JOLTS now rests on about a third of its sample. Widen the error bars.

On March 13, 2026, the Bureau of Labor Statistics began publishing a detailed series of response rates for the Job Openings and Labor Turnover Survey alongside the January 2026 data release, with total nonfarm history running back to January 2015 ^[1]. That decade-long view shows the response rate falling from the high 60s in percentage terms to around 30% in recent periods ^[2][3], which means roughly two-thirds of the businesses the agency contacts are not answering ^[4].

The mechanics matter here because JOLTS is not a census. It is built from a sample of approximately 21,000 establishments and reports job openings, hires, quits, layoffs and total separations ^[5]. At a response rate near 30%, the published series is being estimated from on the order of 6,300 responding establishments ^[6]. Collection runs about 45 days across computer-assisted telephone interviews, web submissions, email and fax, and the breadth of channels has not arrested the decline ^[7]. The BLS describes the fall as structural rather than a drift, and consistent with what is happening across its other establishment surveys ^[8].

Two consequences for anyone who prices off this data. The first is precision. Sampling error scales with the inverse square root of the number of respondents, so a move from roughly 67% response to roughly 30% implies standard errors about 50% wider, holding sample design and everything else constant ^[9]. That is an arithmetic illustration rather than a restatement of the BLS published variances, which incorporate weighting and imputation, but the direction is not in dispute: a month-to-month wobble in openings that would have been marginally meaningful in 2015 is closer to noise now.

The second is bias, which is the harder problem. The concern flagged in the reporting is that non-respondents may differ systematically from respondents, and that if smaller firms or particular industries are disproportionately dropping out, statistical adjustments may not fully compensate ^[10]. Non-response bias does not average out over three-month windows the way sampling noise does. It shifts the level, and it is difficult to detect from outside the agency.

To its credit, the BLS is making the problem legible. Starting with the January 2026 data it also began breaking response rates down by industry, region and establishment size ^[11], which is exactly the cut needed to test the small-firm hypothesis. The response-rates page was last updated on May 20, 2026 with additional methodological context ^[12]. Less helpfully for anyone building regional models, some state-level JOLTS data moved from monthly to annual publication beginning in January 2026 ^[13].

The practical adjustment for operators is unglamorous. Treat single-month JOLTS prints, and especially the openings-to-unemployed ratio built on top of them, as a low-resolution signal; lean on the quits rate over quarters rather than months; and stop treating a revision as a surprise when a third of the intended sample is doing the talking.

What to watch: the industry, region and size breakdowns as they accumulate over 2026, and specifically whether small-establishment response sits materially below the 30% headline ^[11][2]. If it does, the level of reported openings, not just its variance, is in question. Watch too whether other establishment surveys get the same disclosure treatment, given that the BLS says the decline is not confined to JOLTS ^[8].