Causal Inference Considering Neighborhood Treatment Under Network Observation Data

The development and evaluation of public policy requires data statistical analysis as a support.Central to policy assessment is the study of how the implementation of a policy relates to causal effects that affect outcomes.The implementation allocation of real-life social policies is not random,the allocation is affected by the characteristic nature of individuals or cities,and the observed data also have network characteristics such as information interactions,social behaviors between people and others that may affect the intervention outcomes,which is a big statistical challenge in using causal inference to evaluate the intervention effects of policies.There are two issues worth considering for policy analysis research using network observational data: first,spillover effects arising from correlating individuals;second is that complex policies should not simply be classified as binary treatments when considering neighborhood treatment effects to be analyzed in the research.The data used in policy research are interactive,traditional causal policies assessment method are not applicable to the network observation data,which bring challenges and difficulties to the persuasiveness of the policy assessment results.Previous studies on the spillover effect of observation data mostly focused on binary treatment variables.This thesis further studied the causal inference model considering neighborhood effect in network observation data,and proposed a causal effect estimation method using the individual and neighborhood joint tendency score under multi-level treatment.This model can be used to evaluate complex policy treatment(such as multiple levels of single treatment or multiple continuous treatment)under the consideration of neighborhood treatment Unlike previous studies on epidemic prevention measures against COVID-19,this thesis also considers the impact of epidemic prevention policies of neighboring cities to improve the credibility of policy evaluation.In view of the above situation,this thesis develops a multi-level causal inference model and estimation method considering the network effect brought by neighborhood treatment based on the previous research on the spillover effect of observational data,and evaluates the closure policy.The specific work is as follows:First,the intervention effects considered in the policy evaluation involve multilevel individual treatment and neighborhood treatment.Different from the previous scholars’ research on spillover effects based on the stable unit treatment value assumption(SUTVA),the basic assumption in causal inference is relaxed,a hypothesis of unit treatment stability on the neighborhood considering causal effects arising from neighborhood nodes is proposed.Based on this,a theoretical framework for causal inference considering neighborhood treatment effects suitable for multilevel treatment is built.Second,a new generalized propensity score is proposed based on a proposed causal inference model that considers neighborhood treatment.This generalized propensity score can be regarded as the joint of individual propensity scores as well as neighborhood propensity scores.Under the joint propensity score,an individual propensity score is defined as the propensity score under the multilevel treatment.Demonstrated that the unconfounded assumption holds separately for the proposed individual,neighborhood,as well as joint propensity scores,and took a suitable approach to estimate their corresponding propensity scores for different properties of individual treatment and neighborhood treatment.In addition,numerical simulations were conducted for the method proposed in this thesis,to investigate the effect of neighborhood treatment on individual outcomes by using different treatment allocation mechanisms,and to evaluate the performance of this model.Through the causal policy evaluation model proposed in this thesis,the causal effect policy evaluation of the closure measures of 11 cities in Zhejiang Province at the beginning of 2020 was carried out.Considering two different definitions of neighbor cities,the empirical analysis shows that the closure measures of neighbor cities will have an impact on the newly increased epidemic situation in the city.Compared with the estimation results without considering neighborhood treatment,the deviation is small and the estimation results are better.The last part of the thesis is the summary and future work directions.