Semiparametric panel data models using neural networks

Abstract

This paper presents an estimator for semiparametric models that uses a feed-forward neural network to fit the nonparametric component. Unlike many methodologies from the machine learning literature, this approach is suitable for longitudinal/panel data. It provides unbiased estimation of the parametric component of the model, with associated confidence intervals that have near-nominal coverage rates. Simulations demonstrate (1) efficiency, (2) that parametric estimates are unbiased, and (3) coverage properties of estimated intervals. An application section demonstrates the method by predicting county-level corn yield using daily weather data from the period 1981-2015, along with parametric time trends representing technological change. The method is shown to out-perform linear methods such as OLS and ridge/lasso, as well as random forest. The procedures described in this paper are implemented in the R package panelNNET.