Nonignorable nonresponse models for longitudinal categorical data.

Abstract

In longitudinal studies the data for many cases may be incomplete even when the proportion of missing data at any time point is relatively small. When the cause for missing or incomplete data is related to the expected outcomes, the missing data mechanism is called nonignorable. Missing data also occur when subjects drop out of a study. Dropouts may no longer be interested in participating in the study due to their perceived evaluation of efficacy or because they have side effects that preclude their continued participation. Ignoring nonrandomly missing data or assuming an inappropriate missing data mechanism can bias the inference from models fitted to the data. In this dissertation we consider studies in which the primary outcome is categorical, data are collected from a single cohort at several time points, and the response for any subject may be missing at any time point except the first. We develop several nonignorable nonresponse models in which the mechanism for missing values may be nonignorable and the mechanism for dropouts may also be nonignorable. We also assume that the probability of the observation being "missing" depends on the current outcome level, previous outcome level, and whether the observation was missing at the previous time point; the odds ratio of "missing" at one level and "missing" at a base level is independent of time t. Three models are developed: (1) the first with only a single missing data mechanism; (2) the second with both a missing data mechanism and a dropout mechanism and (3) the last with two missing data mechanisms. Simulation is used to study the properties of these models. From the simulation studies, we learn that these models are appropriate for data sets in which the current outcome strongly depends on the previous outcomes. These models are fitted to data set from the Medical, Epidemiological and Social aspects of Aging (MESA) study. We compute the prevalence, incidence and remission of urinary incontinence across time intervals and compare the results to that in Brown et al. (1988). Since an ignorable missing data mechanism is appropriate for these data, the estimates of incidence, prevalence and remission rates from our model do not differ greatly from the estimates in Brown et al. (1988).