Populace vs per-state ECPS parity: transfers agree, baseline child poverty diverges up to 11.7pp

[DTrim99]

Posted for review before any dataset migration decision (task: populace parity harness). Run completed 2026-07-01 on Modal: restore_2021_expanded_ctc computed for all 51 states on both datasets — the per-state calibrated ECPS files CPID uses in production, and the national populace_us_2024.h5 filtered to each state (per-state filtering to avoid the full-frame OOM).

TL;DR: do not switch CPID to populace yet. Demographics and reform dollar flows agree well (child counts mean 1.5% apart; 2021-CTC transfer totals mean 5.1% apart), but baseline child poverty levels diverge systematically — mean 5.7pp, max 11.7pp — with populace lower in every large-gap state. Switching would change every state's headline child-poverty rate, some by half (AL 23.5% → 11.8%). Since counts and transfers line up, this points at income/benefit imputation or calibration-target differences between the datasets rather than the state filtering itself.

Related evidence from the state-credit validation work (populace#295, incl. three-way cross-check): populace reproduces official state EITC outlays within ~10%, but under-6 credits (CA YCTC +38.7%, CO CTC +30.4%) run hot — a separate, smaller-magnitude issue.

cc @MaxGhenis @PavelMakarchuk

---

Populace vs per-state ECPS parity — 2021 CTC restoration, 2026

Test reform: restore_2021_expanded_ctc across all 51 states. Populace = national populace_us_2024.h5 filtered per state; ECPS = the 51 state-calibrated files CPID uses today.

Headline agreement

• Baseline child poverty rate: mean |gap| 5.67pp, median 5.12pp, max 11.70pp
• 2021-CTC child-poverty reduction: mean |gap| 1.22pp, median 0.97pp, max 5.15pp
• CTC transfer totals: mean |rel diff| 5.1%, max 19.1%
• Child counts: mean |rel diff| 1.5%, max 4.0%

Largest divergences (by baseline child-poverty gap)

State	Child pov (populace)	Child pov (ECPS)	Gap	dCTC pp (pop)	dCTC pp (ECPS)	Transfer pop	Transfer ECPS
AL	11.8%	23.5%	-11.7pp	-1.86	-4.64	$1.55B	$1.48B
AR	6.8%	17.6%	-10.8pp	-2.47	-3.51	$0.80B	$0.86B
AK	4.6%	15.2%	-10.6pp	-1.46	-2.43	$0.17B	$0.21B
WY	3.3%	13.7%	-10.4pp	-0.75	-1.90	$0.10B	$0.12B
SC	9.0%	18.8%	-9.8pp	-2.50	-3.04	$1.35B	$1.39B
DE	10.9%	20.6%	-9.7pp	-2.33	-3.97	$0.21B	$0.25B
MS	14.4%	23.7%	-9.3pp	-4.25	-4.50	$0.89B	$0.93B
NH	3.2%	12.5%	-9.3pp	-0.18	-2.51	$0.20B	$0.20B
TX	15.8%	24.9%	-9.2pp	-4.78	-4.10	$8.79B	$9.07B
TN	11.5%	20.6%	-9.1pp	-5.82	-3.72	$1.84B	$1.89B
GA	17.0%	25.7%	-8.7pp	-3.05	-4.06	$3.07B	$3.17B
ID	4.6%	13.1%	-8.5pp	-1.30	-2.96	$0.44B	$0.49B

Full 51-state table

State	Child pov pop/ECPS	dCTC pp pop/ECPS	Transfer pop/ECPS ($B)
AK	4.6% / 15.2%	-1.46 / -2.43	0.17 / 0.21
AL	11.8% / 23.5%	-1.86 / -4.64	1.55 / 1.48
AR	6.8% / 17.6%	-2.47 / -3.51	0.80 / 0.86
AZ	11.9% / 20.0%	-3.76 / -3.13	1.73 / 1.83
CA	22.8% / 23.9%	-4.83 / -3.40	8.86 / 8.94
CO	12.5% / 13.9%	-3.74 / -2.51	1.11 / 1.21
CT	10.5% / 15.1%	-3.15 / -2.27	0.73 / 0.70
DC	17.3% / 25.6%	-6.63 / -7.34	0.16 / 0.17
DE	10.9% / 20.6%	-2.33 / -3.97	0.21 / 0.25
FL	22.6% / 27.0%	-8.68 / -3.53	5.28 / 5.44
GA	17.0% / 25.7%	-3.05 / -4.06	3.07 / 3.17
HI	9.3% / 16.2%	-1.97 / -2.66	0.32 / 0.31
IA	4.6% / 10.6%	-0.76 / -1.69	0.72 / 0.75
ID	4.6% / 13.1%	-1.30 / -2.96	0.44 / 0.49
IL	12.5% / 15.2%	-4.61 / -2.51	2.81 / 2.77
IN	8.5% / 15.6%	-1.88 / -3.49	1.64 / 1.78
KS	16.6% / 15.3%	-5.96 / -2.94	0.82 / 0.76
KY	10.3% / 17.0%	-2.87 / -3.18	1.28 / 1.25
LA	17.0% / 20.8%	-4.28 / -4.48	1.43 / 1.37
MA	13.1% / 18.2%	-2.93 / -2.90	1.21 / 1.30
MD	13.1% / 19.7%	-2.86 / -3.11	1.25 / 1.37
ME	6.6% / 13.2%	-1.06 / -3.71	0.28 / 0.27
MI	10.3% / 12.9%	-4.03 / -2.35	2.34 / 2.35
MN	3.3% / 10.7%	-0.62 / -1.71	1.15 / 1.22
MO	12.2% / 15.6%	-2.84 / -2.72	1.56 / 1.55
MS	14.4% / 23.7%	-4.25 / -4.50	0.89 / 0.93
MT	10.0% / 13.7%	-3.86 / -1.85	0.23 / 0.26
NC	13.0% / 17.7%	-3.62 / -3.01	2.62 / 2.74
ND	6.9% / 9.4%	-2.95 / -0.90	0.18 / 0.16
NE	5.6% / 9.7%	-1.81 / -1.67	0.43 / 0.46
NH	3.2% / 12.5%	-0.18 / -2.51	0.20 / 0.20
NJ	12.2% / 18.3%	-3.58 / -3.09	1.82 / 1.93
NM	15.0% / 18.7%	-6.64 / -3.67	0.58 / 0.60
NV	15.7% / 23.1%	-3.37 / -4.51	0.78 / 0.80
NY	21.0% / 22.6%	-4.39 / -3.49	4.50 / 4.56
OH	12.1% / 14.7%	-3.67 / -2.54	3.00 / 2.91
OK	12.4% / 17.6%	-3.75 / -2.85	1.17 / 1.19
OR	19.0% / 19.7%	-2.05 / -3.36	0.88 / 0.90
PA	13.1% / 14.7%	-5.05 / -2.23	2.92 / 2.79
RI	12.7% / 16.6%	-3.23 / -3.02	0.22 / 0.22
SC	9.0% / 18.8%	-2.50 / -3.04	1.35 / 1.39
SD	7.5% / 13.0%	-1.48 / -1.49	0.22 / 0.22
TN	11.5% / 20.6%	-5.82 / -3.72	1.84 / 1.89
TX	15.8% / 24.9%	-4.78 / -4.10	8.79 / 9.07
UT	5.5% / 12.5%	-1.15 / -1.44	0.83 / 0.87
VA	13.8% / 17.5%	-2.01 / -2.37	1.94 / 1.91
VT	10.0% / 12.3%	-4.39 / -2.62	0.12 / 0.12
WA	12.0% / 16.2%	-2.45 / -3.09	1.42 / 1.60
WI	8.3% / 11.6%	-2.78 / -1.86	1.29 / 1.29
WV	13.7% / 18.1%	-3.69 / -2.95	0.44 / 0.45
WY	3.3% / 13.7%	-0.75 / -1.90	0.10 / 0.12

National CTC transfer: populace $79.7B vs ECPS $81.4B (-2.1%)

[DTrim99]

Addendum: anchoring to Census, not just to each other

Fair challenge from review: the tables above only show the two datasets disagreeing — they don't say which one is wrong. Anchoring the implied national rates to the official Census SPM ("Poverty in the United States: 2024", P60-287, Sept 2025):

	Overall SPM poverty	Child SPM poverty
Census 2024 (actual)	12.9%	13.4%
Populace (national file, 2026 sim)	13.3%	14.6%
Per-state ECPS files (2026 sim)	19.2%	19.5%

Populace is the one near the Census anchor (within ~0.4pp overall / ~1.2pp child, and 2026-vs-2024 timing explains some of that). The per-state ECPS files run ~6pp hot on both measures — so the state-level gaps in the tables above are better read as "the per-state files overstate poverty" than "populace understates it." That reframes the migration question: the concern is less about populace's levels and more that switching would drop CPID's headline poverty rates toward the Census-consistent level, a large but defensible change that needs to be communicated deliberately.

State-level anchoring is coming next: we're adding Census state SPM 3-year-average rates (overall + child) as out-of-sample validation rows in populace's reform-validation payload — same pattern as the state-credit outlays in PolicyEngine/populace#295 — so the calibration dashboard will show per-state poverty error for every build going forward.

Caveat on state comparisons: state SPM 3-year averages carry wide CPS margins of error for small states; treat single-state gaps under a few pp as noise.

[DTrim99]

Decision: CPID switched to Populace (2026-07-07)

PolicyEngine is standardizing all projects on Populace, and the Census anchoring above settled the levels question in its favor, so CPID made the switch (PR #56) — with the state-rate gaps acknowledged as work-in-progress in the rewritten methodology and tracked per release by the Census state SPM validation rows (populace#348).

Two implementation notes for anyone following:

1. The pinned revision is the L0-pruned 57k-household frame (revision 053baf6c, 2026-07-01) — not the 224k full frame the parity run above measured. Spot rates moved accordingly (e.g. CA child 22.8% full-frame → 17.15% pruned, vs Census 18.2% — better). Fresh per-state gaps on the pruned frame from slice verification: TX 13.7% (Census 15.8%) and ID 5.25% (5.8%) are close; NY 23.9% (15.5%) runs hot; WY 0.58% (6.4%) is implausibly low — the pattern the dashboard rows will make visible.

2. Runtime: a national Populace simulation costs ~5.3 min regardless of the requested state, so statewide requests initially ran 10–12 min. CPID now pre-slices each pinned revision into 51 per-state files (PR Pre-slice Populace into per-state files: statewide sims 10–12 min → 1–2 min #57), verified to reproduce the national state-masked values exactly — statewide sims dropped to 1–2 min (ID 707s→115s, CA 597s→70s) with identical results.