NEWS.md

In wdsmatch: Weighted Double Score Matching for Survey-Weighted Causal Inference

wdsmatch 0.2.0

This release corrects the numerical implementation. Point estimates, standard errors, and confidence intervals can differ from 0.1.1. Analyses made with 0.1.1 should be rerun before numerical results are reused.

• Match on standardized propensity probabilities and arm-specific prognostic scores. Replace the old logit and joint-score polynomial with a complete quadratic basis in each arm's own two-dimensional double score.
• Correct PATE's replication residual contributions using the original weighted matching reuse coefficients. Use the corresponding fixed-reuse PATT formula.
• Apply survey-weighted propensity fitting for retrospective sampling and unweighted propensity fitting for prospective sampling, including replication refits. Use finite-checked, zero-initialized logistic fits.
• Use centered normal Wald intervals based on replication variance with divisor B, replacing the previous percentile intervals and B-1 variance convention.
• Respect disabled bias correction in both the point estimate and replication. This option retains matching discrepancies and does not provide the same asymptotic bias-correction guarantee.
• Hold supplied propensity/prognostic scores fixed in replication; corresponding standard errors are conditional on those supplied scores. PATT only requires the control-side prognostic model.
• Reject invalid inputs and unsuccessful numerical fits explicitly. Do not insert zero replicates or silently discard failed replicates.
• Retain the existing public arguments and result components, and add variance, confidence-level and procedure metadata. Print the requested confidence level rather than always labeling an interval as 95%.
• Add regression checks for the corrected calculations, improve the method and inference documentation, and add checks on Linux, Windows, and macOS.

wdsmatch 0.1.1

• Add missing return-value documentation for the print and summary methods for CRAN resubmission. No numerical estimator changes from 0.1.0.

wdsmatch 0.1.0

• Initial package version.
• wdsmatchATE(): Population average treatment effect estimation.
• wdsmatchATT(): Population average treatment effect on the treated.
• Supports retrospective and prospective sampling designs.
• Polynomial sieve bias correction with the original 10-term logit basis.
• Linearization-based multinomial bootstrap variance estimation.
• Bundled example dataset survey_obs.

wdsmatch documentation built on Sept. 12, 2026, 1:06 a.m.