Nicholson et al. conclude that the ancestral ecomode for anoles was a crown-giant anole, and that anole evolution was characterized by a general movement from up in the trees down toward the ground (e.g., from more arboreal to more terrestrial ecomodes). Unfortunately, even accepting ecomode assignments at face value, methodological flaws render this conclusion unreliable (my previous post discusses problems with the manner in which Nicholson et al. assign species to ecomode categories; for the purposes of this post, I accept the ecomode designations they provided). Two main problems plague the analysis. First, Nicholson et al. fail to estimate uncertainty in their ancestral state reconstructions, now a standard and expected method. Had they done so, they would have found that most nodes deep in the tree cannot be reconstructed confidently as a particular ecomode. Moreover, second, independent of this problem, had ecomode state of outgroup taxa been correctly categorized, the ancestral ecomode of the anole radiation would not be unambiguously reconstructed as an arboreal species.
Problems with Ancestor Character State Estimation
The field of comparative biology has advanced greatly in the last 20 years, and it is no longer acceptable to simply reconstruct character states using parsimony. The reason is that such reconstructions provide no indication of how much confidence we may place in these reconstructions; indeed, as methods have been developed to estimate error bars around ancestral reconstructions, we have found that in many cases, the uncertainty is enormous, so great that we cannot state with any confidence that the most parsimonious reconstruction is better supported than other possible ancestral character states (see figure below for an example). The reason this occurs is that when we are dealing with traits that are very labile evolutionarily—i.e., that have evolved back-and-forth many times—there is little phylogenetic consistency in those traits, and thus the underlying assumption of ancestral reconstruction, that close relatives are likely to be similar in character state, does not hold.
An example of the uncertainty in ancestor reconstruction. The black dot represents the reconstruction of an ancestral ecomorph on Puerto Rico, inferred by parsimony. This species was inferred to be a generalist, lying between the ecomorphs in morphological space determined by principal component scores. However, when error bars are calculated for the esimtate, it can be seen that the ancestor could have been almost any of the ecomorphs. Figure from Lizards in an Evolutionary Tree, adapted from Schluter et al., (Evolution, 1997).
I discuss this issue at length in Chapter 5 of Lizards in an Evolutionary Tree, which I have excerpted here. Consider this: the most parsimonious reconstruction of ecomorph evolution in Greater Antillean anoles indicates that 19 transitions have occurred from one ecomorph to another. But, can we really strongly prefer a scenario implying 19 transitions from another scenario implying 20, especially if the 20-transition scenario yields very different reconstructions of ancestral states? Although those of a particular philosophical bent may disagree, I would argue that it’s hard to say with a confidence that reconstructions from a 19-transition scenario are much more reliable than reconstructions requiring 20 transitions.
The figure below estimates the likelihood of different ancestor character reconstructions of ecomorph of anoles—you’ll see that when all descendants of a node are the same ecomorph type, then we can have high confidence that the ancestor was that same ecomorph (the pie chart at a node is all one color); however, for most nodes, particularly further down the tree, this is not the case, and multiple ancestral character states are approximately equally likely.
Ancestor reconstruction of ecomorph state for Greater Antillean anoles from Lizards in an Evolutionary Tree. The likelihood that an ancestral node was a particular ecomorph type is represented by the proportion of the circle that is filled by that ecomorph’s color. None of the deeper nodes in the phylogeny can be confidently assigned to a single ecomorph category.
In others words, we can have little confidence in our reconstructions of the ecomorph/ecomode state of early ancestral species (Nicholson et al.’s ecomode designations are the same as previous ecomorph categorizations). Note in particular that not only is the base of the Caribbean anole radiation ambiguous, but that ambiguity results because there is some likelihood that the ancestral species could be trunk-ground, grass-bush or twig, but not trunk-crown or crown-giant. It thus seems extremely unlikely that the the ancestral ecomode node would have been reconstructed unambiguously as a crown-giant.
And, indeed, the Nicholson et al. analysis does not find unequivocal support that the ancestor of the Caribbean radiation was a crown-giant anole. Nicholson et al. state (p.54): “Our analysis indicates multiple equally parsimonious reconstructions of the ecomode of this northern ancestor. However, this uncertainty is derived from a transition from the crown giant ecomode for the ancestor of all anoles to a grass-bush common ancestor of Chamaelinorops, Audantia, Anolis, Ctenonotus, and Norops (hereafter derived anoles; Fig. 29). This transition represents a third major revision of the anole niche from one focused towards the canopy to one focused towards the ground and this transition makes the crown giant and grass-bush ecomodes equally parsimonious reconstructions of the northern ancestor as well as the ancestors of Deiroptyx and Xiphosurus. Because the majority of species of Deiroptyx (53%) and Xiphosurus (67%) included in our analysis have their habitat focused towards the canopy (crown giant, trunk crown, or trunk ecomorph), we suspect that the ancestors of both lineages, as well as the northern ancestor, were crown giants and not grass-bush anoles.”
But this argument is misguided. Ancestor reconstruction with parsimony is based on the the phylogenetic arrangement of taxa with different character states; not the number of species exhibiting a character state. That is, whether a clade with a particular character state contains two species or 20 is irrelevant. The parsimony analysis clearly indicates that one cannot distinguish on the basis of parsimony between a grass-bush and crown-giant ancestor for the Caribbean anole clade—the reasoning employed by Nicholson et al. violates a basic tenet of parsimony analysis. The conclusion of this analysis is straightforward: if the ancestor of this clade cannot be assigned unambiguously, then the argument that the Caribbean radiation progressed from in the crown to more terrestrial cannot be supported. The Nicholson et al. analysis does not support that conclusion, and had uncertainty in character reconstructions been calculated, the lack of clarity would have been even more apparent.
Norops ecomode evolution from Nicholson et al. 2012. Different colors represent different ecomodes. Species in black are those for which no data are available; rainbow multi-colors are for “polymodal” species.
Is the story for mainland anole ecomodes likely to be different? No. Check out the willy-nilly assortment of ecomode types across the phylogeny of mainland Norops. If anything, ecomode evolution seems more evolutionarily labile in this clade than in the West Indian species. My eyeball estimate is that minimally somewhere around 20 evolutionary transitions are required in this part of the tree. If the degree of uncertainty had been calculated for ancestral character states in this part of the tree, I have no doubt that the reconstructions for most of the deeper Norops nodes would have been as ambiguous as they are for the Caribbean ecomorphs.
Ecomode evolution in Dactyloa from Nicholson et al.
What about for the Dactyloa clade? There things are slightly better, in that only eight character state transitions are required if one accepts that crown-giant is the ancestral ecomode state. Because these transitions are to five different states (including “polymorphic” as a character state, which probably isn’t appropriate procedurally because polymorphic is both a combination of multiple other states and a different combination of ecomodes in different species), the ancestral state of most nodes (other than for the phenacosaurs, which are the green and red species) may be fairly strongly supported as crown giant. Moreover, if one were to forget about body size and combine crown-giant (dashed blue lines) and trunk-crown (solid blue) into one “crown” ecomode (see next section), then this ecomode would probably be very strongly supported as the ancestral state throughout the clade.
Some might argue that phylogenetic inference of ancestral states is always about putting forth the best estimate as a hypothesis for further testing, and in this light, one might consider the most parsimonious reconstruction of ancestral states to be that best hypothesis. That may well be so, but what this sort of analysis indicates is that even the best hypothesis may not be very strongly supported, and thus not suitable for drawing strong conclusions about evolutionary patterns.
Problems with the Outgroup Character State
It is odd that Nicholson et al. distinguish between trunk-crown and crown-giant anoles given that ecomode categorization is explicitly based on habitat use and not on morphology (Nicholson et al.’s justification for this is a tad ironic given their taxonomic proposal to split Anolis: “We retain a category for canopy giant despite the obvious drawback that this category retains a feature of morphology by referring to body size. We do this because of the wide use of this category in past literature.” p.50).
This procedure seems mostly harmless in that it appears to have only been used to distinguish trunk-crown vs. crown-giant anoles in most cases; i.e., a species wasn’t classified as a crown-giant due to its body size if its habitat use wasn’t in the crown. With one big exception. The analysis includes two outgroups that form a polytomy with anoles, and one of them is Basiliscus, representing the Corytophanidae (see Dactyloa phylogeny figure above). Everyone knows that basilisks run across streams, not canopies, and yet it is coded as a crown-giant, presumably because of its size (from what is known, neither of the two other corytophanids–Corytophanes or Laemanctus–is a crown animal, either, at least from what is known of their natural history).
Would changing the ecomode character state of the Basiliscus clade to something more terrestrial—trunk-ground is perhaps the closest anole analogue—have made a difference? Absolutely. Given the huge heterogeneity in character states in the sister group to Dactyloa (incidentally, we need a name for this clade; perhaps Caribanolis recognizing that six of the seven major clades are exclusively Caribbean in geography?), crown-giant would no longer be the clear reconstruction for the ecomode of the ancestral anole. One possibility would still be that crown-giant was the ancestral character state, as exhibited by Polychrus and retained in anoles, with a more terrestrial state in both Basiliscus and Caribanolis being derived. But a plausible alternative would be that being terrestrial is the ancestral state, as exhibited by Basiliscus and, potentially, by Caribanolis, and that the crown-giant character state was independently derived in Polychrus and the ancestor of Dactyloa. More generally, re-casting corytophanids as some other ecomode would most likely have led to an ambiguous reconstruction for the ancestral node if uncertainty were calculated.
Bottom line: If appropriate outgroup characterizations and methods for estimating uncertainty in ancestor character states were used, the results of this analysis would be highly uncertain. of the major ancestral nodes in the phylogeny, probably only the ancestor of Dactyloa would be reconstructed with high certainty. Because ecomode has evolved back-and-forth so many times, phylogenetic approaches simply cannot provide a clear view of the history habitat use evolution. The conclusion that anole history documents a directional trend from a highly arboreal ancestor to more terrestrial descendants is premature.
- Diet Notes on Beautiful Blue Knight Anole - September 4, 2024
- Anoles Provide Ecosystem Services - September 2, 2024
- Mississippi Kite Eats Green Anole - August 6, 2024
Daniel Scantlebury
Great post, Jonathan. I hope people can appreciate that grievances with this paper are not simply a matter of taxonomic convenience, there’s also a many methodological concerns.
Rich Glor
Is there any justification for avoiding use of the maximum likelihood or Bayesian methods for character reconstruction that have been favored by most comparative biologists for the last decade? As Jonathan mentions, these methods have provided for major advances in our ability to understand character evolution, and uncertainty in ancestral reconstructions in particular. I know some people are philosophically opposed to the use of ML or Bayesian methods, but it seems odd that the Nicholson et al. study uses these methods for phylogeny reconstruction and time calibration but then abandons them for reconstruction of ecomodes and biogeography.
Liam Revell
Rich, Dan, & Jonathan.
The ancestral state estimation performed in this paper is problematic, but not only for the reasons highlighted by Jonathan.
Indeed parsimony ancestral state estimation is in many ways “non-statistical” in that there is no straightforward procedure to weigh the evidence supporting alternative character histories on the tree, and we cannot compute the probability of our favored history conditioned on an explicit model. If we used an explicit model, we would find (as Jonathan points out) that the states at internal nodes in the tree (particularly those deep in the tree) con only be estimated with considerable uncertainty, conditioned on our model.
A second problem associated with ancestral character estimation (particularly in this case) is the inherent silliness of our model. Statistical methods for ancestral character estimation explicitly assume a discrete Markov process of trait evolution on the tree. (It has been effectively proven elsewhere that parsimony implicitly assumes this model as well, just with a very low transition rate between states.) A discrete state Markov process (chain) is “memoryless” – meaning that the probability of changing state depends only on the current state, and not at all on prior states of the chain nor on the time spent since the last change of state. This might be quite a reasonable** model for nucleotide substitution, because it effectively mimics the mutational process in which we think that the probability of changing from an A->T or vice versa does not depend on how long the sequence has been in state A or T. (**But is unrealistic in that it ignores the fact that populations can be polymorphic for a nucleotide state. If polymorphism is extremely transient when compared to the evolutionary time spanned by our tree, this is probably not that important.) However, this model seems quite unsuitable to more complex characters that most likely have polygenic bases (such as habitat use), even if they can be measured and categorized discretely. The “nucleotide model” implies, in this case, that no matter how long a lineage has specialized on an arboreal niche (for instance), it has an equal probability of shifting back to a terrestrial niche (or vice versa).
A more reasonable model, in this case, might be something like the threshold model of quantitative genetics (e.g., Felsenstein 2012). Under the threshold model, discrete traits are underlain by continuous liability. When the liability exceeds a threshold, the discrete character changes state. This type of model (although still a simplification of reality) seems much better suited to reconstruct ancestral states of discrete characters with complex, polygenic basis than a nucleotide model. It has some properties that make it seem much more realistic – for instance, once a lineage changes states it is much more likely to change back immediately than after considerable time has passed. Unfortunately, ancestral character estimation under the threshold model has not yet been developed (but is relatively straightforward, in my opinion, so perhaps look to phytools soon).
That’s my two cents!
– Liam
P. S. the plot below is a visualization of evolution under the threshold model on the tree. The vertical axis is time and the horizontal axis is liability. Color is used to represent changes in the discrete character.
Julián Velasco
Great post.
Some assignments of ecomorphs for mainland species are very dubious. For example, Nicholson et al. assigned Anolis princeps as a crown-giant ecomorph, but wI would think in a trunk-crown ecomorph for this species and related species (e.g., Anolis latifrons, Anolis frenatus) and the same with Anolis tropidogaster which was assigned to the “ground” and I would think that this species is more related to a grass-bush ecomorph. Also, the sample size used for this ancestral reconstruction is very low, principally for species from Dactyloa clade where there is a large ecological variation in habitat use (e.g., semiaquatic, grass-bush, trunk, etc.).
I think that to make a clear and fair comparison between island and mainland is necessary compare both radiations with the same set of ecological, morphological and behavioral data, on the contrary the comparisons would be flawed as in this particular case. Definitively, we need a lot of fieldwork for mainland anoles.