% WARNING: This file may contain UTF-8 (unicode) characters.
% While non-8-bit characters are officially unsupported in BibTeX, you
% can use them with the biber backend of biblatex
%    usepackage[backend=biber]{biblatex}

@techreport{LaborHubDP16,
 title = "fevc: Leave-out bias-corrected variance decompositions in Stata",
 author = "Schmieder, Johannes",
 institution = "Cornell University ILR School",
 type = "Discussion Paper",
 series = "Labor Hub Discussion Paper Series",
 number = "16",
 year = "2026",
 month = "October",
 URL = "https://laborhub.ilr.cornell.edu/publications/26/",
 abstract = {Variance decompositions from regressions with worker and firm fixed effects are widely used to measure heterogeneity and sorting in linked data. Variances computed directly from estimated effects overstate heterogeneity because they include estimation noise, and the bias can be large when few workers move between firms. This article introduces fevc, a Stata command that implements the leave-out bias correction of Kline, Saggio, and Sølvsten (2020, “Leave-Out Estimation of Variance Components,” Econometrica 88: 1859–1898). fevc reports corrected variances of worker and firm effects, their covariance, and the variance of their sum. Researchers choose the unit left out, the treatment of controls, and the weights that define the population being described. The command also reports standard errors for projections of fixed effects on observed characteristics. It has two backends: a source-only Mata implementation that runs wherever Stata runs, and a parallel Rust plugin for large datasets. Calculations based on random projections report Monte Carlo standard errors that measure their numerical precision. In a public extract of Italian social-security records, the correction halves the estimated variance of firm effects and reverses the sign of the worker–firm covariance. In synthetic benchmarks with 6.4 million observations and 28 threads, the Rust backend is 1.3 to 3.5 times as fast as the Matlab package of Kline, Saggio, and Sølvsten. Under match deletion, its measured peak memory is about 30 percent of the Matlab package’s.},
}