Model averaging for spatial autoregressive panel data models

Abstract

The spatial autoregressive panel data models are widely employed in regional economics to capture spatial dependencies, but conventional specifications rely on a single spatial weight matrix, heightening the risk of model misspecification. Current research lacks systematic model averaging methods for integrating multiple weight matrices and addressing spatial effect uncertainty. This study proposes a novel model averaging framework for spatial autoregressive panel data models with fixed effects, extending model averaging methodology to the spatial panel context and enabling flexible combinations of multiple weight matrices for both dependent variables and error terms. An adaptive Mallows-type criterion is developed, dynamically adjusting to the presence or absence of spatial effects, with its asymptotic optimality established. Monte Carlo simulations confirm robustness across scenarios with no, single, or mixed spatial dependencies. An empirical application to Chinese provincial housing prices identifies economic adjacency as the key spatial dependence driver, validating the method’s predictive accuracy and policy utility for spatiotemporal data analysis.