Treatment Effect Measures - Usage Guide

Overview

Computes and interprets treatment effect measures (OR, RR, RD, HR, NNT) for clinical trial data with calibrated confidence intervals (Wilson, Newcombe, Miettinen-Nurminen, MOVER, profile likelihood, Bender NNT) and explicit declaration of whether the estimand is marginal or conditional under ICH E9(R1) and FDA 2023 covariate adjustment guidance.

Prerequisites

pip install statsmodels numpy pandas matplotlib

R is recommended for production:

install.packages(c('ratesci', 'exact2x2', 'marginaleffects', 'RobinCar', 'riskCommunicator'))

Quick Start

Tell your AI agent what you want to do: - "Compute OR and RR from my 2x2 with Wald, profile-likelihood, and Miettinen-Nurminen score CIs" - "Extract marginal RD via g-computation from a covariate-adjusted logistic regression per FDA 2023" - "Calculate NNT with Bender 2001 convention (NNTB-infinity-NNTH when RD CI crosses zero)" - "Forest plot of subgroup ORs on log scale with reference line" - "Modified Poisson with HC1/HC3 sandwich SE to estimate RR directly when prevalence > 10%"

Example Prompts

Crude effect measures

"I have a 2x2: 45 treated responders, 55 treated non-responders, 30 control responders, 70 control non-responders. Compute OR, RR, RD with calibrated CIs (Wald + Newcombe-Wilson hybrid)."

"Use Miettinen-Nurminen score CI for the RD as the regulatory standard for NI margin assessment."

Marginal vs conditional (FDA 2023)

"Per FDA 2023 covariate adjustment guidance, my primary estimand should be marginal RD. Fit logistic adjusted for age + sex + baseline severity, then compute marginal RD via g-computation with HC3 sandwich SE. Report conditional OR as supportive."

"My adjusted and unadjusted ORs differ -- is it confounding or non-collapsibility? Cite Permutt 2020."

NNT with proper convention

"Compute NNT from my trial with treated 30/200 events, control 50/200. Use Bender 2001 method; if RD CI crosses zero, report NNTB-infinity-NNTH convention."

"Compare NNT at baseline risks 5%, 20%, 50% for an OR of 0.5 to show stakeholders how benefit varies."

Modified Poisson for RR

"Outcome prevalence is 25% -- OR will overstate RR. Use modified Poisson regression with HC1 sandwich SE to estimate RR directly (Zou 2004)."

Forest plots and visualisation

"Create a forest plot of subgroup ORs (sex, age <65/>=65, race, severity, region) on log scale with reference line at 1.0."

Hauck-Donner pathology

"Small cell counts; Wald p is non-significant but coefficient is large. Switch to profile-likelihood CI (R MASS::confint.glm); cite Yee 2022."

What the Agent Will Do

1. Construct the 2x2 table with proper column ordering (event first per Table2x2)
2. Compute OR, RR, RD using calibrated CI methods (Wilson, Newcombe-Wilson, MN, profile likelihood)
3. For binary outcomes, fit logistic regression and compute BOTH conditional OR AND marginal RD via g-computation per FDA 2023
4. Apply modified Poisson with sandwich SE when prevalence > 10% and RR is the policy quantity
5. Convert effect to NNT with Bender 2001 disjoint-interval convention when RD CI crosses zero
6. Generate forest plots on log scale for OR/RR/HR; linear for RD/RMST

Tips

• OR is non-collapsible. Adjusted vs unadjusted ORs differ even without confounding (Permutt 2020). Don't interpret the difference as confounding.
• Per FDA 2023, marginal RD via g-computation is the primary estimand for binary endpoints with covariate adjustment. Conditional OR is a different parameter, not a different estimate.
• Wald CI has well-documented coverage failures (Brown-Cai-DasGupta 2001). Wilson for single proportions; Newcombe-Wilson or MN for differences/ratios.
• Miettinen-Nurminen score CI is the regulatory standard for RD/RR -- consistent with Pearson chi-square; preferred for NI margins.
• Bender 2001 NNT convention: when RD CI crosses zero, NNT CI is disjoint (NNTB-infinity-NNTH). Report as "NNTB X (NNTB Y to inf to NNTH Z)" -- standard in BMJ/Lancet/Cochrane.
• NNT requires baseline risk to be meaningful -- same OR gives dramatically different NNTs at different baseline risks.
• OR overstates RR when prevalence > 10%. Use modified Poisson (Zou 2004) with HC1/HC3 sandwich SE to estimate RR directly. Log-binomial is an alternative but often fails to converge.
• HC3 sandwich is recommended for n <=250 (Long-Ervin 2000). HC1 (Stata default) and HC3 (R sandwich default) can differ enough to flip NI p-values.
• Hauck-Donner effect (small cell, large coefficient, small Wald chi-square) requires profile-likelihood inference. Detect with R VGAM::hdeff().
• Cornfield 1956 exact OR CI is over-conservative. Mid-p (Berry-Armitage 1995) is less conservative; profile likelihood is transformation-invariant.
• Forest plot reciprocal effects (OR 0.5 vs 2.0) are equidistant on log scale only. Always use log scale.

Related Skills

• clinical-biostatistics/logistic-regression - Adjusted OR + g-computation (in depth)
• clinical-biostatistics/categorical-tests - 2x2 testing producing the ORs
• clinical-biostatistics/subgroup-analysis - Forest plots and stratified effects
• clinical-biostatistics/survival-analysis - HR and RMST as effect measures
• clinical-biostatistics/trial-reporting - CONSORT 2025 + ICH E9(R1) effect reporting
• clinical-biostatistics/missing-data-sensitivity - Effect-measure CIs under MI pooling
• clinical-biostatistics/multiplicity-graphical - Effect reporting post-multiplicity adjustment
• machine-learning/survival-analysis - Predictive survival