Document UC deductions review findings: floor scope, validated years, bias direction

changelog.d/uc-deductions-review-notes.changed.md

@@ -0,0 +1 @@
+UC deductions documentation: promote the aggregate caveat - reform aggregates (poverty counts, floor-reform costs) run roughly 40% low because the model's UC caseload falls short of administrative counts (policyengine-uk-data#452 tracks the calibration fix; #450 tracks moving deduction imputation to the dataset build); per-household statistics are the validated, quotable layer. Also note the rate distribution is national (regional factors scale incidence only, so regional composition differences are not modeled) and that annual per-household amounts are upper bounds for spell-limited deduction types.

policyengine_uk/parameters/gov/simulation/uc_deductions/latent_rate_distribution.yaml

@@ -1,4 +1,4 @@
-description: Distribution of latent Universal Credit deduction demand, as a share of all UC claims in each deduction-rate band (rates as a fraction of the standard allowance). Simple average of the observed March to May 2025 distribution (Table 2), the last full months under the 25% cap, rounded to three decimals; the cap parameter reproduces observed post-cap distributions when applied. Bands sum to the share of UC claims with any deduction (0.467). Observed 2025 and applied to all model years as a constancy assumption (published deductions statistics begin March 2025). Treating the at-25% pileup as latent demand of at least 25%, and observations above 25% as above-cap deductions (last resort and child maintenance categories) with a 30% modeled upper edge, are modeling assumptions, not source facts. Validated for 2024 (25% cap regime) and 2025 (15% cap regime) only - earlier years apply this 2025-shaped demand distribution under the higher caps then in force (40% to October 2019, 30% to April 2021), which the published statistics cannot check. Demand is held fixed when the cap changes, so reforms that lower the cap below 15% overstate the number of constrained households and hence the gains from raising it back, while reforms that raise the cap toward 25% sit inside the observed window and are on firmer ground.
+description: Distribution of latent Universal Credit deduction demand, as a share of all UC claims in each deduction-rate band (rates as a fraction of the standard allowance). Simple average of the observed March to May 2025 distribution (Table 2), the last full months under the 25% cap, rounded to three decimals; the cap parameter reproduces observed post-cap distributions when applied. Bands sum to the share of UC claims with any deduction (0.467). Observed 2025 and applied to all model years as a constancy assumption (published deductions statistics begin March 2025). Treating the at-25% pileup as latent demand of at least 25%, and observations above 25% as above-cap deductions (last resort and child maintenance categories) with a 30% modeled upper edge, are modeling assumptions, not source facts. Validated for 2024 (25% cap regime) and 2025 (15% cap regime) only - earlier years apply this 2025-shaped demand distribution under the higher caps then in force (40% to October 2019, 30% to April 2021), which the published statistics cannot check. Demand is held fixed when the cap changes, so reforms that lower the cap below 15% overstate the number of constrained households and hence the gains from raising it back, while reforms that raise the cap toward 25% sit inside the observed window and are on firmer ground. The rate distribution is national - regional factors scale incidence only, so regional composition differences (e.g. advance shares) are not modeled.
 UNDER_5:
   description: Share of UC households with deductions between 0% and 5% of the standard allowance.
   values:

policyengine_uk/variables/gov/dwp/universal_credit/deductions/uc_deductions.py

@@ -11,7 +11,9 @@ class uc_deductions(Variable):
         "SI 2013/380). Per-household statistics are validated against the "
         "DWP deductions statistics; weighted aggregates are not, because "
         "they scale with the model's UC caseload, which falls short of the "
-        "administrative count."
+        "administrative count. Annualized from the surveyed monthly rate: "
+        "for deduction types with spells shorter than a year (e.g. advance "
+        "recovery), annual per-household amounts are upper bounds."
     )
     entity = BenUnit
     definition_period = YEAR