A distributional regression approach to income-related inequality of health in Australia

Abstract

Background: Several studies have confirmed the existence of a significant positive relationship between income and health. Conventional regression techniques such as Ordinary Least Squares only help identify the effect of the covariates on the mean of the health variable. In this way, important information of the income-health relationship could be overlooked. As an alternative, we apply and compare unconventional regression techniques. Methods: We adopt a distributional approach because we want to allow the effect of income on health to vary according to people's health status. We start by analysing the income-health relationship using a distributional regression model that falls into the GAMLSS (Generalized Additive Models for Location, Scale and Shape) framework. We assume a gamma distribution to model the health variable and specify the parameters of this distribution as linear functions of a set of explanatory variables. For comparison, we also adopt a quantile regression analysis. Based on predicted health quantiles, we use both a parametric and a non-parametric approach to estimate the lower tail of the health distribution. Results: Our data come from Wave 13 of the Household, Income and Labour Dynamics in Australia (HILDA) survey, collected in 2013-2014. According to GAMLSS, we find that the risk of ending up in poor, fair or average health is lower for those who have relatively high incomes ($80,000) than for those who have relatively low incomes ($20,000), for both smokers and non-smokers. In relative terms, the risk-lowering effect of income appears to be the largest for those who are in poor health, again for both smokers and non-smokers. The results obtained on the basis of quantile regression are to a large extent comparable to those obtained by means of GAMLSS regression. Conclusions: Both distributional regression techniques point in the direction of a non-uniform effect of income on health, and are therefore promising complements to conventional regression techniques as far as the analysis of the income-health relationship is concerned.

Income trajectories and self-rated health status in the UK

“…Besides, recent studies that extended the conventional regression approaches in evaluating the health effects of income using distributional regressions techniques found that households with poor income are particularly faced with more significant health risks. Also, they are at the lower end of the health distribution (see Kessels, Hoornweg, Thanh Bui, & Erreygers, 2020 ; Silbersdorff, Lynch, Klasen, & Kneib, 2018 ). Using the health and income distribution data in 2019, Fig.…”

Spatial patterns of avoidable fetal mortality and social deprivation

“…For the construction of the ICS, Pearson's correlation of 64 social indicators related to demography (13), education (10), sanitation (9), health (12), work (8), and vulnerability (12) with the preventable fetal mortality rate (Supplementary Material Chart 1). Indicators with p <0.05 remained in the process.…”

“…To assess the relationship between socioeconomic conditions and public health problems, statistical models such as the Generalized Linear Models (GLM) have been used to model the mean of the response variable 7 , however many phenomena require that the modeling of other distribution parameters be considered 7 . Thus, the use of the framework for the Generalized Additive Models for Location, Scale and Shape (GAMLSS) stands out for allowing to model the response variable and to specify all its parameters as linear functions of a set of explanatory variables 10,11 .…”

“…Studies on social deprivation and fetal mortality can use spatial analysis techniques for visualization, exploratory analysis, and modeling of georeferenced data [9][10][11][12] . Such analysis makes it possible to identify the factors in the territory that hinder access to basic goods and services and favor the occurrence of preventable deaths 8,13 .…”

Distributional causal effects: Beyond an “averagarian” view of intervention effects.

“…The call for evaluating effects that go beyond the mean of an outcome variable is certainly not new. Various previous studies in econometrics (Koenker & Bassett, 1978), psychology (Balota & Yap, 2011;Campitelli et al, 2017;Wiedermann & Alexandrowicz, 2007), education (Costanzo & Desimoni, 2017;Konstantopoulos et al, 2019;Petscher & Logan, 2014), and health sciences (Hart, 2001;Kessels et al, 2020;Rousselet et al, 2017) documented advantages of statistical methods that go beyond standard conditional mean-testing. In general, one can distinguish two approaches to modeling data beyond conditional means, distribution free models and complete distribution models (Kneib, 2013).…”