Assessing the Effect of Panel Attrition on Log-linear Model Estimation

Abstract

Attrition is a common problem for analysts of panel data. In this paper, we assess the impact of panel attrition on the association between categorical or cross-classified data via a log-linear model. We devise a simulation study in which the effects of the three common missingness mechanisms – missing completely at random, missing at random, and missing not at random – are assessed for the three different sample sizes of 200, 500, and 1,000, allowing for varying attrition rates from 10 to 50 percent. Increase in attrition rates are found to decrease efficiency of estimation while sample size tends to improve efficiency. When missing data mechanisms are not known, missing-at-random models are preferred to the other missingness assumptions in terms of model fitting, unbiased estimation, and efficiency. An analysis of real-world data further supports this conclusion.