A causal partition of trait correlations: using graphical models to derive statistical models from theoretical language
Recent studies hypothesize various causes of species‐level trait covariation, namely size (e.g., metabolic theory of ecology and leaf economics spectrum), pace‐of‐life (e.g., slow‐to‐fast continuum; lifestyle continuum), evolutionary history (e.g., phylogenetic conservatism), and ecological conditions (e.g., stabilizing selection). Various methods have been used in attempts to partition trait correlation among these influences (e.g., univariate analysis, principal components analysis, and factor analysis). However, it is not clear that the implied causal structure assumed by these methods matches the hypothesized causal structure driving trait correlations, a situation that can potentially lead to biased estimates and incorrect partitioning among mechanisms. Here, we propose the application of graphical causal models (GCM) for across‐kingdom synthesis and to aid researchers in their selection of correct analytical strategies. Graphical causal models use causal diagrams (i.e., box‐and‐arrow graphs) to represent expert knowledge of the data‐generating processes to analytically investigate the possibility of identifying hypothesized causal associations. We developed a causal diagram that synthesizes prominent hypotheses of trait covariation. Using the causal diagram, we (1) derived a quantitative expression to partition trait covariance among its hypothesized causal elements (i.e., size, pace‐of‐life, evolutionary history, and ecological conditions) and (2) developed analytic strategies to attribute trait covariance among the hypothesized causal elements under real‐world data availability, namely unobserved variables (i.e., pace‐of‐life) and confounding variables (i.e., evolutionary history and ecological conditions). Finally, we tested each analytic strategy by simulating trait datasets and, after incorporating the data limitations, tested their ability to correctly partition trait covariance. The analytical strategies were able to correctly partition trait covariance into the hypothesized causal elements of size, pace‐of‐life, and the historical effects of evolutionary history and ecological conditions. We demonstrate the efficacy of these strategies by applying them to a widely used trait dataset. Overall, the application of GCM revealed that researchers have used inappropriate measures to represent their theoretical constructs and have relied on analytical strategies that violated their causal assumptions, likely resulting in biased estimates. We discuss how this mismatch between theoretical language and statistical methods is prevalent in species‐level, trait‐based research and call for future studies to address these limitations.