research paper · preprint

Weight-aware imputation for policy microsimulation

The paper evaluates populace's regime-gated, sequentially-chained, weighted-bootstrap quantile-regression-forest imputation estimator against unweighted and weighted QRF, OLS, linear quantile regression, and hot-deck statistical matching — with ablations attributing the gains to each design choice under a population-view scoring harness.

authors

Max Ghenis, María Juaristi — PolicyEngine.

headline finding

Weighting is the whole effect: the unweighted ablation posts 6× worse Wasserstein-1 on the within-SCF wealth task, and inflates the population-view holdout's tail (q99 ratio 1.9–2.3×) even where marginal geometry metrics see a tie.

scope

21 pages. Committed run artifacts back every number; the one remaining TODO before submission is the funding statement.