The impact of measurement error on the comparison of two treatments using a responder analysis

Investigators often report results of studies comparing the proportions of subjects who have clinically meaningful responses to various therapeutic regimens. When the outcome variable is a continuous measure this involves dichotomizing the observed response based on a predefined threshold value. The effect of this strategy is examined, paying particular attention to the impact of measurement error on the resulting estimates of treatment effect (difference in the proportion responding). Expressions are obtained for quantifying the magnitude and direction of the resulting bias, and these are illustrated in a study evaluating a pharmaceutical treatment for osteoporosis.