## _wp1519 ## Source details **Canonical URL:** [_wp1519](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2015/_wp1519.pdf) ## Other formats - [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2015/_wp1519.pdf.md) - [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2015/_wp1519.pdf.json) --- ### 1. Introduction - Historical scope and IMF mission - IMF's original mission (1945): promotion of exchange rate stability and adjustment of external imbalances between domestic currencies and the U.S. dollar. - Post-1973 expansion: preserving economic and financial stability through continuous country surveillance and stabilization programs. - New lending programs in last fifteen years: precautionary credit lines granting upfront access to IMF resources to mitigate potential liquidity crises and bank runs. - Research question and dataset - Research focus: effects of IMF lending programs on risk of banking crisis in borrowing countries and channels of Fund influence. - Dataset: covers 113 low- and middle-income countries over 1970-2010. - Events observed: - 108 banking crises - 686 IMF lending arrangements - Stylized timing: concentration of bank distress in early 1980s, the 1990s, and in the two years following the US subprime crisis. - Comparative frequencies: - 3.1% of country-year observations with an IMF program in the previous five years experienced a banking crisis. - 1.4% of country-years without IMF loans experienced a banking crisis. - Theoretical channels - Credit availability channel - IMF disbursements increase emergency liquidity provision; catalytic effect as seal of approval; possible adverse effect if lenders recoup credits from limited liquidity. - Conditionalities and reform channel - Post-1990s conditionalities increasingly concern financial reforms and capital account liberalization; compliance can strengthen regulatory frameworks; poorly designed conditionalities may aggravate crises. - Bad signals and moral hazard - Debtor moral hazard, creditor moral hazard and runs, and stigma effects described. - Empirical strategy and identification - Primary focus: determinants of banking crises with emphasis on IMF program presence in previous years. - Endogeneity addressed via IV strategy: - Instruments: degree of political similarity with G-7 and presence of past elections for likelihood of signing IMF arrangement in previous five years. - Additional methods: propensity score matching and robustness checks. - Main summarized findings (introductory) - Countries participating in IMF-supported programs are significantly less likely to suffer a future banking crisis than comparable non-borrowers. - Negative correlation significant only above a given loan threshold. - Stronger stabilizing effect when recipient countries comply with conditionalities. - World Bank structural adjustment loans show similar positive effect. - Robust to propensity score matching, crisis-subsample restrictions, and excluding programs associated with prior financial crises. ### 2. IMF support and financial crises (literature, data, stylized facts) - Literature context and prior findings - No prior econometric studies specifically on IMF programs and banking crises identified by authors. - Related findings: - Eichengreen et al. (2008): IMF participation associated with lower likelihood of sudden stops. - Dreher and Walter (2010): IMF program in previous five years reduces probability of future currency crisis; driven by lending agreement per se. - Jorra (2012): IMF programs increase average probability of subsequent sovereign defaults by 1.4 percentage points (sample frequency of defaults 4.8 percent). - Presbitero and Zazzaro (2012): IMF lending during 2008-10 targeted contagion prevention; political similarity with G7 predicts program signing. - Data and stylized facts (sample 113 countries, 1970-2010) - Dependent variable: dummy for systemic banking crisis (Laeven and Valencia (2013)). - Banking crisis episodes: - 108 systemic banking crises observed. - First crisis: 1976. - Last crisis: 2009. - Concentration: early 1980s, 1990s, 2008. - Conditional frequencies: - Conditional on no banking crisis in previous eight years: 3.3%. - Conditional on at least one banking crisis in previous eight years: 1.8%. - Conditional on no sovereign crisis in previous eight years: 2.8%. - Conditional on sovereign crisis in previous eight years: 4.8%. - Conditional on currency crisis in previous eight years: 3.7%; absence of past currency crisis: 2.5%. - IMF programs in sample: - 686 IMF-supported programs (276 low-income; 410 middle-income). - ~70% of programs do not follow a financial crisis in prior three years. - Frequency of signing Fund arrangement conditional on past financial crisis in previous three years: 52.2%. - Frequency of signing in tranquil times: 39.8%. - Raw association (cross-tabulation): - Likelihood of banking crisis with IMF lending: 5%. - Likelihood without IMF lending: 2%. - Difference statistically significant in whole sample and by income group. - Empirical specification overview - Linear probability model estimated on pooled panel (no country fixed effects): - Pr(BANKING CRISIS_i,t) = α IMF PRESENCE_i,(t−1,t−5) + sum β_j PAST CRISES_i,(t−1,t−3) + sum γ_j CONTROLS_j,i,t−1 + ε_i,t - IMF presence measured over five-year window t−5 to t−1. - Alternative continuous measure: IMF LOAN/GDP = logarithm of ratio between amount agreed in t−5 to t−1 and GDP (set to zero if no agreement). Loan-to-GDP range: 0.1% to 44.8%. Average (median) loan-to-GDP ratio: 3.2% (2.2%). Half of loans between 1.2% and 4.1% of GDP. - Compliance proxied by amount undrawn at program expiration with threshold 25% (compliant if ≤25% undrawn). - Controls: lagged real per capita GDP, real GDP growth, share of short-term debt over total external debt, credit growth, exchange rate depreciation, M2/reserves, openness, and de facto financial integration ratios (portfolio equity, debt, FDI liabilities) and institutional/banking indicators. - Post-crisis bias addressed by dropping years during which crisis spreads. - Identification and interpretation caveats - Reverse causality: past crises prompt IMF interventions and affect future crisis likelihood. - α interpreted as cross-sectional association between IMF presence in t−5 to t−1 and probability of banking crisis in t (pooled model omits country fixed effects). ### 3. Identification strategy (IV, instruments, diagnostics) - Endogeneity problem and goal - IMF presence endogenous: more likely in crisis years or before crisis onset. - Estimation target: average treatment effect of IMF programs on banking crisis probability. - Primary identification methods - Instrumental Variable (IV) over-identified 2SLS linear probability model. - Robustness: propensity score matching estimator. - Instruments (political variables, measured over (t−6,t−10)) - PROXIMITY: average alignment score with G7 computed from UNGA votes, averaged over (t−6,t−10). - ELECTION: dummy = 1 if at least one executive election between t−10 and t−6. - USAID (robustness): share of US aid in total foreign aid inflows, averaged over (t−6,t−10). - Rationale and timing - Political proximity predicts IMF intervention via political influence/reward mechanisms. - Election timing predicts arrangements as governments may strike deals early to minimize stigma. - Instruments lagged (t−6,t−10) to predate IMF presence window (t−1,t−5). - Validity and diagnostics - First-stage F-statistic (baseline): 18.5 (above Stock and Yogo rule-of-thumb 10). - Sargan-Hansen overidentification test: does not reject null of instrument exogeneity. - Kleibergen-Paap rk LM-statistic: model not underidentified. - Falsification test: no direct correlation between instruments and banking crisis likelihood. - No evidence that US foreign aid share differs significantly between crisis and non-crisis years (average about 21%). - No evidence of strategic voting behavior around IMF arrangements in sample. - Potential exogeneity concerns discussed - Political instruments could affect financial stability directly (e.g., G7 support ≡ liquidity support) but no sample evidence found. - Strategic friendship behavior would bias against finding a positive IMF effect; no evidence in sample. - Election variable lagged at least six years mitigates contemporaneous election effects. - Implementation details - Treatment: IMFARRANGEMENTt−1,t−5 = 1 if at least one IMF arrangement in t−5 to t−1. - Standard errors clustered at country level; bootstrapped standard errors with 2,000 repetitions reported. - Main IV versus OLS findings - OLS: IMF arrangement not statistically correlated with incidence of future banking crises. - IV (2SLS with PROXIMITY and ELECTION): coefficient on IMFARRANGEMENTt−1,t−5 becomes negative and statistically significant. - Magnitude reported: Fund’s assistance lowers probability of incurring a banking crisis from about 6 to 0.5 percent (IV estimates). - First-stage F-test: 18.5 for excluded instruments. ### 4. Credit channel and non-linearity by loan size - Hypothesis: IMF effects operate via credit availability/liquidity provision and require sufficiently large loans. - Empirical approach: split sample by IMF LOAN/GDP thresholds in range [0.5−7.5] with 0.1 percentage point increments; thresholds correspond roughly to 0.5% (5th percentile) to 7.5% (95th percentile) of IMF LOAN/GDP distribution. - Key findings on loan size - Loans below 2% of GDP: IMF lending does not significantly reduce banking crisis probability. - Loans equal or larger than 3% of GDP: IMF participation effect is negative and statistically significant. - Coefficient stability: - Coefficients above thresholds stable around −0.8 and statistically significant (except at very high thresholds with small sample). - Coefficients below thresholds imprecise and smaller in absolute value up to threshold of 3.2% of GDP; above 3.2% even below-threshold coefficients stabilize near −0.6. - Interpretation: sufficiently large IMF loans (around 3% of GDP and above) systematically reduce future banking crisis probability, consistent with liquidity/credit availability channel rather than pure “seal of approval” independent of loan size. ### 5. The reform channel (compliance and World Bank comparison) - Compliance measurement - Compliant when at most 25% of amount agreed remained undrawn at program expiration (MONA data classification not used due to sample reduction). - In sample, almost 40% of programs with data are non-compliant. - Compliance empirical findings - Compliant countries: probability of a banking crisis is 20.9% lower for countries signing and complying with conditionality. - Non-compliers: lending agreement followed by non-compliance lowers probability by 7.4%. - Results robust when measuring IMF participation by loan size. - First-stage F-statistics below critical value 10 in these specifications; LIML estimates reported and yield similar results. - World Bank structural adjustment lending comparison - WB LENDING dummy = 1 if country received at least one World Bank structural adjustment loan in five years before crisis onset. - Instruments for WB LENDING: ELECTION and MOVEMENT (change in average alignment score with G7 between t and t−1), averaged over (t−6,t−10). - Sample truncated to 2005 for WB data comparability. - Findings: World Bank involvement associated with lower probability of future banking crisis, similar in magnitude and significance to IMF interventions—suggestive of reforms as mechanism. - F-statistics indicate risk of weak identification; LIML supports 2SLS findings. - Key numeric/factual points (section 5.3) - Compliance threshold: ≤25% undrawn at expiration defines “compliant.” - Non-compliance observed in almost 40% of programs with data. - Compliance effect on crisis probability: compliers −20.9%; non-compliers −7.4%. - Sample truncated to 2005 for WB comparisons. - Historical average WB adjustment loan example: USD 160 million; average IMF-supported program in same period: USD 345 million. ### 6. Robustness checks and extensions - 6.1 Past crises conditioning - Restrict sample to IMF programs that did not follow any financial crisis in previous three years. - IMF participation effect remains statistically significant and of similar magnitude as whole sample (Table 7). - 6.2 Additional controls and triggers - Regional dummies: Eastern European and Central Asian countries more likely to experience banking crises; Fund intervention effect persists. - Institutional measures: - COUNTRY RISK (ICRG 0–100), SOUND MONEY (Fraser), POLITY (Polity IV), DEMOCRACY dummy. - Only SOUND MONEY mitigates banking-system vulnerability. - Macroeconomic controls: debt service ratio significant and positive; other macro variables do not affect IMF participation effect. - Financial openness/reforms indices: IMF effect confirmed; crisis likelihood does not differ across degree of openness or liberalization. - DEPOSIT INSURANCE dummy positively correlated with future banking crisis likelihood (moral hazard). - 6.3 Policies and institutions (sample splits) - IMF intervention effect significant only in countries with: - Sound institutional framework (ICRG above median). - SOUND MONEY above median. - Democracy splits yield less robust results; IMF coefficient negative and significant with high POLITY score but not significant when splitting by DEMOCRACY dummy (democracies sub-sample p-value = 0.12). - Many regressions suffer weak identification; LIML supports main results. - 6.4 Instruments and sub-samples - Additional instrument: share of US aid confirms negative and significant IMF participation coefficient. - Sub-samples: - Restricting to 83 countries with at least one systemic banking crisis: IMF effect stronger. - Excluding low-income countries (75 middle-income countries): IMF programs still associated with lower subsequent incidence of banking crises. - Loan-type split (GRA vs PRGT): no significant difference across arrangement types; both concessional and non-concessional programs associated with similar lower probabilities. - 6.5 Propensity score matching (PSM) - PS estimated conditioning on past crises, explanatory variables in t−1, and instruments averaged over previous five years. - Matching algorithms: five nearest neighbors, radius matching (caliper = 0.1), kernel matching. - Findings: - All algorithms: frequency of banking crises significantly lower in treated countries than matched non-treated. - PSM effect magnitude lower than IV estimates but supports association of IMF programs with lower incidence of systemic banking crises. - Diagnostics: - Mean and median absolute standard percentage biases dramatically reduced; mean absolute bias below 5 for all algorithms. - Some significant differences persist after matching in pc GDP, inflation, ratio of M2 over reserves. - Pseudo-R2 from probit selection very low in matched group. - PS selection equation results: IMF programs more likely in countries with lower pc GDP, less financial/trade integration, after sovereign debt crises, with higher debt liabilities and lower portfolio/FDI liabilities, depreciating currencies, and higher real interest rates; political similarity and past elections significant predictors. ### 7. Alternative lag structures - IMFARRANGEMENT redefined over alternative windows: previous 3, 4, 6, or 7 years. - Findings: - Effect statistically significant when assessed over shorter horizons (3 and 4 years). - Expanding window to 6 and 7 years: coefficient negative but smaller in absolute value and less precisely estimated. - Interpretation: program effects manifest in short- and medium-run; they diminish over longer horizons as other developments intervene. ### 8. Conclusions - Main empirical conclusion - OLS biased by reverse causality: higher probability of banking crisis increases likelihood of IMF intervention. - After accounting for endogeneity, countries signing IMF lending programs are less likely to experience a future banking crisis than comparable countries. - Net interpretation: any bad-signal or moral hazard effects are offset by credit availability provided by IMF-supported programs. - Mechanisms and heterogeneous effects - Crisis prevention role linked to macroeconomic policies, financial reforms, and direct liquidity support. - Borrowers compliant with IMF conditionality have lower likelihood of banking crises than non-compliers. - Negative association significant only when loan sufficiently large, consistent with liquidity support channel. *Source: _wp1519 — canonical PDF chapter/section (excerpts provided).* ### 1. Introduction....................................................................................... 3 ### 1 Introduction ### Historical scope and IMF mission - The IMF’s original mission (1945) was limited to the promotion of exchange rate stability and the adjustment of external imbalances between domestic currencies and the U.S. dollar. - After the demise of the dollar exchange standard in 1973, the Fund’s scope expanded to preserving economic and financial stability in member countries through continuous country surveillance and stabilization programs. - Over the last fifteen years new lending programs were introduced to grant precautionary credit lines to pre-qualified countries, providing upfront access to IMF resources to mitigate potential economic vulnerability to sudden liquidity crises and self-fulfilling bank runs. ### Research question and dataset - Research focus: the effects of IMF lending programs on the risk of banking crisis in borrowing countries and the channels of influence of Fund interventions. - Dataset: covers 113 low- and middle-income countries over the period 1970-2010. - Events in dataset used for empirical analysis: - 108 banking crises - 686 IMF lending arrangements - Observed stylized timing: concentration of bank distress episodes in the early 1980s, in the 1990s and in the two years following the US subprime crisis. - Comparative frequency of banking crises: - 3.1% of country-year observations in which an IMF program was in operation in the previous five years - 1.4% of country-years in which there was no IMF loans ### Theoretical channels through which IMF involvement may affect banking crises - Credit availability channel: - IMF disbursements increase emergency liquidity provision, reducing crises caused purely by illiquidity (Haldane, 1999; Miller and Zhang, 2000). - Catalytic effect: the existence of an IMF program can act as a seal of approval, reassuring investors and depositors and influencing other official and private lenders (Bird and Rowlands, 2002; Cottarelli and Giannini, 2006). - Counterpoint: IMF loans might drive creditors to recoup credits from a limited pool of fresh liquidity, potentially triggering a banking panic (Zettelmeyer, 2000). - Conditionalities and reform channel: - Since the 1990s IMF conditionalities increasingly concern financial reforms and capital account liberalization (Joyce and Noy, 2008). - Well-designed and implemented conditionalities can strengthen regulatory and supervisory frameworks, improving banking sector resilience; compliance should reduce likelihood of banking crises. - Counterpoint: poorly designed conditionalities or biased advice could aggravate crises (Radelet and Sachs, 1998; Feldstein, 1998; Stiglitz, 2002). - Bad signals and moral hazard: - Debtor moral hazard: IMF support could substitute for domestic adjustment effort, inducing laxer policies and IMF dependency (Vaubel, 1983; Goldstein, 2001). - Creditor moral hazard and runs: anticipation of IMF bailouts or limited IMF resources can induce investors to withdraw funds or underprice sovereign risk, encouraging excessive risk-taking (Zettelmeyer, 2000; Jeanne and Wyplosz, 2003). - Stigma effect: turning to the IMF for prevention might signal more severe troubles than previously recognized. ### Empirical strategy and identification - Primary empirical focus: determinants of banking crises with emphasis on IMF program presence in previous years. - Endogeneity addressed via an instrumental variable (IV) strategy: - Instruments: degree of political similarity between IMF borrowers and the G-7, and the presence of past elections, used as main instruments for the likelihood of a country signing an IMF lending arrangement in the five years before the crisis. - Additional methods: propensity score matching estimator and robustness checks including sub-samples and alternative specifications. ### Main empirical findings (summarized) - Main finding: countries participating in IMF-supported programs are significantly less likely to suffer a future banking crisis than other comparable non-borrowing countries. - The negative correlation between IMF interventions and the likelihood of a future banking crisis is statistically significant only above a given loan threshold (threshold value not specified in this section). - The stabilizing effect is stronger when recipient countries are compliant with IMF conditionalities. - Supplementary evidence: World Bank structural adjustment loans, which have equivalent conditionalities but smaller loan size, also show a similar positive effect on banking sector stability. - Results are robust to: - Propensity score matching estimators - Restricting the sample to countries that suffered at least one systemic crisis over the sample period - Considering IMF programs that were not associated with financial crises (to exclude reverse causality from past financial crisis leading to lower future crisis probability) ### Paper organization (as presented) - Section 2: Related literature framing theoretical and empirical channels. - Section 3: Main variables and stylized facts. - Section 4: Empirical model and identification strategy. - Sections 5 and 6: Baseline results and robustness exercises. - Section 7: Summary and conclusions. *Source: _wp1519 - 1. Introduction (excerpt).* ### 2.2 IMF support and financial crises ### 2.2 IMF support and financial crises ### Literature review and prior findings - No previous econometric studies specifically analyze the association between IMF-supported programs and banking crises were identified by the authors. - Evidence on IMF effects on financial stability: - Eichengreen et al. (2008): participation in an IMF program is associated with lower likelihood of sudden stops in international capital flows in the years following program participation; the stabilizing role of IMF emergency liquidity provision holds after controlling for reverse causality. - Dreher and Walter (2010): existence of an IMF-supported program in the previous five-year period reduces the probability of a future currency crisis; the lending agreement per se drives the result rather than amount disbursed or degree of compliance. - Jorra (2012): focusing on 57 developing and emerging economies over 1975-2008, IMF-supported programs significantly increase the average probability of subsequent sovereign defaults by 1.4 percentage points (sample frequency of defaults 4.8 percent); result appears to reflect effects of IMF interventions per se and may suggest debtor moral hazard. - Presbitero and Zazzaro (2012): during the 2008-10 financial crisis IMF lending was directed at preventing contagion; participation in IMF programs was sensitive to political-economic interests of IMF’s main shareholders. Findings are mixed: political similarity with G7 countries is positively correlated with probability of signing a loan agreement; IMF channeled more financial resources to countries where the economic crisis was more severe independent of balance of payments imbalances. ### Data and stylized facts - Sample and time frame: - Panel dataset covering 113 developing countries over the period 1970-2010. - Dependent variable: dummy equal to one for country-year observations with a systemic banking crisis (BANKING CRISIS) following Laeven and Valencia (2013) definition of systemic banking crisis. - Banking crisis episodes in sample: - 108 systemic banking crises observed. - First crisis in sample: 1976. - Last crisis in sample: 2009. - Concentration of episodes in early 1980s, in the 1990s, and in 2008 (global financial crisis). - Conditional frequencies and correlations with past crises: - Frequency of a banking crisis conditional on not having experienced banking crises in the previous eight years: 3.3%. - Frequency of a banking crisis conditional on having experienced at least one banking crisis in the past eight years: 1.8%. - Frequency of a banking crisis conditional on not having experienced a debt (sovereign) crisis in previous eight years: 2.8%. - Frequency of a banking crisis conditional on having experienced a debt (sovereign) crisis in previous eight years: 4.8%. - Frequency of a banking crisis conditional on occurrence of a currency crisis in previous eight years: 3.7%. - Frequency of a banking crisis in absence of a past currency crisis: 2.5%. - IMF programs in sample: - 686 IMF-supported programs agreed for low- and middle-income countries over sample period. - 276 programs in low-income countries; 410 programs in middle-income countries. - IMF activity generally peaked around crisis years; Fund’s financial support remained sustained (but declining) after late 1990s until 2007 financial crisis. - Approximately 70% of the IMF programs in the sample do not follow a financial (banking, currency or sovereign) crisis that happened in the previous three years. - Frequency of signing a Fund lending arrangement conditional on having gone through a financial crisis in the previous three years: 52.2%. - Frequency of signing a Fund lending arrangement in tranquil times: 39.8%. - Association between IMF-supported programs and banking crises (raw cross-tabulation): - Likelihood of a banking crisis in country-year observations with IMF lending: 5%. - Likelihood of a banking crisis in country-year observations without IMF lending: 2%. - Difference between these two probabilities is statistically significant in the sample (whole sample and when split by income group). ### Empirical strategy and model specification - Main goal: assess effect of IMF lending agreements on probability of occurrence of systemic banking crises. - Estimation approach: - Linear probability model estimated on pooled panel (no country fixed effects) to retain cross-sectional variation and interpret coefficient α in a cross-sectional dimension. - Baseline equation: - Pr(BANKING CRISIS_i,t) = α IMF PRESENCE_i,(t−1,t−5) + sum β_j PAST CRISES_i,(t−1,t−3) + sum γ_j CONTROLS_j,i,t−1 + ε_i,t - Rationale for pooled model: country-fixed effects wash out much cross-sectional variation; authors follow prior literature (Demirguc-Kunt and Detragiache, 1998; Becket al., 2006; De Haan and Shehzad, 2009; Jorra, 2012). - Key explanatory variables and timing: - IMF presence measured on a five-year window t−5 to t−1 to capture medium-term program effects and time needed to implement conditionality reforms. - Baseline IMF measure: dummy I M F ARRANGEMENT = 1 if country signed at least one IMF lending arrangement in t−5 to t−1. - Alternative continuous IMF measure: I M F LOAN/GDP = logarithm of ratio between amount of loan agreed in t−5 to t−1 and country GDP. - Loan-to-GDP ratio in sample ranges from 0.1% to 44.8%. - Average (median) loan-to-GDP ratio: 3.2% (2.2%). - Half of loans are between 1.2% and 4.1% of GDP. - I M F LOAN/GDP set to zero if no agreement in five-year period. - For multiple arrangements in window, loan amounts (as share of GDP) are summed. - Compliance / reform channel: degree of compliance proxied by amount of agreed IMF loan undrawn at program expiration, with threshold at 25% (distinguish compliant vs non-compliant countries). - Controls and past crises: - Past crises (PAST CRISES): three dummies equal to one if country experienced at least one banking crisis, one sovereign crisis (DEBT CRISIS), or one currency crisis (CURRENCY CRISIS) between t−8 and t−1. - Macroeconomic controls (all one-year lagged): real per capita GDP, real GDP growth, share of short-term debt over total external debt, credit growth, exchange rate depreciation, ratio of M2 to reserves, measure of openness, and three indicators of de facto financial integration measured as ratios over GDP of: (i) portfolio equity liabilities, (ii) debt liabilities, (iii) foreign direct investment liabilities. - Extensions include institutional characteristics and domestic banking and financial system indicators. - Addressing post-crisis bias: - To minimize impact of crisis on right-hand-side variables (post-crisis bias), years during which the crisis is spreading are dropped from the sample following Gourinchas and Obstfeld (2012) and Catão and Milesi-Ferretti (2014). ### Identification challenges and interpretation - Reverse causality and endogeneity concerns: - Close interactions between financial crises and subsequent IMF interventions, and between past and current financial crises, complicate identifying IMF participation effect on banking crisis probability. - Past financial crises can both prompt IMF interventions and affect future crisis likelihood by making policies more conservative or by increasing fragility, creating confounding dynamics that must be disentangled. - Interpretation of estimated α: - Because pooled estimation omits country fixed effects, α is interpreted in a cross-sectional dimension as the association between IMF presence in t−5 to t−1 and the probability of a banking crisis in t, rather than as a within-country causal effect over time. * _wp1519 - 2.2 IMF support and financial crises_* ### 4.2 The identification strategy ### 4.2 The identification strategy ### Endogeneity concern and estimation goal - Problem: Fund’s presence is more likely in crisis years or just before crisis onset, so a simple linear probability model can be biased because IMF arrangements are endogenous. - Estimation target: the average treatment effect of IMF programs on the probability of banking crisis (the “IMF participation effect”). - Identification challenge: treatment (“the IMF loan agreement”) is not randomly assigned and is correlated with country characteristics; selection bias must be addressed. ### Identification methods used - Primary strategy: Instrumental Variable (IV) approach (over-identified 2SLS linear probability model). - Robustness check: propensity score matching estimator (reliance on different exogeneity assumptions; see Section 6.5 in source). ### Instruments (political variables) - PROXIMITY: average alignment score with G7 countries computed on all UNGA regular votes; political similarity with the G7 is built by averaging pairwise annual alignment scores. PROXIMITY is averaged over the five-year period (t−6,t−10) before the five-year period (t−1,t−5) in which a Fund lending program is allowed. - ELECTION: dummy equal to one if the country had at least an executive election between t−10 and t−6; measured over the same time frame (t−6,t−10). - USAID (used in robustness analysis): share of US aid in total foreign aid inflows, averaged over (t−6,t−10). ### Rationale and timing of instruments - Political proximity with IMF major shareholders (G7/US) predicts IMF intervention because of political influence and reward mechanisms; voting similarity in the UNGA is used as proxy. - Election timing predicts IMF arrangements because governments have incentive to strike Fund deals early in their terms to minimize stigma. - Timing: all three instruments measured over (t−6,t−10) to capture lagged effects on participation in programs in (t−1,t−5). ### Validity and diagnostic evidence for instruments - Relevance: First-stage F-statistic in baseline is 18.5 (above Stock and Yogo (2005) rule of thumb of 10), indicating instruments are not weak. - Overidentification: Sargan-Hansen test of overidentifying restrictions does not reject the null that instruments are uncorrelated with the error term. - Underidentification: Kleibergen-Paap rk LM-statistic suggests the model is not underidentified. - Falsification: a simple falsification test shows no direct correlation in the sample between each of the three instruments and the likelihood of a banking crisis. - No evidence in sample that past US foreign aid share (U S AI D) differs significantly between crisis and non-crisis years (average about 21%). - No evidence in sample of strategic voting behavior around IMF loan agreements: sample used for Figure 2 — original sample 2,250 country-year observations, dropped arrangements after 2003 leaving 308 arrangements; control group N=368 randomly selected country-year observations with no IMF lending arrangement — shows a small increase in PROXIMITY in the 11-year window common to both groups (no downward trend after lending). ### Potential exogeneity concerns discussed - Political instruments could conceivably affect financial stability directly (e.g., G7 support could translate into direct liquidity support), but no evidence found in sample to support such a direct linkage. - Strategic friendship (countries moving closer to US/G7 in expectation of trouble) would bias results against finding a positive IMF effect on stability; no evidence of such behavior in the sample. - Election variable is lagged at least six years before crisis year, mitigating direct contemporaneous election effects (noting Dinç (2005) documents lending spikes in election years, but those are contemporaneous). ### Implementation details - Main binary treatment: presence of at least one IMF lending arrangement between t−1 and t−5 (IMFARRANGEMENTt−1,t−5). - Standard errors clustered at country level in columns 1-7. - Bootstrapped standard errors with 2,000 repetitions reported in last two columns — results qualitatively identical. --- ### Main empirical findings (IV versus OLS) - OLS: having signed an IMF lending agreement is not statistically correlated with incidence of banking crises in subsequent years (column 1). - IV (2SLS, instrumenting IMFARRANGEMENTt−1,t−5 with PROXIMITY and ELECTION): coefficient on IMFARRANGEMENTt−1,t−5 becomes negative and statistically significant (column 3). - Magnitude reported: Fund’s assistance lowers the probability of incurring a banking crisis from about 6 to 0.5 percent. - First-stage diagnostics: excluded instruments (foreign policy similarity and past elections between t−10 and t−6) show positive and significant correlation with IMF presence in (t−1,t−5); F-test = 18.5. ### Robustness checks and alternative specifications - Dropping the global recession 2008-2010 period (columns 4-5): results unaffected; first-stage F-statistics lower but still above 10. - Using size of the loan (share of GDP) as treatment (columns 6-7): results unaffected; first-stage F-statistics still above safe threshold of 10. - Bootstrapped standard errors (2,000 rep.) yield qualitatively identical results. ### Control variable evidence (consistent with literature) - Banking crisis occurrence is robustly associated with: - Low per capita GDP growth. - Rapid credit growth. - Financial integration: - Foreign direct investment liabilities reinforce banking stability. - More volatile debt inflows raise probability of banking crisis. - Trade integration associated with lower incidence of banking crises. - Association found between sovereign and debt crises; no similar pattern for currency crises. - No statistical association found between probability of banking crises and: - Fraction of short-term external debt. - Currency depreciation. - Ratio of M2 over international reserves. - Real GDP per capita. --- ### Credit channel and non-linearity by loan size - Hypothesis: positive IMF effects on banking sector stability may operate via credit availability/liquidity provision and may require sufficiently large loans. - Empirical approach: split sample by loan size (IMF loan as percentage of GDP) across thresholds in the range [0.5−7.5] with 0.1 percentage point increments; report coefficients α on IMFARRANGEMENTt−1,t−5 for selected subsamples (Table 5). - Note: the threshold range chosen corresponds roughly to 0.5% (5th percentile) to 7.5% (95th percentile) of IMF LOAN/GDP distribution. - Findings: - Loans below 2% of GDP: IMF lending does not significantly reduce probability of banking crises. - Loans equal or larger than 3% of GDP: IMF participation effect is negative and statistically significant both below and above the threshold. - Figure 3: coefficients estimated in samples above thresholds are stable around −0.8 and statistically significant except at very high thresholds where sample size is small. - Coefficients estimated in samples below thresholds are imprecise, less stable, and smaller in absolute value up to a loan size threshold of 3.2 percent of GDP; above 3.2% even below-threshold coefficients become stable and close to −0.6. - Interpretation: only significantly large IMF loans are associated with a lower probability of banking crises across countries, suggesting importance of a credit availability (liquidity) channel; small loans below about 2% of GDP do not produce the same stabilizing effect. ### Summary conclusion on identification strategy and evidence - Using an IV strategy with political instruments measured over (t−6,t−10) and exploiting timing of IMF arrangements in (t−1,t−5), the analysis finds that correcting for endogeneity changes the sign and significance of the IMF participation effect: IMF intervention significantly lowers banking crisis probability. - Diagnostic tests support instrument relevance and overidentifying restrictions; falsification tests find no direct instrument–crisis correlation in the sample. - Non-linearity analysis by loan size indicates that only sufficiently large IMF loans (around 3% of GDP and above) systematically reduce future banking crisis probability, consistent with a liquidity/credit availability channel rather than a pure “seal of approval” independent of loan size. *Source: 4.2 The identification strategy — _wp1519 - 4.2 The identification strategy* ### 5.3 The reform channel ### 5.3 The reform channel ### 5.3.1 Compliance with conditionality - Concept: - IMF intervention may affect banking sector stability via conditions and policy advice attached to IMF-supported programs (financial-sector conditionalities can make the financial sector more resilient to international capital flows and external shocks). - Measurement of compliance: - Following Killick (1995) and Dreher (2003), Dreher and Walter (2010): a country is coded as compliant when at most 25% of the amount agreed under an IMF arrangement remained undrawn at program expiration. - MONA data classification of conditionality categories starts from 1992; using MONA for financial-sector conditionalities would dramatically reduce sample size, so the undrawn-loan indicator is used instead. - In the available sample, almost 40% of the programs for which data exist are considered non-compliant with Fund conditionality. - Empirical findings: - The effect of IMF-supported programs on the probability of a future banking crisis critically depends on conditionality compliance. - Baseline estimates (column 1): the probability of a banking crisis is 20.9% lower for countries that sign an IMF lending agreement and comply with Fund conditionality. - A lending agreement followed by non-compliance lowers the probability of a banking crisis only by 7.4%. - Results are robust when measuring IMF participation with loan size (column 3). - First-stage F-statistics are below the critical value of 10 in both specifications, suggesting possible weak identification; Limited Information Maximum Likelihood (LIML) estimates (columns 2-4) are reported and yield similar results (LIML is less precise but less biased than 2SLS). ### 5.3.2 World Bank lending - Rationale: - Compare IMF and World Bank (WB) structural adjustment loans to disentangle credit vs. reform channels. - Structural adjustment loans typically have similar conditionalities to IMF programs but involve more limited funds. - If WB adjustment loans are less effective than IMF arrangements, this would support a credit channel; if equally effective despite smaller loan size, this would support the reform channel. - Measurement and instruments: - WB LENDING dummy = 1 if country received at least one World Bank structural adjustment loan in the five-year period before the onset of the banking crisis; 0 otherwise. - Instrumental variables for WB LENDING: (1) indicator for elections, and (2) change in the average alignment score with G7 countries between t and t−1 (MOVEMENT), averaged over the five-year period (t−6,t−10) before (t−1,t−5). - For comparability, same instruments and same sample (truncated to 2005 because of data availability on World Bank lending arrangements) are used when considering IMF lending. - Empirical findings: - Results (Table 6, columns 5-8): World Bank involvement is associated with a lower probability of a future banking crisis, with impact similar in significance and magnitude to IMF interventions. - This pattern is suggestive of reforms as a mechanism through which multilateral loans enhance financial stability. - F-statistics indicate risk of weak identification, but LIML estimates support 2SLS findings. ### Key numeric and factual points in section 5.3 - Compliance threshold for “compliant”: at most 25% undrawn at program expiration. - Non-compliance observed in almost 40% of programs with data. - Compliance effect on banking-crisis probability: - Compliers: 20.9% lower probability. - Non-compliers: 7.4% lower probability. - Sample truncated to 2005 for World Bank lending comparisons. - Average WB adjustment loan size cited (historical example): USD 160 million (Dollar and Svensson, 2000); average IMF-supported program in sample over same period: USD 345 million. --- ### 6 Robustness checks and extensions #### 6.1 Past crises, IMF arrangement and future banking crises - Concern: - IMF participation effect might be confounded by the occurrence of past crises, which could induce more conservative policies and lower future crisis probability independently of IMF involvement. - Approach: - Restrict sample to IMF programs that did not follow any financial crisis in the previous three years. - Finding: - IMF participation effect remains statistically significant and of the same order of magnitude as in the whole sample (Table 7), whether conditioning on any past financial crisis (column 1) or exclusively on past banking, currency, or sovereign crises (columns 2-4). #### 6.2 Other potential triggers of banking crises - Concern: - Unobserved country-specific characteristics could influence both financial stability and IMF intervention. - Controls and findings (Table 8): - Regional dummies: Eastern European and Central Asian countries are more likely to experience banking crises, but the positive effect of Fund intervention in forestalling future crises remains when controlling for regional factors (column 1). - Institutional measures included: - COUNTRY RISK: Composite Index from ICRG (0–100; higher = lower risk). - SOUND MONEY: Index from Fraser Institute (measures growth rates of money supply and inflation, freedom to own foreign currency accounts). - POLITY: Polity IV score (autocratic to democratic). - DEMOCRACY: dummy per Cheibub et al. (2010). - Only SOUND MONEY mitigates banking-system vulnerability to systemic crises (column 3); other institutional measures do not significantly alter crisis probability. - Macroeconomic controls: inflation (GDP deflator), financial depth (credit to private sector/GDP), debt service (share of exports), current account balance (share of GDP). - Debt service ratio is significant and positive (lack of liquidity raises systemic crisis likelihood) (column 8); other macro variables do not affect IMF participation effect. - Financial openness and liberalization: - FINANCIAL OPENNESS: Chinn and Ito (2010) index. - FINANCIAL REFORMS: Abiad et al. (2008) index. - IMF lending effect confirmed; banking-crisis likelihood does not differ across degree of capital account openness or financial liberalization (columns 10-11). - Deposit insurance: - DEPOSIT INSURANCE dummy: presence of an explicit deposit insurance scheme is positively correlated with likelihood of a future banking crisis (moral hazard), consistent with Demirguc-Kunt and Detragiache (1998, 2002) (column 12). #### 6.3 The role of policies and institutions - Sample splits (Table 9) by institutional and policy quality: - IMF intervention effect significant only in countries with: - Sound institutional framework (ICRG score above sample median) (columns 1-2). - Access to sound money above median (columns 3-4). - Interpretation: catalytic effect of the Fund stronger and moral hazard risks lower in countries with sound institutions and monetary environment. - Democracy measures: - Splitting by POLITY and DEMOCRACY yields less robust results. - IMF arrangement coefficient is negative and significant only when associated with a high POLITY score; coefficients not significant when splitting by DEMOCRACY dummy (in democracies sub-sample p-value = 0.12). - Magnitude of coefficients across sub-samples generally similar, except when splitting by DEMOCRACY (columns 7-8); many regressions suffer weak identification but LIML supports main results. #### 6.4 Instruments and sub-samples - Additional instrument: - Share of US aid in total aid inflows used as additional instrument; it is relevant and confirms negative and significant coefficient on IMF participation binary indicator (Table 10, columns 1-2). - Sub-sample analyses: - Restricting to 83 countries that had at least one systemic banking crisis: IMF participation effect is even stronger (Table 10, column 3). - Excluding low-income countries (keeping 75 middle-income countries): IMF programs still associated with subsequent lower incidence of banking crises (column 4); low-income sub-sample too small for precise estimates but sign remains negative. - Loan-type split: - GRA (non-concessional) vs PRGT (concessional, replacing PRGF): no significant difference in IMF effect across arrangement types (columns 5-6); both concessional and non-concessional programs associated with similar lower probabilities of banking crises. #### 6.5 Propensity score matching results - Motivation: - Provide an alternative to IV if instrument exogeneity is doubted. - Propensity score (PS) specification (equation reproduced in text): - PS estimated probability that a country agrees to sign at least one IMF arrangement in a five-year window, conditional on: - (i) dummies for occurrence of banking, currency, or sovereign crises in previous three years; - (ii) almost all explanatory variables used in baseline and robustness regressions measured in t−1; - (iii) three instruments used in the analysis averaged over previous five years. - Matching algorithms used: - Five nearest neighbors with replacement. - Radius matching with caliper = 0.1. - Kernel matching. - Findings (Table 11): - All three algorithms show frequency of banking crises is significantly lower in countries with at least one IMF-supported program in previous five years than in matched non-treated countries. - Magnitude of PSM effect is lower than IV estimates but consistently supports association between IMF programs and lower incidence of future systemic banking crises. - Diagnostics: - Mean and median absolute standard percentage biases are dramatically reduced by matching. - Mean absolute bias below critical value of 5 for all algorithms, indicating matching removed significant differences in observables on average. - Some significant differences persist after matching in per capita GDP, inflation, and ratio of M2 over reserves. - Pseudo-R2 from the probit selection equation is very low in the matched group, as expected. - Supporting detail: - Estimates of the PS selection equation (reported in Appendix Table A3) show IMF programs are more likely in countries with lower per capita GDP, less financial and trade integration, after episodes of sovereign debt crises, with higher debt liabilities and lower portfolio and FDI liabilities, depreciating currencies, and higher real interest rates; degree of political similarity with US/G7 and past elections are significant predictors of IMF programs. *Italic source: _wp1519 - 5.3 The reform channel (PDF chapter/section) — canonical source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2015/_wp1519.pdf* ### 6.6 Alternative lag structures ### 6.6 Alternative lag structures ### Robustness check: alternative lag definitions of IMF participation - IMF participation dummy (IMFARRANGEMENT) is redefined to equal one if the country signs at least one Fund-supported program in the previous 3, 4, 6 or 7 years. - Column 1 reports the baseline specification when the IMF participation is assessed over the 5-year window. - Findings: - The positive effect of Fund-supported programs on banking stability materializes in the short- and medium-run, as the IMF participation effect is statistically significant even when the Fund presence in a country is evaluated over shorter horizons (3 and 4 years). - When the time window is expanded (6 and 7 years), the coefficient, although negative, becomes smaller in absolute value and is less precisely estimated. - These findings suggest the effects of the programs tend to vanish over time, when many other macroeconomic and institutional developments could offset the positive effect of the IMF-supported program. ### Interpretation of lag-structure results - Short- and medium-run significance implies either immediate stabilization effects from IMF programs or effects that manifest within a few years, consistent with liquidity support or near-term reform implementation. - Diminishing magnitude and precision with longer windows (6–7 years) indicate that program effects are not permanent and may be overtaken by subsequent developments. ### 7 Conclusions ### Main empirical conclusion - Standard OLS estimates are biased by reverse causality: a higher probability of a banking crisis increases the likelihood of IMF intervention. - After accounting for endogeneity of the IMF’s presence, results consistently show that countries which signed IMF lending programs are less likely to experience a banking crisis in future years than other comparable countries. - Interpretation: bad signals and moral hazard effects, if present, are offset by the credit availability provided by the IMF-supported program. ### Mechanisms and heterogeneous effects - The crisis prevention role of the Fund is connected to: - Macroeconomic policies and financial reforms associated with the loan. - A direct liquidity support effect. - Additional findings: - Other things being equal, the likelihood of a banking crisis is lower for borrowers which are compliant with IMF conditionalities than for non-compliant ones. - The negative association between IMF-supported programs and the probability of future banking crises is significant only when the loan is sufficiently large—consistent with benefits operating through liquidity support. *Source: _wp1519 - 6.6 Alternative lag structures* ### REFERENCES ### _wp1519 - REFERENCES ### Major cited literature - Extensive references on IMF programs, financial crises, banking crises determinants, catalytic effects of IMF lending, moral hazard, and empirical methods. Representative authors and works include ABIAD and MODY (2005), DEMIRGUC-KUNT and DETRAGIACHE (1998, 2005), KAMINSKY and REINHART (1999), LAEVEN and VALENCIA (2013), GOURINCHAS and OBSTFELD (2012), DREHER (2004, 2009), and many IMF Working Papers and NBER chapters. - Methodological references include CALIENDO and KOPEINIG (2008) on propensity score matching, ROSENBAUM and RUBIN (1985) on matched sampling, and STOCK and YOGO (2005) on weak-instrument testing. ### Determinants of banking crises (Table 1: selected review) - Empirically identified covariates with reported sign of association: - Growth: Growth (-) - Real interest rate (RIR): RIR (+) - Inflation: Inflation (+) - M2/reserve: M2/reserve (+) - Private credit: Private credit (+) - Deposit insurance: Deposit insurance (+) in several studies - Law & order: Law & order (-) - pc GDP: pc GDP (-) - RER depreciation / Overvalued ER / RER depreciation: mixed (RER depreciation (+) in several) - Credit growth: Credit growth (+) - Capital inflows: Capital inflows (+) - Financial liberalization / reforms: Financial liberalization (+/- depending on conditioning) - Banking sector concentration and restrictions: concentration (-), restrictions (+) - Reserves: Reserves (-) - Notes: RIR = real interest rate; RER = real exchange rate; FE = Fixed effects; RE = Random effects. Xt and Xt−n indicate contemporaneous or lagged covariates. ### Data, variables, and sample (Table 2: definitions and summary statistics) - Key variable definitions and sample means (on sample of 113 countries used in empirical analysis): - BANKING CRISIS: Laeven and Valencia (2013); Mean = 0.03; s.d. = 0.17; Obs. = 2,250 - CURRENCY CRISIS: Mean = 0.03; s.d. = 0.18; Obs. = 2,250 - SOVEREIGN CRISIS: Mean = 0.01; s.d. = 0.10; Obs. = 2,250 - FINANCIAL CRISES (any): Mean = 0.07; s.d. = 0.29; Obs. = 2,250 - IMF ARRANGEMENT t−1,t−5: Dummy for IMF loan in previous five years; Mean = 0.60; s.d. = 0.49; Obs. = 2,250 - IMF LOAN/GDP t−1,t−5: Log(1 + amount approved as share of GDP); Mean = 0.84; s.d. = 0.85; Obs. = 2,250 - NON COMPLIANCE t−1,t−5: Dummy coded 1 if compliant (≤25% undrawn at expiration); Mean = 0.50; s.d. = 0.50; Obs. = 2,250 - PROXIMITY t−6,t−10: Average alignment with G7 votes; Mean = 0.62; s.d. = 0.07; Obs. = 2,250 - ELECTION t−6,t−10: Executive election dummy; Mean = 0.48; s.d. = 0.50; Obs. = 2,250 - GDP PC (log real GDP per capita): Mean = 6.79; s.d. = 1.08; Obs. = 2,248 - CREDIT GROWTH (Δ domestic credit to private sector % GDP): Mean = 0.04; s.d. = 0.16; Obs. = 2,225 - M2/RESERVE: Mean = 0.07; s.d. = 0.17; Obs. = 2,236 - OPENNESS (merchandise trade % GDP): Mean = 0.56; s.d. = 0.28; Obs. = 2,243 - DEPOSIT INSURANCE dummy: Mean = 0.21; s.d. = 0.41; Obs. = 1,600 - FINANCIAL REFORMS index (0–1): Mean = 0.50; s.d. = 0.25; Obs. = 947 ### Empirical findings — IMF lending and banking crises (selected coefficients and tests) - Table 3 (descriptive): - Conditional on no past crisis: Probability of banking crisis = 3.25 (banking crises column) vs conditional on past crises = 1.82; reported T-test p-values include 0.060*, 0.021**, 0.047**, 0.037** for various comparisons. - Panel B: In whole sample, tranquil years without IMF program = 3,673 observations (98.03%), with IMF program tranquil = 295 (95.04%); crisis years without IMF = 74 (1.97%), with IMF = 34 (4.96%); p-value = 0.000*** on mean comparison. - Table 4 (Baseline regressions; Prob(BANKING CRISISt)): - IMF ARRANGEMENT t−1,t−5 coefficient reported in various specifications: 0.010 (OLS, col 1); -0.076**; -0.110***; -0.076** in 1st-stage/2SLS columns (bracketed clustered s.e. reported). Exact bracketed standard errors shown in table. - IMF LOAN/GDP t−1,t−5: -0.064** [0.028] (reported) - GDP PC t−1: coefficients include 0.002, -0.077***, -0.003, -0.063**, -0.005, -0.160***, -0.007, -0.067***, -0.003 with associated standard errors in brackets. - GDP GROWTH t−1: -0.153* [0.086] (among other column estimates) - SHORT TERM DEBT t−1: 0.071* [0.041]; other columns show large negative 2SLS estimates (e.g., -0.679***) - CREDIT GROWTH t−1: 0.034* [0.018] (positive association) - OPENNESS t−1: -0.033*** [0.013] (negative association) - DEBT LIABILITIES t−1: 0.007; in some columns 0.136*** [0.043] - FDI LIABILITIES t−1: -0.035*** [0.010] - Banking, sovereign, and currency crisis lags enter with various coefficients (e.g., BANKING CRISIS t−1,t−8: -0.025*** [0.007]) - 1st-stage instrument coefficients: - PROXIMITY t−6,t−10: 1.113*** [0.295] - ELECTION t−6,t−10: 0.155*** [0.038] - Instrument diagnostics (selected): - K-P weak identification (F-test): 18.552, 15.072, 12.859 (depending on columns) - C-D weak identification (F-test): 72.449 (for bootstrapped) - K-P underidentification (p-value): 0.000 (multiple columns) - Overidentification test (p-values): 0.650, 0.284, 0.852, 0.622 - Table 5 (The credit channel; Dep Var: Loan size as share of GDP; IMF ARRANGEMENT t−1,t−5 coefficients by crisis probability buckets): - Columns (1)-(6) <1% to >=3%: coefficients include -0.027 [0.045], -0.118*** [0.042], -0.045 [0.032], -0.134** [0.059], -0.063* [0.037], -0.110* [0.056] - Columns (7)-(12) <4% to >=6%: coefficients include -0.068* [0.038], -0.102* [0.058], -0.067* [0.037], -0.124* [0.075], -0.067* [0.037], -0.120 [0.081] - 1st-stage: PROXIMITY and ELECTION strong and significant across specifications (e.g., PROXIMITY 1.227*** [0.359]) - Table 6 (The reform channel; compliance and WB lending): - IMF ARRANGEMENT t−1,t−5: -0.206** [0.094] (2SLS); -0.206*** [0.079] (LIML) in columns (1)-(2) - IMF LOAN/GDP t−1,t−5: -0.112* [0.057] and -0.121** [0.055] in columns (3)-(4) - WB LENDING t−1,t−5: -0.152** [0.075] and -0.154** [0.077] in columns (5)-(6) - NON COMPLIANCE t−1,t−5: 0.132** [0.058], 0.132*** [0.047], 0.086** [0.041], 0.092** [0.037] (positive association with crisis risk) - 1st-stage instrument coefficients: PROXIMITY 0.614** [0.250]; ELECTION 0.055* [0.031]; MOVEMENT 1.771*** [0.667] - Table 7 (IMF programs not following financial crises; conditional on no past crisis): - IMF ARRANGEMENT t−1,t−5: -0.080** [0.037], -0.086** [0.038], -0.082** [0.038], -0.069** [0.032] across conditional specifications (banking, currency, sovereign) - 1st-stage: PROXIMITY 1.325*** [0.255]; ELECTION 0.139*** [0.036] - Table 8 (Additional controls; Prob(BANKING CRISISt)): - IMF ARRANGEMENT t−1,t−5 coefficients across specifications range from -0.285* [0.156] to -0.047* [0.027] with multiple reported standard errors. - SOUND MONEY t−1: -0.013*** [0.005] in column (5) - DEPOSIT INSURANCE t−1: 0.031* [0.018] in column (11) - 1st-stage: PROXIMITY and ELECTION coefficients reported across columns (e.g., PROXIMITY 1.081*** [0.326]; ELECTION 0.123*** [0.034]) - Table 9 (Sample splits by institutional setting): - IMF ARRANGEMENT t−1,t−5 coefficients vary by sub-sample (examples): - COUNTRY RISK Low: -0.146 [0.258]; High: -0.135** [0.061] - SOUND MONEY Low: -1.091 [1.793]; High: -0.124* [0.074] - POLITY Low: -0.142 [0.199]; High: -0.123** [0.060] - DEMOCRACY No: -0.240 [0.205]; Yes: -0.062 [0.039] - Heterogeneity indicates varying significance and magnitudes across institutional contexts. - Table 10 (US AID instrument and sub-samples): - IMF ARRANGEMENT t−1,t−5: -0.066** [0.033] (column 1 with additional instrument), -0.187** [0.077] (column 2 crisis subsample), other sub-sample estimates include -0.058* [0.035], -0.075** [0.035], -0.073** [0.034] - 1st-stage: PROXIMITY 0.902** [0.388]; ELECTION 0.181*** [0.041]; US AID t−6,t−10 coefficient = 0.262** [0.102] reported as instrument in one specification. - Table 11 (Propensity score matching — IMF participation effect): - ATT (IMF participation effect on treated): - 5-nearest neighbor: -0.032; s.e. = 0.016; t-Statistic = -2.030** - Radius: -0.039; s.e. = 0.015; t-Statistic = -2.570*** - Kernel: -0.035; s.e. = 0.014; t-Statistic = -2.490*** - Matched group bias and pseudo-R2 indicators reported (e.g., Matched group Mean bias = 4.4, Median bias = 4.4 for 5-nearest neighbor). - Observations: Treatment = 994; Control varies (e.g., 469 for nearest neighbor). - Table 12 (Different lag structures for IMF ARRANGEMENT): - IMF ARRANGEMENT t−1,t−5: -0.076** [0.035] (5-year) - IMF ARRANGEMENT t−1,t−3: -0.062* [0.037] (3-year) - IMF ARRANGEMENT t−1,t−4: -0.072* [0.037] (4-year) - IMF ARRANGEMENT t−1,t−6: -0.049 [0.033] (6-year) - IMF ARRANGEMENT t−1,t−7: -0.027 [0.035] (7-year) - Robustness across lag choices with variation in magnitude and significance. ### Additional appendices and sample - Appendix tables include country-level sample listing with counts of banking crises and IMF programs (Table A1; total: Banking crises = 653, IMF programs = 903, LIC indicator total = 38). - LIML robustness and extended diagnostics reported in Tables A2–A4 and probit selection equation for PSM (Table A3) with reported coefficients (e.g., PROXIMITY t−1,t−5 = 3.000*** [0.817]; US AID t−1,t−5 = 0.652*** [0.235]). *Source: _wp1519 - REFERENCES (PDF)._ --- _Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2015/_wp1519.pdf_