Multistate Markov chain transition models for clustered longitudinal categorical data: application to a knee pain severity study

Abstract

In longitudinal biomedical research, outcome data are often collected on a categorical scale from multiple locations of the same individual over time. For example, an important feature of knee osteoarthritis (OA) is that severity of knee pain often fluctuates over time. Identifying risk factors that contribute to the fluctuation of knee pain severity can help elucidating the pathology of the disease hence is of great clinical and epidemiological value. Measurements of knee pain are usually knee-specific. This challenges data analysis since an approach needs to model the longitudinal transitions of the outcome and to account for correlation between a person's two knees at the same time. Here I propose a multistate Markov chain transition model with extensions that acconnt for longitudinal and within-cluster correlations. The model assumes discrete time and allows transitions over time between any two states of pain severity. The model is based on a generalized linear regression framework assuming a multinomial distribution for the outcome. First, marginal model based approaches were proposed. These approaches take advantage of the robust sandwich vanance estimator and within-cluster resampling techniques to account for correlation within each cluster. Then, Bayesian random effect model based approaches were employed. Correlations among observations within each cluster are accounted for by including random effects into the model. Both the marginal model based approaches and the random effect models based approaches were evaluated by simulation studies. The models were then used to assess the effect of depression on transitions of knee pain severity in the Osteoarthritis Initiative study. A proportional odds model was further developed to deal with situations where the outcomes are ordinal. The models proposed in this dissertation extend the existing literatures in handling multilevel correlation when estimating effects of risk factors on longitudinal transitions of a categorical outcome. They can be especially helpful for investigation of disease progression or for prognostic purpose.