On interpretations of tests and effect sizes in regression models with a compositional predictor

Compositional data analysis is concerned with the relative importance of positive variables, expressed through their log-ratios. The literature has proposed a range of manners to compute log-ratios, some of whose interrelationships have never been reported when used as explanatory variables in regre...

Descripción completa

Detalles Bibliográficos
Autores: Coenders, Germà, Pawlowsky-Glahn, Vera
Tipo de recurso: artículo
Fecha de publicación:2020
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/362101
Acceso en línea:https://hdl.handle.net/2117/362101
https://dx.doi.org/10.2436/20.8080.02.100
Access Level:acceso abierto
Palabra clave:compositional regression models
CoDa
composition as explanatory
centred log-ratios
pivot coordinates
pairwise log-ratios
additive log-ratios
effect size
Anàlisi multivariable
Estadística matemàtica
Classificació AMS::62 Statistics::62J Linear inference, regression
Classificació AMS::62 Statistics::62H Multivariate analysis
Àrees temàtiques de la UPC::Matemàtiques i estadística::Estadística matemàtica
Descripción
Sumario:Compositional data analysis is concerned with the relative importance of positive variables, expressed through their log-ratios. The literature has proposed a range of manners to compute log-ratios, some of whose interrelationships have never been reported when used as explanatory variables in regression models. This article shows their similarities and differences in interpretation based on the notion that one log-ratio has to be interpreted keeping all others constant. The article shows that centred, additive, pivot, balance and pairwise log-ratios lead to simple reparametrizations of the same model which can be combined to provide useful tests and comparable effect size estimates.