Parametric Estimation of Poverty in Data-Poor Countries

Technical issues with income data and heterogeneous statistical approaches for addressing them lead to discrepancies in poverty estimates across studies. This study assesses how alternative parametric models perform at estimating poverty headcount ratios under various degrees of data granularity, and various thresholds for poverty. We use 982 surveys from 57 countries and years 1963–2023, across most world regions, spanning low‑income conflict‑affected to high‑income contexts. The regimes of data availability include individual-level microdata, grouped data at the level of income deciles, and a pair of basic distributional statistics – mean and Gini. Our findings show that model flexibility enhances our ability to capture income distributions accurately, with three- and four-parameter models generally providing the closest poverty estimates. However, even simpler two-parameter models, such as lognormal and Fisk, sometimes perform well on grouped data or basic distributional data, for instance in high poverty-line, broad poverty, low-income, and low inequality settings. The analysis highlights that additional data or higher model complexity does not always improve poverty estimation, and in some cases, grouped data can yield more reliable point estimates than raw microdata. Higher-parametric models do not always outperform parsimonious models, particularly when it comes to the precision of estimates in limited-data environments. GB2 and beta 2 estimates exhibit inflated standard errors when estimated on grouped data, while parsimonious models – e.g., Dagum and Singh–Maddala – are often more balanced. These results offer practical guidance to practitioners for selecting appropriate models according to data availability, balancing model complexity, and ensuring robust poverty measurement.