Guoliang Lu, Xinyun Tang
The posterior-mean plug-in estimate was b = -0.0157 (95% CI, -0.0286 to -0.0027; p = 0.018), whereas the SWIID multiple-imputation estimate was b = -0.00831 (95% CI, -0.01688 to 0.00026; p = 0.057; fraction of missing information = 0.173). This sensitivity bracket was not treated as formal bounds and straddled the post hoc smallest effect size of interest of |b| = 0.010. Population- and log-population-weighted estimates were -0.00861 (95% CI, -0.02166 to 0.00443; p = 0.193) and -0.00864 (95% CI, -0.01718 to -0.00010; p = 0.047), respectively. Country-trend, region-year, first-difference, and dynamic estimates were smaller and less precise. All 100 completed-data coefficients and all 170 leave-one-country-out posterior-mean coefficients were negative, while subsets with narrower WHO uncertainty intervals.
INTRODUCTION: Income inequality is a distal population exposure, and ecological panel data cannot identify individual-level suicide mechanisms. We examined whether within-country changes in national disposable-income inequality were associated with changes in suicide mortality and assessed how the estimated association depended on exposure uncertainty, outcome measurement, sample composition, population weighting, and temporal identifying assumptions.
METHODS: We linked World Health Organization (WHO) age-standardized suicide rates with Standardized World Income Inequality Database (SWIID) disposable-income Gini estimates for 170 countries from 2000 to 2021 (3,257 country-years). The primary specification was a fully adjusted, unweighted two-way fixed-effects model with country-clustered standard errors. All 100 completed SWIID version 9.92 series were pooled using Rubin's rules and Barnard-Rubin degrees of freedom. Sensitivity analyses examined population weighting, temporal specifications, alternative outcomes, reporting and measurement-quality subsets, and country influence. We also conducted an observed-panel simulation and equivalence testing within ±0.010 log-rate units per Gini point.
RESULTS: The posterior-mean plug-in estimate was b = -0.0157 (95% CI, -0.0286 to -0.0027; p = 0.018), whereas the SWIID multiple-imputation estimate was b = -0.00831 (95% CI, -0.01688 to 0.00026; p = 0.057; fraction of missing information = 0.173). This sensitivity bracket was not treated as formal bounds and straddled the post hoc smallest effect size of interest of |b| = 0.010. Population- and log-population-weighted estimates were -0.00861 (95% CI, -0.02166 to 0.00443; p = 0.193) and -0.00864 (95% CI, -0.01718 to -0.00010; p = 0.047), respectively. Country-trend, region-year, first-difference, and dynamic estimates were smaller and less precise. All 100 completed-data coefficients and all 170 leave-one-country-out posterior-mean coefficients were negative, while subsets with narrower WHO uncertainty intervals.