Answers are grounded in this paper and cite their sources.
Small community physical-activity evaluations may combine unequal variances, unbalanced allocation, and selective follow-up, obscuring the distinction between uncertainty distortion and selection bias. We compared five analysis strategies using synthetic randomized pre-post data: follow-up and change-score Welch tests, conventional and HC3-corrected ANCOVA, and multiple imputation with HC3 inference. The primary scenarios fixed total sample size at 60 while varying allocation, baseline slope, residual heterogeneity, treatment effect, and follow-up mechanism. Across 184 scenarios, 884,000 datasets were generated. Under baseline-dependent missing at random (MAR), conventional ANCOVA null coverage ranged from 86.30% to 97.26%, compared with 94.44–96.00% for HC3. Unadjusted contrasts exhibited selection bias. One diagnostic missing-not-at-random (MNAR) condition biased both HC3 and the evaluated imputation procedure. Imputation findings apply to arm-specific normal regressions with 20 imputations followed by common-slope ANCOVA. Within these linear, independent-observation settings, robust inference addressed variance-estimation problems but did not repair selection on unobserved outcomes. Analysis planning should align the estimand, design, outcome model, and missingness assumptions.
No takes yet. Share an insight, caveat, or question.
Mendieta et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: