On the Use of $L$-functionals in Regression Models

TL;DR

This paper surveys and unifies a broad class of $L$-functionals in regression, generalizing concepts like $L$-moments and introducing order numbers via orthogonal series, applicable to various models including censored data.

Contribution

It introduces a unified framework for $L$-functionals in regression, including their order numbers and construction from orthogonal polynomials, with applications to transformed linear models and censored data.

Findings

Unified asymptotic theory for $L$-functionals estimates.

Application to bird migration arrival time data.

Introduction of a novel $L$-based coefficient of determination.

Abstract

In this paper we survey and unify a large class or $L$-functionals of the conditional distribution of the response variable in regression models. This includes robust measures of location, scale, skewness, and heavytailedness of the response, conditionally on covariates. We generalize the concepts of $L$-moments (Sittinen, 1969), $L$-skewness, and $L$-kurtosis (Hosking, 1990) and introduce order numbers for a large class of $L$-functionals through orthogonal series expansions of quantile functions. In particular, we motivate why location, scale, skewness, and heavytailedness have order numbers 1, 2, (3,2), and (4,2) respectively and describe how a family of $L$-functionals, with different order numbers, is constructed from Legendre, Hermite, Laguerre or other types of polynomials. Our framework is applied to models where the relationship between quantiles of the response and the covariates follow a transformed linear model, with a link function that determines the appropriate class of $L$-functionals. In this setting, the distribution of the response is treated parametrically or nonparametrically, and the response variable is either censored/truncated or not. We also provide a unified asymptotic theory of estimates of $L$-functionals, and illustrate our approach by analyzing the arrival time distribution of migrating birds. In this context a novel version of the coefficient of determination is introduced, which makes use of the abovementioned orthogonal series expansion.