基于倾向得分广义线性模型的非概率抽样统计推断研究

Research on Statistical Inference of Non-probability Sampling Based on Propensity Score Generalized Linear Model

候选者数据库网络调查下非概率抽样的统计推断问题是网络调查发展中迫切需要解决的问题.提出基于倾向得分广义线性模型的非概率抽样推断方法：将网络候选者数据库的调查样本与参考样本结合,建立Logistic、Probit、C-log-log三种广义线性模型来估计倾向得分,并对网络候选者数据库的调查样本进行倾向得分未加权比例的分组调整与倾向得分加权比例的分组调整来估计总体.研究结果表明：基于倾向得分广义线性模型的总体估计效果较好,并且使用调查权数的Logistic与C-log-log倾向得分未加权比例的分组调整估计最为稳健.

How to solve the statistical inference problem of non-probability sampling under web surveys of candidate database is an urgent problem to be solved in the development of web survey. The inference method of non-probability sampling based on propensity score generalized linear model is proposed. A survey sample of the web candidate database and a reference sample are firstly combined. Three generalized linear models, logistic, probit and c-log-log, are then built up to estimate propensity scores. Class adjustments of propensity score unweighted proportion and weighted proportion are lastly applied in the survey sample of the web candidate database to estimate the population. The research results show that the population estimators based on propensity score generalized linear model are good. The class adjustment estimators of logistic and c-log-log propensity score unweighted proportion using survey weights are most robust.