Efficient and proper generalised linear models with power link functions

The generalised linear model is a flexible predictive model for observational data that is widely used in practice as it extends linear regression models to non-Gaussian data. In this paper, we introduce the concept of a properly defined generalised linear model by requiring the conditional mean of the response variable to be properly mapped through the chosen link function and the log-likelihood function to be concave. We provide a comprehensive classification of proper generalised linear models for the Tweedie family and its popular subclasses under different link function specifications. Our main theoretical findings show that most Tweedie generalised linear models are not proper for canonical and log link functions, and identify a rich class of proper Tweedie generalised linear models with power link functions. We provide a novel interpretability methodology for power link functions that is mathematically sound and very simple, which could help the adoption of such a link function that has not been used much in practice for its lack of interpretability. Using self-concordant log-likelihoods and linearisation techniques, we provide novel algorithms for estimating several special cases of proper and not proper Tweedie generalised linear models with power link functions. The effectiveness of our methods is determined through an extensive numerical comparison of our estimates and those obtained using three built-in packages, MATLABfitglm, Rglm2 and Pythonsm.GLM libraries, which are all implemented based on the standard Iteratively Reweighted Least Squares method. Overall, we find that our algorithms consistently outperform these benchmarks in terms of both accuracy and efficiency, the largest improvements being documented for high-dimensional settings. This is concluded for both simulated data and real data, which shows that our optimisation-based GLM implementation is a good alternative to the standard Iteratively Reweighted Least Squares implementations available in well-known software.