THE EFFECT OF DIFFERENT EFFECT SIZE ASSUMPTIONS ON SAMPLE SIZE AND POWER: AN R-BASED APPROACH FOR CLINICAL RESEARCH

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Introduction: In clinical research, sample size and power are usually calculated from closed-form analytic formulas based on a single assumed effect size. This study maps where analytic and Monte Carlo simulation power converge and diverge across clinical designs, and compares assurance (expected power) with classical nominal power. Materials and Methods: Six designs were examined: two-group mean difference, two proportions, one-way ANOVA, logistic regression, survival (log-rank/Cox), and diagnostic accuracy (ROC-AUC). For each, analytical power (pwr, power.prop.test, Hsieh, Schoenfeld, Hanley-McNeil) and simulation power (10,000 repetitions/cell) were compared under matched assumptions, together with assumption-violation scenarios for every design (unequal variance, skewed distribution, non-proportional hazards, non-binormal ROC, rare-event logistic regression, and heteroscedastic ANOVA). For each divergent scenario we additionally computed the correctly-specified analytical method (Welch; Lakatos-type average log-hazard-ratio; Welch ANOVA). The simulation engine's validity was verified via Type-I error calibration under H0, and the effect of optimistic effect-size specification on sample size was quantified. All powers are reported with Monte Carlo standard error (MCSE). Analyses used R 4.6.1 with per-cell deterministic seeds and sessionInfo. Results: Type-I error matched the nominal level in all primary tests (0.046–0.052; continuity-corrected proportion test conservative at 0.030), with two exceptions identified in this revision: heteroscedastic ANOVA was mildly liberal (0.068) and, under 2:1 allocation with unequal variances, the pooled t-test was markedly liberal (0.127). Where assumptions held, analytical and simulation power closely overlapped (e.g., two-group d=0.50: 0.801 vs 0.806; logistic OR=1.50: 0.896 vs 0.888; survival HR=0.70: 0.734 vs 0.731). Under assumption violation, divergence was marked: with unequal variance, simulation power was 34–64% below the naïve analytical estimate (d=0.50: 0.801 vs 0.315), and with a delayed treatment effect 79–83% below the value assuming proportional hazards (HR=0.60: 0.959 vs 0.181). In both cases the correctly-specified method reproduced the simulated power to within Monte Carlo error (Welch 0.312; Lakatos-type 0.187). Under effect-size uncertainty, assurance remained below nominal power, and the gap widened with uncertainty (prior SD=0.15: 0.801→0.740; SD=0.30: 0.801→0.674), a pattern reproduced in the survival and logistic designs. Overestimating the planning effect by 30% dropped a targeted 0.80 power to ~0.50. Conclusion: When assumptions are satisfied, analytical power calculation is sufficient and simulation-verifiable. When applied outside their stated assumptions, analytical formulas overestimate power in most of the violation scenarios examined — with the exception of the skewed-distribution case at small effect sizes, where the direction of the divergence reversed. The divergences reflect misapplication rather than a deficiency of analytical theory: in each case the correctly-specified formula recovered the simulated power. The study provides a reproducible R framework for practitioners. Keywords: Sample size; statistical power; effect size; Monte Carlo simulation; assurance