Corrections
Every correction published on this site, collected from the pieces that carry them. Each says what was wrong, what changed, and what still stands. The corrected text stays on the original page; nothing is rewritten silently.
36 findings published · 6 carry a correction · 10 corrections in total · model v8.0
Birmingham: 43% White British today, projected under 15% by 2051
This piece credited the projection to a fertility gap, Pakistani TFR 2.52 against White British 1.31, and said the true outcome "sits between" the two models. Neither holds. The model does not use total fertility rates: it runs on Census-derived child-woman ratios, which fold fertility, age structure and migration into one observed ratio per cohort, so the TFRs are context from ONS births data and not an input. And two models agreeing does not bracket the truth, because both can share a bias. Here they did: the v8.0 recalibration on an out-of-sample test moved Birmingham's 2051 share from 10.7% to 14.9%, outside the range the earlier text implied. The Census figures, the direction of travel and the 2051 projection stand.
The most deprived fifth of England records 25.5% of births. The least deprived records 15.4%.
This piece was headlined "Women in the most deprived areas have 66% more births. Poverty drives fertility, not just ethnicity," and gave 1.66x as a fertility ratio. It is not one. The figure is 144,522 births against 87,216, and ONS Table 8 publishes no count of women aged 15-44 per decile to divide either number by. A birth count is not a birth rate without the population at risk underneath it, which is the same rule this site applies to asylum support and did not apply here. The gradient is real and the confounding argument survives; the causal headline and the rate claim do not.
Corrections: the privacy policy on analytics, the Birmingham finding on fertility, and a corrections page
The privacy policy said this site uses no analytics and loads no third-party scripts. Both statements were false: every page has carried the Cloudflare Web Analytics beacon, a cookieless script that reports aggregate page views. The policy now says so, names the postcode lookup that sends a typed postcode to postcodes.io, and adds the retention statement it lacked. The footer no longer says no personal data is collected, since the hosts process IP addresses to serve pages; it now says no cookies and no advertising or behavioural tracking. The footer's claim that every number on the site links to its source was wider than the implementation, where sourcing sits at section level and on the sources page, so it now says every headline figure is traceable to its source and methodology. The accessibility statement now names WCAG 2.2 AA as the target and lists the target-size and reflow checks still to do. The Birmingham finding credited its projection to a fertility-rate gap the model does not use, since it runs on Census child-woman ratios, and said the truth sits between two models that had in fact shared a bias; both sentences are corrected on the page. There is now a corrections page, /corrections/, generated from every dated correction in the findings and the diary, so the record is in one place and cannot drift from the pieces that carry it. Ten duplicate redirect entries that Astro warned would become a hard error are reduced to one per path. Raised by the same external audit as the entry below, which missed the analytics beacon; found while checking its privacy findings.
Corrections from an external audit: the 2051 count and chart, the audited share, the crossing years, and a birth count read as a rate
An external audit of the live site found three defects and all three are fixed here. /national/ published 93 local authorities below a 50% White British share by 2051 where the homepage and the finding said 86: that page kept its own inline filter from April and was missed by the commit that carried the v8.0 recalibration through every other surface, so it counted Sheffield under both its retired and its current code and counted the six London boroughs whose projection the plausibility guard withholds. It now reads the shared helper, and the test that checks the count on each page now checks every page that prints it rather than three of them. The true cost article described the audited portion of the £7,973M central estimate as "about 60%"; its own table puts £4,363M on an audited basis, which is 55%, and audited plus attributed is 63%. Both figures are now stated, derived from the cost categories and asserted against them. Threshold years for the Lancashire towns shipped as bare years and now carry a sensitivity band from the model's measured out-of-sample error: Blackburn 2026 to 2028, Pendle 2033 to 2035, Preston 2035 to 2039, Burnley 2048 to 2053. The homepage hotel headline now names the year it covers, because a 2024/25 average sat undated above figures from the latest quarter. Two further defects were found while fixing those. The national ethnic trajectory chart on /national/ aggregated the projection file itself rather than reading the shared helper, so it counted Sheffield and Barnsley under both their retired and current codes and plotted 74.5% for 2021 against the 74.4% published elsewhere; it now derives every series from one place, and that page no longer imports the projection file at all. Separately, the deprivation fertility finding was headlined "Women in the most deprived areas have 66% more births" and gave 1.66x as a fertility ratio. It is a ratio of birth counts: ONS Table 8 publishes births by IMD decile and no count of women aged 15-44 to divide by, and deprived deciles have younger age structures, so an unknown part of the gradient is age composition rather than fertility. The same rule this site applies to asylum support, that a count is not a rate without the population at risk underneath it, was not applied to its own demographic data. The causal headline is withdrawn and the confounding argument, which the data does support, is kept.
Blackpool is worst on three of the four pressures we track. Asylum is no longer one of them.
This page previously said Blackpool was worst or near-worst on every domain, and gave its asylum figure as 577 people at 40.0 per 10,000. That was the position at 31 December 2025. At 30 June 2026 Blackpool has 236 people on asylum support at 16.5 per 10,000. Blackpool's own place page was already showing the newer figure while this page showed the older one.
Academic ethnic projections over-predicted White British share in 95% of areas
This article previously reported NEWETHPOP's error as 2.58pp and our own as 1.71pp, a 33 per cent improvement. Both figures are withdrawn. The 2.58pp had no backing in our published validation data, which records NEWETHPOP at 3.95pp across 296 areas, four paragraphs above where the 2.58pp appeared. The 1.71pp came from a backcast that fitted its ratios on the same two Censuses it was then tested against, so it measured the model's own guardrails rather than its accuracy, and it had the direction of the bias backwards. The claim to be the most accurate model publicly available is withdrawn with it. The head-to-head above is the replacement, on out-of-sample figures for both models. Full limitations are documented on our methodology page, and only a one-decade horizon has been validated.
Model v8.0: recalibrated on an out-of-sample test, and the counts that follow from it
The projection model was re-fitted and tested by fitting cohort change ratios on Census 2001 to 2011, forecasting 2021, and scoring against the actual Census 2021, so the fitting window never touches the target. On the White British share it scores mean absolute error 1.56pp with a bias of +0.05pp across 285 areas. The settings it replaces scored 2.82pp with a bias of -2.13pp on the same test, meaning the previous model projected change too fast rather than too slow. Shrinkage toward the national ratio is K=25 and the growth ceiling is 1.65 per decade, both chosen on that test. Counts move with it: 86 local authorities are projected below a 50% White British share by 2051 rather than 109, 59 of them currently majority, and Birmingham's 2051 share is 14.9% rather than 10.7%. The backcast comparison with NEWETHPOP is withdrawn; the defensible comparison is out-of-sample against out-of-sample, this model at 1.56pp across 285 areas against NEWETHPOP at 3.95pp across 296. Only a one-decade horizon has been tested.
Estate design system, five claim corrections, and a data refresh that had been failing since July
The site moved onto the shared visual system used across the network. Five published claims were corrected: the 190,809 safe-and-legal humanitarian grants for year ending March 2026 are 77 per cent in-country extensions rather than arrivals, and out-of-country grants fell from 58,459 to 43,620; the 107,003 people in supported accommodation was the 31 December 2025 quarter end, not the current 97,519 at 31 March 2026, and it had been understating every area's share by about 9 per cent; and the claim that 70 nationalities see grant rates rise after appeal was never right, the figure being 101 of 105. Separately, the scheduled data refresh had failed on every run since at least 13 July, across five distinct faults, and now runs Monday and Thursday and deploys what it commits.
Asylum grant rates rise sharply on appeal, across 101 nationalities
Two figures in this article were wrong and one that looked wrong was not. The count was wrong. This article said 70 nationalities; the correct figure is 101 rising out of 105 tracked, on the method set out above. The initial grant rate column was wrong and could not be reproduced from Asy_D04. It read 31.0 percent for Sri Lanka against 26.9 percent, 37.7 percent for Iraq against 30.9 percent, and 37.7 percent for Vietnam against 49.0 percent, an error in the opposite direction, so it was not one systematic slip. The "after appeals" column was correct on every row and is unchanged. Both columns are now computed the same way from the same dataset, and the uplift column is their difference rather than a separately carried figure.
86 local authorities projected minority White British by 2051
This piece originally said 109 areas and cited the model's backcast score as validation. Both were wrong. The backcast fitted its ratios on the same two Censuses it was tested against, so it measured the model's own guardrails rather than its accuracy, and those guardrails were letting projections compound into arithmetic rather than demography. The model has since been recalibrated against a genuine out-of-sample test that fits on Census 2001 to 2011 and forecasts 2021, scored on the actual Census 2021. That test showed the old settings were projecting change too fast, not too slow as previously stated here. On the recalibrated model the count is 86, and the currently-majority subset is 59, not 76. See the methodology for the test and the numbers.
How corrections work here
A correction is added to the top of the piece it concerns, dated, with the original claim quoted so the change can be judged. Withdrawn figures are named rather than deleted, and a test blocks each withdrawn figure from returning. The release diary records what changed and when, and the methodology carries the reasoning behind every model change. To report an error: info@asylumstats.co.uk.