Assessing Reproducibility of Novometric Bootstrap Confidence Interval Analysis Using Multiple Seed Numbers (Invited)

Nathaniel J. Rhodes

Chicago College of Pharmacy, and the Pharmacometrics Center of Excellence, Midwestern University

I study the role of the random seed number in affecting the reliability of a statistical finding, which in turn determines upper and lower bounds of statistical confidence in expected gain or loss yielded from associated decision-making. Simulation research reveals that obtaining a bootstrap solution consistent with an effect (e.g., the lower exact, discrete 2.5th percentile bound of the model bootstrap exceeds the upper exact, discrete 97.5th percentile bound of the chance bootstrap)—more than once using any two independent seeds—supports the hypothesis that the effect is not attributable to random chance because it was replicated vis-à-vis an independent seed. Based upon this finding I suggest consistent use of a primary seed and confirmation using a secondary seed when undertaking model qualification via simulation. This procedure reduces doubt about the veracity of borderline statistically significant findings, and eliminates false positive results identified by replication failure.

View journal article