There are numerous ways to characterize community diversity; however, there is little theoretical basis for selecting diversity measures. For instance, diversity measurements can incorporate species identity, species abundances, evolutionary relationships, functional groups, functional traits, or combinations thereof. Many of these variants build off each other and provide new insights into community ecology. However, this large number of metrics also presents a problem: if most of these metrics are not highly correlated, then this large number of metrics may increase the odds of finding a metric that confirms the expected pattern, potentially increasing the odds of self-confirmatory bias. Biomass and niche metrics face a similar challenge.
We addressed this problem using a grassland experiment implemented in 30 sites around the world. Because of the differences in traits and evolutionary history among functional groups, we tested whether litter accumulation at the beginning of the experiment can be predicted better by splitting the community into functional groups than by treating the community as a whole. We used as predictors biomass, richness and phylogenetic diversity by functional group, or the same metrics applied to the whole community, respectively. We replicated the analysis with soil properties after 2-4 years of nutrient application.
Results/Conclusions
Using Bayesian Information Criterion, we show that splitting the community into functional groups and using multiple biomass and biodiversity metrics simultaneously improved the predictive power of each model. Among the best models, we found that the best diversity metric was not always the same: for instance, legume’s richness and forb’s mean phylogenetic distance (MPD) were relevant to predict changes in soil sulphur, but forb’s richness and legume’s MPD were not. These differences suggest that different aspects of the community drive each of these processes. However, it is also possible that we increased the odds of self-confirmatory bias with the extra variables.
We discuss the risk of self-confirmatory bias faced in each one of these analyses and propose criteria that can be used to guide the selection of metrics in ecological studies. We discuss how experimental design, comprehensive indices (indices that incorporate several aspects of a biological phenomenon in a single metric), and statistical controls (e.g., randomization, structural equation models) can be mixed to improve the selection of adequate metrics for different studies. Finally, we discuss if using sub-sampling can help to compare the relevance of the multiple aspects of the community represented by different metrics.