## sipea2026057

## Source details

**Canonical URL:** [sipea2026057](https://www.imf.org/-/media/files/publications/selected-issues-papers/2026/english/sipea2026057.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/selected-issues-papers/2026/english/sipea2026057.pdf.md)
- [Structured JSON version](/-/media/files/publications/selected-issues-papers/2026/english/sipea2026057.pdf.json)

---

### A. Introduction — context and headline household outcomes
- Kiribati: population of 127 thousand people located in islands spanning 3.4mn km2 of ocean; shipping costs and lack of connectivity limit market sizes and service delivery.
- Economy dependent on public spending funded by large but volatile fishing license revenues.
- Key social safety nets provided: free education up to age 15, health care, unemployment, senior citizen and disability benefits.
- Vulnerabilities: weather shocks and long-run threat of rising sea levels; high cost of building resilience (IMF 2025, Selected Issues Paper).
- COVID-19 response and FY2025 spending:
  - Kiribati more than doubled social protection programs; total benefits in 2025 amounted to 18 percent of GDP (27 percent of government recurrent expenditure) versus 6 percent of GDP (12 percent of recurrent expenditure) in 2019.
- Household outcomes (2019/20 vs 2023/24 HIES):
  - Poverty rate fell from 21.9 to 5.5 percent.
  - Gini coefficient fell from 27.8 to 24.7.
  - Extreme poverty (US$2.15/day, 2017 PPP) was 0.04 percent in 2023/24.
  - Sanitation deprivation fell from 56 percent to 35 percent.
  - Improvements in access to electricity and water and asset ownership (e.g., vehicles).
- Between 2019 and 2023, social transfers drove most income growth for bottom-quartile households; formal labor income (excluding copra) broadly unchanged for bottom quartile.
- Resulting spending patterns: all households increased spending on food and substances (tobacco, alcohol, and kava); limited growth in other categories.

### B. Social protection programs and fiscal magnitudes (primarily 2025 unless noted)
- Main programs and fiscal magnitudes:
  - Copra subsidies: copra purchased at AUD 4/kg (doubled from AUD 2/kg in 2022), about four times the global market price; copra subsidy is the largest benefit at 9 percent of GDP.
  - Unemployment benefits (Support Fund for Unemployment, SFU, introduced 2022): AUD 50 per person per month for people aged 18-60 not formally employed; close to universal coverage—about 95 percent of people in 2023/24 lived in a household with at least one SFU recipient; about 2.4 beneficiaries per recipient household; paid quarterly due to payment capacity constraints; unemployment benefits about 5 percent of GDP.
  - Senior citizen allowance: age requirement reduced from 65 to 60 in 2020 and amount approximately tripled to AUD 200-300 per month per person; senior citizen benefits at 4 percent of GDP.
  - Disability allowance: introduced in 2020; amounts to 1 percent of GDP.
  - Contributory pension system via the Kiribati Provident Fund.
- Delivery mode and targeting:
  - Highly cash-dependent; gradual transition to digital payments. Ministry started monthly digital payments in Tarawa in 2026.
  - Unemployment benefits, senior citizen and disability allowances decline with household consumption, though unemployment benefits are less well targeted (flatter decline).
  - Copra benefits and contributory pensions increase with household consumption.
  - Around 94 percent of copra subsidy recipient households also receive unemployment benefits.
  - About 30 percent of senior citizen and disability transfers accrue to households in the bottom consumption quintile; almost 50 percent of people in the bottom quintile live in a household where at least one member receives these transfers.
  - Traditional social assistance spending higher than in comparator countries and other Pacific Islands; unemployment benefits much higher than in the OECD.

### C. Empirical approach and key estimation details
- Data and estimators:
  - Uses 2019/20 and 2023/24 HIES waves.
  - Poisson Pseudo Maximum Likelihood (PPML) estimator used to estimate semi-elasticities in presence of heteroskedasticity and zeros.
  - Specification includes village fixed effects; household head ethnicity, education, gender, age, marital status, labor force status; indicators for home ownership, church donations, car ownership; controls for non-transfer income, household size, number of children and seniors.
- Hurdle approach to zero spending:
  - Extensive margin: logit model for decision to spend.
  - Intensive margin: log-level regression conditional on spending.
- Extensive zero-spending shares (2023/24 HIES):
  - Substances: about 11 percent of households did not report any substance spending.
  - Education: 28 percent did not report spending.
  - Health: 79 percent did not report spending.
  - Business expenditure: 51 percent did not report spending.

### D. Main empirical findings — 2019 vs 2023 and margins
- 2019 findings:
  - Social protection receipts positively associated with spending on food, health, and business expenditure.
  - Copra income associated with higher spending on food and business, and increased expenditure on alcohol/tobacco.
  - Senior citizen allowance receipts associated with higher spending on food, health and business, and lower spending on alcohol/tobacco.
  - Pension income associated with higher spending on education and health, as well as kava.
- 2023 findings:
  - Link between social protection and human development spending less clear.
  - Unemployment benefits (now the largest) associated with lower spending on education and higher spending on substances (alcohol, tobacco, kava).
  - Copra income remains positively associated with food and business expenditure.
  - Pension income associated with higher spending on education and kava.
  - Senior and disability allowances no longer statistically significantly associated with specific spending categories.
- Extensive-margin (probability of spending):
  - Unemployment receipts associated with lower probability of spending on education (and small negative magnitudes for health and business) and higher probability of spending on substances.
  - Bottom-quartile households significantly less likely to spend on alcohol and tobacco given unemployment receipts: interaction between SFU receipts and bottom-quartile dummy is statistically significant and negative.
- Intensive-margin (amount conditional on spending):
  - Average marginal effect of unemployment receipts on substance spending smaller among bottom-quartile households.
  - On average, an additional AUD in SFU receipts is associated with 34 cents in alcohol/tobacco spending in the top three quartiles and 25 cents in the bottom quartile.
- Causal analysis:
  - Fuzzy regression discontinuity exploited via age eligibility for SFU; instruments constructed from number of household members eligible for unemployment benefits by age (19-59).
  - Estimation via control-function (two-stage residual inclusion) with PPML in second stage.
  - Results (Tables 5-6) confirm a statistically significant causal effect of unemployment receipts on substance spending, with larger estimated semi-elasticities than baseline regressions.
  - Control functions are highly significant, indicating endogeneity in SFU receipts unaddressed by regressions without instruments.

### E. Robustness checks and diagnostics (selected numeric diagnostics preserved verbatim)
- PPML with control function diagnostics:
  - J statistic: 0.424; 0.250; 4.568; 1.289; 3.537; 1.117
  - J test p-value: 0.809; 0.883; 0.102; 0.525; 0.171; 0.572
  - First-stage joint Wald chi2: 63.740 (for all six outcomes)
  - First-stage joint test p-value: 0.000 (for all six outcomes)
- Hurdle model diagnostic statistics (Table 9):
  - Extensive J statistic: 0.939; 2.479; 1.869; 3.066; 0.057
  - Extensive J p-value: 0.625; 0.289; 0.393; 0.216; 0.972
  - Intensive J statistic: 0.128; 0.009; 0.164; 3.489; 1.289
  - Intensive J p-value: 0.938; 0.995; 0.921; 0.175; 0.525
  - First stage (unemp. income) joint Wald chi2: 11.575; First stage joint test p-value: 0.009; First stage joint F: 3.858
  - First stage (unemp. income × bottom quartile) joint Wald chi2: 495.570; First stage joint test p-value: 0.000
- Robustness to alternative specifications:
  - Results robust to additional controls (value of household durables, number of rooms), different functional forms (linear levels, probit for extensive margin), and redefinition of instruments as indicators for exact number of age-eligible people.

### F. Robustness check — spending patterns by age-eligibility (main finding)
- Comparison between (potential) SFU recipient households and senior citizen allowance recipient households, based on number of age-eligible household members:
  - Relative to an average non-recipient household, spending on kava and tobacco on average increases markedly faster than spending on food as the number of SFU-age-eligible people in the household increases.
  - Spending on substances does not increase substantially more than spending on food with the number of senior citizens in the household.
- Figures (preserved chart scales as in source):
  - Cumulative percent change series for Food, Alcohol, Tobacco, Kava with SFU-age-eligible members (horizontal axis: 0–5; vertical axis range: 0 to 800).
  - Cumulative percent change series for Food, Alcohol, Tobacco, Kava with Senior Citizens in household (horizontal axis: 0–2; vertical axis range: 80 to 150).
  - Spending per Additional AUD of SFU Receipts (AUD; by household consumption quartile): vertical axis range: 0 to 0.45.
  - Error bands: 95% confidence intervals with standard errors clustered at the village level.

### G. Policy conclusions and recommendations (summary of paragraphs 16–19 and specific agenda)
- Overarching assessment:
  - Social protection benefits have substantially and rapidly reduced poverty in Kiribati and ensure most I-Kiribati have access to basic income given limited education access, difficulties reaching formal labor markets, and lack of scale economies, especially in rural areas and outer islands.
  - Current approach effective during shocks: COVID-19 pandemic in 2020 and the cost-of-living crisis in 2022 and 2023.
  - Need to enhance efficiency of social benefits while maintaining fiscal sustainability and addressing climate-resilient infrastructure investment needs.
  - Some instruments (notably unemployment benefits and the copra subsidy) operate at very large scale with limited targeting, leading to leakage to better-off households and potential distortions.
- Efficiency assessment criteria:
  - Coverage (whether beneficiaries and benefit shares are concentrated among poorer households).
  - Efficiency (how much poverty and inequality are reduced per dollar spent), while ensuring fiscal sustainability.
- Specific reform agenda items:
  - Increasing excises on tobacco and alcohol and introducing an excise on kava:
    - Purpose: curb spending with negative public health externalities; improve human development; supplement expenditure measures.
  - Continuing copra subsidy reforms:
    - Purpose: avoid distorting agricultural production.
    - Options: shift more towards low-income households by introducing caps or progressively replacing broad price support with targeted cash transfers.
  - Reviewing the adequacy and efficiency of social benefits:
    - Purpose: rationalize overlaps between programs (e.g. copra subsidy and unemployment benefits), reduce potential benefit leakage, and improve targeting.
- Implementation and administrative recommendations:
  - Improve delivery and administration systems to enhance efficiency and equity:
    - Reduce cash dependence.
    - Strengthen beneficiary identification and records.
    - Modernize payment channels.
    - Expected benefits: lower administrative costs, improved timeliness and transparency, enable agile support without expanding untargeted transfers.
  - Targeting considerations for Kiribati:
    - True targeting (means testing) has prohibitively high administrative costs in Kiribati.
    - Targeting could be improved by tightening program eligibility, e.g., capping the copra subsidy and preventing copra and unemployment benefits from accruing to the same households.
- Overall policy objective:
  - Avoid introducing market distortions, supporting spending with negative externalities, or providing higher benefits to richer households.
  - Redirect scarce public resources toward high-impact services and well-targeted transfers that primarily benefit lower-income households.

*Source: sipea2026057 — KNSO HIES; and IMF staff estimation / IMF staff calculations.*

### References ____________________________________________________________________________ 11

### sipea2026057 - References

### A. Introduction
- Kiribati: population of 127 thousand people located in islands spanning 3.4mn km2 of ocean; shipping costs and lack of connectivity limit market sizes and service delivery.
- Economy dependent on public spending funded by large but volatile fishing license revenues.
- Key social safety nets provided: free education up to age 15, health care, unemployment, senior citizen and disability benefits.
- Vulnerabilities: weather shocks and long-run threat of rising sea levels; high cost of building resilience (IMF 2025, Selected Issues Paper).
- COVID-19 response: Kiribati more than doubled social protection programs; total benefits in 2025 amounted to 18 percent of GDP (27 percent of government recurrent expenditure) versus 6 percent of GDP (12 percent of recurrent expenditure) in 2019.
- Household outcomes (2019/20 vs 2023/24 HIES):
  - Poverty rate fell from 21.9 to 5.5 percent.
  - Gini coefficient fell from 27.8 to 24.7.
  - Extreme poverty (US$2.15/day, 2017 PPP) was 0.04 percent in 2023/24.
  - Sanitation deprivation fell from 56 percent to 35 percent.
  - Improvements in access to electricity and water and asset ownership (e.g., vehicles).
- Between 2019 and 2023, social transfers drove most income growth for bottom-quartile households; formal labor income (excluding copra) broadly unchanged for bottom quartile.
- Resulting spending patterns: all households increased spending on food and substances (tobacco, alcohol, and kava); limited growth in other categories.

### B. Social Protection Programs in Kiribati
- Main programs and fiscal magnitudes (2025 unless otherwise noted):
  - Copra subsidies: copra purchased at AUD 4/kg (doubled from AUD 2/kg in 2022), about four times the global market price; copra subsidy is the largest benefit at 9 percent of GDP.
  - Unemployment benefits (Support Fund for Unemployment, SFU, introduced 2022): AUD 50 per person per month for people aged 18-60 not formally employed; close to universal coverage—about 95 percent of people in 2023/24 lived in a household with at least one SFU recipient; about 2.4 beneficiaries per recipient household; paid quarterly due to payment capacity constraints; unemployment benefits about 5 percent of GDP.
  - Senior citizen allowance: age requirement reduced from 65 to 60 in 2020 and amount approximately tripled to AUD 200-300 per month per person; senior citizen benefits at 4 percent of GDP.
  - Disability allowance: introduced in 2020; amounts to 1 percent of GDP.
  - Contributory pension system via the Kiribati Provident Fund.
- Delivery mode: highly cash-dependent; gradual transition to digital payments. Ministry started monthly digital payments in Tarawa in 2026.
- Progressivity and targeting:
  - Unemployment benefits, senior citizen and disability allowances decline with household consumption, though unemployment benefits are less well targeted (flatter decline).
  - Copra benefits and contributory pensions increase with household consumption.
  - Around 94 percent of copra subsidy recipient households also receive unemployment benefits.
  - About 30 percent of senior citizen and disability transfers accrue to households in the bottom consumption quintile; almost 50 percent of people in the bottom quintile live in a household where at least one member receives these transfers.
  - Traditional social assistance spending higher than in comparator countries and other Pacific Islands; unemployment benefits much higher than in the OECD.

### C. Social Protection Benefits and Spending: 2019 vs 2023
- Methodology:
  - Uses 2019/20 and 2023/24 HIES waves.
  - Poisson Pseudo Maximum Likelihood (PPML) estimator to estimate semi-elasticities in presence of heteroskedasticity and zeros.
  - Specification includes village fixed effects; household head ethnicity, education, gender, age, marital status, labor force status; indicators for home ownership, church donations, car ownership; controls for non-transfer income, household size, number of children and seniors.
- Key findings:
  - 2019:
    - Social protection receipts positively associated with spending on food, health, and business expenditure.
    - Copra income associated with higher spending on food and business, but also with increased expenditure on alcohol/tobacco.
    - Senior citizen allowance receipts associated with higher spending on food, health and business, and lower spending on alcohol/tobacco.
    - Pension income associated with higher spending on education and health, as well as kava.
  - 2023:
    - Link between social protection and human development spending less clear.
    - Unemployment benefits (now the largest) associated with lower spending on education and higher spending on substances (alcohol, tobacco, kava).
    - Copra income remains positively associated with food and business expenditure.
    - Pension income associated with higher spending on education and kava.
    - Senior and disability allowances no longer statistically significantly associated with specific spending categories.
- Causal analysis:
  - Fuzzy regression discontinuity (Annex I) used to establish causal effect of unemployment benefits on alcohol, tobacco, and kava spending.
  - Instruments constructed from number of household members eligible for unemployment benefits by age (19-59).
  - Estimation via control-function (two-stage residual inclusion) with PPML in second stage.
  - Results (Tables 5-6) confirm a statistically significant causal effect of unemployment receipts on substance spending, with larger estimated semi-elasticities than baseline regressions.

### D. What to Spend on, and How Much? Extensive and Intensive Margins
- Rationale: substantial shares of households report zero spending in many categories; separate analysis of:
  - Extensive margin: decision to spend or not (logit model).
  - Intensive margin: amount spent conditional on spending (log-level regression).
- Zero-spending shares (2023/24 HIES):
  - About 11 percent of households did not report any substance spending.
  - Education: 28 percent did not report spending.
  - Health: 79 percent did not report spending.
  - Business expenditure: 51 percent did not report spending.
- Findings on margins:
  - Extensive-margin (probability of spending):
    - Unemployment receipts associated with lower probability of spending on education (and small negative magnitudes for health and business) and higher probability of spending on substances.
    - Bottom-quartile households significantly less likely to spend on alcohol and tobacco given unemployment receipts: interaction between SFU receipts and bottom-quartile dummy is statistically significant and negative.
  - Intensive-margin:
    - Overall average marginal effect of unemployment receipts on substance spending smaller among bottom-quartile households.
    - On average, an additional AUD in SFU receipts is associated with 34 cents in alcohol/tobacco spending in the top three quartiles and 25 cents in the bottom quartile.
  - Robustness:
    - Causal relation between unemployment benefits and substance spending confirmed by fuzzy regression discontinuity; results robust to additional controls and alternative functional forms (Annex I; Tables 7-9).

*Source: https://www.imf.org/-/media/files/publications/selected-issues-papers/2026/english/sipea2026057.pdf*

### 15.      Finally, as a robustness check, spending across categories is compared for households

### 15.      Finally, as a robustness check, spending across categories is compared for households

### Robustness check — main finding
- Comparison between (potential) SFU recipient households and senior citizen allowance recipient households, based on the number of age-eligible household members for each benefit.
- Relative to an average non-recipient household, spending on kava and tobacco on average increases markedly faster than spending on food as the number of SFU-age-eligible people in the household increases.
- Spending on substances does not increase substantially more than spending on food with the number of senior citizens in the household.

### Methodology — combining extensive and intensive margins
- To quantify and combine the estimates from the two margins, the average marginal effects of SFU receipts are calculated as the average across the marginal effect of an additional AUD of SFU receipts through both the probability of spending and the amount spent in expectation.
- The formula presented in the source is:
  ∂E(y_i)/∂SFUIncome_i = ∂P(y_i>0)/∂SFUIncome_i × exp((log y_i)^ ) + P(y_i>0)^ × ∂E(y_i|y_i>0)/∂SFUIncome_i,
  where:
  - ∂P(y_i>0)/∂SFUIncome_i is the marginal effect from the extensive margin estimation.
  - ∂E(y_i|y_i>0)/∂SFUIncome_i is the marginal effect from the intensive margin estimation.

### Figures and statistical notes (as described in the source)
- Spending on Food and Substances (Cumulative percent change relative to households with 0 member aged 19-59):
  - Horizontal axis: Number of SFU-Age-Eligible Household Members (0, 1, 2, 3, 4, 5).
  - Series reported: Food, Alcohol, Tobacco, Kava.
  - Vertical axis values shown in the chart range from 0 to 800 (preserved as in the figure).
- Spending on Food and Substances (Cumulative percent change relative to households with 0 member aged 60+):
  - Horizontal axis: Number of Senior Citizens in the Household (0, 1, 2).
  - Series reported: Food, Alcohol, Tobacco, Kava.
  - Vertical axis values shown in the chart range from 80 to 150 (preserved as in the figure).
- Spending per Additional AUD of SFU Receipts (AUD; by household consumption quartile):
  - Separate series for Bottom quartile and Top three quartiles for Alcohol/Tobacco and for Kava.
  - Vertical axis values shown in the chart range from 0 to 0.45 (preserved as in the figure).
- Notes: The error bands show 95% confidence intervals for the average overall marginal effects, with standard errors clustered at the village level.
- Sources cited in the charts: KNSO HIES; and IMF staff estimation / IMF staff calculations.

### Conclusions and policy recommendations (summary of paragraphs 16–19)
- Social protection benefits have substantially and rapidly reduced poverty in Kiribati and ensure most I-Kiribati have access to basic income given limited education access, difficulties reaching formal labor markets, and lack of scale economies, especially in rural areas and outer islands.
- The current approach has been effective during shocks: COVID-19 pandemic in 2020 and the cost-of-living crisis in 2022 and 2023.
- Need to enhance efficiency of social benefits while maintaining fiscal sustainability and addressing climate-resilient infrastructure investment needs.
- Some instruments (notably unemployment benefits and the copra subsidy) operate at very large scale with limited targeting, leading to leakage to better-off households and potential distortions.
- Efficiency assessment criteria: coverage (whether beneficiaries and benefit shares are concentrated among poorer households) and efficiency (how much poverty and inequality are reduced per dollar spent), while ensuring fiscal sustainability.
- IMF 2022 referenced for overview on designing and institutionalizing a spending review; IMF, 2019 referenced for trade-offs in targeting approaches (references noted in source).

### Specific reform agenda items (as listed)
- Increasing excises on tobacco and alcohol and introducing an excise on kava:
  - Purpose: curb spending with negative public health externalities; improve human development; supplement expenditure measures.
- Continuing copra subsidy reforms:
  - Purpose: avoid distorting agricultural production.
  - Options: shift more towards low-income households by introducing caps or progressively replacing broad price support with targeted cash transfers.
- Reviewing the adequacy and efficiency of social benefits:
  - Purpose: rationalize overlaps between programs (e.g. copra subsidy and unemployment benefits), reduce potential benefit leakage, and improve targeting.

### Implementation and administrative recommendations
- Improve delivery and administration systems to enhance efficiency and equity:
  - Reduce cash dependence.
  - Strengthen beneficiary identification and records.
  - Modernize payment channels.
  - Expected benefits: lower administrative costs, improved timeliness and transparency, enable agile support without expanding untargeted transfers.
- Targeting considerations for Kiribati:
  - True targeting (means testing) has prohibitively high administrative costs in Kiribati.
  - Targeting could be improved by tightening program eligibility, e.g., capping the copra subsidy and preventing copra and unemployment benefits from accruing to the same households.
- Overall policy objective:
  - Avoid introducing market distortions, supporting spending with negative externalities, or providing higher benefits to richer households.
  - Redirect scarce public resources toward high-impact services and well-targeted transfers that primarily benefit lower-income households.

*Sources: KNSO HIES; and IMF staff estimation.*

### References

### sipea2026057 - References

### References list
- Doherty, L. and Sayegh, A. How to Design and Institutionalize Spending Reviews, IMF How to Notes, 2022, 004 (2022).
- International Monetary Fund. Fiscal Affairs Dept. Kiribati: Agricultural Subsidies for Copra: Improving Efficiency and Equity and Balancing Trade-offs, High-Level Summary Technical Assistance Reports, 2024, 042 (2024).
- International Monetary Fund. Asia and Pacific Dept. Kiribati: Selected Issues. IMF Staff Country Reports, 2024, 104 (2024).
- International Monetary Fund. Asia and Pacific Dept. Kiribati: Selected Issues, IMF Staff Country Reports, 2025, 173 (2025).
- Terza, J. V., Basu, A., & Rathouz, P. J. (2008). Two-stage residual inclusion estimation: Addressing endogeneity in health econometric modeling. Journal of Health Economics, 27(2), 531-543.
- Santos Silva, J. M. C., & Tenreyro, S. (2006). The log of gravity. Review of Economics and Statistics, 88(4), 641–658.
- Sargan, J. D. (1958). The estimation of economic relationships using instrumental variables. Econometrica, 26(3), 393–415.
- World Bank. 2024. Beyond Copra.

### Tables and empirical outputs (selected numeric results preserved verbatim)
- Table 1 & Table 2: Kiribati: Poisson Pseudo Maximum Likelihood Estimates of Spending Semi-Elasticities in 2019/20 HIES and 2023/24 HIES
  - Senior Allowance (2019/20 or 2023/24 entries shown across columns):
    - 6.35e-5**; -0.0002; 0.0006***; 0.0003**; -0.0002**; -0.0002
    - Standard errors: (3.03e-5); (0.0003); (0.0002); (0.0001); (0.0001); (0.0001)
  - Copra Income:
    - 1.03e-5***; 2.09e-5; 2.37e-6; 6.33e-5***; 1.78e-5***; -2.8e-5***
    - Standard errors: (2.95e-6); (2.3e-5); (2.06e-5); (1.25e-5); (6.82e-6); (1.04e-5)
  - Pension Income:
    - 2.21e-5; 0.0003***; 0.0002***; -0.0004***; -2.73e-5; 0.0001**
    - Standard errors: (1.53e-5); (7.62e-5); (7.07e-5); (5.32e-5); (3.37e-5); (5.05e-5)
  - Other Income:
    - 1.08e-5***; 1.02e-5***; 1.29e-5***; 3.54e-5***; 1.17e-6; 1.31e-5***
    - Standard errors: (6.8e-7); (3.28e-6); (4.26e-6); (2.24e-6); (1.95e-6); (2.27e-6)
  - Household Size:
    - 0.0713***; 0.1342***; 0.0127; -0.0497***; 0.0927***; 0.0643***
    - Standard errors: (0.0034); (0.0214); (0.0233); (0.0159); (0.0095); (0.0099)
  - Number of Children:
    - 0.0035; 0.0816**; -0.0412; -0.0465*; 0.0134; -0.0309
    - Standard errors: (0.0055); (0.0370); (0.0450); (0.0257); (0.0158); (0.0202)
  - Number of Seniors:
    - 0.0467**; 0.1001; -0.6857***; 0.1447; 0.0348; 0.0055
    - Standard errors: (0.0189); (0.1929); (0.1765); (0.1215); (0.0552); (0.0715)
  - Model fit and sample sizes (selected):
    - Observations: 3,432; 2,865; 1,969; 3,288; 3,432; 3,423
    - Squared Cor.: 0.53587; 0.37799; 0.32759; 0.44273; 0.28715; 0.28014
    - Pseudo R2: 0.54441; 0.50003; 0.31230; 0.45884; 0.33681; 0.29082
    - BIC: 3,264,337.1; 531,112.1; 80,646.1; 2,245,623.9; 2,176,131.1; 3,462,870.2

- Alternative PPML estimates (2023/24 subset shown)
  - Unemployment Benefits coefficients (row showing six dependent vars):
    - -2.77e-5; -0.0004***; -0.0001; -0.0001; 0.0002***; 0.0003***
    - Standard errors: (2.08e-5); (7.3e-5); (0.0001); (0.0001); (5.59e-5); (5.44e-5)
  - Copra Income (alternative rows):
    - 8.03e-6***; 8.25e-6; 1.59e-5; 2.1e-5*; -6.24e-6; 8.36e-6
    - Standard errors: (2.32e-6); (1.05e-5); (1.75e-5); (1.09e-5); (5.5e-6); (5.67e-6)
  - Other Income (alternative rows):
    - 5.66e-6***; 7.32e-7; 1.03e-5***; 2.56e-5***; 7.36e-6***; 7.55e-6***
    - Standard errors: (6.41e-7); (2.32e-6); (2.79e-6); (2.8e-6); (1.7e-6); (1.48e-6)
  - Household Size (alternative rows):
    - 0.0793***; 0.3274***; 0.1023**; -0.0760; 0.0877***; 0.0928***
    - Standard errors: (0.0100); (0.0354); (0.0514); (0.0625); (0.0267); (0.0284)
  - Observations (alternative PPML rows): 1,679; 1,676; 1,290; 1,646; 1,679; 1,668
  - Squared Cor.: 0.51670; 0.39183; 0.29135; 0.44920; 0.31078; 0.25244
  - Pseudo R2: 0.54426; 0.47665; 0.33806; 0.50624; 0.35479; 0.25187

- Table 3: Kiribati: Logit Estimates of Spending Odds Semi-Elasticities in 2023/24 HIES (selected coefficients)
  - Unemployment benefits:
    - -0.0010***; -0.0003*; -0.0002*; 0.0010***; 0.0007***
    - Standard errors: (0.0002); (0.0001); (0.0001); (0.0002); (0.0002)
  - Copra income:
    - 4.56e-5**; 2.93e-5; 1.08e-6; -8.78e-6; -1.01e-5
    - Standard errors: (1.82e-5); (2.07e-5); (1.37e-5); (1.77e-5); (1.48e-5)
  - Other income:
    - -4.77e-7; 1.17e-5***; 1.81e-5***; 6.92e-7; 1.56e-6
    - Standard errors: (5.48e-6); (4.08e-6); (4.56e-6); (5.57e-6); (4.23e-6)
  - Household size:
    - 0.8914***; 0.2573***; 0.1379**; 0.3818***; 0.2636***
    - Standard errors: (0.1186); (0.0750); (0.0644); (0.0953); (0.0703)
  - Observations: 1,628; 1,287; 1,591; 1,542; 1,619
  - Squared Cor.: 0.48784; 0.23433; 0.22829; 0.19464; 0.22857
  - Pseudo R2: 0.42690; 0.20307; 0.18266; 0.18780; 0.18557

- Table 4: Conditional-on-spending>0 log regressions (2023/24 HIES, selected)
  - Unemployment benefits:
    - -0.0004***; -6.88e-6; 8.8e-6; 0.0002***; 8.03e-5*
    - Standard errors: (6.66e-5); (0.0001); (0.0001); (4.89e-5); (4.65e-5)
  - Copra income:
    - 4.62e-7; -4.75e-6; 4.15e-5***; -3.28e-6; 1.07e-5**
    - Standard errors: (6.13e-6); (2.35e-5); (1.36e-5); (4.86e-6); (5.15e-6)
  - Other income:
    - -8.27e-7; 3.04e-6; 2.05e-5***; 4.66e-6***; 4.41e-6***
    - Standard errors: (2.04e-6); (2.77e-6); (3.27e-6); (1.57e-6); (1.66e-6)
  - Household size:
    - 0.2887***; 0.0626; -0.0731; 0.1452***; 0.1048***
    - Standard errors: (0.0330); (0.0494); (0.0510); (0.0258); (0.0264)
  - Observations: 1,204; 365; 877; 1,379; 1,110
  - R2: 0.40879; 0.44265; 0.35613; 0.38343; 0.30655
  - Within R2: 0.10972; 0.05085; 0.09911; 0.15454; 0.09718

- Table 5 & Table 6: PPML with Control Function Estimates and Diagnostics (2023/24 HIES)
  - PPML with control function: selected coefficients for Unemployment Benefits and First-Stage Residual
    - Unemployment Benefits: 0.0002**; -0.0003; -8.9e-5; -0.0007; 0.0010***; 0.0013***
      - Standard errors: (8.47e-5); (0.0005); (0.0007); (0.0007); (0.0002); (0.0002)
    - First-Stage Residual: -0.0002***; -9.17e-5; -4.45e-5; 0.0006; -0.0008***; -0.0011***
      - Standard errors: (8.53e-5); (0.0004); (0.0007); (0.0007); (0.0002); (0.0002)
  - Copra Income (PPML CF):
    - 6.26e-6***; 7.5e-6; 1.56e-5; 2.49e-5*; -1.28e-5**; 1.23e-7
    - Standard errors: (2.31e-6); (1.13e-5); (2.22e-5); (1.42e-5); (6.46e-6); (6.09e-6)
  - Diagnostics (Table 6):
    - J statistic: 0.424; 0.250; 4.568; 1.289; 3.537; 1.117
    - J test df: 2; 2; 2; 2; 2; 2
    - J test p-value: 0.809; 0.883; 0.102; 0.525; 0.171; 0.572
    - First-stage joint Wald chi2: 63.740 (for all six outcomes)
    - First-stage joint test p-value: 0.000 (for all six outcomes)
  - Observations (PPML CF): 1,679; 1,676; 1,290; 1,646; 1,679; 1,668
  - Squared Cor.: 0.51967; 0.39239; 0.29159; 0.45087; 0.32002; 0.27559
  - Pseudo R2: 0.54645; 0.47671; 0.33807; 0.50721; 0.36359; 0.26530

- Table 7 & Table 8: Logit with Control Functions and Control Function estimates conditional on spending>0 (selected)
  - Logit with control functions (Table 7) — Unemployment benefits:
    - -0.0026; -0.0008; -0.0007; 0.0041*; 0.0038*
    - Observations: 1,628; 1,287; 1,591; 1,542; 1,619
    - Squared Cor.: 0.48988; 0.23600; 0.22847; 0.19670; 0.23077
    - Pseudo R2: 0.42794; 0.20479; 0.18276; 0.18987; 0.18808
  - Control Function conditional-on-spending>0 (Table 8) — selected:
    - Copra income: 1.33e-5; 2.31e-5; 5.05e-5**; -5.44e-6; 8.92e-6
      - Standard errors: (1.42e-5); (3.68e-5); (2.02e-5); (9.54e-6); (1.01e-5)
    - Other income: -3.81e-6; -3.95e-6; 1.85e-5***; 5.16e-6**; 4.82e-6*
      - Standard errors: (3.28e-6); (6e-6); (4.99e-6); (2.25e-6); (2.48e-6)
    - Household size: 0.7951**; 1.172*; 0.2924; 0.0539; 0.0303
      - Standard errors: (0.3953); (0.6852); (0.5663); (0.3106); (0.3063)
    - Observations: 1,204; 365; 877; 1,379; 1,110
    - Squared Cor.: 0.41342; 0.45694; 0.35721; 0.38372; 0.30660
    - Pseudo R2: 0.11668; 0.07519; 0.10063; 0.15494; 0.09725

- Table 9: Diagnostic Statistics for the Hurdle Model with Control Function Approach (2023/24 HIES)
  - Extensive J statistic: 0.939; 2.479; 1.869; 3.066; 0.057
  - Extensive J df: 2; 2; 2; 2; 2
  - Extensive J p-value: 0.625; 0.289; 0.393; 0.216; 0.972
  - Intensive J statistic: 0.128; 0.009; 0.164; 3.489; 1.289
  - Intensive J df: 2; 2; 2; 2; 2
  - Intensive J p-value: 0.938; 0.995; 0.921; 0.175; 0.525
  - First stage (unemp. income) joint Wald chi2: 11.575 (for each outcome)
    - First stage joint test p-value: 0.009 (for each outcome)
    - First stage joint F: 3.858 (for each outcome)
  - First stage (unemp. income × bottom quartile) joint Wald chi2: 495.570 (for each outcome)
    - First stage joint test p-value: 0.000 (for each outcome)

*Source: sipea2026057 - References (PDF chapter/section).*

### Annex I. Robustness and Causal Inference

### Annex I. Robustness and Causal Inference

### Identification strategy and instruments
- Exploits discontinuity in age eligibility for SFU to address endogeneity concerns.
- Constructs instrumental variables indicating whether the number of people aged 19-59 in each household is no less than one, two, and so on.
- Identification relies on imperfect discontinuity (fuzzy regression discontinuity) in SFU receipt (treatment intensity) across otherwise similar households differing in eligibility solely due to age differences.
- Notes potential concern that HIES uses five-year grouping for age (five years in age difference may not be small enough for instruments to be truly exogenous).
- Sargan overidentification test cannot reject the null that the instruments are endogenous, alleviating concern that coarse age grouping contaminates instruments with age-varying omitted factors that correlate with both SFU receipts and the dependent variable (Tables 6 and 9).

### Estimation and inference approach
- Uses control function approach (two-stage residual inclusion) to handle model nonlinearity.
- First stage: linear fixed-effect regressions estimate relationship between SFU receipts and age-eligibility instruments; residuals capture endogenous component of SFU receipts.
- Second stage: includes first-stage residuals as a control while keeping SFU receipts unchanged from the equation without instruments; estimated with same estimator as estimation without instruments.
- Notes that in a linear regression model, this is mathematically identical to using the fitted value from the first stage to proxy the endogenous variable.
- Control functions are highly significant, indicating endogeneity in SFU receipts unaddressed by regressions without instruments.
- Inference: standard errors, p-values, and confidence intervals obtained by repeated estimation on village-clustered bootstrapped samples.

### Causal evidence on SFU receipts and spending
- Confirms causal effect of SFU receipts on spending patterns.
- Table 3: estimated semi-elasticity of spending on substances with respect to SFU receipts remains highly significant in the control-function approach, and the magnitude increases.
- Discusses counterintuitive sign: omitted factors might bias estimates upward if positively correlated with both SFU receipts and substance spending; geographic concentration of employment and supply in Tarawa could generate negative correlation between unemployment and substance spending due to geographical factors.

### Causality of spending decision (extensive vs intensive margins)
- Applies two-stage residual inclusion strategy for dichotomous spending decision: first-stage prediction of SFU receipts using age-eligibility instruments; include first-stage residual in second-stage logit equation alongside SFU receipts.
- Conditioning on residual nets out endogenous component of SFU receipts correlated with unobserved determinants of household spending, restoring exogeneity for causal interpretation.
- Inference: village-clustered bootstrap standard errors to account for non-random generated regressor (control function).
- Results: Both SFU receipts and its interaction with bottom quartile dummy have unchanged signs and remain significant (Tables 7-8).

### Robustness checks
- Estimated effects on spending level and spending decision regarding substances are robust to:
  - Additional controls such as value of household durables and number of rooms in the dwelling (no measurable differences in key estimates in PPML regression or hurdle model).
  - Different functional forms (simple linear regression in levels and probit model for the extensive margin do not materially change results).
  - Redefinition of instruments as indicators for the exact number of age-eligible people.

*Source: Annex I. Robustness and Causal Inference, sipea2026057.*

---


_Source: https://www.imf.org/-/media/files/publications/selected-issues-papers/2026/english/sipea2026057.pdf_