The lay abstract featured today (for Bayesian transition models for ordinal longitudinal outcome by Maximilian D. Rohde,  Benjamin French,  Thomas G. Stewart,  Frank E. Harrell Jr.) is from Statistics in Medicine with the full Open Access article now available to read here.

Abstract

Many COVID-19 clinical trials collect ordinal longitudinal outcome data, which is ordered data collected at multiple points in time. As an example, the ACTT-1 clinical trial collected ordinal patient status every day for 28 days. The ordinal states included asymptomatic, at home with mild symptoms, at home with moderate symptoms, hospitalized (with or without ventilation), recovered, or deceased. However, many of these clinical trials only analyze part of the data, such as looking at time-to-recovery, rather than analyzing the full ordinal longitudinal data.

Ordinal longitudinal outcomes capture disease progression more fully than outcomes such as time-to-recovery and can also incorporate symptom severity, which is an important measure of treatment effectiveness. The longitudinal dimension of the data captures disease progression over time and can quantify when the treatment is most effective. Furthermore, ordinal longitudinal outcomes can accommodate terminal events such as death and recurrent events such as hospitalization.

This tutorial shows how Bayesian ordinal transition models are a flexible modeling framework for analyzing ordinal longitudinal outcomes. These models can estimate clinically interpretable estimands such as the treatment difference in the mean number of days recovered over the course of the trial. Ordinal transition models can also accommodate covariate adjustment to increase statistical power. The authors develop the theory from first principles and provide an application using data from ACTT-1 with code examples in R.