盗窃犯罪易发地的构成、影响及其区域异质性：基于地理探测器方法

本文旨在识别盗窃犯罪易发地的主要类型、影响强度及其区域异质性特征，同时厘清地理探测器在犯罪地理研究中的适用性。利用北京中心城区的多源大数据，研究发现：① 在全域范围内，37类设施均属显著的盗窃易发地，道路边界、行政边界、用地边界也是盗窃高发的边缘地带，影响力突出的盗窃易发地更具备犯罪模式理论提及的犯罪引致条件。② 如社会解组理论所言，随着社会解组程度减弱，犯罪易发地的类型数量及其对盗窃分布的影响逐渐下降。除低解组区外，各类设施及边界的相对影响排序基本相同。③ 因未考虑混淆因素干扰，地理探测器容易高估解释变量的作用，与负二项回归结果的吻合度不高。而结合排除控制法，将混淆因素同质化的样本进行分区分析是规避外生变异和稳健运用地理探测器的可行路径。

Composition,influence, and regional heterogeneity of theft places: Based on geographical detector

This study aims to identify the main types, their magnitudes of influence, and their characteristics of regional heterogeneity of theft places, and take the lead in discussing the application of geographical detector in the field of crime geography analysis. Using criminal judgement, Points of Interest, Location-based Service data, road network, census data, housing data, and other data in the central urban area of Beijing and taking the 1 km×1 km square grid as our spatial analysis unit, this research demonstrates that thirty-seven types of urban facilities belong to the major crime generators, crime attractors or crime enablers at the global level, and the edge areas of roads, administrative districts and land uses are also proved to be evident theft locations. Among them, the more influential theft locations are more aligned with the crime generating conditions proposed by crime pattern theory, including accessible crime target and lower risk of arrest. Second, as stated by social disorganization theory, if the level of social disorganization composed by mobility factor and housing factor decreases, the number of theft location types as well as their impacts will gradually diminish. But except for the low social disorganization districts, the rankings of relative influence of various facilities and edge areas on theft crime remain relatively stable. Third, geographic detector has the advantages of no linear hypothesis, no collinearity problem, clear physical meaning and so on. But owing to the neglect of the potential interference by confounding factors, geographical detector would probably overestimate the impacts of independent variables. It can be testified by the moderate correlation between the rankings of the overall influence of theft locations obtained from the negative binomial regression and those calculated by the geographical detector. We suggest that combining elimination control method with geographical detector is a feasible path to achieve robust results, so classification discussion after grouping units with similar confounding factors is essential to overcome the potential endogeneity problem if we use geographical detector in crime geography as well as other sub-fields of human geography.