Bayesian meta-analysis of tree functional traits

How do tree species differ in their physiological and structural traits? Answering this question will improve our ability to predict forest dynamics. Towards this goal, we are compiling a database of physiological and structural traits for 305 tree species in the US. Trait data obtained from published studies include sample means, standard errors, sample sizes, and covariate data (e.g., tree age, light, soil, and climate information). However, studies vary in the types, quantity, and quality of data reported. Covariate information, sample sizes, and standard errors are often missing and only a subset of species is represented. How can we synthesize such data to learn about species-specific traits? We describe a hierarchical Bayesian meta-analysis that accommodates missing data, multiple levels of uncertainty, and taxonomic relationships that facilitate estimation of traits for all species. We apply this approach to a specific leaf area (SLA) dataset derived from 178 published studies, containing 1711 records for 152 species. The analysis models mean SLA as a function of species, genera, tree age, and light effects; it estimates the missing age, light, standard error, and sample size data; and it yields posterior distributions of SLA values for all species and genera. For example, predicted (mean ± sd, cm2/g) species-specific SLA spans an order of magnitude (32.6 ± 3.1 for Pinus Palustris to 343.7 ± 23.5 for Fagus grandifolia) and genera-specific values range from 54.0 ± 7.8 (Juniperus) to 275.9 ± 51.4 (Fagus). Across all species, tree age and light are critical determinants such that SLA is 1.47 times greater for shade vs. sun leaves and 1.64 times greater for seedlings vs. mature trees. This hierarchical Bayesian meta-analysis is broadly applicable to a variety of problems aimed at synthesizing plant functional trait data obtained from the literature or multiple studies.