Identification of phenotypic modules, semiautonomous sets of highly correlated traits, can be accomplished through exploratory (e.g., cluster analysis) or confirmatory approaches (e.g., RV coefficient analysis). Although statistically more robust, confirmatory approaches are generally unable to compare across different model structures. For example, RV coefficient analysis finds support for both two- and six-module models for the therian mammalian skull. Here, we present a maximum likelihood approach that takes into account model parameterization. We compare model log-likelihoods of trait correlation matrices using the finite-sample corrected Akaike Information Criterion, allowing for comparison of hypotheses across different model structures. Simulations varying model complexity and within- and between-module contrast demonstrate that this method correctly identifies model structure and parameters across a wide range of conditions. We further analyzed a dataset of 3-D data, consisting of 61 landmarks from 181 macaque (Macaca fuscata) skulls, distributed among five age categories, testing 31 models, including no modularity among the landmarks and various partitions of two, three, six, and eight modules. Our results clearly support a complex six-module model, with separate within- and intermodule correlations. Furthermore, this model was selected for all five age categories, demonstrating that this complex pattern of integration in the macaque skull appears early and is highly conserved throughout postnatal ontogeny. Subsampling analyses demonstrate that this method is robust to relatively low sample sizes, as is commonly encountered in rare or extinct taxa. This new approach allows for the direct comparison of models with different parameterizations, providing an important tool for the analysis of modularity across diverse systems.
Keywords: Mammals; model selection; phenotypic integration; trait correlations.
© 2016 The Author(s).