Assessment of robustness of Markov model to missing data in repeated measures studies.

摘要

The purpose of many longitudinal studies is the evaluation of a disease process over time. Transitional models are used to characterize disease processes. This research assesses the robustness of Markov models for analysis of incomplete longitudinal repeated measures data.;The primary objective is to assess the applicability and robustness of the Markov model to longitudinal data subject to intermittent missing and dropout data. A complete dataset containing daily drinking data from an alcohol treatment study is used to address the objective. Parameter estimates from the complete dataset are compared to estimates obtained from datasets containing various amounts of simulated missing data that are missing at random or nonrandom. For comparison, these datasets are applied to mixed effects models and results are compared among the datasets. For the Markov model, results show increased bias and variability in parameter estimates with increasing amounts of missingness. However, parameter estimates are similar between types of missing data mechanisms. In general, results suggests that amount of missing data, rather than type of missing data, has the greatest effect on the Markov model. In contrast, the mixed effects models show sensitivity to amount as well as type of missing data.;The application of the Markov model to a longitudinal dataset with nonrandom missing data is illustrated with a Markov model with states based on quality of life and relapse status for modeling quality adjusted survival given some missing data. Data from a study of time to relapse in breast cancer patients undergoing chemotherapy are evaluated. A patient's QOL is categorized based on a coping score that measures how well a patient is coping with her illness. Results show no difference in relapse rates between the two treatments, however, those undergoing more intense chemotherapy spend significantly longer duration in the poor QOL state.;In summary, the Markov model may be a useful tool for analyzing longitudinal data containing various types of missing data. Results from this research suggest that given the amount of missing data, analysis and estimation are not affected by the type of missing data.