王 彤-回归分析中的高维变量选择.ppt

Extensions of Marginal Correlation Extensions of Model The idea of penalized regression and ISIS is widely applicable. It can be applied to Classification Categorical data analysis Survival analysis Nonparametric learning Robust and quantile regression Work of Our Team Big Data in Medicine Interesting Questions: Disease classification, such as for diagnosis Predict clinical outcome, such as for personalized treatment Association between phenotypes and SNPs or eQTL Analysis of interaction, such as gene-gene, gene-environment ... Our team focus on predicting clinical outcome of cancer and make comparative studies on prediction of different approaches. 1.SPCA PLS Regression Supervised principle component analysis (SPCA) Step1: find variables with high correlation to survival by likelihood ratio test Step2: successively extracting factors to capture the maximum of the variance of the variables in step 1 Step3: fit a Cox model for extracted component and survival Partial least square (PLS) regression Successively extracting linear combinations of variables which best model survival. Fit a Cox model for extracted component and survival Simulation Study Criteria Small deviance and large R-square indicate better prediction. Conclusion R2 increase as variance increase when correlation equal R2 increase as correlation increase when variance equal R2 decrease as censor rate increase R2 of SPCA is similar to that of PLS regression Simulation design n=100, p=1000 includes 10 equal blocks, distribute N(0,Σ) Independent between blocks and equal correlation within block : ρ=0.3,0.6,0.9 Variance of first two blocks are 0.5,1,2 Coefficients of 1st block are 0.01, Coefficients of 2nd block range from 0.01 to 0.1 by 0.01 Variance of last 8blocks are 1,coefficients of them are 0, Censor rate: 0.2,0.5 training set and test set in 2:1 ratio, simulate 200 times Real Data Analysis Breast cancer data (Van ‘t Veer et al, 2002) n=78,p=4751, censor=56.4% Lung cancer data (Beer e