populace_us_2024 under-1 geography: CA carries 21.7% of US infants (~1.9x its population share), MI 1.0% (a third of its share)

Found while adjudicating analysis-notebooks#135 (MI under-age-1 reproducer). The state-file undercount that issue reported was fixed upstream (policyengine-us-data 1.115.4-patch promotion re-uploaded states/*.h5 on 2026-05-20; states/MI.h5 now totals 10.12M ≈ Census), but rerunning the reproducer surfaced a different, current defect on the populace side: the default national dataset's under-age-1 population has a badly miscalibrated state distribution.

Artifact

policyengine-us ≥1.78x DEFAULT_DATASET:
hf://datasets/policyengine/populace-us/populace_us_2024.h5@populace-us-2024-c86a631-6e1bcd0271a5-20260619T002242Z
(measured with policyengine-us 1.782.0, 2026-07-23).

Measured (2024 base year)

The 2026 layer has identical shares (CA 21.7%, MI 1.0%) — this is the base artifact's geography, not the multi-year projection.

from policyengine_us import Microsimulation
sim = Microsimulation()  # default = populace_us_2024 pin above
age = sim.calculate("age", 2024).values
w = sim.calculate("person_weight", 2024).values
sc = sim.calculate("state_code", 2024, map_to="person").values
# w[(age < 1) & (sc == "MI")].sum() -> 39,552 ; sc=="CA" -> 874,630

Why this is wrong

• National total is roughly right: 4.02M under-1 vs 3,596,017 US births in 2023 (NCHS, Births: Final Data for 2023) — +12%, fine at this altitude. The distribution is not.
• Implied crude birth rates: CA 22.3/1,000 vs MI 3.9/1,000 — a 5.7x spread across large states (in-dataset US average 11.8/1,000). Real state fertility differentials are tens of percent, not multiples.
• External anchor for MI: ~104K births/year (MDHHS, receipts in analysis-notebooks#135) vs 39.6K modeled under-1s (−62%).
• Cross-artifact disagreement: the us-data per-state file states/MI.h5 (1.115.4-patch) puts MI under-1 at 61,938 — itself low vs births, but 1.5x the populace national slice for the same state/year.

Downstream

Any under-1/infant-targeted analysis inherits 2–3x state-level errors: baby-bonus and newborn-credit scoring, under-1 poverty on the child-poverty dashboard (its state audit carries pe_pop_under_1 per state), and MT/other newborn provisions. CA-targeted infant analyses run ~2x hot; MI ~3x cold.

Question for the calibration surface

Is single-year age 0 (or any age-by-state cut finer than under-5) in the target surface for populace_us_2024? If under-1 is uncontrolled per state, this is the expected failure mode; a state-level under-1 (or births) target would pin it.

Reproducer notebook: us/poverty/michigan_under_age_1.ipynb in analysis-notebooks (PR link to follow in first comment).

🤖 Generated with Claude Code

Start by running the reproducer in us/poverty/michigan_under_age_1.ipynb and comparing the default populace_us_2024 output with the states/MI.h5 artifact. Then trace the DEFAULT_DATASET calibration surface to determine whether age-by-state targets are controlled; done should include an agreed correction or calibration decision for the under-1 state distribution.