中国省域犯罪率影响因素的空间非平稳性分析

收入差距和流动人口是影响犯罪率的两个重要因素, 以往研究基于OLS模型, 在假设地域空间为均质的前提下分析其对犯罪率的影响, 但现实世界的空间单元往往难以满足“均质”的假设, 多数表现为“空间异质”。以OLS计量空间异质会造成计量结果出现偏差, 同时无法了解不同空间单元的不同影响。而地理加权回归模型通过将空间结构嵌入线性回归模型中, 很好的解决了空间异质的计量问题。利用地理加权回归模型研究2008 年中国大陆省域单元犯罪率的影响因素, 结果表明:① 犯罪率的影响因素表现出空间非平稳性, 流动人口与犯罪率显著相关, 但各个省份相关程度并不相同, 影响关系随空间位置变化而变化;② 地理加权回归模型的计量精度和拟合度比OLS模型有大幅提高

Spatial non-stationarity of the factors affecting crime rate at province scale in China

Income inequality and floating population are two important factors affecting crime rate. One major problem of the previous studies is that they were all based on ordinary least squares (OLS) estimation with constant coefficients. OLS estimation presumes that the individuals are homogeneous and the relationship between the crime rate and the two affecting factors do not change over spatial units, which contradicts the fact that significant differences exist among the 31 provinces of China. In other words, the relationship between crime and income inequality and floating population is too complicated to be explained by ordinary least squares estimation with constant coefficients. Geographically weighted regression (GWR) is a powerful tool for exploring spatial heterogeneity. GWR recognizes that relationships between variables are likely to vary across space. Instead of estimating one parameter for each independent variable, GWR estimates local parameters. A parameter is estimated for each data location in the study area. In a GWR model, parameters are estimated using a weighting function based on distance so that locations closest to the estimation point have more influence on the estimate. Using geographically weighted regression model, this paper analyzes the local relationship between crime rate and income inequality and floating population in 31 provinces of China. The results show that: (1) The effects on crime rate are spatially non-stationary. The correlation between crime rate and income inequality is significant in some provinces, but not significant in some other provinces. The correlation between crime rate and floating population is significant in all provinces, but not with the same degree. (2) GWR model is more suited than OLS model, the AIC and R square are both improved in GWR model. This study demonstrates the usefulness of GWR for exploring local processes that drive crime rates and for examining the misspecifications of a global model of crime rate. The practical implication of GWR analysis is that different crime prevention policies should be implemented in different regions of China. Because of such a heterogeneity, criminal policy needs to suit the local situations.