| Summary: | 碩士 === 國立中正大學 === 政治學所 === 93 === Abstract
This research applied the data from Taiwan’s Election and Democratization Study, 2004: The Presidential Election (TEDS 2004P) for studying “item nonresponse”. The main purpose is to establish standard procedures for discovering how “nonresponse” formulated and handling these nonresponses in survey. Firstly, by using the probit model, it discusses the formulation of item nonresponse. Secondly, it manages the item nonresponse with applications of voting models and compares the different prediction results from the three models: probit model, selection bias model and multiple imputation.
The research shows that, in the study of TEDS2004P, the factors causing nonresponse for interviewees on answering the presidential votes are “media use” and “party identification”. These two variables serve as the variables in selection equation for selection bias model. In addition, on processing the nonresponse, the probit model is used for models without further modifications and election bias model deals with the nonresponses that are not appeared at random. As for the multiple imputation, it assumes the nonresponses are all at random and takes further steps. The overall research design compares to process or not, the results of the nonresponses whether are missing at random, MAR, or not missing at random, NMAR.
However, it is impossible to know the parameter in survey. Therefore, it uses the prediction of candidates’ voting rates as comparison standards. The results show that selection bias model can best reveal the nonresponses and, with the assumption of NMAR, it gets the closest prediction. The results are not surprising at all. According to the survey time of TEDS2004P and the social environment at the time, there are great possibilities to cause the nonresponses not at random. For this reason, the time and spatial factors of a survey could influence the prediction results of models.
|
|---|
