populace

strategies / support

strategy 01 · the support

The support.

A population estimate runs on a set of records: which households exist in the file, and what values each one carries. That set is the support. populace builds it by composing observed survey records — a spine from one survey, variables from all of them — with each record completed by drawing from the population's conditional distribution. Everything downstream can reweight the support; nothing downstream can repair it.

paper: preprint Read the paper Repository
01 — the commitments

Four distributional features, each measured against the method that drops it.

Whatever builds the support must hand downstream consumers a file that behaves like a draw from the population's joint distribution. Four features carry that requirement, and each one is measured in the accompanying paper against an otherwise-identical method that drops it:

the zero mass 0.6ppvs 4.4pp ungated dividend zero-share error; 25–84pp for linear and Gaussian baselines
dependence between filled variables 5.8AUC points cost of independent draws on a four-part balance-sheet identity
the survey design 6×worse Wasserstein-1 unweighted within-SCF net worth; population-view p99 inflated 1.9–2.3×
the extreme tail 0.70–0.81forest p99 ratio; donation ≈1 the axis where donation methods beat conditional draws — kept on the scoreboard

All figures from the imputation paper's committed experiment runs (draft; regenerates from run configs). The fourth row is a limit of the current method, not a win: forest draws do not extrapolate beyond training support, so donation-based methods hold the extreme tail better.

02 — the current method

A spine of observed records, completed by sequential zero-inflated QRF.

compositionshipped

Multispine

Every record in the support is an observed survey record, never a synthesized one: a CPS ASEC spine carries the demographics, a PUF support channel carries tax detail, and an ACS channel is planned. Cloned records keep their source identity, so every value traces to the survey that observed it.

completionshipped

Sequential zero-inflated QRF

Variables a record's own survey didn't observe are drawn, not predicted: a gate model settles the zero/participation regime, a quantile regression forest draws the amount within it, and multi-component targets are drawn in sequence, each conditioned on the ones already drawn.

weightsshipped

By construction

Design weights enter the estimator itself — bootstraps drawn proportional to weight, gates fit on weighted samples — and the interface requires an explicit weights decision, so an unweighted fit is a recorded choice rather than a silent default.

benchmarked against unweighted & weighted QRF, OLS, linear quantile regression, hot-deck statistical matching, and an unconditional weighted draw

How the support is built Three observed-survey channels — the CPS ASEC spine, the PUF donor channel, and a planned ACS channel — feed a clone step that preserves each record's source identity, then a sequential zero-inflated quantile-regression-forest completion step draws every variable a record's own survey didn't observe. The output is the support. spine CPS ASEC demographics donor channel IRS SOI PUF tax detail donor channel · planned Census ACS not yet integrated clone source identity kept per record completion · gate zero / participation regime completion · draw quantile regression forest sequenced multi-component targets, in order output the support throughout design weights enter the estimator: weighted bootstraps, weighted gate fits
The support's build path: each channel is an observed survey record, never a synthesized one. Cloning keeps every record's source identity, so a value can always be traced back to the survey that measured it. Completion runs a gate model (zero/participation regime) then a quantile regression forest (the amount), sequencing multi-component targets so each is conditioned on the ones already drawn.
03 — against the alternatives

Scored like any other candidate.

The support is judged by the population-view harness: any weighted population file — composed, donated, or generated — is projected through each survey's view and scored against that survey's own holdout, with sampling-noise floors marking what an independent draw from the same population would score. The current method's results against the standard alternatives, from the paper's committed runs:

Hot-deck matching is a genuine near-peer — its classifier AUC (0.768) sits almost on top of the candidate's (0.761), and its donated values hold the extreme tail better than any conditional method. Ordinary least squares collapses both failure axes at once: 84 points of dividend zero-share error, and a reweighting stress test shows it understating the truth's tail exposure by factors of 7 to 11 — missing tails, not robustness. An unconditional weighted draw that discards every predictor posts competitive marginal-fit numbers and ties on energy distance, then gets caught by the classifier at AUC 0.968: marginal metrics alone cannot distinguish a method that draws from the correct weighted marginal from one that also gets the conditionals right.

A fully generative synthesis is an admissible occupant of this slot — it would replace both the spine and the completion step, producing records outright. The harness exists so that choice stays a measured one: the scoreboard above is method-agnostic, and a generated candidate would be scored on exactly the same axes against exactly the same holdouts.

Classifier AUC: candidate vs. hot-deck vs. an unconditional weighted draw Bar chart of classifier two-sample-test AUC, scaled from 0.5 (indistinguishable, marked as the reference floor) to 1.0 (perfectly separable, a failure). The current method scores 0.761 AUC, hot-deck statistical matching scores 0.768 — a near-peer, essentially tied. An unconditional weighted draw that discards every predictor scores 0.968 AUC — caught almost perfectly by the classifier despite competitive marginal-fit numbers. 0.5 0.75 1.0 indistinguishable perfectly separable candidate 0.761 hot-deck matching 0.768 unconditional weighted draw 0.968 0.5 = the reference floor
Classifier two-sample-test AUC (lower is better; 0.5 is indistinguishable from the holdout). Hot-deck statistical matching (0.768) is a genuine near-peer to the current method (0.761) — the two sit almost on top of each other. An unconditional weighted draw that discards every predictor (0.968) is caught by the classifier despite tying on energy distance, which is why the harness scores four blocks rather than one.

Installable on its own.

populace-fit on GitHub Read the paper