Extensions of the Dynamic Propensity Score Adjustment for Longitudinal Data

Dynamic propensity adjustment models are extensions of the propensity score adjustment strategy to longitudinal observational study data, in which observations on the same subjects are gathered repeatedly over time. If the treatment or exposure is ordinal, random intercept ordinal regression models are used to model the propensity for ordinal treatment with measured covariates, both baseline and time-varying, as predictors in the model. Level-1 observations are stratified based on quintiles of their estimated dynamic propensity score, and mixed effects regression models are fit within each propensity stratum, with treatment and measured covariates as predictors for the outcome of interest. In the absence of significant propensity quintile-by-treatment interaction, a common treatment effect is estimated using a weighted, pooled estimator. In this dissertation we describe several extensions of the dynamic propensity score methodology, including a method which employs a model that is robust to misspecification of the random effects distribution. We develop this model and analyze its performance using a Monte Carlo simulation study. We find that the dynamic propensity adjustment is robust to violations of random effect distributional assumptions, and its performance is not improved by using more flexible, but computationally demanding methods of modeling the dynamic propensity score. In an effort to understand these results, we conduct a simulation study to assess the degree to which ordinal random effects models are robust to misspecification of the random effects distribution. In addition, we develop methods that use the dynamic propensity score as an index for matching, weighting, and direct adjustment and evaluate each methods performance using simulation studies. We find that these methods perform well when the effect size of confounders is modest, but no better than the adjustment method based on stratification. When the effect size is large, stratification outperforms these methods. Finally, we analyze data from The National Institute of Mental Health Collaborative Depression Study, a longitudinal observational study of the treatment of major affective disorders, using stratification on the estimated dynamic propensity score.