Don't Trust the Survey? Fine. Here Are the Numbers Nobody Can Argue With.

When I wrote about the six unemployment measures last month, the most common response was not about U-3 versus U-6. It was more fundamental: the whole thing is a survey, response rates are collapsing, and nobody I know answers an unknown number anymore. So why should anyone trust it?

I am not going to wave that away, because the skeptics are partly right. The household survey's response rate has fallen from the low 90s to the upper 60s in 15 years. And the bias is not hypothetical. Alan Krueger and his coauthors documented something called rotation group bias: households report unemployment at meaningfully higher rates in their first interview than in later rounds, roughly half a point higher, and the gap has grown over time. If the first answer is closest to the truth, the official rate runs a little low. The Census Bureau's own researchers found nonresponse bias crept into income measures during the pandemic. This is all documented, published, and openly debated. That transparency is itself worth something. But the objection stands: the headline number rests on people agreeing to talk.

So let me answer the question I keep getting. Are there hard numbers, actual counts rather than sampled opinions? There are. Most people have just never been told where to look.

The hardest labor market number in America is one almost nobody cites: the Quarterly Census of Employment and Wages. It is not a survey. It is built from the unemployment insurance tax filings that virtually every employer is legally required to submit, roughly 12 million workplaces covering more than 95 percent of all jobs. Nobody opts out of it the way they opt out of a phone call, because opting out of a tax filing is called something else.

The QCEW is effectively a census of American employment, and it is the referee of the entire system. Once a year, the monthly payroll survey gets benchmarked against it, and when the two disagree, the tax records win. If you want proof of how much that referee matters, look at September 2025: the preliminary benchmark showed the payroll survey had overcounted employment by 911,000 jobs in the year through March 2025, a 0.6 percent miss and the largest downward revision on record. The survey told one story for a year. The tax records corrected it. That correction is not a scandal. It is the system working exactly as designed, with hard data overruling sampled data on a schedule.

The catch is the lag. QCEW data arrives five to six months after the quarter closes. It is truth, delivered late.

At the other end of the speed spectrum sits the most underrated release in economics: weekly initial unemployment insurance claims. Every Thursday at 8:30 a.m., the Department of Labor publishes a count of actual claims filed with actual state agencies by actual people. Nobody is sampled. Nobody is interviewed. For the week ending July 25, initial claims came in at 197,000, with the four-week average around 203,000.

Claims undercount total joblessness, because not everyone who loses a job files and not everyone qualifies. Independent contractors, for instance, are largely invisible here, a gap my industry knows intimately. But as a layoff signal, claims are unmatched: administrative, weekly, and boring right up until the moment they are not. Every recession of the modern era announced itself in the claims data before it showed up anywhere else. When the four-week average starts climbing week after week, believe it over any survey.

Between the weekly count and the quarterly census sits a middle tier of harder-than-survey evidence. ADP publishes monthly employment figures built from its actual payroll processing records covering more than 25 million private sector workers. That is not a poll; it is paychecks. Its limitation is scope: ADP sees its own client base, private sector only. State tax withholding receipts track paycheck volumes in near real time for anyone willing to dig into state revenue reports. And job postings data from platforms like Indeed measures labor demand directly: a posting is a real action by a real employer spending real money, not an answer to a questionnaire.

None of these is complete. That is the point. Every labor market number is a partial view from a different angle, and the professionals who get paid to be right never rely on one instrument. They triangulate. The household survey provides breadth and the only measure of people, not just jobs. Claims provide speed. Payroll records provide hardness. Tax records provide the final word.

So here is the practical framework for anyone who distrusts the survey but still needs to read the labor market. Watch weekly claims for turning points. Read the monthly report for breadth, but treat it as a draft. Check the payroll records as a second opinion. And once a year, let the tax records tell you how wrong everything else was. When the soft data and the hard data agree, be confident. When they diverge, the divergence is the signal.

Sources: BLS QCEW; BLS CES Preliminary Benchmark Revision (Sept 2025); DOL ETA Weekly Claims (July 25, 2026); Krueger, Mas & Niu (2017), Review of Economics and Statistics; U.S. Census Bureau SEHSD-WP2020-10; ADP Research Institute.