## wpiea2019245-print-pdf

## Source details

**Canonical URL:** [wpiea2019245-print-pdf](https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019245-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2019/wpiea2019245-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2019/wpiea2019245-print-pdf.pdf.json)

---

### I. Purpose and research questions
- Fund resources are to be made available temporarily and under adequate safeguards to help members address balance of payments (BoP) problems.
- Core research questions:
  - What distinguishes programs that are not followed by a successor arrangement from those that are?
  - Which country-specific, global, program-specific, and IMF-history variables relate to the probability of (no) repeated use of IMF GRA resources?
  - Can (no) repeated use be predicted out-of-sample?

### II. Sample, definitions, and measurement
- Sample coverage:
  - 890 GRA-supported programs starting between January 1952 and January 2012.
  - Program types: 700 Stand-By Arrangements (SBA), 81 Extended Fund Facilities (EFF), 109 First-Credit Tranche Arrangements (FCTA).
  - Programs requested by 138 IMF member countries.
  - Number of programs per country: range 1 to 23; average 6.4 programs per country.
  - Exclusions: precautionary arrangements without disbursements; certain country cases (Cuba, former Czechoslovakia, former Serbia & Montenegro, former Yugoslavia).
- Dependent variable:
  - Binary indicator equal to 1 if an arrangement is not succeeded by another arrangement within a threshold period; 0 otherwise.
  - Baseline threshold period: three years after expiration or cancellation.
  - Alternative threshold tested: five years (reduces unconditional success rate from 22 percent to 18 percent).
- Variable groups:
  - (i) Domestic conditions: GDP growth, primary balance, public debt, current account balance, international reserves (initial level and change), trade openness, exchange rate regime, inflation, institutional quality (bureaucracy quality index).
  - (ii) Global environment: global economic growth, world real interest rates (U.S. federal funds rate adjusted by CPI inflation), market volatility (VIX).
  - (iii) IMF- and program-specific: fraction of previous five years under an IMF program, long-run share of successful predecessors (share of past programs without successors), outstanding loans as share of quota, quota share, exceptional access dummy, program duration dummies (short/long), blended arrangements.

### III. Stylized facts (unconditional and descriptive)
- Unconditional probability of 'no successor arrangement' (sample average): approximately 22 percent.
  - By program type: SBA 22 percent, EFF 25 percent, FCTA 17 percent.
- Concentration:
  - Around 30 out of 138 countries had more than 10 GRA arrangements since 1952.
  - Top 20 percent of countries in terms of number of programs account for approximately 50 percent of all GRA programs.
- Key unconditional mean differences (programs without successors vs with successors):
  - Average real growth: about 1 percentage point higher for programs without successors.
  - Initial conditions more favorable for programs without successors: lower public debt-to-GDP, larger current account balance (or smaller deficit), larger stock of reserve assets at program start.
  - Programs without successors on average show stabilized debt, improved current account balance, improved primary balance, and increased reserves over the program period.
  - Trade openness higher and institutional quality better on average for programs without successors.
  - Inflation, exchange rate adjustment, and exchange rate regime do not differ significantly on average between the two groups.
- Global variables:
  - No clear-cut differentiation between arrangements with and without successors for world growth and market uncertainty; global interest rates somewhat lower around programs without successors.
- IMF/program variables:
  - Programs without successors tend to have: larger quota share, spent a smaller fraction of the previous five years under a program, and higher share of past successful programs.
  - Exceptional access cases have an unconditional probability of 'no successor' that is six percentage points higher than normal access counterparts (36 exceptional access cases in sample).
  - Blended arrangements in sample: 23.

### IV. Econometric approach and handling missing data
- Model: probit binary response linking probability of no successor to explanatory variables.
- Missing data strategy: Conniffe and O’Neill (2011) “probit-miss” approach under the missing at random (MAR) assumption.
  - Steps: regress x2 on x1 by OLS (complete cases); estimate probit on complete cases; estimate probit on incomplete cases using x1 only; combine outputs for efficient estimator and covariance matrix.
  - Limitation: original method assumes same missing pattern across incomplete observations; dataset violates this strictly, but most variables available and missingness concentrated in earlier years, so efficiency gains still achieved.
- Goodness-of-fit and predictive performance metrics:
  - Pseudo R-squared (Efron, 1978) for fit.
  - AUROC (area under the ROC curve) for classification/predictive accuracy; AUROC > 0.5 indicates better-than-random forecasting.

### V. Key estimation results (baseline and subsamples)
- Baseline (country-specific variables):
  - Average growth during and after the program: positive and significant.
    - Average marginal effect (AME) of growth: around 0.7 (an increase in average growth by one percentage point corresponds to a 0.7 percentage points higher probability of avoiding repeated use).
  - Current account balance at program start: positive sign and statistically significant.
  - Change in current account over program: positive sign (less consistently significant in full sample).
  - International reserves (initial level and change over program): positive and economically meaningful; initial reserves more statistically significant than adjustment.
  - Trade openness: positive association with avoiding repeated use.
  - Public debt level and its adjustment: not statistically significant in baseline.
- Global and IMF/program-specific variables in baseline:
  - Global GDP growth and real U.S. interest rate: generally not significant in full sample baselines.
  - IMF-related variables: loans outstanding (share of quota) small marginal effect; share of previous five years under a program shows negative sign (not always significant), suggesting short-run persistence of program use.
- Subsample starting in 1990:
  - Improved goodness-of-fit.
  - AME of domestic growth more than doubled relative to full sample (growth more decisive to prevent repeated use).
  - Current account adjustment over program period gains importance.
  - World real interest rate shows anticipated negative sign (higher real interest rates associated with higher probability of repeated use) in the 1990 subsample.
  - Program duration dummies: longer (shorter) than planned programs associated with higher probability of repeated use (AMEs imply about 10 and 12 percentage points respectively), though coefficients remain statistically insignificant.
- Institutional quality and governance (subsample since 1990):
  - Bureaucracy quality index (ICRG, scale 0–4): higher bureaucracy quality significantly increases probability of no successor.
    - AME for bureaucracy around 12 (implying a one-unit improvement raises probability of no repeated use by about 12 percentage points on average).
    - Example: improving bureaucracy score from 2 to 4 would increase probability of no repeated use by around 24 percentage points (ceteris paribus).
  - Price stability (inflation) not significant.
- Country’s long-run track record with Fund:
  - Share of past programs without successors over the entire past: positive and statistically significant (1 percent level), supporting that past avoidance of repeated use predicts future avoidance.

### VI. Robustness checks (summary)
- Tests performed: adding quota share, VIX, program type dummy (EFF), alternative governance indicators (government stability, corruption), external debt-to-GDP, interaction terms, time trend, excluding FCTAs, blended arrangement dummy, varying threshold period for no successor.
- Main findings robust:
  - Quota share and VIX not significant.
  - Program type dummy (EFF) not significant.
  - Government stability significant in some specifications; corruption index not significant.
  - External debt-to-GDP not significant.
  - Interaction terms and linear time trend do not alter main conclusions.
  - Excluding FCTAs or including blended dummy does not materially change baseline results.
  - Extending threshold from 3 to 5 years reduces unconditional success rate from 22 percent to 18 percent; current account adjustment gains importance and reserves lose significance; world real interest rate becomes significant.

### VII. Out-of-sample forecasting (design, limitations, and stylized results)
- Design:
  - Split-sample forecasting: estimate probit on an estimation subsample, predict on a later prediction subsample, dropping overlapping programs to mimic forecasting environment.
  - In practice, realized values of explanatory variables during the prediction period are used (limitation for true ex-ante forecasting).
  - AUROC used to evaluate predictive performance; statistical test of H0: AUROC = 0.5 versus H1: AUROC > 0.5 reported.
  - Classification threshold chosen to minimize number of falsely predicted outcomes (simple unweighted loss).
- Limitations:
  - Use of realized predictor values instead of genuine ex-ante forecasts may overstate achievable forecasting accuracy.
  - For averaged explanatory variables computed over differing periods for programs with and without successors, a forecaster would need to know succession timing ex ante to compute correct averages; infeasible in practice.
- Stylized forecasting exercise results:
  - Out of the total 52 programs to be predicted, the model correctly calls the outcome of 35.
  - The model correctly predicts the majority of:
    - ‘no successor’ cases: 79 percent
    - repeated use cases: 57 percent
  - AUROC value: 0.6994.
  - Focused exercise (estimate on 1990–2002, predict 2005 onwards) achieves a share of 77 percent correctly predicted outcomes and AUROC 0.7870 (p-value: 0.00) when institutional quality is included.
  - Parsimonious probit specifications tend to have superior out-of-sample forecasting performance relative to larger models.

### VIII. Main policy-relevant takeaways
- Domestic macroeconomic conditions matter most for avoiding repeated use of IMF GRA resources:
  - Stronger average growth, better current account position and improvement during programs, and larger international reserves correlate with higher probability of no successor.
- Institutional capacity is important:
  - Higher bureaucracy quality materially increases probability of avoiding repeated use.
- Country history with the Fund is informative:
  - Countries that avoided repeated use in the past are more likely to avoid it again.
- Global environment and some program-specific features:
  - Evidence on the role of global variables is ambiguous and varies over time; IMF-specific variables generally less influential once domestic conditions and institutions are controlled for.
- Predictability:
  - Parsimonious models including the most important determinants can have reasonable out-of-sample predictive power, but accurate ex-ante forecasts require realistic assumptions about macroeconomic variables and program developments.

### IX. Recommendations and avenues for future research
- Potential extensions and improvements:
  - Extend the analysis to concessional lending programs under the PRGT, noting that repeated use would need to be redefined to reflect the longer-term nature of PRGT arrangements.
  - Evaluate in more detail the role of program-specific characteristics, for example:
    - Define a measure of realism of program targets and investigate the impact of over-ambitious targets on the probability of repeated use by comparing program targets to benchmarks based on country-specific or cross-country experiences.
  - Improve the stylized forecasting exercise by:
    - working with data vintages to properly mimic a true out-of-sample forecasting situation, and
    - including measures of forecasting uncertainty surrounding the point estimates.

*Source: Annexes and References of wpiea2019245-print-pdf*

### References27

### wpiea2019245-print-pdf - References27

### References
- Section title: "References"
- Page indicator: 27

### Annex 1: Figures and Tables
- Section title: "Annex 1: Figures and Tables"
- Page indicator: 28

*Source: wpiea2019245-print-pdf - References27*

### Annex 2: Other Dataset Figures and Tables39

### Annex 2: Other Dataset Figures and Tables39

### I. Purpose and research questions
- Fund resources are to be made available temporarily and under adequate safeguards to help members address balance of payments (BoP) problems.
- Core research questions:
  - What distinguishes programs that are not followed by a successor arrangement from those that are?
  - Which country-specific, global, program-specific, and IMF-history variables relate to the probability of (no) repeated use of IMF GRA resources?
  - Can (no) repeated use be predicted out-of-sample?

### II. Sample, definitions, and measurement
- Sample coverage:
  - 890 GRA-supported programs starting between January 1952 and January 2012.
  - Program types: 700 Stand-By Arrangements (SBA), 81 Extended Fund Facilities (EFF), 109 First-Credit Tranche Arrangements (FCTA).
  - Programs requested by 138 IMF member countries.
  - Number of programs per country: range 1 to 23; average 6.4 programs per country.
  - Excluded: precautionary arrangements without disbursements and certain country cases (Cuba, former Czechoslovakia, former Serbia & Montenegro, former Yugoslavia) due to data limitations or changes in country existence.
- Dependent variable:
  - Binary indicator equal to 1 if an arrangement is not succeeded by another arrangement within a threshold period; 0 otherwise.
  - Baseline threshold period: three years after expiration or cancellation.
  - Alternative threshold tested: five years (reduces unconditional success rate from 22 percent to 18 percent).
- Variable groups:
  - (i) Domestic conditions: GDP growth, primary balance, public debt, current account balance, international reserves (initial level and change), trade openness, exchange rate regime, inflation, institutional quality (bureaucracy quality index).
  - (ii) Global environment: global economic growth, world real interest rates (U.S. federal funds rate adjusted by CPI inflation), market volatility (VIX).
  - (iii) IMF- and program-specific: fraction of previous five years under an IMF program, long-run share of successful predecessors (share of past programs without successors), outstanding loans as share of quota, quota share, exceptional access dummy, program duration dummies (short/long), blended arrangements.

### III. Stylized facts (unconditional and descriptive)
- Unconditional probability of 'no successor arrangement' (sample average): approximately 22 percent.
  - By program type: SBA 22 percent, EFF 25 percent, FCTA 17 percent.
- Concentration:
  - Around 30 out of 138 countries had more than 10 GRA arrangements since 1952.
  - Top 20 percent of countries in terms of number of programs account for approximately 50 percent of all GRA programs.
- Key unconditional mean differences (programs without successors vs with successors):
  - Average real growth: about 1 percentage point higher for programs without successors.
  - Initial conditions more favorable for programs without successors: lower public debt-to-GDP, larger current account balance (or smaller deficit), larger stock of reserve assets at program start.
  - Programs without successors on average show stabilized debt, improved current account balance, improved primary balance, and increased reserves over the program period.
  - Trade openness higher and institutional quality better on average for programs without successors.
  - Inflation, exchange rate adjustment, and exchange rate regime do not differ significantly on average between the two groups.
- Global variables:
  - No clear-cut differentiation between arrangements with and without successors for world growth and market uncertainty; global interest rates somewhat lower around programs without successors.
- IMF/program variables:
  - Programs without successors tend to have: larger quota share, spent a smaller fraction of the previous five years under a program, and higher share of past successful programs.
  - Exceptional access cases have an unconditional probability of 'no successor' that is six percentage points higher than normal access counterparts (36 exceptional access cases in sample).
  - Blended arrangements in sample: 23.

### IV. Econometric approach and handling missing data
- Model: probit binary response linking probability of no successor to explanatory variables.
- Missing data strategy: application of the Conniffe and O’Neill (2011) “probit-miss” approach to gain efficiency while allowing for missing observations under the missing at random (MAR) assumption.
  - Steps: regress x2 on x1 by OLS (complete cases), estimate probit on complete cases, estimate probit on incomplete cases using x1 only, combine outputs for efficient estimator and covariance matrix.
  - Limitation: original method assumes same missing pattern across incomplete observations; dataset violates this strictly, but most variables available and missingness concentrated in earlier years, so efficiency gains still achieved.
- Goodness-of-fit and predictive performance metrics:
  - Pseudo R-squared (Efron, 1978) for fit.
  - AUROC (area under the ROC curve) for classification/predictive accuracy; AUROC > 0.5 indicates better-than-random forecasting.

### V. Key estimation results (baseline and subsamples)
- Baseline (country-specific variables):
  - Average growth during and after the program: positive and significant.
    - Average marginal effect (AME) of growth: around 0.7 (an increase in average growth by one percentage point corresponds to a 0.7 percentage points higher probability of avoiding repeated use).
  - Current account balance at program start: positive sign and statistically significant.
  - Change in current account over program: positive sign (less consistently significant in full sample).
  - International reserves (initial level and change over program): positive and economically meaningful; initial reserves more statistically significant than adjustment.
  - Trade openness: positive association with avoiding repeated use.
  - Public debt level and its adjustment: not statistically significant in baseline.
- Global and IMF/program-specific variables in baseline:
  - Global GDP growth and real U.S. interest rate: generally not significant in full sample baselines.
  - IMF-related variables: loans outstanding (share of quota) small marginal effect; share of previous five years under a program shows negative sign (not always significant), suggesting short-run persistence of program use.
- Subsample starting in 1990:
  - Improved goodness-of-fit.
  - AME of domestic growth more than doubled relative to full sample (growth more decisive to prevent repeated use).
  - Current account adjustment over program period gains importance.
  - World real interest rate shows anticipated negative sign (higher real interest rates associated with higher probability of repeated use) in the 1990 subsample.
  - Program duration dummies: longer (shorter) than planned programs associated with higher probability of repeated use (AMEs imply about 10 and 12 percentage points respectively), though coefficients remain statistically insignificant.
- Institutional quality and governance (subsample since 1990):
  - Bureaucracy quality index (ICRG, scale 0–4): higher bureaucracy quality significantly increases probability of no successor.
    - AME for bureaucracy around 12 (implying a one-unit improvement raises probability of no repeated use by about 12 percentage points on average).
    - Example: improving bureaucracy score from 2 to 4 would increase probability of no repeated use by around 24 percentage points (ceteris paribus).
  - Price stability (inflation) not significant.
- Country’s long-run track record with Fund:
  - Share of past programs without successors over the entire past: positive and statistically significant (1 percent level), supporting that past avoidance of repeated use predicts future avoidance.

### VI. Robustness checks (summary)
- Tests performed include: adding quota share, VIX, program type dummy (EFF), alternative governance indicators (government stability, corruption), external debt-to-GDP, interaction terms, time trend, excluding FCTAs, blended arrangement dummy, varying threshold period for no successor.
- Main findings robust:
  - Quota share and VIX not significant.
  - Program type dummy (EFF) not significant.
  - Government stability significant in some specifications; corruption index not significant.
  - External debt-to-GDP not significant.
  - Interaction terms and linear time trend do not alter main conclusions.
  - Excluding FCTAs or including blended dummy does not materially change baseline results (number of blended arrangements small).
  - Extending threshold from 3 to 5 years reduces unconditional success rate from 22 percent to 18 percent; current account adjustment gains importance and reserves lose significance; world real interest rate becomes significant.

### VII. Out-of-sample forecasting (design, limitations, and stylized results)
- Design:
  - Split-sample forecasting: estimate probit on an estimation subsample, predict on a later prediction subsample, dropping overlapping programs to mimic forecasting environment.
  - In practice, realized values of explanatory variables during the prediction period are used (forecasters would need predictions of these variables in real applications, which is a limitation).
  - AUROC used to evaluate predictive performance; statistical test of H0: AUROC = 0.5 versus H1: AUROC > 0.5 reported.
  - Classification threshold chosen to minimize number of falsely predicted outcomes (simple unweighted loss).
- Limitations:
  - Use of realized predictor values instead of genuine ex-ante forecasts may overstate achievable forecasting accuracy.
  - For averaged explanatory variables computed over differing periods for programs with and without successors, a forecaster would need to know succession timing ex ante to compute correct averages; infeasible in practice.
- Stylized forecasting exercise:
  - Parsimonious probit estimated on programs starting since 1980 and ending before 1997.
  - Specification includes essential country-specific characteristics, global variables, and long-run track record variable (“share of past programs without successors”).
  - Parsimonious specifications tend to have superior out-of-sample forecasting performance relative to larger models.

### VIII. Main policy-relevant takeaways
- Domestic macroeconomic conditions matter most for avoiding repeated use of IMF GRA resources:
  - Stronger average growth, better current account position and improvement during programs, and larger international reserves correlate with higher probability of no successor.
- Institutional capacity is important:
  - Higher bureaucracy quality materially increases probability of avoiding repeated use.
- Country history with the Fund is informative:
  - Countries that avoided repeated use in the past are more likely to avoid it again.
- Global environment and some program-specific features:
  - Evidence on the role of global variables is ambiguous and varies over time; IMF-specific variables generally less influential once domestic conditions and institutions are controlled for.
- Predictability:
  - Parsimonious models including the most important determinants can have reasonable out-of-sample predictive power, but accurate ex-ante forecasts require realistic assumptions about macroeconomic variables and program developments.

*Source: Annex 2 of the IMF paper (sample of 890 GRA arrangements starting 1952–2012; baseline “no successor” threshold: three years).*

### 8. Out of the total 52 programs to be predicted, the model

### 8. Out of the total 52 programs to be predicted, the model

### Forecasting performance and exercise results
- Out of the total 52 programs to be predicted, the model correctly calls the outcome of 35.
- The model correctly predicts the majority of:
  - ‘no successor’ cases: 79 percent
  - repeated use cases: 57 percent
- AUROC value: 0.6994, indicating reasonable forecasting performance clearly superior to the naive forecast.
- A focused forecasting exercise (Table 9) that:
  - estimates the model using programs between 1990 and 2002 and
  - applies it to predict program outcomes from 2005 onwards,
  - changes the set of predictors to account for the importance of institutional factors,
  - achieves a share of 77 percent of correctly predicted outcomes for this period.
- For the 2005-onwards period, institutional quality serves as a valuable predictor to assess the potential risk of repeated use.
- The model performs significantly better in predicting no repeated use cases than programs with a successor.

### Key empirical findings on determinants of repeated use
- Data and sample:
  - Large sample of GRA arrangements approved between 1952 and 2012.
  - A restricted sample starting in 1990 is also considered to account for structural changes over time.
- Methodology:
  - Probit models are used.
  - Missing observations of explanatory variables are accounted for.
- Country-specific variables associated with a higher likelihood of implementing a Fund-supported program that does not require a successor:
  - higher economic growth
  - a larger current account surplus (or a smaller deficit)
  - a larger stock of international reserves
  - more trade openness
  - higher institutional quality
- Program-specific variables and a country’s history with the IMF partly explain repeated use.
  - Some evidence that a country’s long-run track record of implementing programs without successors is a relevant factor for the risk of repeated use.

### Policy implications and limitations
- Parsimonious model specifications can help successfully predict repeated use of IMF-supported programs and offer reasonable predictive power for ex ante risk assessment.
- Forecasting performance is far from perfect and subject to limitations:
  - The set-up of a successor arrangement is a notoriously difficult variable to predict and is influenced by factors outside the empirical model, notably political economy considerations.
- From a policy perspective, the outlined approach could serve as one tool within a larger toolbox to assess risks to program success and the Fund’s financial resources.

### Recommendations and avenues for future research
- Potential extensions and improvements:
  - Extend the analysis to concessional lending programs under the PRGT, noting that repeated use would need to be redefined to reflect the longer-term nature of PRGT arrangements.
  - Evaluate in more detail the role of program-specific characteristics, for example:
    - Define a measure of realism of program targets and investigate the impact of over-ambitious targets on the probability of repeated use by comparing program targets to benchmarks based on country-specific or cross-country experiences.
  - Improve the stylized forecasting exercise by:
    - working with data vintages to properly mimic a true out-of-sample forecasting situation, and
    - including measures of forecasting uncertainty surrounding the point estimates.

*Source: wpiea2019245-print-pdf*

### Annex 1: Figures and Tables

### Annex 1: Figures and Tables

### Figures (listed)
- Figure 1: IMF-supported GRA programs per year
- Figure 2: Key macroeconomic variables: pre-program, program, and post-program period for programs without successor (left column) and with successor (right column) within three years

### Table 1: Baseline model specifications including public debt, full sample
- Report: estimated coefficients of probit regressions and corresponding standard errors; AME = average marginal effect.
- Selected AME (column ordering preserved as in source):
  - GDP_growth: 0:70 0:031(0:015); 0:82 0:036(0:015); 0:75 0:033(0:015)
  - Public_debt: 0:03 0:001(0:003); 0:03 0:001(0:003); 0:04 0:002(0:003)
  - Public_debt_adj: 0:08 0:004(0:006); 0:06 0:003(0:006); 0:10 0:004(0:006)
  - Primary_balance: 0:89 0:039(0:027); 1:20 0:052(0:027); 1:31 0:057(0:028)
  - CA_balance: 0:75 0:033(0:010); 0:84 0:036(0:011); 0:83 0:036(0:011)
  - Reserves: 2:54 0:112(0:038); 2:34 0:101(0:038); 2:22 0:097(0:038)
- Sample and fit:
  - Obs. (complete): 890 (263) for all three columns
  - R2: 0.1232; 0.1288; 0.1410
  - AUROC: 0.7077; 0.7147; 0.7141

### Table 2: Baseline model specifications, full sample
- Selected AME highlights:
  - GDP_growth: 0:81 0:037(0:012); 0:89 0:039(0:013); 0:89 0:040(0:013)
  - Primary_balance: 0:62 0:028(0:024); 0:86 0:038(0:024); 0:82 0:037(0:025)
  - CA_balance: 0:73 0:033(0:010); 0:81 0:036(0:010); 0:80 0:036(0:010)
  - Reserves: 2:78 0:127(0:037); 2:60 0:116(0:037); 2:54 0:114(0:037)
- Sample and fit:
  - Obs. (complete): 890 (279)
  - R2: 0.1233; 0.129; 0.1405
  - AUROC: 0.7204; 0.7295; 0.7279

### Table 3: Baseline model specifications, sample since 1990
- Selected AME highlights:
  - GDP_growth: 2:72 0:098(0:027); 2:15 0:080(0:029); 1:92 0:072(0:031)
  - CA_balance: 1:27 0:046(0:018); 1:18 0:044(0:019); 1:20 0:045(0:020)
  - Reserves: 3:98 0:143(0:056); 3:03 0:113(0:058); 2:81 0:105(0:058)
- Sample and fit:
  - Obs. (complete): 226 (173)
  - R2: 0.1821; 0.2392; 0.2601
  - AUROC: 0.7469; 0.7978; 0.8133

### Table 4: Model specs including exchange rate (regime) and alternative measure of track record, full sample
- Selected AME highlights:
  - GDP_growth: 0:72 0:033(0:013); 0:81 0:036(0:014); 0:74 0:033(0:018)
  - NEER_adj: 0:16 0:007(0:005); 0:15 0:007(0:005); 0:17 0:007(0:005)
  - ER_regime: 0:44 0:020(0:014); 0:48 0:021(0:014); 0:50 0:023(0:016)
  - Share_suc_past: 0:20 0:009(0:003)
- Sample and fit:
  - Obs. (complete): 890 (279); last column 752 (255)
  - R2: 0.1310; 0.1377; 0.1760
  - AUROC: 0.7302; 0.7381; 0.7584
- Note: last-column specification drops first program of each country to compute share of successfully implemented programs in the past.

### Table 5: Model specs including exchange rate (regime) and alternative measure of track record, sample since 1990
- Selected AME highlights:
  - GDP_growth: 2:69 0:097(0:027); 2:18 0:081(0:029); 2:60 0:092(0:039)
  - ER_regime (last column): 1:01 0:036(0:031)
  - Share_suc_past: 0:29 0:010(0:006)
- Sample and fit:
  - Obs. (complete): 226 (173); last column 187 (155)
  - R2: 0.18; 0.2447; 0.2207
  - AUROC: 0.75; 0.8031; 0.7932

### Table 6: Model specs including inflation, institutions and exceptional access dummy, sample since 1990
- Selected AME highlights:
  - GDP_growth: 2:66 0:115(0:032); 2:39 0:111(0:033); 2:28 0:104(0:036)
  - Institutions: 10:91 0:471(0:176); 11:53 0:533(0:188); 11:86 0:541(0:194)
  - Dum_access: 6:04 0:262(0:446)
- Sample and fit:
  - Obs. (complete): 226 (139)
  - R2: 0.2407; 0.3058; 0.3154
  - AUROC: 0.7995; 0.8420; 0.8456

### Table 7: Robustness checks (multiple specifications)
- Columns (1)–(12) report AME and (se) for many variables across robustness specifications.
- Selected recurring AME magnitudes (examples across columns):
  - GDP_growth ranges reported: 0:76 … 2:72 … 2:67 … 1:20 etc., with corresponding standard errors as listed per column.
  - CA_balance and CA_balance_adj show positive AME across many specifications (examples: CA_balance 0:78 0:035(0:010); 1:34 0:056(0:018); 1:13 0:044(0:023)).
  - Reserves and Reserves_adj show positive AME in many specifications (examples: Reserves 2:54 0:115(0:037); Reserves_adj 2:14 0:096(0:067); Reserves 3:59 0:159(0:041)).
  - Institutions appears positive where included: 11:68 0:551(0:201)
  - Public_debt and Public_debt_adj listed in some columns: Public_debt 0:06 0:003(0:003); Public_debt_adj 0:38 0:016(0:012)
- Model features varied across robustness columns: inclusion of time effects, exclusion of FCTA, five-year threshold, alternative variable sets (Quota_share, Vola, Dum_EFF, Dum_blend, Inflation, Gov_stability, Corruption, External_debt interactions).
- Observations, R2, AUROC vary by column; examples:
  - Obs. (complete): 890 (279); 226 (139); 781 (269); 890 (276) depending on column
  - R2 examples: 0.1411; 0.2751; 0.3140; 0.1372
  - AUROC examples: 0.7296; 0.8145; 0.8485; 0.7539
- Note: IA refers to interaction term (product of initial condition and adjustment over the program period).

### Table 8: Out of sample forecasting (period: 01/2000 - 01/2012)
- Confusion counts:
  - True (no repeated use) / Pred. (no repeated use): 191 / 231
  - True (repeated use) / Pred. (repeated use): 51 / 621
  - Totals: Pred. (no repeated use) 191 231; Pred. (repeated use) 51 621; Total 242 852
- Performance:
  - Share (correct): 0.79 0.57 0.67 (preserve formatting as in source)
  - Obs. (est. / pred.): 280 / 52
  - Cut-off probability: 0.28
  - AUROC (p-value): 0.6994 (0.00)
- Included predictors: GDP_growth, Primary_balance_adj, CA_balance, CA_balance_adj, Reserves, Openness, World_GDP_growth, US_IR, Share_suc_past.
- Estimation period: 01/1980 - 01/1997 (latest completion date).

### Table 9: Out of sample forecasting (period: 01/2005 - 01/2012)
- Confusion counts:
  - True (no repeated use) / Pred. (no repeated use): 125 / 17
  - True (repeated use) / Pred. (repeated use): 189
  - Totals: Pred. (no repeated use) 125 17; Pred. (repeated use) 189; Total 131 326
- Performance:
  - Share (correct): 0.92 0.62 0.77
  - Obs. (est. / pred.): 134 / 26
  - Cut-off probability: 0.43
  - AUROC (p-value): 0.7870 (0.00)
- Included predictors: GDP_growth, Primary_balance_adj, CA_balance, CA_balance_adj, Institutions, US_IR, Share_suc_past.
- Estimation period: 01/1990 - 01/2002 (latest completion date).

*Source: Annex 1: Figures and Tables (from wpiea2019245-print-pdf).*

### Annex 2: Other Dataset Figures and Tables

### Annex 2: Other Dataset Figures and Tables

### Figures: GRA programs and distribution
- Figure A-1: GRA programs starting between 1952 and 2012 per country (A-Gab) with (red) and without (blue) successor within three years.
- Figure A-1 (continued): GRA programs per country (Gam-New) with (red) and without (blue) successor within three years.
- Figure A-1 (concluded): GRA programs per country (Nic-Z) with (red) and without (blue) successor within three years.
  - Note: Red bars not followed by another bar indicate GRA programs that were succeeded by a PRGT program within three years.
- Figure A-2: Distribution of GRA programs across countries I
- Figure A-3: Distribution of GRA programs across countries II

### Variable definitions and sources (Table A-1)
- (No) repeated use
  - Description: Binary variable (1 if a program was not followed by a successor program within three years, 0 otherwise)
  - Measurement: ———-Internal IMF program database, authors’ calculations
  - Share missing: 0%
- GDP_growth
  - Description: Real gross domestic product (GDP), percentage change
  - Measurement: Average
  - Source: World Economic Outlook (WEO), Penn World Table (PWT)
  - Share missing: 0%
- Public_debt
  - Description: General government debt, % of GDP
  - Measurement: Initial level at year of program start; change over program period (adjustment)
  - Source: WEO, IMF Historical Debt Database
  - Share missing: 22%
- External_debt
  - Description: Total external debt, % of GDP
  - Measurement: Initial level at year of program start; change over program period (adjustment)
  - Source: WEO
  - Share missing: 42%
- Primary_balance
  - Description: General government primary balance, % of GDP
  - Measurement: Initial level at year of program start; change over program period (adjustment)
  - Source: WEO, OECD, Inter-American Development Bank, IMF Historical Public Finance Dataset, IMF Government Finance Statistics
  - Share missing: 51%
- CA_balance
  - Description: Current account balance (for 30% of the dataset only trade balance), % of GDP
  - Measurement: Initial level at year of program start; change over program period (adjustment)
  - Source: WEO, PWT, Worldbank World Development Indicators (WDI)
  - Share missing: 0%
- Reserves
  - Description: Total reserves in months of imports
  - Measurement: Initial level at year of program start; change over program period (adjustment)
  - Source: Worldbank WDI
  - Share missing: 48%
- Inflation
  - Description: Percentage change of the Consumer Price Index (CPI)
  - Measurement: Average
  - Source: WEO, International Financial Statistics (IFS), Worldbank WDI
  - Share missing: 6%
- Openness
  - Description: Share of total trade (EX+IM) in GDP
  - Measurement: Year of program start
  - Source: PWT, Worldbank WDI
  - Share missing: 0%
- NEER_nominal effective exchange rate
  - Description: Nominal effective exchange rate
  - Measurement: Percentage change over program period (adjustment)
  - Source: Darvas(2012)
  - Share missing: 11%
- ER_regime
  - Description: Exchange rate regime
  - Measurement: Year of program start
  - Source: Ilzetzki et al.(2019)
  - Share missing: 0%
- Institutions
  - Description: Bureaucracy quality index (scale from 0 to 4)
  - Measurement: Average
  - Source: International Country Risk Guide (ICRG)
  - Share missing: 70%
- World_GDP_growth
  - Description: Real World GDP growth, prior to 1963: real U.S. GDP growth
  - Measurement: Average
  - Source: WEO
  - Share missing: 0%
- Vola
  - Description: VXO index, prior to 1986: 12-month centered rolling window standard deviation of S&P500, level difference removed when merging both series
  - Measurement: Average
  - Source: Cboe Global Markets, FRED Database, authors’ calculations
  - Share missing: 0%
- US_IR
  - Description: Real U.S. interest rate (FED Funds rate adjusted by CPI inflation)
  - Measurement: Average
  - Source: FRED Database
  - Share missing: 0%
- Quota_share
  - Description: Share of country’s quota in total IMF quota (in %)
  - Measurement: Year of program start
  - Source: IFS, authors’ calculations
  - Share missing: 0%
- Loans_out
  - Description: IMF GRA credit outstanding relative to quota (in %)
  - Measurement: Year of program start
  - Source: IFS, authors’ calculations
  - Share missing: 0%
- Dum_access
  - Description: Binary variable (1 if program is an exceptional access case, 0 otherwise)
  - Measurement: ———-Internal IMF program database
  - Share missing: 0%
- Share_suc_past
  - Description: Share of a country’s successful preceding GRA programs (in %)
  - Measurement: Year of program start
  - Source: Internal IMF program database, authors’ calculations
  - Share missing: 16%
- Share_cov_past
  - Description: Fraction of previous five years spend under a GRA program (in %)
  - Measurement: Year of program start
  - Source: Internal IMF program database, authors’ calculations
  - Share missing: 0%
- Dum_long
  - Description: Binary variable (1 if program lasted longer than originally planned, 0 otherwise)
  - Measurement: ———-Internal IMF program database, authors’ calculations
  - Share missing: 0%
- Dum_short
  - Description: Binary variable (1 if program lasted shorter than originally planned, 0 otherwise)
  - Measurement: ———-Internal IMF program database, authors’ calculations
  - Share missing: 0%
- Dum_EFF
  - Description: Binary variable (1 if program is an EFF arrangement, 0 otherwise)
  - Measurement: ———-Internal IMF program database, authors’ calculations
  - Share missing: 0%
- Dum_blend
  - Description: Binary variable (1 if program is a ‘blended’ arrangement, 0 otherwise)
  - Measurement: ———-Internal IMF program database, authors’ calculations
  - Share missing: 0%

### Measurement and timing (Figure A-4)
- A: period of the program under consideration
- B: period between program start and the start of the successor program (if existing)
- C: program period + 3-year threshold period (relevant if no successor program exists)
- Measurement conventions:
  - Initial level (year of program start): T0
  - Adjustment over program period: A (program with successor), Average BC (program without successor)
- Notes: Following Larch et al.(2017), in cases where the program ended between the first and eighth month of a year, the previous year has been used as the year of program end (T1). If the adjustment program only lasted one year, this rule was not applied, i.e. the actual year of program end has been used to compute the relevant adjustments and averages of the explanatory variables.

### Summary statistics (Table A-2)
- GDP_growth
  - mean (total): 3.14
  - mean (no repeated use): 3.96
  - mean (repeated use): 2.91
  - p-value: 0.01
- Public_debt
  - mean (total): 57.26
  - mean (no repeated use): 52.23
  - mean (repeated use): 58.78
  - p-value: 0.16
- Public_debt_adj
  - mean (total): 1.61
  - mean (no repeated use): -0.26
  - mean (repeated use): 2.16
  - p-value: 0.13
- Primary_balance
  - mean (total): -0.21
  - mean (no repeated use): -0.27
  - mean (repeated use): -0.18
  - p-value: 0.88
- Primary_balance_adj
  - mean (total): 0.05
  - mean (no repeated use): 0.77
  - mean (repeated use): -0.22
  - p-value: 0.00
- CA_balance
  - mean (total): -4.24
  - mean (no repeated use): -2.40
  - mean (repeated use): -4.76
  - p-value: 0.00
- CA_balance_adj
  - mean (total): 0.10
  - mean (no repeated use): 0.57
  - mean (repeated use): -0.03
  - p-value: 0.14
- Reserves
  - mean (total): 3.14
  - mean (no repeated use): 3.83
  - mean (repeated use): 2.90
  - p-value: 0.01
- Reserves_adj
  - mean (total): 0.03
  - mean (no repeated use): 0.35
  - mean (repeated use): -0.09
  - p-value: 0.14
- Openness
  - mean (total): 52.61
  - mean (no repeated use): 58.23
  - mean (repeated use): 51.04
  - p-value: 0.01
- NEER_adj
  - mean (total): -7.05
  - mean (no repeated use): -6.46
  - mean (repeated use): -7.23
  - p-value: 0.62
- ER_regime
  - mean (total): 7.16
  - mean (no repeated use): 7.32
  - mean (repeated use): 7.11
  - p-value: 0.60
- Inflation
  - mean (total): 40.10
  - mean (no repeated use): 35.30
  - mean (repeated use): 41.52
  - p-value: 0.64
- Institutions
  - mean (total): 1.85
  - mean (no repeated use): 2.01
  - mean (repeated use): 1.79
  - p-value: 0.06
- External_debt
  - mean (total): 56.17
  - mean (no repeated use): 58.93
  - mean (repeated use): 55.27
  - p-value: 0.59
- External_debt_adj
  - mean (total): 0.83
  - mean (no repeated use): -0.50
  - mean (repeated use): 1.26
  - p-value: 0.54
- World_GDP_growth
  - mean (total): 3.94
  - mean (no repeated use): 3.90
  - mean (repeated use): 3.95
  - p-value: 0.70
- Vola
  - mean (total): 20.87
  - mean (no repeated use): 21.29
  - mean (repeated use): 20.75
  - p-value: 0.16
- US_IR
  - mean (total): 5.92
  - mean (no repeated use): 5.50
  - mean (repeated use): 6.03
  - p-value: 0.04
- Quota_share
  - mean (total): 0.61
  - mean (no repeated use): 0.84
  - mean (repeated use): 0.54
  - p-value: 0.07
- Loans_out
  - mean (total): 101.04
  - mean (no repeated use): 116.70
  - mean (repeated use): 96.68
  - p-value: 0.06
- Share_suc_past
  - mean (total): 11.25
  - mean (no repeated use): 17.10
  - mean (repeated use): 9.70
  - p-value: 0.00
- Share_cov_past
  - mean (total): 40.11
  - mean (no repeated use): 34.29
  - mean (repeated use): 41.73
  - p-value: 0.00
- Dum_access
  - mean (total): 0.04
  - mean (no repeated use): 0.09
  - mean (repeated use): 0.03
  - p-value: 0.00
- Dum_long
  - mean (total): 0.10
  - mean (no repeated use): 0.13
  - mean (repeated use): 0.09
  - p-value: 0.09
- Dum_short
  - mean (total): 0.13
  - mean (no repeated use): 0.11
  - mean (repeated use): 0.14
  - p-value: 0.26
- Dum_EFF
  - mean (total): 0.09
  - mean (no repeated use): 0.10
  - mean (repeated use): 0.09
  - p-value: 0.51
- Dum_blend
  - mean (total): 0.03
  - mean (no repeated use): 0.01
  - mean (repeated use): 0.03
  - p-value: 0.04

### Correlation matrix highlights (Table A-3)
- Correlations are reported pairwise; statistical significance at the 5% level is indicated by *.
- Selected pairwise correlations (value, significance if indicated):
  - GDP_growth with World_GDP_growth: 0.32*
  - GDP_growth with NEER_adj: 0.28*
  - Public_debt_adj with GDP_growth: -0.31*
  - Primary_balance_adj with Primary_balance: -0.54*
  - CA_balance with Primary_balance: 0.44*
  - CA_balance_adj with CA_balance: -0.42*
  - Reserves_adj with CA_balance_adj: 0.39*
  - NEER_adj with Openness: 0.27*
  - ER_regime with NEER_adj: -0.46*
  - NEER_adj with Inflation: -0.56*
  - Inflation with ER_regime: 0.25*
  - US_IR with Primary_balance: 0.22*
  - Quota_share with CA_balance: 0.20*
  - Loans_out with US_IR: -0.27*
  - Share_suc_past with Loans_out: 0.17*
  - Share_cov_past with Primary_balance: 0.31*
  - Share_cov_past with Share_suc_past: -0.33*
- Notes: Statistical significance of the pairwise correlations at the 5% level is indicated by *.

*Source: Annex 2: Other Dataset Figures and Tables (wpiea2019245-print-pdf)*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2019/wpiea2019245-print-pdf.pdf_