An investigation of sliced inverse regression with censored data

An Investigation of Sliced Inverse Regression with Censored Data Daniel Riggs August,62010 The complexity of high-dimensional data creates a number of concerns when trying to analyze it. This data often consists of a response or survival time and potentially thousands of predictors. These predictors can be highly correlated, and the sample size is often very small and right censored. Sliced inverse regression(SIR) is a method of reducing the dimension of the data, while preserving all the regression information. Sliced inverse regression with regularizations was developed to work when the number of predictors exceeds the sample size, and to deal with highly correlated predictors as well. In this study we investigated the performance of Sliced inverse regression with regularizations using three different approaches for handling right censored data. The methods of reweighting, mean imputation, and multiple imputation were analyzed. Based on the simulation scenarios, the mean imputation method performs the best in regards to fitting the data as well as prediction. The method of reweighting appears inadequate when combined with SIR.