## 1. Identification of Natural Disaster Episodes

## Source details

**Canonical URL:** [1. Identification of Natural Disaster Episodes](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2016/_wp1665.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2016/_wp1665.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2016/_wp1665.pdf.json)

---

### Identification of Large Natural Disasters
- A large natural disaster shock is identified when its impact is above the country-specific 50th percentile (right-tail) by either:
  - cost of damages in percent of GDP; and/or
  - people affected in percent of the population.
- Each natural disaster shock is counted only in the first year of occurrence, yielding 166 shocks (out of 989 observations).
- The upper-percentile shocks show a much larger ratio of people affected than non-shock episodes, indicating the cut-off captures severe natural disasters.

### Dependent Variable: Food Crisis Definition
- A food crisis is defined by two simultaneous conditions:
  - (i) the two year average level of daily food supply per capita after the shock (t and t+1) falls below the pre-shock three-year average; and
  - (ii) year-on-year growth of food supply per capita is negative at time t.
- Observations not meeting both criteria are labeled normal episodes.
- This definition targets severe events that both produce negative food supply growth and lower the level of food supply for the following two years relative to the prior three.

### Sample, Episodes, and Key Statistics
- Data cover 55 LICs during 1990–2009, producing 1,180 observations; effective sample varies between 45–47 countries across estimations.
- Identified episodes:
  - 166 natural disaster shocks (first-year occurrences).
  - 218 food crisis episodes in the full sample.
  - 42 food crisis episodes coincide with large natural disaster episodes.
- Probability and distribution notes:
  - The probability of a food supply decline within the shock sample is 25 percent.
  - The overall sample has a positive median growth number; not all natural disasters cause food supply declines.
- Table 1 reported figures (as presented):
  - Median0.63-2.521.220.25
  - Observations16642124

### Persistence of Food Supply Declines
- Negative food supply growth episodes tend to be persistent with no quick recovery, a key vulnerability in LICs.
- Food decline episodes: 28; Normal episodes: 101 (Table 2).

---

### Econometric Framework and Main Results

### Econometric Framework: Probit Model
- Model: panel probit conditioning food crisis episodes on natural disasters using Φ (cumulative normal density) with country-specific effects.
- Estimators evaluated: fixed effects (FE), random effects (RE), Chamberlain-Mundlak (CM).
  - FE estimator was rejected.
  - RE and CM not rejected.
  - CM estimator adopted for performance and accounting for correlation between unobserved country effects and regressors.
- Initial variable set: 22 potential variables narrowed to seven significant indicators via a general-to-specific approach.
- Explanatory variables are lagged by one year except for exogenous-shock variables.

### Key Explanatory Variables (identified)
- Policy and external variables:
  - Ratio of government balance to GDP.
  - Reserve coverage (months of imports of goods and services).
  - Exchange market pressure index (composite of official exchange rate depreciation, change in reserves in months of imports, and parallel market premium).
  - Change in export prices and growth in trading partners (weighted by lagged exports to GDP).
- Structural and institutional variables:
  - World Bank’s CPIA.
  - Country-specific average of food supply per capita growth (kilo calories).

### Benchmark Probit Estimates (selected coefficients, standard errors in parentheses)
- Sample variants: All LICs (Column 1), With People Affected (Column 2), Excluding Fragile States (Column 3), All LICs Logit (Column 4).
- Selected coefficient estimates:
  - CPIA (t-1): -0.584* (0.265); Column 2: -0.683* (0.304); Column 3: -0.235 (-0.451); Column 4: -0.964* (0.489)
  - Government balance, (% of GDP) (t-1): -0.0431* (0.0199); Column 2: -0.0405* (0.0199); Column 3: 0.016 (-0.028); Column 4: -0.0714* (0.0357)
  - Food supply growth average (cal): -1.090*** (0.234); Column 2: 0.131*** (0.0336); Column 3: -1.36*** (0.348); Column 4: 0.180*** (0.0529)
  - Reserve coverage (months) (t-1): -0.153* (0.0623); Column 2: -0.781** (0.263); Column 3: -0.291*** (0.102); Column 4: -0.970** (0.337)
  - Export prices percent change: 0.105*** (0.0315); Column 2: -1.213*** (0.286); Column 3: 0.165*** (0.053); Column 4: -1.819*** (0.409)
  - Growth in trading partners (weighted): -0.552** (0.185); Column 2: -0.222** (0.0745); Column 3: -0.525** (0.235); Column 4: -0.272* (0.112)
  - Exchange market pressure index: -0.135* (0.0584); Column 2: -0.200* (0.0912); Column 3: -0.113 (-0.093); Column 4: -0.233* (0.116)
  - People affected (% of population): 0.00685 (0.0220) (Column 2)
- Model diagnostics (Table 2):
  - Observations: Column 1: 129; Column 2: 108; Column 3: 82; Column 4: 129
  - Pseudo R-sq: 0.266; 0.324; 0.371; 0.261
  - Number of countries: 45; 45; 29; 45
  - Food decline episodes: 28; 24; 15; 28
  - Normal episodes: 101; 84; 67; 101
  - Wald test (Chi-square): 39.263; 33.993; 39.153; 6.72
  - Sample probability: 0.22; 0.22; 0.18; 0.22

### Interpretation of Benchmark Results
- Weaker institutions (lower CPIA) and low food supply per capita growth trend increase probability of a food crisis.
- Strong policy fundamentals (higher fiscal balance and reserve coverage) reduce likelihood.
- Increase in exchange market pressure index tends to reduce probability, possibly reflecting depreciation and reserve drawdown encouraging domestic production/exports.
- Growth in trading partners reduces probability; high export prices increase probability (potential diversion of food to external markets).
- Including people affected increases predicted probability but is not statistically significant.
- Excluding fragile states makes CPIA, fiscal balance, and EMPI insignificant, indicating their effects are stronger in fragile states.

### Correlated Pooled Probit (Appendix Table 2 excerpts)
- Selected estimates (standard errors in parentheses) and marginal effects:
  - CPIA (t-1): (1) -0.257** (0.0888); average marginal effects -0.063** (0.0216)
  - Government balance, (% of GDP) (t-1): (1) -0.0228* (0.0108); average marginal effects -0.006* (0.0027)
  - Food supply growth average (cal): (1) -0.783*** (0.0708); average marginal effects -0.194*** (0.0131)
  - Reserve coverage (months) (t-1): (1) -0.0143 (0.0255); average marginal effects -0.004 (0.0063)
  - Export prices percent change: (1) 0.0195* (0.00901); average marginal effects 0.005* (0.0022)
  - Growth in trading partners (weighted): (1) -0.230*** (0.0520); average marginal effects -0.057*** (0.0127)
  - Exchange market pressure index: (1) -0.0178 (0.0218); average marginal effects -0.004 (0.0054)
  - People affected (% of population): (4) 0.0722 (0.130)
  - Constant examples: (1) 0.617* (0.267); average marginal effects 0.597* (0.269)
- Observations reported: 743 743 129 (as presented)
- Pseudo R-sq: 0.122 0.122 0.261
- Number of countries: 47 45 47
- Food decline episodes: 151 151 28
- Normal episodes: 592 592 101
- Wald test (Chi-square): 171.481 171.483 6.72
- Sample probability: 0.20 0.20 0.22
- Note: Significance denoted at 10 percent *, 5 percent **, 1 percent ***; standard errors in parentheses.

---

### Marginal Effects, Thresholds, and Model Performance

### Average Marginal Effects (Table 3 highlights)
- Average marginal effects (column 1), median LIC (column 2), LICs with weak fundamentals (column 3):
  - CPIA (t-1): -0.1244; -0.1398; -0.1713
  - Government balance, (% of GDP) (t-1): -0.0092; -0.0103; -0.0126
  - Food supply growth average (cal): -0.2321; -0.2609; -0.3195
  - Reserve coverage (months) (t-1): -0.0327; -0.0367; -0.0449
  - Export prices percent change: 0.0224; 0.0252; 0.0309
  - Growth in trading partners (weighted): -0.1177; -0.1323; -0.1620
  - Exchange market pressure index: -0.0287; -0.0323; -0.0395
- Structural variables (average food supply per capita and CPIA) are the best predictors, especially for countries with weaker fundamentals.

### Probability Threshold and In-/Out-of-Sample Performance (Table 4)
- Threshold determined by minimizing loss function (equal weights to Type I and Type II errors):
  - Overall probability threshold: 0.19
  - Type I error: 14 percent
  - Type II error: 27 percent
- Distribution of predicted probabilities (Food Crisis vs Normal Episodes, percentiles):
  - 1%: Food Crisis 0.07 / Normal 0.00
  - 5%: 0.08 / 0.00
  - 10%: 0.12 / 0.00
  - 25%: 0.21 / 0.04
  - 50% (median): 0.38 / 0.10
  - 75%: 0.60 / 0.20
  - 90%: 0.83 / 0.42
  - 95%: 0.86 / 0.54
  - 99%: 0.86 / 0.64
- In-sample:
  - Observations: Food Crisis 28; Normal Episodes 101; Sample probability 0.22
  - Type I: 0.14; Type II: 0.27; Overall threshold: 0.19
- Out-of-sample (estimate 1990–2004, predict 2005–2009):
  - Correctly identified 7 out of 10 crises.
  - False alarms in 31 percent of non-crisis episodes (11 out of 35 non-crisis episodes classified as crisis).
  - Out-of-sample performance slightly worse than in-sample, possibly due to low crisis counts.

### Robustness Check
- Regressed every food decline episode (151) on identified explanatory variables (no restriction to natural disaster shocks).
- Results broadly unchanged; most variables remain significant except reserves and exchange market pressure index.
- Conclusion: findings generally hold independent of natural disaster definition.

---

### Signaling Approach and Food Decline Vulnerability Index (FDVI)

### Signaling Methodology
- Each indicator receives an optimal threshold minimizing total misclassification error TME(T) = Type I + Type II.
- A breach of an indicator threshold signals an upcoming food crisis.
- Overall FDVI constructed by grouping variables into three clusters and weighting indicators by predictive power:
  - Clusters: (i) overall economy and institutions; (ii) external sector; (iii) fiscal sector.
  - Overall Index = weighted sum of cluster indices; indicators converted to dummies based on thresholds.

### FDVI Composition and Cluster Weights (Table 5 highlights)
- Analysis uses 12 variables (including share of agriculture in GDP; life expectancy; population growth; government tax revenue; government debt).
- Sample: 55 LICs, 1990–2008 (food crisis identification requires post two-year average).
- Food crisis episodes matched with natural disasters: 42.
- Cluster weights:
  - Overall economy and institutions: 0.42 (top predictor).
    - CPIA (t-1) threshold > 38; Type I: 0.84; Type II: 0.66; Index weight: 0.05
    - Life expectancy > 38: Type I: 0.45; Type II: 0.26; Index weight: 0.13
    - Food supply growth average (cal) > 42: Type I: 0.80; Type II: 0.28; Index weight: 0.16
    - Agriculture value added (% of GDP) (t-1) < 37: Type I: 0.25; Type II: 0.60; Index weight: 0.07
    - Population growth > 33: Type I: 0.12; Type II: 0.72; Index weight: 0.03
  - External sector: 0.34
    - Reserve coverage (months) (t-1) > 35: Type I: 0.12; Type II: 0.95; Index weight: 0.11
    - Growth in trading partners (weighted) > 36: Type I: 0.12; Type II: 0.40; Index weight: 0.09
    - Exchange market pressure index > 31: Type I: 0.15; Type II: 0.61; Index weight: 0.06
    - Export prices percent change < 35: Type I: 0.12; Type II: 0.49; Index weight: 0.08
  - Fiscal sector: 0.24
    - Government balance (% of GDP) (t-1) > 42: Type I: 0.13; Type II: 0.34; Index weight: 0.09
    - Government tax revenue (t-1) > 28: Type I: 0.93; Type II: 0.07; Index weight: 0.03
    - Government debt (t-1) < 20: Type I: 0.81; Type II: 0.49; Index weight: 0.09
    - Real government revenue (% change over 2 yrs) (t-1) > 25: Type I: 0.98; Type II: 0.11; Index weight: 0.02
- Overall index threshold and fit:
  - Overall Index threshold: 0.35
  - Proportion of Crisis Missed: 0.12
  - Proportion of Non-crisis mis-specified (false alarms): 0.45
  - Overall error: 0.37
- Note: Sample includes 63 PRGT-eligible countries for 1990–2009, including climatic shocks.

### Predictive Performance and Goodness of Fit
- Univariate probit regressions show overall FDVI predicted probabilities are better distributed than individual subcomponents.
- Economy and institutions cluster shows more dispersed predicted probabilities, justifying higher weight.
- External sector cluster assigned slightly higher weight than fiscal cluster.
- Median predicted probability (probit): 0.38 for crisis vs 0.10 in normal times (Table 4).
- In-sample FDVI performance:
  - Most food crisis events (75 percent) have a food decline index > 0.49.
  - Only 10 percent of food crisis episodes report an index < 0.10.
  - The index correctly calls 88 percent of the food crisis events.
  - Overall model misclassification error: 37 percent.
- Threshold trade-offs:
  - Overall index threshold 0.35 (asymmetric loss with higher weight on Type I errors) → Type I error 12 percent.
  - Medium vs low threshold 0.21:
    - < 10 percent of food crises missed.
    - False alarms increase to about 73 percent.
  - A threshold corresponding to Type I error of 8 percent is used to differentiate medium vs low risk in other contexts.

### FDVI and Subcomponent Percentiles (selected)
- Vulnerability Index / Economy and Institutions Index / External Index / Fiscal Index — percentiles for Food Crisis vs Normal Episodes (excerpted):
  - 25%: Food Crisis 0.23 / Normal Episodes 0.08; Economy 0.21 / 0.11; External 0.22 / 0.16; Fiscal 0.18 / 0.15
  - 50%: Food Crisis 0.34 / Normal Episodes 0.16; Economy 0.33 / 0.21; External 0.28 / 0.18; Fiscal 0.24 / 0.18
  - 75%: Food Crisis 0.49 / Normal Episodes 0.28; Economy 0.44 / 0.27; External 0.33 / 0.26; Fiscal 0.29 / 0.28
- Food Decline Vulnerability Index percentiles (Table 7 excerpt):
  - 1%: Food Crisis 0.00 / Normal Episodes 0.00
  - 5%: Food Crisis 0.10 / Normal Episodes 0.10
  - 10%: Food Crisis 0.25 / Normal Episodes 0.15
  - 25%: Food Crisis 0.41 / Normal Episodes 0.21
  - 50%: Food Crisis 0.51 / Normal Episodes 0.33
  - 75%: Food Crisis 0.63 / Normal Episodes 0.46
  - 90%: Food Crisis 0.69 / Normal Episodes 0.55
  - 95%: Food Crisis 0.75 / Normal Episodes 0.59
  - 99%: Food Crisis 0.90 / Normal Episodes 0.65

### Classification Errors (reported)
- Type I: 0.12 (In sample), 0.50 (Out of sample)
- Type II: 0.45 (In sample), 0.31 (Out of sample)
- Sample probability: 1/0.23 (In sample), 0.13 (Out of sample)
- Note: "1/ Number of growth crisis divided by total observations."

---

### FDVI Evolution, Country Groups, and Projections

### FDVI Time Profile and Drivers
- FDVI evolution 1991–2013:
  - Sharp increase in 1995, decline in 1997, steady increase until 2003 peak.
  - Mid-1990s spike driven by high changes in export prices, exchange rate pressure, and fiscal deficits.
  - CPIA and population growth, together with the above, drove the 1997–2003 steady increase.
  - Increased frequency of natural disasters during the rise to 2003.
  - After 2003, lower disaster frequency and improvements (lower agricultural vulnerabilities, lower fiscal deficits, higher reserves, reduced exchange rate pressure) reduced food supply vulnerabilities.
  - FDVI reached lowest point in 2013; 2009 outlier due to sudden drop in export prices and exchange rate pressures.
- Visualization notes: median, 25th percentile, 75th percentile lines for 1991–2013; risk categories High/Medium/Low shown.

### Country Group Disaggregation and Projections
- Fragile States:
  - More prone to food supply declines.
  - In 2015: share of Fragile States with high risk not increasing; share with medium risk edging up.
  - 2015 medium-risk increase due partly to higher vulnerabilities in previously low-risk countries and partly to decreased vulnerabilities in high-risk countries.
  - For 2016 the FDVI projects a substantial increase in food insecurities with high and medium risk rising across the board.
- Figures referenced (not reproduced here) show natural disaster counts and percent of LIDC population affected by year, and group comparisons for LIDCs and Fragile States in 2014–2016.

---

### Conclusions and Policy Recommendations

### Main Findings
- Exposure of LICs to natural disasters significantly impacts food production and food security.
- Macroeconomic and structural indicators important for resilience to external shocks are also crucial to minimize food crises associated with natural disasters.
- Probability of experiencing a food crisis increases with weaker institutions (lower CPIA) and lower food supply per capita growth.
- Fragile States are more prone to food crisis episodes.

### Policy Recommendations and Mitigation Options
- Strengthen pre-existing macroeconomic conditions to improve capacity to avert food crises following large natural disasters.
- Secure supplementary buffers through insurance schemes, including:
  - Self-insurance via contingency reserves.
  - Rapid access to disaster support lines from donors.
- Free up resources for disaster preparedness.
- LICs particularly vulnerable to natural disasters and food production shocks should:
  - Lower trade barriers.
  - Improve road networks to create larger regional markets.
  - Allow faster and more efficient responses to localized food shortages.

---

### Methods, Variables, and Appendix Materials

### Methods
- Two complementary approaches:
  - Multivariate regression analysis (correlated pooled probit model with cluster-robust standard errors).
  - Univariate signaling approach.

### Variables (Appendix Table 3 sectors and sources)
- Economy and Institutions:
  - CPIA (t-1) — World Bank
  - Food supply growth average (cal) — Food and Agriculture Organization
- Fiscal sector:
  - Government balance, (% of GDP) (t-1) — World Economic Outlook
  - Government taxes, (% of GDP) (t-1)
  - Public debt, (% of GDP) (t-1)
  - Government tax revenue (t-1); Real government revenue (% change over 2 yrs) (t-1)
- External sector:
  - Reserve coverage (months of imports) (t-1) — World Economic Outlook, IFS, Staff Reports
  - Export prices percent change — World Economic Outlook
  - Growth in trading partners (weighted) — World Economic Outlook
  - Exchange market pressure index — World Economic Outlook
- Shock variables:
  - Number of affected from natural disasters — CRED EM-DAT
  - Cost of natural disasters (in percent of GDP) — EM-DAT
- Additional controls:
  - Agriculture value added (% of GDP) (t-1)
  - Life expectancy (t-1)
  - Population growth (t-1)

### Appendix materials (selected)
- Appendix Table 1: List of countries in sample — 74 LICs listed; 55 countries used in regressions (listed verbatim in source).
- Appendix Figure 1: FDVI by sub-indicator — share of countries breaching thresholds for CPIA, ExportPricesPercentChange, Fiscal Balance (In percent of GDP), Exchange Market Pressure Index, Reserve coverage (months), External Demand, Life Expectancy, Agriculture ValueAdded, Population Growth (1991–2013).

*Source: _wp1665 - 1. Identification of Natural Disaster Episodes; FAO; World Bank; EM-DAT; IMF WEO and authors' calculations (as presented in the supplied content).*

### 1. Identification of Natural Disaster Episodes _____________________________________7

### 1. Identification of Natural Disaster Episodes

### Identification of Large Natural Disasters
- A large natural disaster shock is identified if its impact on the economy is above the norm: specifically, if the cost of damages in percent of GDP and/or the amount of people affected in percent of the population is above the 50th percentile (i.e., in the right-tail) of the country-specific distribution.
- Each natural disaster shock is considered only in the first year of its occurrence, resulting in a total of 166 shocks (out of 989 observations).
- Comparing the average of people affected and the cost of damages for the upper percentile of the distribution shows the ratio of people affected by natural disasters in shock episodes (natural disasters above the 50th percentile) is much larger than in non-shock episodes, suggesting the shock cut-off captures severe natural disaster shocks.

### Identification of the Dependent Variable: Food Crisis
- A food crisis is defined as a large drop in daily food supply (crops and livestock) measured in kilocalories per capita. A food crisis occurs when both conditions hold:
  - (i) the two year average level of daily food supply per capita after the shock (t and t+1) falls below the pre-shock three-year average; and
  - (ii) year-on-year growth of food supply per capita is negative at time t.
- All other observations that do not meet these criteria are considered as normal episodes.
- This definition follows the example of a growth crisis and is designed to identify severe food crisis events that both lead to negative food supply growth and alter the level of food supply such that the following two years after the event are substantially below the previous three years.

### Sample, Episodes, and Key Statistics
- Data span 55 LICs during 1990–2009, yielding 1,180 observations; the sample shrinks to 45–47 countries in various estimations depending on data availability.
- Identified episodes:
  - 166 natural disaster shocks (first-year occurrences).
  - 218 food crisis episodes in the full sample.
  - 42 food crisis episodes coincide with large natural disaster episodes as defined above.
- Probability and distributional notes:
  - The probability of a food supply decline within the shock sample is 25 percent.
  - The positive median growth number for the overall sample means not all natural disasters cause a decline in food supply.
- Table 1 (Median of Food Supply Per Capita Growth, 1990–2009) — key reported figures as presented:
  - Median0.63-2.521.220.25
  - Observations16642124

*Source: _wp1665 - 1. Identification of Natural Disaster Episodes _____________________________________7*

### 3.74 percentage points between

### _wp1665 - 3.74 percentage points between

### Persistence of Food Supply Declines
- The drop in food supply growth (negative value) is often associated with a persistent decline in the level of food supply; once food supply growth is negative, food supply declines tend to be persistent with no quick recovery, a key vulnerability in LICs.
- Figure 2 presents the distribution of food supply per capita growth for crisis and non-crisis episodes (1990–2009).  
- Food decline episodes: 28; Normal episodes: 101 (as reported in Table 2).

### Econometric Framework: Probit Model
- Model: panel probit model conditioning food crisis episodes on natural disasters; specification uses Φ as the cumulative normal density function and includes country-specific effects.
- Estimators evaluated: fixed effect (FE), random effect (RE), and Chamberlain (1982)-Mundlak (1978) (CM) approach.
  - Fixed effects estimator was rejected.
  - Random effect estimator and CM approach were not rejected.
  - CM estimator adopted as it performed better and accounts for correlation between unobserved country specific variable and explanatory variables.
- Initial variable set: 22 potential variables; narrowed to seven significant indicators through a general-to-specific approach.
- Explanatory variables are lagged by one year, except variables capturing exogenous shocks.

### Key Explanatory Variables (identified)
- Policy and external variables:
  - Ratio of government balance to GDP.
  - Reserve coverage (in months of imports of goods and services).
  - Exchange market pressure index (composite of official exchange rate depreciation, change in stock of international reserves in months of imports, and parallel market premium).
  - Change in export prices and growth in trading partners (weighted by lagged exports to GDP).
- Structural and institutional variables:
  - World Bank’s CPIA.
  - Country-specific average of food supply per capita growth over the sample period (measured in kilo calories).

### Estimation Results: Benchmark Probit Specifications (Table 2 highlights)
- Sample variants: All LICs (Column 1), With People Affected (Column 2), Excluding Fragile States (Column 3), All LICs Logit (Column 4).
- Selected coefficient estimates (standard errors in parentheses):
  - CPIA (t-1): -0.584* (0.265); Column 2: -0.683* (0.304); Column 3: -0.235 (-0.451); Column 4: -0.964* (0.489)
  - Government balance, (% of GDP) (t-1): -0.0431* (0.0199); Column 2: -0.0405* (0.0199); Column 3: 0.016 (-0.028); Column 4: -0.0714* (0.0357)
  - Food supply growth average (cal): -1.090*** (0.234); Column 2: 0.131*** (0.0336); Column 3: -1.36*** (0.348); Column 4: 0.180*** (0.0529)
  - Reserve coverage (months of imports) (t-1): -0.153* (0.0623); Column 2: -0.781** (0.263); Column 3: -0.291*** (0.102); Column 4: -0.970** (0.337)
  - Export prices percent change: 0.105*** (0.0315); Column 2: -1.213*** (0.286); Column 3: 0.165*** (0.053); Column 4: -1.819*** (0.409)
  - Growth in trading partners weighted by lagged exports to GDP: -0.552** (0.185); Column 2: -0.222** (0.0745); Column 3: -0.525** (0.235); Column 4: -0.272* (0.112)
  - Exchange market pressure index: -0.135* (0.0584); Column 2: -0.200* (0.0912); Column 3: -0.113 (-0.093); Column 4: -0.233* (0.116)
  - People affected from natural disasters (in percent of population): 0.00685 (0.0220) (included in Column 2)
- Model diagnostics (Table 2):
  - Observations: Column 1: 129; Column 2: 108; Column 3: 82; Column 4: 129
  - Pseudo R-sq: 0.266; 0.324; 0.371; 0.261
  - Number of countries: 45; 45; 29; 45
  - Food decline episodes: 28; 24; 15; 28
  - Normal episodes: 101; 84; 67; 101
  - Wald test (Chi-square): 39.263; 33.993; 39.153; 6.72
  - Sample probability: 0.22; 0.22; 0.18; 0.22

### Interpretation of Benchmark Results
- Probability of experiencing a food crisis increases for countries with weaker institutions (CPIA) and low food supply per capita growth trend.
- Sound policy fundamentals (high fiscal balance and reserve coverage) link to lower likelihood of food crisis.
- Increase in exchange market pressure index tends to reduce probability of food supply crisis, suggesting depreciation and lower reserves may encourage domestic food production and exports.
- Growth in trading partners reduces probability; high export prices increase probability (possible incentive to shift food supply to external markets).
- Inclusion of people affected from natural disasters increases predicted probability but is not statistically significant.
- Excluding fragile states renders CPIA, fiscal balance, and EMPI insignificant, suggesting these matter more for fragile states.

### Marginal Effects (Table 3 highlights)
- Average marginal effects (column 1), median LIC (column 2), LICs with weak fundamentals (column 3):
  - CPIA (t-1): -0.1244; -0.1398; -0.1713
  - Government balance, (% of GDP) (t-1): -0.0092; -0.0103; -0.0126
  - Food supply growth average (cal): -0.2321; -0.2609; -0.3195
  - Reserve coverage (months of imports) (t-1): -0.0327; -0.0367; -0.0449
  - Export prices percent change: 0.0224; 0.0252; 0.0309
  - Growth in trading partners weighted by lagged exports to GDP: -0.1177; -0.1323; -0.1620
  - Exchange market pressure index: -0.0287; -0.0323; -0.0395
- Structural variables (average food supply per capita and CPIA) predict food crisis best, with greater prominence for countries with weaker fundamentals.

### Probability Threshold and Model Performance (Table 4)
- Threshold determination via loss function minimization (equal weights to Type I and Type II errors) yields:
  - Overall probability threshold: 0.19
  - Type I error: 14 percent
  - Type II error: 27 percent
- Distribution of predicted probabilities (Food Crisis vs Normal Episodes, percentiles):
  - 1%: Food Crisis 0.07 / Normal 0.00
  - 5%: 0.08 / 0.00
  - 10%: 0.12 / 0.00
  - 25%: 0.21 / 0.04
  - 50% (median): 0.38 / 0.10
  - 75%: 0.60 / 0.20
  - 90%: 0.83 / 0.42
  - 95%: 0.86 / 0.54
  - 99%: 0.86 / 0.64
- In-sample performance:
  - Observations: Food Crisis 28; Normal Episodes 101; Sample probability 0.22
  - Type I: 0.14; Type II: 0.27; Overall threshold: 0.19
- Out-of-sample (estimate 1990–2004, predict 2005–2009):
  - Correctly identified 7 out of 10 crises.
  - False alarms in 31 percent of non-crisis episodes (11 out of 35 non-crisis episodes misspecified as crisis).
  - Out-of-sample performance slightly worse than in-sample, possibly due to low number of crisis observations.

### Robustness Check
- Robustness test regressed every food decline episode (151) on the identified explanatory variables (not restricting to natural disaster shocks).
- Results broadly unchanged; most variables remain significant except for reserves and exchange market pressure index.
- Concludes results generally hold independent of the definition of natural disasters in Section II B.

### Signaling Approach: Methodology
- Uses signaling approach to estimate thresholds for each indicator separately; a breach signals upcoming food crisis.
- Optimal threshold balances Type I (missed crises) and Type II (false alarms) errors.
- Thresholds determined by minimizing total misclassification error TME(T) = Type I + Type II.
- Overall vulnerability index constructed by grouping variables into three clusters and weighting indicators by predictive power:
  - Clusters: (i) overall economy and institutions; (ii) external sector; (iii) fiscal sector.
  - Overall Index = weighted sum of cluster indices; indicators converted to dummies based on thresholds.

### Signaling Approach: Results and FDVI Composition (Table 5 highlights)
- Analysis uses 12 variables (including share of agriculture in GDP; life expectancy; population growth; government tax revenue; government debt).
- Sample: 55 LICs, time period 1990–2008 (food supply crisis identification requires post two-year average).
- Food crisis episodes matched with natural disasters: 42 (overall sample).
- Cluster weights in FDVI:
  - Overall economy and institutions: 0.42 (top predictor of food crisis).
    - Indicators and index weights:
      - CPIA (t-1) threshold > 38; Type I: 0.84; Type II: 0.66; Index weight: 0.05
      - Life expectancy > 38: Type I: 0.45; Type II: 0.26; Index weight: 0.13
      - Food supply growth average (cal) > 42: Type I: 0.80; Type II: 0.28; Index weight: 0.16
      - Agriculture value added (in percent of GDP) (t-1) < 37: Type I: 0.25; Type II: 0.60; Index weight: 0.07
      - Population growth > 33: Type I: 0.12; Type II: 0.72; Index weight: 0.03
  - External sector: 0.34
    - Indicators and index weights:
      - Reserve coverage (months of imports) (t-1) > 35: Type I: 0.12; Type II: 0.95; Index weight: 0.11
      - Growth in trading partners weighted by lagged exports to GDP > 36: Type I: 0.12; Type II: 0.40; Index weight: 0.09
      - Exchange market pressure index > 31: Type I: 0.15; Type II: 0.61; Index weight: 0.06
      - Export prices percent change < 35: Type I: 0.12; Type II: 0.49; Index weight: 0.08
  - Fiscal sector: 0.24
    - Indicators and index weights:
      - Government balance (% of GDP) (t-1) > 42: Type I: 0.13; Type II: 0.34; Index weight: 0.09
      - Government tax revenue (t-1) > 28: Type I: 0.93; Type II: 0.07; Index weight: 0.03
      - Government debt (t-1) < 20: Type I: 0.81; Type II: 0.49; Index weight: 0.09
      - Real government revenue (% change over 2 yrs) (t-1) > 25: Type I: 0.98; Type II: 0.11; Index weight: 0.02
- Overall index threshold and fit:
  - Overall Index threshold: 0.35
  - Proportion of Crisis Missed: 0.12
  - Proportion of Non-crisis mis-specified (false alarms): 0.45
  - Overall error: 0.37
- Note: 1/ Sample includes 63 PRGT-eligible countries for the time period 1990–2009, including climatic shocks.

### Predictive Performance of FDVI and Subcomponents
- Univariate probit regressions for FDVI and subcomponents show overall index predicted probabilities are better distributed than individual subcomponents.
- The economy and institutions cluster displays more dispersed predicted probabilities, supporting its higher weight.
- External sector cluster assigned slightly higher weight than fiscal cluster based on predictive dispersion.

### Goodness of Fit: Signaling Approach
- Table 7 (not fully reproduced in the excerpt) presents distribution of FDVI for crisis and normal episodes.
- Median predicted probability for a food crisis (from probit model): 0.38 versus 0.10 in normal times (Table 4).
- In-sample and out-of-sample distributions indicate the empirical models (probit and FDVI) are able to distinguish between crisis and non-crisis states, with dispersion of predicted probabilities in [0,1] signaling informativeness.

*Source: FAO; World Bank; EM-DAT; IMF WEO and author’s calculations (from the supplied content).*

### 0.51 against 0.33 for normal episodes. Most

### _wp1665 - 0.51 against 0.33 for normal episodes. Most

### Index performance and classification thresholds
- The food decline index: most food crisis events (75 percent) have a food decline index higher than 0.49 and only 10 percent of food crisis episodes report an index below 0.10.
- In-sample performance:
  - The index correctly calls 88 percent of the food crisis events.
  - Overall model misclassification error (sum of Type 1 and Type 2 error as percent of total observation) is 37 percent.
- Thresholds and trade-offs:
  - To distinguish high versus medium risk, the overall index threshold selected is 0.35 (obtained from minimizing an asymmetrically weighted loss function putting higher weight on Type I errors). This threshold corresponds to a Type I error of 12 percent.
  - To distinguish medium versus low probability of a food supply crisis, a threshold of 0.21 is selected. At this threshold:
    - Less than 10 percent of food supply crisis are missed.
    - False alarms increase to about 73 percent.
  - A threshold from the minimization function corresponding to a Type I error of 8 percent is used to differentiate between medium and low risk.

### Distribution of Food Decline Vulnerability Index: percentiles (selected percentiles and sectors)
- Vulnerability Index / Economy and Institutions Index / External Index / Fiscal Index — percentiles for Food Crisis vs Normal Episodes (excerpted percentiles and observations preserved):
  - 1%: Food Crisis 0.02, Normal Episodes 0.02, Food Crisis 0.05, Normal Episodes 0.05, Food Crisis 0.10, Normal Episodes 0.10, Food Crisis 0.15, Normal Episodes 0.15
  - 5%: Food Crisis 0.04, Normal Episodes 0.04, Food Crisis 0.05, Normal Episodes 0.05, Food Crisis 0.10, Normal Episodes 0.10, Food Crisis 0.15, Normal Episodes 0.15
  - 10%: Food Crisis 0.10, Normal Episodes 0.06, Food Crisis 0.11, Normal Episodes 0.06, Food Crisis 0.10, Normal Episodes 0.10, Food Crisis 0.15, Normal Episodes 0.15
  - 25%: Food Crisis 0.23, Normal Episodes 0.08, Food Crisis 0.21, Normal Episodes 0.11, Food Crisis 0.22, Normal Episodes 0.16, Food Crisis 0.18, Normal Episodes 0.15
  - 50%: Food Crisis 0.34, Normal Episodes 0.16, Food Crisis 0.33, Normal Episodes 0.21, Food Crisis 0.28, Normal Episodes 0.18, Food Crisis 0.24, Normal Episodes 0.18
  - 75%: Food Crisis 0.49, Normal Episodes 0.28, Food Crisis 0.44, Normal Episodes 0.27, Food Crisis 0.33, Normal Episodes 0.26, Food Crisis 0.29, Normal Episodes 0.28
  - 90%: Food Crisis 0.57, Normal Episodes 0.40, Food Crisis 0.49, Normal Episodes 0.41, Food Crisis 0.44, Normal Episodes 0.33, Food Crisis 0.41, Normal Episodes 0.29
  - 95%: Food Crisis 0.65, Normal Episodes 0.45, Food Crisis 0.49, Normal Episodes 0.44, Food Crisis 0.48, Normal Episodes 0.34, Food Crisis 0.41, Normal Episodes 0.41
  - 99%: Food Crisis 0.81, Normal Episodes 0.52, Food Crisis 0.52, Normal Episodes 0.52, Food Crisis 0.48, Normal Episodes 0.48, Food Crisis 0.41, Normal Episodes 0.41
  - Observations: 42 14 14 21 4 14 2 14 1 42 14 1 (as presented: Observations42141421414214142141)
- Food Decline Vulnerability Index (Table 7 excerpt percentiles):
  - 1%: Food Crisis 0.00, Normal Episodes 0.00, Food Crisis 0.33, Normal Episodes 0.03
  - 5%: Food Crisis 0.10, Normal Episodes 0.10, Food Crisis 0.33, Normal Episodes 0.03
  - 10%: Food Crisis 0.25, Normal Episodes 0.15, Food Crisis 0.33, Normal Episodes 0.09
  - 25%: Food Crisis 0.41, Normal Episodes 0.21, Food Crisis 0.33, Normal Episodes 0.19
  - 50%: Food Crisis 0.51, Normal Episodes 0.33, Food Crisis 0.43, Normal Episodes 0.23
  - 75%: Food Crisis 0.63, Normal Episodes 0.46, Food Crisis 0.53, Normal Episodes 0.37
  - 90%: Food Crisis 0.69, Normal Episodes 0.55, Food Crisis 0.53, Normal Episodes 0.44
  - 95%: Food Crisis 0.75, Normal Episodes 0.59, Food Crisis 0.53, Normal Episodes 0.45
  - 99%: Food Crisis 0.90, Normal Episodes 0.65, Food Crisis 0.53, Normal Episodes 0.45
  - Observations: 42 14 14 12 13
- Classification errors (reported):
  - Type I: 0.12 (In sample), 0.50 (Out of sample)
  - Type II: 0.45 (In sample), 0.31 (Out of sample)
  - Sample probability: 1/0.23 (In sample), 0.13 (Out of sample)
  - Note: "1/ Number of growth crisis divided by total observations."

### Assessment of vulnerabilities since the 1990s (FDVI evolution and drivers)
- Time profile (FDVI):
  - The FDVI increased sharply in 1995, declined in 1997, then steadily increased until 2003 where it reached its peak.
  - The sharp increase in the mid-nineties was mainly caused by high changes in export prices, pressure on the exchange rates and to some extent high fiscal deficits.
  - CPIA and population growth, together with the above, drove the steady increase of the FDVI from 1997 until the 2003 peak.
  - During the rise to 2003, many LICs experienced an increase in the frequency of natural disasters.
  - After 2003, with the decline in the frequency of natural disasters, food supply vulnerabilities receded due to:
    - Lower vulnerabilities in agricultural production.
    - Lower fiscal deficits.
    - Higher reserves.
    - Receding pressure on exchange rates.
  - The FDVI reached its lowest point in 2013, with a notable outlier in 2009 caused by a sudden drop in export prices and pressures on the exchange rate.
- Visualization notes (Figure 3):
  - Median, 25th percentile, and 75th percentile lines are shown for 1991–2013.
  - Risk categories shown: High, Medium, Low (timeline 1991–2013).

### Country group disaggregation and projections
- Fragile States:
  - Fragile States are more prone to a decline in food supply.
  - In 2015: the share of Fragile States with high risk is not increasing, but the share with a medium risk rating is edging up.
  - The increase in medium risk share in 2015 is partly due to higher vulnerabilities in previously low risk countries and partly due to a decrease in vulnerabilities in high risk countries.
  - For 2016 the FDVI projects a substantial increase in food insecurities with high and medium risk rating increasing across the board.
- Figures referenced for natural disasters and affected populations (Figure 4) show:
  - Number of Natural Disaster per year and percent share of LIDC population affected (bubble size).
  - Time series of natural disasters and people affected with annual counts from 1988 onward.
  - Country group comparisons for LIDCs and Fragile States in 2014, 2015, and 2016 (share of countries with low, medium and high vulnerability).

### Conclusions and policy recommendations
- Main findings:
  - Exposure of LICs to natural disasters has a significant impact on food production and food security.
  - Macroeconomic and structural indicators crucial for resilience to adverse external shocks are equally important for minimizing occurrence of food crisis associated with natural disasters.
  - Probability of experiencing a food crisis increases for countries with weaker institutions and low food supply per capita growth trend.
  - Fragile States are found to be more prone to food crisis episodes.
- Policy recommendations and mitigation options:
  - Strengthen pre-existing macroeconomic conditions to improve a country's ability to avert a food crisis in the context of a large natural disaster.
  - Secure supplementary buffers through insurance schemes, including:
    - Self-insurance via contingency reserves.
    - Rapid access to disaster support lines from donors.
  - Free up resources for disaster preparedness.
  - LICs particularly vulnerable to natural disasters and food production shocks should:
    - Seek to lower trade barriers.
    - Improve road networks to create a larger regional market.
    - Allow a faster and more efficient response to localized food shortages.

### Methods, regressions, and variables
- Two complementary approaches used:
  - Multivariate regression analysis (correlated pooled probit model with cluster-robust standard errors).
  - Univariate signaling approach.
- Appendix Table 2: Correlated Pooled Probit Regression (1990-2009) — selected reported estimates and model statistics (preserved exactly as in source):
  - CPIA (t-1): (1) -0.257** (0.0888); average marginal effects -0.063** (0.0216); (3) -0.259** (0.0897); (4) -0.964* (0.489)
  - Government balance, (% of GDP) (t-1): (1) -0.0228* (0.0108); average marginal effects -0.006* (0.0027); (3) -0.0235* (0.0110); (4) -0.0714* (0.0357)
  - Food supply growth average (cal): (1) -0.783*** (0.0708); average marginal effects -0.194*** (0.0131); (3) -0.784*** (0.0711); (4) 0.180*** (0.0529)
  - Reserve coverage (months of imports) (t-1): (1) -0.0143 (0.0255); average marginal effects -0.004 (0.0063); (3) -0.0135 (0.0255); (4) -0.970** (0.337)
  - Export prices percent change: (1) 0.0195* (0.00901); average marginal effects 0.005* (0.0022); (3) 0.0195 (0.00896); (4) -1.819*** (0.409)
  - Growth in trading partners weighted by lagged exports to GDP: (1) -0.230*** (0.0520); average marginal effects -0.057*** (0.0127); (3) -0.230*** (0.0520); (4) -0.272* (0.112)
  - Exchange market pressure index: (1) -0.0178 (0.0218); average marginal effects -0.004 (0.0054); (3) -0.0185 (0.0219); (4) -0.233* (0.116)
  - People affected from natural disasters (in percent of population): (4) 0.0722 (0.130)
  - Constant: (1) 0.617* (0.267); average marginal effects 0.597* (0.269); (4) 3.975* (1.784)
  - Observations: 743 743 129 (as presented)
  - Pseudo R-sq: 0.122 0.122 0.261
  - Number of countries: 47 45 47
  - Food decline episodes: 151 151 28
  - Normal episodes: 592 592 101
  - Wald test (Chi-square): 171.481 171.483 6.72
  - Sample probability: 0.20 0.20 0.22
  - Note: Significance denoted at 10 percent *, 5 percent **, 1 percent ***; standard errors in parentheses.
- Appendix Table 3: Variables used in probit regressions and signaling approach — sectoral listing preserved (Economy and Institutions, Fiscal sector, External sector, Shock variables) with variable names and sources:
  - Economy and Institutions: CPIA (t-1) — Lag of country policy and institutions assessment — World Bank
  - Food supply growth average (cal) — Average growth of food supply in kilo calories — Food and Agriculture Organization
  - Fiscal sector: Government balance, (% of GDP) (t-1) — World Economic Outlook
  - External sector: Reserve coverage (months of imports) (t-1) — World Economic Outlook, International Finance Statistics, Staff Reports
  - Export prices percent change — World Economic Outlook
  - Growth in trading partners weighted by lagged exports to GDP — World Economic Outlook
  - Exchange market pressure index — World Economic Outlook
  - Shock variables: Number of affected from natural disasters — CRED International Disaster Database (EM-DAT); Cost of natural disasters (in percent of GDP) — EM-DAT
  - Additional controls: Government taxes, (% of GDP) (t-1); Public debt, (% of GDP) (t-1); Agriculture value added (in percent of GDP) (t-1); Life expectancy (t-1); Population growth (t-1).

### Appendix materials (selected)
- Appendix Table 1: List of countries in the sample — 74 LICs listed; 55 countries used in the regressions (countries listed verbatim in source).
- Appendix Figure 1: FDVI by sub-indicator — share of countries with low and high probability of breaching indicator threshold for indicators including CPIA, ExportPricesPercentChange, Fiscal Balance (In percent of GDP), Exchange Market Pressure Index, Reserve coverage (In months of imports), External Demand, Life Expectancy, Agriculture ValueAdded, Population Growth (time series 1991–2013 or as noted).

*Source: Authors' calculations (as presented in the supplied content).*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2016/_wp1665.pdf_