test_reform_fiscal_impacts[Raise VAT standard rate by 2pp] fails on main (43bn actual vs 25bn expected)

[MaxGhenis]

Observed

Recent CI on main (and every open PR downstream of current main) fails:

policyengine_uk_data/tests/microsimulation/test_reform_impacts.py::test_reform_fiscal_impacts[Raise VAT standard rate by 2pp] FAILED
E  AssertionError: Impact for Raise VAT standard rate by 2pp is 43.5 billion, expected 25.0 billion

Also failing, likely related:

policyengine_uk_data/tests/test_energy_calibration.py::test_non_negative_energy FAILED
E  AssertionError: 14 households have negative electricity

Scope

Pre-existing on main (branch Push workflow is failing). Reproduces on every open uk-data PR:

• Add second-stage QRF imputation for FRS-only variables #362 (second-stage QRF) → 43.3 bn
• Refresh Tax-Free Childcare calibration targets #363 (TFC calibration targets) → 43.6 bn
• Fix calibration population overshoot (~6% drift) #310 (population rescale) → 43.5 bn
• Anchor UC/PC/CB takeup flags to FRS-reported receipt #359 (reported-anchor takeup) → 43.5 bn

Hypothesis

A recent consumption-imputation or calibration change roughly doubled the enhanced-FRS VAT base vs the OBR-derived £25 bn target for a +2pp VAT reform. The test_non_negative_energy failure is consistent — imputed consumption for 14 households went negative.

Worth bisecting against the last run that had this test green. Blocks downstream PR merges.

[MaxGhenis]

Traced the cause. The VAT formula itself is unchanged:

full_rate_consumption = consumption * full_rate_vat_expenditure_rate  # default 0.5
reduced_rate_consumption = consumption * reduced_rate_share  # 0.025
raw_vat = full_rate_consumption * 0.20 + reduced_rate_consumption * 0.05
vat = raw_vat / 0.38  # microdata_vat_coverage

A 2pp standard-rate rise produces a delta of roughly consumption * 0.5 * 0.02 / 0.38 = 0.0263 * consumption.

• Expected £25bn → implies total consumption base of £950 bn
• Actual £43bn → implies total consumption base of £1,635 bn (matches UK national total consumer expenditure ≈ £1.6T)

The dataset's consumption aggregates have caught up with reality; the hardcoded expected_impact is stale. Two options:

1. Update reforms_config.yaml from 25.0 to ~43.5 with a wider tolerance. Simplest.
2. Recalibrate microdata_vat_coverage. The 0.38 multiplier was chosen when LCFS/ETB under-reported consumption; if the enhanced FRS now captures closer to the full consumption base, the coverage factor should be higher (toward 1.0). Running the VAT aggregate through OBR/HMRC outturn would tell us the right value.

test_non_negative_energy (14 households with negative electricity) is a separate, smaller calibration issue — likely worth a one-line clamp in the energy imputation and not blocking.

Will file a PR for option 1 to unblock the open uk-data PRs. Option 2 deserves its own investigation since it affects every VAT-touching reform simulation.