R-lunches - Topics in machine learning - Piecewise linear regression with splines

03.12.2019 12:15 – 13:45

A common problem in data analysis is how to model nonlinear associations. The standard linear regression model can be extended by adding curvilinear transformations to the data (e.g., a quadratic effect) but researchers often find such effects difficult to interpret. An alternative to a curve is to build a piecewise linear model, consisting of several linear slopes joined by knot points. These connected slopes retain their linear interpretation locally, and are referred to as "splines". In this presentation, I will introduce spline models and illustrate their application to modelling of bipolar variables, threshold effects, and longitudinal change. Following this, I present multivariate adaptive regression splines (MARS), a popular machine learning variant of spline models (Friedman, 1991), which automates modelling choices such as which variables to include and how many splines to use. The R package "earth" (Milborrow, 2019) currently offers the most comprehensive version of MARS and will be a main focus of the presentation.

Lieu

Bâtiment: Uni Mail

M 1150

entrée libre