Inference after discretizing unobserved heterogeneity

We consider a linear panel data model with nonseparable two-way unobserved heterogeneity corresponding to a linear version of the model studied in Bonhomme et al. (2022). We show that inference is possible in this setting using a straightforward two-step estimation procedure inspired by existing discretization approaches. In the first step, we construct a discrete approximation of the unobserved heterogeneity by (k-means) clustering observations separately across the individual (i) and time (t) dimensions. In the second step, we estimate a linear model with two-way group fixed effects specific to each cluster. Our approach shares similarities with methods from the double machine learning literature, as the underlying moment conditions exhibit the same type of bias-reducing properties. We provide a theoretical analysis of a cross-fitted version of our estimator, establishing its asymptotic normality at parametric rate under the condition max(N, T) = o(min(N, T) 3 ). Simulation studies demonstrate that our methodology achieves excellent finite-sample performance, even when T is negligible with respect to N.