A new working paper by economists Raj Chetty, Jamie Fogel, Lawrence F. Katz, Kadeem Noray, Sonya Porter, and James Reisinger reports that machine learning models built on large-scale administrative data can estimate the earnings effects of workforce training programs with accuracy comparable to randomized controlled trials (RCTs). The full paper is available as a PDF via Opportunity Insights.

From small surveys to national tax data

Modern program evaluation has long prioritized evidence from RCTs, in which participants are randomly assigned to treatment and control groups so that a program's effect can be isolated from other factors. Governments and researchers have leaned on such experiments because observational estimates of program impacts have often been found to diverge substantially from experimental benchmarks — a problem economists have flagged since the 1980s. Those earlier findings, the authors of the new paper note, were based on small survey datasets. The study revisits the question in the era of large-scale administrative data, at a time when running an experiment for every program a government operates is rarely feasible for reasons of cost, logistics, or ethics. Administrative data offer an alternative evidence base: records generated in the course of routine government functions — tax filing among them — that cover entire populations rather than the small survey samples behind the earlier divergence findings.

Using anonymized income tax records covering all American adults between 1999 and 2025, the researchers estimate the effects of workforce training programs on earnings. Machine learning models predict counterfactual outcomes — what participants would have earned in the absence of enrollment — from rich data on prior work histories, demographics, and residential locations.

Matching the experimental benchmark

Rather than testing whether the observational estimates are unbiased, the authors use 37 RCTs to estimate the distribution of their bias — a validation spanning dozens of separate experiments rather than a single benchmark trial. The machine learning estimates closely match the RCT estimates, with a root mean squared error (RMSE) — a standard measure of how far estimates fall from the values they aim to approximate — comparable to that of the RCTs themselves. According to the paper, the approach reduces RMSE by 60 percent relative to the parametric estimators used in prior work, because the larger samples and richer covariates permit more flexible, higher-dimensional models of the counterfactual.

Where the gains concentrate

Applying the method to a broader set of sites and enrollment years beyond the experimental sample, the paper finds that treatment effects on earnings vary substantially across locations and over time, even within a given program — meaning the same program may deliver very different returns depending on where and when it operates. Programs that place more participants into target-sector jobs generate the largest earnings gains, with impacts that persist even after participants leave their initial jobs.

The findings, the authors write, provide guidance for improving job training programs and, more broadly, illustrate a scalable approach to program evaluation: using observational data to construct experimentally validated "digital twins." Under this framework, experiments retain a central role as the benchmark against which data-driven estimates are calibrated and checked, while administrative data extend evaluation to sites, years, and programs beyond those where randomized trials were run. The paper applies the method to workforce training, and how far it travels into other areas of policy evaluation remains to be demonstrated by subsequent research.

The paper was highlighted on Marginal Revolution, the economics blog founded by Tyler Cowen and Alex Tabarrok, in a post published on November 11, 2026. Its accompanying commentary was unequivocal: "Work of this sort is going to revolutionize economic method."

As a working paper, the study is circulating in the format researchers typically use to release findings ahead of journal publication and the peer review that accompanies it, placing these results at an early stage of the research process.

Chetty and Katz are both economists at Harvard University. Chetty directs Opportunity Insights, a research organization based at Harvard focused on economic opportunity and mobility, whose website hosts the working paper. Katz, a labor economist, has long studied wages, education, and technological change.