基于倾向得分匹配的缺失数据插补方法

The Missing Data Imputation Method Based on Propensity Score Matching

针对预测均值匹配中相近性刻画较为单一的问题,考虑多种相近性刻画方法,同时结合倾向得分可将多个协变量降维的特点,提出采用倾向得分匹配来对缺失数据进行插补的新方法:首先估计倾向得分,然后可选择最近邻、卡钳与半径、分层或区间等多种匹配方法进行匹配,最后利用匹配单元的目标变量来对数据缺失单元进行插补.进一步采用蒙特卡罗模拟和实际数据证实方法是有效的,且在均值插补、回归插补、随机插补、最近邻倾向得分匹配插补、卡钳与半径倾向得分匹配插补、分层或区间倾向得分匹配插补方法中分层或区间倾向得分匹配插补效果最好.

Since the measure of proximity in predictive mean matching was single, through considering various measure methods of proximity and the characteristic that propensity score could be used to reduce the dimension of multiple covariates, a new method to impute missing data using propensity score matching was proposed. The first step was estimating the propensity score, then choosing many matching methods like nearest neighbor, caliper and radius, stratification or interval, and finally imputing missing data using the target variable of the matching unit. Moreover, the results of a Monte Carlo simulation and a real dataset confirm that the proposed method is effective, and the results of stratification or interval propensity score matching are best among mean imputation, regression imputation, random imputation, nearest neighbor propensity score matching imputation, caliper and radius propensity score matching imputation as well as stratification or interval propensity score matching imputation.