LoginSign Up
← Back to all sparks
P

probmed

ANALYTICS
Velocity0.0

R package probmed by Data-Wise — release notes from GitHub.

github.com/Data-Wise/probmed ↗

probmed went from one probabilistic effect size to a family of them in sixteen days.

causal mediationeffect sizesemiparametric inferencecross-fittingmediationverser package
◆Current state
probmed computes P_med, a scale-free probabilistic effect size for causal mediation, as part of the Data-Wise mediationverse alongside medfit, medsim and RMediation. Three releases in three weeks took it from a single estimator to four additional families built on a shared cross-fitted corner-EIF core, covering gauge-calibrated, incremental-elasticity and Sobol variance-share versions of the proportion mediated. Distribution is GitHub and r-universe rather than CRAN, with a load-bearing Remotes pin on medfit.
◆Where it's heading
The pace is manuscript-driven — estimators arrive with their citations attached and vignettes alongside, and the 0.1.0 notes correct the estimand itself against a manuscript definition rather than fixing a bug in code. Each release adds inference machinery as well as point estimates: percentile-bootstrap intervals and Fieller sets in 0.3.0, a deterministic MBCO interval in 0.2.0 that avoids resampling entirely. The gauge residual and the pmed_sensitivity() helper suggest a growing concern with when the estimand does not decompose at all.
◆Prediction
0.3.0 shipped a sensitivity helper for shared mediator-outcome confounding and a diagnostic that flags non-decomposability, so the next release most likely extends that diagnostic side rather than adding a fifth estimator family.

◆Recent moves

  1. 1mo ago

    Four estimator families on a cross-fitted corner-EIF core

    ⚡ SPARK

    Adds four estimator families on a shared cross-fitted corner-EIF core: gauge-calibrated proportion mediated, incremental mediated elasticity with a Neyman-orthogonal standard error, a Sobol variance-share version, and a sensitivity helper for shared mediator-outcome confounding. This turns a single-estimator package into a suite covering several definitions of proportion mediated.

    View source ↗
  2. 2mo ago

    Parallel mediators and a resampling-free MBCO interval

    ⚡ SPARK

    Extends P_med to k parallel mediators moved together from control to treated levels, recovering the single-mediator formula at k=1, and adds a deterministic MBCO interval obtained by inverting the constrained-likelihood test instead of resampling. All four estimation methods support the parallel case.

    View source ↗
  3. 2mo ago

    First release: pmed() with the estimand corrected

    The founding release, shipping pmed() with plugin and bootstrap methods — and correcting the estimand in the same breath. The previous computation returned a direct-effect contrast that depended on c' and could land on the wrong side of 0.5; it now computes the mediation estimand, verified against a closed form and a worked manuscript example.

    View source ↗