Introduction

This chapter proposes a new approach to tackling attrition problems in panel studies. Panel data are obtained by observing cross-sectional units such as individuals or households more than once over an extended period of time. Panel data sets are typically unbalanced due to attrition. In the first year of a panel survey, researchers pay full attention to making the sample as representative as possible. Therefore, the first year observations can be regarded as forming a random sample, able to represent the underlying population. However, over time, initial cross-sectional units are dissipated, possibly in a nonrandom fashion. Attrition may negate the initial randomization. If the variable of interest is stochastically related to the factors governing attrition process, then the attrition is nonrandom and potentially causes bias. In this sense, attrition bias can be regarded as a sample selection bias. A major difference between the attrition bias and the typical sample selection bias is that the attrition occurs through a multistage process.

Fuller and Battese (1974), Biørn (1981), Baltagi (1985), and Wansbeek and Kapteyn (1989) have discussed incomplete panel data models, with data unbalanced due to attrition. However, they all assumed that the attrition process is exogenous, and thus assumed away the possible attrition bias.

Attrition bias in panel studies was first considered by Hausman and Wise (1979) using a random effects model for the individual response. The existence of attrition bias was postulated as due to nonrandomness of the remaining crosssectional units after the initial survey years. In other words, attrition bias is simply a sample selection bias.