Sieve Estimation of Time-Varying Panel Data Models with Latent Structures
Abstract
We propose a heterogeneous time-varying panel data model with a latent group structure that allows the coefficients to be varying over both individuals and time. We assume that the coefficients change smoothly over time and form different unobserved groups. When treated as smooth functions of time, the individual functional coefficients are heterogeneous across groups but homogeneous within a group. We propose a penalized-sieve-estimation-based classifier-Lasso (C-Lasso) procedure to identify the individuals’ membership and to estimate the group-specific functional coefficients in a single step. The C-Lasso serves to shrink individual functional coefficients to the unknown group-specific functional coefficients. The classification exhibits the desirable property of uniform consistency. The C-Lasso estimators and their post-Lasso versions achieve the oracle property so that the group-specific functional coefficients can be estimated as well as if the individuals’ membership were known. Simulations demonstrate excellent finite-sample performance of the approach in both classification and estimation. We apply our method to study the heterogeneous trending behavior of GDP per capita across 91 countries for the period 1960-2012 and find four latent groups.