UK spine: child disability under-identification blocks the TCL disability cells (−56.5% / −40.0% across three spine vintages) #808

Activity

1. juaristi22 commented on Aug 28, 2026

   

   juaristi22

   on Aug 28, 2026

   CollaboratorAuthor

   More actions

   Diagnosis revision + measurement: the two blocking cells were mis-bound, and are now fixed and green — but the child-level identification gap is real, at ~4×.

   Part 1 — the binding. These two targets come from DWP sheet 04B, which counts every child in the affected household (the 3- and 4-children sibling rows divide to exactly 3.00 and 4.00 children per household; these cells run 3.6–3.7). Our bindings — uniquely among the sheet's children targets — summed uc_is_child_limit_affected (~1.27 affected children per household) against that all-children concept: a structural ~2.8× shortfall no disability flagging could close. #813 re-binds them to is_child like every sibling. v13 measurement (spine-j):

   cell	before	after	initial support
   children_disabled_child_element	−56.5%	+0.1%	72,118 → 190,414
   children_claimant_pip	−40.2%	−0.3%	72,943 → 182,933

   Both cells hold green across the #809 lever sweep (0.55/0.70/0.85). The TCL household cells and children_in_* stayed within 0.4%. These cells no longer block the #796 capstone.

   Part 2 — the identification audit (the issue's original concern, measured against the unused chronicle marginal uc_tcl_children_by_disability_child_dla_pip_or_blind = 436,780). On spine-l055 (current main + #813, design weights, proxy universe = children in would_claim_uc benunits with 3+ dependent children — if anything broader than DWP's affected_3plus_children universe, since it ignores exceptions):

   • children with their own DLA/PIP reported > 0: 109,031 weighted, 191 records — ~25% of DWP's 436,780
   • across all would_claim_uc benunits: 326,361 weighted, 629 records
   • DWP's implied prevalence among children in affected households: 436,780 / 1,665,540 ≈ 26%; ours ≈ 3.5%

   The under-identification is invariant to the binding fix: the calibration now hits the published counts by up-weighting the households that do carry flagged children (190k initial vs 479k target still solves), but 191 records carrying that mass is thin support, and any reform analysis conditioning on child DLA/PIP receipt inherits the 4× gap. The frs_disability stage flags a child only when the FRS benefits tape reports an actual DLA/PIP amount for that child — no age branch, no top-up.

   Proposed second increment (the fix the issue originally suggested, now scoped): an anchored-residual child-disability draw — assign_binary_with_anchored_residual keeps every tape-reported recipient true and tops up to a DWP-implied rate, conditioned on household receipt, with the rate translated from this same crosstab (26% among affected-household children) or from the DWP DLA/PIP child caseload statistics as the adjudicated source. Machinery, seed pattern, and test template all exist (stochastic_assignment.py:57-84, take_up_contract.json source blocks, test_uk_frs_stochastic.py). Needs a rate-source adjudication before implementation — the rate must come from a published DWP fact, not from tuning to fit.

   Keeping this issue open for that increment; it is a quality item, not a capstone blocker.
2. juaristi22

   added a commit that references this issue on Aug 28, 2026

   Merge pull request #813 from PolicyEngine/808-809-rebind

   ...

   e0de620
3. juaristi22

   closed this as completedin #813on Aug 28, 2026