## wpiea2022253-print-pdf

## Source details

**Canonical URL:** [wpiea2022253-print-pdf](https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf.json)

---

### Sample Characteristics and Damage Distribution
- An average country in the global sample rarely experienced a disaster whose damage exceeds 10 percent of GDP, whereas a quarter does so in our sample.
- For our sample:
  - Real GDP (USD in billion): 462 252
  - Real GDP per capita (USD in thousand): 5.312
  - Land area (in million of square kms): 0.32
  - Population (in million of persons): 8.6195
  - Pecuniary damage (USD in billion): 0.2 0.8
  - Damage ratio (in percent of GDP): 1.8 0.3
- For our sample, only around a quarter had their damage falling into the 0.5 percent range; the sample has a much thicker right tail of the distribution, implying larger average damage and greater exposure to extremely severe disasters.

### High-Damage Events and ECCU Concentration
- Five of the top ten high-damage events and nine of the top twenty between 1970 and 2020 occurred in the Eastern Caribbean Currency Union (ECCU).
- The most extreme listed: Hurricane Maria hit Dominica in 2017, causing an immediate damage around 224 percent of GDP.
- Except for an episode in Mongolia and in Honduras, all highly damaging disasters occurred either in the Caribbean or small island developing countries.

### Synthetic Controls Evidence of Scarring (Selected Episodes)
- Good pre-event fit for Antigua and Barbuda (1995) and St. Kitts and Nevis (1998); treated and synthetic series diverge after events and take around 5 to 10 years to return to the counterfactual—consistent with medium-term scarring.
- Extremely large damages as share of GDP: 60 percent for Antigua and Barbuda in 1995, and 49 percent for St. Kitts and Nevis in 1998.
- Dominica episodes: 1995 damage-to-GDP ratio was 64 percent; 2007 episode was 5 percent. The 1995 episode shows a stronger divergence consistent with larger damage.
- Long-run widening of the gap in real GDP per capita between affected countries and synthetic controls observed in some episodes; could reflect scarring or other shocks (left for future research). Current focus is on medium-term impact (five years after shocks).

### Quantifying Medium-Term Effects of Natural Disasters on Real GDP per Capita
- Estimation results for impulse responses to a 1 percent larger decline in real GDP per capita on impact (Table 3):
  - Low-damage events (cost of damage below 10 percent of GDP):
    - Year 1: 0.841 percent
    - Year 2: 0.769 percent
    - Year 3: 0.574 percent
    - Year 4: 0.478 percent
    - Year 5: 0.203 percent
    - Interpretation: Pattern suggests little scarring for small disaster shocks.
  - High-damage events (cost of damage exceeds 10 percent of GDP):
    - Year 1: 1.357 percent
    - Year 2: 1.049 percent
    - Year 3: 0.943 percent
    - Year 4: 0.891 percent
    - Year 5: 0.911 percent
    - Interpretation: Effects are more persistent and one cannot reject constancy over time.
- Table 3 additional statistics:
  - High-Damage Dummy coefficients by year: 0.0226, 0.00929, 0.00513, 0.00537, 0.0117
  - Constant terms by year: -0.0145*; -0.0253***; -0.0400***; -0.0506***; -0.0577***
  - Observations: 71 for each column (Years 1–5).
  - R-squared: 0.786, 0.677, 0.442, 0.311, 0.181 (Years 1–5 respectively).
  - Significance notation: *** p<0.01, ** p<0.05, * p<0.1.

### Impulse Response Illustration: Severe Shock Example
- Example shock: economic damage around 30 percent of GDP causing a 2 percent decline in real GDP per capita in the year of the disaster.
- Elasticity estimated by regressing the decline in real GDP per capita on reported pecuniary damage (no intercept); impulse responses combine coefficients of change in real GDP per capita on impact with intercepts to capture other recovery-influencing factors.
- Resulting impulse response (Figure 5):
  - Five years after the shock, real GDP per capita remains around 8 percent below the trend, composed of:
    - Around 2.3 percent from the contribution of the decline in real GDP per capita on impact.
    - Around 5.7 percent from the intercept.
- The large intercept contribution (5.7 percent) is unexplained within this paper and left for future research.

### Placebo Tests and Robustness
- Placebo regressions: regress the GDP gap in years -5 to -1 on the GDP gap in year -6 (no disasters occurred in these years) using the same specification as equation (2).
- Table 4 suggests placebo regressions do not replicate the significant patterns in Table 3; the only significant effect found is on the GDP gap of year - (text truncated in source).
- Additional placebo-test results (Table 4, sample Observations = 35 for each specification):
  - Change in GDPPC in Year 6 × Low-damage dummy — Year 5: 0.691*** (0.197); Year 4: 0.0309 (0.223); Year 3: -0.0153 (0.232); Year 2: -0.0711 (0.208); Year 1: 0.308 (0.227).
  - Change in GDPPC in Year 6 × High-Damage Dummy — Year 5: -0.0266 (0.244); Year 4: -0.2990 (0.276); Year 3: 0.00288 (0.287); Year 2: -0.1870 (0.258); Year 1: 0.227 (0.280).
  - High Damage Dummy main effect — Year 5: 0.0496* (0.0260); Year 4: 0.0611** (0.0295); Year 3: 0.0388 (0.0306); Year 2: 0.0426 (0.0275); Year 1: 0.0223 (0.0299).
  - Constants (per year): Year 5: 0.00563 (0.0139); Year 4: -0.00678 (0.0157); Year 3: -0.0128 (0.0163); Year 2: -0.0145 (0.0147); Year 1: -0.0241 (0.0160).
  - R-squared: Year 5: 0.366; Year 4: 0.123; Year 3: 0.069; Year 2: 0.074; Year 1: 0.131.
- Note: Using a coarser SCA fit or the same sample as Table 3 yields broadly similar placebo-test results.
- It is also useful to recognize that the intercept is insignificantly different from zero in this exercise.

### Exploring Underlying Channels — Overview and Hypotheses
- Potential mechanisms for persistent output losses considered:
  - Stochastic trend in real GDP per capita (implies persistent decline after adverse shocks) — inconsistent with small disasters having only transitory effects.
  - Radical political evolution (political revolution) — deemed unlikely as a major source given frequency of severe disasters in sample.
  - Lack of resources to recover damaged infrastructure — plausible channel for persistent suppression of output per capita in sample countries that face binding resource constraints.
- Supporting evidence: Von Peter, Von Dahlen, and Saxena (2012) find uninsured losses drive macroeconomic costs; Duarte et al. (2022) shows limited disaster insurance in Caribbean countries.
- Hypotheses tested:
  - Hypothesis I: High indebtedness of a country can exacerbate the medium-term output loss of a severe natural disaster.
  - Hypothesis II: During a severe natural disaster episode, tourism and agriculture sectors have stronger medium-term output loss than other sectors (implicit assumption: physical assets of tourism and agriculture are more exposed).

### Testing Hypothesis I: Role of Initial Conditions (Equation (3))
- Structural indicators considered (SS_{i,j,T0_i-1}) — four variables:
  1. Dummy: public debt over GDP exceeds 70 percent.
  2. Political stability indicator.
  3. Government effectiveness indicator.
  4. Tourism dependency (tourism receipts over GDP).
- Threshold choice: 70 percent for high-debt dummy; results not sensitive to threshold choice.
- Findings (Table 5 summary):
  - ∆ real GDP per capita main effect:
    - Debt exceeding 70 percent of GDP: 0.403*** (0.148)
    - Tourism dependency: 0.539*** (0.196)
    - Political stability: 0.413** (0.189)
    - Government effectiveness: 0.422* (0.216)
  - Triple interaction term (∆ real GDP per capita × Dummy_damage × Structural indicator):
    - Debt exceeding 70 percent of GDP: 1.147* (0.598)
    - Tourism dependency: 0.0628 (0.130)
    - Political stability: 17.70 (54.72)
    - Government effectiveness: -0.991 (18.37)
  - ∆ real GDP per capita × Structural indicator:
    - Debt: -0.913*** (0.254)
    - Tourism: -0.0540* (0.0273)
    - Political stability: -0.0154 (0.344)
    - Government effectiveness: 0.125 (0.421)
  - Dummy_damage × Structural indicator and Dummy_damage main effect reported (see source for values).
  - Constants:
    - Debt: -0.0599*** (0.0123)
    - Tourism: -0.0756*** (0.0159)
    - Political stability: -0.0505*** (0.0135)
    - Government effectiveness: -0.0447*** (0.0147)
  - Observations:
    - Debt model: 68
    - Tourism model: 53
    - Political stability model: 42
    - Government effectiveness model: 42
  - R-squared:
    - Debt: 0.345
    - Tourism: 0.341
    - Political stability: 0.393
    - Government effectiveness: 0.285
- Interpretation:
  - Highly indebted countries exhibit stronger medium-term output loss after a severe natural disaster.
  - Illustrative example in text: For a 1 percent decline in real GDP per capita on impact (from a high-damage event with damage-to-GDP ratio exceeding 10 percent), the fifth-year impact would be stronger by 1.24 percent of GDP for countries whose public debt exceeds 70 percent of GDP than the rest of the sample.
  - Political stability, government effectiveness, and tourism dependency do not show robust significant roles in affecting medium-term output effects.

### Testing Hypothesis II: Heterogeneous Output Effects across Sectors
- Data: Sectoral GDP for 15 Caribbean economies and five sectors: tourism, agriculture, construction, manufacturing, high-skilled services (including real estate and financial services).
- Sample selection: Only disasters with damage-to-GDP ratio exceeding the median within this sectoral sample are kept.
- Empirical method:
  - Step 1: Estimate event-study specification for log sectoral GDP with episode-sector fixed effects (equation (4)).
  - Step 2: Fit linear trend to pre-event β_k in the 10-year pre-event window, extrapolate to post-event, bootstrap pre-event β_k to estimate confidence bands (bootstrap ignores cross-β correlations across episodes).
- Findings:
  - Tourism sector: natural disasters have significantly negative impact on GDP for ECCU countries and countries with debt-to-GDP ratio above the sectoral-sample median.
  - Pattern not present for non-ECCU countries or countries with debt-to-GDP ratio below the median.
  - Agriculture shows a similar but less salient pattern.
  - Construction, manufacturing, high-skilled services: weaker or no clear patterns.

### Impact on Import of Machinery and Equipment (M&E)
- Method: Replace dependent variable in equation (4) with import of M&E as share of GDP to test whether post-disaster repair crowds out M&E imports.
- Finding: Some weak evidence for ECCU countries that imports of M&E may be affected, suggesting potential crowding out of capital goods investment that could slow capital accumulation and TFP growth.

### Conclusion — Key Findings and Policy Implications
- Main result: For extremely severe events, real GDP per capita is persistently lower than the counterfactual trend even five years after the shock.
- Evidence suggests lack of resources to repair damaged assets is a plausible cause for persistent medium-term output losses.
- Policy implication: Strong support for strengthening resilience building ex-ante to reduce scarring effects of severe natural disasters.
- Suggested future research: Compare post-disaster economic dynamics of small island developing economies after severe shocks with disaster-prone regions in large economies to further understand mechanisms.

### Data Sources and Donor Pool Construction
- Primary datasets: April 2021 vintage of IMF World Economic Outlook (WEO) database; EM-DAT (CRED).
- Supplemental datasets: Penn World Tables 10.0; World Bank WGI and WDI; Google’s DSPL; sectoral GDP from Caribbean country authorities; UN COMTRADE import data.
- Donor pool construction algorithm (Annex II):
  - Step 1: Keep countries with non-missing predictor variables for at least one year between 1970 and T0_i.
  - Step 2: Keep countries whose nominal GDP is between 10 percent and 15 times that of affected country; select ten countries whose real GDP per capita is closest (aim for five richer, five poorer when possible).
  - Step 3: From those ten, select six countries whose real GDP per capita growth rate between T0_i-15 and T0_i is closest to that of the affected country.
  - Step 4: Define 31 country sets (one set of six; six five-country sets excluding each of the six; twenty-four sets removing each of the six and adding each of the four eliminated in Step 3).
  - Donor pool choice: country set with the lowest pre-event MSPE for its synthetic control. Results robust to choosing the second-lowest MSPE set and other variations.

*Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf*

### 0.5 percent.  For our sample, only around a quarter had their damage falling into this range. Meanwhile,

### wpiea2022253-print-pdf

### Sample Characteristics and Damage Distribution
- An average country in the global sample rarely experienced a disaster whose damage exceeds 10 percent of GDP, whereas a quarter does so in our sample.
- For our sample:
  - Real GDP (USD in billion): 462 252
  - Real GDP per capita (USD in thousand): 5.312
  - Land area (in million of square kms): 0.32
  - Population (in million of persons): 8.6195
  - Pecuniary damage (USD in billion): 0.2 0.8
  - Damage ratio (in percent of GDP): 1.8 0.3
- For our sample, only around a quarter had their damage falling into the 0.5 percent range; the sample has a much thicker right tail of the distribution, implying larger average damage and greater exposure to extremely severe disasters.

### High-Damage Events and ECCU Concentration
- Figure 3 (events with damage larger than 10 percent of GDP) highlights:
  - Five of the top ten high-damage events and nine of the top twenty between 1970 and 2020 occurred in the Eastern Caribbean Currency Union (ECCU).
  - The most extreme listed: Hurricane Maria hit Dominica in 2017, causing an immediate damage around 224 percent of GDP.
- Note: Except for an episode in Mongolia and in Honduras, all highly damaging disasters occurred either in the Caribbean or small island developing countries.

### Synthetic Controls Evidence of Scarring (Selected Episodes)
- Synthetic control analysis (SCA) for selected ECCU episodes shows:
  - Good pre-event fit for Antigua and Barbuda (1995) and St. Kitts and Nevis (1998); solid and dashed lines match well before events and diverge after, taking around 5 to 10 years for the treated series to return to the counterfactual—pattern consistent with medium-term scarring.
  - Extremely large damages as share of GDP: 60 percent for Antigua and Barbuda in 1995, and 49 percent for St. Kitts and Nevis in 1998.
  - Long-run widening of the gap in real GDP per capita between affected countries and synthetic controls observed in some episodes; could reflect scarring or other shocks (left for future research). Current focus is on medium-term impact (five years after shocks).
  - Dominica episodes: 1995 damage-to-GDP ratio was 64 percent; 2007 episode was 5 percent. The 1995 episode shows a stronger divergence consistent with larger damage.

### Quantifying Medium-Term Effects of Natural Disasters on Real GDP per Capita
- Estimation results (Table 3) for impulse responses to a 1 percent larger decline in real GDP per capita on impact:
  - Low-damage events (cost of damage below 10 percent of GDP): Year 1: 0.841 percent; Year 2: 0.769 percent; Year 3: 0.574 percent; Year 4: 0.478 percent; Year 5: 0.203 percent. Pattern suggests little scarring for small disaster shocks.
  - High-damage events (cost of damage exceeds 10 percent of GDP): Year 1: 1.357 percent; Year 2: 1.049 percent; Year 3: 0.943 percent; Year 4: 0.891 percent; Year 5: 0.911 percent. Effects are more persistent and one cannot reject constancy over time.
- Table 3 additional reported statistics:
  - High-Damage Dummy coefficients by year: 0.0226, 0.00929, 0.00513, 0.00537, 0.0117 (corresponding standard errors in parentheses).
  - Constant terms by year: -0.0145*; -0.0253***; -0.0400***; -0.0506***; -0.0577*** (standard errors in parentheses).
  - Observations: 71 for each column (Years 1–5).
  - R-squared: 0.786, 0.677, 0.442, 0.311, 0.181 (Years 1–5 respectively).
  - Significance notation: *** p<0.01, ** p<0.05, * p<0.1.

### Impulse Response Illustration: Severe Shock Example
- Example shock: economic damage around 30 percent of GDP causing a 2 percent decline in real GDP per capita in the year of the disaster.
- Calculation approach:
  - Elasticity estimated by regressing the decline in real GDP per capita on reported pecuniary damage (no intercept); impulse responses combine coefficients of change in real GDP per capita on impact with intercepts to capture other recovery-influencing factors.
- Resulting impulse response (Figure 5):
  - Five years after the shock, real GDP per capita remains around 8 percent below the trend, composed of:
    - Around 2.3 percent from the contribution of the decline in real GDP per capita on impact.
    - Around 5.7 percent from the intercept.
- The large intercept contribution (5.7 percent) is noted as unexplained within this paper and left for future research.

### Placebo Tests
- To check robustness against poor pre-event SCA fit, authors regress the GDP gap in years -5 to -1 on the GDP gap in year -6 (no disasters occurred in these years) using the same specification as equation (2).
- Table 4 suggests that placebo regressions do not replicate the significant patterns in Table 3; the only significant effect found is on the GDP gap of year - (text truncated in source).

*Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf*

### 5. It is also useful to recognize that the intercept is insignificantly different from zero in this exercise.

### 5. It is also useful to recognize that the intercept is insignificantly different from zero in this exercise.

### Placebo Test: Limited Response in GDP per Capita to a Year -6 Shock (Table 4)
- Sample: Observations = 35 for each specification (Columns (1)–(5)).
- Key coefficient estimates (Change in GDPPC in Year 6 × Low-damage dummy):
  - Year 5: 0.691*** (standard error 0.197)
  - Year 4: 0.0309 (0.223)
  - Year 3: -0.0153 (0.232)
  - Year 2: -0.0711 (0.208)
  - Year 1: 0.308 (0.227)
- Key coefficient estimates (Change in GDPPC in Year 6 × High-Damage Dummy):
  - Year 5: -0.0266 (0.244)
  - Year 4: -0.2990 (0.276)
  - Year 3: 0.00288 (0.287)
  - Year 2: -0.1870 (0.258)
  - Year 1: 0.227 (0.280)
- High Damage Dummy main effect:
  - Year 5: 0.0496* (0.0260)
  - Year 4: 0.0611** (0.0295)
  - Year 3: 0.0388 (0.0306)
  - Year 2: 0.0426 (0.0275)
  - Year 1: 0.0223 (0.0299)
- Constants (per year):
  - Year 5: 0.00563 (0.0139)
  - Year 4: -0.00678 (0.0157)
  - Year 3: -0.0128 (0.0163)
  - Year 2: -0.0145 (0.0147)
  - Year 1: -0.0241 (0.0160)
- R-squared:
  - Year 5: 0.366
  - Year 4: 0.123
  - Year 3: 0.069
  - Year 2: 0.074
  - Year 1: 0.131
- Standard errors in parentheses. Significance: ***p<0.01, **p<0.05, *p<0.1.
- Note: Using a coarser SCA fit or the same sample as Table 3 yields broadly similar placebo-test results.

### V. Exploring Underlying Channels — Overview
- Mechanisms considered for persistent output losses:
  - Stochastic trend in real GDP per capita (implies persistent decline after adverse shocks), but inconsistent with finding that small disasters have only transitory effects.
  - Radical political evolution (political revolution) — deemed unlikely as a major source given frequency of severe disasters in sample.
  - Lack of resources to recover damaged infrastructure — plausible channel for persistent suppression of output per capita in sample countries that face binding resource constraints.
- Supporting evidence: Von Peter, Von Dahlen, and Saxena (2012) find uninsured losses drive macroeconomic costs; Duarte et al. (2022) shows limited disaster insurance in Caribbean countries.

### Hypotheses Tested
- Hypothesis I: High indebtedness of a country can exacerbate the medium-term output loss of a severe natural disaster.
- Hypothesis II: During a severe natural disaster episode, tourism and agriculture sectors have stronger medium-term output loss than other sectors.
  - Implicit assumption: physical assets of tourism and agriculture are more exposed to severe natural disasters in the sample.

### B. Testing Hypothesis I: Role of Initial Conditions (Equation (3))
- Structural indicators (SS_{i,j,T0_i-1}) considered (four variables):
  1. Dummy: public debt over GDP exceeds 70 percent.
  2. Political stability indicator.
  3. Government effectiveness indicator.
  4. Tourism dependency (tourism receipts over GDP).
- Threshold choice: 70 percent for high-debt dummy; results not sensitive to threshold choice. Authors note 70 percent is not asserted as an appropriate level for other considerations (reference to Greenidge et al. (2012) estimating a 55-56 percent threshold for Caribbean countries).
- Findings (Table 5):
  - ∆ real GDP per capita main effect:
    - Column (1) Debt exceeding 70 percent of GDP: 0.403*** (0.148)
    - Column (2) Tourism dependency: 0.539*** (0.196)
    - Column (3) Political stability: 0.413** (0.189)
    - Column (4) Government effectiveness: 0.422* (0.216)
  - Triple interaction term (∆ real GDP per capita × Dummy_damage × Structural indicator):
    - Debt exceeding 70 percent of GDP: 1.147* (0.598)
    - Tourism dependency: 0.0628 (0.130)
    - Political stability: 17.70 (54.72)
    - Government effectiveness: -0.991 (18.37)
  - ∆ real GDP per capita × Dummy_damage:
    - Debt: 0.378 (0.316)
    - Tourism: 0.402 (0.961)
    - Political stability: -18.901 (59.84)
    - Government effectiveness: 0.98 (9.735)
  - ∆ real GDP per capita × Structural indicator:
    - Debt: -0.913*** (0.254)
    - Tourism: -0.0540* (0.0273)
    - Political stability: -0.0154 (0.344)
    - Government effectiveness: 0.125 (0.421)
  - Dummy_damage × Structural indicator:
    - Debt: 0.0189 (0.0470)
    - Tourism: 0.00246 (0.00775)
    - Political stability: -1.762 (4.404)
    - Government effectiveness: -1.755 (2.315)
  - Dummy_damage main effect:
    - Debt: -0.00129 (0.0354)
    - Tourism: -0.0101 (0.0639)
    - Political stability: 1.877 (4.773)
    - Government effectiveness: -0.161 (0.250)
  - Structural indicator main effect:
    - Debt: 0.0227 (0.0253)
    - Tourism: 0.00292 (0.00202)
    - Political stability: 0.0396** (0.0159)
    - Government effectiveness: 0.0285 (0.0185)
  - Constants:
    - Debt: -0.0599*** (0.0123)
    - Tourism: -0.0756*** (0.0159)
    - Political stability: -0.0505*** (0.0135)
    - Government effectiveness: -0.0447*** (0.0147)
  - Observations:
    - Debt model: 68
    - Tourism model: 53
    - Political stability model: 42
    - Government effectiveness model: 42
  - R-squared:
    - Debt: 0.345
    - Tourism: 0.341
    - Political stability: 0.393
    - Government effectiveness: 0.285
- Interpretation:
  - Highly indebted countries exhibit stronger medium-term output loss after a severe natural disaster.
  - Example: For a 1 percent decline in real GDP per capita on impact (from a high-damage event with damage-to-GDP ratio exceeding 10 percent), the fifth-year impact would be stronger by 1.24 percent of GDP for countries whose public debt exceeds 70 percent of GDP than the rest of the sample. (This numerical example is provided in the text as an illustrative magnitude.)
  - Political stability, government effectiveness, and tourism dependency do not show significant roles in affecting medium-term output effects.

### C. Testing Hypothesis II: Heterogeneous Output Effects across Sectors
- Data: Sectoral GDP for 15 Caribbean economies (ECCU members plus Haiti, The Bahamas, Barbados, Jamaica, Belize, Guyana, Trinidad and Tobago) and five sectors: tourism, agriculture, construction, manufacturing, high-skilled services (including real estate and financial services).
- Sample selection: Only natural disasters with damage-to-GDP ratio exceeding the median within this sectoral sample are kept.
- Empirical method:
  - Step 1: Estimate event-study specification for log sectoral GDP with episode-sector fixed effects (equation (4)).
  - Step 2: Fit a linear trend to pre-event β_k in the 10-year pre-event window, extrapolate to post-event, bootstrap pre-event β_k to estimate confidence bands (bootstrapping ignores cross-β correlations across episodes).
- Findings (Figure 6 summary in text):
  - Tourism sector: natural disasters have significantly negative impact on GDP for ECCU countries and countries with debt-to-GDP ratio above the sectoral-sample median.
  - Pattern not present for non-ECCU countries or countries with debt-to-GDP ratio below the median.
  - Agriculture: similar but less salient pattern.
  - Other sectors (construction, manufacturing, high-skilled services): weaker or no clear patterns.

### D. Impact on Import of Machinery and Equipment (M&E)
- Method: Replace dependent variable in equation (4) with import of M&E as share of GDP to test whether post-disaster repair crowds out M&E imports.
- Finding: Figure 7 shows some weak evidence for ECCU countries that imports of M&E may be affected, suggesting potential crowding out of capital goods investment that could slow capital accumulation and TFP growth.

### VI. Conclusion — Key Findings and Policy Implications
- Main result: For extremely severe events, real GDP per capita is persistently lower than the counterfactual trend even five years after the shock.
- Evidence suggests lack of resources to repair damaged assets is a plausible cause for persistent medium-term output losses.
- Policy implication: Strong support for strengthening resilience building ex-ante to reduce scarring effects of severe natural disasters.
- Suggested future research: Compare post-disaster economic dynamics of small island developing economies after severe shocks with disaster-prone regions in large economies to further understand mechanisms.

### Annex I. Data Sources and Definitions — Key datasets
- Primary: April 2021 vintage of IMF World Economic Outlook (WEO) database; EM-DAT (Emergency Events Database) by CRED.
- Supplemental: Penn World Tables 10.0; World Bank’s World Governance Indicators and World Development Indicators; Google’s DSPL; sectoral GDP from Caribbean country authorities; UN COMTRADE import data.
- Indicator sources (selected):
  - Natural disaster data: EM-DAT (CRED).
  - Cross-country GDP, population, inflation, public debt: IMF WEO (April 2021 vintage).
  - Latitude data: Google DSPL Guide.
  - Land area and tourism expenditures: World Bank WDI.
  - Institutional quality indicators: World Bank WGI.
  - Missing historical GDP: Penn World Tables 10.0.
  - Sectoral GDP: Caribbean country authorities.
  - Import data: United Nations' COMTRADE.

### Annex II. Donor Pool Construction Algorithm — Steps
- Step 1: Keep countries with non-missing predictor variables for at least one year between 1970 and T0_i.
- Step 2: Keep countries whose nominal GDP is between 10 percent and 15 times that of affected country; select ten countries whose real GDP per capita is closest (aim for five richer, five poorer when possible).
- Step 3: From those ten, select six countries whose real GDP per capita growth rate between T0_i-15 and T0_i is closest to that of the affected country.
- Step 4: Define 31 country sets:
  - One set: the six countries from Step 3.
  - Six sets: five-country sets excluding each of the six one at a time.
  - Twenty-four sets: remove each of the six one at a time but add each of the four countries eliminated in Step 3.
- Donor pool choice: country set with the lowest pre-event MSPE for its synthetic control. Results robust to choosing the second-lowest MSPE set and other variations.

*Source: IMF Working Paper "Natural Disasters and Scarring Effects" (Working Paper No. WP/2022/253), authors' estimates.*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2022/english/wpiea2022253-print-pdf.pdf_