Bayesian Adaptive Phase II Oncology Trial Simulator

Response-adaptive randomization simulator for a Phase II oncology trial with a time-to-event primary endpoint and one event-driven interim futility analysis, plus a parallel survival analysis on TCGA-BRCA. End-to-end R + Stan + SAS, with a Quarto report and a mock SAP section in ICH E9(R1) estimand language.

📄 Read the report:

• Rendered PDF — full main report, rendered inline by GitHub
• HTML version — same content as the PDF, served via GitHub Pages
• Source: report/index.qmd (main), report/sap_section.qmd (mock SAP)

Headline result

12,000 simulated trials (1,000 per scenario × 2 designs, ~100 s wall time, 4 furrr workers). The Bayesian adaptive design controlled Type I at 0.019 (vs 0.021 for fixed; both ≤ 0.025 target) and stopped early for futility in 48% of trials under harmful HR and 36% under the null, saving 5–8% of expected sample size in those scenarios. Across non-null effects (HR 0.85 → 0.55), adaptive ceded 0.6–5.6 percentage points of power to the fixed design — the standard adaptive-design trade-off between operational efficiency under no/harmful signal and peak power under strong signal.

What this demonstrates

• Adaptive trial design with an event-driven interim (30% information time under H1, fires when 12 events accumulate) and post-interim Thompson-style response-adaptive randomization (sqrt damping + 20/80 caps, refit every 20 enrollees).
• Frequentist group-sequential boundary via {rpact} (O'Brien-Fleming alpha + beta spending), with both fixed and adaptive designs applying the same final-stage z-boundary (z = 1.969) for apples-to-apples comparison.
• Bayesian exponential survival at the interim, fit in Stan with a baseline-rate-centered Gamma prior on the control hazard. Compiled once, cached on disk, parallelized across furrr workers.
• Parallel survival analytics on n = 1,002 TCGA-BRCA patients: Kaplan-Meier (with log-rank + risk table), Cox PH with Schoenfeld diagnostic, Bayesian Weibull AFT with posterior-predictive KM overlay.
• Two complementary survival models (Cox PH non-parametric vs Weibull AFT parametric) directionally agree on the protective effect of hormone-receptor positive status and the risk-amplifying effect of age, despite a flagged proportional-hazards violation.
• Mock SAP section in ICH E9(R1) estimand language and CDISC-aligned analysis-population structure.
• Reproducibility-as-code: Makefile, config.yml, single-seed RNG, GitHub Actions CI running a reduced 100-sim version of the pipeline + the full TCGA analysis + unit tests + Quarto render on every push.

Stack

R 4.3 · Stan / {rstan} / {rstanarm} · {rpact} · {gsDesign} · {survival} · {survminer} · {furrr} · {bayesplot} · {posterior} · {flextable} · Quarto · SAS OnDemand for Academics

Repository structure

config.yml             all simulation + design parameters (no hard-coded numbers in code)
Makefile               make sims | tcga | report | test | all
R/
  00_setup.R           libs, paths, future multisession, helpers
  01_design_params.R   rpact OBF group-sequential design + interim event target
  02_sim_fixed.R       fixed-design simulator (1:1, log-rank + Cox PH on OBF z)
  03_sim_adaptive.R    adaptive simulator with event-driven Bayesian interim + RAR
  04_run_all_sims.R    furrr orchestrator (--n-sims, --workers CLI flags)
  05_oc_compute.R      operating-characteristics aggregator
  06_oc_plots.R        4 OC figures (power, E[N], futility, summary heatmap)
  07_tcga_data.R       TCGA-BRCA pull + clean (RTCGA::survivalTCGA)
  08_tcga_km.R         Kaplan-Meier by hormone-receptor status
  09_tcga_cox.R        Cox PH + Schoenfeld + stratified sensitivity
  10_tcga_bayes_aft.R  Bayesian Weibull AFT in Stan + PP KM overlay
  11_futility_sensitivity.R  futility-threshold sweep for the SAP sensitivity
stan/
  exp_survival.stan    interim Bayesian model (adaptive design)
  weibull_aft.stan     Weibull AFT model (TCGA case study)
sas/
  seqdesign.sas        PROC SEQDESIGN cross-check of rpact boundaries
  tcga_lifetest.sas    PROC LIFETEST cross-check of survminer
  tcga_phreg.sas       PROC PHREG cross-check of coxph (incl. PH assessment)
report/
  index.qmd            full Quarto report (10 sections, embeds 10 figures)
  sap_section.qmd      standalone mock SAP, ICH E9(R1) estimand structure
tests/                 testthat unit tests, run via per-file Rscript subprocesses
.github/workflows/
  ci.yml               full pipeline + tests + report render on every push
outputs/               figures (PDF + PNG) and tables (CSV)

Reproducing

First-time setup (restores all R + Stan package versions pinned in renv.lock):

# from R, in the project directory
renv::restore()

macOS note. R's "recommended" packages (Matrix, MASS, survival, etc.) ship with R itself rather than being installable from CRAN. renv::restore() on macOS sometimes errors with package 'Matrix' is not available if a CRAN/PPM mirror doesn't list them separately. The renv settings in this project (renv/settings.json → external.libraries and ignored.packages) tell renv to source these from your R installation. If you still hit the issue, install R 4.3.1 from CRAN (pkg.r-project.org) first — it bundles all the recommended packages — then re-run renv::restore(). CI on Linux uses Posit Package Manager and doesn't see this issue.

Then:

make sims         # 12,000 trial simulations (~100 s @ 4 workers)
make tcga         # TCGA KM + Cox + Bayesian AFT pipeline (writes sas/data/tcga_brca.csv too)
make sas-data     # just the CSV export, for SAS-only users
make sensitivity  # futility-threshold sweep used in SAP §10.5 (~90 s)
make report       # render Quarto report and SAP section
make test         # testthat suite (per-file subprocess isolation)
make all          # sims + tcga + report (sensitivity not included by default)
make clean        # remove all generated outputs

All randomness is seeded from CONFIG$simulation$seed = 20260513 with per-sim seeds derived as seed + sim_id * 10 + as.integer(factor(design)). Stan models compile once and cache to stan/*.rds (gitignored). Package versions are pinned via renv (R 4.3.1, lockfile in renv.lock).

Output policy. PNG figures (used by this README) and CSV tables are committed. PDF figures (used only by the rendered Quarto report) and the SAS-input CSV (regenerated by R/07) are gitignored to keep the working tree clean across re-runs. make all regenerates the PDFs and the SAS CSV.

Operating-characteristics summary

Scenario	Design	True HR	Reject rate	95% CI	P(futility)	E[N]	E[events]	Mean HR
null	adaptive	1.00	0.019	(0.011, 0.030)	0.36	113.5	32.8	1.16
null	fixed	1.00	0.021	(0.013, 0.032)	—	120.0	45.0	1.04
harmful	adaptive	1.15	0.008	(0.003, 0.016)	0.48	110.8	30.4	1.33
harmful	fixed	1.15	0.010	(0.005, 0.018)	—	120.0	47.4	1.20
mild	adaptive	0.85	0.075	(0.059, 0.093)	0.22	116.3	35.3	0.98
mild	fixed	0.85	0.069	(0.054, 0.087)	—	120.0	42.3	0.89
moderate	adaptive	0.75	0.139	(0.118, 0.162)	0.16	117.7	35.5	0.86
moderate	fixed	0.75	0.146	(0.125, 0.169)	—	120.0	40.3	0.78
strong	adaptive	0.65	0.238	(0.212, 0.266)	0.11	118.6	35.2	0.74
strong	fixed	0.65	0.263	(0.236, 0.291)	—	120.0	38.3	0.68
very_strong	adaptive	0.55	0.387	(0.357, 0.418)	0.06	119.2	34.6	0.61
very_strong	fixed	0.55	0.443	(0.412, 0.474)	—	120.0	36.2	0.57

Full table: outputs/tables/oc_table.csv. Plots: outputs/figures/oc_*.{pdf,png}.

TCGA-BRCA case study (n = 1,002, 97 events)

Term	Cox PH (HR, 95% CI)	Bayes Weibull AFT (1/TR, 95% CrI)
HR+ vs HR-	0.58 (0.40, 0.84)	0.70 (0.54, 0.93)
Age (per decade)	1.28 (1.10, 1.49)	1.17 (1.07, 1.29)

Both models agree directionally: HR+ is significantly protective and each decade of age significantly increases hazard. The point-estimate gap (Cox 0.58 vs Bayes 0.70 on the HR scale) is the expected behavior when proportional hazards is violated — Cox estimates a time-averaged HR while the Weibull AFT's 1/TR correspondence to HR holds strictly only under both PH and a Weibull baseline. Schoenfeld test flags PH violation for hr_status (p = 0.013), so the disagreement is informative rather than reassuring. Bayes diagnostics: max R̂ = 1.003, min bulk ESS = 1,598, 4 chains × 2,000 iter in ~5 s each.

Limitations and design choices

• Phase II screening design. Maximum n = 120 with 24-month follow-up is deliberately small for a Phase II go/no-go trial; rpact's sample-size calculator says n ≈ 791 would be needed for 80% power at HR = 0.70 with this alpha-spending. Power at smaller effect sizes (HR 0.75 / 0.85) is correspondingly modest. This is by design, not a misconfiguration.
• Interim timing chosen at 30% information under H1. A 50% interim was originally specified but, at this n and event-rate envelope, almost never accumulated the target event count before end-of-study; 30% places the interim inside the practical event-accrual window. Sensitivity analysis across alternative information fractions is on the roadmap.
• Futility threshold (P(HR < 0.7 | data) < 0.20) is operator-defined. A formal sensitivity analysis across alternative thresholds (0.10, 0.15, 0.30) is on the roadmap; the current threshold has not been swept.
• Adaptive HR estimates are biased upward (mean log-HR bias +0.03 to +0.09, decreasing with stronger true effect) due to two mechanisms: (1) futility-stopped trials report the interim Bayesian posterior median, which is pulled toward HR = 1 by the weakly informative prior; (2) RAR-induced allocation imbalance modestly inflates the Cox HR estimate under benefit. The fixed design's bias is the standard small-sample Cox attenuation (-0.01 to -0.03), opposite direction. Quantified in report/index.qmd §7 (bias table by scenario × design); see outputs/tables/oc_table.csv bias_log_hr column. The Cox PH analysis does not adjust for RAR-induced imbalance — an IPTW-weighted sensitivity would be the standard regulatory companion.
• TCGA-BRCA is a survival toolkit validation, not a data-generating-model validation. The endpoint (overall survival in breast cancer) differs from the simulator's (time-to-progression in a hypothetical oncology indication). The TCGA section demonstrates that the same Stan / KM / Cox / AFT pipeline works on real, messier data — not that the simulator's exponential data-generating model matches breast cancer biology.

Background

• ICH E9(R1) — Statistical Principles for Clinical Trials, Addendum on Estimands, 2019.
• FDA — Adaptive Designs for Clinical Trials of Drugs and Biologics, 2019.
• O'Brien PC, Fleming TR (1979) — A multiple testing procedure for clinical trials. Biometrics 35: 549–556.
• Cox DR (1972) — Regression models and life-tables. J R Stat Soc B 34: 187–220.
• Wassmer G, Brannath W (2016) — Group Sequential and Confirmatory Adaptive Designs in Clinical Trials. Springer.