Note:

Background: Response bias can distort treatment effect estimates and inferences in clinical trials. Although prevention, quantification, and adjustments have been developed, current methods are not applicable when subject-level reliability is used as the measure of response bias. Thus, the objective of the current study is to develop, test, and recommend a series of bias correction strategies for use in these cases. Methods: Monte Carlo simulation and logistic regression modeling were used to develop the strategies, examining the collective impact of sample size (N), effect size (ES), reliability distribution, and response style on estimating the treatment effect size in a series of hypothetical clinical trials. The strategies included a linear (LW), quadratic (QW), or cubic weight (CW) applied to the subject-level reliability; a reliability threshold (%); or a combination of the two (W-%). Bias and percent relative root mean square error (RRMSE (%)) were calculated for each treatment effect estimate and RRMSE (%) was compared to inform the bias correction recommendations. Results: The following recommendations are made for each N and ES combination: N=200/ES=small: no adjustment, N=200/ES=medium: 40%-LW, N=200/ES=large: 40%-QW, N=2000/ES=small: 40%-LW, N=2000/ES=medium: 55%-CW, N=2000/ES=large: 75%-CW, N=20000/ES=small: 70%-CW, N=20000/ES=medium: 85%-CW, N=20000/ES=large: 95%-CW. Conclusion: Employing these bias correction strategies in clinical trials where subject-level reliability can be calculated will decrease error and increase accuracy of estimates and validity of inferences.