Direct marginal model analyses of longitudinal, ordinal outcomes with absorbing states
Jonathan Schildcrout, PhD Vanderbilt University Medical Center
Studies of critically ill, hospitalized patients often follow participants over a fixed follow-up period and characterize daily patient health with an ordinal scale. While, at first glance, longitudinal proportional odds models appear to be a natural choice in these setting, a challenge arises when one or more of the outcome states is absorbing (i.e., no one exits from the state once they enter it). In these settings, the proportional odds assumption for outcome associations with follow-up day are likely to be violated, and so partial proportional odds models may be required to ensure valid inferences. Motivated by the Violet 2 Study, a randomized clinical trial of Vitamin D on critically ill patients, we describe modeling approaches that capture intervention effects over time with a four-level ordinal outcome (not alive, on mechanical ventilator or with acute respiratory distress syndrome, in hospital, at home) where the lowest (not alive) state is absorbing and the highest (at home) state is nearly absorbing. For this purpose, we will discuss extensions of the proportional odds model for longitudinal data. We will focus on direct estimation of marginal mean model parameters using likelihood-based analysis procedures that can naturally handle absorbing states. We illustrate the utility of the modeling procedure through simulation studies and analysis of the Violet 2 study data.