Membership grew faster than employment: A sixteen-year panel of inside electrical construction in hyperscale data center metros, 2010 to 2025

Research Note, version 2.0, 20 September 2026. Supersedes the working paper at 10.5281/zenodo.22310706 and is deposited under the same concept DOI. The data and code are deposited separately at 10.5281/zenodo.22862472. This version supersedes 10.5281/zenodo.22861977, published earlier the same day, which carried the note and the full replication package in one record. The data now have their own DOI and this record carries the note alone. No figure, table or sentence of the analysis changed. A prior working paper measured inside electrical union membership in metropolitan areas where hyperscale data center operators built campuses, using an eight-year panel opening in 2018. Several treated metros already had operating campuses in that base year, so the design could observe one usable pre-treatment period and could not test the parallel-trends assumption it required. This note rebuilds the panel over sixteen years, 2010 to 2025, on the same specification and sources, and adds an independent employment series covering every year. Seven coefficients for 2010 through 2016 are individually insignificant and jointly indistinguishable from zero, chi2(7) = 9.40, p = 0.225; four treated metros carry a site dated before 2017, so those seven test whether the two groups were already diverging rather than bounding a clean pre-treatment window. The treated contrast is 27.4 percentage points of membership growth, 95 percent interval -1.4 to +56.1, clustered on 33 states, which does not clear conventional significance, against randomization p-values of 0.0136 permuting whole metropolitan areas and 0.0102 permuting whole states, which do. Quarterly Census of Employment and Wages counts of electrical-contractor employment, summed to the same statistical areas over the same years, show a treated contrast of 9.1 to 11.9 percent by 2025 against 17.3 percent in membership. Conditioning the membership estimate on employment growth in the same rows removes 7.3 to 8.9 points of a 26.6 to 27.5 point unconditional contrast and leaves a positive residual that does not clear conventional significance on either industry code. The note is descriptive: no instrument, no runner-up control, and the employment series it benchmarks against is not itself flat before treatment. This version corrects three things in the superseded working paper. It supplies the parallel-trends test the earlier design could not run. It replaces a randomization p-value computed by permuting individual locals, which was too tight by a factor of about three, with permutations at the metropolitan area and the state. And it corrects a treatment classification error: Local 150 of Libertyville, Illinois, whose county sits in the Chicago metropolitan area, was counted as a control because the crosswalk matched a city-name prefix. The eight-year estimates the working paper reported are unaffected. LM2_CLASSIFICATION_ERROR_2026-09-20.md records the sweep and the effect on every specification. Data and code. Not in this record. The 11,645-row local-union panel, its codebook, the sixteen OLMS yearly extracts, both jurisdictional roster vintages, thirty analysis scripts, every deposited log and JSON, every retrieved source and the classification-error record are deposited as a separate citable dataset at 10.5281/zenodo.22862472, cited by this note and by its companion alike. That record reproduces from its own inputs; its README.md gives the run order and names the one step that cannot be run from the package and why. JEL: J51, J21, R23, C81.