Measuring the Labor Market in a Decade of Shocks

I recently delivered the keynote at the NBER Conference on Innovation in Measuring Employment, and International Business Times covered the talk https://www.ibtimes.com/adriana-kugler-what-americas-jobs-data-misses-how-fix-it-3807869. I want to lay out the argument here in my own words because it gets to the heart of how economic policy decisions are made.

Why timing matters more now

Since 2020, the U.S. economy has faced a pandemic, disrupted supply chains, a war in Europe, tariff increases on a scale not seen since Smoot-Hawley, oil disruptions reminiscent of the early 1980s, and the emergence of AI. Any one of these developments would challenge policymakers. When they arrive on top of one another and in rapid succession, the speed and accuracy of the information we rely on become decisive. Yet the agencies responsible for producing official statistics are operating with fewer resources at the same time that private companies are collecting more data than ever before.

What I learned from two dashboards

In both of my most recent public roles, I needed answers that standard data could not yet provide. At the World Bank, when food prices rose after Russia invaded Ukraine, we asked staff to create a dashboard that ranked countries by the severity of food insecurity every two weeks. This allowed the Board to determine where action was needed first.

At the Federal Reserve, unemployment had begun to edge higher from 3.4 percent in April 2023, its lowest level in half a century, before eventually reaching a peak of 4.5 percent in November 2025. Because recessions often bring sharp increases in joblessness, I wanted an early warning. The dashboard we developed brought together weekly jobless claims and WARN notices with private signals, including announced job cuts, layoff plans discussed on corporate earnings calls, and internet searches related to layoffs. It detected the cooling in early 2024 and the increase in layoffs that summer.

Four questions our data should answer

Is the labor market beginning to loosen before unemployment reflects it? Employers generally make adjustments within the firm first. People work at a slower pace, absences increase, and the match between skills and jobs becomes weaker. Official statistics capture very little of the first two, while workplace software that monitors hours can. Combining those records with other measures could identify weakness while policymakers still have time to respond.

Why are people outside the labor force? Surveys tell us who is not working and can identify illness or caregiving as reasons, but after the pandemic we were still left trying to understand early retirements, adults leaving the workforce, and young people taking longer to enter it. Asking people directly about pay, flexibility, and how they are treated at work would provide policymakers with more useful information for designing policy.

What is AI doing to the work itself? Most tracking to date focuses on which businesses use AI and how much they invest in it. We know much less about which tasks disappear, which become easier, and which new tasks emerge. Researchers applying natural language processing to job listings are beginning to answer those questions, and this remains an area that official data cannot yet address on their own.

How good are the jobs? The federal survey of contingent and alternative work has existed since 1995, but after 2005 it went twelve years without being fielded, just as gig work was shifting from staffing agencies toward apps. Connecting tax records with platform data could help close that gap. Many aspects of job quality are not counted at all, including dental, vision, disability, and life insurance, as well as whether a job provides opportunities to advance.

Where each kind of data is strongest

Official data are representative, carefully processed, and consistent across decades. That consistency allows us to recognize patterns when they appear again. The drawback is the delay. Monthly household and payroll surveys take two to three weeks, job openings and turnover data take about five weeks, and information on gross job creation and destruction arrives roughly seven months later. By that point, decisions may already have been made using an outdated picture.

Private data can move much faster. ADP releases payroll figures almost as soon as a month ends, although its clients employ only about one in five private-sector workers. That is why ADP benchmarks its figures against the Quarterly Census of Employment and Wages.

Each type of data can strengthen the other. Fewer households and businesses respond to government surveys than they once did, partly because people are more reluctant to share information and partly because agencies face resource constraints. In-person visits are more effective than sending reminders, while modest payments work better than generous ones, which can make people wary. Private payroll information could help fill the remaining gaps, much as private scanner data are being tested alongside official price statistics. The practical challenge is that agencies must obtain permission from each company whose data they use.

Getting faster without cutting corners

Automated checks using machine learning could verify incoming data as it arrives, helping shorten release times without lowering standards. The models used to estimate business openings and closings could also be updated more frequently than twice a year. The larger point is straightforward: private data provide speed and new detail, public data provide representativeness and historical depth, and policy is better served when both are used together."