Bayesian Extrapolation of Treatment Effects • BExTE

A collection of R tools to :

Study frequentist and Bayesian operating characteristics of clinical trial designs leveraging Bayesian partial extrapolation (also known as borrowing).
Analyze clinical trial data using Bayesian partial extrapolation methods.

📚 Vignettes

Step-by-step guides to simulation studies, data generation, borrowing methods, and case-study replications.

🔧 API reference

Every exported function and R6 class, grouped by topic: data, methods, plots, tables, operating characteristics.

🖥️ Browser app

Configure, run, and analyze a simulation study from BExTE-app, a separate Shiny app, without hand-editing config files.

📝 Changelog

What changed in each release.

Core contributor: Tristan Fauvel
Contributors: Pascal Godbillot
Project management: Marie Génin
Leads: Julien Tanniou, Billy Amzal
Sponsors: European Medicines Agency, Quinten Health

Quick start

Installation

BExTE needs R >= 4.0 and, for the MCMC-based methods, a C++ toolchain and a CmdStan installation - the Stan models are compiled to native binaries the first time they are used. install.R handles all three.

Download BExTE_0.0.2.tar.gz and install.R from the latest release into the same directory, then run:

Rscript install.R

That installs the dependencies, BExTE itself, and CmdStan if it is missing. Load the package with:

library(BExTE)

Installing by hand

cmdstanr is not on CRAN, so its repository has to be declared before the dependencies will resolve:

options(repos = c(
  CRAN = "https://cloud.r-project.org",
  stan = "https://stan-dev.r-universe.dev"
))
install.packages("/path/to/BExTE_0.0.2.tar.gz", repos = NULL, type = "source")
cmdstanr::check_cmdstan_toolchain(fix = TRUE)
cmdstanr::install_cmdstan()

Working from a source checkout

renv.lock pins every dependency. From the checkout:

renv::restore()
devtools::load_all()

Running and analyzing simulations

To launch a simulation study, run main.R
To plot the results of the simulation study, run plots.R
To create the results tables, run tables.R

Browser-based interface

Instead of hand-editing config files and running the scripts above, you can configure, run, and analyze a simulation study from a browser with BExTE-app, a Shiny app built on this package:

remotes::install_github("TristanFauvel/BExTE-app")
BExTEapp::run_bexte_app()

Design logic

There are three main objects in the simulation framework: source data, target data, and models. Source data are defined based on the case study configuration. Target data depend on the source data, and the specific scenario (drift, sample size). Models correspond to specific methods, and depend on the source data. They implement methods to estimate their operating characteristics. The target data objects implement a sample method, allowing to sample many replicates of the target study.

Access documentation

The full documentation website is available at tristanfauvel.github.io/BExTE.

To access it locally instead : browseURL(“docs/index.html”)

Methods and configurations requiring MCMC inference

MCMC inference with Stan is used in the following scenarios :

RMP x aprepitant
separate x aprepitant
pooling x aprepitant
conditional_power_prior x aprepitant

When using MCMC inference, the method inherits from the MCMCModel class. The initialize() method contains a string with the Stan code. This code is saved in a temporary Stan file which is used to compile the model (using self$stan_model <- cmdstanr::cmdstan_model(stan_file_path)). At inference

Comparison with existing packages

psborrow2 implements Bayesian dynamic borrowing in scenarios where a two-arm RCT is supplemented with external data on the control arm.
RBEsT focuses on meta-analytic and mixture models

Contributing

New contributors are always welcome. Please have a look at the contribution guidelines on how to get started and to make sure your code complies with our guidelines.

Rationale of the design

The core of the code implements the inference logic: a source data class represents source data, which are observed. Target data are represented in a different class. They different as we can sample target data replicates. When initializing a model, source data are provided to it, as well as methods parameters and MCMC configuration (if applicable), so as to define a prior. Inference is performed by using the inference method on target data, which updates properties of the model object with moments of the treatment effect posterior distribution and, if applicable, computes the posterior distribution of borrowing parameters. Note that inference proceeds in two steps: first, if the method uses empirical Bayes, prior parameters are updated based on the target data sample, second Bayes” rule is applied.

When launching a simulation study, all scenarios are generated based on the provided configuration files. These scenarios are then sequentially treated: source data and target data are initiated, and a model is defined. Target data replicates are then generated depending on the target data characteristics. The model is then fitted on each data replicate, and the corresponding inference metrics and frequentist metrics are computed. The data from the simulation are then analyzed: the sweet spot is computed for each metric of interest, and the Bayesian operating characteristics are estimated.