Research Seminar
Clustering Covariates RegressionEva Vande GaerKU Leuven | |
| Abstract: | Linear regression is a much applied technique in many research fields. Its aim is to predict one or more dependent variables on the basis of a number of independent variables. However, when analyzing data sets with very many independent variables, some of which are highly correlated, one may face the bouncing beta problem: Regression weights obtained for such data sets tend to be unstable, in that small changes in the data under study can lead to completely different regression weights. To solve the bouncing beta problem, many solutions have already been suggested. Roughly, two types of solutions can be distinguished: variable selection and dimension reduction methods. Examples of the former include ridge regression, the LASSO, and elastic nets, whereas the latter entails methods like principal component regression, reduced rank regression and principal covariates regression. However, the interpretation of the solutions obtained by these methods is not always straightforward. As a possible alternative, we therefore propose the Clustering Covariates Regression method (CCovR). This method simultaneously partitions the independent variables into a few predictor types and regresses the dependent variable(s) on these types. In this talk, we first introduce the CCovR method and evaluate its performance by means of a small simulation study. Next, we compare CCovR and some variable selection and dimension reduction methods by applying them to the same data set. |
| Date: | Tue Mar 15, 12:15 pm - 1:15 pm |
| Place: | room 02.60 (Department of Psychology, Tiensestraat 102, 3000 Leuven) |