Skip to main content
QUICK REVIEW

[论文解读] Modelling Network Interference with Multi-valued Treatments: the Causal Effect of Immigration Policy on Crime Rates

Costanza Tortú, Irene Crimaldi|arXiv (Cornell University)|Feb 28, 2020
Advanced Causal Inference Techniques参考文献 52被引用 5
一句话总结

本文将网络干扰下的因果推断扩展至多值处理,采用加权网络结构,引入联合多重广义倾向得分(JMGPS)以估计直接效应并考虑溢出效应。研究发现,限制性移民政策会提高犯罪率,且干扰效应放大了该影响;而欢迎性政策则通过促进融合降低犯罪率。

ABSTRACT

Policy evaluation studies, which intend to assess the effect of an intervention, face some statistical challenges: in real-world settings treatments are not randomly assigned and the analysis might be further complicated by the presence of interference between units. Researchers have started to develop novel methods that allow to manage spillover mechanisms in observational studies; recent works focus primarily on binary treatments. However, many policy evaluation studies deal with more complex interventions. For instance, in political science, evaluating the impact of policies implemented by administrative entities often implies a multivariate approach, as a policy towards a specific issue operates at many different levels and can be defined along a number of dimensions. In this work, we extend the statistical framework about causal inference under network interference in observational studies, allowing for a multi-valued individual treatment and an interference structure shaped by a weighted network. The estimation strategy is based on a joint multiple generalized propensity score and allows one to estimate direct effects, controlling for both individual and network covariates. We follow the proposed methodology to analyze the impact of the national immigration policy on the crime rate. We define a multi-valued characterization of political attitudes towards migrants and we assume that the extent to which each country can be influenced by another country is modeled by an appropriate indicator, summarizing their cultural and geographical proximity. Results suggest that implementing a highly restrictive immigration policy leads to an increase of the crime rate and the estimated effects is larger if we take into account interference from other countries.

研究动机与目标

  • 解决现有因果推断方法的局限性,即假设处理为二值且忽略现实政策评估中的干扰。
  • 在单位相互影响的网络化人群中,对复杂多值处理(如不同水平的移民政策)进行建模。
  • 开发一种统计框架,利用加权网络结构同时考虑个体和网络层面的暴露。
  • 在控制文化与地理接近性作为干扰来源的前提下,估计移民政策对犯罪率的直接因果效应。
  • 通过分析各国移民政策与犯罪率的关系,展示该方法的实证相关性。

提出的方法

  • 提出一个多值处理框架,其中每个单位的处理沿多个维度定义,例如移民政策限制程度的水平。
  • 引入邻域处理暴露矩阵(NTEM),以加权文化与地理接近性的方式表示每个单位对其邻居处理的多变量暴露。
  • 开发联合多重广义倾向得分(JMGPS),以建模个体处理与多变量邻域暴露的联合分布。
  • 基于JMGPS使用逆概率加权法估计潜在结果,并计算不同处理对比下的直接因果效应。
  • 在网络环境中应用Horvitz-Thompson估计量,以在存在干扰的情况下获得直接效应的无偏估计。
  • 采用多值暴露映射,允许非二值、类连续的处理水平,并捕捉异质性溢出效应。

实验结果

研究问题

  • RQ1在考虑网络干扰的情况下,国家移民政策的限制程度在多大程度上影响犯罪率?
  • RQ2国家之间的文化与地理接近性在多大程度上调节移民政策对犯罪率的溢出效应?
  • RQ3当同时建模多值处理与网络干扰时,直接因果效应的大小如何?
  • RQ4当忽略干扰与正确建模干扰时,政策效应的估计值有何变化?
  • RQ5在观察性政策影响研究中,包含多值网络暴露是否能提高因果推断的准确性?

主要发现

  • 实施高度限制性的移民政策与犯罪率的统计显著上升相关,高限制性与低限制性对比的点估计为0.24439(95%置信区间:0.24246–0.24618)。
  • 欢迎性移民政策(LH或HL)对降低犯罪的直接效应显著,点估计分别为0.21145和0.25819,两者均高度显著(p < 0.001)。
  • 忽略干扰会导致处理效应的低估;纳入干扰后,估计效应增大,尤其在考虑文化接近性时更为明显。
  • 当将文化相似性作为溢出机制纳入时,估计效应进一步增强,表明跨国影响更强。
  • 仅以地理接近性作为溢出机制时,效应仍强于无干扰情况,但弱于纳入文化接近性时。
  • 稳健性检验表明,即使将LH与HL类别合并为单一的“中等”类别,结果依然成立,支持研究发现的稳定性。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。