Bootstrap On Confounder Seletion

• Main Authors: Hong, Jian-Xiang, 洪建祥
• Other Authors: Lian Ie-Bin
• Format: Others
• Language: zh-TW
• Published: 1997
• Online Access: http://ndltd.ncl.edu.tw/handle/43381991455306475819

碩士 === 國立彰化師範大學 === 數學系 === 85 === Minimizing mean squared error ( MSE ) or expected prediction error ( PE) is a frequently used criterion for variable selection procedure in linear model. MSE or PE can be estimated by Cp method ( Mallows, 1973) bootstrap procedure (Efron, 1982 ), and some modified bootstrap procedures (Shao, 1996 ). These methods may perform well in prediction. However, sometimes economists and epidemiologists are more interested in estimating and making inferences about the effect of certain risk factor. In this case, it is important to select the confounders of the risk factor into the model for controlling, rather than just to select the important covariates of the response. In this work, we propose a "best subset" criterion for confounder selection procedure by minimizing the mean squared error of the risk factor. This value is estimated by both bootstrap and naive methods. We compare the selection procedures based on these two methods with conventional Cp method by simulation.