OBJECTIVES:

Trial planning, recruitment, and subsequent results interpretation can be improved through consideration of patient characteristics that may modify treatment effect. Current approaches typically rely on patient observations collected at a single timepoint, but longitudinal approaches may afford more statistical power and/or the ability to consider full trajectories of response.

METHODS:

Using simulated data we model treatment effects of a continuous variable across time with a mixed model for repeated measures and an interaction effect between treatment, characteristic, and a flexible spline on time. This allows treatment effect to vary arbitrarily across time. A second approach is developed that first clusters patient characteristics using PCA before fitting an interaction of treatment X time X (component 1 + component 2).

RESULTS:

Predicted treatment effects are plotted for each patient and the subsequent output is analyzed using methods for clustering of longitudinal data. The resulting clusters are presented alongside a table of patient characteristics.

CONCLUSIONS:

Longitudinal clustering of treatment effects allows for focus on identification of patient characteristics that may be associated with desirable/undesirable longitudinal patterns of response. This method could be used to aid trial planning for enrichment purposes and to assess potential sensitivity of indirect treatment comparison results to differences in patient characteristics alongside differing durations of follow-up.