Preprint L1-penalised regression

A preprint titled “Natural coordinate descent algorithm for L1-penalised regression in generalised linear models” is available on the arXiv. L1-penalised regression refers to the use of a penalty term that sums the absolute effect sizes of all parameters in a high-dimensional regression model to select the most important variables in the model. This paper uses a simple result from convex analysis to show that a solution mechanism (“soft-thresholding”) previously known for least-squares regression is in fact generic and also holds for regression with for instance binary or count-based data.