Model-Averaged Propensity Scores Selected by Prognostic-Score Balance

Constructs a model-averaged propensity score as a convex combination of candidate propensity score models, with mixing weights selected on a simplex grid to optimize covariate or prognostic-score balance, implementing the method of Kabata, Stuart and Shintani (2024) . Prognostic scores follow Hansen (2008) : outcome models are fit on untreated units only. The resulting score is designed to be supplied directly to the matchit() function of 'MatchIt' as a distance measure or to the weightit() function of 'WeightIt' as a propensity score, with balance assessment via 'cobalt'.

psAve

Model-averaged propensity scores selected by prognostic-score balance.

psAve implements the propensity score model averaging method of Kabata, Stuart & Shintani (2024, BMC Medical Research Methodology). Instead of committing to a single propensity score model, psave() fits several candidate models (logistic regression, CART, random forest, gradient boosting, or any SuperLearner wrapper), and combines them as a convex mixture whose weights are chosen on a simplex grid to optimize balance on the prognostic score — the predicted outcome under the untreated condition, estimated from untreated units only (Hansen 2008). Because the prognostic score summarizes the covariates as they relate to the outcome, balancing it targets exactly the covariate directions that drive bias in the treatment effect; in the paper's simulations this "Prog (Ave)" strategy achieved the lowest and most robust bias and RMSE across 16 scenarios. Covariate-balance (SMD, KS) and prediction-accuracy (log loss) criteria from the paper are also available.

psAve is a companion to the MatchIt/WeightIt/cobalt ecosystem, not a replacement for any part of it. The deliverable is a plain numeric vector of propensity scores, handed to MatchIt::matchit() as a distance measure or to WeightIt::weightit() as a propensity score (thin psave_match() / psave_weight() wrappers do this without retyping the formula); balance assessment — including prognostic-score balance — works out of the box via cobalt::bal.tab(); effect estimation stays where it belongs, in MatchIt/WeightIt/survey/marginaleffects. Estimands: ATT (default) and ATE.

Installation

# install.packages("remotes")
remotes::install_github("kabajiro/psAve")

Quick example

library(psAve)
data("lalonde", package = "MatchIt")

set.seed(1234)
fit <- psave(treat ~ age + educ + race + married + nodegree + re74 + re75,
             data = lalonde, outcome = ~ re78)      # defaults: criterion = "prog", ATT
m <- psave_match(fit, method = "nearest")           # a genuine matchit object
cobalt::bal.tab(m, distance = data.frame(prog = fit$prog))

Learn more

Getting Started with psAve — motivation, the lalonde matching workflow, interpreting output, and why using the outcome in the design stage does not bias the analysis.
Reproducing the published IPW workflow — ATT and ATE weighting, and the paper's survey::svyglm() estimator.
Method details and design decisions — all formulas, the documented fixes relative to the paper's reference code, relation to other software, and limitations.

Citation

If you use psAve, please cite the paper:

Kabata, D., Stuart, E. A., & Shintani, A. (2024). Prognostic score-based model averaging approach for propensity score estimation. BMC Medical Research Methodology, 24, 228. doi:10.1186/s12874-024-02350-y

@article{kabata2024prognostic,
  author  = {Kabata, Daijiro and Stuart, Elizabeth A. and Shintani, Ayumi},
  title   = {Prognostic score-based model averaging approach for propensity score estimation},
  journal = {BMC Medical Research Methodology},
  year    = {2024},
  volume  = {24},
  pages   = {228},
  doi     = {10.1186/s12874-024-02350-y}
}