[MaxGhenis]

We add experiments/moop_spm, which measures how the choice of method for imputing medical out-of-pocket spending (MOOP) changes Supplemental Poverty Measure (SPM) poverty. The American Community Survey (ACS) does not ask about MOOP, so researchers who build an SPM on the ACS impute it. The 2026 Current Population Survey Annual Social and Economic Supplement (CPS ASEC) asks about MOOP, and its public-use file covers calendar-year 2025. We impute MOOP within that file and check each method against the file's own MOOP. We compare logit models with cell means, OLS means or donor draws, predictive mean matching (PMM), and microimpute's quantile regression forests, alone (QRF) and zero-inflated (ZI-QRF). We summarize the results in NOTE.md and give the full tables in TABLE.md.

Results

We use the file's 134,729 people in 55,401 SPM units and 53,344 households.

• With reported MOOP, we reproduce Census's published 2025 SPM rate (13.11%) and margins of error, for all people and each age group, to within 0.01 percentage point. Subtracting MOOP raises the rate by 2.25 points; Census reports 2.28.
• We estimate that logit plus cell means raises the national SPM rate by 0.30 points (SE 0.08), the OLS variant by 0.34, logit draws with donors by 0.39 and PMM by 0.32. QRF changes it by -0.07 (SE 0.07) and ZI-QRF by -0.05 (SE 0.07), though their MOOP totals fall 3.6% and 3.7% short of the reported total.
• For people whose resources before MOOP run 1.0 to 1.25 times their threshold, logit plus cell means raises the rate by 14.69 points, PMM by 4.19 and QRF by 0.55. Reported MOOP rises with resources ($2,385 on average in that band, $6,086 at 2 to 3 times the threshold), and cell means assign the band $3,189, PMM $2,799 and QRF $2,352.
• Logit draws with donors match the reported share at zero (6.5% against 6.4%) and 99th percentile ($33,147 against $33,640), yet raise the rate by 0.39 points. Cell means cut the 99th percentile to $15,585.
• Census imputes MOOP for 43% of the file's people and the Medicare premium for 13%, so part of our benchmark comes from its imputation. In the 44% of SPM units where it imputes neither for any member, the three logit methods and PMM raise the rate by 0.26 to 0.52 points. QRF and ZI-QRF change it by +0.01 (SE 0.09) and +0.07 (SE 0.10).

What the code does

• cps_asec.py downloads the Census zip, checks its pinned SHA-256 and re-extracts any CSV that differs from the zip. It merges household state and the 160 replicate weights and flags people whose MOOP or Medicare premium Census imputed. It then builds SPM-unit predictors an ACS-style file carries: ages, sex, disability, current coverage, cash income, family type, tenure and state.
• methods.py implements (a) logit plus cell or OLS means, (b) logit draws with donors, PMM within coverage groups and (c) microimpute QRF and ZeroInflatedImputer. It also implements the decomposition variants that cross the participation rule with the amount rule.
• run.py assigns whole households to five folds, imputes each fold from models it fits on the rest (ten draws for stochastic methods) and writes results/results.json. It caches each fold under a key that covers the settings, the experiment's and microimpute's source files, package versions and the model inputs.
• evaluate.py recomputes person-level SPM poverty for each method and computes rates, replicate-weight SEs, gross flows and MOOP distributions.
• render.py writes TABLE.md and NOTE.md from results.json and fails if a placeholder in NOTE.template.md has no value.

Invariants

tests/test_moop_spm_experiment.py checks these on synthetic data:

• With reported MOOP, recomputed poverty equals SPM_POOR for every person.
• Raising any unit's MOOP never moves anyone out of poverty, and people whose unit's resources before MOOP fall below its threshold stay poor under any nonnegative imputation.
• Folds partition households, and each household stays in one fold.
• Cell pools hold only same-cell donor spenders, and each recipient uses the finest cell with enough donors.
• Every logit method that uses the 0.5 cutoff gives zero to units below it, and no logit method imputes negative MOOP.
• logit_draw_hotdeck and logit_threshold_hotdeck impute positive amounts only from the recipient's own cell pool.
• For each recipient, the nearest-donor search returns min(k, number of donors) distinct donors, and their sorted distances match the smallest ones a brute-force search finds.
• Weighted group rates, the replicate SE formula and the ratio-band index match direct reference computations, and weighted quantiles with equal weights match NumPy's.
• allocation_flags marks a person's MOOP as Census-imputed when I_MOOP runs 1 to 3, and their Medicare premium when I_MCPREM equals 2.
• A new microimpute version or source file, or changed model inputs or row keys, change the cache key, so run.py refits every fold.
• extract_verified replaces any extracted CSV whose size or CRC-32 differs from the zip's member.

On every run, run.py also checks the identities that rebuild SPM_RESOURCES, SPM_POOR and SPM_MEDXPNS from their components. It raises if any method leaves a unit without an imputation or imputes negative MOOP.

Changes after review

GPT-6.1 Sol, reviewing da0144c, asked for these changes, which we make in 754e518:

• We now also leave units where Census imputed a Medicare premium (I_MCPREM of 2) out of the comparison among units with no imputed MOOP, since SPM_MEDXPNS includes the premium. The subset falls from 52% to 44% of SPM units. Logit plus cell means' change there moves from +0.65 to +0.36, the OLS variant's from +0.57 to +0.26, PMM's from +0.56 to +0.48 and QRF's from +0.10 to +0.01. In every record, I_MOOP flags an imputation if and only if I_PHIPVAL, I_POTCVAL or I_PMEDVAL does, so it already covers MOOP's three components.
• We now include microimpute's source files, package versions and the model inputs in the fold cache key, so run.py refits every fold when any of them changes.
• In the note, we describe PMM's log(1 + MOOP) regression across all donors, including zero spenders, and split a 47-word sentence.

Rerunning from an empty cache at 754e518, we reproduced every other number in results.json to full precision.

Reviewing 754e518, Sol found no correctness, claim or reproducibility problems. Sol asked us to drop the bold labels that opened the results above and to give the README's predictor sentence an actor. We make both changes in 2a92477 and rewrite the other label openers and reduced passives in the README, note and table. No number changes.

Reviewing e283c4c, Sol found no computation or reproducibility defect and asked for three prose fixes, which we make in 12fd398. We now state the nearest-donor invariant as its test checks it, with matching distances instead of identical donors. In the note, we credit María Juaristi with the microimpute paper's finding, and in this section we drop a reduced passive. We also give the verb to a person or program wherever a document or commit held it, and narrow two more invariants to what their tests check.

Reviewing 12fd398, Sol asked us to narrow the donor-draw and nearest-donor invariants to what their tests check and to name Sol as the actor in this section. We make those changes in this description, with no new commit, and Sol approved them in a follow-up review at the same head.

Two microimpute behaviors we leave unchanged

• ZeroInflatedImputer.fit passes weight_col only to the auxiliary imputer for non-numeric targets, so it fits numeric targets without weights. We fit every method without weights, so our results don't depend on it.
• QRF's stochastic predict maps each draw onto a 101-point quantile grid from 1/102 to 101/102, so no draw exceeds the 101/102 quantile of a unit's predicted distribution.

Other changes

• We add hypothesis to the dev extras for the property tests.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
microimpute-dashboard	Ready	Preview	Nov 6, 2026 4:47am UTC