## wpiea2025006-print-pdf

## Source details

**Canonical URL:** [wpiea2025006-print-pdf](https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025006-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/wp/2025/english/wpiea2025006-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/wp/2025/english/wpiea2025006-print-pdf.pdf.json)

---

### 1. Introduction — background, motivation, and research questions
- Foreign aid defined as transfers provided by governments in concessional terms to achieve economic development goals.
- Allocation drivers include donor characteristics, recipient characteristics, and political alignment between donor and recipient.
- Natural disasters (earthquakes, landslides, hurricanes) treated as natural experiments: timing considered arguably exogenous; paper chiefly exploits timing for identification.
- Haiti (2010 earthquake) motivating example:
  - Death toll: an estimated 220,000 individuals.
  - Injured: 300,000.
  - Homeless: 1.5 million.
  - Losses and damages estimated at US$7.3 billion, equivalent to 120 percent of GDP.
  - Initial donor pledges: US$9.9 billion.
- Core research questions:
  - Is the volume of aid allocated following natural disasters significant?
  - How is aid composition impacted (humanitarian versus structural)?
  - Are aid commitments disbursed faster following natural disasters?
  - Do poorest countries or countries facing the most damaging natural disasters receive the most aid?
  - Do recipient characteristics—state capacity and political alignment—matter for the aid response?
- Data and empirical approach highlights:
  - Gravity model of bilateral aid for 1995 to 2021; exploits EM-DAT timing of disasters for identification.
  - ODA from OECD DAC members; exclusively bilateral aid flows used (bilateral aid accounts for about 75 percent of overall aid globally).
  - Focus on the 20 largest donors during the study period, which together accounted for 99 percent of aid commitments.
  - 47-year panel dataset, 189 recipients in total, average of 165 recipients per donor-year.
  - Inclusion of AidData’s Global Chinese Development Finance Dataset: 13,427 development projects valued at $843 billion during 2000 to 2017.
  - EM-DAT (CRED): data from 1900 to present, over 26,000 mass disasters recorded.
  - Political alignment constructed from UN General Assembly votes (session 49 to session 75); alignment equals average fraction of concordant votes between recipient and G7 donors.
  - State capacity index normalized to range between 0 and 1.
  - Additional controls from World Bank WDI: population and GDP for recipient and donor countries.

### 3. Empirical strategy and identification
- Baseline reduced-form specification:
  - ODA_drt = θ NAT_rt + X'_drt β + γ_dr + ε_drt
  - ODA_drt is the logarithm of commitment ODA from donor d to recipient r at time t.
  - NAT_rt is either a dummy (1 if country r has any natural disasters in year t) or the logarithm of the overall damage value in US dollars at time t in recipient r.
  - X'_drt includes time-varying donor or recipient controls (GDP, population), year or donor-year fixed effects; γ_dr denotes donor-recipient pair fixed effects.
- Mechanism exploration via interactions:
  - Example: ODA_drt = θ1 NAT_rt + θ2 UN_drt + θ3 (NAT_rt * UN_drt) + X'_drt β + γ_dr + ε_drt
  - Political alignment proxied by UN voting closeness; absorptive capacity proxied by state capacity.
- Treatment of zero aid observations:
  - Zeros set to one and natural logarithm taken for baseline presentation.
  - Alternative estimators for robustness: Poisson Pseudo-Maximum Likelihood (PPML), Zero-Inflated Poisson, Heckman selection model, Probit model.
- Inference:
  - Standard errors clustered on donor-recipient pairs.

### 4. Main empirical findings (summary)
- Aggregate aid response to disasters:
  - Occurrence of a natural disaster increases bilateral aid commitment by about 69 percent (Column III benchmark).
  - Damage-dollar measure: a 1 percent increase in damages increases bilateral aid by about 0.05 percent (Column VI benchmark).
  - Bilateral aid constitutes 75 percent of ODA; multilateral response excluded, so estimates are a lower bound of the overall aid response.
- Composition: humanitarian versus non-humanitarian aid:
  - Humanitarian aid increases significantly and earlier than structural aid.
  - Column III benchmark: natural disaster increases humanitarian bilateral aid commitment by about 39 percent.
  - Non-humanitarian aid estimates are statistically insignificant; point estimates small or negative.
  - Humanitarian aid data available only from 2005 to 2021; interpret with caution.
- Income-level heterogeneity:
  - Natural disaster increases bilateral aid commitment for the 50th–75th income level by about 150 percent (Column III benchmark).
  - Interaction coefficient for the lowest-income group is about 0.5 (i.e., less than half the size of the 50th–75th effect).
  - Finding: lowest-income countries do not receive proportionally the most aid following natural disasters.
- Damage-intensity heterogeneity:
  - Occurrence of a natural disaster increases bilateral aid commitment for 25h–50th damage per GDP by about 180 percent (Column III benchmark).
  - Interaction coefficient for the highest damage group is about 0.6 (less than half the size of the 25h–50th effect).
  - Finding: countries with the highest damage do not receive proportionally the most aid.
- Disbursement speed:
  - Across models and horizons (horizons 0 to 3 years examined), cannot reject the null that aid disbursement response is ≤ 0; commitments do not translate into quicker disbursement on average.
  - Example: response of aid disbursement between the year of the disaster and the following year has p-value = 0.99 (not significantly positive).
- Mechanisms — political alignment and state capacity:
  - Interactions of NAT with political alignment (UN voting closeness) and state capacity are statistically significant and positive at multiple horizons.
  - Column III benchmark (three-year horizon): both NAT * political alignment and NAT * state capacity are positive and large; coefficient for state capacity is slightly larger.
  - Double interaction (NAT * political alignment * state capacity) is large and negative, indicating limited quantitative significance for the triple interaction.
  - Interpretation: political alignment and state capacity largely operate independently; at low/medium state capacity and high political alignment, political alignment can dominate the aid response.

### 5. Extensions, heterogeneity, dynamics, and robustness checks
- Alternative estimators and zeros:
  - Coefficient on natural disaster damage is statistically significant across OLS, PPML, Zero-Inflated Poisson, Heckman, and Probit.
- Missing damage information:
  - Augmenting specifications with a dummy for missing damage information yields statistically significant natural disaster damage coefficients similar to main results; missingness dummy is itself statistically significant.
- Heterogeneity by disaster type:
  - Column (III) benchmark: larger statistically significant coefficients associated with extreme temperatures, wildfire, landslide (wet), and infestation.
  - Column (VIII) benchmark (damage as proxy): larger statistically significant coefficients associated with extreme temperatures, flood, landslide (dry), and wildfire.
- Temporal dynamics (local projections and LP-DiD):
  - Aggregate aid peaks on impact and declines slowly; response becomes statistically insignificant after five years.
  - Humanitarian aid peaks on impact; non-humanitarian aid peaks after impact and is more muted.
  - LP-DiD results robust under treatment-absorbing assumption.
- Grants versus loans:
  - Response of grants to natural disasters is overwhelmingly higher than loans, including when using damage information.
  - Statistical significance observed across most specifications for both grants and loans, except Column (III) of loans.
- Sample-period and sample-composition robustness:
  - Results hold for sample periods 1990-2021 and 2000-2021.
  - Results robust to including only small island countries and excluding them; effects more pronounced for non-island countries.
  - Removing disasters that immediately follow a previous disaster and restricting to disasters after a three-year gap yield similar results.
- Selected numerical estimates from tables:
  - Natural disaster (dummy) coefficients on log bilateral ODA: 0.952***, 0.939***, 0.527*** (standard errors (0.051), (0.048), (0.043)).
  - Damage coefficients (log damage value): 0.048***, 0.047***, 0.045*** (standard errors (0.003)).
  - Natural disaster (dummy) on log humanitarian aid: 0.373***, 0.369***, 0.331*** (standard errors (0.047), (0.046), (0.047)).
  - Natural disaster (dummy) on log aid excluding humanitarian aid: 0.049, 0.049, -0.004 (standard errors (0.047), (0.045), (0.045)) — not consistently significant.
  - Heterogeneity by income interactions (examples): Natural disaster × 50th-75th income: 1.454***, 1.439***, 0.918*** (standard errors (0.120), (0.114), (0.092)).
  - Heterogeneity by damage-per-GDP quartile (examples): 25th-50th damage per GDP: 1.038***, 1.012***, 1.031*** (standard errors (0.079), (0.074), (0.076)).
- Data and descriptive statistics (selected):
  - ODA (Donor/Recipient/Year Level): N = 151,076; Mean = 25.76; SD = 402.79 (ODA is total aid commitments in millions of US dollars).
  - Natural disaster (Recipient/Year Level): N = 8,854; Mean = 0.45; SD = 0.50 (dummy).
  - Political alignment: N = 4,336; Mean = 0.53; SD = 0.13.
  - State capacity: N = 6,521; Mean = 0.56; SD = 0.16.
  - GDP (Recipient): N = 7,109; Mean = 89,848; SD = 539,508 (millions of US dollars).
  - Population (Recipient): N = 8,351; Mean = 28.67; SD = 122.37 (millions).
  - GDP (Donor): N = 906; Mean = 1,472,285; SD = 2,841,205 (millions of US dollars).
  - Population (Donor): N = 911; Mean = 70.67; SD = 189.53 (millions).

### 6. Conclusion — implications and policy messages
- Summary of core conclusions:
  - Aid commitments statistically and economically significantly increase following natural disasters.
  - Humanitarian aid precedes structural (non-humanitarian) aid; humanitarian aid drives the observed effect.
  - The poorest countries or countries facing the most damaging natural disasters do not receive the most aid proportionally.
  - No evidence that foreign aid commitments disburse faster following natural disasters.
- Mechanisms and policy implications:
  - Low state capacity and political alignment with major donors are important determinants of the aid allocation after disasters.
  - Shift toward reactive, humanitarian-focused aid may crowd out structural aid and could increase recipient countries’ exposure to future disasters.
  - Allocative inefficiencies identified: concentrating aid on countries with the greatest damage or poorest countries could yield greater benefits.
  - Investing in resilience and state capacity in low-income countries could improve aid disbursement and access to capital and insurance markets.
  - Future research direction: explore state capacity as a mechanism for macroeconomic "insurance" in response to natural disasters.

*Source: wpiea2025006-print-pdf — IMF Working Paper (selected sections: 1. Introduction; 3. Empirical Strategy; 5. Extensions and Robustness Checks; 6. Conclusion)*

### 1. Introduction

### 1. Introduction

### Background and motivation
- Foreign aid is generally defined as transfers provided by governments in concessional terms to achieve economic development goals.
- The allocation of aid may be driven by donor characteristics, recipient characteristics, or both; political alignment between donor and recipient has been identified as important by prior literature.
- Exogenous shocks (natural experiments) can help reveal mechanisms driving aid allocation by providing exogenous variation to explore donors’ responses.

### Natural experiments and disasters
- Disasters such as earthquakes, landslides, and hurricanes are treated as natural experiments because the timing of natural disasters is arguably exogenous.
- The extent of damages (intensity of shocks combined with exposure) is endogenous; hence this paper chiefly exploits the timing of natural disasters for identification.
- Natural disasters are salient shocks with significant consequences on physical infrastructure and human capital.

### Haiti (2010 earthquake) as a motivating example
- Death toll: an estimated 220,000 individuals.
- Injured: 300,000.
- Homeless: 1.5 million.
- Losses and damages estimated at US$7.3 billion, equivalent to 120 percent of GDP.
- Initial donor pledges: US$9.9 billion, with most intended to be spent in the first few years; a sizable share of pledges was cancelled.
- Observed weaknesses in the aid response included lack of coordination between organizations and failure to involve local authorities; donor leverage was used to dictate terms of engagement because of alleged corruption and weak state capacity.

### Research questions
- Is the volume of aid allocated following natural disasters significant?
- How is the composition of aid impacted (humanitarian versus structural)?
- Are aid commitments disbursed faster following natural disasters?
- Do poorest countries or countries facing the most damaging natural disasters receive the most aid?
- Do recipient characteristics—such as state capacity and political alignment with donors—matter for the aid response?

### Data and empirical approach
- Empirical approach: gravity model of (bilateral) aid from donors for the period 1995 to 2021; exploits EM-DAT timing of disasters for identification.
- Aid data:
  - ODA from OECD DAC members; exclusively bilateral aid flows are used.
  - Bilateral aid accounts for about 75 percent of overall aid globally (per Dube and Naidu (2022)).
  - Focus on the 20 largest donors during the study period, which together accounted for 99 percent of aid commitments.
  - ODA cover period noted elsewhere from 1975 to 2021.
- Dataset dimensions:
  - 47-year panel dataset, featuring 189 recipients in total, with an average of 165 recipients per donor-year.
- Inclusion of non-traditional donor data:
  - AidData’s Global Chinese Development Finance Dataset documents 13,427 development projects valued at $843 billion, financed by over 300 Chinese government institutions and state-owned entities across 165 countries during 2000 to 2017.
- Natural disaster data:
  - EM-DAT (CRED): data from 1900 to present, over 26,000 mass disasters recorded.
- Political alignment:
  - Constructed using UN General Assembly votes from session 49 (1995/1996) to session 75 (2020/2021) compiled by Voelsen et al. (2021); alignment equals the average fraction of concordant votes between the recipient and G7 donors, averaged across G7 donors.
- State capacity:
  - Index of state capacity (O’Reilly, Colin, and Ryan H. Murphy (Forthcoming)) captures legal and fiscal elements; normalized to range between 0 and 1.
- Additional controls: demographic and economic variables from World Bank WDI (World Bank, 2023), including population and GDP for both recipient and donor countries.

### Key findings (as stated)
- Aid commitment statistically and economically significantly increases following natural disasters.
- Humanitarian aid precedes structural aid in the post-disaster response.
- Poorest countries or countries facing the most damaging natural disasters do not receive the most aid.
- No evidence that foreign aid commitments disburse faster following natural disasters.
- Mechanisms: low state capacity and political alignment with donors are important determinants of the aid response in disaster-affected countries.

### Contribution to literature and implications
- Connects to literature on aid allocation, aid effectiveness, and climate dynamics:
  - Documents that aid response to natural disasters is economically significant and that foreign aid is becoming more humanitarian and reactive.
  - Suggests that increased humanitarian focus could crowd out structural aid, potentially aggravating countries’ exposure to natural disasters and posing implications for the sustainability of the aid industry.
- Relates to prior findings on political drivers of aid (e.g., Faye and Niehaus (2012)), critiques of development aid effectiveness, and evidence on climate change increasing the intensity/frequency of certain disaster types.

### Organization of the paper
- Section 2: data sources.
- Section 3: empirical strategy.
- Section 4: results.
- Section 5: extensions and robustness checks.
- Section 6: conclusion.

*Source: wpiea2025006-print-pdf - 1. Introduction*

### 3. Empirical Strategy

### 3. Empirical Strategy

### Empirical design and identifying variation
- Exploit the timing of natural disasters, treated as arguably exogenous, to test whether foreign aid commitments change; approach akin to difference-in-difference using time variation within donor-recipient pairs.
- Primary reduced-form specification (following Faye and Niehaus (2012)):
  - ODA_drt = θ NAT_rt + X'_drt β + γ_dr + ε_drt
  - ODA_drt is the logarithm of commitment ODA from donor d to recipient r at time t.
  - NAT_rt is either a dummy (1 if country r has any natural disasters in year t) or the logarithm of the overall damage value in US dollars at time t in recipient r.
  - X'_drt includes time-varying donor or recipient controls such as GDP and population, or year or donor-year fixed effects.
  - γ_dr denotes donor-recipient pair fixed effects.
- To explore mechanisms, augment with interactions (example shown):
  - ODA_drt = θ1 NAT_rt + θ2 UN_drt + θ3 (NAT_rt * UN_drt) + X'_drt β + γ_dr + ε_drt
  - Political alignment proxied by UN voting closeness (UN_drt); absorptive capacity proxied by state capacity.
- Include donor-recipient fixed effects to remove time-invariant recipient and bilateral characteristics; include year fixed effects, donor-year fixed effects, and time-varying controls (population, GDP) to account for concurrent trends.

### Treatment of zero bilateral aid observations and estimator choice
- Many donor-recipient-year observations have zero bilateral aid.
- Instead of dropping zeros, zeros are set to one and the natural logarithm is taken for ODA commitments in baseline presentation.
- Recognizing potential bias/heteroskedasticity with OLS and log-linear specification, alternative estimators used for robustness:
  - Poisson Pseudo-Maximum Likelihood (PPML)
  - Zero-Inflated Poisson
  - Heckman selection model
  - Probit model
- References noted: Silva and Tenreyro, 2006; Herrera, 2010; Martin and Hall, 2017.

### Inference and clustering
- Even after fixed effects, conditional uncorrelation of errors is unlikely.
- Cluster standard errors on donor-recipient pairs as the most general and restrictive clustering approach (Bertrand et al., 2004).

### Main empirical findings (summary of Section 4 results)

- 4.1 Does foreign aid increase following natural disasters?
  - Statistically significant and positive relationship between aid commitment and natural disasters across specifications controlling for donor-year fixed effects, year fixed effects, or macro controls.
  - Using Column III as benchmark: occurrence of a natural disaster increases bilateral aid commitment by about 69 percent.
  - Using damage-dollar measure (Column VI benchmark): a 1 percent increase in damages caused by natural disaster increases bilateral aid by about 0.05 percent.
  - Bilateral aid constitutes 75 percent of ODA; multilateral donor response excluded, so results are a lower bound of average overall aid response.

- 4.2 Composition: humanitarian vs. non-humanitarian aid
  - Humanitarian aid:
    - Statistically significant and positive relationship with natural disasters across fixed-effect specifications.
    - Column III benchmark: natural disaster increases humanitarian bilateral aid commitment by about 39 percent.
  - Non-humanitarian aid:
    - Estimates become statistically insignificant; point estimates small or negative.
  - Data limitation: humanitarian aid data available only from 2005 to 2021 (aggregate aid earlier used 1975–2021); interpret with caution.
  - Robustness checks confirm humanitarian aid predominates in the aggregate disaster response.

- 4.3 Income-level heterogeneity: do poorer countries receive more aid?
  - Interaction of natural disasters with income-level categories yields statistically significant and positive interaction terms.
  - Column III benchmark: natural disaster increases bilateral aid commitment for the 50th–75th income level by about 150 percent.
  - For the lowest income group, the interaction coefficient is about 0.5 (i.e., less than half the size of the 50th–75th effect).
  - Finding: lowest-income countries do not receive proportionally the most aid following natural disasters.

- 4.4 Damage-intensity heterogeneity: do more-damaged countries receive more aid?
  - Interaction of natural disasters with damage-intensity categories yields statistically significant and positive interaction terms.
  - Column III benchmark: occurrence of a natural disaster increases bilateral aid commitment for 25h–50th damage per GDP by about 180 percent.
  - The interaction coefficient for the highest damage group is about 0.6 (less than half the size of the 25h–50th effect).
  - Finding: countries experiencing the highest damage do not receive proportionally the most aid following natural disasters.

- 4.5 Disbursement speed: are disbursements faster after disasters?
  - Tests compare predictive power of committed aid on disbursed aid in presence vs. absence of disaster across horizons (horizons 0 to 3 years examined).
  - Across models and horizons, cannot reject the null that aid disbursement response is less than or equal to 0 in favor of the prediction that it is positive.
  - Example: response of aid disbursement between the year of the disaster and the following year has p-value = 0.99 (not significantly positive).
  - Conclusion: aid disbursement is not significantly faster in the first three years following disasters; commitments do not translate into quicker disbursement on average.

- 4.6 Mechanisms: political alignment versus absorptive capacity
  - Regression includes interactions among natural disasters, political alignment (UN voting closeness), state capacity, and double interactions.
  - Interaction terms for natural disasters with political alignment and with state capacity are statistically significant and positive at multiple horizons.
  - Column III benchmark (three-year horizon):
    - Political alignment and state capacity interact with disasters to drive bilateral aid commitment three years after disasters.
    - Both simple interactions (NAT * political alignment; NAT * state capacity) are positive and large; coefficient for state capacity is slightly larger.
    - Double interaction (NAT * political alignment * state capacity) is large and negative, indicating limited quantitative significance for the triple interaction.
  - Interpretation: political alignment and state capacity operate mostly independently; at low/medium state capacity and high political alignment, political alignment can dominate the aid response.
  - Policy implication discussed: countries with low state capacity (low absorptive capacity) receive much less aid following disasters, raising concerns for efficient allocation and resilience-building; donors sometimes circumvent governments (example cited qualitatively) to deliver aid when state capacity is low.

### Methodological considerations highlighted
- Use of donor-recipient fixed effects to eliminate time-invariant confounders.
- Year and donor-year fixed effects and time-varying controls to mitigate trend confounding.
- Robustness checks across estimators (PPML, Zero-Inflated Poisson, Heckman, Probit).
- Clustering on donor-recipient pairs for inference.

*Source: wpiea2025006-print-pdf — 3. Empirical Strategy (IMF Working Paper).*

### 5. Extensions and Robustness Checks

### 5. Extensions and Robustness Checks

### Alternative estimators and zero observations
- Benchmark estimator: OLS (Equation (1)).
- Alternative estimators tested: Poisson Pseudo-Maximum Likelihood, Zero-Inflated Poisson, Heckman selection model, Probit model.
- Finding: The coefficient associated with natural disaster damage (Columns (I-V) in Table 8, Supplementary Appendix) is statistically significant across different estimators.
- Conclusion: Main results are robust to using different estimators that account for the presence of many zero observations.

### Missing damage information
- Specification: Equation (1) augmented with a dummy for missing damage information (Table 9, Supplementary Appendix).
- Findings:
  - Coefficients in Columns (I-III in Table 9) show the natural disaster damage coefficient is statistically significant across specifications.
  - Quantitatively similar to main results in Table 2.
  - The dummy for missing information is also statistically significant, indicating missingness carries meaningful variation.
- Conclusion: Robustness checks support reliance on the missing-data dummy to identify the effect of natural disaster shocks on bilateral aid.

### Heterogeneity by disaster type
- Approach: Use dummy variables for the timing of natural disasters (Columns (I)-(III)) and damage as a proxy (Columns (IV)-(VI), and Column (VIII) as benchmark) in Equation (1) (Table 10, Supplementary Appendix).
- Findings:
  - Coefficients for "all disasters" are almost always statistically significant but vary in size across specifications.
  - Column (III) benchmark: larger statistically significant coefficients associated with extreme temperatures, wildfire, landslide (wet), and infestation.
  - Column (VIII) benchmark (damage as proxy): larger statistically significant coefficients associated with extreme temperatures, flood, landslide (dry), and wildfire.
- Conclusion: There is heterogeneity in bilateral aid responses across disaster types.

### Dynamics: local projections and LP-DiD
- Methods: Impulse response functions using local-projection method (Jordà, 2005); alternative LP-DiD method (Dube et al, 2023) to address staggered treatment concerns.
- Findings:
  - Figure 5 (Supplementary Appendix): aggregate aid shows a peak on impact and a slow decline; response becomes statistically insignificant after five years.
  - Figures 6 and 7: humanitarian aid peaks on impact; non-humanitarian aid peaks after the impact and its response is more muted (smaller peak).
  - Overall: humanitarian aid drives the aid response following natural disasters (consistent with Table 3 main paper).
  - LP-DiD results are robust, especially under the assumption that once a unit is treated, the unit remains treated (see Figure 8).
- Conclusion: Temporal dynamics show immediate humanitarian responses dominate and effects dissipate within five years.

### Grants versus loans
- Specifications: Equation (1) estimated separately for grants (Table 11, Supplementary Appendix) and loans (Table 12, Supplementary Appendix).
- Findings:
  - Statistical significance across most specifications for both grants and loans, except Column (III) of loans.
  - Quantitatively, response of grants to natural disasters is overwhelmingly higher than loans, including when using damage information.
  - Overall magnitude of the aid response is relatively small compared to the damage caused by natural disasters.
- Conclusion: Humanitarian-style grants drive the aid response more than loans.

### Sample period and sample composition robustness
- Alternative sample periods tested: 1990-2021 and 2000-2021 (instead of 1975-2021) — results hold (Tables 13 and 14).
- Sample composition tests:
  - Samples including only small island countries and samples excluding them — results robust; effects more pronounced for non-island countries (Tables 15 and 16).
  - Removing natural disasters that immediately follow a previous disaster — results remain robust (Tables 17 and 18).
  - Using disasters that occurred when no disasters happened in the past three years — results remain largely robust (Tables 19 and 20).
- Conclusion: Main findings are robust across alternative periods, island/non-island samples, and restrictions excluding clustered disasters.

### 6. Conclusion (summary of implications)
- Main empirical findings:
  - Aid commitments statistically and economically significantly increase following natural disasters.
  - Humanitarian aid precedes structural (non-humanitarian) aid; humanitarian aid drives the observed effect.
  - The poorest countries or countries facing the most damaging natural disasters do not receive the most aid.
  - No evidence that foreign aid commitments disburse faster following natural disasters.
- Mechanisms highlighted:
  - Importance of low state capacity and political alignment with (major) donors in recipient countries as drivers of aid allocation.
- Policy and research implications:
  - Increasing frequency and intensity of natural disasters shifts aid composition toward a more "reactive" form, reducing structural, developmental aid and increasing vulnerability.
  - Aid allocation appears inefficient: concentrating aid on countries with the greatest damage or poorest countries could yield greater benefits.
  - Investing in resilience and state capacity in low-income countries could improve aid disbursement and access to capital and insurance markets.
  - Future research direction: explore state capacity as a mechanism for "insurance" at the macroeconomic level in response to natural disasters.

*Source: wpiea2025006-print-pdf - 5. Extensions and Robustness Checks*

### 2012. doi: 10.1257/aer.102.7.3516. URL https://www.aeaweb.org/articles? id=10.1257/aer.102.7.3516.

### wpiea2025006-print-pdf - 2012. doi: 10.1257/aer.102.7.3516. URL https://www.aeaweb.org/articles?id=10.1257/aer.102.7.3516.

### Key empirical findings on natural disasters and foreign aid
- Natural disasters are associated with positive and statistically significant increases in bilateral Official Development Assistance (ODA) commitments:
  - Natural disaster (dummy) coefficients on log bilateral ODA: 0.952***, 0.939***, 0.527*** (Table 2, columns (I)-(III)) — standard errors (0.051), (0.048), (0.043).
- Measured damage (log of overall damage value in US dollars) is positively and significantly associated with larger bilateral ODA commitments:
  - Damage coefficients: 0.048***, 0.047***, 0.045*** (Table 2, columns where Damage is included) — standard errors (0.003) in each column.
- Recipient and donor macro controls show consistent associations in the bilateral ODA regressions (Table 2):
  - Recipient GDP: -1.310*** and -1.329*** (standard errors (0.083) and (0.083))
  - Donor GDP: 2.865*** and 2.884*** (standard errors (0.117) and (0.116))
  - Recipient population: 1.319*** and 1.414*** (standard errors (0.300) and (0.300))
  - Donor population: -1.245*** and -1.244*** (standard errors (0.156) and (0.155))
- Natural disasters primarily increase humanitarian aid flows rather than non-humanitarian aid (Table 3):
  - Natural disaster (dummy) on log humanitarian aid: 0.373***, 0.369***, 0.331*** (standard errors (0.047), (0.046), (0.047))
  - Natural disaster (dummy) on log aid excluding humanitarian aid: 0.049, 0.049, -0.004 (standard errors (0.047), (0.045), (0.045)) — not consistently significant.
- Heterogeneity by recipient income group (interaction effects, Table 4):
  - Natural disaster × Bottom 25th income: 0.886***, 0.903***, 0.519*** (standard errors (0.112), (0.098), (0.092))
  - Natural disaster × 25th-50th income: 0.943***, 0.947***, 0.365*** (standard errors (0.105), (0.096), (0.079))
  - Natural disaster × 50th-75th income: 1.454***, 1.439***, 0.918*** (standard errors (0.120), (0.114), (0.092))
  - Natural disaster × Top 25th income: 0.052, -0.016, 0.203 (standard errors (0.127), (0.129), (0.131)) — not significant.
- Heterogeneity by damage-per-GDP quartile (Table 5):
  - Bottom 25th damage per GDP: 0.677***, 0.650***, 0.719*** (standard errors (0.073), (0.068), (0.070))
  - 25th-50th damage per GDP: 1.038***, 1.012***, 1.031*** (standard errors (0.079), (0.074), (0.076))
  - 50th-75th damage per GDP: 0.811***, 0.797***, 0.684*** (standard errors (0.078), (0.073), (0.076))
  - Top 25th damage per GDP: 0.975***, 0.965***, 0.641*** (standard errors (0.065), (0.060), (0.062))

### Dynamic responses and timing of aid
- Impulse-response / event-study evidence (figures summarized in text/captions):
  - Dynamic response of bilateral aid commitments to natural disaster shocks shows point estimates (blue line) with 90% confidence bands (gray areas); vertical axis reports percentage changes (Figures 5–7).
  - Separate dynamic responses are shown for total bilateral aid (Figure 5), bilateral humanitarian aid (Figure 6), and bilateral non-humanitarian aid (Figure 7).
  - An event study using local projection difference-in-difference (LP-DiD) under the treatment-absorbing assumption is presented for bilateral aid commitments (Figure 8).
- Tests of speed of disbursement following natural disasters (Table 6) report hypothesis-test P-values equal to 0.996, 0.999, 0.999, and 1.0 across alternative models and horizons for Net disbursement and 1.0 or 1.0-equivalent values for Gross disbursement:
  - Net disbursement P-values reported: .996, .999, .999, .999 (multiple model/horizon specifications).
  - Gross disbursement P-values reported: 1.0, 1.0, 1.0, 1.0 (multiple model/horizon specifications).
- Models used in disbursement tests:
  - Model 1 (with DR-Y Fixed Effects): Disbursement_dr,t = sum_h β_h * Commitment_dr,t−h * NoDisaster_r,h + sum_h θ_h * Commitment_dr,t−h * Disaster_r,h + γ_dr + γ_t + ε_dr,t
  - Model 2 (with DR-DY Fixed Effects): Disbursement_dr,t = sum_h β_h * Commitment_dr,t−h * NoDisaster_r,h + sum_h θ_h * Commitment_dr,t−h * Disaster_r,h + γ_dr + γ_dt + ε_dr,t
  - Model 3 (with DR Fixed Effect and GDP and Population): Disbursement_dr,t = sum_h β_h * Commitment_dr,t−h * NoDisaster_r,h + sum_h θ_h * Commitment_dr,t−h * Disaster_r,h + γ_dr + δ * GDP/Population + ε_dr,t

### Data and descriptive statistics (Table 1)
- Sample and variable means/standard deviations:
  - ODA (Donor/Recipient/Year Level): N = 151,076; Mean = 25.76; SD = 402.79 (ODA is total aid commitments in millions of US dollars).
  - Natural disaster (Recipient/Year Level): N = 8,854; Mean = 0.45; SD = 0.50 (dummy).
  - Political alignment: N = 4,336; Mean = 0.53; SD = 0.13 (average probability of agreement between the G7 donor and recipient at the United Nations).
  - State capacity: N = 6,521; Mean = 0.56; SD = 0.16 (index from 0 to 1).
  - GDP (Recipient): N = 7,109; Mean = 89,848; SD = 539,508 (expressed in millions of US dollars).
  - Population (Recipient): N = 8,351; Mean = 28.67; SD = 122.37 (millions of inhabitants).
  - GDP (Donor): N = 906; Mean = 1,472,285; SD = 2,841,205 (millions of US dollars).
  - Population (Donor): N = 911; Mean = 70.67; SD = 189.53 (millions of inhabitants).

### Model specifications and fixed effects
- Regressions use logarithms for dependent variables and controls; standard errors clustered at donor-recipient pair level.
- Fixed effects employed across specifications include:
  - DR: donor-recipient pair fixed effects
  - Y: year fixed effects
  - DY: donor-year fixed effects
  - Combinations reported as DR-Y, DR-DY, and DR in various columns (Tables 2–5).
- Sample sizes vary across specifications when macroeconomic indicators are not available for all countries/years; columns (III) and (VI) in several tables have smaller observations (e.g., 126,810 vs 151,076).

*Source: wpiea2025006-print-pdf - 2012. doi: 10.1257/aer.102.7.3516. URL https://www.aeaweb.org/articles?id=10.1257/aer.102.7.3516. (Canonical source PDF: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025006-print-pdf.pdf)*

---


_Source: https://www.imf.org/-/media/files/publications/wp/2025/english/wpiea2025006-print-pdf.pdf_