bio-clinical-biostatistics-subgroup-analysis

What it does

Performs subgroup and heterogeneous treatment effect (HTE) analyses for clinical trials, covering Mantel-Haenszel pooling, Breslow-Day, interaction tests in regression, RERI for additive interaction, modern data-adaptive HTE methods (STEPP, SIDES, causal forests, X/R-learners), Bayesian shrinkage (Dixon-Simon, MAP, EXNEX), graphical multiplicity (Bretz-Maurer), and credibility frameworks (Sun BMJ, EMA 2019). Use when analyzing treatment effects across patient subgroups for regulatory submissions or precision-medicine claims.

How it works

The skill explicitly instructs step-by-step methods and examples:

• Mantel-Haenszel pooled odds ratio calculation across strata with CMH test; Breslow-Day test for homogeneity; cautions on power and interpretation. Includes Python snippet using StratifiedTable to compute pooled OR, confidence interval, and p-values for CMH and Breslow-Day.
• Interaction terms in a single regression model as the regulatory standard; provides code for logistic regression with interaction, LR test against an additive model, and extraction of subgroup-specific ORs with confidence intervals.
• RERI for additive interaction, describing the computation and interpretation in terms of ORs and the delta-method/bootstrap for CIs.
• Modern data-adaptive HTE methods: STEPP, SIDES/SIDEScreen, causal forests, Meta-learners (X/R-learner, MOB). It shows conceptual workflows and references, with Python/R code hints:
  • STEPP: sliding-window approach with permutation supremum tests; notes on correlation of estimates and Naive CI limitations.
  • SIDES/SIDEScreen: recursive partitioning with permutation-adjusted subgroup p-values; multiplicity absorption.
  • Causal forests: use grf in R or econml in Python; emphasis on honest splitting, ATE/CATE estimation, and calibration testing (test_calibration).
  • Meta-learners: S-, T-, X-, and R-learners with cross-fitting and propensity considerations; example code using econml XLearner or related classes.
• Bayesian shrinkage: describes Dixon-Simon 1991 prior for exchangeable interaction effects, EXNEX approach, and notes on borrowing across baskets with caution about prior influence.
• Diagnostic and regulatory context: references to credibility frameworks (Sun BMJ, EMA 2019) and considerations for pre-specified vs discovery subgroups, multiplicity adjustments, and reporting practices.

When to use it

Use when: analyzing treatment effects across subgroups, including regulatory submissions or precision-medicine claims; when planning subgroup analyses with pre-specified or data-discovered subgroups; when employing either classical stratified methods or modern HTE approaches and Bayesian borrowing.

What it can touch

Tooling is restricted to the primary tool: statsmodels for Python-based analyses and R packages (STEPP, SIDES, grf) as indicated; executable scripts ship with the skill folder and are meant for Python usage with statsmodels and related libraries as stated.

Caveats

Notes include power limitations for interaction detection, potential over- or under-estimation in heterogeneity, need to declare confirmatory vs exploratory intent, and sensitivity to prior choices in Bayesian shrinkage; also cautions about honest-splitting validation and calibration checks in causal forests. Licensing is NOASSERTION.