missingmed

Overview

missingmed runs regression (GLM)-based mediation analysis across incomplete data and pools with Rubin's rules. It is the missing-data middle of the mediationverse: a thin orchestration layer that fits each analysis with medfit and delegates inference to RMediation.

Two estimators share one S7 pipeline:

• Multiple imputation (method = "mi") — pools per-imputation fits with Rubin's rules.
• Inverse-probability weighting (method = "ipw") — reweights complete cases (stabilized weights, trimming, HC sandwich SEs).
• Structural equation models (engine = "lavaan") — give lavaan syntax in model, including a latent mediator; works with both estimators.
• Sensitivity to missing not at random (sensitivity_mnar()) — repeats the analysis with the imputed values shifted, giving a sensitivity curve.

New to the package? Start with Choosing an analysis on the website.

For the indirect effect it provides both a Monte-Carlo confidence interval and a D4-stacked MBCO likelihood-ratio test (which, unlike pooling, respects the union-null geometry of H0: ab = 0).

Installation

From the Data-Wise R-universe (binaries, no compilation):

install.packages(
  "missingmed",
  repos = c("https://data-wise.r-universe.dev", "https://cloud.r-project.org")
)

Or the development version from GitHub:

# install.packages("pak")
pak::pak("Data-Wise/missingmed")

The pipeline

set_md_mediation()  ->  run()          ->  pool()             ->  infer()
   MDMediationData       MDMediationFit     MDMediationResult      CI / MBCO

library(missingmed)

# `imp` is a mice::mids object; X -> M -> Y with confounder C
md  <- set_md_mediation(imp, Y ~ X + M + C, M ~ X + C,
                        treatment = "X", mediator = "M")  # method = "mi" (default)
res <- pool(run(md))

infer(res, type = "mc")    # Monte-Carlo CI for the indirect effect
infer(run(md), type = "mbco", ariv = "fixed")  # D4-stacked MBCO test of H0: ab = 0

Inverse-probability weighting takes a raw data.frame:

md_ipw <- set_md_mediation(df, Y ~ X + M + C, M ~ X + C,
                           treatment = "X", mediator = "M", method = "ipw")
infer(pool(run(md_ipw)), type = "mc")

See vignette("missingmed") to get started, vignette("mbco-mi") for why MBCO needs the per-imputation fits, and vignette("technical") for the full design, contracts, and methodology.

Ecosystem

missingmed → {medfit (fitting), RMediation (inference)}; simulation via medsim. All are part of the mediationverse.

Citation

Cite the version you used:

citation("missingmed")

License

GPL (>= 3) · Davood Tofighi (dtofighi@gmail.com) · ORCID 0000-0001-8523-7776