Information criteria on multimodel selection of parametric regression: Biological applications

Abstract

When carrying out modelling analysis using experimental data is important to obtain a measure of the relative fit of the each model as a primary selection criterion. In this sense, there are few studies based on multi-model selection techniques for the theoretical representation of data sets, so it is common to incur in a misinterpretation of the existing patterns, or even more, the incorrect extrapolation and prediction based on the wrong model. This paper is intended to evaluate in 40 datasets from various ecological published studies the effectiveness of the linear regression model designated by the authors by contrasting with six regression models using the Akaike and Bayesian information criteria, and furthermore to discuss its implications on subsequent interpretations made. It was found that the linear regression model was successful only in 13.35% of the datasets (15% of datasets for AIC and 11.7% of datasets for BIC ), but in the other hand, the logarithmic model was the most successful model in 38.5% of the cases (35% of datasets for AIC and 41.1% of the datasets for BIC), casting doubts on the efficiency of the linear regression over other types of regression under biological data.