Revisiting the random shift approach for testing in spatial statistics

TL;DR

This paper revisits the random shift method for independence testing in spatial statistics, proposing improved permutation strategies that reduce liberality and enhance test power, especially in the presence of spatial correlation and clustering.

Contribution

It introduces variance correction strategies for the random shift method, improving its accuracy and power in spatial and point process independence tests.

Findings

Variance correction with factor 1/n improves test accuracy.

New permutation strategies reduce liberality in the random shift approach.

Proposed methods perform well in clustered point patterns.

Abstract

We consider the problem of non-parametric testing of independence of two components of a stationary bivariate spatial process. In particular, we revisit the random shift approach that has become a standard method for testing the independent superposition hypothesis in spatial statistics, and it is widely used in a plethora of practical applications. However, this method has a problem of liberality caused by breaking the marginal spatial correlation structure due to the toroidal correction. This indeed causes that the assumption of exchangability, which is essential for the Monte Carlo test to be exact, is not fulfilled. We present a number of permutation strategies and show that the random shift with the variance correction brings a suitable improvement compared to the torus correction in the random field case. It reduces the liberality and achieves the largest power from all investigated variants. To obtain the variance for the variance correction method, several approaches were studied. The best results were achieved, for the sample covariance as the test statistics, with the correction factor $1/n$. This corresponds to the asymptotic order of the variance of the test statistics. In the point process case, the problem of deviations from exchangeability is far more complex and we propose an alternative strategy based on the mean cross nearest-neighbor distance and torus correction. It reduces the liberality but achieves slightly lower power than the usual cross $K$-function. Therefore we recommend it, when the point patterns are clustered, where the cross $K$-function achieves liberality.