On the effect of flexible adjustment of the p value significance threshold on the reproducibility of randomized clinical trials.

Abstract

Background: Reproducibility crisis is among major concerns of many scientists worldwide. Some researchers believe that the crisis is mostly attributed to the conventional p significance threshold value arbitrarily chosen to be 0.05 and propose to lower the cut-off to 0.005. Reducing the cut-off, although decreases the false-positive rate, is associated with an increase in false-negative rate. Recently, a flexible p significance threshold that minimizes the weighted sum of errors in statistical inference tests of hypothesis was proposed.

Methods: The current in silico study was conducted to compare the error rates under different conditions assumed for the p significance threshold-0.05, 0.005, and a flexible threshold. Using a Monte Carlo simulation, the false-positive rate (when the null hypothesis was true) and false-negative rate (when the alternative hypothesis was true) were calculated in a hypothetical randomized clinical trial.

Results: Increasing the study sample size was associated with a reduction in the false-negative rate, however, the false-positive rate occurred at a fixed value regardless of the sample size when fixed significance thresholds were used; the rate decreased, however, when the flexible threshold was employed. While employing the flexible threshold abolished the reproducibility crisis to a large extent, the method uncovered an inherent conflict in the frequentist statistical inference framework. Calculation of the flexible p significance threshold is only possible a posteriori, after the results are obtained. The threshold would thus be different even for replicas, which is in contradiction to the common sense.

Conclusions: It seems that relying on frequentist statistical inference and the p value is no longer a viable approach. Emphasis should be shifted toward alternative approaches for data analysis, Bayesian statistical methods, for example.