B1435
Title: Expanding the boundaries of generalized additive modelling with shape constraints in R
Authors: Natalya Pya Arnqvist - Umea University (Sweden) [presenting]
Per Arnqvist - Umea University (Sweden)
Abstract: Exploring the relationships between a response variable and multiple predictors through flexible semi-parametric regression modelling approaches can sometimes lead to excessive flexibility and implausible results. When analyzing such relationships, it is often reasonable to assume that some adhere to specific shape constraints, like monotonicity or convexity. A past study introduced a comprehensive framework for shape-constrained generalized additive models known as SCAM. This framework demonstrated its effectiveness and utility across diverse application domains, spanning ecological and environmental studies, medicine, genetic research, biotechnology, public health, and sustainability analysis. Expansions to the SCAM framework are introduced and implemented within the R package scam. Scam empowers the imposition of various constraints on smooth model components beyond monotonicity and convexity. The framework now accommodates the inclusion of linear functionals of smooth terms, either with or without shape constraints. This is known as a scalar-on-function regression in functional data analysis. The extended SCAM framework readily handles short-term temporal and spatial autocorrelation in the residuals and linear random effects terms. Furthermore, more robust schemes for smoothing parameter estimation for SCAM will also be presented.
