| Summary: | 碩士 === 國立臺北大學 === 統計學系 === 102 === With more and more emphasis on handling missing data, missing data can not only be rounded up or replaced by mean, mode, instead, should be to select the appropriate imputation through the missing pattern. It will not only solve the affect by missing data, but also to enhance the credibility and efficiency of the analysis results.This study use the first wave to the third wave of survey data in "Taiwan Education Panel Survey" to investigate the effect of missing treatment.Since the third wave of student data in the "Core Panel", the proportion of missing data samples are relatively high.Using this type of the missing data to analyze, would better reveal the importance of missing data processing methods.
In the imputation, we start to find significant variables that may affect the missingness, and then we use of these significant variables to construct 50 groups imitation missing data sets from the complete data set.Then we compare the four missing treatment methods (list-wise deletion, discriminant function imputation, logistic regression imputation , Monte Carlo - Markov chain single imputation) regarding the changes and influence in variables and coefficients of the binary logistic regression model.The results show that Monte Carlo - Markov chain single imputation method has relatively better performance and more stable imputation methods.Finally, we use Monte Carlo - Markov chain single imputation to imputate the original data and analysis, and discuss whether the length of the tutorial can enhance learning through multiple regression analysis and the average treatment effect on the treated (ATT) by completed data after imputated.
|
|---|
