Are Administrative Union Records Reliable for Substate Estimates of US Union Membership?

Testing LM filings against a survey benchmark

The Current Population Survey (CPS) has supplied the headline US union density numbers since 1973, but its sample is too small to be reliable below the state level. Researchers responded by turning to administrative data: unions representing private-sector, federal, or postal workers must file annual LM forms with the Department of Labor's Office of Labor-Management Standards, and these forms have asked for membership numbers since 2000. A common practice is to filter to local unions, geolocate the union office, and assign reported members to that office's county, commuting zone (CZ), or congressional district (CD).

This paper asks whether that practice is sound. Because the CPS cannot serve as a fine-grained benchmark on its own, the authors build one: multilevel regression with poststratification (MrP) on the CPS, which models union membership from individual and geographic predictors and then reweights to known population totals. They then validate LM-based density against this benchmark, descending from the state level to CDs, CZs, and counties.

Two LM membership definitions are compared throughout. The inclusive version keeps all reported members, including retirees and unemployed members the CPS would miss. The exclusive version drops extremely large locals and those sharing addresses. The distinction matters because it changes both the level and the geography of the resulting estimates.

State-level agreement is strong, then breaks down

At the state level in 2007, the three measures line up closely: CPS, MrP, and LM density show pairwise correlations of r = 0.99, 0.88, and 0.89, with Spearman rank correlations of 0.99, 0.86, and 0.86 under the inclusive LM definition. This is the result earlier work relied on when treating LM-based estimates as credible.

The correspondence weakens sharply one level down. Averaged across 2000, 2007, 2014, and 2019, the MrP-LM Pearson correlation at substate levels hovers around 0.3, less than half its state-level value, ranging from 0.40 for CZs in 2000 to 0.22 for CDs in 2019 under the inclusive definition. Spearman correlations show a similar pattern, so the drop is not an artifact of the correlation metric. The paper reports no obvious trend over time or across geographic scale, though 2019 produced the weakest correspondences across all years and levels considered.

The exclusive definition does not rescue the substate estimates. At the state level it actually performs worse than the inclusive version against both CPS (2007 r = 0.7, rho = 0.8) and MrP (2007 r = 0.6, rho = 0.7), yet at the county level it correlates marginally better with MrP than the inclusive version does. That reversal is a clue to what is driving the divergence.

Reporting lumpiness produces impossible densities

The breakdown is driven in part by lumpiness in LM reporting. Under the inclusive definition, several counties and congressional districts show union densities near or above 1.0, meaning aggregated reported members exceed the area's entire employed population. Pooled over the four years, 13 CDs and 8 counties cross that threshold, with maximum values of 2.63 and 2.10 respectively; no state or CZ exceeds 1.0, and their maxima are 0.65 and 0.63. Under the exclusive definition, no unit at any geography exceeds 1.0, and the maxima fall to 0.35 for CDs and 0.36 for counties.

The paper's explanation is organizational rather than demographic. Many unions have mailing addresses in moderate-population counties that are state capitals or former industrial zones with a history of unionization. The largest county-level value, 210%, is Albany County, New York, home to the NYSUT; other examples include Dauphin County, Pennsylvania (Harrisburg) and St. Louis City, the small former industrial core surrounded by the separate and much more populous St. Louis County. The authors note that densities above 1.0 are conceptually possible if locals report large numbers of retired or unemployed members, but they judge this very unlikely here. For comparison, the largest county-level density from the MrP benchmark is 0.30 (Honolulu County, Hawaii, 2007).

The forces producing this lumpiness are organizational factors and decisions, such as how union locals are structured, rather than the geographic dispersal of members around a central location. That is why the pathology disappears when extremely large locals and address-sharing locals are excluded.

The density gap is concentrated within states

A second diagnostic decomposes the signed MrP-LM density gap at the congressional-district level into between-state and within-state components. Under the inclusive definition, variation across districts in the same state accounts for about 95% of total variation in the gap in every year, and LM estimates differ from MrP by roughly 11 percentage points within a state on average. Under the exclusive definition, within-state variation falls sharply: the typical district differs from others in the same state by about three to four percentage points, and across-state differences now account for the majority of the gap.

The two LM versions therefore paint very different pictures of where union membership variation lives. The inclusive version implies that most of the action is local, within states; the exclusive version shifts the weight back toward differences between states. Because the choice between them is a data-cleaning decision rather than a substantive one, this is a warning about how much conclusions can depend on filtering rules.

Spatial correlation points the same way. Using a row-standardized queen contiguity matrix for 2014, the MrP benchmark shows a high degree of spatial correlation, meaning nearby CZs and CDs tend to have similar density levels. The paper reports that LM-based measures do not reproduce this pattern as well, consistent with members being assigned to office locations rather than residential geographies.

Missing public-sector members explain only a small share

Unions representing only state and local public-sector workers are exempt from filing LM forms, so LM-based estimates should systematically undercount those members. The paper tests how much of the MrP-LM gap this explains by regressing the signed gap in percentage points on the state and local government employment share, with year fixed effects, standard errors clustered on the geographic unit, and controls for urbanization and log employment.

At the state level there is no evidence that public-sector size accounts for the gap. At the county level, using the exclusive LM data, a 1 percentage point higher public-sector employment share is associated with about a 0.55 percentage point larger MrP-LM gap. That relationship all but disappears in states where public-sector collective bargaining is effectively prohibited, consistent with part of the gap coming from missing public-sector union members. On a model R-squared basis, state and local public-sector employment explains about 12% of the MrP-LM gap under the exclusive definition. The authors describe this as substantial but by no means overwhelming, so missing public-sector unions cannot be the main story.

A China shock regression flips with the density measure

The stakes become concrete in a standard commuting-zone China shock shift-share regression for 2000-2014, following Autor, Dorn, and Hanson. The same regression design, run on the same data, produces import-exposure coefficients that change magnitude, cross conventional significance thresholds, and switch sign depending only on which density measure is used to capture local unionization.

This is the paper's strongest argument that the measurement problem is not academic. A researcher choosing between CPS-based, MrP-based, and LM-based density, or between the inclusive and exclusive LM definitions, can reach qualitatively different conclusions about how trade exposure relates to union density. The authors stress that simply smoothing LM-reported members in space is not a good solution, and that any work using LM-reported quantities should report how conclusions are sensitive to filtering and cleaning decisions.

The paper also cautions against extending the same geolocation-and-aggregation approach to other LM-reported quantities such as expenditures and assets. The authors lack data for a comparable validation of financial variables, but the lumpiness in membership gives pause. They note that financial quantities are subject to auditing and investigation by OLMS and the Department of Justice in ways membership reports are not, so the caution is specifically about attributing members or resources to small geographies by treating the filing address as where resources are deployed.

Use LM records for organizations, not for mapping members

The paper's conclusion is not that LM data are worthless. LM forms are an organization-level census: they record the birth, death, location, affiliation, leadership, and finances of local unions with a coverage and consistency no survey matches. The authors point to work using LM forms to study local union formation and growth, expenditures and dues revenues, and links between locals and campaign donations, and they suggest future research on how economic shocks affect the birth and death of union organizations or the turnover and career paths of union officers.

The boundary is geographic. A union's office location is arguably a better proxy for where work happens than for where members live and vote, and the concerns are especially pronounced when studying social or political outcomes in which workers' geography of residence is of prime importance. The paper's own framing is that LM records remain a rich resource for studying unions as organizations, but aggregating LM-reported membership is not a reliable way to measure union density at fine geographic resolution.

The MrP benchmark comes with its own limits, which the authors state plainly. It inherits the known limitations of the CPS, including recall and misreporting bias and the absence of retired and unemployed members. Because it uses a shrinkage estimator, it will by design somewhat understate within-state variation in union membership relative to the CPS. It is computationally expensive, and the detailed ACS five-year tables needed for poststratification only become available in 2007, so extending the approach to the 1990s or filling gaps between the decennial census and the ACS era would require additional modeling assumptions the authors have not investigated. Smaller geographies such as counties would be even harder.

The filtering decision is the one to report

The most decision-relevant result for anyone using LM data is not the headline correlation but the sensitivity to inclusion rules. The inclusive and exclusive definitions reverse which geographic level looks more variable, change the maximum observed density from 2.63 to 0.35 at the congressional-district level, and move the within-state share of the MrP-LM gap from about 95% to roughly a third. Because the exclusive definition performs worse at the state level but better at the county level, there is no single cleaning rule that is uniformly correct.

That is why the paper's practical priority is transparency about filtering rather than a recommendation to abandon LM data. Any analysis using LM-reported quantities should show how its conclusions respond to decisions about extreme values, missing values, and unions sharing addresses. The China shock exercise demonstrates that these choices are not hypothetical: they can change the magnitude, significance, and sign of a published-style coefficient.

For readers weighing whether to trust a substate union density estimate, the paper's threshold is geographic. State-level LM estimates track the CPS and MrP closely enough to be useful; below the state level, the correlation falls to roughly 0.3 and the estimates carry lumpiness that does not average out. The appropriate use of LM forms is questions about union organizations themselves, where the filing is the unit of interest, rather than questions that require assigning members to the small geographies where they live.

Sources used in this answer