Kruskal-Wallis Power Studies Utilizing Bernstein Distributions; preliminary empirical studies using simulations/medical studies

TL;DR

This study evaluates the effectiveness of Bernstein distribution-based Monte-Carlo Kruskal-Wallis power studies, showing they often outperform ANOVA methods in accuracy, especially when ANOVA assumptions are violated, across simulated and medical datasets.

Contribution

Introduces Bernstein distribution-based Monte-Carlo Kruskal-Wallis power studies and compares their performance with traditional methods using simulations and real medical data.

Findings

Kruskal-Wallis predictions were more accurate than ANOVA in most simulated runs.

Monte-Carlo Kruskal-Wallis provided more reliable power estimates when ANOVA assumptions failed.

Analytical methods were less accurate than Monte-Carlo approaches.

Abstract

Bernstein fits implemented into R allow another route for Kruskal-Wallis power-study tool development. Monte-Carlo Kruskal-Wallis power studies were compared with measured power, with Monte-Carlo ANOVA equivalent and with an analytical method, with or without normalization, using four simulated runs each with 60-100 populations (each population with N=30000 from a set of Pearson-type ranges): random selection gave 6300 samples analysed for predictive power. Three medical-study datasets (Dialysis/systolic blood pressure; Diabetes/sleep-hours; Marital-status/high-density-lipoprotein cholesterol) were also analysed. In three from four simulated runs (run_one, run_one_relaxed, and run_three) with Pearson types pooled, Monte-Carlo Kruskal-Wallis gave predicted sample sizes significantly slightly lower than measured but more accurate than with ANOVA methods; the latter gave high sample-size predictions. Populations (run_one_relaxed) with ANOVA assumptions invalid gave Kruskal-Wallis predictions similar to those measured. In two from three medical studies, Kruskal-Wallis predictions (Dialysis: similar predictions; Marital: higher than measured) were more accurate than ANOVA (both higher than measured) but in one (Diabetes) the reverse was found (Kruskal-Wallis: lower; Monte-Carlo ANOVA: similar to measured). These preliminary studies appear to show that Monte-Carlo Kruskal-Wallis power studies based on Bernstein fits might perform better than ANOVA equivalents in many settings (and provide reasonable results when ANOVA cannot be used); and both Monte-Carlo methods appeared considerably more accurate than the analysed analytical version.