US PUF imputation predicts each income item from the recipient's own survey value, so the PUF clone half reproduces ASEC income

[MaxGhenis]

Summary

The PUF tax-detail imputation predicts each major income item from the recipient's own survey value of that same item. The model therefore learns "output ≈ input", and the PUF-clone half of the pool comes out as ASEC income with PUF-style variation around it. This caps the top of the income distribution at ASEC topcodes and carries survey underreporting into the half of the pool that is supposed to carry tax-return income.

Found while diagnosing #958. Read in code at origin/main 2b85b7b22 and measured on the 2026-09-12 base-build checkpoints.

Mechanism (read in code)

• The predictor set is filing status, tax-unit size, and six income items: wages, self-employment, taxable interest, dividends, short-term gains, long-term gains (us_runtime/puf_support.py:199-208).
• The same items are imputation targets (PUF_TAX_DETAIL_DEFAULT_PERSON_OUTPUTS, puf_support.py:210).
• On the donor side, each predictor column is a copy of the donor's own PUF amount (_add_predictor_aliases on the donor tax units, puf_support.py:1664 and :1769, via _predictor_source_column). So in training, the wage predictor and the wage target are the same number.
• On the recipient side, the predictors are the household's survey amounts (_strict_recipient_predictor_surface, puf_support.py:2693; ASEC wages come from WSAL_VAL, cps_carried.py:172).

Measured: PUF clone vs its own ASEC parent

432,523 PUF-clone persons compared with their clone-0 parent, after 004_qrf_finalization. Before the QRF (002_clone_feature_extraction) every clone value equals its parent exactly; after it, none of the positive values are exact copies, but they stay next to the survey value.

Item	Zero/nonzero status agrees with ASEC	Rank correlation with ASEC	Within 10% of the ASEC value, among ASEC > 0
Wages	100.0%	1.000	100.0%
Self-employment	99.8%	0.979	97.6%
Taxable interest	100.0%	0.999	99.9%
Dividends	99.0%	0.961	82.1%
Long-term gains	100.0%	1.000	90.7%
Short-term gains	100.0%	1.000	78.8%

Consequences:

1. The top is capped at survey topcodes. No clone's imputed wages exceed the ASEC wage maximum ($2,099,999, the WSAL_VAL topcode) by more than a few percent, while the processed PUF has about 25,000 weighted returns above $10M of income. This is the root cause behind the QRF row of the stage table in US top-income tail is missing above $5M of AGI: Table 1.1 size-of-AGI facts never bind, and the pool's tail stratum carries capital gains only #958 (18 of 231,007 clones at or above $5M).
2. Survey underreporting carries into the PUF half. A household that reports no dividends, interest or capital gains in the ASEC gets none in its PUF clone 99-100% of the time. Calibration then has to hit SOI income totals with records whose incomes keep the survey's shape. The over-weighting of $200k-$2M returns measured in US top-income tail is missing above $5M of AGI: Table 1.1 size-of-AGI facts never bind, and the pool's tail stratum carries capital gains only #958 is consistent with this, but that link is a hypothesis, not measured.

History

• The archived eCPS imputed PUF income onto its CPS clone from demographic predictors only: age, sex, joint filing, dependent count, and tax-unit roles (policyengine-us-data, calibration/puf_impute.py, DEMOGRAPHIC_PREDICTORS), with a stratified PUF subsample that preserves the top percentile.
• Microcosm's first PUF commit (63b818730, 2026-06-19) introduced the income predictors. I found no design note in DESIGN.md or docs/ explaining the change.

Status elsewhere

• The graph-native path does not change this. The native branches (e.g. 516e6e8fe, b2d7a57c1) reuse the same tuple in full_puf_enrichment.py (PREDICTORS = tuple(support.PUF_TAX_DETAIL_DEFAULT_PREDICTORS); the PUF59 variant keeps the six income predictors). Their native_puf_tail.py ports the Carry PUF high-AGI donors into the own-tail stratum with their full income vector #964 selection, bounded to invented fixtures.
• Carry PUF high-AGI donors into the own-tail stratum with their full income vector #964 bypasses it only at the top. It copies whole donor records for proxy AGI of $5M or more. Everything below that still comes through the self-predicting imputation.

Fix options, for decision

1. Demographic predictors, as in the eCPS. The PUF half then carries the PUF's own joint income distribution attached to survey demographics.
2. Rank bridge. Condition on the household's survey income rank instead of income levels, and draw the PUF value at the matching rank, so top-ranked survey households can receive PUF-scale values (compare the UK rank-preserving approach in UK spine: composition-faithful SPI income draw — bridge predictors and rank-preserving replacement (with-children UC residual) #840).
3. Drop each target from its own predictors while keeping the other income items. Smaller change, but the other items are highly correlated with it, so it may not break the identity by much.

Each option changes the whole PUF half and needs a base build and a scorecard comparison against the current pool. Option 1 is the known design from the eCPS and is the natural first comparison.

[MaxGhenis]

Fix is in #1033.

The PUF QRF now conditions on demographics, the unit's weighted mid-rank income share within its own population, and an earnings-participation flag. No survey income level is a predictor any more. The flag is there because the PUF covers filers only: without it, the rank alone gave PUF wages to 48% of survey units with no earnings.

The table below compares the two designs before calibration, on the 2026-09-12 inputs at 8 trees. Receipts are in experiments/982-puf-self-prediction/ on the branch.

Measure	Old	New
Rank correlation, imputed vs survey wages	1.000	0.906
$5M–10M returns (SOI Table 1.1: 49,262)	23,600	48,629
$10M+ returns (SOI: 30,382)	8,186	33,920
Capital gains share of $10M+ AGI (SOI Table 1.4: 39.5%)	0.1%	42.6%
Survey non-earners given earnings	0	7 of 67,908 units

State top-tail totals are still noisy. In Colorado, one survey-weighted record carries 84% of the pre-calibration income above $1M, so #940's state × AGI-band amount targets are still needed.

[MaxGhenis]

Correction to my earlier numbers. The top-tail figures I posted weighted the PUF-clone half ×2, "as if this half were the whole population", and compared that with SOI. The real pool gives the survey half the other 50% of the weight, and the survey half has no units at $5M or more. Both halves at frame weight, before calibration and before main's capital-gains tail stage (8 trees, 15-item proxy AGI):

Band	Old design	#1033	SOI Table 1.1
$1M–1.5M returns	701k (1.90×)	483k (1.31×)	369k
$5M–10M returns	11.8k (0.24×)	24.3k (0.49×)	49.3k
$10M+ returns	4.1k (0.13×)	17.0k (0.56×)	30.4k
$10M+ AGI	$92B	$519B	$908B
Colorado income above $1M	$20.2B	$66.3B (2.1×)	$31.7B

#1033 roughly quadruples the pool's mass at $5M+, but on its own it reaches about half of SOI there, not SOI. Colorado is about 2× SOI above $1M, not 4×. One record still carries 81% of that. The script and output are at ~/PolicyEngine/_recovered/lane958-958-964-982-archive-20260926/982/actual_pool_bands.{py,json}. A follow-up will add them to experiments/982-puf-self-prediction/.