Making sense of of community phylogenetic metrics and null modelsSubmitted by editor on 20 March 2017. Get the paper!
A rainbow lorikeet (Trichoglossus haematodus) and a helmeted friarbird (Philemon buceroides) eye each other warily at a flowering Schefflera actinophylla in Port Douglas, Queensland, Australia. Photo by Bryan Suson. http://bryansuson.com/
By Eliot Miller and Christopher Trisos
Editor's Choice April 2017
Measures of phylogenetic community structure are frequently used by conservation biologists, ecologists, and evolutionary biologists to address a variety of questions. Examples include prioritizing areas for conservation, describing patterns of community assembly, and testing for phylogenetic niche conservatism.
A wide variety of metrics and null models have been introduced to address such questions. Which are best? Our paper “Phylogenetic community structure metrics and null models: a review with new methods and software” addresses that question. This is a big field, and we encountered a lot of confusion in the literature while reviewing these approaches (Box 1 in our paper details some of the history of community phylogenetic metrics). Allow us to clarify some of that confusion here.
Phylogenetic community structure methods can be broken into two parts: the metric and the null model. A given method is, most frequently, a combination of these two things. Misunderstanding this seemingly simple fact is responsible for much of the confusion in the literature. For instance, the commonly used "metric" NRI (net relatedness index) is actually a combination of a metric (MPD = mean pairwise phylogenetic distance) and a null model. Even more confusingly, the choice of null model doesn't matter in the definition of NRI. That is, MPD + any null model = NRI. Frequently, we see papers in the literature using multiple measures to address their question. For instance, authors of a scientific paper might use NRI and PSV (phylogenetic species variability) to look at their study question. However, MPD and PSV are equivalent (more details below), so if the authors standardized PSV with the same null model as they used to create their NRI measure, then they would in fact be using identical measures. If practitioners think of phylogenetic community structure methods as a two-part package, the metric + the null model, this sort of confusion can be entirely avoided.
Because a wide variety of both metrics and null models have been introduced, the number of combined possible approaches is truly daunting, and is compounded by confusion over metric and null model combinations, as well as the inadequacy of recent classification systems. We reviewed 278 metric and null model combinations in our paper. When these are appropriately separated into their constituent metrics and null models, the problem appears less intimidating: only 32 metrics and nine null models. Furthermore, we showed that many approaches are in fact equivalent or at least nearly so, making the choice of which approach to use less daunting.
For the users of these methods there's more good news: some performed much better than others, further constraining the choice of methods. Averaging across all null models, some community phylogenetic metrics exhibited false positive (type I error) rates of greater than 15%, and many failed to detect significant community assembly patterns over 50% of the time. Null model choice was also an important influence on error rates, particularly type I errors. However, combining the best metrics with the best null models resulted in notably improved performance. For instance, intraspecific abundance-weighted MPD used with the regional null model exhibited an overall error rate of just 1.8%. Furthermore, metrics of phylogenetic beta diversity showed better performance than alpha diversity metrics for detecting habitat filtering and interspecific competition.
In general, the field of phylogenetic community structure methods is one that has seen both tremendous growth and widespread confusion. As the field continues to grow, we strongly advocate that newly introduced methods be tested across a varied phylogenetic, null model, and community assembly parameter space. To that end, we wrote an R package metricTester, available from https://cran.rstudio.com/web/packages/metricTester or https://github.com/eliotmiller/metricTester, which can be used to test the behavior and performance of user-defined metrics and null models. Examples and tutorials for how to harness metricTester methods are available both in the official package documentation, and on the GitHub readme page. Related to all of this, recent work by Constantinos Tsirogiannis and Brody Sandel has elucidated exciting and quick methods for developing null-model-standardized metric expectations, and we encourage future work along these lines!