Is infrastructure capital really productive? Non-parametric modeling and data-driven model selection in a cross-sectionally dependent panel framework

This paper provides a broad replication of Calderón et al. (2015). We address some complex and relevant issues, namely functional form, non-stationary variables and cross-sectional dependence. In particular, by adopting the CCE framework, we consider both parametric – static and dynamic - and non-parametric specifications, thus allowing for different degrees of flexibility. Contrary to Calderón et al. (2015), we find a lack of significance of the infrastructure index, with an estimated elasticity very close to zero for all estimates. Moreover, by employing the data-driven model selection procedure proposed by Gioldasis et al. (2021), it is found that non-parametric specifications provide the best predictive performance and that CCE models always overperform with respect to traditional panel data methods that employ cross-sectional demeaning to account for cross-sectional dependence.