A model for eliminating aggregate and specification bias in multivariate fine-scale urban scenarios

Abstract

The spatial relationships between predictors and responses are influenced by their frequency and spatial distribution. Ecological bias in regression models can occur due to the aggregate frequency and clustering of independent variables, leading to false, over-, or underestimations. This can be exacerbated by an increase in data resolution, complexity, and variable count, as is often the case in urban research scenarios. To address this issue, a new relationship-estimation model called the Ecologically Corrected Spatial Relationship Estimator (ECSRE) was proposed and compared to Geographically Weighted Regression (GWR). The results showed that ECSRE outperformed GWR by correctly revealing pre-planned relationships in simulated data, presenting a lower influence of aggregate frequencies on the outcome, better suppression of specification errors, higher R2 scores, and better randomness of residuals.