Britain's replacement labour force survey classifies fewer people as long term sick than the survey it is meant to replace, and the statistics office says the order of the answers may be part of the reason

The two surveys do not agree about who is too sick to work.

That is the finding buried in the middle of the Office for National Statistics update published on Tuesday morning, and it comes from a study designed to catch exactly this. In 2025 the office ran a crossover survey: people who had just completed the interviewer led Labour Force Survey were asked to complete the online Transformed Labour Force Survey as well. More than 5,000 of them provided usable data. Because it is the same people answering both, the comparison strips out the sampling and non-response differences that usually muddy any comparison between two surveys.

On the big questions the two agree. Employed, unemployed or inactive, and employee, self-employed or unpaid family worker, both surveys put people in the same box most of the time.

Where they part

The largest gaps, and the lowest levels of agreement, are in the types of inactivity. Fewer respondents are classified as long-term sick on the new survey. Some of them instead report that they are retired, looking after family or home, or inactive for some other reason.

The office does not present this as a discovery about health. It offers a mechanical explanation. The wording of some response options changed, and so did the order in which they appear, and it gives one example: on the old questionnaire, retired is the ninth option in the list, while on the new one it comes first, because it is the most common answer.

Further analysis of reported disability is under way.

Not this year, and not as research either

The July readiness assessment, the first of its kind, reconfirmed that this is not the moment to move the headline statistics across. That much had been signalled. The harder line is what follows it: the office will not publish research outputs from the new survey at this stage either, because design and methods changes coming shortly would force revisions to anything it put out now.

So the timetable stretches again. A progress check in January 2027, which the office is careful to say is not a full readiness assessment. A detailed transition plan in the first quarter of 2027. Then July 2027, described as an important decision point, when, if the office and its users are ready, it starts the process for a transition in November 2027. That remains the most likely path, and the article gives the reason for not letting it slide indefinitely, which is the cost and the quality implication of running two surveys side by side.

The numbers that did improve

The April 2026 changes look to have worked. Data rotation, so that household and education questions are not asked again in full at every wave, better pay and earnings questions, and the conditional incentive doubled from £10 to £20.

Core Wave 1 return rates went from 37.1 percent in the first quarter of 2026 to 39.9 percent in the second. The longitudinal rate across Waves 1 to 5 went from 33.0 percent to 36.7 percent. The office notes that the two surveys calculate longitudinal response differently and cannot be compared directly on that measure, because the new one follows up only a subsample.

The burden numbers are the striking ones. Median household completion time on the new survey is about 17 minutes at Wave 1, against 40 to 45 minutes on the old one, and between 9 and 12 minutes at the later waves. All of this rests on the first nine weeks of data, and the office says it will not have two full quarters collected on the new basis until November.

One admission, in a single line

The old survey is still the lead source and its response levels have been recovering since 2023, helped by higher incentives and more interviewers. Waves 2 to 5 are now slightly above where they were before the pandemic.

Then this. An operational issue in May 2026 left telephone collection temporarily under-resourced. The office says it acted immediately, ran a lessons learned review, and that its analysis indicates the effect on the headline estimates was minimal.