|
Abstract:
|
Linear regression models are widely used in all fields to describe the relationship between a dependent response variable and a group of p independent predictor variables. In practice, choosing one final model from the 2^p possible models is often done via model-selection criteria or a greedy algorithm such as stepwise regression. This choice, although statistically "the best," does not take into account the variables' practicality (cost, availability, ease of measurement, etc.). We present Cluster-Based Modeling, a method that explores and visualizes the linear regression model space to find groups of models that predict similarly. Instead of one final model, we return a group of candidate models (that could contain the stepwise model choice) from which a proxy model more suitable to the application could be chosen. Results with real and simulated data sets will be presented.
|
