(Debiased) Inference for Fixed Effects Estimators with Three-Dimensional Panel and Network Data

Abstract

Inference for fixed effects estimators is often unreliable due to Nickell- and incidental parameter biases. While these issues are well understood for classical two-dimensional panels, little is known about three-dimensional panel structures (e.g., sender x receiver x time). We develop inferential theory for a broad class of linear and nonlinear fixed effects M-estimators in this setting, covering bipartite, directed, and undirected network panel data, multiple specifications of additively separable unobserved effects, and both strictly exogenous and predetermined regressors. Our analysis reveals fundamentally different asymptotic properties compared to two-dimensional panels. In particular, we find a sharp dichotomy across specifications: (i) when unobserved effects vary along a single panel dimension, estimators are asymptotically unbiased, (ii) when they vary along two panel dimensions, estimators may suffer from a severe inference problem characterized by a degenerate asymptotic distribution. We resolve the latter by deriving explicit bias formulas and proposing analytically debiased estimators with nondegenerate, correctly centered asymptotic distributions. An empirical application studies dynamic network formation in a directed panel of bilateral trade relationships.