A comparison of five statistical methods used to analyze longitudinal EORTC QLQ-C30 quality of life scores in...

Quality of life (QoL) is an important component of cancer care. The European Organization for Research and Treatment of Cancer Quality of Life Questionnaire Core 30 (EORTC QLQ-C30) is widely used to evaluate patient-reported QoL over time. A scoping review found a wide range of statistical methods are used to analyze longitudinal QLQ-C30 scores in randomized controlled trials (RCTs), and many of the methods ignore the longitudinal nature of the data, potentially impacting the conclusions drawn. We planned a simulation study to evaluate the performance of five recommended and/or commonly applied methods for analyzing longitudinal QLQ-C30 scores.

Using data from two RCTs to inform the design, we simulated three hypothetical parallel-group RCTs assessing QoL using the EORTC QLQ-C30 global health status score measured at baseline and four postrandomization time points. We varied the sample sizes, follow-up patterns, treatment effects, death rates, hazard ratios between groups, levels of missing data, and covariance structures. The estimand of interest was the difference between groups at each time point and we compared the t-test, linear mixed effects model (LMM), generalized estimating equations (GEEs), joint longitudinal survival model (JLSM), and repeated measures analysis of variance (RMA) considering performance measures: bias, coverage, empirical and model standard errors, mean squared error, and power.

The JLSM provided the least biased estimates of treatment effect with other performance measures similar for the JLSM, LMM, and GEE across all scenarios. In all scenarios the t-test and RMA had considerably higher mean squared error, empirical and model standard errors than other methods.

The JLSM is the recommended method, as it performed at least as well as the LMM and GEE, however, its sensitivity to misspecification of time may be of concern and the LMM or GEE with time fitted categorically may be preferable when researchers are uncertain about the best fitting parametrization of time. We do not recommend the t-test or RMA, as even in scenarios where bias was comparable to that of the JLSM, GEE, and LMM, performance was poor for the other performance measures.

Quality of life in cancer clinical trials is often measured several times for each patient to understand how their quality of life changes over their time in the study. In practice, many different statistical methods are used to compare patient's quality of life between the treatment groups being investigated. We carried out a simulation study to see how well five different statistical methods performed when estimating the difference in average quality of life score between two treatment groups. We investigated whether factors such as the number of patients in the study, the amount of missing data, the number of deaths in each group, the true difference between the groups, and the timing of follow-up measurements affected how the methods performed. We found that statistical methods that considered the relationship between the quality of life scores from the same patient gave more reliable results than methods that ignored this information. The joint longitudinal survival model was the best model overall. The t-test and repeated measures analysis of variance were frequently used in practice but led to unreliable results.