Frequentist Dynamic Borrowing for Hybrid-Control Survival Trials

Implements a class of likelihood-informed frequentist dynamic borrowing methods for hybrid-control survival trials based on penalized Cox partial likelihood estimation. Implements four likelihood-informed penalty structures (precision-weighted L1, smoothed integrated-gate, information-adaptive minimax concave penalty (MCP), and likelihood-ratio-weighted L1), together with the adaptive lasso borrowing approach of Li et al. (2023, ). Provides conditional model-based standard errors and local plug-in sandwich variance approximations, with smoothed penalties. Tools for design-stage lambda calibration via simulation, including a two-stage coarse-fine grid search, drift-level early stopping, and per-method tuning under both inference types, are also provided. A simulation harness for evaluating type I error and statistical power across population drift scenarios is included.

fdb

Frequentist Dynamic Borrowing for Hybrid-Control Survival Trials

The fdb package implements a class of likelihood-informed frequentist dynamic borrowing methods for hybrid-control survival trials, based on penalized Cox partial likelihood estimation. The borrowing strength is controlled by a single tuning parameter (\lambda) acting on a drift parameter that captures population shift between concurrent and external controls. Both standard model-based standard errors and penalized estimating-equation sandwich standard errors are provided.

Penalty methods

The package implements four likelihood-informed penalties together with the adaptive lasso comparator of Li et al. (2023):

All penalties operate on a smoothed absolute-value primitive (|\delta|_\varepsilon = \sqrt{\delta^2 + \varepsilon^2}), so the resulting penalized objective is twice differentiable in (\delta). This makes plug-in sandwich variance estimation possible as a local plug-in approximation. First-stage estimates are held fixed; conditional or unconditional validity is not guaranteed. Negative P2/P3 curvature is clipped only for the variance calculation; raw curvature remains available as pen_curv_raw.

P2 integrates the gate from eps to sqrt(delta^2 + eps^2). MCP integrates a C1 slope from zero to sqrt(delta^2 + eps^2) - eps, smoothing both the origin and the flat-tail transition. Set rho_mcp in direct fits, calibration, or run_fdb_study, or in the tuning list passed to run_simulation. Its software default is 0.1; it is not a calibrated manuscript value. Existing function names and method tags are retained for compatibility. Both P2 and P3 results and calibrated lambdas from the old definitions need recomputation. No-covariate studies are supported with p = 0.

Installation

From a local source build:

# install.packages("remotes")
remotes::install_local("fdb_0.2.0.tar.gz", dependencies = TRUE)

Quick start

library(fdb)

# Simulate a hybrid-control trial with HR = 0.8 and no population drift
set.seed(1)
sim <- simulate_hybrid_cox(
  nI1 = 150, nI0 = 150, nE = 300,
  theta0 = log(0.8), delta0 = 0
)

# Fit all methods on this dataset
fit_all_methods(sim$data)

# Fit a single method
fit_one_penalized_method(sim$data, method = "P1", lambda = 0.2)

Design-stage calibration and operating characteristics

For confirmatory use, the penalty strength (\lambda) should be calibrated to target an error-rate threshold over a prespecified drift grid. Finite Monte Carlo calibration does not guarantee control between grid points or outside that grid:

study <- run_fdb_study(
  scenario_base   = scenario_S1,
  drift_hr_range  = c(0.8, 1.2),
  drift_by_hr     = 0.05,
  alpha           = 0.025,
  alt_hr          = 0.8,
  do_calibration  = TRUE,
  nsim_cal        = 500,
  nsim_curve      = 1000,
  parallel        = TRUE,
  ncores          = 2,
  seed            = 1
)

study$type1_curve
study$power_curve
study$calibration$lambda_star

References

• Li, R., Lin, R., Huang, J., Tian, L., and Zhu, J. (2023). A frequentist approach to dynamic borrowing. Biometrical Journal 65(7), 2100406.
• Zhang, C.-H. (2010). Nearly unbiased variable selection under minimax concave penalty. The Annals of Statistics 38(2), 894-942.
• Andersen, P. K. and Gill, R. D. (1982). Cox's regression model for counting processes: A large sample study. The Annals of Statistics 10(4), 1100-1120.

License

MIT (see the LICENSE file).

Interpretation and reproducibility

The model-based SE conditions on the estimated drift as a fixed offset and can substantially underestimate uncertainty. The sandwich SE is a local plug-in approximation holding adaptive weights fixed; nominal coverage is not guaranteed. Report observed coverage, bias, RMSE and valid-fit counts. An alpha_cal threshold above alpha permits error inflation relative to the nominal test level. ESS is a variance-equivalent gain, not a literal number of borrowed controls; negative values indicate a precision loss and positive values do not establish low bias.

Calibration never substitutes an uncalibrated default when it fails. Refine the prespecified design or candidate grid and repeat confirmation. An explicit drift_set_confirm is honored by the calibration functions; when omitted, confirmation uses the calibration grid. The one-stop study wrapper confirms on its calibration grid and evaluates performance on its wider curve grid.

run_simulation() always returns replicate estimates in $raw. For curves:

# Larger nsim is needed for scientific conclusions.
curve <- run_drift_curve(theta0 = 0, drift_set = log(c(1, 1.1)),
                         scenario_base = scenario_S1, lambdas = lambdas_default,
                         nsim = 2, keep_raw = TRUE)
curve$summary
head(curve$raw)

The default keep_raw = FALSE preserves the summary-only curve interface. run_fdb_study(keep_raw = TRUE) also returns raw_type1 and raw_power. Save these with saveRDS() for paired ESS uncertainty and normality diagnostics. Simulation summaries include n_valid, n_missing, and Monte Carlo standard errors. Check those counts before interpreting rates. Parallel execution is opt-in and defaults to two workers. Install the package before starting workers; fixed seeds are reproducible for a fixed worker count and environment, not necessarily between serial runs and different parallel configurations.