Optimizing Variable Selection in Phylogenetic Eigenvector Regression for Trait Correlation Analysis

Abstract

Phylogenetic autocorrelation among samples is a critical problem in evolutionary analysis across multiple species. To address this challenge, methods like phylogenetic eigenvector regression (PVR) have been developed. This study investigates the impact of variable selection on PVR outcomes in evolutionary biology research. Through simulations and model comparisons, we explore how different selections of dependent and independent variables influence phylogenetic eigenvector selection and correlation outcomes. Our results reveal significant discrepancies in phylogenetic eigenvector selection and correlation outcomes when swapping dependent and independent variables. R2has been identified as an effective criterion for selecting variables in PVR analysis. Additionally, we compare the performance of three non-sequential phylogenetic eigenvector selection methods. The first is to select phylogenetic eigenvectors significantly related to the response variable. The second and third methods involve forward stepwise regression, and their model fit is assessed by minimizing the Akaike Information Criterion value and by minimizing Moran's I of model residuals, respectively. We find no significant difference between the first and third methods, but both are superior over the second method. Overall, our findings highlight the critical role of variable selection, particularly based on R2, in enhancing the efficacy of PVR analysis and improving correlation analyses in evolutionary biology.