Group structure estimation for panel data -- a general approach

Yu, Lu, Gu, Jiaying, Volgushev, Stanislav

arXiv.org Machine Learning 

Panel data models are a standard empirical tool in statistics, economics, marketing, and financial research. The conventional modeling approach is to assume that all individual heterogeneity can be summarized by an individual specific intercept, often known as the fixed effects, while assuming all covariates have a common effect among all the individuals, such that information can be pooled across individuals to gain efficiency of these common parameters. However, heterogeneous responses towards observed control variables are often better supported by empirical evidence, especially as detailed individual level data becomes more available. An increasingly popular approach to model unobserved heterogeneity in the effects of covariates on individual responses is to assume the existence of a finite number of homogeneous groups.