Modeling nonresponse in multiwave panel studies using discrete-time Markov models

Nonresponse is of major concern to social scientists, due to the possibility of selectivity: not all groups in the population are equally represented in the final sample, when some groups have a larger probability to be in the sample than others. It is dangerous to base conclusions on such biased samples. Therefore, it is of importance to study nonresponse patterns. First it is shown that a decreasing nonresponse for every successive wave indicates that nonresponse is selective to a degree. Successively we discuss how Markov models can be used to get some idea of the seriousness of this bias in the sample, by examining how many chains are needed to reproduce the observed pattern of nonresponse acceptably well, and what the probability is that members of these chains will participate in a particular wave of the study. A small application is given, after which the implications of the findings are discussed.