Trial Reporting - Usage Guide

Overview

Prepares statistical reports for clinical trials following CONSORT 2025 (Hopewell et al Lancet 2025), SPIRIT 2025, ICH E9(R1) Addendum on Estimands (2019), and FDA 2023 Final Guidance "Adjusting for Covariates in RCTs." Covers Table 1 generation, analysis populations (ITT vs FAS vs PP vs Safety), the 5 ICH E9(R1) intercurrent-event strategies, MMRM under MAR with Kenward-Roger via mmrm package, reference-based MI via rbmi (J2R/CR/CIR), Permutt tipping-point sensitivity, and the Cro vs Bartlett information-anchored vs frequentist variance debate.

Prerequisites

pip install tableone statsmodels scikit-learn pandas numpy

R is strongly recommended for confirmatory regulatory work (Python lacks Kenward-Roger):

install.packages(c('mmrm', 'rbmi', 'gMCP', 'rpact', 'gsDesign'))

Quick Start

Tell your AI agent what you want to do: - "Define the ICH E9(R1) estimand for my primary endpoint (5 attributes including ICE strategy)" - "Fit MMRM with Kenward-Roger via the mmrm R package; UN covariance with pre-specified fallback hierarchy" - "Run J2R reference-based MI via rbmi as MNAR sensitivity for treatment-policy estimand; report BOTH Rubin's and CMI+jackknife variance" - "Compute Permutt tipping-point delta that flips primary p > 0.05; report in residual SD units" - "Generate Table 1 with TableOne; pre-specify CONSORT 2025 + SPIRIT 2025 reporting per item 4 (data sharing) + item 21c (missing data) + item 15 (harms)"

Example Prompts

Estimand specification

"My primary endpoint is change in HbA1c at 24 weeks. ICE is treatment discontinuation due to AE. Define the ICH E9(R1) estimand with 5 attributes and choose between treatment policy, hypothetical, composite, while-on-treatment, or principal stratum."

"Cite Olarte Parra 2022 to argue MMRM-MAR IS a causal hypothetical estimator under specific identifying assumptions. Articulate the hypothetical scenario in the SAP."

MMRM with Kenward-Roger

"Implement MMRM in R mmrm package with UN covariance + Kenward-Roger-Linear (matches SAS PROC MIXED bit-for-bit). Pre-specify fallback: UN+KR -> UN+Satterthwaite -> heterogeneous Toeplitz -> AR(1) -> CS."

Reference-based MI

"Apply J2R imputation in rbmi for treatment-discontinuation ICEs in my treatment-policy estimand. Use both Bayesian MI + Rubin's rules (Cro 2019 information-anchored) AND CMI+jackknife (Wolbers 2022 frequentist). Justify discrepancy in CSR."

Tipping-point sensitivity

"Permutt 2016 tipping-point analysis: apply delta only to active-arm imputed values; scan delta from 0 to 20 in units of residual SD. Report the tipping delta that flips primary p > 0.05. Judge clinical plausibility against MCID."

Aducanumab/Aprocitentan-style differential dropout

"My trial has 35% dropout in active arm vs 12% in placebo, predominantly due to AEs. MAR is implausible. Switch from MMRM-MAR primary to treatment-policy with retrieved-dropout MI as primary, J2R as sensitivity (Aprocitentan 2024 precedent)."

Table 1 and analysis populations

"Generate Table 1 with age, sex, race, BMI, baseline severity. Group by ARM. Use SMD (NOT p-values) to assess balance. Distinguish ITT, FAS, PP, Safety populations with explicit pre-specified criteria."

CONSORT 2025 reporting

"Review my analysis against CONSORT 2025 (Hopewell 2025 Lancet 405:1633). Verify items 4 (data/code sharing), 8 (PPI), 15 (harms), 21c (missing data), 24a/24b (TIDieR intervention description)."

"Estimands did NOT make consensus for mandatory CONSORT 2025 inclusion (Box 1 terminology only). Report the estimand per ICH E9(R1) directly."

What the Agent Will Do

Define the ICH E9(R1) estimand with 5 attributes BEFORE choosing analysis method
Examine DS (Disposition) domain to characterise dropout patterns by arm
Choose estimand strategy based on clinical reasoning (treatment policy / hypothetical / composite / WoT / principal stratum)
Generate Table 1 with SMD-based balance assessment
Define ITT, FAS, PP, Safety populations with explicit pre-specified criteria
Execute MMRM-MAR (R mmrm) OR reference-based MI (R rbmi) per pre-specified estimand
Run MNAR sensitivity analyses including Permutt tipping-point in residual SD units
Apply graphical multiplicity (Bretz-Maurer via gMCP) for primary + key secondary
Format reporting per CONSORT 2025 30-item checklist + ICH E9(R1) estimand statement

Tips

The estimand comes FIRST. Kahan 2023 Am J Epidemiol 192:987: 98% of published trials don't articulate the estimand. Define 5 ICH E9(R1) attributes BEFORE choosing the method.
Aducanumab (2021) is the textbook case where MAR-primary in trial with high differential missingness was regulator-divisive (the Nov 2020 AdCom voted overwhelmingly against; FDA approved in June 2021 anyway). Switch to treatment-policy + reference-based MI when differential dropout suggests informative missingness.
Aprocitentan (2024) precedent is now the de facto FDA standard: hybrid imputation -- J2R for treatment-discontinuation ICEs, MMRM-MAR for other missingness.
Wegovy/Ozempic STEP precedent: retrieved-dropout MI for treatment-policy when post-ICE data collected (FDA 2025 obesity guidance endorses).
MMRM with Kenward-Roger is the FDA-favoured continuous-endpoint analysis. mmrm package method = "Kenward-Roger-Linear" matches SAS PROC MIXED bit-for-bit. Pre-specify the convergence fallback hierarchy in the SAP.
Reference-based MI variance debate (Cro vs Bartlett) is unsettled. Report both Rubin's (information-anchored per Cro 2019) AND frequentist (CMI+jackknife per Wolbers 2022) for safety.
Permutt tipping-point delta should be in residual SD units (FDA preference) for cross-trial comparison.
LOCF is biased even under MCAR (Mallinckrodt 2008). NRC 2010 Rec 10 rejects LOCF. Don't use as "conservative" sensitivity.
Selection models (Diggle-Kenward) should be sensitivity only. FDA prefers pattern-mixture (reference-based MI) because it is clinically articulable.
sklearn.IterativeImputer caveats: sample_posterior=True only works with BayesianRidge (silently ignored otherwise). For confirmatory work, use R rbmi or mice.
Imputation model must be congenial with analysis model (Meng 1994). If analysis includes interactions, imputation must too.
FAS vs ITT distinction matters: FAS may exclude eligibility failures or no-post-baseline. ITT cannot. Pre-specify both with explicit FAS exclusion criteria.
Senn 1994 baseline balance testing is incoherent. Don't condition adjustment on observed imbalance -- destroys nominal Type-I error.
CONSORT 2025 (Hopewell 2025) is 30 items, not 25. 7 new items, 3 revised. Estimands in Box 1 only (consensus failed for mandatory). No DOORS framework.
SPIRIT 2025 has 34 items (was 33). New PPI item (11) and trial-monitoring item (29).

clinical-biostatistics/cdisc-data-handling - SDTM/ADaM data preparation
clinical-biostatistics/logistic-regression - Primary analysis with marginal-vs-conditional
clinical-biostatistics/effect-measures - Effect-measure CIs under MI pooling
clinical-biostatistics/subgroup-analysis - Pre-specified subgroup analyses for CSR
clinical-biostatistics/missing-data-sensitivity - MMRM/rbmi/tipping-point in depth
clinical-biostatistics/multiplicity-graphical - Bretz-Maurer graphs for co-primary
clinical-biostatistics/survival-analysis - Time-to-event estimand framing
clinical-biostatistics/power-and-sample-size - Sample size justification per CONSORT 2025 item 16a
clinical-biostatistics/adaptive-designs - Adaptive trial reporting
clinical-biostatistics/bayesian-trials - Bayesian inference reporting
reporting/rmarkdown-reports - Formatted statistical report generation
experimental-design/multiple-testing - General multiplicity correction