New methodology
This contribution derives from careful examination and analysis of the growing suite of papers that analyze niche evolution across phylogenies (e.g., Peterson et al. 1999; Wiens and Graham 2005; Knouft et al. 2006; Losos 2008; Evans et al. 2009; Vieites et al. 2009; Nyári and Reddy 2013; Meseguer et al. 2015). Of concern is that estimates of fundamental niches based on species’ realized niches are not equivalent to their true fundamental niches, owing to constraints imposed by the set of environments that can be observed within the area accessible to a species (M ). The effect of limited available environments in M lies in adding variation to niche estimates that bias analyses of niche evolution toward concluding increased niche lability. In other words, estimated fundamental niches include elements of biogeographic limitations and environments associated with the accessible areas of the different species. Indeed, use of summary statistics to characterize species’ niches introduces further variation related to the environmental vagaries of sampling, which has its own intrinsic biases (Kadmon et al. 2004) that are—again—reflected in the environmental signature of the occurrence data that derive from the process.
Previous work has illustrated the precarious nature of correlational ecological niche estimates in situations in which species’ responses to environmental variables are not fully observable across M(Elith et al. 2011; Owens et al. 2013; Araújo et al. 2013). Such situations are, at best, ones in which modeling algorithms are forced to extrapolate. At worst, these situations are ones in which significant, predictive models are impossible to obtain (Saupe et al. 2012; Qiao et al. 2015). For analyses of niche evolution in a phylogenetic context, however, environmentally-limitedM areas introduce consistent directional biases in the analyses: estimated fundamental niches will vary more than does a species’ true fundamental niche.
Analyzing ecological niche change on a phylogenetic tree without considering uncertainty produces less-ambiguous conclusions (as illustrated in Fig. 2), and certainly has the advantage of ease of implementation (e.g., simply calculating the median or mean of environmental values across all occurrences for a species). Our explorations, presented here and previously (Ribeiro et al. 2016; Saupe et al. 2017), indicate that this approach comes with a significant cost: niche change may be over- or under-estimated, introducing substantial biases in reconstructing evolutionary change in niches through time. The simple simulation results reported herein mirror the results of more sophisticated simulations that incorporate dispersal and temporal components (Saupe et al. 2017). Our simulation and empirical case further develops the body of evidence suggesting that bin-based coding accounting for uncertainty explicitly is a valuable methodological improvement. First, bin-based coding is free of the assumption that more occurrences at a particular environmental variable value indicate that that value is more suitable than another—indeed, it may just be more common within a species’ M or more likely to be sampled by researchers. Second, as we noted above in the Methods section, σ2 describes degrees of change among traits or groups (Revell et al. 2008; Cooper et al. 2011); when these changes are small, as in our simulation and for temperature in the orioles example, traditional coding methods and our new bin-based method (incorporating the known breadth of species’ niches), perform similarly. However, when variation is high, such as for precipitation in the orioles example, traditional coding methods appear to grossly inflate estimates of the rate of niche evolution across the phylogeny.