Stress-Testing Meta-Analyses: A Reproducible Python Framework for Robustness Assessment of Continuous Outcomes

Meta-analyses of continuous outcomes often summarize evidence as a single pooled effect estimate, but the stability of that estimate can depend on the effect-size calculation, heterogeneity estimator, confidence-interval method, and influence of individual studies. This technical report presents a reproducible Python framework for stress-testing pairwise meta-analyses of continuous outcomes from independent two-group comparisons. The workflow accepts study-level means, standard deviations, and sample sizes; validates the input; calculates Hedges g and its approximate sampling variance; fits fixed-effect and random-effects models; compares DerSimonian-Laird and Paule-Mandel estimates of between-study variance; calculates a prediction interval independently of the pooled-effect confidence interval (CI); applies a conservative Hartung-Knapp-Sidik-Jonkman (HKSJ)-style confidence-interval adjustment; performs leave-one-out influence analysis; identifies the study producing the largest absolute change in the pooled estimate; and performs an Egger-type regression for funnel-plot asymmetry when at least 10 studies are available. The demonstration uses a synthetic dataset of 12 independent studies. The primary random-effects estimate was Hedges g = -0.249 (95% CI, -0.455 to -0.043; tau-squared = 0.091), with a 95% prediction interval of -0.961 to 0.463. The HKSJ-style CI was wider (-0.494 to -0.003), while the Paule-Mandel model produced a similar pooled estimate (Hedges g = -0.248; 95% CI, -0.467 to -0.028). All leave-one-out estimates retained a negative direction, and Study 07 produced the largest absolute change (0.074 Hedges g). The Egger-type intercept test was not statistically significant (p = 0.352). The framework is intended to make statistical fragility explicit and reproducible while complementing, rather than replacing, established meta-analysis software.