Britain's jobs survey undercounted employees by 365,000 a quarter for three years. The office that runs it has just published that itself

Nearly every figure in the argument about AI and British jobs starts in the same place: the Labour Force Survey. Employment, unemployment, economic inactivity. All of it rests on knocking on doors and asking people what they did last week.

Fewer and fewer people answer. Response has fallen from around 80 per cent of sampled households in the early 1990s to under 25 per cent by the end of 2025. In Great Britain it bottomed out at 12.7 per cent in the third quarter of 2023, when roughly 14,000 households were completing interviews.

That on its own is not bias. It becomes bias only if the people who stop answering are different, in ways the weighting does not already fix, from the people who keep answering. Until now, nobody could say whether they were.

How they checked

On Monday morning the Office for National Statistics published research that links survey responses to PAYE tax records, person by person, from 2019 to the end of 2025. The tax data covers nearly everyone paid through a payroll, respondents and non-respondents alike. If working people are under-represented among those who answer, the survey will count fewer employees than the tax records show, once you have made the two definitions line up.

The office did that work first, adjusting for people taxed here but living abroad and for employees in care homes, hotels and barracks, who are outside what the survey is trying to measure.

What it found

The gap moved three times.

In 2019 the survey counted about 217,000 fewer employees a quarter than the tax data. Working people were slightly harder to reach, which is what you would expect of people who are out of the house.

Through the pandemic that reversed. From spring 2020 to the end of 2021 the survey counted 332,000 more employees a quarter than the tax records showed, peaking at 644,000 in the first quarter of 2021. Everyone was at home and interviews moved to the telephone, which made employed people easier to reach, not harder.

Then it flipped back, and further. From the start of 2022 to the end of 2024 the survey counted 365,000 fewer employees a quarter than the tax data. That is the largest and longest gap in the series, and it lines up with the period when response was at its lowest.

By 2025 it is 10,000 a quarter. Near enough nothing.

What that does to the numbers

Reweighting the survey to the tax totals moves the headline rates. For 2022 to 2024, the employment rate comes out 0.42 percentage points higher, about 174,000 more people in work each quarter. Economic inactivity comes out 0.37 points lower, and unemployment 0.09 points lower. Through the pandemic the corrections run the other way, with employment 0.57 points lower and 230,000 fewer people in work.

For 2025 the adjustment is 0.08 points on employment. The office says the improvement follows work to lift response: the sample boost restored, more interviewers, more chasing, and payments for taking part that are substantially higher than before the pandemic.

The cautions, which are the office's own

This is not a revision. The ONS calls it proof of concept research, says the results are not official statistics, and tells users to carry on using the published figures.

The unemployment correction is the weakest part of it, and the office says so. Tax records tell you whether someone was paid, which is a good guide to employment and a poor one to unemployment. The two are only weakly correlated. So a small adjustment to the unemployment rate does not mean unemployment was measured well.

The findings are also subject to change when the survey is reweighted to newer population estimates.

Who was missed

The office published the gap by age and sex. In the years after the pandemic the biggest shortfalls were among people aged 50 to 64, and among men. During the pandemic the overcounting was concentrated among people aged 25 to 49, and among women.

For this paper that is worth saying plainly, because it points away from a story we have been circling. We asked the ONS whether drop-out was skewing youth unemployment. It said it has made no assessment. This new work does not find young people at the centre of the drift.

What happens next

The office wants feedback before deciding whether to build this into how the figures are produced. It also has to redo the work on the Transformed Labour Force Survey, the replacement now being brought in, before any of it can be used routinely.