Issues in the determination of 'responders' and 'non-responders' in physiological research

What is the topic for this review? We discuss the dichotomization of continuous-level physiological measurements into 'responders' and 'non-responders' when interventions/treatments are examined in robust parallel-group studies. What advances does it highlight? Sample responder counts are biased by pre-to-post within-subject variability. Sample differences in counts may be explained wholly by differences in mean response, even without individual response heterogeneity and even if test-retest measurement error informs the choice of response threshold. A less biased and more informative approach uses the SD of individual responses to estimate the chance a new person from the population of interest will be a responder.

As a follow-up to our 2015 review, we cover more issues on the topic of 'response heterogeneity', which we define as clinically important individual differences in the physiological responses to the same treatment/intervention that cannot be attributed to random within-subject variability. We highlight various pitfalls with the common practice of counting the number of 'responders', 'non-responders' and 'adverse responders' in samples that have been given certain treatments or interventions for research purposes. We focus on the classical parallel-group randomized controlled trial and assume typical good practice in trial design. We show that sample responder counts are biased because individuals differ in terms of pre-to-post within-subject random variability in the study outcome(s) and not necessarily treatment response. Ironically, sample differences in responder counts may be explained wholly by sample differences in mean response, even if there is no response heterogeneity at all. Sample comparisons of responder counts also have relatively low statistical precision. These problems do not depend on how the response threshold has been selected, e.g. on the basis of a measurement error statistic, and are not rectified fully by the use of confidence intervals for individual responses in the sample. The dichotomization of individual responses in a research sample is fraught with pitfalls. Less biased approaches for estimating the proportion of responders in a population of interest are now available. Importantly, these approaches are based on the SD for true individual responses, directly incorporating information from the control group.