This paper examines bias in fixed effects (FE) estimators when the absorbed grouping structure is generated by both the explanatory variable and unobserved determinants of the outcome. This problem is relevant across management research settings, for example, where FE groups are formed through selection, sorting, rankings, thresholds, or other organizational and market processes. In such settings, FE estimation implicitly conditions on a collider-defined partition of the data, creating spurious within-group dependence between the regressor and latent outcome determinants. We derive the resulting bias as a function of the induced within-group covariance between the regressor and the latent factor, the within-group identifying variance of the regressor, and the weighting structure of the FE estimator. The analysis shows that local collider bias can be severe in groups with limited within-group variation, while the aggregate FE bias depends on how identifying variation is distributed across groups. We propose diagnostics based on variance concentration, FE weights, effective identifying groups, and high-weight leave-out checks. The results caution against treating FE as a routine adjustment and instead frame FE selection as a causal modeling decision in empirical management research.
Contact person: Anastasiia Lutsenko
Subscription to the invitation mailing list and more information on the seminar page.