Pervasive impact of spatial dependence on predictability

TL;DR

This paper demonstrates that spatial dependence significantly affects the accuracy of predictions in various domains, highlighting its pervasive influence beyond geographic spaces and across different types of data.

Contribution

It introduces a comprehensive analysis of spatial dependence in both geographic and non-geographic spaces, proposing new methods to capture its impact on prediction models.

Findings

Spatial dependence influences prediction outcomes across diverse datasets.

Geographic space shows stronger spatial dependence for environmental variables.

Spatial dependence also exists in non-geographic, attribute-based dimensions.

Abstract

Understanding the complex nature of spatial information is crucial for problem solving in social and environmental sciences. This study investigates how the underlying patterns of spatial data can significantly influence the outcomes of spatial predictions. Recognizing unique characteristics of spatial data, such as spatial dependence and spatial heterogeneity, we delve into the fundamental differences and similarities between spatial and non-geospatial prediction models. Through the analysis of six different datasets of environment and socio-economic variables, comparing geospatial models with non-geospatial models, our research highlights the pervasive nature of spatial dependence beyond geographical boundaries. This innovative approach not only recognizes spatial dependence in geographic spaces defined by latitude and longitude but also identifies its presence in non-geographic, attribute-based dimensions. Our findings reveal the pervasive influence of spatial dependence on prediction outcomes across various domains, and spatial dependence significantly influences prediction performance across all spaces. Our findings suggest that the strongest spatial dependence is typically found in geographic space for environment variables, a trend that does not uniformly apply to socio-economic variables. This investigation not only advances the theoretical framework for spatial data analysis, but also proposes new methodologies for accurately capturing and expressing spatial dependence under complex conditions. Our research extends spatial analysis to non-geographic dimensions such as social networks and gene expression patterns, emphasizing the role of spatial dependence in improving prediction accuracy, thereby supporting interdisciplinary applications across fields such as geographic information science, environmental science, economics, sociology, and bioinformatics.