LoginSign Up
← Back to all sparks
B

BayesianMCPMod

INFRA · APIS
Velocity0.0

Simulate, Evaluate, and Analyze Dose Finding Trials with Bayesian MCPMod

github.com/boehringer-ingelheim/bayesianmcpmod ↗

A Bayesian dose-finding package extends from continuous endpoints to binary ones

clinical-trialsdose-findingbayesian-statisticsr-packagebiostatistics
◆Current state
BayesianMCPMod implements the Bayesian form of MCP-Mod for dose-finding trials, combining a multiple-comparison test for dose-response signal with model fitting for dose selection. Version 1.3.0 added functions and vignettes for the binary endpoint case, opening the package beyond the continuous endpoints it was built around, and 1.3.2 followed with Firth's penalized regression to handle separation in those binary fits. The same 1.3.0 release let assessDesign() accept custom simulated data and custom model estimates, which moves simulation control out of the package and into the user's hands.
◆Where it's heading
Each release has widened the estimands and data shapes the framework accepts rather than changing its statistical core. 1.0.2 added non-monotonic beta and quadratic model shapes; 1.1.0 introduced getMED() for the minimally efficacious dose and parallel execution through the future framework; 1.2.0 switched the posterior and contrast functions from a standard deviation vector to a full covariance matrix and supported non-zero off-diagonals in the MCP step. The binary endpoint work is the same pattern applied to the outcome type, and the Firth addition shows the follow-through of a maintainer who has hit the separation problem in practice.
◆Prediction
Expect the binary endpoint arm to keep filling in - more diagnostics and design assessment coverage matching what the continuous case already has - since 1.3.2 addressed a specific estimation failure rather than adding a new capability.

◆Recent moves

  1. 3mo ago

    Firth penalized regression handles separation in binary endpoints

    Adds Firth's penalized regression for binary endpoints under separation, the case where standard maximum likelihood estimates diverge. A narrow addition, but it is the practical follow-up to the binary endpoint support introduced three months earlier and suggests real trial data has been run through it.

    View source ↗
  2. 5mo ago

    Regression fix for missing future.apply, plus credible band options

    Repairs a regression from the previous release that only appeared when the optional future.apply package was absent, and loosens the bootstrapped credible band settings in plot.modelFits(). A patch closing out 1.3.0 rather than adding to it.

    View source ↗
  3. 5mo ago

    Binary endpoint support opens the framework past continuous outcomes

    ⚡ SPARK

    The release that changed what kind of trial this package can analyse. Everything before it extended model shapes, estimands and computation within continuous endpoints; this adds a second outcome type, and 1.3.2's Firth regression is already servicing it.

    View source ↗
  4. 11mo ago

    Covariance matrices replace standard deviation vectors in the MCP step

    Switches getPosterior(), getCritProb() and getContr() to accept a covariance matrix and supports non-zero off-diagonals, meaning correlated dose-level estimates can be handled properly. Also fixes a significance-assignment bug between the MCP and Mod steps that could attach the wrong status to a fitted model, which is the kind of error that silently changes conclusions.

    View source ↗
  5. 1y ago

    Minimally efficacious dose estimation and parallel execution

    Adds getMED() and threads it through assessDesign() and performBayesianMCPMod, giving the package a second decision quantity alongside dose-response model fitting. Parallel processing via the future framework arrives here too, which is what later made future.apply an optional dependency worth managing.

    View source ↗
  6. 1y ago

    Non-monotonic beta and quadratic dose-response shapes

    Adds beta and quadratic model shapes, letting the package fit dose-response curves that are not monotonic - relevant where efficacy falls off at high doses. The vignette comparing frequentist and Bayesian MCP-Mod under vague priors is a positioning move as much as documentation.

    View source ↗