bio-clinical-biostatistics-effect-measures

What it does

Computes and interprets treatment effect sizes (OR, RR, RD, HR, NNT) with confidence intervals calibrated to sample size, and reports whether the estimand is marginal or conditional according to FDA 2023 covariate adjustment guidance. It is intended for use in reporting treatment effects in confirmatory trials, comparing effect sizes across studies, or constructing forest plots.

How it works

The skill presents an algorithmic taxonomy and decision-tree guidance for selecting estimands and corresponding CI methods across scenarios (binary outcomes in RCTs, case-control settings, time-to-event data, noninferiority on different scales, stratified designs). It prescribes specific CI methods for each measure (e.g., MN score, Newcombe-Wilson, profile likelihood, Wald, MOVER) and includes practical code patterns for:

• Unadjusted 2x2 tables to compute OR, RR with appropriate CIs and calibrations.
• RD confidence intervals via Newcombe-Wilson hybrid or MN score for regulatory contexts.
• Absolute and relative measures (RD, RR, OR) with recommended methods depending on design and prevalence.
• NNT computation using RD CI and handling of RD=0 with Bender 2001 methodology, including NNTB/NNTH reporting when CI crosses zero.
• Marginal RD via g-computation (FDA 2023 recipe) using logistic regression with covariates, predicting under treatment and control, averaging predictions, and obtaining SE via bootstrap or influence-function approaches. R equivalents via marginaleffects are provided.
• Distinguishing marginal versus conditional estimands in logistic models, with examples showing how to report marginal RD as primary and conditional OR as supportive.

The skill also includes concrete code snippets (Python and R) demonstrating how to implement these calculations, including: fit/logistic models, counterfactual predictions, standardisation to obtain marginal effects, and bootstrapSE calculations.

When to use it

Use when reporting treatment effects in confirmatory trials, when comparing effect sizes across studies, or when constructing forest plots that require calibrated CIs for OR, RR, RD, HR, or NNT, and when FDA 2023 covariate-adjustment guidance applies (report marginal estimands for primary endpoints and conditional estimands as supportive).

What it can touch

• Python tools: statsmodels (for contingency tables, logit models, Wald/score-type CIs), numpy.
• R tools: marginaleffects (for marginal RD with HC3), base glm/logit for conditional estimates.
• It references methods like Miettinen-Nurminen, Newcombe-Wilson, MOVER, Bender NNT, profile likelihood, and g-computation via standardisation.

Caveats

• License: NOASSERTION.
• Requires correct version compatibility (e.g., statsmodels 0.14+, numpy 1.26+, pandas 2.1+, etc.).
• The conditional vs marginal estimand distinction is emphasized; marginals are primary per FDA 2023 guidance for binary endpoints, with conditional estimates as supportive.
• The methods assume appropriate model specification for marginal RD via g-computation; bias can arise if models are misspecified (robustness note included).