An Empirical Comparison of Statistical Methods for Estimating Treatment Effects on EQ-5D in Randomized Clinical...

This study aimed to empirically compare commonly used statistical models for estimating treatment effects using EuroQol 5-Dimension (EQ-5D) data from randomized clinical trials (RCTs).

We identified eligible RCTs through Vivli using National Clinical Trial numbers. Trials reporting EQ-5D at baseline and at least 2 follow-ups were included. Treatment effects on EQ-5D utilities or EQ Visual Analog Scale (VAS) were estimated as change from baseline both at the final visit and averaged across visits. We compared 4 commonly used models: linear mixed-effects model (LMM), generalized estimating equations (GEE), mixed-effects Tobit model, and mixed-effects Beta model, in terms of model diagnostics on distributional assumptions, agreements in statistical significance (P < .05), and clinical relevance using the minimally important difference.

Thirteen RCTs (n = 120-1841) were included. Baseline mean EQ-5D utilities ranged from 0.596 to 0.772, and EQ VAS scores from 44.5 to 75.5. For EQ-5D utilities, all 4 models agreed on statistical significance at the final visit. For average effects, 1 comparison in each model differed in statistical significance. LMM and GEE showed 95.7%-100% agreement in clinical relevance, whereas Tobit and Beta showed 3-5 disagreements. For EQ VAS, GEE and Tobit agreed on statistical significance in 95.7% of comparisons at the final visit; LMM showed 1 disagreement for average effects. All models agreed on clinical relevance, except for 1 LMM disagreement at the final visit.

Our analyses showed high agreement in statistical significance and clinical relevance across models. Considering model diagnostics, robustness, and practical usability, the LMM might be a reasonable and pragmatic option.