Reference Limits for Outlier Analyses in Randomized Clinical Trials.

Background: Reference limits used in clinical medicine to screen and manage patients are typically developed nonparametrically using reference values from a limited number of healthy subjects using a 95th percentile reference interval. We have evaluated alternative methods of computation and the resulting limits for use in the analyses of treatment-emergent outliers in clinical trials.

Methods: We developed a set of alternative reference limits for 38 laboratory analytes based on alternative statistical methods and assessed their relative performance in clinical trial analysis. Performance assessment was based on the clinical credibility of the limits, inferential statistical performance, consideration of incidences for the test drug and control (placebo) in cases where the drug was reasonably believed to be associated with a change in an analyte (positive cases), and in cases where prior analyses failed to demonstrate a change associated with the drug (negative cases).

Results: Based on consideration of these cases, no single method resulted in optimal limits for all cases considered. However, with the limits developed using clinical trial subjects' values at baseline as reference values, excluding outliers, the robust method and the 98th percentile interval appeared to produce optimal limits across the greatest number of cases considered.

Conclusion: Although no single method of limit computation will result in optimal limits for all outlier analyses for all analytes across all clinical trials, the 98th percentile reference interval robust limits based on clinical trial reference values appeared superior to multiple alternatives considered for such analyses.