Panel data often provide an understanding of household behavior not possible with cross-sectional information alone. However, a disturbing feature of such data is that there can be substantial, nonrandom attrition and many analysts share the
concern that this inhibits the ability to make accurate inferences. The author examines attrition in the KwaZulu–Natal Income Dynamics Study 1993–1998, assesses the extent of attrition bias for a specific empirical example, and proposes and implements a selection correction methodology using quality of first round interview variables as identifying instruments. The results show that attrition does lead to statistical bias in the “behavioral” coefficients in estimation of household
-level expenditure functions. Since it is typically difficult to determine the bias for a particular analysis a priori, and such bias is by its nature model-specific, it behooves researchers using panel data to evaluate the effects of attrition
in their analyses.