Adjusting for Nonresponse Bias Using Logistic Regression
提出一种仅利用应答者信息、通过逻辑回归模型估计应答概率的方法，并用Horvitz-Thompson型估计量减少样本矩估计的偏差，模拟表明该方法在应答概率与调查变量强相关时能降低均方误差。
A method of adjustment for nonresponse is considered. No information about the nonrespondents is required, but the respondents are assumed to answer all questions of interest. One or more call-backs are assumed. The probabilities of response are represented by a logistic regression model, in which variables in the survey are explanatory variables. The probabilities are estimated using a modified conditional maximum likelihood approach, based on the respondents. The estimated probabilities are used in a Horvitz-Thompson type estimator to reduce bias in the estimation of sample moments. Simulations suggest that the method reduces the mean squared error of estimation when response probabilities depend strongly on variables in the survey.