Abstract:
The fixed effects (FE) panel model is one of the main econometric tools in empirical economic research. A major practical limitation is that the parameters on time-constant covariates are not identifiable. This paper presents a new approach to grouping FE in the linear panel model to reduce their dimensionality and ensure identifiability. By using unsupervised nonparametric density based clustering, cluster patterns including their location and number are not restricted. The approach works with large data structures (units and groups) and only clusters units that are sufficiently similar, while leaving others as unclustered atoms. Asymptotic theory and rates of convergence are presented. With the help of simulations and an application to economic data it is shown that the suggested method performs well and gives more insightful and efficient results than conventional panel models.