Lasso-penalized clusterwise linear regression modeling with a two–step approach

Available approaches for the computation of lasso-penalized estimators of clusterwise linear regression models are time consuming and/or require approximate schemes. This makes the computation of the final solution, as well as the tuning of the penalty necessary to select the model, particularly cumbersome. To ease such computation, we introduce: 1) an expectation maximization algorithm with closed-form updates, that uses efficient formulas that are available for linear regression; 2) a new strategy to select the model based on a Least-Angle-RegreSsion (LARS) grid using standard information criteria.