网络观测数据下考虑邻域处理的因果推断模型研究

Causal Inference Models for Neighborhood Treatment under Network Observation Data

因果效应评估是理解变量之间相互作用及因果机制的理论基础,然而在社交网络中进行因果效应评估一直是一个备受挑战的研究领域。因为网络中个体间的相互影响会使得传统的因果推断方法变得具有挑战性。文献中考虑邻域处理产生网络效应的研究较少,而且仅仅考虑二元处理下的因果推断,未能充分应对多重水平的处理情境。但现实观察性研究中普遍存在多重水平处理。文章提出了网络观测数据下考虑邻域处理及其影响的策略评估的模型和方法。首先,放松SUTVA条件,提出考虑邻域处理影响的个体处理稳定性假设。基于此构建适用于多重水平处理和邻域处理效应的因果推断理论框架。其次,改进原先使用倾向得分进行因果推断的模型,提出包含邻域处理的广义倾向得分。最后,通过数值模拟和理论分析,研究不同处理分配机制对个体结果的影响,并证明了该模型的稳定性和有效性。该模型为网络数据中的因果推断提供了一种有效方法,并具有广泛的应用前景和实际应用价值。

Causal effect estimation forms the theoretical foundation for understanding interactions between variables and underlying causal mechanisms.However,conducting causal inference in social networks remains a challenging research area due to the interdependence among individuals,which complicates the application of traditional causal inference methods.Existing studies considering network effects from neighborhood treatments are limited,and many focus solely on binary treatments,failing to adequately address multi-level treatment scenarios.However,multi-level treatments are common in observational studies.This paper proposes a model and method for evaluating strategies under network observational data,accounting for neighborhood treatment effects.Firstly,relax the traditional Stable Unit Treatment Value Assumption(SUTVA)and introduce a modified individual treatment stability assumption that incorporates neighborhood treatment effects.On the basis,a causal inference framework is constructed,which is suitable for multi-level treatments and neighborhood effects.Secondly,improve the original propensity score method by introducing a generalized propensity score model that accounts for neighborhood treatment effects.Finally,numerical simulations and theoretical analysis are conducted to study the impact of different treatment assignment mechanisms on individual outcomes.The simulation results demonstrate that the proposed model exhibits good stability and robustness in handling both neighborhood effects and multi-level treatments.The model offers a robust method for causal inference in network data and shows broad potential for practical application.