A critical evaluation of longitudinal proportional effect models

TL;DR

This paper critically evaluates nonlinear longitudinal proportional effect models, highlighting their biases and limitations in clinical trial analysis, especially when assumptions are violated or labels are misassigned.

Contribution

It provides a comprehensive assessment of the biases and limitations of existing proportional effect models in longitudinal clinical trial data.

Findings

Models assume fixed effects, leading to bias when violated.

Bias favors active treatment, inflating Type I error.

Bias complicates detection of treatment harm.

Abstract

Nonlinear longitudinal proportional effect models have been proposed to improve power and provide direct estimates of the proportional treatment effect in randomized clinical trials. These models assume a fixed proportional treatment effect over time, which can lead to bias and Type I error inflation when the assumption is violated. Even when the proportional effect assumption holds, these models are biased, and their inference is sensitive to the labeling of treatment groups. Typically, this bias favors the active group, inflates Type I error, and can result in one-sided testing. Conversely, the bias can make it more difficult to detect treatment harm, creating a safety concern.