LoginSign Up
← Back to all sparks
M

medrobust

ANALYTICS
Velocity0.0

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

github.com/Data-Wise/medrobust ↗

medrobust made its partial-identification bounds usable by giving them confidence intervals.

causal mediationpartial identificationmisclassificationsensitivity analysismediationverser package
◆Current state
medrobust computes partial-identification bounds for mediation effects when exposure or mediator is differentially misclassified, part of the Data-Wise mediationverse. Its 0.2.0 release corrected three estimator defects against population oracles and added Imbens-Manski confidence intervals for the bounds; the two releases since have paired each identification path with a real public-domain dataset and a worked vignette. CRAN is deferred, with distribution through GitHub and r-universe.
◆Where it's heading
The pattern is deliberate and symmetric: 0.3.0 shipped the mediator-side example on NCHS natality data, 0.4.0 its exposure-side mirror on NHANES, each demonstrating what the bounds do when reporting accuracy is allowed to depend on the outcome. Alongside that runs a consistent concern with failing usefully rather than loudly — bound_ne() returns NA bounds with a machine-readable reason and a typed condition instead of aborting, so a simulation replicate is recorded rather than lost, and non-finite endpoint standard errors produce a documented NA rather than a silent one. That is a package expecting to be run thousands of times inside someone else's loop.
◆Prediction
Both identification paths now have a dataset, a vignette and interval coverage, so the next release is most likely the deferred CRAN submission rather than new methodology.

◆Recent moves

  1. 2mo ago

    NHANES exposure-side misclassification example dataset

    Adds nhanes_pa, a pooled NHANES 2015-2018 sample of 9,906 records, as the exposure-side mirror of the mediator-side gesthtn example — self-reported physical inactivity misclassified against laboratory-measured hs-CRP. The worked result is the point: the natural direct effect survives mild differential reporting, but its Imbens-Manski interval covers the null once reporting accuracy is allowed to depend strongly on the outcome.

    View source ↗
  2. 2mo ago

    Natality example dataset; bounds degrade instead of aborting

    Pairs the mediator path with a real-data example — a 5,000-row NCHS Natality 2021 sample for gestational hypertension and preterm birth — and makes the bounds degrade usefully. bound_ne() now returns NA bounds with a machine-readable reason and a typed medrobust_infeasible condition instead of aborting when no compatible parameter sets exist, so simulation replicates are recorded rather than lost.

    View source ↗
  3. 2mo ago

    Bounds corrected against oracles; Imbens-Manski intervals added

    ⚡ SPARK

    Corrects three estimator defects — a mis-specified 3x3 mediator solve, a plug-in-mean true-effects estimand that should have been g-computation, and an exposure NIE assembly treating a conditional as joint — all verified against population oracles to 5e-17. It then adds Imbens-Manski confidence intervals, lifting coverage from a raw-bound range of 0.00-0.42 to 0.90-1.00.

    View source ↗