An Analytical Approach Morphology, Shape and Phylogeny by Peter J. Wagn

Palaeontologia Electronica http://palaeo-electronica.org

Fossils, Phylogeny, and Form: An Analytical Approach by Jonathan M. Adrain, Gregory D. Edgecombe, and Bruce S. Lieberman (editors) Topics in Geobiology vol. 19. Kluwer Academic / Plenum Publishers, New York, 2001, 402 p. ISBN: 0-306-46721-6. $130.00.

Morphology, Shape and Phylogeny by Norman MacLeod and Peter Forey (editors) Systematics Association Special Volume no. 64. Taylor and Francis, London and New York, 2002, 308 p. ISBN: 0-415-24074-3. £70.00/$125.00

by Peter J. Wagner

In the eleven conference of the years since Harvey Systematics Associa- and Pagel (1991), tion with the same evolutionary biolo- name as the book. gists committed con- This excellent volume siderable includes papers deal- methodological effort ing with both aspects into the relationship of morphometric data between phylogeny and phylogeny: how and numerous types one can infer phylog- of patterns. Morphol- enies or even test ogy, Shape and Phylogeny and Fossils, phylogenetic hypotheses from morpho- Phylogeny, and Form both concern metric data, and how one can reconstruct themselves with these issues. The former continuous character evolution over model book deals primarily with the relationship phylogenies. Of course, these concepts between phylogeny and continuous mor- are inexorably linked and several of the phometric data, whereas the latter deals papers treat both issues. The authors with a variety of other issues also. include many of the movers and shakers MacLeod and Forey's text stems from of morphometric, phylogenetic and tree- the 1999 Second Biennial International based methodology. Whether continuous

Wagner, Peter J., 2003. FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH and MORPHOLOGY, SHAPE AND PHYLOGENY. Palaeontologia Electronica, vol. 5, issue 2, book review 3: 15pp., 170KB; http://palaeo-electronica.org. WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

data contribute to phylogenetic studies Felsenstein also notes that the mathemati- (one of contention not long ago) receives cal intractability of juggling so many little attention - the authors unanimously parameters means that we will have to use agree that continuous characters can do Markov Chain Monte Carlo (MCMC) meth- so. The closest thing to a dissenting voice ods (e.g., Larget and Simon 1999) to find is Humphries' short paper, and even he and "integrate" over likely parameter val- backs off from his previous stand (e.g., ues. (Notably, Felsenstein does not com- Cranston and Humphries 1988) and con- ment on using MCMC to estimate cludes that any type of data is suitable for Bayesian probabilities of parameters.) phylogenetic study. Humphries does Other papers focus on recognizing lament that ideas about transformation states within continuous characters. Rae's and character evolution have become chapter is concerned (in large part) with muddle with the concepts of sorting the correlation between size and shape homologies and classification. However, and how that affects residuals and cen- one thing made clear by many of the troids. However, as I note above, corre- papers in the volumes is that the former lated change is a potential problem for all concept is inseparable from the two ideas. character data (including molecular data). (The relationship of classification to any of Rae advocates the use of ratios, but I am this is arbitrary philosophy, of course.) skeptical of such approaches simply Two distinct concepts accompany because there are numerous transforma- reconstructing phylogeny from morpho- tions than can result in similar ratios metric data - estimating phylogeny directly among features, and because many fea- from morphometric data, and inferring tures that one might describe with ratios characters and character states from mor- will themselves not be independently phometric data. Given the choices of qual- evolving ones. Reid and Sidwell summa- ifying continuous data or analyzing it rize methods commonly used to divide directly, Felsenstein opts for the latter. As continua into discrete characters and note he warns, it will not be easy. Of particular that all suffer from problems. As Felsen- importance here are the effects of corre- stein notes, variation in peak conditions lated change. This is critical when examin- among closely related species would blur ing multiple landmarks associated with the boundaries even if state transformations same homology, but really represents an were fairly discrete, so perhaps these under-addressed problem that is not problems should not be surprising. unique to morphometric data (e.g., Perhaps the most interesting papers in McCracken et al. 1999.) He also briefly this section are the ones dealing with geo- outlines the role of fossils in such analy- metric morphometric methods for inferring ses, noting two ways in which they can be characters and states. Chapters by Swid- handled: using model molecular trees to erski, Zelditch and Fink and by MacLeod estimate covariance and rate patterns offer interesting points and counterpoints among landmarks, and then assigning fos- concerning landmark versus outline analy- sils to their most likely positions, or trying ses. Swiderski et al. continue earlier to simultaneously estimate (or test) mor- papers (e.g., Zelditch et al. 1995; Swider- phological parameters and phylogenetics ski et al. 1998) exploring how one might topology. (Although Felsenstein does not use Bookstein's thin-plate splines methods mention it, this is the only option when to infer characters. In particular, they dealing with completely extinct taxa.) examine partial warps, which are analo-

2 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

gous to Factor Analysis scores, with the and larger samples might have relative eigenvectors (principal warps) derived warps confounded by phylogenetics auto- from a description of non-uniform differ- correlation. Although MacLeod dismisses ences between observed forms and some Swiderski et al.'s criticism of using multi- reference form. The authors repeat some variate summaries of distributions to infer of their concerns about using "traditional" characters and states, eigenvectors might methods (e.g., principal components anal- blend multiple warps diagnosing particular ysis) to identify multivariate characters on clades, and distort warps that change the grounds that phylogenetic autocorrela- occur with numerous combinations of tion has strong effects on the relative other warps. Similarly, eigenshape analy- importance of eigenvectors. Also ses (which use eigenvectors derived from (although the authors do not state this angular differences between specimen explicitly), phylogenetic autocorrelation will outlines and some standard outline) run a cause eigenvectors to summarize suites of similar risk of blurring/distorting indepen- states separating clades and paraclades, dent derivations because phylogeny can- and thus can miss biologically indepen- not be removed. Still, eigenshape dent features. More notably, the authors analyses are able to replicate (and back away from some of their previous improve upon) nominal characters and positions. For example, they do not rec- states for trilobite pygidia. Given the pau- ommend using partial warps as a means city of character states for trilobites (and, of exploring for characters rather than to a lesser extent, other commonly fossil- diagnosing them. They also acknowledge ized taxa), this is at least a start. For their that outline methods are needed to part, Swiderski et al. offer measures such describe differences in shapes between as smoothness, sinuosity, etc. as possible homologous landmarks. MacLeod illus- measures of outline shapes rather than trates some of the weaknesses of partial multivariate reductions such as eigen- warps using empirical and simulated shape axes. However, this will reintroduce examples. Unfortunately, the empirical tri- the problem of categorizing continua into lobite example is not a good test case - tri- discrete characters if one is to use such lobite matrices (including the one used by features as character states. MacLeod) have notoriously poor resolu- Bookstein offers a different take on tion of states (Wagner 2000b) and the dis- identifying character states, summarizing agreement between partial warps and his creases technique. A "crease" is the nominal characters might be telling us apex of a function in morphospace more about the inadequacy of the nominal describing an axis of deformation between characters than of partial warps. However, two types and thus the transition boundary the inability of partial warps to replicate between two states. Visually, this is the known transformation patterns in simula- point where a D'Arcy Thomson-esque grid tions is damning. MacLeod advocates shows directly abutting curves at some using relative warps, i.e., the eigenvector extrapolation angle. Although the concept summaries of multivariate distributions is a little hairy, Bookstein offers a simple among partial warps. This does separate exercise involving basic geometric fea- out state types among simulated fish tures that illustrate it nicely. Bookstein pre- nicely, especially when the analyses are sents three empirical examples, two confined to particular organ systems. involving intraspecific variation (schizo- However, this is only a 10-taxon example phrenic vs. non-schizophrenic and male

3 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

vs. female brain dimensions) and one netic structure present. Of course, this involving interspecific differences (homi- really should be a concern for any charac- noid species). Although Procrustes, thin- ter set. Cole, Lele and Richtsmeir discuss plate splines and other methods do not parametric bootstrapping techniques for clearly separate intraspecific differences, evaluating how well some observed mea- prominent creases exist separating males sure of structure (in this case, phenogram from females, and schizophrenics from metrics) match real data. Parametric boot- non-schizophrenics. Similarly, two creases strapping is a cross between standard are found for hominoids (chimps vs. homi- bootstrapping and typical simulations. nids, and archaic vs. modern Homo sapi- Unlike standard bootstrapping, which res- ens) that are more prominent than amples from an "observed" distribution (in differences found by other methods. this case, the data are branch lengths Creases obviously are very new and it is derived from a model tree and thus mod- not clear yet how important the extrapola- els rather than data), a distribution is fit to tion angles will prove to be when interpret- the data and then used in the simulations. ing these functions. Still, they do offer Using an empirically derived tree as a some hope that discrete sets do exist in model, best phenograms from simulated morphometric data. data are compared to the original tree to Rohlf's chapter is in some ways a examine how often that shape is repli- compliment to Felsenstein's, with the cated. With structured data, this will be emphasis on morphometrics and phylog- common. Unfortunately, the authors eny somewhat reversed. Rohlf notes fitting present only one example using only four shape descriptors to phylogeny requires taxa, so the utility is difficult to assess. that factors such as specimen orientation Also, the approach would be most inter- be factored out, something that geometric esting if one is comparing two or more methods using reference shapes do. How- morphometric data sets, or contrasting ever, the method used to infer ancestral continuous data to discrete or molecular conditions also must be independent of data. For example, one could measure reference shape, which is true of likelihood matrix structure using something like aver- and squared change parsimony, but not age squared Procrustes distances and true of traditional linear parsimony. The then simulate continuous evolution at dif- chapter initially focuses on reconstructing ferent rates (I have used similar evolution on model trees, but it briefly approaches to evaluate the likelihood of addresses inferring phylogeny from mor- amounts of change given observed con- phometric data. Rohlf briefly suggests like- gruence and compatibility; Wagner 1998, lihood, squared-change parsimony or 2001). This then could be used to test con- even neighbor-joining. However, he cau- trast rates of change among morphometric tions against attempting this unless many and discrete data sets. Also, although the landmarks and multiple systems are being authors present this as a model-tree test, used. This latter introduces new problems, they do note that MCMC methods could however, as it will be important to not allow be used to explore structure over a range structures with numerous landmarks to of trees (which in turn are weighted by the swamp the signal of structures with few tree likelihoods) and thus make the test of landmarks. structure among continuous characters One concern about morphometric assume only that there was a phylogeny, data is that there is little if any phyloge- not a particular phylogeny.

4 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

The remaining papers focus on using closest to matching observed ancestral model trees to test hypotheses about con- sizes (based on the sum of squared differ- tinuous character evolution. These will be ences between observed and expected). of especial interest to paleontologists as However, the authors note that there often the empirical examples include fossils and were multiple equally parsimonious recon- fossil information about temporal branch structions, which means that the method lengths figures prominently. The latter had multiple chances to get a correct issue figures prominently in Polly's paper, answer. I did find it surprising that the two- which examines rates of rates of mamma- parameter likelihood model made the lian tooth evolution. Two issues are impor- same reconstructions that the one-param- tant here. One is simply the length of the eter likelihood model did for the conodont branches. The second is the phylogeny data. The two-parameter model includes used in the analysis, as there are multiple both a rate parameter (β) and a constraint phylogenies consistent with any given cla- parameter (α), with the latter affecting dogram ("Eldredge's Enigma") and these rates of within-lineage change relative to different trees will imply different rates. between lineage change β. However, I The importance of not just a correct cla- suspect that this was due to the method dogram, but of a correct phylogeny not taking into account the stratigraphic (including ancestor-descendant models) ranges of the conodont species. Having a therefore is critical. Polly's analysis sug- range of nearly static values would require gests extremely high rates of change for a high a and thus necessarily change . mammal teeth, almost rivaling that of lit- Also, the authors only briefly note that the torinind snail shells. The above being said, β parameters almost certainly vary across MCMC Bayesian methods might offer a both phylogenies. If so, then two, three, way around this, as one could sum the etc., b parameters increase the probability probabilities of observed morphometric of the observed data significantly. In addi- distributions over multiple phylogenies tion to allowing tests for a variety of mac- (including those from different cla- roevolutionary hypotheses, this should dograms) and thus estimate the condi- also improve the ancestral reconstruc- tional likelihoods of rates assuming only tions. that there was an underlying phylogeny Pagel's chapter actually addresses rather than assuming a particular phylog- this and other issues when examining eny. hominid brain size evolution. In addition to Webster and Purvis examine the abil- rate and constraint parameters, Pagel also ity of several methods to reconstruct tests for biased versus unbiased change ancestral conditions. The chapter presents in brain size and shifting rates of brain size a very lucid review of what the different evolution. (Unfortunately, Pagel uses dif- methods (e.g., linear parsimony, squared- ferent Greek letters for the parameters change parsimony, different likelihood Webster and Purvis describe, and α and β methods) do and how they are related. for different parameters, which makes the Instead of simulations, the authors used first read of the paper difficult!) Pagel finds two model trees (a conodont family and that a positive bias parameter was signifi- primates) including inferred ancestors, cantly more likely than an unbiased and asked how well inferred ancestral size parameter, and also that an increasing matches "observed" ancestral size. In both rate parameter is significantly more likely cases, linear parsimony actually comes than an unchanging rate parameter. That

5 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

is, brain size tended to increase over time, ters in no way detracts from what is truly and the amount of increase itself excellent volume, filled with very good to increased over time. Pagel does make excellent (and frequently cutting edge) one conclusion against which I would (in papers. general) caution. The most likely branch Fossils, Phylogeny, and Form is in scaling parameter (κ, essentially equiva- some ways a more specialized volume lent to Webster and Purvis' a parameter) is than is Morphology, Shape and Phylog- close to 1.0, which means that the amount eny in that most of the papers deal with tri- of change is proportional to the branch lobites and in that most of the authors are length. Purvis asserts that if evolution is trilobite specialists. Fossils, Phylogeny, punctuated, then κ should be close to and Form also is a more conservative vol- zero. However, Felsenstein notes in his ume than is Morphology, Shape and chapter that unsampled extinct ancestors Phylogeny, being content to apply tradi- invalidate this assumption. This might not tional analyses rather than to push enve- be a major problem on hominid phylogeny, lopes. Wills' "primer" stands out as one of which has been intensely sampled for self- the excellent papers, one that would have ish reasons. However, in examples such made an excellent Annual Reviews as Webster and Purvis' primate tree, the paper. Although very long, Wills deftly number of unsampled extinct ancestral reviews examples of empirical mor- species along branches will be propor- phospaces for both discrete and morpho- tional to the temporal length of the branch. metric data sets. He then provides a (This will be especially true on longer tem- similar review of theoretical mor- poral branches.) Simulations by Marcot phospaces. In doing so, Wills provides (2000) show that this results in punctua- nice descriptions of methods ranging from tion and gradualism having nearly identical simple phenetic dissimilarity to thin-plate expectations. splines. Wills does an especially good job A question that could have made a of describing (and illustrating) what differ- fascinating paper in its own right stands ent disparity metrics such as variances the original question on its head - do and nearest neighbor distances tell us shape characters have discrete states? about morphospace, and how contrasting Felsenstein suggests "no", noting that metrics often is more informative than sin- even seemingly discrete features such as gle metrics. Similarly, Wills illustrates numbers of digits might represent simply nicely how different morphospace patterns thresholds along continua. He suggests create different rarefaction patterns. that morphometrics might allow us to Finally, Wills discusses the use of mor- "uncode" seemingly discrete states into phospace studies in morphological con- their constituent continuous factors. On straint analyses and other the other hand, Bookstein suggests "yes", macroevolutionary studies, and clearly that there are discrete boundaries in mor- delimits between situations where morphospace (i.e., his creases). Also, the phospaces can test hypotheses, and book assumes a certain level of knowl- where morphospace patterns only allow edge about morphometric techniques. inferences. However, the book clearly is aimed at fairly Hughes and Chapman's case study advanced researchers planning to exam- with Silurian trilobites also stands out. In ine morphometric data in phylogenetic particular, they wish to address the rela- contexts. Thus, the absence of such chap- tionship between developmental styles

6 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

and phylogeny by combining Procrustes authors "continue" their debates about the analyses and simply phenetic analyses of relative merits of different methods, both trilobite populations. Hughes and Chap- chapters are more general than the man report several interesting findings authors' contributions in the MacLeod and showing that species are fairly discrete Forey volume. Thus, they should provide and that nuisance parameters such as easier (and initially more informative) for sampling and ontogeny contribute little to readers unfamiliar with morphometrics. perceived variation. One especially However, readers already familiar with remarkable finding is that morphological morphometrics probably will get more from variation does vary in different organ sys- the authors' chapters in the MacLeod and tems, at least where one species is con- Forey volume. cerned. That species, Aulacopleura Eldredge addresses how the diagnos- konincki, shows extreme variation in tho- tic characters of higher taxa might evolve. racic segment numbers. Although this is He focuses on the expectations of the partly correlated, the other species are "Sloshing Bucket" hypothesis, named after notable for there complete lack of varia- organisms' existences in two parallel bio- tion. Notably, similar types of variation are logical hierarchies (genealogical and eco- seen in some Cambrian species. How- logical). Eldredge summarizes the ever, estimated phylogenies imply that expectations of different levels on ecologi- Aulacopleura is much more closely cal processes on the derivation and fixa- related to static Silurian species than to tion of major morphological innovations. the variable Cambrian ones. Also, thoracic He then reviews how well trilobites fit variation in Aulacopleura is not accompa- those expectations. Eldredge expects that nied by variation elsewhere - in fact, ecological processes ranging from busi- cranidial features tend to be on the static ness-as-usual to protracted long-term side (likely due to enrolling). Hughes and change all will do little to induce either Chapman note that increasing the thoracic speciation or major morphological change. segments increases the number of gills Eldredge expects only the highest two lev- (and thus gill surface area) and thus might els, i.e., ecologic change being too rapid have served as an adaptation for life in low for habitat-tracking to maintain species O2 conditions. Hughes and Chapman note and (especially) full-blown mass extinc- that this indicates that intrinsic constraints tions (and their subsequent radiations) to were not inviolable (although they are induce macroevolutionary patterns. careful to note that this is very different Eldredge further proposes than interplay from saying that intrinsic constraints did minor turnovers (e.g., substage- or stage- not exist). An interesting follow-up to this level) during radiations create a relay pat- study would be to estimate how frequently tern that amplifies the affects of those radi- such relaxations occurred in response to ations. Trilobites appear to meet the extreme ecological environments, and expectations in some cases. Hamiltonian whether this might have showed any tem- species ranges and Cambrian biomeres poral trend. both are consistent with pulsed turnovers The sole redundancy between Fos- and the morphologic diversification of sils, Phylogeny, and Form and Morphol- Ordovician trilobites is consistent with the ogy, Shape and Phylogeny exists in alleged role of radiations. However, other chapters by Zelditch, Swiderski and Fink, putative corroborating examples are prob- and MacLeod. However, although the lematic. The last radiation of trilobites, the

7 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

calymeniids, show the relay pattern of struct phylogeny (see Wagner 2000a). morphologic innovation Eldredge pre- Thus, it is difficult to ascertain whether the dicts. However, Eldredge acknowledges lack of biogeographic signal reflects flaws that sampling of the clade is not good and in the component-based methods or sim- the relay pattern might reflect periodic ply inaccurate assumptions about trilobite sampling of a clade continuously accumu- history. Moreover, at least one of the meth- lating synapomorphies. Eldredge suggests ods (Tree Mapping) allows only for vicari- that apparently high early rates of specia- ance, even though we know that dispersal tion among olenelloid trilobites docu- is very important in evolution, and it is mented by Lieberman (2001) support the questionable how well the other methods role of radiation in clade innovation, but can accommodate dispersal. However, that analysis probably is hindered by a methods that explicitly look for this pattern highly inaccurate model phylogeny (Web- (e.g., Alroy 1995) are not even discussed. ster 2002 - also, see below). Also, the What might have been more useful is relay pattern that olenelloids show could a simulation study that could examine the easily be an artifact of temporally varying robustness of component-based methods preservation. Thus, it is not clear that trilo- to geographic distribution assumptions bites corroborate the Sloshing Bucket and model tree error. However, hypothesis particularly well. approaches such as these are rapidly Ebach and Edgecombe provide an becoming outdated. "Bayesian" MCMC empirical review of paleobiogeographic techniques provide tree-based tests with- reconstruction. They focus on component- out assuming a model tree (Huelsenbeck based methods including Tree Mapping, et al. 2000b) and such tests have been Three Area Statements and Paralogy-free applied to the nearly identical problem of Subtree analysis. Again, their examples cospeciation (Huelsenbeck et al. 2000a). focus on trilobites, using numerous model With the development of likelihood tests phylogenies for Silurian and Devonian tri- given morphological data (Wagner 2000c; lobites. Because disagreement exists con- Lewis 2001), one could sum the likeli- cerning the information that widespread hoods of trilobite trees from different species present for such analyses, the clades that are consistent with different authors repeat each analysis using differ- biogeographic histories (including both ent assumptions about those species. The vicariance and dispersal). The summed results are quite discouraging, as the likelihoods (which approximate Bayesian methods produce very different results not posterior probabilities if we assume the only from one another but depending on prior probabilities to be identical) then can the assumptions about widespread spe- be used to evaluate alternative biogeo- cies. Unfortunately, these methods use graphic hypotheses directly from character phylogenetic models in lieu of data, and data rather than from intermediate phylo- the extent to which model error might genetic models. Again, it bears emphasiz- affect results is not discussed. This is ing that such analyses test hypotheses especially critical when dealing with trilo- about geographic dispersal assuming only bites, for which minimum rates of that there were phylogenies, not particular homoplasy are extremely high (Wagner phylogenies. 2000b) and often well beyond the rate at Lieberman applies tree-based tests of which parsimony methods (whence the speciation rates to Cambrian olenelloid tri- models were derived) accurately recon- lobites. He emphasizes three approaches

8 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

for estimating speciation rates, Yule's pure expected relationship between speciation birth (PB) model, Feller's birth-death (BD) and extinction rates (e.g., Walker and Val- model, and the Bienyamé-Galton-Watson entine 1984). However, this reveals a fun- discrete time birth-death model. Although damental shortcoming of estimating not described as such, the first model is a extinction rates using the methods Lieber- special case of the second model (i.e., man reviews: pure birth hypotheses pure birth is simply birth-death with an always will yield higher likelihoods than extinction rate of zero) and the second will birth-death models when making X to model is a special case of the third model Y over Z Ma comparisons whenever Y (where there are "turnover" and "back- >zero. That is because P[Y=zero | PB] = ground" speciation rates happen to be 0.0 whereas P[Y=zero | BD] > 0.0, which equal. Beginning with the pure birth means that the maximum P[Y=1 | BD] is model, Lieberman finds that a 90th per- always less than the maximum P[Y=1 | centile speciation rate for the entire Phan- PB]. Thus, the maximum L[PB | Y=1] will erozoic is not particularly unlikely for always be greater than the maximum olenelloids, given a model phylogeny that L[Y=1 | PB]. (Modifications to these suggests that olenelloids went from 2 to 20 approaches by Nee et al. [1994] do pro- species in approximately six million years. duce ML extinction rates >0, but these Lieberman then adds a "moderate" Phan- rates still are unrealistically low.) erozoic extinction rate to that speciation One could compensate for this by esti- rate and finds that 2 to 20 in 6 Ma still is mating the likelihood of extinction rates not improbable. From this, he concludes given the stratigraphic ranges of species that speciation rates might not have been that go extinct over an interval (Foote unusually high among Cambrian trilobites. 1997) and incorporating that into the equa- Unfortunately, the approach taken is fun- tions. However, this still leaves an impor- damentally flawed. What should instead tant biasing model: the phylogeny that is have been done is to take post-Cambrian the basis for the 2 to 20 diversification, as trilobite clades as well as the Cambrian most of the "observations" are range clade and determined the likelihood of the extensions (Smith 1988) implicit to the phylogenies given the same speciation tree. Even among trilobites, olenelloids and extinction parameters. The null stand out as homoplasy-ridden and lack- hypothesis (courtesy of Ockham's Razor) ing hierarchical structure in character state is that they are equal, as this represents a distributions (Wagner 2000b). Simulated special case of the test hypotheses (i.e., matrices with similar rates of homoplasy separate speciation/extinction rates for inevitably produce highly erroneous trees Ordovician and Cambrian clades that hap- and erroneous trees are strongly biased pen to be equal). Unless the topologies toward exaggerating range extensions are identical (including the temporal rather than underestimating them. Notably, lengths of branches), likelihoods will Webster's reanalysis of olenelloid phylog- always be maximized using different rates. eny, cited above (using morphometrics to Instead of using a model extinction rate for weed out suites of correlated characters the entire Phanerozoic, the best extinction and populations to identify traits varying rate could be calculated. This would seem within species) suggests that far fewer ole- sensible given that extinction rates for nelloid range extensions stretch into the Cambrian trilobites seem to be unusually early Cambrian. Thus, the "data" in this high (Foote 1988) and also given the

9 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

study represent a model that appears to diluvian introduction to cladistics, citing not be highly unsound. a single methods paper published after Even if one uses an accurate phylog- 1985. The simplistic summary of cladistics eny and even if one accounts for extinction discusses neither the conditions under properly, there are additional potential which cladograms will replicate basic phy- problems. Contrary to common assertion logenetic structure, nor the numerous sim- (e.g., Adrain and Westrop below), the dis- ulation studies showing the variety of tribution of phylogenetic range extensions circumstances sufficient to mislead cladis- depends on numerous sampling factors. tics, nor the numerous empirical studies For example, parameters such as variable showing that different data sets evolved on sampling over time systematically the same phylogeny produce markedly dif- decrease the number of expected range ferent cladograms. The methods that have extensions (Wagner 1998). Second, it largely supplanted traditional cladistics are must be remembered that speciation and not even mentioned. This guide might extinction parameters make predictions illustrate how phylogenetics used to be about true richness, not sampled richness. done, but one would see little resembling Even if phylogeny is accurately recon- this approach at modern systematics structed, the implied richnesses depend meetings. heavily on sampled richness, which in turn Adrain and Westrop take issue with depend not simply on sampling but also on recent methods designed to test phyloge- abundance distributions (Hurlbert 1971). netic hypotheses with morphological and Richness can be estimated only by extrap- (especially) stratigraphic data. Indeed, olation (e.g., Efron and Thisted 1976) much of the paper is an attempt to dis- whereas phylogeny can only interpolate a credit my work, which the authors minimum number of unsampled ancestors (mis)characterize as "voicing as many (which probably are few relative to unsam- objections as possible" to the use of tradi- pled side branches). The most obvious tional parsimony. Aside from the fact that solution is one to which I already have my research program actually is about referred. Methods developed by Foote testing macroevolutionary hypotheses, I (2001) can test speciation, extinction and have yet to voice a novel objection to par- sampling rate parameters simultaneously. simony. Most of the objections to parsi- Modifications to these techniques could mony raised by systematists in issue after account for abundance distributions when issue of Systematic Biology predate me, evaluating sampling parameters. Now, this never mind my research career. The initial is not to say that phylogeny will be irrele- presentation of parsimony (Edwards and vant - for example, one could use general Cavalli-Sforza 1964) noted that under very phylogenetic relationships in the manner particular circumstances (i.e., such low to test for differences among clades or rates of change that the probability of sta- over time (e.g., Magallón and Sanderson sis along each branch is nearly 1.0 and 2001). However, even there different gen- with no variation in those low rates across eral phylogenies should be used (a la branches) the most probable character Sanderson and Donoghue 1994) if only to matrix is one in which the shortest network provide sensitivity analysis. among the character states (i.e., a cla- Two chapters deal with the topic of dogram) matches the basic phylogenetic phylogenetic reconstruction. McLennan structure. Edwards and Cavalli-Sforza and Brooks contribute a remarkably ante- also noted that this probably never hap-

10 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

pens in the real world. (Edwards [1996] assumption that cladistic topology essen- provides an interesting account of the tially determines character state distribu- early history of parsimony and phylogenet- tions is false is a necessary conclusion of ics.) empiricism and simple deduction. That is, Perhaps Adrain and Westrop's misun- if cladistic topology alone determines the derstandings of modern systematics is shortest network among character states best summarized by their claim that parsi- and assuming that there is only one phy- mony is empirical and pattern-based logeny linking a set of taxa, then we whereas methods such as likelihood are should never see different cladograms for modeling and process-based. This claim is the same taxa given different data parti- fundamentally incorrect in multiple ways. tions (e.g., two different sets of morphol- First, if we restrict ourselves to character ogy, morphology and molecules, etc.) data, then both approaches are equally Instead, not only are different cladograms empirical-the empiricism begins and ends the norm, but radically different cla- with the character data. This, too, is the dograms for the same taxa are far from beginning and end of the pattern- hypothe- uncommon. There now are only two possi- ses have likelihoods only given observed ble conclusions: either the creationists are patterns, after all. Where cladistics and correct or the basic premise of cladistics is likelihood differ is in the parameters used false. Given the preponderance of data to explain/predict the character patterns. showing that the first idea is false, this Cladistics assumes that factors other than essentially falsifies the "cladistic topology cladistic topology are (for all intents and 'determines' character state distributions" purposes) irrelevant to basic character idea and seriously diminishes the trust we state distributions. Likelihood assumes can place in a method requiring that to be that phylogeny (including branch lengths) true. This also falsifies Patterson's "all and rates (including relative rates of homologies will be synapomorphies (con- change, differences in transition rates gruent)" premise unless we are willing to among states, correlated change, varia- assume that some data sets completely tion in rates across the tree, etc.) all affect lack homologies for certain portions of the character state distributions and thus that tree. These two details lead to the reason we have no expectations given a hypothe- why the simulations by myself and other sis about one parameter without reference workers are important - given that phylog- to many others. Process does not enter eny certainly affects (but does not deter- into the picture, except insofar that the mine) character state distributions, when demonstration that two parameters are does the shortest network among charac- significantly better than one (say, two rates ter states begin to reflect factors other instead of one) indicates that more than than cladistic topology? one process (whatever those might be) Still, Adrain and Westrop state that were at work. because we cannot know the actual Their misunderstanding of the param- amount of change and homoplasy without eter issue becomes important when knowing the phylogeny (a statement that Adrain and Westrop question the rele- is itself false unless every species is sam- vance of these simulations and new meth- pled!), simulations cannot be used to eval- ods, noting that simulations are based on uate real data sets (a la Wagner 1998). the premise that parsimony is often (or However, we do not need to know abso- even always) wrong. However, the lute rates to evaluate data - we need to be

11 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

able to measure data properties such as determines whether observations deviate minimum rates of change in real and simu- from those expectations). The only crite- lated matrices. We have amply demon- rion for whether data of any sort test a strated that simulated matrices showing hypothesis is: does a hypothesis make the same minimum rates of change as real predictions about what that data should matrices yield erroneous cladograms. look like? If so, and if the observations are Indeed, many empirical matrices show met, then the hypothesis passes this test minimum rates exceeding the highest min- and is considered (for now) to be one of imum rates of published simulations. the likely explanations. If not, the hypothe- Adrain and Westrop's additional claim that sis fails the test and is considered unlikely. such assessments are somehow subjec- (Note that Boolean modus tollens deductive because matrix structure metrics are tion is a special case of a likelihood argu- affected by how characters evolved com- ment, in which the likelihood is zero and pletely misses the point; these only dem- the hypothesis must be false - thus, falsifi- onstrate that nuisance parameters such as cation also is a special case of a likelihood variable rates, correlated change, etc., argument.) Adrain and Westrop advocate have such a strong effect on character an orthogonal logic known as abduction state distributions that we cannot evaluate (see Sober 1988). Abduction is exempli- phylogeny without also evaluating those fied by the saying that (given that ducks parameters. have flat bills, waddle and quack) if it Adrain and Westrop's discussion looks, walks and quacks like a duck, then about the utility of stratigraphic data and it is a duck. Essentially, abduction is how workers are using it reveals additional unquantified Bayesian probability in which misconceptions concerning phylogenetic we assert that the probability of a waddling methods. They distinguish between intrin- quacking billed-thing being a duck is 1.0. sic information (e.g., morphology and Of course, without Bayesian formalization, other things that are a heritable properties this "common sense" really is logical fal- of individuals) and extrinsic information lacy. (For those keeping score, induction is (e.g., stratigraphic ranges and other things different yet again, representing argu- that are not directly properties of individu- ments to support generalizations such as als) and the relevance of such data to phy- "ducks quack", which is an assumption for logenetic issues. Because phylogeny either deduction or abduction.) concerns inheritance, they argue that The likelihood:Bayesian dichotomy shared morphologies offer evidence of exemplifies the difference in logic applied relationships whereas absences of strati- by workers such as myself and that graphic gaps do not. This is entirely true applied by workers such as Adrain and and it entirely misses the point. We are (or Westrop. If we can assume that most spe- at least I am) not interested in gathering cies sharing a trait do so because of com- evidence to support phylogenetic infer- mon ancestry, then it is probable that they ences. Instead, we are interested in test- are closely related if they share traits. ing ideas about phylogeny and other Thus, if two species waddle, quack and aspects of evolution with evidence. otherwise look like ducks, they probably Hypothesis testing can be done only in a belong to a clade of ducks. Of course, if deductive framework (i.e., where one rates of homoplasy are high enough, then explicitly states the expectations of a most species sharing a trait probably do hypothesis or set of hypotheses and then not do so because of common ancestry,

12 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

and the probability that the species share environmental differences among species the trait by common ancestry actually is within the same clade mean that we do not less than the probability that they share it expect to sample members of subclade A because of convergence. And this is just because we are sampling members of exactly why stratigraphy does not provide subclade B. Adrain and Westrop docu- information about how species are related ment such environmental heterogeneity - most contemporaneous species in the among species within a clade of Cambrian fossil record are not close relatives. How- trilobites. What would have been interest- ever, we expect close relatives to be con- ing here is quantitative tests of species temporaneous or nearly so. Thus, it is durations within these different sampling likely that contemporaneous species are units, using the protocol outlined by Mar- close relatives and unlikely that species shall (1995; Marshall and Ward 1996) and separated by large gaps in sampling are myself (Wagner 1995). Unfortunately, no close relatives. It is critical to understand such calculations are made. Nor are any what I have just written: stratigraphy pro- analyses presented that show how strati- vides no evidence against the idea that graphic data might have misled analyses. contemporaneous species are relatives - Indeed, a more interesting question of how the likely hypotheses pass this test. (The changes in environment across phylogeny vast majority of phylogenies passing a affect this - if changes are common, then stratigraphic test will fail subsequent tests we actually expect few gaps in phylogeny using morphology: it is improbable that even if sampling only one environment close relatives will be as different as bra- simply because lineages would keep chiopods and trilobites even under high "criss-crossing" through that environment. rates of change, and thus phylogenies Conversely, if the differences typified sub- suggesting this are unlikely.) However, the clades, then we would expect simply a few failure to sample a Silurian link between long (and not unlikely) range extensions an Ordovician and a Devonian species is spanning the gaps in facies sampling. Do unexpected if we have sampled numerous we need to account for more parameters species from the same clade and from the when evaluating hypotheses given strati- same environments and geographic units. graphic data? Almost certainly. However, Thus, trees positing such a gap given such the same statement is true for morpho- data fail our test and are dismissed as logic data. unlikely. To a hypothetico-deductive All of this could be rendered moot if, worker such as myself, the concern never as Adrain and Westrop imply, the failure is is about what inferences data support not with parsimony but with comparative unless we can deductively eliminate all data. Students of mammal teeth and snail rival inferences. Instead, it is about what shells should either get their acts together ideas we can reject, and which ones or get different groups, they say. This is require more data and better methods quite an audacious statement coming from before we can reject them. trilobite workers given that trilobite matri- Of course, we need to know the ces often look almost as much like the expectations for stratigraphic distributions work of random number generators as of given hypothesized durations to test the work of Markov processes (and almost hypotheses. Adrain and Westrop's one never so well structured as most gastro- potentially useful contribution concerns pod matrices, I might add!). Moreover, stratigraphic sampling regimes and how advances in comparative data often add

13 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY homoplasy as well as reduce it. For exam- Edwards, A. W. F. 1996. The origin and early develop- ple, Webster's above-cited analyses ment of the method of minimum evolution for the reconstruction of phylogenetic trees. Systematic reduced homoplasy among olenelloid trilo- Biology 45:79 - 91. bites substantially for many characters, Edwards, A. W. F., and Cavalli-Sforza, L. L. 1964. but also added homoplasy for others by Reconstruction of evolutionary trees. p. 67 - 76 In Heywood, J. H. and McNeil, J. (eds.) Phenetic and showing that states linking supraspecific phylogenetic classification. Systematic Associa- OTU's in early analyses varied homoplas- tion, London. tically among species within those OTU's. Efron, B., and Thisted, R. 1976. Estimating the number My own survey of published studies (Wag- of unseen species: how many words did Shakes- peare know? Biometrika 63:435 - 447. ner 2000b) found that further study added Foote, M. 1988. Survivorship analysis of Cambrian and homoplasy quite frequently. Given this, I Ordovician trilobites. Paleobiology 14:258 - 271. have little hope that comparative analyses Foote, M. 1997. Estimating taxonomic durations and will ever become sophisticated enough preservation probability. Paleobiology 23:278 - 300. Foote, M. 2001. Inferring temporal patterns of preserva- that they will reduce coded homoplasy to tion, origination, and extinction from taxonomic survi- levels where parsimony has much of a vorship analysis. Paleobiology 27:602 - 630. chance to recover phylogeny. It is not time Harvey, P. H., and Pagel, M. D. 1991. The comparative method in evolutionary biology. Oxford Press, to get new groups; instead, it is time to Oxford. adopt methods that use logical formalisms Huelsenbeck, J. P., Rannala, B., and Larget, B. 2000a. A and dismiss those based solely on philos- Bayesian framework for the analysis of cospeciation. ophy. Evolution 54:352 - 364. Huelsenbeck, J. P., Rannala, B., and Masly, J. P. 2000b. In summary, MacLeod and Forey have Accommodating phylogenetic uncertainty in evolu- produced a masterful volume that repre- tionary studies. Science 288:2349 - 2350. sents an important read for workers inte- Hurlbert, S. H. 1971. The nonconcept of species diver- grating morphometric and phylogenetic sity: a critique and alternative parameters. Ecology 52:577 - 586. analyses in any capacity. The papers deal Larget, B., and Simon, D. L. 1999. Markov Chain Monte with a wide range of issues pertinent to Carlo algorithms for the Bayesian analysis of phylo- paleontologists in particular and evolution- genetic trees. Molecular Biology and Evolution 16:750 - 759. ary biologists in general, and it offers the Lewis, P. O. 2001. Maximum likelihood phylogenetic potential to provide both solutions to prob- inference: modeling discrete morphological charac- lems and inspirations for new analyses. ters. Systematic Biology 50:913 - 925. However, I cannot recommend Adrain et Lieberman, B. S. 2001. A test of whether rates of speciation were unusually high during the Cambrian radia- al.'s volume. Although there are some out- tion. Proceedings of the Royal Society of London, standing papers, too many cling to yester- Series B. Biological Sciences 268:1707 - 1714. day's methods and it offers little Marcot, J. D. 2000. Evaluating tests of evolutionary advancement. Readers would do better to mode using simulation. Geological Society of Amer- ica - Abstracts with Program 32:A445 - A446. devote the same time to catching up on Magallón, S., and Sanderson, M. J. 2001. Absolute the advances in analytic methods regularly diversification rates in angiosperm clades. Evolution presented in Systematic Biology, Evolu- 55:1762 - 1780. Marshall, C. R. 1995. Stratigraphy, the true order of spe- tion and Paleobiology. cies' originations and extinctions, and testing ancestor-descendant hypotheses among Caribbean REFERENCES bryozoans. p. 208 - 236 In Erwin, D. H. and Anstey, Alroy, J. 1995. Continuous track analysis: a new phylo- R. L. (eds.) New approaches to studying specia- genetic and biogeographic method. Systematic tion in the fossil record. Columbia University Press, Biology 44:152 - 178. New York. Cranston, P. S., and Humphries, C. J. 1988. Cladistics Marshall, C. R., and Ward, P. D. 1996. Sudden and grad- and computers: a chironomid conundrum? Cladis- ual molluscan extinctions in the latest Cretaceous of tics 4:72 - 92. the western European Tethys. Science 274:1360 - 1363.

14 WAGNER, PETER J.: FOSSILS, PHYLOGENY, AND FORM: AN ANALYTICAL APPROACH MORPHOLOGY, SHAPE AND PHYLOGENY

McCracken, K. G., Harshman, J., McClellan, D. A., and Wagner, P. J. 2000a. The quality of the fossil record and Afton, A. D. 1999. Data set incongruence and corre- the accuracy of phylogenetic inferences about sam- lated character evolution: an example of functional pling and diversity. Systematic Biology 49:65 - 86. convergence in the hind-limbs of stifftail diving ducks. Wagner, P. J. 2000b. Exhaustion of cladistic character Systematic Biology 48:683 - 714. states among fossil taxa. Evolution 54:365 - 386. Nee, S., Holmes, E. C., May, R. M., and Harvey, P. H. Wagner, P. J. 2000c. Phylogenetic analyses and the fos- 1994. Extinction rates can be estimated from molecu- sil record: tests and inferences, hypotheses and lar phylogenies. Philosophical Transactions of the models. p. 341 - 371 In Erwin, D. H. and Wing, S. L. Royal Society of London Series B 344:77 - 82. (eds.) Deep time - Paleobiology's perspective. Sanderson, M. J., and Donoghue, M. J. 1994. Shifts in Paleobiology Memoir 26 (Suppl. to No. 4). diversification rate with the origin of angiosperms. Wagner, P. J. 2001. Gastropod phylogenetics: progress, Science 264:1590 - 1593. problems and implications. Journal of Paleontology Smith, A. B. 1988. Patterns of diversification and extinc- 75:1128 - 1140. tion in early Palaeozoic echinoderms. Palaeontol- Walker, T. D., and Valentine, J. W. 1984. Equilibrium ogy 31:799 - 828. models of evolutionary species diversity and the Sober, E. 1988. Reconstructing the past. MIT Press, number of empty niches. The American Naturalist Cambridge. 124:887 - 899. Swiderski, D. L., Zelditch, M. L., and Fink, W. L. 1998. Webster, M. 2002. Ontogeny and paleobiology of olenel- Why morphometrics is not special: coding quantita- loid trilobites. Geological Society of America - tive data for phylogenetic analysis. Systematic Biol- Abstracts with Program 34:266 - 267. ogy 47:508 - 519. Zelditch, M. L., Fink, W. L., and Swiderski, D. L. 1995. Wagner, P. J. 1995. Stratigraphic tests of cladistic Morphometrics, homology and phylogeny: quantified hypotheses. Paleobiology 21:153 - 178. characters as synapomorphies. Systematic Biology Wagner, P. J. 1998. A likelihood approach for estimating 44:179 - 189. phylogenetic relationships among fossil taxa. Paleo- biology 24:430 - 449.