Covariate-distance Weighted Regression (CWR): A Case Study for Estimation of House Prices
Authors:
Hone-Jay Chu,
Po-Hung Chen,
Sheng-Mao Chang,
Muhammad Zeeshan Ali,
Sumriti Ranjan Patra
Abstract:
Geographically weighted regression (GWR) is a popular tool for modeling spatial heterogeneity in a regression model. However, the current weighting function used in GWR only considers the geographical distance, while the attribute similarity is totally ignored. In this study, we proposed a covariate weighting function that combines the geographical distance and attribute distance. The covariate-di…
▽ More
Geographically weighted regression (GWR) is a popular tool for modeling spatial heterogeneity in a regression model. However, the current weighting function used in GWR only considers the geographical distance, while the attribute similarity is totally ignored. In this study, we proposed a covariate weighting function that combines the geographical distance and attribute distance. The covariate-distance weighted regression (CWR) is the extension of GWR including geographical distance and attribute distance. House prices are affected by numerous factors, such as house age, floor area, and land use. Prediction model is used to help understand the characteristics of regional house prices. The CWR was used to understand the relationship between the house price and controlling factors. The CWR can consider the geological and attribute distances, and produce accurate estimates of house price that preserve the weight matrix for geological and attribute distance functions. Results show that the house attributes/conditions and the characteristics of the house, such as floor area and house age, might affect the house price. After factor selection, in which only house age and floor area of a building are considered, the RMSE of the CWR model can be improved by 2.9%-26.3% for skyscrapers when compared to the GWR. CWR can effectively reduce estimation errors from traditional spatial regression models and provide novel and feasible models for spatial estimation.
△ Less
Submitted 14 May, 2023;
originally announced May 2023.
