Analysis Plan

The AI read your question, dataset profile, and variable definitions and proposed the plan below — including outcome, predictors, and model type. Review and confirm before running.

How the system interpreted your question

AI Parsed

Outcomesurvival_90d (binary: 1 = survived at 90 days)

Primary exposuretreatment_group (A = reference, B = intervention)

DirectionSuperiority: B > A

Covariatesage · sex · baseline_score · comorbidity_index · site

ObjectiveMeasure the adjusted treatment effect on 90-day survival; test whether Group B is superior to standard care.

If this doesn't match your intent, go back and revise your question or variable definitions — then regenerate the plan.

Model Selection

AI Generated

Model Family

Logistic regression (binomial family, logit link)

Formula

survival_90d ~ treatment_group + age + sex + baseline_score + site + comorbidity_index + treatment_group:baseline_score

Why this model

Binary outcome (0/1) with ~12% event rate — logistic regression is the canonical choice.
Interaction term treatment_group × baseline_score included per protocol hypothesis (does benefit differ by baseline severity?).
Site included as fixed effect (8 levels) given study-level sample size; mixed model explored as sensitivity analysis.
Complete-case analysis as primary; multiple imputation included as pre-specified sensitivity for age / baseline_score missingness.

Planned Diagnostics

1Complete-separation detection (Hauck–Donner effect scan)

2Variance Inflation Factor (VIF) for all predictors

3Hosmer–Lemeshow goodness-of-fit (decile groups)

4ROC / AUC with 1 000-replicate bootstrap 95% CI (seed 42)

5Calibration plot (predicted vs observed decile frequency)

6Influence analysis: Cook's distance · dfBetas

Safeguards & Fallbacks

Active

Separation detected → refit with Firth-penalised logistic regression (logistf / statsmodels penalized).

Max VIF > 10 → flag, suspend interaction term, re-evaluate collinear predictor.

EPP = 7.8 — advisory flag raised (below recommended threshold of 10). Analysis proceeds with caution; interpret coefficient estimates carefully.

All random operations seeded: seed = 42.

Reproducibility

Locked

Random seed

Python

3.11.8

statsmodels

0.14.1

pandas

2.2.0

numpy

1.26.4

matplotlib

3.8.3

scikit_learn

1.4.1