When tax records and surveys describe different populations

Around 2015 a widely repeated claim held that alternative work arrangements in the United States had expanded sharply. Its main source was a paper by Lawrence Katz and Alan Krueger, who used a survey fielded through the RAND American Life Panel and reported a rise in such arrangements between 2005 and 2015. The figure travelled quickly, in part because it arrived while platform companies were growing.

In 2019 Katz and Krueger published a reassessment in the ILR Review. Comparing their earlier survey against the Bureau of Labor Statistics Contingent Worker Supplement, which returned in 2017, they concluded that their original estimate had overstated the increase, and they attributed part of the discrepancy to differences between their instrument and the official one. The episode is unusual and worth noting on its own terms: the authors of a widely cited number revisited it and revised it downward.

What administrative records saw

A separate line of work started from tax filings rather than from questionnaires. Brett Collins, Andrew Garin, Emilie Jackson, Dmitri Koustas and Mark Payne, working with United States Internal Revenue Service data, examined information returns — the forms platforms issue to the people they pay — to trace participation over time. Their analysis describes participation in the online platform economy rising from a negligible base in the early 2000s to roughly one per cent of the workforce by the mid-2010s, with most participants reporting modest amounts alongside other income.

Administrative data has a specific advantage: it does not depend on anyone remembering anything. It also has specific blind spots. Reporting thresholds determine which payments generate a form at all, and those thresholds have changed. Payments routed outside the reporting system, or received in cash, leave no trace. And a form records a payment, not an hour, so nothing in the data speaks to time worked.

What bank records saw

Fiona Greig and Diana Farrell, working at the JPMorgan Chase Institute, took a third route: transaction records from a large set of checking accounts, identifying deposits originating from platform companies. That approach captured participation and volatility with unusual granularity, including the finding that platform income for many account holders was episodic rather than continuous.

Its limits are equally specific. The sample consists of customers of one bank, in one country, who receive deposits electronically into an account that the bank can observe. People paid through other channels, or banking elsewhere, are outside the frame entirely.

Why the counts diverge

Survey estimates have consistently run higher than administrative ones. The Pew Research Center, surveying United States adults in 2021, reported that roughly sixteen per cent had at some point earned money through an online gig platform, with a considerably smaller share doing so in the previous year. That kind of figure and a one-per-cent participation rate from tax records are not necessarily in conflict: they answer different questions. "Ever" is not "currently". "Earned money" is not "filed a return above a threshold". A single afternoon of activity three years ago registers in one instrument and not in the other.

Question wording contributes as well. Respondents who complete tasks online for payment do not always describe that activity as a job, and surveys that ask about employment last week can miss it. Instruments that name platforms explicitly tend to record more activity than instruments that ask in general terms.

None of this makes any of the numbers wrong. It makes them non-interchangeable. A figure quoted without its instrument — who was asked, over what period, and what counted as participation — cannot be compared with another figure quoted the same way, and the difference between two such figures is often a difference in definition rather than in the world.

Return to the register