Jayson D. Wilbur (1), Manish Gupta (2) Chaitali Passey (2), and Amit Roy (2)
(1) Metrum Research Group, Tariffville CT, (2) Bristol-Myers Squibb, Lawrenceville NJ
Objectives: Accurate characterization of exposure-response (E-R) relationships can be challenging in the presence of confounding factors that affect both pharmacokinetic (PK) properties as well as response. In such situations, virtual randomization using case-matching of treatment arm subjects has been proposed to select control arm subjects for inclusion in the E-R analysis [1]. We present two approaches to evaluate the effectiveness of the virtual randomization by case-matching with respect to PK properties (that are not observable in control arm subjects).
Methods: The proposed case-matching evaluation methods are illustrated for a 2-arm clinical trial of drug (treatment) vs placebo (control), in which treatment arm subjects with exposures in the lowest quartile are matched to control arm subjects by propensity score matching. The effectiveness of the matching with respect to drug clearance (CL) is assessed by: (1) Holding out half the subjects in the treatment arm and attempting to match within the treatment arm and (2) Reverse matching the identified control subjects back to the treatment arm; and comparing exposure to what would be expected. . The validity of these methods were assessed for several simulated scenarios with varying sample sizes (N=100,200,500), covariates (p=5,10,20), and correlation among covariates and exposure (r=0.0,…,0.99).
Results: The effectiveness of case-matching improved with increasing correlation among exposure and matched covariates, and with sample size; and both evaluation methods were useful for assessing the effectiveness of the case-matching. Specifically, the percentage of held-out or reverse-matched treated subjects with exposure in the lowest quartile was predictive of the percentage of matched controls expected to have exposure in the lowest quartile. Likewise, the standardized difference in mean exposure between subjects in the lowest quartile and the held-out subjects was predictive of the standardized difference in mean exposure expected with matched controls.
Conclusions: It is recommended that effectiveness of case-matching be evaluated prior to performing exposure-response analysis on non-randomized subjects, to ensure that the matching results in balanced distributions of observed and unobserved factors that may affect both exposure and response.
References:
[1] Yang, J. et al., Combination of Exposure-Response and Case-Control Analyses in Regulatory Decision Making. JCP, 53: 160–166 (2013).
Reference: PAGE 25 (2016) Abstr 5982 [www.page-meeting.org/?abstract=5982]
Poster: Methodology - Model Evaluation