A comparison of spline methods in R for building explanatory models

Splines offer high flexibility for modelling complex variable forms for continuous covariates but this flexibility requires the user to have a good understanding of how to select an appropriate spline function and how to tune parameters to obtain an optimal fit. In a previous project we have identified all available R packages that can be used to fit splines within a regression model. R packages often come with examples and vignettes, that offer users a cookbook approach in spline fitting. That may add to the confusion of non-experts.

Topic group 2 of the STRATOS initiative is working on a project to thoroughly compare approaches and provide guidance on best practice for using splines to correctly identify functional forms and variable selection in model building. In this work we will present a comparison of the most commonly used spline procedures and their corresponding packages. We will focus on thin plate regression splines, natural splines, b-splines  and p-splines under mgcv, splines, rms and gamlss packages.  We will work on simulated and real datasets in order to investigate under which conditions each one of these approaches should be preferred, we will evaluate their ability to identify the correct functional form, consider prediction errors and investigate how these approaches can be used in a multivariable setting when selection of variables is of interest. Finally, we will discuss ease of use, computational speed and efficiency.

This seminar is present on behalf of TG2 of the STRATOS initiative.