IMF Policy Paper: A Strategy For IMF Engagement on Social Spending—Background Papers

Background papers to the "Strategy for IMF Engagement on Social Spending"

Abstract

Background papers to the "Strategy for IMF Engagement on Social Spending"

Background Paper II. Impact of IMF Programs on Social Spending: Empirical Evidence1

This paper analyzes whether levels of public education and health spending have been safeguarded in IMF-supported programs. The analysis addresses various methodological and data challenges present in existing studies. It confirms the findings of earlier studies: on average there is no difference between spending trends in program countries compared to similar countries without a program. However, in a significant number of instances spending decreased in program countries. High GDP growth increases the probability of a decline in spending as a share of GDP but lowers the probability of a decline in real per capita spending. The probability of a decline is greater where the magnitude of short-term fiscal consolidation is high, and especially where consolidation is achieved through expenditure reductions. Therefore, strengthening measures that reinforce growth and revenue mobilization can help to avoid short-term declines in social spending. Declines are much more likely where initial spending is high, which may reflect spending inefficiencies.

A. Introduction

1. The IMF has long recognized the importance of protecting social spending in -all Fund-supported programs (“programs” thereafter), particularly in low-income countries (LICs). A significant step in this direction was the introduction in 1999—jointly with the World Bank—of the Poverty Reduction Strategy Paper (PRSP) approach to assist LICs with the development of their poverty reduction strategies. This approach was supplemented by the transformation of the Enhanced Structural Adjustment Facility (ESAF) into the Poverty Reduction and Growth Facility (PRGF), which strengthened the inclusion of pro-poor growth considerations into the design of Fund-supported programs (“programs” hereafter). The design of the IMF’s concessional lending was further reformed in July 2009 (effective January 7, 2010) by replacing the PRGF with the Extended Credit Facility (ECF) supported by the Poverty Reduction and Growth Trust (PRGT) that enhanced the monitoring of social and other priority spending, including through incorporating explicit performance criteria (such as social spending “floors”) into program design. Programs under the General Resources Account (GRA) also include spending indicative targets (ITs) involving funding to specific social assistance programs, while they more often rely on structural conditionality to strengthen social safety nets (IMF 2017, pp. 30).

2. This paper intends to answer two important questions:

  • Have IMF-supported programs safeguarded the level of public education and health spending?2

  • What is the effect of IMF-supported programs on public education and health spending, relative to the counterfactual under which they had not engaged in a program?

B. Data and Measurement

3. Government spending on education refers to all levels of education.3 The data are collected by the UNESCO Institute for Statistics and are mapped to the International Standard Classification of Education (ISCED) using the method adopted in 2011 by the UNESCO General Conference to ensure the comparability of education programs at the international level. Total general government spending on education captures current and capital spending, and includes spending funded by transfers from international sources to government.

4. Public health spending is defined as the sum of domestic general government health spending and external health spending channeled through government.4 This definition refers to current health spending given the lack of a decomposition of capital spending into public and private components in the WHO’s updated database that uses the framework of System of Health Accounts 2011 (SHA 2011). According to the WHO (2017), current government spending on health financed from domestic sources provides a more precise measure by which to evaluate health policy analysis as capital spending tends to fluctuate and does not finance access to health services.5 Adding external transfers channeled via the government is useful in our case, as it is of interest whether IMF programs catalyze donor assistance. Existing studies have typically used total public health spending data that were previously compiled by the WHO using the National Health Accounts approach, which did not distinguish between current and capital spending. The SHA 2011 improved on the previous methodology by classifying country-specific health spending financing flows in a uniform way to produce comparable results, increasing in this way accuracy in the tracking of spending.

5. Three measures are used to evaluate the impact of a program. The paper focuses on the evolution of spending in countries with programs approved after 2000 using data over the period 2000–2016: (i) for health spending, and (ii) for education spending.6 Government spending on health and education can be expressed in real per capita terms and as a share of GDP or total government spending. To comprehensively evaluate the impact of programs on spending in health and education we use all three measures in our analysis. The pros and cons of each measure are:

  • Spending as a share of GDP helps assess whether spending fluctuates in line with general economic conditions. However, changes in this measure could simply reflect changes in GDP rather than spending levels. Thus, spending as a share of GDP could increase even when real spending declines because GDP declines by a greater proportion.

  • Real per capita spending allows for comparison of the level of resources allocated by the government to these sectors. Nevertheless, real per capita spending estimates can suffer from measurement error as population data are revised infrequently and nominal spending is often deflated using a GDP deflator in the absence of sector-specific deflators. Moreover, as is the case with all three indicators, real spending could simply reflect changes in wages rather than changes in service provision, although competitive wages are required to avoid staffing problems.

  • Spending in percent of total government spending can be used to evaluate whether it is protected relative to other spending. However, increases in the share of government spending allocated to social sectors could coincide with declines in overall spending when spending reductions in these sectors are less pronounced.

6. Approval and expiration dates of programs are collected for all arrangements approved and ongoing during the period 2000–2016. Appendix Table 1 provides an overview of the 283 programs approved over the period 2000–2016. Blended programs that combine PRGT with GRA resources are treated as PRGT programs in the descriptive and empirical analyses that follow, with the program length being the longer duration among both program types. The sample includes 123 GRA and 160 PRGT programs, of which 13 are blended arrangements and 18 are Policy Support Instrument (PSIs).7

C. Social Spending Trends During Programs

7. Both health and education spending are, on average, protected during program years. This is demonstrated using boxplots of trends in health and education spending during program years expressed as changes in their share of GDP and government spending, as well as percent changes in real per capita spending (Figure 1). Changes in spending in a program year are calculated relative to spending in the year prior to the program approval date. The analysis indicates that changes are generally positive (whether focusing on mean or median change), with a notable exception being the reduction in education spending as a share of GDP and as a share of spending observed in countries with GRA programs. Reassuringly, however, real per capita education spending is on average protected across facilities, suggesting that GDP increases may drive downward the median change in education spending as a share of GDP—even when real spending is increasing. The change in the share of spending on each sector as a share of total government spending provides a measure to evaluate the priority each government gives to that sector over others. On average, using this measure, health spending has been increasing in programs whereas there is some evidence that education spending as a share of total spending fell in GRA programs.

Figure 1.
Figure 1.

Trends in Public Spending on Health and Education during an IMF Program

Citation: Policy Papers 2019, 017; 10.5089/9781498318907.007.A002

Sources: WHO, UNESCO, and IMF staff calculations.Note: Bars represent cross-country interquartile ranges of changes in spending in any program year after the approval date relative to spending in the year prior to the approval date. Only countries with observations in the year before approval and during program years are included.

8. While spending was protected on average, it declined in over a quarter of program years (Figure 1). This holds for both education and health spending and regardless of the measure of spending used. It raises the questions of what factors are behind these declines and, where warranted, how program design could be further strengthened to avoid them.

9. For real per capita education and health spending, there is evidence that, on average, spending continued to increase over the course of the program. A simple regression of spending changes in program countries on dummy variables reflecting the program years, also controlling for country fixed effects, suggests that on average real per capita health and education spending increases were higher in the later years in the program relative to the approval year (Figure 2). Similar regressions for spending expressed as share of GDP or total government spending do not point to statistically significant differences relative to the approval year, except for programs that lasted five years which had significant positive changes in health spending as a share of GDP.

Figure 2.
Figure 2.

Comparative Trends in Public Spending on Health and Education

Citation: Policy Papers 2019, 017; 10.5089/9781498318907.007.A002

Sources: WHO, UNESCO, and IMF staff calculations.Note: Bars represent 90 percent confidence intervals of point estimates from regressing changes in spending during a program controlling for country fixed effects and dummy variables representing each year in the program.

10. What factors contribute to declining social spending in program countries? To empirically explore this question, we use probit regressions where the dependent binary variable equals one for large declines in spending, with large defined as spending decreases greater than the median decline in spending when measured as shares of GDP and in real per capita terms. Appendix Table 2 presents the range of spending values captured in each dummy variable. Macroeconomic conditions can affect the probability of large changes in social spending. To capture the broader economic developments that may trigger large drops in social spending, we include as regressors: (i) real GDP growth and inflation (CPI index growth) observed during program years since nominal and real changes in GDP are expected by construction to reduce spending ratios; and (ii) proxies for the evolution of fiscal space during program years as reflected in the changes in cash balances and revenues in percent of GDP.8 Initial health and education spending in percent of GDP are also added as a regressor to explore whether countries with higher spending tend to have lower changes during program years. Figure 3 shows the average marginal effects of an increase in the regressors from their 25th to their 75th percentile value in the sample (see Appendix Table 3 for underlying coefficient estimates).

Figure 3.
Figure 3.

Average Marginal Effects on Probability of Large Declines in Social Spending

Citation: Policy Papers 2019, 017; 10.5089/9781498318907.007.A002

Sources: WHO, UNESCO, and IMF staff calculations.Note: Bars represent of average marginal effect estimates from a probit model of large negative changes in spending during a program when each regressor ranges from the 25th to 75th percentile.

11. Results suggest that higher real GDP growth is positively correlated to the likelihood of large drops in education spending when the latter is expressed as a share of GDP (Figure 3). Controlling for inflation, higher real GDP growth translates into higher nominal GDP growth that is (mechanically) negatively correlated to the ratio of spending over GDP. This correlation points to the advantages of also assessing spending changes in real per capita terms. We note that higher real GDP growth is associated with a lower probability of a sharp decline in real social spending per capita. Similarly, the probability of a sharp decline is higher for countries with higher inflation and therefore also higher growth in nominal GDP. Measures that reinforce growth can therefore help to protect (or even enhance) real social spending.

12. The probability of a decline in spending (both as a share of GDP and in real per capita terms) is also greater where the magnitude of short-term fiscal consolidation is high. The negative impact of revenue increases on the probability of a decline indicates that this effect is attenuated the more fiscal consolidation (i.e., improvements in the fiscal balance) is achieved through enhanced revenue mobilization as opposed to through expenditure consolidation. Therefore, program measures that strengthen revenue mobilization (e.g., stronger revenue administration and higher taxes) can help to avoid short-term declines in social spending.

13. Declines in spending are much more likely where initial spending is high (Figure 4). High spending can be associated with spending inefficiencies (such as high wages or other costs) and addressing these inefficiencies can reduce the likelihood of decreases in efficient spending components. For instance, reforms in Greece during the Troika programs explicitly aimed at lowering the cost of pharmaceuticals and other medical goods, as well as containing wages (October 2013 Fiscal Monitor; pp. 50).

Figure 4.
Figure 4.

Probability of Large Declines in Social Spending

Citation: Policy Papers 2019, 017; 10.5089/9781498318907.007.A002

Sources: WHO, UNESCO, and IMF staff calculations.Note: Probit model probability predictions of large negative changes in spending during a program.

D. The Impact of IMF Programs on Social Spending

Previous studies

14. Most previous empirical studies on the effect of IMF-supported programs on social spending find that social spending trends have been on average similar to those in comparable non-program countries (Appendix Table 4). The availability of comparable data on health and education spending for a broad sample of countries has facilitated the empirical examination of the average impact of programs across countries. However, credibly estimating the effects of programs on social spending is challenging. The fundamental empirical hurdle is that the correlation between the presence of a program and the size of government spending could reflect causation in either direction, or the effect of a third factor influencing both the likelihood of a program and the level of spending. The literature has applied a variety of empirical strategies to overcome these identification concerns that often include the system-GMM and Heckman correction approaches. These methods typically rely on exclusion restrictions to tackle program endogeneity requiring variables that are strongly correlated with the likelihood of having a program but not correlated with government spending on health and education. Appendix Table 4 critically discusses the findings of the main papers in this area. While most studies find a significantly positive or no average impact on education and health spending in program countries, although two studies find a negative impact for health.

15. The strategies of existing studies to address the identification problem suffer from various shortcomings. These studies have attempted to assess the causal effect of IMF programs, while controlling for variables measured after the program was approved assuming these can also affect social spending directly—such as fiscal balances, per capita GDP, and lagged dependent variables. However, IMF programs can have an indirect impact on social spending through their impact on these variables, so that controlling for such intermediate outcomes confounds the estimate on the total impact of IMF programs and complicates the interpretation of the IMF program dummy. Another important methodological shortcoming relates to the use of variables that do not satisfy the exclusion restrictions required to identify the average effect of programs on spending. For example, proxies for external conditions—exchange rates, reserves, and exchange rate regimes—are likely correlated with fluctuations in the cost of imported goods used in the social sectors (such as drug costs), and therefore are not suitable as instrumental variables. Other strategies rely on instruments such as the total number of countries with programs, and United Nations General Assembly voting similarity with key donor countries to proxy for the Fund’s willingness to lend. However, the number of programs in a given year could also be a proxy for omitted variables (such as commodity price shocks or international financial crises) that are positively correlated to the probability of having a program. Such shocks could both increase the likelihood of a specific country requesting a program and the total number of countries that are approved for a program. More importantly, these shocks could affect government financing and should therefore also be included in the spending equation.9 Similarly, political proximity to the key IMF shareholders is significantly correlated with the degree of openness and other characteristics of member countries that in turn are correlated with the level of government spending yet are excluded from the spending equation.10 Empirical strategies that make use of such instrumental variables risk biasing their results in unpredictable directions.

16. Using the same empirical strategy as existing studies and the data constructed for this paper for the period 2000–2016, this paper finds that spending trends are not statistically different in program and non-program countries. Despite the shortcomings of the Heckman and system-GMM approaches used by previous studies, for comparison purposes we present updated estimates over the period 2000–2016 for PRGT-eligible countries noting that health spending data in this paper refer to current spending rather than total spending as used in the studies reviewed in Appendix Table 4.11 For education spending, the results point to sign differences across the Heckman and system-GMM estimates, with a significant positive impact found in the former and an insignificant negative impact in the latter (Appendix Table 5). For health spending, both approaches yield negative—albeit statistically insignificant—estimates of the program impact. But, in general, the results suggest that it is hard to reject the hypothesis that spending trends have on average been similar in both program and non-program countries.

A new approach to measuring the impact of IMF-supported programs on social spending

17. Doubly robust estimators can address some of the shortcomings of previous identification strategies. Essentially, under this approach, non-program countries that look similar to program countries in terms of pre-program macroeconomic conditions are assigned a higher weight, based on their probability of being in a program, when estimating differences in average spending trends between program and non-program countries. These estimators rely on inverse propensity weighting to proxy random allocation in a two-stage procedure, where the probability of a country engaging in a program is estimated in a separate first stage regression (participation equation) that controls for pre-treatment observable sources of endogeneity (Box 1). Pre-treatment characteristics include: (i) real economy variables to proxy for the comparable level of development and growth in the year prior to program approval (e.g., per capita GDP in purchasing power parity terms and real per capita GDP growth); (ii) proxies for the fiscal pressures in the economy captured by the cash balance and government debt in percent of GDP; (iii) external sector proxies, including reserves in months of imports, external debt in percent of GDP, as well as trade and capital account balances; and (iv) eligibility for concessional lending (PRGT dummy variable), which may render IMF lending an attractive alternative to other sources of financing. In the second stage regression (outcome equation), observations are weighted inversely to the estimated probability of engaging in a program, thus allocating greater weight to those that mimic a random allocation. Year fixed-effects in this second stage can control for cohort specific effects in the outcome equation where the dependent variable is changes in spending in that year relative to the year before the program approval. The initial spending in percent of GDP is also added as a regressor in an alternative set of the estimations to explore whether countries with higher spending tend to have lower changes during program years.

18. The construction of the dummy variable characterizing the presence of an IMF-supported program is critical. Previous studies do not differentiate across arrangements approved for each country that overlap in a given year. As a result, the estimates may be driven by those countries that have longer-term engagement with the Fund, as greater variation in spending is likely observed in programs with longer duration. In what follows, each lending arrangement is considered as a separate program given that its objectives or conditionality may differ from the previous one. Constructing the program dummy in this way significantly increases the sample size of program cases, providing further insight into the heterogeneity of outcomes within the typical duration of arrangements that range between 2–5 years.

19. The dataset assigns all program cases to the “treated” group, while countries not having an IMF program form the “control” group. In our regressions all programs that have completed at least one year of the program following the program approval are included. The dependent variable captures changes in social spending in a program year, other than the program approval year, relative to the year prior to approval. The control group is constructed by including all countries that did not have a program during the time interval we are considering. For example, the sample of program and non-program countries in the case of programs that lasted two years is constructed by first dividing the period 2000–2016 into rolling three-year periods, where t-1 refers to the year prior to the program approval (Figure 5). The treated group contains countries with IMF programs that lasted two years (i.e., an IMF program dummy equal to one in the years t and t+1), whereas all countries that did not undergo any IMF program over the course of three consecutive years (i.e., including the pre-approval year) form the control group (i.e., an IMF program dummy equal to zero for all years considered). Treated countries could have another IMF arrangement in the year prior to the approval of the arrangement under consideration since we consider consecutive programs as two separate programs. Pre-treatment characteristics in period t-1 for the treated and control group are included as regressors in our participation and spending equations. The full sample is constructed by repeating this process for all available durations of IMF programs over 2000 to 2016.

Inverse Probability-Weighted Regression Adjustment (IPWRA) Estimation

The IPWRA estimator provides a “doubly robust” approach in estimating the effect of programs. The IPWRA estimator includes both outcome and participation (or treatment) equations to account for the non-random treatment assignment. Inverse-probability weighting is applied to estimate corrected regression coefficients that are subsequently used to perform regression adjustment. By modeling both the outcome and the treatment probability, the method is robust to misspecification of at most one of the underlying outcome or participation equations (Wooldridge, 2007; Wooldridge and Słoczyński, 2018). This is contrary to the Heckman approach that requires specific distributional assumptions, and at least one selection variable not affecting the outcome equation. We estimate the average treatment effect on the treated (ATT), i.e., the effect of the program on social spending of those countries that receive the treatment (the control group consists of all countries for which there is no program over a similar duration):

τIMF=E(ΔGIMFΔGnoIMF\DIMF=1)=1P(DIMF=1)E[DIMFΔGIMF1DIMF1p(X)p(X)ΔGnoIMF](1.1)

where τIMF refers to the average impact of the IMF program on social spending; ΔτIMF denotes the change in social spending observed in year t of the program relative to the year prior to program approval; ΔτnoIMF refers to the counterfactual outcome that would be observed if the country had not received an IMF program; DIMF is a binary indicator for treatment with an IMF program; X is a vector of observed pretreatment regressors that predict participation in an IMF program and have explanatory power for the counterfactual change observed during an IMF program; P(DMF =1)is the unconditional probability of participating in an IMF program; p(X) is the propensity score of participating in an IMF program. Identification of the IMF program effect on the treated countries relies on the assumption of unconfoundedness between the treatment assignment and the counterfactual changes in social spending in the control state, ΔτnoIMF. The unconfoundedness assumption requires that conditional on the regressors, the treatment and potential outcomes are independent.

Figure 5.
Figure 5.

Sample Construction Example

Citation: Policy Papers 2019, 017; 10.5089/9781498318907.007.A002

Note: Dataset construction of the treated and control groups considered for IMF programs with duration of two years. Cells contain the values of a dummy variable indicating the presence of a specific IMF arrangement over the period of interest.

20. The first stage estimates suggest a correct specification of the program participation equation. A desirable property of the “doubly robust” estimators is their flexibility in obtaining consistent results if either the linear model for potential outcomes or the probit model used for program participation is valid. The null hypothesis that the first stage model balances the covariates cannot be rejected, thus lending support to the specification of the program participation equation. The macroeconomic variables used in the first stage have the expected sign and most of them are statistically significant, with more developed and faster growing countries being less likely to engage in a program. On the other hand, countries with weaker fiscal balances, lower foreign reserves, weaker trade positions and higher external debt are more likely to request an IMF program.

21. On average, Fund-supported programs do not appear to have a statistically significant effect on social spending trends. The estimation results using the IPWRA estimates of the impact of the program suggest that the point estimate is statistically insignificant across the three years considered after program approval in the sample of 101 lending arrangements with duration of at least two years over the period of 2000–2016 (Appendix Tables 6 and 7). There is weak evidence suggesting that the impact on health spending as a share of GDP is greater in program countries where initial spending is relatively low (Columns 2 and 4; Table 7). Similar insignificant effects are found when education spending is considered in a sample with 81 program cases, with the point estimate having a negative sign in the early years of the program, which turns positive for those programs with longer duration when initial spending is considered (Appendix Tables 8 and 9).

E. Main Implications and Conclusions

22. Consistent with previous studies, this study finds that while education and health spending has been protected on average in programs, spending decreases are sizeable in a large share of program countries.

  • On average, education and health spending has not declined in program countries. This holds regardless of the measure of spending used and for both GRA and PRGT programs. When spending is measured in real per capita terms, there is evidence that on average spending continues to increase over the life of the program.

  • However, spending declined in over one-quarter of countries with the reduction being large in some cases. Again, this holds regardless of the measure of spending used. This raises the issue of the factors behind these decreases and the potential for strengthening program design to prevent declines where warranted.

  • On average, spending on education and health in program countries are similar to those in otherwise comparable non-program countries. New econometric analysis undertaken in this paper aims at addressing shortcomings of existing studies related to the identification of program impact and the definition of programs. The results of the new analysis confirm those of previous studies that, on average, spending trends are similar in program and non-program countries. There is also evidence that spending is better protected in countries with relatively low initial spending levels.

References

Appendix I. Selected Tables

Appendix Table 1.

IMF Lending Arrangements Approved, 2000–16

(number of arrangements)

article image
Sources: IMF Monitoring of Fund Arrangements database and IMF Financial database.Note: Data refer to the arrangements approved over the period 2000–16.
Appendix Table 2.

Binary Variable for Large Declines in Health and Education Spending

article image
Sources: WHO; UNESCO; and IMF staff calculations.Note: Dummy variable construction for country-years with declines in spending that are greater than the median negative changes observed in health and education spending.
Appendix Table 3.

Probit Analysis of Large Reductions in Social Spending

article image
Source: IMF staff calculations.Note: Standard errors are shown in parentheses. Three, two, one asterisks indicate significance levels of 99 percent, 95 percent, and 90 percent, respectively. Panel A shows the results from the probit regression of the dummy of a reduction in large social spending over the period (t-h) and (t). Regressors include real GDP growth and inflation, initial social spending, and changes over the period in revenue and fiscal balances as shares of GDP. Panel B refers to the average marginal effects of increasing the regressors from their 25th to their 75th percentile value in the sample.
Appendix Table 4.

Summary of Studies on the Impact of IMF Programs on Social Spending, 2010–2018

article image
article image
article image
Appendix Table 5.

Summary of Regression Results of the Effect of IMF Programs on Social Spending Replicating Past Methodologies

article image
Source: IMF staff calculations.Note: Standard errors are shown in parentheses. Three, two, one asterisks indicate significance levels of 99 percent, 95 percent, and 90 percent, respectively. In the fixed effect estimation, the inverse Mills ratio variable controls for selection bias (the tendency of countries with macroeconomic imbalances to have Fund-supported programs). This is calculated from a probit regression of the Fund program dummy on lagged Fund program dummy, government balance (% of GDP, lagged), international reserves (in months of imports, lagged), and the exchange rate to the U.S. dollar (% change, lagged), and an index of exchange rate regime (lagged). In the system GMM estimation, real GDP per capita and the government balance are assumed to be endogenous and instruments include only one lag of endogenous variables as well as international reserves (in months of imports, lagged), and the exchange rate to the U.S. dollar (% change, lagged), and an index of exchange rate regime (lagged).
Appendix Table 6.

Government Expenditure on Health (in Percent of GDP)

article image
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Appendix Table 7.

Government Expenditure on Health (in Constant NCU per capita)

article image
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Appendix Table 8.

Government Expenditure on Education (in Percent of GDP)

article image
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
Appendix Table 9.

Government Expenditure on Education (in Constant NCU per capita)

article image
Robust standard errors in parentheses*** p<0.01, ** p<0.05, * p<0.1
1

Prepared by Emmanouil Kitsios and Baoping Shang. Research assistance was provided by Nghia Piotr Le.

2

The focus on education and health spending reflects the absence of reliable time-series data on social protection spending, especially in emerging and developing economies. Data limitations also prevent an empirical analysis on the quality of spending, as captured by changes in actual service delivery and social outcomes.

3

The focus on all levels reflects data availability. Subject to improvements in data availability, future analysis could analyze trends in basic education (primary and secondary).

4

Health spending, henceforth, refers to all levels of healthcare provision. Future analysis could explore basic health package spending, subject to data availability.

5

Capital spending rather improves future resilience of the health sector.

6

The year 2000 is the starting year in the current Global Health Expenditure Database provided by the WHO.

7

For the purposes of this study, LIC facilities—the Structural Adjustment Facility (SAF), Enhanced Structural Adjustment Facility (ESAF), and Poverty Reduction and Growth Facility (PRGF)—available prior to the establishment of the PRGT in 2009 are also labeled as PRGT-supported programs. The PSI offers LICs Fund support without a borrowing arrangement.

8

Similar results were obtained in robustness checks that included the debt-to-GDP ratio among the control variables.

9

Even if the validity of the exclusion restrictions were established, the studies reviewed in this note do not correct standard errors for using predicted regressors. Thus, the statistical significance of the estimates obtained in the outcome equations is not appropriately determined.

10

Such measures of political proximity to the U.S. have been found strongly correlated to democratization, financial openness, and government ideology (Bailey, Strezhnev, and Voeten; 2017).

11

The updated estimates are presented in the last four columns of Appendix Table 5.

A Strategy for IMF Engagement on Social Spending: Background Papers
Author: International Monetary Fund. Fiscal Affairs Dept., International Monetary Fund. Strategy, Policy, &, and Review Department