Inferencia multimodelo en ciencias sociales y ambientales

Abstract

Professionals of the social and environmental sciences must solve problems (answer questions) based on data sampling and analyses. Commonly, all professionals face similar challenges: they need to take decisions on a population (e.g., all the trees of a region), but only have data from a sample (some trees of that region). A key tool in this process is to propose population models for the response variable (tree growth as a function of tree age and climatic conditions) and then use model predictions to take decisions (e.g., when to cut trees according to climatic conditions). In this paper we discuss how to propose, estimate, and select models of a population based on sampling data. We put special emphasis in proposing several alternative models (hypotheses) to solve one problem (e.g., different tree growth functions for age), which must be proposed before data sampling, including a null model (tree growth does not depend on tree age or climatic conditions). Models guide us on how data must be sampled for a valid contrast (growth measurements in trees of different age and under contrasting climates). Then, the Akaike information criterion (AIC) can be employed to sort the most parsimonious models, selecting those with the best goodness of fit (likelihood) and the lowest number of parameters (model complexity). Along the text, we introduce basic notions of multimodel inference and discuss common user mistakes. We provide real examples, and share their data and the analyses code in R, a free and open source software. In addition to be useful to professionals from different sciences, we expect our paper to promote the teaching of multimodel inference in graduate courses.