2025-05-13-世界银行-地方福利的非正态经验贝叶斯预测（英）_30页_1mb

Summary

This paper addresses the challenge of estimating local welfare, such as poverty and inequality, at the small area level using household survey and census data. It proposes a method called "Non-Normal Empirical Bayes Prediction" to improve estimation precision by accounting for deviations from normality in error distributions, without sacrificing the computational efficiency of Empirical Best (EB) estimation.

Key findings include:

1. Deviations from Normality: Empirical data from 142 household surveys across 16 countries show significant non-normality in both area and household errors, especially in household idiosyncratic errors.
2. EB Advantage: When the area error (location effect) constitutes ≥5% of total error, standard Normal-EB estimation outperforms Non-EB methods due to the reliability of normality assumptions under high area variance. Below this threshold, Non-EB methods perform better.
3. Non-Normal-EB Superiority: This new approach maintains the best performance across all scenarios, offering improvements over both Normal-EB and Non-EB methods without any drawbacks. The gains in precision (reductions in Root Mean Squared Error) range from marginal to substantial (up to 15–25%).
4. Implementation: The method uses finite normal mixtures for flexible modeling, making it easy to implement and suitable for real-world applications. Simulation studies confirm its practical value, regardless of skewness or kurtosis in errors.

The approach provides a robust solution for small area estimation, enhancing accuracy in poverty mapping and welfare analysis, particularly in data-constrained environments.