Model Selection Methods for Social Interaction and Spatial Models

Abstract

In this project, we not only develop model selection (MS) methods for social interaction models, but also revisit Manski's (1993) fundamental problem of identifying different social effects. We argue that MS is particularly important in applying social interaction models by proving three theorems about model uncertainty (formal statements will be provided). 

Based on the misspecification resistant information criteria (MRIC) by Hsu et al. (2018, 2019), we develop predictive-efficient MS criteria, for finite sets of parametric, potentially misspecified, and non-nested social interaction and spatial models. We develop consistent nonlinear least squares and novel penalty estimators under general covariance or network structures. Whether the true model is included, the MRIC are asymptotically efficient, are consistent, and satisfy a parsimony principle (formal statements will be provided). From this principle, we derive a empirical modeling strategy under which the most parsimonious prediction-error-minimizing model is selected with probability approaching 1.

線上視訊請點選連結