This study proposes the adoption of a neural network as an alternative to logistic regression analysis, which is conventionally used to estimate the propensity score (Rosenbaum & Rubin, 1983). Moreover, covariates that are frequently obscured are presented.
Considering the response pattern to a mail survey by random sampling as a criterion, we examined how is the response pattern to a Web survey by purposive selection rectified using the propensity score. The propensity score was estimated using the subjects' demographic variables as covariates.
The results of adopting a neural network were compared with those of the logistic regression analysis. As a result, the accuracy of bias reduction by the threelayer neural networks was found to be greater than that by the logistic regression analysis.
In addition, detailed contents of the covariates were presented, and a decision tree was produced to examine the influence of covariates on allocation of the subjects to survey forms.
抄録全体を表示