## wpiea2026128-source-pdf

## Source details

**Canonical URL:** [wpiea2026128-source-pdf](https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026128-source-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2026/english/wpiea2026128-source-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2026/english/wpiea2026128-source-pdf.pdf.json)

---

### Introduction — climate shock–sovereign risk nexus and research agenda
- Climate disasters amplify sovereign risk through multiple channels: (i) macroeconomic impacts, (ii) fiscal impacts of disasters and fiscal pressures from adaptation and mitigation policies, (iii) climate related risks and financial sector stability, (iv) depletion of natural capital, and (v) impacts of climate change on trade and political stability.
- Debt typically rises after disasters via financing recovery and reconstruction costs, which can push climate vulnerable countries toward default.
- Adaptation reduces future damages but involves high upfront costs; if financed by debt at market terms it may increase debt ratios today and accentuate debt vulnerabilities, while delaying adaptation may preserve debt ratios today but increase future vulnerabilities — an “impossible trilemma.”

Illustrative disaster magnitudes (selected):
- Grenada, Hurricane Ivan (2004): total losses about US$1.1 billion representing 200 percent of GDP; debt rose from 80 percent of GDP in 2003 to 95 percent of GDP in 2004.
- Grenada, Hurricane Beryl (2024): estimated to have cost about a third of Grenada’s GDP.
- Dominica, Hurricane Maria (2017): about US$931 million in damages and US$380 million in losses, representing about 226 percent of 2016 GDP; projected 21 percent external current account deficit for 2018.
- Cyclones Idai and Kenneth (2019): losses over $873 million across Mozambique, Malawi and Zimbabwe; World Bank recovery operation of $191.7 million.
- Europe: estimated €822 billion in economic losses across Europe between 1980 and 2024.
- U.S.: more than 400 climate-related events since 1980, with total damages exceeding US$2.9 trillion.
- Global infrastructure need: roughly US$90 trillion in sustainable climate resilient infrastructure needed by 2030.

Core research question and empirical approach:
- Can investing in climate adaptation using concessional financing build resilience, reduce disaster-induced damages and lower default probability without imperiling debt sustainability?
- Three-step empirical approach:
  1. Estimate impact of shocks on default probabilities using a rare-events adjusted methodology and a cross-country database of climate shocks merged with default episodes.
     - Finding: climate shocks have a positive statistically and economically significant impact on default probability, robust across estimations and income groups, and robust to an alternative default dataset split by creditor type.
  2. Assess whether adaptive capacity mitigates economic damages from climate disasters.
     - Finding: higher adaptive capacity is associated with a reduced marginal economic impact of additional disasters.
  3. Estimate whether access to concessional financing (ODA) contributes to building adaptive capacity.
     - Finding: ODA (nominal US$ and percent of GDP) tends to be positively associated with improvements in adaptive capacity (associational evidence; ODA likely endogenous).
- Back-of-the-envelope mapping translates estimated coefficients into US dollar magnitudes and ODA-to-GDP ratios required to offset climate disaster–induced increases in default probability.
  - Conclusion from mapping: ODA is associated with stronger adaptive capacity, which in turn tends to attenuate damages and lower default probability.

Contributions:
- Empirical innovations: use event-level macroeconomic loss measures, address rare-event bias in sovereign default estimation, and translate estimates into US dollar and ODA-to-GDP policy yardsticks.
- Policy relevance: evaluates how concessional financing can strengthen adaptive capacity and attenuate climate-shock effects on sovereign default probabilities.

### Data and methodology — construction of variables and panel
- Sample: unbalanced country-year panel 1995–2020 for 178 countries.

Sovereign default measures:
- Primary: Asonuma and Trebesch (2016) comprehensive default database (monthly converted to annual); total default events: 196.
  - Default dummy = 1 during default spell, 0 otherwise.
- Complementary: Bank of Canada/Bank of England (BoC/BoE) sovereign default database:
  - Country-year considered in default when stock of arrears increases by more than 100 percent vs previous year and previous year’s arrears positive.
  - Creditor splits: official (IMF, IBRD, IDA, Paris Club, China, other official), private (FX banks, FX bonds, other private), domestic (LC debt, domestic arrears).
- Merged default episodes with EM-DAT disaster-level economic damages to construct panel linking climate shocks to sovereign defaults.

Disaster events and damage metric (EM-DAT):
- EM-DAT records natural disasters; inclusion criteria: 10 fatalities, 100 affected people, state of emergency, or call for international assistance.
- Disasters of interest: drought, extreme temperature, flood, landslide, storms.
- Main variable: “Total Economic Damage” aggregated across events per country-year and scaled by nominal GDP → damage as a share of GDP.
  - Disaster counts range 0 to 34 events per country-year.
  - About 44 percent of country-year observations record no disasters; a small share record more than 19 in a year.
  - EM-DAT coverage for Total Damage: 183 countries and 1406 country–year datapoints.
  - Insured Damage available for 86 countries and 402 country–years (about 71 percent missing among those observations).
  - Note: EM-DAT economic losses are systematically underreported; damage/GDP should be interpreted as a lower bound.
- Robustness: constructed precision-weighted climate shock index à la Felbermayr & Gröschl (2014).

Adaptive capacity proxy and index:
- Adaptive capacity proxied using ND-GAIN’s adaptive capacity component indicators across:
  - Food (Agricultural capacity);
  - Water (Dam capacity, Access to reliable water);
  - Ecosystems (Ecological footprint, Engagement in international environmental conventions);
  - Habitat (Paved roads);
  - Infrastructure (Projected change of hydropower generation capacity).
- Adaptive capacity index constructed by averaging indicators; higher values denote more adapted countries.

Concessional financing proxy (ODA):
- Proxy: Official Development Assistance (ODA) disbursements from OECD-DAC CRS.
  - Since 2018, OECD-DAC reports ODA on grant equivalent basis.
  - DAC statistics: about 85 percent of ODA and 64 percent of bilateral concessional financing are provided as grants.
- Caveats: ODA imperfect proxy for adaptation finance; climate-tagged ODA not used as main measure because adaptation marker available only post 2009.
- Constructed for each country-year:
  - (i) cumulative flow of ODA up to year t (millions USD);
  - (ii) cumulative ODA stock as a share of GDP.

Macroeconomic controls and panel:
- Controls: real GDP growth, log real GDP per capita, inflation, debt/GDP, financial development (domestic credit to private sector/GDP), terms of trade, REER (where available), disaster frequency.
- Final panel dataset: 1995–2020 for 178 countries.

Stylized facts (descriptive):
- Stylized Fact 1: Higher ND-GAIN vulnerability associated with sharply higher predicted default probability (model-implied probability rises from about 0.2 to over 0.8 as vulnerability rises from 5 to 40 percent on 0–100 scale).
- Stylized Fact 2: Defaults are rare (~4 percent of country-year datapoints); a small set of sovereigns account for roughly half of episodes. In BoC/BoE data, official creditors most commonly involved via “other official” (~59.4 percent) and Paris Club (~9 percent).
- Stylized Fact 3: Disaster losses highly clustered; most countries lose well under 1 percent of GDP in a typical year, a small subset repeatedly faces large losses.
- Stylized Fact 4: ODA modest on average, concentrated in LIDCs:
  - Median ODA/GDP (1995–2020): All 1.72 percent, EMEs 0.59 percent, LIDCs 4.85 percent.
  - Mean ODA/GDP: All 4.45 percent, EMEs 2.27 percent, LIDCs 7.16 percent.
  - Median (mean) annual ODA in USD: All US$231 million (US$655 million), EMEs US$158 million (US$543 million), LIDCs US$386 million (US$770 million).
  - Maximum annual ODA (USD millions): All 21,700, EMEs 21,700, LIDCs 12,400.

Summary statistics (Table A.1, full sample 1995–2020):
- Disaster damage (% GDP): Observations 4574; Mean 0.30; Std. Dev. 3.53; Min 0.00; Max 148.41
- GDP growth (%): Observations 4556; Mean 3.50; Std. Dev. 5.98; Min -50.34; Max 147.98
- log (GDP per capita): Observations 4500; Mean 9.36; Std. Dev. 1.16; Min 6.30; Max 11.87
- Inflation rate (%): Observations 4528; Mean 29.40; Std. Dev. 1019.19; Min -72.73; Max 6537.08
- Debt-to-GDP ratio (%): Observations 4182; Mean 55.60; Std. Dev. 45.11; Min 0.00; Max 657.84
- Private credit (% GDP): Observations 3713; Mean 48.25; Std. Dev. 42.64; Min 0.00; Max 304.58
- log (Terms of Trade): Observations 4059; Mean 4.59; Std. Dev. 0.22; Min 3.00; Max 5.61
- log (REER): Observations 2334; Mean 4.59; Std. Dev. 0.18; Min 3.76; Max 6.61
- Default dummy: Observations 4628; Mean 0.041; Std. Dev. 0.198; Min 0.00; Max 1

### Empirical strategy and main results
Empirical strategy (three-step chain):
- Step 1: Estimate Pr(default_{i,t}=1) = Λ(α + β damage_{i,t-1} + γ′ X_{i,t-1} + μ_i + λ_t) with damage lagged one year; use Firth’s bias reduced logit for rare-event correction.
- Step 2: damage_{i,t} = α_i + λ_t + β dis_{i,t} + θ adapt_{i,t} + φ·(adapt_{i,t} × dis_{i,t}) + γ′ X_{i,t} + ε_{i,t} to test attenuation (φ expected < 0).
- Step 3: adapt_{i,t} = α_i + λ_t + β tot_oda_{i,t-1} + γ′ X_{i,t-1} + ε_{i,t} (tot_oda in millions USD) and adapt_{i,t} = α_i + λ_t + β tot_oda_GDP_{i,t-1} + γ′ X_{i,t-1} + ε_{i,t} (tot_oda_GDP in pp of GDP). Map ODA → adaptation → damages → default using chain rule/back-of-the-envelope.

Effect of disaster damages on sovereign default (selected estimates):
- Sovereign defaults are rare (~4 percent of the sample).
- A one percentage point increase in disaster-related losses raises the odds of default by approximately 2–3 percent (robust across logit and rare-logit).
- Selected coefficient estimates (statistical significance preserved):
  - Column (1) Logit ALL: damage (% GDP) = 0.025 ** (standard error 0.010)
  - Column (2) Rare Logit ALL: damage (% GDP) = 0.026 ** (standard error 0.011)
  - Column (3) Rare Logit EME: damage (% GDP) = 0.021 * (standard error 0.012)
  - Column (4) Rare Logit LIDC: damage (% GDP) = 0.097 (standard error 0.120)
- Other controls:
  - debt (% GDP) example: 0.051 *** (SE 0.010) in Column (1).
  - GDP growth example: -0.119 * (SE 0.070) in Column (1).
- Robustness: excluding pre-2000 observations leaves results essentially unchanged for full sample and EMEs.
- Comparison context: ICE Physical Risk Score predicts about 3.7 percent increase in sovereign default probability; Fitch (2025) and other literature referenced for benchmark default rates.

Adaptation attenuates disaster damages (Table 2, interaction results):
- Main interaction estimates (statistical significance preserved):
  - adaptive_capacity:
    - All: 0.008 (SE 0.010)
    - EME: 0.033 (SE 0.029)
    - LIDC: -0.004 (SE 0.005)
  - disaster frequency:
    - All: 0.346 *** (SE 0.103)
    - EME: 0.513 *** (SE 0.161)
    - LIDC: 0.275 (SE 0.187)
  - interaction (adapt × disaster frequency):
    - All: -0.004 *** (SE 0.002)
    - EME: -0.007 *** (SE 0.002)
    - LIDC: -0.005 (SE 0.004)
- Interpretation (margins):
  - An additional disaster increases damages by ≈ 0.35 percentage points of GDP at very low adaptation for all countries; ≈ 0.51 percentage points for EMEs.
  - Attenuation: for the full sample, effect shrinks by 0.04 pp of GDP for every 10-point increase in adaptation index (interaction -0.004 per 1 point → -0.04 per 10 points).
  - In EMEs attenuation stronger: 0.07 pp shrinkage per 10 points (interaction -0.007).

ODA, adaptive capacity, and mapping to default (Tables 3a/3b and chain rule)
- Table 3a (total ODA in millions USD → adaptive capacity):
  - total ODA:
    - All: 0.00013 *** (SE 0.00003)
    - EME: 0.00009 ** (SE 0.00004)
    - LIDC: 0.00015 ** (SE 0.00006)
  - Economic interpretation: US$1 billion (1,000 million) more in cumulative ODA since 1995 associated with a 0.13 point increase in adaptive capacity.
  - Average annual ODA inflow: $632 million.
  - Cumulative ODA received since 1995: $4.3 billion (sample average).
- Table 3b (ODA as share of GDP → adaptive capacity):
  - total ODA/GDP:
    - All: 0.0066 * (SE 0.0038)
    - EME: 0.0049 (SE 0.0041)
    - LIDC: 0.0164 ** (SE 0.0075)
  - Interpretation: 1 pp increase in ODA/GDP associated with a 0.66-point increase in adaptive capacity for full sample; 1.64-point increase in LIDCs.

Back-of-the-envelope mapping (chain rule; ψ ≈ 2.6 odds per 1 pp damage from Table 1):
- Using Table 3a (US$1 billion → κ = 0.13 adaptive capacity increase):
  - κ = 0.13 lowers damages (% GDP) by 4.3pp at disaster frequency = 10 and by 2.1pp at disaster frequency = 5 (from Equation (2) margins).
  - With ψ = 2.6, maps to about 11 pp (10 disasters) and 5.5 pp (5 disasters) reduction in default probability per US$1bn of ODA.
- Using Table 3b (1 pp ODA/GDP → κ = 0.66 adaptive capacity increase):
  - κ = 0.66 lowers damages by 3.9pp (10 disasters) and 1.7pp (5 disasters).
  - With ψ = 2.6, maps into change in default probability of approximately -10 pp (disaster frequency = 10) and -4.4 pp (disaster frequency = 5).
- Note: These back-of-the-envelope calculations are illustrative, using average marginal effects and associational estimates rather than causal claims.

Hazard-specific and robustness findings:
- Precision-weighted climate shock index constructed for robustness; baseline shock measure retained: damage as share of nominal GDP.
- Hazard-specific rare-events logit (Table A.3): only lagged storm intensity raises next-year default odds; other hazard-specific measures not predictive for full sample, EME, or LIDCs. Damage (% of GDP) remains robustly predictive and is kept as baseline.
- Creditor-specific analysis using BoC/BoE database performed for robustness; IMF program dummy included as diagnostic control — disaster shocks remain predictive of higher default probability even after conditioning on IMF program status.
  - Example from Table A.2 (Rare events logit with IMF program dummy, Column (1) ALL):
    - damage (% GDP) coefficient 0.029** (0.011)
    - IMF program dummy coefficient 1.201*** (0.268)
    - Observations 3031

### Policy-relevant conclusions, implications, and caveats
Key empirical conclusions:
- Larger disasters in year t−1 are associated with higher odds of sovereign default in year t.
- Building adaptive capacity attenuates disaster damages; interaction negative and economically significant, strongest in EMEs.
- ODA disbursements are associated with higher adaptive capacity; modest ODA increases correspond to measurable improvements in adaptive capacity that reduce damages and implied default probabilities.

Policy implications:
- Concessional financing can mitigate climate-induced debt distress by enabling adaptation investment that reduces macroeconomic impacts of climate shocks.
- Results provide a practical budgeting yardstick linking targeted reductions in default probability to annual concessional (ODA) needs by country and disaster frequency for fiscal frameworks and donor programs.
- Where concessional financing is constrained, concessional lending instruments that catalyze private climate financing (e.g., RSF) are important to build resilience while maintaining debt sustainability.

Caveats and avenues for future research:
- Estimates are reduced-form associations; sequential estimation chosen due to recursive chain ODA → adaptation → damages → default.
- EM-DAT economic losses likely underreported; LIDC results may be affected by underreporting and small sample size.
- ODA–adaptation link is associational; donor allocation not random and ODA may fund other objectives, raising endogeneity concerns.
- Future work could construct more robust measures combining disaster intensity, frequency, and exposure, and explore exogenous variations in lending terms as instruments for adaptation finance.

*IMF Working Paper — wpiea2026128-source-pdf*

### Introduction ...........................................................................................................

### Introduction

### Climate shock–sovereign risk nexus
- Climate disasters are increasing in both frequency and severity and can amplify sovereign risk through multiple channels: (i) macroeconomic impacts, (ii) fiscal impacts of disasters and fiscal pressures from adaptation and mitigation policies, (iii) climate related risks and financial sector stability, (iv) depletion of natural capital, and (v) impacts of climate change on trade and political stability.
- Debt typically rises after disasters via financing recovery and reconstruction costs, which can push climate vulnerable countries toward default.
- Adaptation can reduce future damages but involves high upfront costs; if financed by debt at market terms it may increase debt ratios today and accentuate debt vulnerabilities, while delaying adaptation may preserve debt ratios today but increase future vulnerabilities — an “impossible trilemma.”

### Illustrative disaster episodes and magnitudes
- Grenada, Hurricane Ivan (2004): total losses about US$1.1 billion representing 200 percent of GDP; debt rose from 80 percent of GDP in 2003 to 95 percent of GDP in 2004, followed by debt restructuring over 2004-2006. Hurricane Beryl (2024) estimated to have cost about a third of Grenada’s GDP.
- Dominica, Hurricane Maria (2017): about US$931 million in damages and US$380 million in losses, representing about 226 percent of 2016 GDP; projected 21 percent external current account deficit for 2018.
- Cyclones Idai and Kenneth (2019): losses over $873 million in buildings, equipment, infrastructure and crops across Mozambique, Malawi and Zimbabwe; World Bank emergency recovery and resilience operation of $191.7 million.
- Europe: estimated €822 billion in economic losses across Europe between 1980 and 2024, with over one quarter of the cost incurred between 2021 and 2024.
- U.S.: more than 400 climate-related events since 1980, with total damages exceeding US$2.9 trillion.
- Global infrastructure need: roughly US$90 trillion in sustainable climate resilient infrastructure needed by 2030 (Global Commission on the Economy and Climate, 2016).

### Macroeconomic and fiscal consequences
- Climate disasters tend to:
  - Depress output growth and cause losses countries do not fully recover from.
  - Raise consumer price inflation with persistent effects on food prices.
  - Widen fiscal gaps through recovery and reconstruction costs.
  - Affect firm-level productivity and the cost of debt.
  - Cause sovereign distress that may lead to debt defaults.

### Research questions and methodological approach
- Core question: Can investing in climate adaptation using concessional financing build resilience, reduce disaster-induced damages and lower default probability without imperiling debt sustainability?
- Three-step empirical approach:
  1. Revisit the relationship between climate shocks and sovereign debt defaults by estimating impact of shocks on default probabilities using a methodology adapted for rare events and a cross-country database of climate shocks merged with default episodes (Asonuma & Trebesch, 2016).
     - Finding: climate shocks have a positive statistically and economically significant impact on default probability, robust across different estimations, income groups and to the use of an alternative default dataset from the Bank of Canada/Bank of England which allows splits by creditor type.
  2. Assess whether a country’s adaptive capacity mitigates economic damages from climate disasters.
     - Finding: holding other factors constant, higher adaptive capacity is associated with a reduced marginal economic impact of additional disasters.
  3. Examine whether access to concessional financing contributes to building adaptive capacity by estimating panel regressions of adaptive capacity on Official Development Assistance (ODA) disbursements.
     - Finding: ODA (in both nominal US$ terms and in percent of GDP) tends to be positively associated with improvements in adaptive capacity.
     - Note: ODA disbursements are likely endogenous; results should be interpreted as associations rather than causal effects.
- Back-of-the-envelope mapping:
  - Estimated coefficients are mapped to implied changes in damages and default probabilities using a partial equilibrium calculation, translating coefficients into budget-relevant US dollar magnitudes and ODA to GDP ratios required to offset climate disaster–induced increases in default probability.
  - Conclusion from mapping: ODA is associated with stronger adaptive capacity, which in turn tends to attenuate damages and lower default probability.

### Contributions of the paper
- Empirical innovations:
  - Replace broad structural vulnerability measures with climate event–level macroeconomic loss measures tying the climate-default link to realized shocks.
  - Address rare event bias in sovereign default estimations by using a methodology tailored to such events.
  - Price the policy lever by translating estimated coefficients into US dollar magnitudes and ODA-to-GDP ratios required to offset the effect of climate disasters on default probability.
- Policy relevance:
  - Evaluates how access to concessional financing (ODA) can strengthen adaptive capacity and attenuate the effect of climate shocks on sovereign default probabilities.

*IMF Working Paper — Introduction section*

### 3. Data and methodology

### 3. Data and methodology

### Sovereign default measures and construction
- Primary default measure: Asonuma and Trebesch (2016) comprehensive default database.
  - Sovereign default defined as a failure to meet debt service obligations and a non-honoring of original debt agreements (includes missed coupon payments and event-type restructurings with either no legal default or no unilateral default episode).
  - Database at monthly frequency converted to annual basis; country marked “in default” throughout duration of the default spell.
  - Default variable = 1 when the country is considered in default, and 0 otherwise.
  - Total default events in this measure: 196 default events.
- Complementary default measure: Bank of Canada/Bank of England (BoC/BoE) sovereign default database.
  - Provides nominal value of government debt in default by type and creditor (foreign and local currency denominated bank loans, bonds and debt by multilateral organizations, Paris Club and other private creditors).
  - Country considered in default in a year when the stock of arrears increases by more than 100 percent compared to the previous year, provided the previous year’s stock of arrears is positive.
  - Default variable set to 1 in those years, and 0 otherwise.
  - Robustness: default dataset also split by creditor:
    - official – IMF, IBRD, IDA, Paris Club, China, other official;
    - private – foreign currency (FX) banks, FX bonds, other private;
    - domestic – local currency (LC) debt, domestic arrears.
- Both databases used as published without extending temporal coverage.
- Merged default episodes with disaster event level economic damages from EM-DAT to construct a panel linking climate shocks to sovereign defaults across 178 countries from 1995 to 2020.

### Disaster events and economic damage metric (EM-DAT)
- Disaster event source: Emergency Events Database (EM-DAT).
  - EM-DAT records natural disasters (meteorological, hydrological, climatological, geophysical and biological) since 1900.
  - A record enters EM-DAT if it has at least one of the following: 10 fatalities, 100 affected people, a state of emergency declared or a call for international assistance.
- Disasters of interest: meteorological and hydrological events — drought, extreme temperature, flood, landslide and storms.
- Multi-year disasters (e.g., droughts) enter the panel according to the year in which EM-DAT records the event and its associated damages.
- Main variable of interest: economic impact variable “Total Economic Damage”.
  - Aggregated across all events in a given country and year and scaled by nominal GDP.
  - This produces a per-country-year severity metric (damage as a share of GDP).
  - Disaster counts range from 0 to 34 events per country-year.
  - About 44 percent of country-year observations record no disasters; a small share record more than 19 in a year.
- EM-DAT coverage for damage variable in sample: Total Damage is available for 183 countries and 1406 country–year datapoints; Insured Damage is available for only 86 countries and 402 country–years (about 71 percent of observations missing among country–year datapoints with Insured Damage).
- Note: EM-DAT notes that Economic losses are systematically underreported; damage/GDP should be interpreted as a lower bound.
- Robustness: constructed an exogenous precision weighted climate shock index à la Felbermayr & Gröschl (2014) (see appendix).

### Adaptive capacity proxy and index construction
- Adaptive capacity proxied using ND-GAIN’s adaptive capacity component indicators from:
  - (i) Food, with Agricultural capacity;
  - (ii) Water, with Dam capacity and Access to reliable water;
  - (iii) Ecosystems, with Ecological footprint and Engagement in international environmental conventions;
  - (iv) Habitat, with Paved roads;
  - (v) Infrastructure, with Projected change of hydropower generation capacity.
- Adaptive capacity index constructed by averaging across the indicators for each country observation.
- Higher index values denote more adapted countries.

### Concessional financing proxy (ODA) and construction
- Concessional financing defined as below market rate financing (grants, low to zero interest loans with longer maturities or grace periods).
- Proxy: Official Development Assistance (ODA) disbursements from OECD-DAC Creditor Reporting System (CRS).
  - ODA mainly composed of government aid grants or soft loans from bilateral and multilateral donors.
  - Since 2018, OECD-DAC reports ODA on a grant equivalent basis (provides a measure of loan concessionality).
  - DAC statistics: about 85 percent of ODA and 64 percent of bilateral concessional financing are provided as grants.
- Caveats:
  - ODA is an imperfect proxy for concessional resources supporting adaptive capacity; not all ODA is adaptation-related and only a part is directly connected to climate resilience.
  - Climate-tagged ODA not used as main measure because adaptation marker available only post 2009, which would truncate the 1995–2020 panel.
  - Green Climate Fund (GCF), Global Environment Facility (GEF), and Adaptation Fund data series are shorter, project based and not reported in OECD-DAC grant equivalent terms; therefore not used.
- For each country-year, constructed:
  - (i) cumulative flow of ODA up to year t;
  - (ii) cumulative ODA stock as a share of GDP (used in step 3 of empirical strategy to scale financing needs to country size).
- Disbursements used rather than commitments to capture actual financial flows.

### Macroeconomic variables and panel construction
- Macroeconomic sources: IMF World Economic Outlook (WEO), International Financial Statistics (IFS), World Bank World Development Indicators (WDI).
- Included macro variables (following Cevik & Jalles (2022a) and solvency literature):
  - real GDP growth and GDP per capita, inflation, debt as a share of GDP, financial development measured by domestic credit to the private sector as a share of GDP, terms of trade, real effective exchange rate (REER) index (where available).
- Also included disaster frequency as a control.
- Final panel dataset: 1995–2020 for 178 countries.

### Stylized Facts (descriptive patterns motivating empirical strategy)
- Stylized Fact 1: Countries with a higher vulnerability measure face significantly higher predicted default probability.
  - Conditional fixed effects logit of default probability on ND-GAIN vulnerability index shows predicted default probability rises steeply as vulnerability increases.
  - Model-implied probability: as vulnerability rises from 5 to 40 percent (0–100 scale), predicted probability rises from about 0.2 to over 0.8.
  - Beyond a vulnerability measure of 45 percent, default probability remains elevated.
  - These are descriptive correlations and not causal effects; conditional logit can suffer small sample bias in rare events.
- Stylized Fact 2: Defaults are rare and differ by creditor type.
  - In the panel, default occurs in about 4 percent of country-year datapoints.
  - A small set of sovereigns (e.g., Cote d’Ivoire, Republic of Congo, Iraq, Tanzania, Serbia etc.) feature repeated defaults, accounting for roughly half of all observed episodes.
  - In BoC/BoE creditor-disaggregated panel (1995–2020), official creditors are most commonly involved in default episodes largely via “other official” creditors (about 59.4 percent) and Paris Club (about 9 percent); IDA, China, IMF and IBRD each less frequent.
  - On the private side, “other private” and foreign currency banks feature more often.
  - Literature comparisons: D’Erasmo and Mendoza (2019) simulated a 1.21 percent long-run frequency of domestic defaults; Reinhart-Rogoff (2008) evidence of 1.1 percent.
- Stylized Fact 3: Disaster losses are highly clustered and unevenly distributed.
  - Most countries lose well under 1 percent of GDP to disasters in a typical year, but a small subset repeatedly faces large macroeconomic losses.
  - Geographic pattern (2000 vs. 2020): Central America and the Caribbean, South and East Asia and parts of Sub-Saharan Africa have disproportionate share of high damage observations.
  - Literature projections (Kahn et al. (2019), Volz et al. (2020)): global GDP per capita projected to be between 1.07 and 7.2 percent lower by 2100 under standard IPCC pathways.
- Stylized Fact 4: ODA is modest on average, concentrated in Low-Income Developing Countries (LIDCs).
  - Over 1995–2020:
    - median ODA to GDP ratio is 1.72 percent for the full sample, 0.59 percent for EMEs and 4.85 percent for LIDCs.
    - means are 4.45, 2.27 and 7.16 percent respectively.
    - median (mean) annual ODA in dollar terms: US$231 million (US$655 million) overall, US$158 million (US$543 million) for EMEs, US$386 million (US$770 million) for LIDCs.
  - Tabulated ODA summary (1995–2020):
    - ODA as Share of GDP (%) — Median: All 1.72, EMEs 0.59, LIDCs 4.85
    - ODA as Share of GDP (%) — Mean: All 4.45, EMEs 2.27, LIDCs 7.16
    - ODA as Share of GDP (%) — Maximum: All 90.99, EMEs 77.13, LIDCs 90.99
    - ODA Disbursements (USD millions) — Median Annual Value: All 231, EMEs 158, LIDCs 386
    - ODA Disbursements (USD millions) — Mean Annual Value: All 655, EMEs 543, LIDCs 770
    - ODA Disbursements (USD millions) — Maximum Annual Value: All 21,700, EMEs 21,700, LIDCs 12,400
  - Notes: Figures summarize country-year observations of ODA from 1995–2020.

*Source: IMF Working Paper — 3. Data and methodology (content unit: wpiea2026128-source-pdf).*

### 5. Empirical Strategy and Results

### 5. Empirical Strategy and Results

### Empirical strategy and identification
- Three-step strategy:
  - Step 1: Estimate how climate shocks affect sovereign debt default probability using a logit specification following Cevik and Jalles (2022a) but measuring climate risk with realized economic damages (total economic damages from drought, extreme temperature, flood, landslide, storm scaled by nominal GDP) and a different sovereign default database.
  - Step 2: Regress economic damages from climate shocks on adaptive capacity, including an interaction between adaptation and disaster frequency.
  - Step 3: Estimate the relationship between ODA concessional finance and adaptive capacity, and map ODA → adaptation → damages → default via back-of-the-envelope calculations.
- Main default specification (Equation (1)):
  - Pr(default_{i,t}=1) = Λ(α + β damage_{i,t-1} + γ′ X_{i,t-1} + μ_i + λ_t)
  - damage_{i,t-1}: total economic damages from climate-related hazards scaled by nominal GDP (lagged one year).
  - X: lagged macro controls (real GDP growth, log real GDP per capita, inflation, debt/GDP, financial development, terms of trade).
  - Country/income group and year fixed effects included where applicable; standard errors clustered at the country level.
- Methodological adjustments for rarity and separation:
  - Sovereign defaults are rare (~4 percent of the sample).
  - Use Firth’s bias reduced logit (Jeffreys prior penalty) to correct small-sample bias and stabilize estimates when quasi-separation is a risk.

### Effect of disaster damages on sovereign default (Table 1)
- Key empirical findings:
  - Sovereign defaults are rare events (~4 percent of the sample).
  - Depending on the sample, a one percentage point increase in disaster-related losses raises the odds of default by approximately 2-3 percent; results are robust across standard logit and rare-events logit models.
  - For EMEs, the results remain significant; for LIDCs the coefficient is positive but not statistically significant.
  - Robustness checks: excluding pre-2000 observations leaves results essentially unchanged for the full sample and EMEs.
  - Comparison references:
    - Long-horizon average global default rates: about 7 percent for advanced economies and about 17 percent for developing countries (Qian & Roch, 2024).
    - Fitch ratings (2025): default rate fell to 1.9 percent in 2024 from 5.5 percent for speculative grade debt; overall sovereign default rate is 0.9 percent from 2.6 percent in 2023.
    - ICE Physical Risk Score predicts an increase in sovereign default probability of about 3.7 percent.
- Selected coefficient estimates from Table 1 (statistical significance levels preserved):
  - Column (1) Logit ALL: damage (% GDP) = 0.025 ** (standard error 0.010)
  - Column (2) Rare Logit ALL: damage (% GDP) = 0.026 ** (standard error 0.011)
  - Column (3) Rare Logit EME: damage (% GDP) = 0.021 * (standard error 0.012)
  - Column (4) Rare Logit LIDC: damage (% GDP) = 0.097 (standard error 0.120)
- Other notable controls (examples from Table 1):
  - debt (% GDP) positive and significant in several specifications (e.g., 0.051 *** in Column (1), SE 0.010).
  - GDP growth tends negative on default probability in several specifications (e.g., -0.119 * in Column (1), SE 0.070).

### Adaptation attenuates disaster damages (Equation (2), Table 2, Figure 2)
- Specification (Equation (2)):
  - damage_{i,t} = α_i + λ_t + β dis_{i,t} + θ adapt_{i,t} + φ·(adapt_{i,t} × dis_{i,t}) + γ′ X_{i,t} + ε_{i,t}
  - dis: disaster frequency across five climate event types; adapt: adaptive capacity index; φ expected < 0 (attenuation).
- Main Table 2 estimates (statistical significance preserved):
  - adaptive_capacity:
    - All: 0.008 (SE 0.010)
    - EME: 0.033 (SE 0.029)
    - LIDC: -0.004 (SE 0.005)
  - disaster frequency:
    - All: 0.346 *** (SE 0.103)
    - EME: 0.513 *** (SE 0.161)
    - LIDC: 0.275 (SE 0.187)
  - interaction (adapt × disaster frequency):
    - All: -0.004 *** (SE 0.002)
    - EME: -0.007 *** (SE 0.002)
    - LIDC: -0.005 (SE 0.004)
- Interpretation and margins:
  - An additional disaster increases damages by approximately 0.35 percentage points of GDP at very low adaptation for all countries, and by approximately 0.51 percentage points of GDP for EMEs.
  - The effect shrinks by 0.04 percentage points for every 10-point increase in the adaptation index for the full sample (i.e., interaction implies attenuation of 0.004 per 1 point of adaptation → 0.04 per 10 points).
  - In EMEs attenuation is stronger: 0.07 percentage points shrinkage per 10 points (interaction -0.007).
  - Figure 2 margins: damages increase with disaster frequency under low adaptation but slopes are flatter under higher adaptation; damages decline monotonically with adaptation, steeper when disaster frequency is higher. For LIDCs differences are not statistically significant.

### ODA, adaptive capacity, and mapping to default (Equations (3a)/(3b), Tables 3a/3b, chain rule)
- Two fixed effects panels estimated:
  - (3a) adapt_{i,t} = α_i + λ_t + β tot_oda_{i,t-1} + γ′ X_{i,t-1} + ε_{i,t}, where tot_oda is cumulative stock of ODA disbursements in millions of USD (lagged one year).
  - (3b) adapt_{i,t} = α_i + λ_t + β tot_oda_GDP_{i,t-1} + γ′ X_{i,t-1} + ε_{i,t}, where tot_oda_GDP is cumulative ODA as share of GDP (lagged one year).
- Table 3a (ODA in millions USD) key estimate:
  - total ODA:
    - All: 0.00013 *** (SE 0.00003)
    - EME: 0.00009 ** (SE 0.00004)
    - LIDC: 0.00015 ** (SE 0.00006)
  - Economic interpretation: a beta of 0.00013 means US$1 billion (1,000 million) more in total ODA accumulated since 1995 is associated with a 0.13 point increase in adaptive capacity.
  - Average annual ODA inflow: $632 million.
  - Cumulative ODA received since 1995: $4.3 billion.
- Table 3b (ODA as share of GDP) key estimate:
  - total ODA/GDP:
    - All: 0.0066 * (SE 0.0038)
    - EME: 0.0049 (SE 0.0041)
    - LIDC: 0.0164 ** (SE 0.0075)
  - Economic interpretation: a 1 pp increase in ODA/GDP is associated with a 0.66-point increase in adaptive capacity for the full sample and a 1.64-point increase in LIDCs.
- Mapping chain rule (Equation (4)) and back-of-the-envelope calculations:
  - Δp(dis) ≈ ψ · Δdam(κ; dis)
    - Δp: change in default probability per US$1 billion of ODA (or 1 pp increase in ODA/GDP).
    - ψ: average marginal effect of disaster damages on default probability from Table 1 (quoted as "2.6 higher odds per 1 percentage point of disaster damages").
    - Δdam(κ; dis): model predicted change in damages (% of GDP) per κ point increase in adaptive capacity from a US$1 billion ODA or 1 pp ODA/GDP, evaluated at a given disaster frequency dis.
  - Using Table 3a (US$1 billion → κ = 0.13 increase in adaptive capacity):
    - Low/high example calculations:
      - κ = 0.13 lowers damages (% of GDP) by 4.3pp at a 10 disaster frequency and 2.1pp at a disaster frequency of 5 (using Equation (2) margins).
      - With ψ = 2.6 pp (higher default probability per 1 pp increase in damages), this maps to about 11 pp (for 10 disasters) and 5.5 pp (for 5 disasters) reduction in default probability per US$1bn of ODA.
  - Using Table 3b (1 pp increase in ODA/GDP → κ = 0.66 point rise in adaptive capacity):
    - κ = 0.66 lowers damages (% of GDP) by 3.9pp (10 disasters) and 1.7pp (5 disasters).
    - With ψ = 2.6 pp, this maps into a change in default probability of approximately -10 pp (disaster frequency = 10) and -4.4 pp (disaster frequency = 5).
  - Notes: All these effects are back-of-the-envelope calculations using estimated average marginal effects and should be read as economically relevant magnitudes rather than causal effects.

### Key policy-relevant conclusions and implications
- Core empirical conclusions:
  - Countries face higher odds of sovereign default following larger disasters in the previous year.
  - Building adaptive capacity attenuates disaster damages; interaction term negative and economically significant, strongest in EMEs.
  - ODA disbursements are associated with higher adaptive capacity; modest increases in ODA correspond to measurable improvements in adaptive capacity that in turn reduce damages and implied default probabilities.
- Policy implications:
  - Concessional financing can mitigate climate-induced debt distress by enabling adaptation investment that reduces the macroeconomic impact of climate shocks.
  - The results provide a practical budgeting yardstick linking target reductions in default probability to annual concessional (ODA) needs by country and disaster frequency for fiscal frameworks and donor programs.
  - In contexts of constrained concessional financing, concessional lending instruments that catalyze private climate financing (e.g., RSF) play an important role in building climate resilience while maintaining debt sustainability.
- Caveats and avenues for future research:
  - Estimates are reduced-form associations; sequential estimation chosen due to recursive chain ODA → adaptation → damages → default.
  - EM-DAT economic losses likely underreported; LIDC results may be affected by underreporting and small sample size.
  - ODA – adaptation link is associational; donor allocation is not random and ODA may be used for other purposes, raising endogeneity concerns.
  - Future work could construct more robust measures combining disaster intensity, frequency, and exposure, and explore exogenous variations in lending terms as instruments for adaptation finance.

*Source: IMF Working Paper — 5. Empirical Strategy and Results (section content provided in the source PDF).*

### 7.  References

### 7. References and 8. Appendix (selected items and empirical summaries)

### Key bibliographic coverage
- References cite literature on sovereign default, climate impacts on macroeconomic and sovereign risk, disaster databases, and climate data sources including:
  - Agarwala et al. (2021); Arellano (2008); Asonuma & Trebesch (2016); Beers et al. (2021a, 2021b); Beguería et al. (2023); Beirne et al. (2021); Bhattacharya et al. (2023); Black et al. (2023); Boehm (2022); Calcaterra et al. (2025a, 2025b); Cevik & Jalles (2022a, 2022b, 2023); Cevik & Miryugin (2023); Choirat et al. (2025); Cruces & Trebesch (2013); D’Erasmo & Mendoza (2021); Delforge et al. (2025); Diallo et al. (2026); Duffy (2025); Felbermayr & Gröschl (2014); Fitch Ratings (2025); Gahtan et al. (2024); Global Commission on the Economy and Climate (2016); Government of the Commonwealth of Dominica (2017); Hatchondo & Martinez (2009); Heinze & Schemper (2002); Hsiang & Jina (2014); ICE Sustainable Finance (2024); IMF (various); Joint Research Centre (2024); Kahn et al. (2019); King & Zeng (2001); Kirschbaum et al. (2010, 2015); Kling et al. (2018, 2021); Klusak et al. (2023); Knapp et al. (2010); Mallucci (2022); Masum (2018); Mejia (2014); Mihalache (2025); Muñoz Sabater (2019); NOAA NCEI (2025); ND-GAIN (2025); Parker (2018); Phan & Schwartzman (2024); Qian (2012); Rauniyar et al. (2025); Reinhart & Rogoff (2008); Reinhart & Trebesch (2016); Seghini (2024a, 2024b); Strobl (2011); UNEP (2025); Volz et al. (2020); World Bank (various).
- Data sources referenced for hazard and disaster measures include: SPEIbase v. 2.7; EM-DAT; IBTrACS (Version 4r01); ERA5-Land; JRC RP50 flood maps; NASA landslide data; Accumulated Cyclone Energy (ACE) from IBTrACS; NOAA NCEI U.S. Billion-Dollar Weather and Climate Disasters; ND-GAIN; World Bank DataBank; IMF World Economic Outlook database (April 2025).

### Appendix A — Sample and summary statistics (full sample, 1995–2020, 178 countries)
- Sample description:
  - The sample is an unbalanced country-year panel over 1995–2020 across 178 countries.
- Table A.1. Summary statistics (full sample, 1995–2020):
  - Disaster damage (% GDP): Observations 4574; Mean 0.30; Std. Dev. 3.53; Min 0.00; Max 148.41
  - GDP growth (%): Observations 4556; Mean 3.50; Std. Dev. 5.98; Min -50.34; Max 147.98
  - log (GDP per capita): Observations 4500; Mean 9.36; Std. Dev. 1.16; Min 6.30; Max 11.87
  - Inflation rate (%): Observations 4528; Mean 29.40; Std. Dev. 1019.19; Min -72.73; Max 6537.08
  - Debt-to-GDP ratio (%): Observations 4182; Mean 55.60; Std. Dev. 45.11; Min 0.00; Max 657.84
  - Private credit (% GDP): Observations 3713; Mean 48.25; Std. Dev. 42.64; Min 0.00; Max 304.58
  - log (Terms of Trade): Observations 4059; Mean 4.59; Std. Dev. 0.22; Min 3.00; Max 5.61
  - log (REER): Observations 2334; Mean 4.59; Std. Dev. 0.18; Min 3.76; Max 6.61
  - Default dummy: Observations 4628; Mean 0.041; Std. Dev. 0.198; Min 0.00; Max 1
- Notes: Table A.1 reports summary statistics for the full sample over 1995–2020.

### Robustness: creditor-specific default probabilities and IMF program diagnostic control
- Creditor-specific analysis:
  - The BoC/BoE sovereign default database is considered for robustness; default probability is estimated by creditor type using creditor-specific rare events logit specification. Controls included but not shown; country income-group and year fixed effects included.
  - Figure A.1 (described): plots predicted default probabilities separately by creditor type for the full sample (sources: BoC/BoE sovereign default database, compiled by Authors).
- Diagnostic control for IMF program episodes (Table A.2 — Rare events logit with IMF program dummy):
  - Dependent variable: Default dummy
  - Column (1) ALL:
    - damage (% GDP) coefficient 0.029** (0.011)
    - IMF program dummy coefficient 1.201*** (0.268)
    - Observations 3031
    - Income Group FE Yes; Year FE Yes
  - Column (2) EME:
    - damage (% GDP) coefficient 0.023* (0.012)
    - IMF program dummy coefficient 1.041*** (0.384)
    - Observations 1469
    - Income Group FE No; Year FE Yes
  - Column (3) LIDC:
    - damage (% GDP) coefficient 0.091 (0.128)
    - IMF program dummy coefficient 1.198*** (0.438)
    - Observations 939
    - Income Group FE No; Year FE Yes
  - Notes: Standard errors in parentheses. Controls included but not shown. * p < 0.10, ** p < 0.05, *** p < 0.01.
  - Interpretation provided in text: "The results showed that disaster shocks continue to predict higher default probability. Disaster losses remain positively associated with next year’s default, even after conditioning on IMF program status in the previous year. The IMF program dummy is included to account for programs that occur in periods of acute stress, sometimes overlapping with restructurings. We interpret this as a diagnostic control and not a causal control."

### Climate shock index construction and hazard-specific measures
- Methodology summary (precision-weighted climate shock index, à la Felbermayr & Gröschl (2014)):
  - Construct exogenous precision-weighted climate shock index by combining hazard-specific measures and indicators on droughts, extreme temperature events, floods, landslides and storms.
  - Each hazard component defined as deviation of an annual climate variable from its long run mean; anomalies aggregated using inverse-variance weights to prevent dominance by any single hazard.
  - Components and data sources:
    - Drought intensity: Standardized Precipitation Evapotranspiration Index (SPEI-12); monthly gridded data at 0.5° spatial resolution; averaged over the year and sign-flipped for larger values to signify drier years.
    - Extreme temperature/heat: 2m temperature anomaly against 1981-2010 climatology baseline from ERA5-Land (European Center for Medium Range Weather Forecasts Reanalysis v5).
    - Flooding: product of long-term floodable land share from Joint Research Centre RP50 map and precipitation anomaly vs. 2000-2020 baseline (ERA5).
    - Landslide intensity: NASA data (primary source).
    - Storm intensity: IBTrACS v4 — constructed annual Accumulated Cyclone Energy (ACE); landslide and storm measures normalized by land area (per 100 km^2) before precision weighting.
  - Baseline shock measure used in main analysis: damage as a share of nominal GDP; the precision-weighted index used in robustness checks.

### Hazard-specific rare events logit results (Table A.3)
- Results using individual disaster intensity measures in place of aggregate damage (% of GDP); dependent variable: default dummy. Country and Year fixed effects included; controls included but not reported.
- Coefficients (standard errors in parentheses) and sample sizes:
  - (1) drought:
    - default dummy coefficient -0.296** (0.131)
    - Observations 2638; Country FE Yes; Year FE Yes
  - (2) heat:
    - default dummy coefficient -0.337 (0.275)
    - Observations 1920; Country FE Yes; Year FE Yes
  - (3) flood:
    - default dummy coefficient 0.009 (0.180)
    - Observations 1862; Country FE Yes; Year FE Yes
  - (4) landslide:
    - default dummy coefficient -0.641** (0.276)
    - Observations 2156; Country FE Yes; Year FE Yes
  - (5) storm:
    - default dummy coefficient 0.306* (0.157)
    - Observations 387; Country FE Yes; Year FE Yes
  - Significance: * p < 0.10, ** p < 0.05, *** p < 0.01
- Text interpretation:
  - "Only lagged storm intensity measure raises next-year default odds. Other hazards and the combined composite index created are not predictive for the full sample, EME or LIDCs. In contrast, damage (% of GDP) is robustly predictive. We keep it as the baseline shock measure."

*Climate Shocks, Debt Defaults and Investment in Climate Adaptation – Squaring an Impossible Trilemma, Working Paper No. WP/2026/128*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2026/english/wpiea2026128-source-pdf.pdf_