Use disclosure-controlled UK geography outputs for constituency and local authority impacts #859

Problem

UK constituency and local-authority results in the economy output are not disclosure-controlled, and some of them are wrong in a way readers can see.

The executor builds them with policyengine.py's deprecated compute_uk_constituency_impacts and compute_longwise_uk_geography_impacts (simulation_output_geographic.py, build_uk_constituency_impact and build_uk_local_authority_impact). Per area, these compute:

• average_household_income_change = Σ(change × w) / Σw
• relative_household_income_change = Σ(reform × w) / Σ(baseline × w) − 1

There is no minimum sample. A small cell with a few negative-income households has a baseline total near zero or below zero, so the relative change explodes or changes sign.

Examples from production output

Reform: gov.hmrc.income_tax.rates.uk[0].rate 0.20 → 0.21, 2026.

• App report 4684, region country/wales. Cardiff West: average_household_income_change ≈ −192, but relative_household_income_change ≈ +13.72 (the app shows +1372%). A tax rise that loses £192 shows a large gain. Both values imply a baseline mean household income of about −£14. Caerphilly shows −5.6%, while neighbouring seats show −0.1% to −2.3%.
• App report 4693, UK-wide. Pendle and Clitheroe shows +1.8%, while every other English constituency loses.

First reported, in the wrong issue, at #858 (comment).

Licensing risk

The UK datasets derive from UK Data Service licensed survey microdata (FRS and others). The End User Licence binds every published number, and a local result that rests on a handful of survey households can breach it. policyengine.py deprecated the two functions above for this reason: their output "is not disclosure-controlled, so it must not be published when the data is licensed survey microdata". The executor still publishes it in every UK economy result, through the API and the app.

Fix

policyengine.py already has the replacement: compute_uk_geography_disclosure (PolicyEngine/policyengine.py#605, closing PolicyEngine/policyengine.py#589), available in the pinned 6.7.5. It applies a count floor of 10 distinct survey households (by source_household_id), a dominance rule, a check on what subtraction between published areas recovers, a bootstrap 90% interval and precision-adaptive rounding. Each cell is reported, approximate or withheld.

Switch both breakdowns to it:

1. Build constituencies and local authorities in one release, nested in regions and the nation, so the check sees both levels.
2. Resolve blocking findings by withholding the suggested cells; never withhold a reference (region or national) cell, because those numbers are public elsewhere.
3. Keep the output contract backward compatible: existing fields keep their types. A withheld area is left out of the list. Published areas gain optional fields for the tier and the 90% interval.
4. Keep source_household_id (and region_code_oa) in the Stage 12 retained household frame.

Known limits

• The enhanced FRS 1.58.1 in the current bundle sets source_household_id after its SPI and capital-gains copies are stacked, so copies of one survey household count separately (eFRS source_household_id doesn't name the survey household: SPI copies and capital-gains clones count as separate households policyengine-uk-data#591, fixed in Prebuild single-year datasets into the Modal simulation image (temporary) #597 for later releases). Support is overstated on that release until the bundle moves to a rebuilt eFRS or to Microcosm (Certify the Microcosm UK national and local-area pair with policyengine-uk 2.132.6 policyengine.py#685).
• The check covers one run. It does not compare results across runs, such as a UK run against a Wales run.

🤖 Generated with Claude Code