How to Analyze Longitudinal Patient-Reported Outcomes in Populations With High Mortality Rates

Analyzing longitudinal patient-reported outcomes (PROs) remains methodologically challenging, particularly in populations with high mortality rates during follow-up. Many PRO studies overlook this issue, potentially leading to biased estimates and limited clinical interpretability. We discuss various methods for analyzing patients who died during follow-up in PRO analyses and examine how these methods affect the interpretation of the study outcomes, using a clinical example in older patients with cancer.

For our clinical example, we included adults aged $\ge$  70 years with cancer from a prospective cohort. Quality of life (QoL) was assessed with the EuroQoL 5-Dimension (EQ-5D) tool at baseline and 6 and 12 months after treatment initiation. We applied six different methods: Four methods calculated absolute changes in mean scores; two methods estimated the proportion of patients with a changed QoL.

Of the 1218 participants, 321 (26%) died within 12 months. Methods restricted to the 897 survivors at the end of follow-up lead to results that are difficult to interpret due to selection bias. In linear mixed models, EQ-5D scores are implicitly reconstructed for the 321 participants that died, creating a hypothetical view of QoL assuming no patients died. Alternative approaches, such as the "while-alive" strategy or composite endpoint strategies, treating death as the worst possible score on the EQ-5D, produce more complete views of the QoL.

The interpretation of longitudinal PROs strongly depends on the chosen method, associated estimands and how patients that died are addressed: different methods lead to different estimates and conclusions. Properly defining the research question and choosing the most appropriate method is essential for meaningful interpretations for clinicians and patients.