OBJECTIVES

There is a growing literature on options for estimating a treatment effect for PROs, where an event censors the measure of the effect of interest. The censoring may stem from an event such as death that makes further collection impossible; censoring may also impinge directly on what it is desired to estimate, and may be required to be taken account of quantitatively. The objective of this project was to evaluate approaches to analysing PRO data in the presence of censoring due to death and other events.

METHODS

A targeted literature review was performed, with advice from subject matter experts (both statistical and PRO experts). Pubmed and Google were used to identify approaches that account for censoring of PRO data. Methods, and their attributes were abstracted and summarised. The approaches identified were further discussed with experts and a short list was prepared that was evaluated with respect to the assumptions required by each method, their limitations, ease of implementation and interpretability.

RESULTS

We reviewed 26 publications and identified 8 approaches: win ratio for a ranked composite endpoint; continuation ratio modelling for a categorized composite endpoint; principal stratification with Rubin’s implementation; responder (binary) analysis; mixed model repeated measures (MMRM) with zero PRO imputation for death; area under the curve (AUC) as a patient-level summary measure; ordinal regression (proportional odds) modelling and joint modelling of survival and longitudinal PRO values. A short list of 4 approaches was assessed in depth and these methods were implemented using two publicly available clinical data sets as test cases.

CONCLUSIONS

This review identified several analysis methods (estimators) for three different types of estimands for inference about PROs when an event impinges on the measurement of the PRO.