Skip to content

Bind England's PIP at 16–19 and 20–24, and impute UC limited capability for claims without a linked DWP record - #1663

Draft
juaristi22 wants to merge 5 commits into
mainfrom
uk-milburn-pip-split-lcwra
Draft

juaristi22 wants to merge 5 commits into
mainfrom
uk-milburn-pip-split-lcwra

Conversation

@juaristi22

@juaristi22 juaristi22 commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator

Why

The benchmarks check of the Milburn review dashboard (PolicyEngine/uk-milburn-review-dashboard#1) found two data gaps behind the young people figures on build fcfac709.

  • PIP. England's 16–24 band is calibrated and fits. But within it, 16–19 ran 17.6% high and 20–24 15.9% low, and the review's measures cut PIP at 18.
  • UC health element. Young people receiving the LCWRA element ran about a third above the Resolution Foundation's count, and nine in ten of them also got PIP, against the RF's 110,000 of 176,000.
    • Claims without a linked DWP record keep policyengine-uk's proxy. Those claims include every one the take-up redraw adds. The proxy gives LCWRA to every claimant on a disability benefit, and to no one else.
    • On linked FRS claims, DWP's own status disagrees with the proxy in both directions. At the spine's design weights, about 73% of linked claims the proxy flags hold DWP's LCWRA status, and about 11% of those it doesn't flag.

María's call of 17 November 2026: "do the pip age-split and lcwra in a single pr".

Changes

  1. PIP age split (uk/calibration_measure_exclusions.json). This now also lands on its own in Bind England's PIP cases at 16–19 and 20–24 in place of 16–24 #1671 (see Next).

    • dwp.pip.cases_england_aged_16_to_19 and dwp.pip.cases_england_aged_20_to_24 now bind.
    • dwp.pip.cases_england_aged_16_to_24 is held out as their sum, as dwp.uc.people_aged_16_to_24 is for its two bound parts.
    • England's eleven other published bands stay held out as checks. Their restated exclusion carries her call, dated 2026-11-17 and expiring with the family on 2027-02-15.
  2. LCWRA for claims without a linked record (uk_runtime/uc_limited_capability.py).

    • Each claim outside the SPI rows that has no linked record now takes an imputed status: LCWRA at the share of linked claims with that status in its cell, and NONE otherwise.
    • The share counts only linked claims in which no claimant has their own ESA award, so each rate is measured on the kind of claim that draws from it (the review's point 1).
    • A cell is the youngest claimant's age band (16–24, 25–49, or 50 and over) and whether the proxy flags any claimant.
    • Each claim draws from an identity-keyed uniform, and the linked rules then assign the flags, with the claim's first claimant as the lead.
    • A cell with fewer than three source households takes its proxy group's rate over every age.
    • Claims in which a claimant has their own ESA award keep the proxy, since the award carries DWP's determination.
    • The stage now asks the engine for has_own_esa_award too.
    • The receipt adds each cell's rate. Beside it is the rate of the own-ESA linked claims the cell leaves out. Each is printed only where its LCWRA claims and the rest each rest on three source households or more.
    • sources.yaml, its schema and source_stages.json carry the new parameters in lockstep.
    • The graph binds uc_take_up (the young ages), frs_take_up (the age source) and stochastic_assignment into the stage's code identity.
    • The release input coverage manifest records source_stages.json's new hash; nothing else in it moves.
  3. Tests and changelog. Six new stage tests, and the first stage test now covers an unlinked claim:

    • the rate each cell draws at;
    • own-ESA linked claims staying out of the rates;
    • own-ESA and SPI claims keeping the proxy;
    • a cell under three households drawing at its proxy group's rate over every age;
    • no rate at all;
    • who takes imputed LCWRA in a couple.

    Two changelog fragments.

Testing

  • Main's head. This branch merges main at f86134f11 (UK #931 prep: count floors in true sources, fold geographic clones in the FRS fits, ready the conditioned area draw #1519).
  • Engine-free tests, before the merge.
    • The stage's tests: 10 pass.
    • The 48 files that read the register, the stage, its manifest or the graph kernels: 3,341 tests, 23 skipped.
    • The run's 3 failures are fixed in these commits: the register's count pin, and the coverage manifest, regenerated (hash only).
  • After the merge.
    • The stage, the register, the manifest, the source stages, the country spec, the receipt counts and the support floor: 362 tests pass.
    • Every test file that main's floor commits change: 861 tests pass, 1 skipped.
    • The stage's engine tests: 3 pass.
  • After the review's fixes (d850c0aac).
    • The stage's tests: 12 pass.
    • The 44 engine-free files that read the stage, its spec or its manifests (38 UK files and the shared spec-engine files): 1,815 tests pass, 6 skipped.
    • The stage's engine tests: 3 pass.
    • The H2 fixture is regenerated, and only this stage's node parameters move.
  • Arms: two national arms on Chronicle 4ce631f5 at review date 2026-11-17. The control is main f86134f11, and the head is this branch's d850c0aac. Results below.

Results

Both arms ran the same recipe, on Chronicle 4ce631f5 at review date 2026-11-17. The control is on main f86134f11, and the head on this branch's d850c0aac. Figures are for 2025.

Health.

  • Both calibrations pass all seven seam gates.
  • The head's final loss is 0.00507, against the control's 0.00512.
  • In each arm, 98.8% of 1,534 rows fit within 10%. No bound row moves newly outside 10%.

PIP split. It works.

  • England's 16–19 band goes from +17.6% to 0.0%, and its 20–24 band from −15.9% to +0.1%.
  • The 16–24 band, now held out as their sum, fits at +0.1%.
  • PIP at 16–24 in England and Wales stays at 407,000.

LCWRA imputation. It brings the mix close to the Resolution Foundation's, but it raises the level.

  • People aged 16–24 in Great Britain receiving the LCWRA element: 233,000 in the control and 312,000 in the head, against the RF's 176,000.
  • The share of them also on PIP: 91% in the control and 66% in the head, against the RF's 63%.
  • The element for units whose claimant is under 25: £1.18bn in the control and £1.58bn in the head, against the review's £0.92bn.
  • The rise comes from the claims the take-up draw adds.
    • Young people with the element in those claims go from 87,000 to 172,000. Of the head's 172,000, 93,000 have no PIP.
    • The claims the proxy doesn't flag far outnumber the ones it flags. Drawing LCWRA for them at 8% to 19% adds more than the flagged side's lower rate removes.

Checks.

  • LCWRA households, a bound row: +1.5% in the control, +5.2% in the head.
  • DWP's UC regimes for young people (held out):
    • no work requirements: +26.5% → +51.9% at 16–21, and +4.0% → +18.4% at 22–24;
    • other regimes: −14.6% → −28.6% at 16–21, and −3.0% → −13.3% at 22–24.
  • NEET at 16–24 improves, from +4.3% to +2.6%.
  • PAYE employees aged 18–24 don't move (−10.4% → −10.5%).

The rates each cell draws at. These are measured on linked claims without their own ESA award, at the spine's design weights.

  • Where the proxy flags a claimant: 66% at 16–24, 67% at 25–49, and 80% at 50 and over.
  • Where it flags no one: 8%, 9% and 19%.
  • The receipt withholds the own-ESA claims' rates in every cell under the three-household floor.
  • The rates with those claims added aren't shown either: beside these, they would give the withheld rates back.

Certification. The criteria fixed before the results asked that young LCWRA move toward 176,000. It moves away, so the head was not certified.

Next

  • The PIP split lands on its own in Bind England's PIP cases at 16–19 and 20–24 in place of 16–24 #1671. That PR has its own arm on its head, and a certification against the current national build. María's call of 17 November 2026, after these results, was "Yes, split it".
  • This PR stays a draft for the LCWRA imputation. Its draw is to be revised before the arms run again. The results above show where the level rises: the claims the take-up draw adds, where the proxy flags no one.

Not in this pull request

Disclosure

Every figure here is national and rests on 30 or more survey households. The new receipt prints a cell's rate only where its LCWRA claims and the rest each rest on three source households or more.

🤖 Generated with Claude Code

juaristi22 and others added 4 commits November 17, 2026 15:34
María's call of 17 November 2026 ("do the pip age-split and lcwra in a
single pr"), after the Milburn review's benchmarks check
(PolicyEngine/uk-milburn-review-dashboard#1). The review's measures cut
PIP at 18, inside the 16 to 24 band, and on build fcfac709 the split
within it ran 17.6% high at 16 to 19 and 15.9% low at 20 to 24 while the
band itself fitted.

The two published bands bind; the 16 to 24 band is held out as their sum,
as dwp.uc.people_aged_16_to_24 is for its two bound parts. England's
eleven other published bands stay held out as checks on the split within
their broad bands; their restated exclusion carries her call, dated
2026-11-17 and expiring with its family on 2027-02-15.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
María's call of 17 November 2026 ("do the pip age-split and lcwra in a
single pr"), after the Milburn review's benchmarks check
(PolicyEngine/uk-milburn-review-dashboard#1). policyengine-uk's proxy
gives LCWRA to every claimant on a disability benefit and to no one
else, while on linked FRS claims DWP's status disagrees with it in both
directions; so the take-up redraw's entrants carried the element only
where someone got PIP, DLA or ESA, and on fcfac709 nine in ten young
people with the element got PIP, against the Resolution Foundation's
110,000 of 176,000.

uc_limited_capability now gives every claim outside the SPI rows that has
no linked record an imputed status: LCWRA at the share of linked claims
with that status in its cell (the youngest claimant's age band, 16 to 24,
25 to 49 or 50 and over, and whether the proxy flags any claimant), at
the stage's household weights, from an identity-keyed uniform; NONE
otherwise. The linked rules then assign the flags, with the claim's first
claimant as the lead. A cell under three source households takes its
proxy group's rate over every age; claims with a claimant's own ESA award
keep the proxy. The receipt adds the cells' rates, printed only where the
LCWRA claims and the rest each rest on three source households or more.

The stage asks the engine for has_own_esa_award too. sources.yaml, its
schema and source_stages.json carry the new parameters in lockstep, the
graph binds uc_take_up (the young ages), frs_take_up (the age source)
and stochastic_assignment into the stage's code identity, and the
release input coverage manifest records source_stages.json's new hash.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
María asked that the pull request target main's head. #1519 (#931's
geography-first prep) merged at 15:33Z; its 22 commits change the UK
build. The one conflict, the release input coverage manifest's
source_stages.json hashes, is resolved by regenerating the manifest on
the merged tree (tools/build_uk_release_input_coverage_manifest.py):
it differs from main's only by that hash.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
tools/graph_uk_spine_fixture.py on the merged tree: only the
uc_limited_capability node's parameters move (the has_own_esa_award
predictor, the age bands, seed, salt and imputation rule, and the
restated scope and notes).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

@vahid-ahmadi vahid-ahmadi left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for this. The cell logic, the disclosure rule and the register change all read cleanly. The stage tests and test_uk_measure_simulation.py pass at b5045212 (77 passed). Two points:

1. Own-ESA claims are in the source rates but not in the targets (medium).

  • uc_limited_capability.py:451-452 puts every linked claim with a claimant into cell_rows, including claims where a claimant has their own ESA award. The imputation at :515-518 then excludes own-ESA claims from the claims that draw.
  • This matches the rule as written ("the share of linked claims"), so it is a question of population alignment rather than a code slip.
  • If own-ESA claims hold LCWRA at a different rate from the rest of their cell, each cell's rate is partly measured on claims that never draw from it. This applies on both the proxy_flags and the proxy_does_not side.
  • A synthetic check at this head shows the mechanism. One 25–49 proxy_flags cell with 3 linked own-ESA LCWRA claims, 3 linked non-ESA NONE claims and 40 unlinked non-ESA flagged claims gives a cell rate of 0.5, and 18 of the 40 take LCWRA, although the non-ESA linked rate is 0.
  • Could the receipt or the PR description show the cell rates with and without own-ESA linked claims? If the gap is material, excluding them from the cell and pooled rates would align source and target. If they are kept on purpose, a line in the imputation rule saying why would help.

2. The pooled-rate draw is not exercised (low).

test_unlinked_claims_draw_their_cells_linked_rate asserts only the "pooled" source label for the 50+ cells, and no imputed claim falls in a pooled cell. A few unlinked claims aged 50+, checked against the proxy side's all-ages rate, would cover the fallback at :488-489.

Also checked, with no issues found:

  • household-weighted rates counted once per claim, with SPI excluded;
  • consistent cell assignment between source and targets, at survey-year age;
  • the <3 rule counted in distinct source households;
  • deterministic, salt-separated identity draws;
  • the first-claimant/carer rule matching the linked rules;
  • the receipt printing a rate only with ≥3 households on both sides;
  • the PIP register: 16–24 held out, 16–19 and 20–24 bound, eleven bands restated to 2026-11-17/2027-02-15, pin 142;
  • the code-identity binding.

The engine tests and arms are noted as still to run.

Review drafted with AI assistance and checked against the code at b5045212. uc_limited_capability.py is unchanged at de21892 (main merge and H2 fixture only).

Linked claims in which a claimant has their own ESA award now stay out of
uc_limited_capability's cell and pooled rates, as such claims stay out of the
draw: a rate measured on them would speak for claims that never draw from it
(the review of microcosm#1663, point 1). Beside each cell's rate and each
pooled rate, the receipt gives the rate of the own-ESA linked claims it leaves
out, under the same three-household rule, so every build shows the gap.

The imputation rule, the stage's notes and the changelog fragment say why.
sources.yaml and source_stages.json carry the new text, the release input
coverage manifest its new hash, and the H2 spine fixture the stage node's new
parameters (no other node moves).

Two stage tests: own-ESA linked claims staying out of the rates, and a cell
under three households drawing at its proxy group's rate over every age
(point 2).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@juaristi22

Copy link
Copy Markdown
Collaborator Author

Thanks, Vahid. Both points are addressed in d850c0aac.

1. Own-ESA claims in the source rates. Agreed: a rate should be measured on the kind of claim that draws from it.

  • Linked claims in which a claimant has their own ESA award now stay out of the cell and pooled rates. They are left out by the same test that keeps such claims out of the draw.
  • Beside each cell's rate and each pooled rate, the receipt now gives the rate of the own-ESA linked claims the cell leaves out (own_esa_award_claims). The same rule applies: a rate is printed only where its LCWRA claims and the rest each rest on three source households or more. So every build shows the gap.
  • The imputation rule in sources.yaml and source_stages.json, the stage's notes and the changelog fragment say why.
  • New test, test_linked_claims_with_their_own_esa_award_stay_out_of_the_rates, in the shape of your synthetic check. The claims without an award set the cell's rate at a third, where every linked claim together would give 0.4. Some of the unlinked claims' draws fall between the two, so the test tells the rates apart.

2. The pooled draw. New test, test_a_cell_under_three_households_draws_at_its_groups_rate_over_every_age.

  • Unlinked claims aged 50 and over sit in a cell that rests on one household, so they draw at the proxy side's rate over every age, 0.4.
  • Their draws fall on both sides of 0.4, and between it and the young cell's 0.5. So the test tells the pooled rate from the cell rates.

Checks at d850c0aac:

  • The stage's 12 tests pass.
  • 1,815 tests pass, 6 skipped. They cover the 44 engine-free files that read the stage, its spec or its manifests: 38 UK files and the shared spec-engine files.
  • The stage's 3 engine tests pass.
  • The H2 fixture is regenerated, and only this stage's node parameters move.

The arms now run from this commit: the control on main f86134f11, and the head on d850c0aac. Each cell's rate, with and without the own-ESA claims, will go in the description with the results.

@vahid-ahmadi vahid-ahmadi left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks, María. Both points are addressed at d850c0a.

  1. Own-ESA claims and the rates. Linked claims with an own ESA award now go to own_esa_rows (uc_limited_capability.py:457-465). They are picked out by the same own_esa[rows].any() test that keeps such claims out of the draw, so the cell and pooled rates are measured only on the kind of claim that draws. The receipt's own_esa_award_claims entry, beside each cell and pooled rate, goes through the same measured_rate. That means the same three-household rule on both sides applies before anything prints, and no combined rate is printed that could be differenced. The rule text in sources.yaml, source_stages.json and the module constant says the same.
  2. The pooled draw. test_a_cell_under_three_households_draws_at_its_groups_rate_over_every_age now drives the 50+ cell's draws against the pooled 0.4 (four of ten flagged linked claims). The draws fall on both sides of 0.4 and between it and the young cell's 0.5, so the test separates the pooled rate from the cell rates. test_linked_claims_with_their_own_esa_award_stay_out_of_the_rates does the same for point 1, separating 1/3 from 0.4.

CI is green at this head, engine-uk included. Nothing further from me. The arms' per-cell rates, with and without the own-ESA claims, will be useful to see when they land.

Review drafted with AI assistance and checked against the code at d850c0a.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants