A simple and robust method of inference for spatial lag dependence. Working paper

A simple and reliable method of inference for the spatial parameter in spatial autore-gressive models is introduced, based on a statistic obtained by centering and rescaling the numerator of the concentrated Gaussian score function. The resulted tests and confi-dence intervals are robust against the distributional misspecifications and are insensitive to the spatial layouts and the error standard deviation. In contrast, the standard meth-ods based on Gaussian score and information matrix may lead to inconsistent inference when errors are nonnormal, and can be quite sensitive to the spatial layouts and the error standard deviation even when errors are normally distributed. Extensive Monte Carlo results are reported and an empirical illustration is given