Outlier Detection for Welfare Analysis

Abstract

Extreme values are common in survey data, and represent a recurring threat to the reliability of both poverty and inequality estimates. The adoption of a consistent criterion for outlier detection is useful in many practical applications, particularly when international and intertemporal comparisons are involved. This paper discusses a simple, univariate detection procedure to flag outliers in the distribution of any variable of interest. It presents outdetect, a Stata command that implements the procedure and provides useful diagnostic tools. The output of outdetect compares statistics – with a focus on inequality and poverty measures – obtained before and after the exclusion of outliers. Finally, the paper carries out an extensive sensitivity exercise, where the same outlier detection method is applied consistently to per capita expenditure across more than 30 household budget surveys. The results are clear-cut and provide a sense of the influence of extreme values on poverty and inequality estimates.

Publication
Work in progress (first draft available soon)