Determining significant covariates is a fundamental problem in spatial regression analysis. However, parametric assumptions limit flexibility and can lead to inaccurate inference when misspecified. To address this, we propose a fully nonparametric testing procedure for spatial regression that does not impose restrictive model assumptions. Our approach follows a Monte Carlo testing framework through random shifts with both torus and variance correction. We construct test statistics based on the correlation between the residuals, where the effects of nuisance covariates have been removed, and the covariate of interest, allowing us to assess the significance of the covariate in the sense of partial correlation. This enables robust inference across various models as it does not require parametric assumptions about the distribution, spatial correlation structure, and form of the dependence or even a closed-form distribution of the test statistics. Our method is computationally stable compared to the parametric approaches that depend on numerical optimisation. Furthermore, we show that the random shift test with variance correction and sample covariance as the test statistic is asymptotically exact. Through extensive numerical experiments, we demonstrate that our method achieves the nominal significance level and exhibits power comparable to that of parametric methods, even when they are correctly specified.

翻译：暂无翻译