contribute to a much stronger first-stage relationship between changes in government spending and changes in predicted disbursements. The weak identification based on loans from bilateral creditors is reflected in much more imprecise estimates of multipliers when only this source of exogenous variation is used. In contrast, the estimates of the multiplier identi- fied from fluctuations in predicted disbursements from multilateral creditors are Vol. 6 No. 4 kraay: government spending multipliers in developing countries 191 Table 5—Estimates of the Government Spending Multiplier— Variants on the Predicted Disbursements Instrument Sample of countries (Dependent variable is change Full IDA Disb/G>10% in real GDP) (1) (2) (3) Panel A. Multilateral creditors only Change in total government spending 0.433* 0.612*** 0.526*** (0.238) (0.187) (0.171) First-stage F-statistic on excluded instrument 11.67 23.62 21.46 Panel B. Bilateral creditors only Change in total government spending 0.293 0.0814 0.247 (0.406) (0.313) (0.346) First-stage F-statistic on excluded instrument 4.11 5.73 6.25 Panel C. Typical disbursement rates based on pooling all loans Change in total government spending 0.532* 0.532* 0.511* (0.317) (0.280) (0.265) First-stage F-statistic on excluded instrument 9.83 19.52 15.95 Panel D. Typical disbursement rates calculated excluding country in question Change in total government spending 0.332 0.339* 0.366* (0.238) (0.189) (0.197) First-stage F-statistic on excluded instrument 14.68 28.11 24.17 Panel E. Predicted disbursements based only on slow-disbursing loans Change in total government spending 0.415 0.426* 0.431* (0.277) (0.224) (0.225) First-stage F-statistic on excluded instrument 12.84 22.64 23.90 Observations 2,804 1,508 1,950 Number of countries 102 60 70 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data and include a full set of country and year fixed effects. Changes in GDP, government spending, and disbursements are all scaled by lagged GDP. Table reports 2SLS estimates of equation (1) using the following variants on the predicted disbursements instrument: (i) using loans from official creditors only; (ii) using loans from bilateral creditors only; (iii) using typical disbursement rates calculated pooling across all loans, rather than within creditor/region/decade bins; and (iv) excluding country in question when calculating typical disburse- ment rates. The three country samples correspond to the full sample of countries, the sample of IDA-eligible coun- tries, and the sample of countries where disbursements on loans from official creditors average at least 10 percent of government spending over the sample period. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. much more precisely estimated, and moreover are somewhat larger than those reported in Table 4, ranging from 0.43 to 0.61.22 Overall, this shows that much estimates comes from the strong first-stage of the identification of my benchmark ­ relationship between government spending and predicted disbursements on loans from multilateral creditors. 22  While these differences are small relative to the estimated standard errors and should not be overinterpreted, one possible explanation for this difference in magnitude is that loan approvals by bilateral donors are more likely to anticipate future negative shocks to growth than loan approvals by multilaterals. If this potential violation of the exclusion restriction were important for the component of the instrument based on bilateral lenders, it would imply a downward bias in the 2SLS estimates of the multiplier, that is corrected when bilateral lending is removed from the instrument. 192 American Economic Journal: Macroeconomics october 2014 The third set of results in Table 5 addresses the possibility of over-fitting the l­oan-level predicted disbursements series. Recall that loan-level p ­redicted disbursements are based on average disbursement rates calculated within ­ 443 creditor/decade/region bins. To the extent that disbursement profiles are more homogenous within more disaggregated bins, a potential concern is that the more disaggregated are these bins, the closer predicted disbursements will be to actual disbursements. This would in turn imply that, the more disaggregated are the bins on which predicted disbursements are based, the less effective are predicted disburse- ments in purging the endogenous component of actual disbursements. To address this concern, I construct an alternative extreme version of the instrument, based on combining all loans from all creditors into a single bin, i.e., applying the overall average disbursement profile shown in the top panel of Figure 1 to all of the loans in my dataset. Naturally, doing so leads to a somewhat weaker first-stage relationship between changes in government spending and changes in predicted disbursements, with first-stage F-statistics ranging from 9.8 to 19.5 (as opposed to 12.6 to 28.1 using the default instrument). However, the first-stage fit remains respectable, and the point estimates of the multiplier change only slightly to around 0.5 as compared with around 0.4 in the default specification. This robustness check suggests that over-fitting of the predicted disbursement instrument is not a major concern in my benchmark results. Another possible objection to the predicted disbursements instrument is it indi- rectly includes some information on future country-specific shocks, as it is based on typical disbursement rates averaging across all loans within the creditor/decade/ region bins, including future loans to the country in question. This potential viola- tion of the exclusion restriction is unlikely to be very important given the large num- ber of loans within each bin (recall that the median bin includes 65 loans). However, as a further robustness check, I reconstruct the instrument, but now excluding all other loans to the country in question when calculating average disbursement rates. This eliminates any potential future country-specific information in the predicted disbursement instrument that comes through the inclusion of the country in question in the calculation of average disbursement rates. The fourth set of results in Table 5 show that this robustness check has only minimal effects on my benchmark esti- mates. The first-stage F-statistics are actually slightly higher than in the benchmark results in Table 4, and the estimates of the multiplier are slightly smaller (ranging from 0.33 to 0.37). In summary, the benchmark results in this section suggest that the one-year government spending multiplier is in the vicinity of 0.4, and moreover is reason- ably precisely estimated. Specifically, I find that the multiplier is in most cases significantly different from zero and also significantly less than one. The statisti- cal identification of these multipliers comes primarily from a strong first-stage relationship between fluctuations in the predetermined component of disburse- ments on loans from multilateral, as opposed to bilateral, creditors. These findings are robust to variants on the instrument designed to address possible concerns about over-fitting of predicted disbursements, and the possible incorporation of future information country-specific information in the calculation of typical dis- bursement profiles. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 193 Table 6A—Robustness Checks Removing influential Adding cubic country-specific observations time trends Sample of countries (Dependent Full IDA Disb/G>10% Full IDA Disb/G>10% variable is change in real GDP) (1) (2) (3) (4) (5) (6) Panel A. OLS estimates Change in government spending 0.313*** 0.266*** 0.275*** 0.263*** 0.220*** 0.233*** (0.0380) (0.0505) (0.0437) (0.0384) (0.0517) (0.0445) Panel B. 2SLS estimates Change in government spending 0.371 0.387* 0.389** 0.260 0.266* 0.291* (0.268) (0.194) (0.191) (0.204) (0.156) (0.166) First-stage F-statistic 11.48 41.20 30.34 17.42 25.23 24.50 Observations 2,786 1,503 1,939 2,804 1,508 1,950 Number of countries 102 60 70 102 60 70 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data. Columns 1–3 include a full set of country and year fixed effects. Changes in GDP, government spending, and disbursements are all scaled by lagged GDP. Panel A reports OLS estimates of equa- tion (1). Panel B reports 2SLS estimates of equation (1). Columns 1–3 report results after using the Hadi (1992) procedure to remove influential observations from the first-stage and reduced-form regressions for the benchmark specifications in Table 4. Columns 4–6 add country-specific cubic time trends to the benchmark specification. Note that since I am now estimating country-specific trend parameters, it no longer is possible to also cluster standard errors at the country level as in the benchmark specification. This is why the first-stage F-statistics are larger, and the estimated standard errors are smaller, than in the benchmark specification. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. IV.  Further Robustness Checks I next address a variety of further potential concerns about the robustness of the benchmark estimates of the multiplier presented in Table 4. Given the noisy and highly-imperfect data on government spending and output in many of the develop- ing countries that comprise my sample, a somewhat generic first concern is that the results in Table 4 might be driven by a small number of influential observations. To investigate this possibility more systematically, I use a procedure suggested by Hadi (1992) to identify influential observations in the reduced-form and first-stage regressions (the ratio of the corresponding two slope coefficients being the 2SLS estimate of the multiplier). I then reestimate the OLS, first-stage, and 2SLS regres- sions, excluding these influential observations. The results of this first robustness check are reported in the first three columns of Table 6A. The OLS estimates of the multiplier change very little relative to the benchmark results. The 2SLS esti- mates of the multiplier are virtually unchanged once influential observations are removed, ranging from 0.37 to 0.39, and moreover they are slightly more precisely estimated than before. This is in part due to an even stronger first-stage relationship after removing influential observations in the IDA and high-disbursement samples, in which the first-stage F-statistics jump to 41.2 and 30.3, respectively.23 23  A related data quality concern has to do with the fact that, as discussed in Section II, the data on changes in ­ traightforward government spending that I use as the dependent variable is assembled from a variety of sources. One s 194 American Economic Journal: Macroeconomics october 2014 A second potential concern is that estimated relationships between changes in output, changes in government spending, and changes in predicted disbursements in part reflect slowly-moving country-specific trends in these variables. Suppose, for example, that a country experiences a decade of steadily increasing government spending, together with aid in the form of concessional loans, and that this decade also happens to be a time of higher-than-average growth. This would bias up my esti- mates of the short-run spending multiplier, by contributing a number of observations corresponding to large changes in predicted disbursements and large changes in government spending, combined with high growth. Of course, the opposite direction of bias is possible as well, the period of increasing loans and government spending happens to be a period of slower average growth. I investigate this potential concern in columns 4–6 of Table 6A, by adding country-specific third-order polynomial time trends to the benchmark specification, to flexibly capture any such slowly-moving trends. I find that the estimated multipliers are slightly smaller than in the bench- mark specification, ranging from 0.26 to 0.29, as compared with 0.38 to 0.42 in the benchmark specification. At the same time, however, these differences are not large relative to the estimated standard errors, suggesting that omitted country-specific trends are not a major consideration in driving my results.24 A third potential objection to my results has to do with the possibility that com- mitments and/or disbursements on loans trigger changes in policy performance that themselves have direct effects on economic growth. If, for example, the approval of a loan from an official creditor such as the World Bank is conditional on spe- cific policy reforms that might themselves affect growth over the next several years, this would violate the exclusion restriction that loan commitments are uncorrelated with future shocks to growth. Another possibility might be that loan approvals are a response to policy reforms which themselves are persistent over time.25 To the extent that the associated reforms are growth-enhancing, this would contribute to a positive correlation between my instrument and the error term in equation (1), which would in turn imply an upward bias in my 2SLS estimates of the multiplier. robustness check is to reestimate the specifications in Table 4, but restricting attention to only those observations for which the government spending data come from the IMF’s World Economic Outlook, the most comprehensive single data source. The 2SLS point estimates of the multiplier are slightly larger using only the WEO data, ranging from 0.46 to 0.69. However, they are much less precisely estimated, and the first-stage fit is much weaker. This is because the sample size is reduced by 40 percent, and moreover is now concentrated in the 1990s and 2000s, a period during which loans from official creditors become a less important source of financing of government spending. 24  I obtain broadly similar results if I replace the third-order polynomial trend with a linear trend, or with a qua- dratic trend. For example, in the full sample these variants give estimated multipliers of 0.28 and 0.38 respectively, as compared with 0.38 in the benchmark specification. It is also important to note that the inclusion of country- specific polynomial trends does not significantly undermine the strength of the estimated first-stage relationship between changes in predicted disbursements and changes in government spending. To see this, note first that in the specifications with polynomial trends, I do not cluster standard errors at the country level as this is no longer feasible when there are country-specific parameters to be estimated. This means that the first-stage F-statistics reported here are not comparable with those in the benchmark specification without country-specific trends reported in Table 4, where the standard errors are clustered at the country level. However, if I reestimate the benchmark specification in Table 4 without clustering and without country-specific trends, the first-stage F-statistics are only slighly larger than those reported in Table 6A for the specifications with cubic country-specific trends. 25  Note however that, since I rely on predicted rather than actual disbursements, I do not need to be concerned with the possibility that disbursements on loans are triggered by policy reforms. This is because any ­country-specific variation in disbursement rates has been eliminated from the predicted disbursements instrument. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 195 Table 6B—Further Robustness Checks Controlling Controlling for approvals of for policy slow-disbursing loans Sample of countries (Dependent Full IDA Disb/G>10% Full IDA Disb/G>10% variable is change in real GDP) (1) (2) (3) (4) (5) (6) Panel A. OLS estimates Change in government spending 0.305*** 0.253*** 0.267*** 0.305*** 0.257*** 0.276*** (0.0390) (0.0508) (0.0433) (0.0375) (0.0498) (0.0429) Change in CPIA policy indicator 0.0109*** 0.00809** 0.00999*** (0.00284) (0.00371) (0.00336) Approvals of slow-disbursing 0.0591** 0.0700** 0.0535*  projects (0.0272) (0.0318) (0.0278) Panel B. 2SLS estimates Change in government spending 0.286 0.373* 0.441* 0.335 0.364* 0.389* (0.301) (0.218) (0.232) (0.249) (0.193) (0.202) Change in CPIA policy indicator 0.0111*** 0.00706 0.00856** (0.00362) (0.00428) (0.00396) Approvals of slow-disbursing 0.0586** 0.0676** 0.0520*  loans (0.0267) (0.0311) (0.0272) First-stage F-statistic 8.7 21.03 16.66 12.20 26.65 22.98 Observations 2,516 1,386 1,755 2,787 1,494 1,935 Number of countries 102 60 70 102 60 70 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data. Columns 1–6 include a full set of country and year fixed effects. Changes in GDP, government spending, and disbursements are all scaled by lagged GDP. Panel A reports OLS estimates of equa- tion (1). Panel B reports 2SLS estimates of equation (1). Columns 1–3 control for changes in the CPIA policy indi- cators. Columns 4–6 control for contemporaneous approvals of slow-disbursing loans. The three country samples correspond to the full sample of countries, the sample of IDA-eligible countries, and the sample of countries where disbursements on loans from official creditors average at least 10 percent of government spending over the sample period. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. I investigate this possibility in columns 1–3 of Table 6B, by controlling for con- temporaneous changes in policy performance, as measured by the World Bank’s Country Policy and Institutional Assessment (CPIA) ratings. These ratings have been produced annually since 1978 by World Bank country economists for all World Bank borrowers, and are scored on a 1–6 point scale, with higher values corre- sponding to better policy performance.26 Annual changes in policy performance are significantly positively correlated with growth in the full sample and high disburse- ments sample in the 2SLS specification. However, controlling for policy does not significantly change the 2SLS estimates multipliers relative to the benchmark speci- fication, with estimates that fall between 0.29 and 0.44. In summary, while ­short-run changes in policy performance do appear to matter for short-run fluctuations in growth, these changes in policy performance are not very strongly c ­ orrelated with 26  The structure of the CPIA has evolved somewhat over time. Currently it is based on a set of 16 indicators cor- responding to different policy areas, each of which is scored on a 1–6 scale in increments of 0.5. The overall CPIA score is a simple average across these component indicators, and varies fairly continuously from one to six. Details on the CPIA assessment criteria can be found at www.worldbank.org/ida. 196 American Economic Journal: Macroeconomics october 2014 past loan commitment decisions, and so do not significantly impact my estimates of the multiplier.27 Another difficulty in interpreting the estimated multipliers in this paper has to do with anticipation effects. I identify the government spending multiplier using fluc- tuations in predicted disbursements on loans from official creditors, which I have argued are plausibly uncorrelated with future macroeconomic shocks. At the same time, however, the spending plans set in motion at the time of loan commitment, as well as the associated burden of future taxes required to eventually repay the loan, are both known at the time of loan approval. As stressed by Ramey (2011b), it is likely that private agents will respond to these anticipated future events at the time that the future spending plans are announced, rather than when the spending actu- ally occurs. For example, one might expect that the standard positive neoclassical labor supply response to the negative wealth effect of an increase in government spending should occur at the time that the spending plans are announced, and not when they are implemented. This in turn suggests that the contemporaneous rela- tionship between changes in spending and changes in output will understate the overall effect, since most of the effect occurs at the time of the announcement. An even richer set of theoretical possibilities emerges when government spend- ing is both anticipated and productive. For example, Leeper, Walker, and Yang (2010) study a calibrated neoclassical model in which productive public capital takes time to build. The announcement of spending plans triggers the usual neo- classical negative wealth effect from higher future taxes, although this is muted because agents understand that their future productivity will be enhanced by public spending. Moreover, the anticipated future productivity of public spending makes it optimal to shift investment towards the future when public capital is complete. In this case, the overall effect of anticipation can be to shift the output response to an announced future increase in spending more towards the subsequent periods in which the spending occurs. Given that my identification strategy relies on the predetermined component of disbursements on loans from official creditors, I cannot use it to contribute to the evidence on the output effects of an unanticipated increase in government spending, and the results in the paper should be interpreted with this qualification in mind. Moreover, given the theoretical ambiguities noted above, it is unclear whether anticipation effects imply that my estimates of the contemporaneous relationship between changes in government spending and changes in output overstate or under- state the overall multiplier. Rather than attempt to sort out these issues with a struc- tural model, my much more modest objective in this section is to provide some suggestive evidence on how output responds to the approval of new loans from offi- cial creditors, and how this matters for my empirical estimates of the contemporane- ous spending multiplier. This addresses the specific concern that if loan approvals are serially correlated within countries over time, then predicted disbursements on 27  There is however some weak evidence that loan approvals might be correlated with lagged improvements in policy. Adding the first lag of changes in the CPIA reduces the estimates of the multiplier somewhat relative to the benchmark, with 2SLS estimates ranging from 0.11 to 0.31 across the three samples (as compared with 0.38 to 0.42 in the benchmark specification). Vol. 6 No. 4 kraay: government spending multipliers in developing countries 197 previously-approved loans in a given year may be correlated with contemporaneous loan approvals in the same year, which themselves may have direct output effects through some combination of the anticipation mechanisms discussed above. In principle, a straightforward way of addressing this issue is to control for the commitment of new loans. Doing so, however, is complicated by the same basic problem that motivates my identification strategy—loan commitment decisions are themselves potentially endogenous responses to contemporaneous macroeconomic events. As a result, I cannot simply include contemporaneous loan approvals as an additional regressor in equation (1). Instead, it is necessary to somehow distin- guish between loans that are committed for cyclical reasons and those that are not. In Kraay (2012), I developed a coding of World Bank projects according to their cyclical motivation based on a reading of project documentation. Not surprisingly, I found that projects approved for cyclical reasons also typically disbursed much more quickly than projects approved for other reasons. Applying the same reasoning in this context, it is plausible that loans that ultimately take many years to disburse are less likely to have been approved for cyclical reasons, whereas it is more likely that fast-disbursing loans are cyclically motivated. Accordingly, I construct a variable containing the total value of new loan com- mitments as a fraction of GDP in a given country-year, restricting attention to loans that ultimately take four or more years to fully disburse.28 Based on the discussion above, I assume that these loans are unlikely to have been approved for cyclical rea- sons, and so I can include this variable as an additional exogenous control variable in equation (1). The results are shown in the last three columns of Table 6B. This variable enters positively in all three samples, and significantly so in the full sample and IDA sample, suggesting that there are some output responses to the approval of loans from official creditors, consistent with the idea that anticipated future changes in spending have some immediate output effects. However, controlling for this effect only slightly reduces my estimates of the contemporaneous spending multiplier, relative to the benchmark estimates in Table 4. Taken together these results suggest some evidence in favor of the hypothesis that output responds to loan approvals, but that this channel has little effect on the estimates of the contemporaneous effects of government spending when the spending eventually occurs. To sum up, in this section I have considered a range of alternative specifications designed to check the robustness of my results to a variety of potential objections to the validity of the exclusion restriction. For the most part, these changes have only small effects on estimated multipliers, relative to the benchmark results of the previ- ous section. Overall, they suggest that the one-year impact of an additional dollar of government spending is to raise GDP in the same year by somewhere between 0.3 and 0.5 dollars in most specifications. 28  This is the same threshold used in Kraay (2012) for World Bank-financed projects. I obtain similar results considering loans that require at least three or at least five years to fully disburse. 198 American Economic Journal: Macroeconomics october 2014 V.  Longer-Run Effects of Government Spending The empirical work thus far has focused on estimating the one-year government spending multiplier. It is plausible, however, that the full effect of an increase in government spending on output takes more than one year to become apparent in the data. In order to empirically document these longer-term output effects of govern- ment spending, I use the local projections approach of Jordà (2005) to estimate the impulse response function of output to an increase in government spending over a multiyear horizon.29 Specifically, I estimate the following series of local projection regressions p (2) Δ​ y​ t+h​ i,  ∑   = ​  ρ​  h  ​   ​ ​  Δ​ s​  ​   + ​ yi​, t−s​ gi​t​ β​  h​  Δ​   + ​ μ​  h   + ​ i​  ​ t​  ​ λ​  h εi​, t+h​   + ​  , s=1 _y​ ​  − ​ it+h​ yi​t−1​ _g​ ​  − ​ it​ g​ it−1​ where Δ​ y​ t+h​  ≡ ​  i,  yi​t−1 ​ ​ ​    and Δ​ g​it​  ≡ ​  ​ yi​t−1 ​ ​     . The coefficients ​ β​  h​ trace out the impulse response function of the change in output at time t + h to a change in government spending at time t, while the cumulative sums of the ​ β​  h​ capture the cumulative impact of an additional dollar of government spending on the level of output after h periods. The parameters μ i​  ​ and λ ​ ​  h t​  ​ capture country- and year-effects ​ ​  h that may vary with the projection horizon, and the error term ε ​​ i,t+h​captures all other sources of variation in output changes h periods into the future. I estimate equa- tion (2) by 2SLS, using the same predicted disbursement instrument as before. Conditional on the identifying assumption that loan approvals do not respond to future shocks to growth, this instrument will be uncorrelated with ​ ε​ t+h​ for h ≥ 0. i,  As before, I also include country and year dummies to sweep out the country and year effects. Note that when h = 0 and p = 0, the local projection regression reduces to the benchmark empirical specification in equation (1). Relative to this benchmark, the addition of lagged changes in output to this specification has implications for the plausibility of my identifying assumption that loan approval decisions are uncorre- lated with future macroeconomic shocks. A potential objection to this assumption is that, while loan commitments are made before subsequent shocks are realized, these shocks may themselves be persistent or otherwise predictable in some way. If, in addition, loan commitments are correlated with contemporaneous shocks, then they will also be correlated with future shocks, in violation of my exclusion restriction. Controlling for lagged growth is a straightforward way of addressing this possibil- ity, and therefore also enhancing the credibility of the identifying assumption. Table 7 reports the cumulative estimated effect of an additional dollar of govern- ment spending on output over a four-year horizon following the initial spending increase, and setting the lag length p = 1. The estimated impact effect of spending in the first year (i.e., at h = 0) is very similar to that reported in the benchmark specification in Table 4, with OLS estimates ranging from 0.25 to 0.29, and the See Auerbach and Gorodnichenko (2012a) and Leduc and Wilson (2012) for applications of this technique to 29  the estimation of fiscal multipliers. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 199 Table 7—Longer-Run Effects OLS 2SLS Full IDA Disb/G>10% Full IDA Disb/G>10% Sample of countries (1) (2) (3) (4) (5) (6) Cumulative effect on output over h years:   h = 0 0.290*** 0.249*** 0.267*** 0.337 0.391** 0.404** (0.0381) (0.0512) (0.0437) (0.249) (0.191) (0.202)   h = 1 0.0854** 0.0589 0.111** 0.561** 0.561*** 0.573*** (0.0405) (0.0537) (0.0521) (0.256) (0.210) (0.212)   h = 2 0.0904 0.0931 0.0978 0.164 0.245 0.352 (0.0575) (0.0732) (0.0668) (0.408) (0.369) (0.352)   h = 3 0.165** 0.167* 0.182** 0.138 0.123 0.310 (0.0726) (0.0914) (0.0843) (0.684) (0.632) (0.608)   h = 4 0.166* 0.185 0.196* 0.580 0.500 0.623 (0.0911) (0.127) (0.101) (0.741) (0.687) (0.656) Lagged change in GDP (from the 0.130*** 0.0987*** 0.105*** 0.0580*** 0.0407** 0.0303**  h = 0 local projection regression) (0.0310) (0.0353) (0.0352) (0.0115) (0.0163) (0.0120) First-stage F-statistic 11.81 26.66 21.39 11.81 26.66 21.39 Observations 2,782 1,493 1,937 2,782 1,493 1,937 Number of countries 102 60 70 102 60 70 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data and include a full set of country and year fixed effects. Changes in GDP, government spending, and disbursements are all scaled by lagged GDP. Columns 1–3 and columns 4–6 report OLS and 2SLS estimates of the sequence of local projection regressions in equation (2), for the three indicated country samples: the full sample of countries, the sample of IDA-eligible countries, and the sample of countries where disbursements on loans from official creditors average at least 10 percent of government spending over the sample period. The table entries report the cumulative sums of the impulse response function for change in output, and so capture the cumu- lative effect on output at t + h of an additional dollar of government spending in year t. To conserve space, the coef- ficient on the lagged change in output is reported only for the h = 0 local projection regression. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. 2SLS estimates slightly larger and ranging from 0.34 to 0.40. The bottom part of the table also reports the coefficient on lagged growth, which is positive and strongly significant across all specifications, indicating a small but significant degree of per- sistence in annual GDP growth rates in these countries. Nevertheless, the similarity of the estimated impact multipliers to those in the benchmark specification (which did not control for lagged growth) suggest that loan approval decisions reflected in the instrument are not correlated even with contemporaneous growth shocks. While clearly not a test of my identification strategy, this observation does provide some support for the plausibility of the exclusion restriction which requires loan approval decisions to be unrelated to future shocks to growth as well. β​  1​, ​ ​  0​  + ​ The cumulative output effect in the first year following disbursement, i.e., β is slightly larger than the initial one-year impact, suggesting some small additional output effects over and above the contemporaneous ones estimated so far in the paper. The 2SLS estimates are in the vicinity of 0.56, and are strongly significantly different from zero across all three samples in the 2SLS specifications. However, over longer horizons than this, the estimated output effects rapidly becomes much less precisely estimated—the cumulative output effect of an increase in government spending after two or more years is statistically indistinguishable from zero. Overall, 200 American Economic Journal: Macroeconomics october 2014 this evidence suggests that my identification strategy is not very informative about the longer-run effects of government spending on output, beyond the initial year or two following the increase in spending. VI.  Heterogeneity in Estimated Multipliers As noted in the introduction, an important benefit of my data and identification strategy is that it is applicable to a very large set of developing countries that rely on loans from official creditors as an important source of financing for govern- ment spending. The large cross-sectional dimension of my dataset makes it feasible to empirically examine a variety of possible hypotheses regarding differences in spending multipliers across countries, and also over time, as I do in this section. A first potential source of heterogeneity in estimated multipliers is the state of the business cycle. A variety of theoretical mechanisms for spending multipliers imply larger short-run effects of government spending during economic downturns when there is greater slack in the economy. Consistent with this view, Auerbach and Gorodnichenko (2012a, 2012b) provide extensive evidence for the United States and for OECD countries that multipliers are indeed smaller during booms and larger during recessions. To investigate this possibility in my sample of developing coun- tries, I classify each country-year observation in my sample as being in a boom or recession based on whether real GDP growth is above or below the corresponding country-decade average. I then estimate multipliers separately for booms and reces- sions. The results are shown in panel A of Table 8. Consistent with the theory, I find that estimated multipliers are much larger dur- ing recessions than during booms in all six specifications (OLS and 2SLS, for the three country samples). Particularly in the case of the 2SLS estimates, the differences are quite dramatic. During recessions, the estimated multipliers range from 0.61 to 0.81, while in booms the multipliers are between 0.01 and 0.15. Although these dif- ferences in point estimates are sufficiently large as to be economically meaningful, they are small relative to estimated standard errors. In none of the six cases do I find a statistically significant difference in the multipliers in booms versus reces- sions.30 This in part reflects the rather weaker identification in several subsamples— first-stage F-statistics are above 10 in only the last two columns, corresponding to booms in the IDA and high disbursement subsamples. Qualitatively, at least, this evidence is broadly consistent with the view that there is a greater scope for spending increases to stimulate economic activity during recessions rather than during booms. A second potential source of heterogeneity in estimated multipliers has to do with the extent of trade openness of the country. For example, simple open-economy Keynesian models imply that when the marginal propensity to import is high, a frac- tion of the increase in income due to an increase in government spending “leaks” into imports, reducing the multiplier. To consider this possibility, I calculate the average trade share of GDP for each country-decade, and then divide my sample in 30  I assess the statistical significance of these estimated differences in multipliers by estimating an equivalent specification in the full sample that includes a dummy variable indicating the two groups, and its interaction with the change in government spending. I then test the significance of this interaction term. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 201 Table 8—Heterogeneity in Estimated Multipliers Full IDA Disb/G>10% Full IDA Disb/G>10% Sample of countries (1) (2) (3) (4) (5) (6) Panel A. State of business cycle Recession Boom OLS estimate Change in government spending 0.195*** 0.186*** 0.204*** 0.101*** 0.0611 0.0796** (0.0365) (0.0457) (0.0456) (0.0326) (0.0432) (0.0384) 2SLS estimate Change in government spending 0.660* 0.614* 0.807** 0.146 0.0398 0.00873 (0.353) (0.328) (0.383) (0.265) (0.171) (0.215) First-stage F-statistic 7.40 7.99 8.01 8.02 18.64 14.76 Observations 1,312 701 919 1,492 807 1,031 Panel B. Trade openness Closed Open OLS estimate Change in government spending 0.337*** 0.274*** 0.319*** 0.281*** 0.236*** 0.243*** (0.0617) (0.0723) (0.0745) (0.0465) (0.0645) (0.0526) 2SLS estimate Change in government spending 0.634** 0.571* 0.712** 0.116 0.180 0.150 (0.295) (0.284) (0.353) (0.491) (0.328) (0.320) First-stage F-statistic 10.23 13.42 8.71 4.42 13.75 10.95 Observations 1,398 750 966 1,406 758 984 Panel C. Exchange rate regime Flexible Fixed OLS estimate Change in government spending 0.320*** 0.301*** 0.308*** 0.269*** 0.209*** 0.244*** (0.0513) (0.0649) (0.0632) (0.0487) (0.0656) (0.0498) 2SLS estimate Change in government spending 0.387 0.482** 0.320 0.306 0.188 0.450 (0.304) (0.199) (0.208) (0.371) (0.280) (0.342) First-stage F-statistic 9.55 25.54 21.88 6.03 11.46 7.25 Observations 1,009 504 592 1,795 1,004 1,358 Panel D. Aid dependence Low High OLS estimate Change in government spending 0.349*** 0.321*** 0.393*** 0.265*** 0.209*** 0.204*** (0.0603) (0.0818) (0.0716) (0.0388) (0.0538) (0.0433) 2SLS estimate Change in government spending −0.146 0.587* 0.275 0.547** 0.430* 0.438** (0.951) (0.343) (0.750) (0.224) (0.255) (0.173) First-stage F-statistic 0.82 8.21 1.26 13.22 16.22 22.92 Observations 1,373 747 970 1,431 761 980 Notes: Heteroskedasticity-consistent standard errors are clustered at the country level. All regressions are estimated using pooled country-year data and include a full set of country and year fixed effects. Changes in GDP, government spending, and disbursements are all scaled by lagged GDP. All four panels report OLS and 2SLS estimates of equa- tion (1) for various sample splits. Panel A distinguishes recessions from booms, defined as growth below/above the country-decade average. Panel B distinguishes countries less and more open to trade, defined as the decade-average trade/GDP being below/above the corresponding sample median. Panel C distinguishes countries with flexible and fixed exchange rates, defined as below/above 2 in the Ilzetzki, Reinhart, and Rogoff (2008) classification. Panel D distinguishes less and more aid dependent countries, defined as the decade-average ODA to GDP ratio below/above the corresponding sample median. The three country samples correspond to the full sample of countries, the sample of IDA-eligible countries, and the sample of countries where disbursements on loans from official creditors average at least 10 percent of government spending over the sample period. *** Significant at the 1 percent level.   ** Significant at the 5 percent level.    * Significant at the 10 percent level. 202 American Economic Journal: Macroeconomics october 2014 half at the median country-decade-average trade ratio. The estimated multipliers in the two subsamples are reported in panel B of Table 8. Qualitatively, the results here are again unambiguous—in all six specifications the estimated multiplier is larger in the more closed half of the sample. For the 2SLS estimates, the differences in esti- mated multipliers are also fairly large: in the more closed subsample, the multipliers range from 0.57 to 0.71 and are statistically significantly different from zero, while in the more open subsamples the multipliers vary between 0.12 and 0.18 and are insignificantly different from zero. While the strength of identification is respectable in several of the subsamples, these differences in estimated multipliers are however small relative to estimated standard errors, and are again not statistically significant. Yet another potential source of heterogeneity has to do with the exchange rate regime. A textbook implication of the basic IS/LM framework with limited capital mobility is that increases in government spending are more expansionary under a flexible-exchange rate regime, as the resulting depreciation of the exchange rate has an additional expansionary effect (at least when capital mobility is limited, which seems a plausible benchmark for the set of developing countries considered here). In panel C of Table 8, I use the Ilzetzki, Reinhart, and Rogoff (2008) classification of exchange rate regimes to identify country-year observations corresponding to flex- ible and fixed exchange rates.31 Qualitatively, the findings are again consistent with this basic theory, with estimated multipliers that are larger in the flexible exchange rate regime in five out of six specifications. In this case, however, the magnitude of the differences is less stark: in the flexible exchange rate regime the 2SLS estimates of the multiplier vary from 0.32 to 0.48 across country samples, while in the fixed exchange rate group the multipliers range from 0.19 to 0.45.32 Finally, in panel D of Table 8 I revisit a potential source of heterogeneity in mul- tipliers discussed in Kraay (2012), based on cross-country differences in the extent to which government spending is aid-financed. An important component of the neo- classical mechanism for a positive government spending multiplier operates through wealth effects: when spending increases, private agents are poorer by the present value of the current and future tax obligations required to finance the increase in spending. According to this mechanism, private agents react by consuming less, investing more, and providing more labor input, so that output increases. The key point here is that the magnitude of this wealth effect depends on the size of the bur- den of current and future tax obligations associated with the increase in government spending. This burden is smaller in countries where a greater proportion of govern- ment spending is financed by aid flows from abroad that do not need to be repaid, and as a result, the neoclassical mechanism predicts a smaller spending multiplier in more aid-dependent countries. 31  I classify the exchange rate regime as fixed where the IRR “coarse” measure is equal to 1 or 2, corresponding to fixed or slowly-crawling pegs, and flexible otherwise. Interestingly, this finding is the opposite of Ilzetzki, Mendoza, and Végh (2013) who work with a set of 32  developed and emerging market economies where financial openness is likely to be greater than in my developing country sample. Consistent with the predictions of the Mundell-Fleming model with capital mobility, they find that multipliers are higher under fixed exchange rate regimes in their sample of countries. They also find higher multipli- ers in countries that are more open to trade, as I do here. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 203 To investigate this possibility in this large set of developing countries, I divide each sample in half at the median level of official development assistance (ODA) as a fraction of GDP, and then estimate spending multipliers separately in the two subsamples. The OLS estimates of the multiplier display the pattern suggested by the theory, and are consistently larger in the low-aid sample. However, the differences are not large. Turning to the 2SLS estimates, in the IDA sample of countries I do again find the same pattern, with a larger multiplier in the less aid-dependent sub- sample of these IDA countries (0.59 versus 0.43 in the more aid-dependent subsam- ple). Unfortunately, however, in the full and high disbursement samples of ­countries, identification is extremely weak in the less aid-dependent groups. As a result, the multipliers are very imprecisely estimated, to the point of being meaningless, and it is not possible to draw any firm conclusions about differences in multipliers by the degree of aid dependence. Summing up, the novel DRS loan-level data on which the instrument developed in this paper is based covers a very large set of developing countries. This makes it feasible to investigate a range of plausible hypotheses regarding various potential sources of heterogeneity in government spending multipliers in developing coun- tries. This section has provided some suggestive (although not statistically signifi- cant) evidence that the short-run effects of government spending do vary with the state of the business cycle, the degree of trade openness, the exchange rate regime, and (to a limited extent) with the degree of aid-dependence of the economy, in ways that are consistent with the implications of some basic theory. VII. Conclusions In this paper, I have used a novel dataset of loan-level transactions covering lend- ing by official creditors to developing country governments to estimate government spending multipliers. My identification strategy exploits the substantial lags that occur between loan commitment and the eventual full disbursement of the loan, a process which typically takes several years. The key identifying assumption is that loan approvals, and the decision to embark on the associated spending plans, do not anticipate future shocks to growth. Given this assumption, fluctuations in disbursements that are attributable to fluctuations in past loan approval decisions are plausibly exogenous to contemporaneous shocks, and can be used as an instru- ment for fluctuations in government spending. Deploying this strategy in a large sample of developing countries, I find reasonably precise estimates of the govern- ment spending multiplier that are in the vicinity of 0.4. These results survive a range of robustness checks designed to address concerns about data quality and potential violations of the exclusion restriction. I find some evidence of heterogeneity in esti- mated multipliers that is consistent with the implications of basic theories. However, these differences typically are not statistically significant. To put these findings in context, it is useful to compare them with estimates of the government spending multiplier in the empirical literature, which has over- whelmingly been based on evidence from developed countries, most notably the United States. For the United States, Hall (2009) suggests that estimates of the fed- eral government spending multiplier are between 0.5 and 1, while Ramey (2011a) 204 American Economic Journal: Macroeconomics october 2014 suggests a somewhat higher range from 0.8 to 1.5. Moreover, many of the studies of subnational spending multipliers cited in the introduction are even higher, rang- ing from 1 to 2. Policy discussions in many countries are also often premised on the assumption that the spending multiplier is substantially above one. The smaller, but nevertheless reasonably-precisely estimated, multipliers estimated in this paper for developing countries stand in sharp contrast to this evidence, and if any­ thing, are more consistent with the modest short-run effects of government spend- ing in emerging markets uncovered by Ilzetzki, Mendoza, and Végh (2013) using VAR-based techniques. The small multipliers estimated in this paper suggest a rather limited output effect of countercyclical responses of government spending in response to economic downturns in developing countries. This finding should, however, be interpreted in light of several qualifications. First, my empirical results can only uncover evidence regarding the average short-run effects of government spending over the large set of countries and years included in my dataset, while the actual effects in particular situations might very well be different. Indeed, the limited evidence on heterogene- ity in multipliers in the last part of the paper suggests that such differences may be important in reality, even if they are difficult to isolate empirically. Second, as noted in the introduction, my empirical estimates of aggregate multipliers are not “deep” structural parameters, knowledge of which would be crucial for understanding likely future effects of any given fiscal policy response to an economic downturn. As such, my estimates are better interpreted as contributing a stylized fact on the correlation between changes in output and a plausibly predetermined component of changes in government spending that can serve as an empirical reference point for further theoretical work on this issue, particularly as it applies to developing countries. And finally, it is worth emphasizing that the absence of evidence in support of a large spending multiplier does not imply there is no role for a fiscal response to economic downturns. For example, in many developing countries, there is a strong rationale and scope for expanding social protections to the most vulnerable during economic crises, independent of any macroeconomic stimulative effects of such policies. In such a case, a countercyclical fiscal response to crises would be warranted even if it had limited short-run effects on aggregate economic growth. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 205 Appendix Table A1—Sources of Government Spending Data Country WEO CGN GFS ADI Angola 2004–2010 1985–2003 Albania 1997–2010 1987–1996 Armenia 2005–2010 1995–2004 Burundi 1994–2010 1981–1973 1993–1982 Benin 1991–2010 1985–1990 Burkina Faso 1986–2010 1973–1985 Bangladesh 1981–2010 1973–1980 Bulgaria 2001–2010 1988–2000 Belize 1996–2010 1977–1995 Bolivia 1981–2010 Bhutan 1981–2010 Botswana 1981–2010 1971–1980 Central African Rep. 1988–2010  o ​te d’Ivoire C​   1997–2010 1979–1985 1986–1996 Cameroon 2000–2010 1994–1999 1975–1993 Congo, Rep. 1990–2010 1970–1983 1984–1989 Colombia 1982–2010 1971–1981 Comoros 1984–2010 Cape Verde 2002–2010 1999–2001 1988–1998 Costa Rica 2000–2010 1972–1999 Djibouti 1991–2010 1979–1988 1990 Dominica 1990–2010 1985–1989 Dominican Rep. 2001–2010 1972–2000 Algeria 1990–2010 1970–1989 Ecuador 1995–2010 1973–1994 Egypt 2003–2010 1975–1991 1992–2002 Eritrea 1992–2010 Ethiopia 1981–2010 1972–1980 Fiji 1992–2010 1970–1991 Gabon 1991–2010 1973–1984 1985–1990 Georgia 2005–2010 1996–2004 Ghana 1981–2010 1972–1980 Guinea 1992–2010 1986–1991 Gambia, The 2000–2010 1990–1999 1973–1982 1983–1989 Guinea-Bissau 2001–2010 1983–2000 Grenada 1990–2010 Guatemala 2000–2010 1984–1999 1972–1983 Guyana 2001–2010 1986–2000 1970–1985 Honduras 2000–2010 1985–1999 Haiti 1997–2010 1972–1986 Indonesia 2000–2010 1972–1999 Jamaica 1991–2010 1975–1985 Jordan 1991–2010 1974–1990 Kenya 1983–2010 1972–1982 Kyrgyz Republic 2000–2010 1993–1999 Cambodia 1996–2010 1994–1995 St. Kitts and Nevis 1980–2010 Lao P.D.R. 2003–2010 1996–2002 Liberia 2000–2010 1974–1988 St. Lucia 1986–2010 1978–1985 Sri Lanka 1990–2010 1970–1989 Lesotho 1982–2010 Latvia 2000–2010 1995–1999 Morocco 1990–2010 1970–1989 Moldova 1991–2010 Madagascar 1980–2010 Maldives 1990–2010 1979–1989 Macedonia 1998–2010 1995–1997 (Continued) 206 American Economic Journal: Macroeconomics october 2014 Table A1—Sources of Government Spending Data (Continued) Country WEO CGN GFS ADI Mali 2000–2010 1987–1999 1976–1986 Mongolia 1981–2010 Mozambique 1980–2010 Mauritania 2007–2010 1985–2006 Mauritius 2001–2010 1973–2000 Malawi 2003–2010 1985–2002 1971–1984 Malaysia 1991–2010 1972–1990 Niger 1995–2010 1985–1994 1984 Nicaragua 2001–2010 1970–2000 Nepal 2001–2010 1972–2000 Pakistan 1994–2010 1973–1993 Panama 2009–2010 1986–2008 1974–1985 Peru 2001–2010 1970–2000 Philippines 1990–2010 1972–1989 Papua New Guinea 1992–2010 1975–1991 Paraguay 1990–2010 1972–1989 Romania 2000–2010 1980–1999 Rwanda 1994–2010 1973–1980 1981–1993 Sudan 1998–2010 1984–1997 Senegal 2000–2010 1986–1999 1970–1985 Solomon Islands 1990–2010 1975–1989 Sierra Leone 2000–2010 1985–1999 1974–1984 El Salvador 1993–2010 1970–1972 Sao Tome and Principe 2001–2010 1988–2000 Swaziland 1982–2010 1971–1981 Seychelles 1984–2010 1972–1983 Syria 1991–2010 1972–1990 Chad 1995–2010 1983–1994 Togo 1997–2010 1977–1981 1982–1996 Thailand 1995–2010 1972–1994 Tajikistan 1998–2010 1995–1997 Tonga 2000–2010 1986–1999 1980–1985 Tunisia 1992–2010 1972–1991 Turkey 2002–2010 1985–2001 Tanzania 1991–2010 1972–1985 Uganda 1997–2010 1972–1986 Uruguay 2002–2010 1972–2001 Uzbekistan 1992–2010 St. Vincent and Gren. 1980–2010 Vietnam 1998–2010 1994–1997 Vanuatu 1991–2010 1982–2010 Yemen, Republic of 1995–2010 1990–1994 Congo, Dem. Rep. 1996–2010 Zambia 2001–2010 1990–2000 1972–1989 Notes: This table shows the data sources used to construct the government spending series used in the paper, as described in Section II of the main text of the paper. The four data sources are the IMF’s World Economic Outlook (WEO), Clemens, Gupta, and Nozaki (2011) (CGN), the IMF’s Government Finance Statistics Yearbook (GFS), and the World Bank’s African Development Indicators (ADI). The table lists all countries included in the largest 1 ­ 02-country sample. References Auerbach, Alan, and Yuriy Gorodnichenko. 2012a. “Fiscal Multipliers in Recessions and Expan- sions.” In Fiscal Policy after the Financial Crisis, edited by Alberto Alesina and Francesco Gia- vazzi, 63–102. Chicago: University of Chicago Press. Auerbach, Alan, and Yuriy Gorodnichenko. 2012b. “Measuring the Output Responses to Fiscal Pol- icy.” American Economic Journal: Economic Policy 4 (1): 1–27. Vol. 6 No. 4 kraay: government spending multipliers in developing countries 207 Barro, Robert J. 1981. “Output Effects of Government Purchases.” Journal of Political Economy 89 (6): 1086–1121. Barro, Robert J., and Charles J. Redlick. 2011. “Macroeconomic Effects from Government Purchases and Taxes.” Quarterly Journal of Economics 126 (1): 51–102. Blanchard, Olivier, and Roberto Perotti. 2002. “An Empirical Characterization of the Dynamic Effects of Changes in Government Spending and Taxes on Output.” Quarterly Journal of Economics 117 (4): 1329–68. Chang, Charles, Eduardo Fernandez-Arias, and Luis Serven. 2002. “Measuring Aid Flows: A New Approach.” Global Economy Journal 3 (2–4): 197–218. Chodorow-Reich, Gabriel, Laura Feiveson, Zachary Liscow, and William Gui Woolston. 2012. “Does State Fiscal Relief during Recessions Raise Employment? Evidence from the American Recovery and Reinvestment Act.” American Economic Journal: Economic Policy 4 (3): 118–45. Clemens, Jeffrey, and Stephen Miran. 2012. “Fiscal Policy Multipliers on Subnational Government Spending.” American Economic Journal: Economic Policy 4 (2): 46–68. Clements, Benedict, Sanjeev Gupta, and Masahiro Nozaki. 2011. “What Happens to Social Spending in IMF-Supported Programs?” International Monetary Fund (IMF) Staff Discussion Note 11/15. Cohen, Lauren, Joshua D. Covall, and Christopher Malloy. 2010. “Do Powerful Politicians Cause Cor- porate Downsizing?” National Bureau of Economic (NBER) Working Paper 15839. Dias, Daniel A., Christine J. Richmond, and Mark L. J. Wright. 2011. “The Stock of External Sover- eign Debt: Can We Take the Data At ‘Face Value’?” National Bureau of Economic (NBER) Work- ing Paper 17551. Dikhanov, Yuri. 2004. “Historical PV of Debt in Developing Countries: 1980–2002.” Unpublished. Driscoll, John C., and Aart C. Kraay. 1998. “Consistent Covariance Matrix Estimation with Spatially Dependent Panel Data.” Review of Economics and Statistics 80 (4): 549–60. Fishback, Price V., and Valentina Kachanovskaya. 2010. “In Search of the Multiplier for Federal Spending on the States During the Great Depression.” National Bureau of Economic (NBER) Working Paper 16561. Fisher, Jonas D. M., and Ryan Peters. 2010. “Using Stock Returns to Identify Government Spending Shocks.” Economic Journal 120 (544): 414–36. Giavazzi, Francesco, and Michael McMahon. 2012. “The Household Effects of Government Con- sumption.” National Bureau of Economic (NBER) Working Paper 17837. Hadi, Ali S. 1992. “Identifying Multiple Outliers in Multivariate Data.” Journal of the Royal Statisti- cal Society 54 (3): 761–71. Hall, Robert E. 2009. “By How Much Does GDP Rise if the Government Buys More Output?” Brook- ings Papers on Economic Activity 39 (2): 183–250. Ilzetzki, Ethan, Enrique Mendoza, and Carlos Végh. 2013. “How Big (Small?) Are Fiscal Multipli- ers?” Journal of Monetary Economics 60 (2): 239–54. Ilzetzki, Ethan, Carmen M. Reinhart, and Kenneth S. Rogoff. 2008. “Exchange Rate Arrangements Entering the 21st Century: Which Anchor Will Hold?” Unpublished. Jordà, Òscar. 2005. “Estimation and Inference of Impulse Responses by Local Projections.” American Economic Review 95 (1): 161–82. Kraay, Aart. 2012. “How Large Is the Government Spending Multiplier? Evidence from World Bank Lending.” Quarterly Journal of Economics 127 (2): 829–87. Kraay, Aart. 2014. “Government Spending Multipliers in Developing Countries: Evidence from Lend- ing by Official Creditors: Dataset.” American Economic Journal: Macroeconomics. http://dx.doi. org/10.1257/mac.6.4.170 Leduc, Sylvain, and Daniel Wilson. 2012. “Roads to Prosperity or Bridges to Nowhere? Theory and Evidence on the Impact of Public Infrastructure Investment.” In National Bureau of Economic Research Macroeconomics Annual 2012, edited by Daron Acemoglu, Jonathan Parker, and Michael Woodford, 89–142. Chicago: University of Chicago Press. Leeper, Eric M., Todd B. Walker, and Shu-Chun S. Yang. 2010. “Government Investment and Fiscal Stimulus.” Journal of Monetary Economics 57 (8): 1000–1012. Moreira, Marcelo J. 2003. “A Conditional Likelihood Ratio Test for Structural Models.” Economet- rica 71 (4): 1027–48. Nakamura, Emi, and Jón Steinsson. 2014. “Fiscal Stimulus in a Monetary Union.” American Eco- nomic Review 104 (3): 753–92. Ramey, Valerie, 2011a. “Can Government Purchases Stimulate the Economy?” Journal of Economic Literature 49 (3): 673–85. Ramey, Valerie. 2011b. “Identifying Government Spending Shocks: It’s all in the Timing.” Quarterly Journal of Economics 126 (1): 1–50. 208 American Economic Journal: Macroeconomics october 2014 Ramey, Valerie A., and Matthew D. Shapiro. 1998. “Costly Capital Reallocation and the Effects of Government Spending.” Carnegie-Rochester Conference Series on Public Policy 48 (1): 145–94. Serrato, Juan Carlos Suárez, and Philippe Wingender. 2011. “Estimating Local Fiscal Multipliers.” Unpublished. Shoag, Daniel. 2010. “The Impact of Government Spending Shocks: Evidence on the Multiplier from State Pension Plans.” http://www.people.fas.harvard.edu/~shoag/papers_files/shoag_jmp.pdf. Staiger, Douglas, and James H. Stock. 1997. “Instrumental Variables Regression with Weak Instru- ments.” Econometrica 65 (3): 557–86. Stock, James, and Motohiro Yogo. 2005. “Testing for Weak Instruments in Linear IV Regressions.” In Identification and Inference for Econometric Models: Essays in Honor of Thomas Rothenberg, edited by Donald W. K. Andrews and James H. Stock, 80–108. Cambridge: Cambridge University Press. Wilson, Daniel. 2012. “Fiscal Spending Jobs Multipliers: Evidence from the 2009 American Recovery and Reinvestment Act.” American Economic Journal: Economic Policy 4 (3): 251–82. Wooldridge, Jeffrey. 2001. Econometric Analysis of Cross Section and Panel Data. Cambridge: MIT Press.