Aims To evaluate the stability of anchor-based responder thresholds.
Methods Change scores of a generic 100-point PRO measure and scores of a 7-point patient-reported global impression of change (PGIC; a retrospective anchor) were simulated based on the Spearman’s rank correlation between the change in PRO and anchor measures using the Iman-Conover method. Simulations were designed to mimic common characteristics of clinical trial data (sample size: n = 50, 100, 200; proportion of patients improving on anchor measure: 30%, 50%, 70%; distribution of PRO change scores: mean = 30, standard deviation (SD) = 10, 25, 50; and Spearman correlations between PRO change scores and anchor measure score: r = 0.3, 0.5, 0.7), yielding 81 permutations. PRO responder definitions were estimated using common methods: mean or median PRO change per anchor level and logistic regression predicting anchor level from the change in PRO. Using 1000 replications for each of the 81 permutations, the variability of the responder threshold estimates were compared across the three anchor-based methods. To provide insight on the most influential clinical trial data characteristics driving the stability of anchor-based responder definitions, a general linear model was fit to identify the characteristics contributing to stability (semipartial ω2).
Results The mean change per anchor level method yielded the most stable estimates compared to the median and logistic regression methods. The most important aspects driving the stability of the responder definitions were the variance of the PRO change variable and sample size.
Conclusions Researchers who design clinical trials and hope to develop stable estimates of clinically meaningful change from the patient perspective should prioritize study designs that include sufficient sample size and incorporate the anchor-based mean change method.
