Housing association rents ~20% above benchmark; housing/rent_social target missed by 1.55x

[vahid-ahmadi]

Warning

This body has been corrected. The original asserted social rents ran 25-35% high against an RSH "general needs" benchmark, and that scaling the housing element by -25% reproduced the administrative capped count. Both were wrong: the benchmark excluded affordable-rent tenancies, and the counterfactual cut private rents that are not overstated. Corrected figures below; the original text is preserved in the first comment.

Problem

Modelled rents in the enhanced FRS exceed published averages, with the error concentrated in the housing association sector. Council and private rents are close to benchmark.

tenure	model (2026)	EHS 2024-25 (England)	uprated ~12% to 2026	gap
council	GBP 145/wk (median 128)	GBP 120	~GBP 135	+7%
housing association	GBP 181/wk (median 161)	GBP 134	~GBP 151	+20%
social, blended	GBP 166/wk	GBP 129	~GBP 145	+14%
private	GBP 261/wk (median 225)	GBP 250	~GBP 275	−5%

Source: English Housing Survey 2024-25, chapter 2. EHS is England-only while the model is UK-wide; devolved social rents are lower, so the true gap is probably slightly larger than shown.

Comparability caveat that is not yet resolved: FRS HHRENT, the basis for the model's rent, is defined net of service charges and water, whereas EHS mean weekly rents for social tenants generally include service charges. That biases the model low in this comparison, which makes the +20% housing association gap harder to explain rather than easier. Quantifying this should be the first step before acting on the gap.

Enhanced FRS 2023/24 v1.40.3, simulated 2026.

The calibration target is being missed

This is the sharper statement of the problem. housing/rent_social already targets this quantity and is not converging, while the private target is:

target (2025)	model	target	ratio
housing/rent_private	GBP 92.9bn	GBP 89.0bn	1.04x
housing/rent_social	GBP 47.6bn	GBP 30.7bn	1.55x

The social miss decomposes as 5.77m households against the target's 5.0m basis (1.15x) times GBP 159/wk against the GBP 118/wk basis (1.35x).

Note the private figures cannot both be right: this issue puts model private rent 5% below the EHS-uprated benchmark, while the target basis (ONS PRHI, GBP 317/wk) puts it 19% below. ONS PRHI mixes new lets into a stock measure, so part of the private "count 30% high / mean 19% low" decomposition may be an artefact of the basis rather than model error. The two comparators need reconciling.

Why it matters

The benefit cap is a threshold effect and unusually sensitive to level errors in its inputs (PolicyEngine/policyengine-uk#1818). Housing is 46% of capped benefits, and the median excess over the cap is only GBP 2,733/yr against a mean housing element of GBP 11,604/yr.

Applying tenure-specific corrections (council −7%, HA −20%, private 0%) moves modelled capping from 222k households to 145k, against DWP's 111k — closing about 69% of the excess. The residual belongs to other causes tracked in #1818.

Downstream this makes benefit-cap-interacting reform estimates roughly 2x too large, blocking JRF-style protected minimum floor analysis (PolicyEngine/policyengine-uk#1814).

Household counts

	model (2025)	UK benchmark
total households	31.76m	29.0m (ONS 2025)
owner-occupied	59.9%	65% (FRS 2024-25)
private rent	22.0%	19%
social rent	18.2%	17%

The household total is ~9.5% high and renting is over-represented. A concrete candidate: targets/sources/ons_tenure.py is documented as ONS "subnational dwelling stock by tenure" but targets/compute/households.py matches it against modelled household counts, and dwellings exceed households by roughly 8% in England.

What would help

• Quantify the service-charge basis difference between FRS HHRENT and EHS rents before treating the +20% as pure model error.
• Diagnose why housing/rent_social does not converge while housing/rent_private does. This is the same class of failure as the benefit cap targets in #1818, which are missed by 2.0x and 3.1x — a shared cause is plausible.
• Reconcile the dwelling-stock vs household-count mismatch in the tenure targets.
• Pair rent totals with household-count constraints, or target mean rent by tenure, so count and mean errors cannot cancel. Note datasets/local_areas/constituencies/devolved_housing.py already does exactly this for Wales and Scotland.
• Add mean rent by tenure to per-release validation against EHS.

Related, found alongside

num_bedrooms is identically zero for every household in the dataset. The variable is declared in the model (variables/household/demographic/num_bedrooms.py) with no formula, nothing populates it, and nothing references it. Harmless today, but it makes the under-occupancy deduction (removal of the spare room subsidy) unimplementable — that policy is absent from the model entirely, and social renters receive full rent as their housing costs element.

[vahid-ahmadi]

Correction and a precise diagnosis

My original figures overstated the problem and used the wrong benchmark. Correcting on both counts.

The benchmark was wrong. I compared against RSH general needs social rent (GBP 113.69/wk), which excludes affordable-rent tenancies let at up to 80% of market. The right comparator is the English Housing Survey, which covers all social tenancies:

tenure	model (2026)	EHS 2024-25	uprated ~12% to 2026	gap
council	GBP 145/wk	GBP 120	~GBP 135	+7%
housing association	GBP 181/wk	GBP 134	~GBP 151	+20%
private	GBP 261/wk	GBP 250	~GBP 275	−5%

So it is a housing-association problem, not the blanket 25–35% social overstatement I claimed. Council rents are close and private rents are slightly low.

The 25% counterfactual was correspondingly wrong. Applying tenure-specific corrections (council −7%, HA −20%, private 0%) moves benefit cap capping from 222k to 145k, not to the 114k I reported. Rents explain roughly 60% of the excess over DWP's 111k, not all of it. The residual belongs to the unmodelled 9-month grace period, the missing exemptions listed in PolicyEngine/policyengine-uk#1818, and annualisation of a monthly test.

Also wrong: my service-charge hypothesis. FRS HHRENT is defined as "the rent eligible for Housing Benefit paid by a household ... after taking off extras such as service charges", with water and sewerage explicitly deducted (FRS derived variable specification, HHRENT). frs.py:1165 uses it correctly. There is no double count analogous to #449.

Where the problem actually is

Calibration hits the private rent target and misses the social one:

target (2025)	model	target	ratio
housing/rent_private	GBP 92.9bn	GBP 89.0bn	1.04x
housing/rent_social	GBP 47.6bn	GBP 30.7bn	1.55x

The social miss decomposes cleanly: 5.77m social renter households against the target's 5.0m basis (1.15x) times GBP 159/wk mean against the GBP 118/wk basis (1.35x) = 1.55x.

Two contributing defects in targets/sources/housing.py:

1. Inconsistent vintages for the same target year. Both targets are assigned to 2025, but private uses ONS PRHI from February 2026 while social uses EHS 2023-24 (GBP 118/wk). EHS 2024-25 has since published GBP 129/wk. The social target is roughly 10% stale before any calibration question.
2. The private target is hit for the wrong reasons. Its basis is 5.4m households at GBP 317/wk; the model has 6.99m at GBP 256/wk. Count and mean are both wrong by ~30% and ~19% in opposite directions, and the total matches by cancellation. A total-only target cannot detect this, so housing/rent_private is currently not validating anything.

Underlying: the household count is 12% high

	model (2025)	UK benchmark
total households	31.76m	~28.4m
owner-occupied	59.9%	~63%
private rent	22.0%	~19%
social rent	18.2%	~17%

Tenure shares are only modestly off; the totals inherit a 12% overshoot on households overall. That accounts for most of the social renter count excess (15% = 12% household overshoot x 3% share), and is likely worth its own issue.

Suggested fix

• Refresh the social rent basis to EHS 2024-25 and state the vintage of each target explicitly, so both housing targets are anchored to the same year.
• Pair the rent totals with household-count constraints, or target mean rent by tenure, so count and mean errors cannot cancel.
• Add mean rent by tenure to the per-release validation output against EHS.

Happy to raise the target changes as a PR; the calibration convergence question and the household count overshoot are separate and larger.