How the Census Model Finds the Employer Component

The Census study addresses a harder question than whether nonprofit workers have a higher or lower average wage: after accounting for the kinds of workers involved, how much of the difference is associated with the employer? Its answer depends on two linked moves. First, the researchers use administrative records to identify workers, employers, and nonprofit status more broadly than a conventional survey can. Then they use workers' movements between firms to separate a worker component from a firm component. This chapter explains that identification story and its limits, so “mostly causal” is understood as a conclusion within a design rather than as a claim from an experiment.

From many records to one labor market

The project began with a linkage among the American Community Survey, IRS Form 990 filings, and the Longitudinal Employer Household Dynamics data. That combination could connect worker characteristics, nonprofit information, and employer–employee earnings records. In 2024, the researchers gained access to administrative tax data covering the near-universe of U.S. workers from 1999 through 2023. The tax records include information that the LEHD does not provide in the same way, including federal workers, independent contractors, sole proprietors, other earners in a household, and demographic and family information.

That expansion matters because a pay comparison can be distorted when the data omit a systematic part of the labor market. LEHD is built from employment covered by unemployment insurance and is recorded quarterly. It excludes federal workers and some private-sector categories. The tax data are collected annually, but their broader coverage lets the researchers observe a wider set of workers and earnings relationships. The gain is not simply a larger sample. It changes which worker histories can be connected and which possible explanations—career paths, household structure, or family circumstances—can be examined.

Administrative records also provide a more direct way to identify the employer. The study uses W-3 firm returns to identify industry and nonprofit status, then assigns those returns to the population of W-2 wage records. The analysis restricts attention to full-time, main-employer worker-years. That restriction gives the comparison a consistent unit, but it also defines the population whose pay is being explained. It is not a promise that the result covers every person who receives some income from a nonprofit, every part-time arrangement, or every form of compensation.

The linked records reveal why classification is part of the economics rather than an administrative footnote. Form 990 information includes employee and volunteer counts. A charity that relies heavily on volunteers may look small if the analyst counts only paid employees, even though its capacity to provide services is much larger. The study also compares nonprofit classifications from Form 990 with industry codes designed primarily to describe for-profit activity. A nonprofit legal-aid organization, for example, may be placed in a broad nonprofit category rather than in legal services. That can make it appear to compete with the wrong set of employers.

The researchers use these comparisons to improve their understanding of the data, not to claim that every classification has been perfected.