Model methodology · LAD only

Coverage-adjusted whole-stock Retrofit Gap

This is a modelled companion to the observed Retrofit Gap. It estimates how the below-C share could change if the rated sample had the published type, age and tenure composition of each local authority. It does not replace the observed figure.

Method lad-whole-stock-calibration-1.0.0 · EPC snapshot 2026-06-30 · May 2026 LAD codes

318

LADs published with explicit estimate state.

318

LADs with a calibrated estimate.

65.8%

Median rated-stock coverage of official stock.

What extra value it adds

The observed metric answers: “What share of homes with a valid EPC is below C?” The companion tests a narrower second question: “What would that share be after adjusting the rated sample to the area's published housing mix?” This exposes where EPC coverage is compositionally skewed without inventing ratings for individual unrated homes.

Publishing both figures shows the size and direction of the adjustment. Coverage, effective sample size, convergence, weight trimming and uncertainty are included so a large adjustment cannot be mistaken for an observed fact.

What it does not add

It does not assign an EPC band, SAP score or retrofit route to an unrated address. It does not estimate constituencies or postcodes. Its interval covers resampling variation within rated homes, not unknown selection bias among homes that have never been certified.

OS Open UPRN counts all addressable locations and lifecycle states. It is displayed only as a diagnostic and is neither the dwelling denominator nor an input to the estimate.

Calculation

  1. Keep the latest valid certificate per UPRN; retain blank-UPRN certificates as distinct dwellings.
  2. Harmonise EPC and VOA records to five dwelling types and common construction periods; harmonise Census tenure to owner, private rented, social rented and unknown.
  3. Pool sparse or absent cells into “other”, then iteratively calibrate separate type, period and tenure margins.
  4. Cap dwelling weights at 0.2–5.0 and require convergence within 100 iterations. If tenure prevents convergence, retry type plus period and record the fallback.
  5. Bootstrap the calibrated LAD sample to produce a 95% sampling interval.

0 LADs used the documented type-plus-period fallback in this release. Failed fits are published as unavailable, never filled with the observed value.

Release review

Calibration raised the estimated gap in 318 LADs, left 0 unchanged and lowered it in 0. The median adjustment is +6.5 percentage points; the 95th percentile absolute adjustment is 9.7 points.

14 LADs move by at least 10 points. The nation medians are +6.6points in England and +5.3 in Wales. Because the direction is systematic and the interval excludes never-certified-home selection bias, this release stays in methodology/download review and is not a map layer.

Largest model adjustments for review

Observed fact and model estimate remain separate columns.

Download all LADs (CSV)
Local authorityCoverageObservedAdjusted estimate95% sampling interval
Mid Suffolk65.4%45.5%58.3%58.0%–58.8%
Dartford69.6%37.7%48.7%48.3%–49.1%
Ribble Valley63.5%51.7%62.7%62.0%–63.2%
South Derbyshire68.8%37.9%48.8%48.3%–49.1%
Tewkesbury67.3%38.2%49.1%48.6%–49.6%
Uttlesford65.8%41.1%51.9%51.5%–52.4%
Vale of White Horse70.8%37.0%47.6%47.2%–48.0%
South Oxfordshire66.1%44.9%55.3%54.9%–55.7%
Rugby66.6%41.5%51.9%51.4%–52.2%
West Oxfordshire65.7%42.7%53.0%52.5%–53.5%
Wokingham64.4%40.5%50.6%50.2%–51.0%
Tonbridge and Malling63.3%43.5%53.6%53.1%–54.0%

Official inputs