Geographic And Ontogenetic Variation InMorphology Of Australian Waratahs(Telopea: Proteaceae)

I look at the term species, as one arbitrarily given for the sake of convenience, to a set of individuals closely resembling each other, and that it does not essentially differ from the term variety.. .. (Dar­win, 1859:52)

The “species problem” has a long history and perhaps is little closer to resolution than when Darwin said the above—a sur­prising statement, given that he set in train the succession of events leading to the cur­rent debate about phylogenetic species concepts. Many species concepts have been proposed, and we will not review them here, but papers in Taxon (41:307-320) in response to Andersson (1990) are a sample of the divergent views held currently.

Abstract.—In this empirical study of species boundaries in a small genus of plants, we take the view that species are ambivalent; some appear to be monophyletic taxa, but some lack autapomorphies and are metataxa. As an operational definition, we recognized species from differentiated clusters in phenetic space whose distinctness was assumed to be the manifestation of underlying, fixed, and qualitative differences following speciation. These units were consid­ered appropriate as terminals for phylogenetic reconstruction. The appropriateness of different phenetic methods in relation to models of infraspecific geographic variation and evolution is discussed. At the population level, ordination was more suitable than either cladistics or cluster analysis because it does not impose a rigidly hierarchical pattern on the data when none is expected. Variation among populations of Telopea was investigated by phenetic analysis of adult morphology. The main questions were whether the conventional distinction of T. mongaensis Cheel from T. oreades F. Muell. could be justified and whether disjunct populations referred to T. speciosissima (Smith) R.Br. in the Gibraltar Range, northern New South Wales, constituted a distinguishable taxon. The Gibraltar Range waratahs were distinguishable from typical T. spe­ciosissima by their abundant ferruginous hairs, elliptic to obovate leaves, and numerous teeth along the lower half of the leaf margin; we propose recognizing them as a new species. Ordination analysis revealed a strong ontogenetic pattern within populations of T. speciosissima sensu lato, indicating that adult plants were retaining lobed intermediate leaves. Canonical variate analysis confirmed that this pattern was distinguishable from the between-population geographic pattern, but cluster analysis confounded the geographic and ontogenetic patterns. Conventional recog­nition of T. oreades and T. mongaensis as distinct species was supported by both ordination and cluster analysis. One population was mixed, with little evidence of hybridization between the sympatric species. Canonical variate analysis of populations was confounded by the heteroge­neous population. [Geographic variation; ontogeny; species problem; morphometries; phenetics; Proteaceae; Telopea.]

analysis, whether these correspond to “species,” and if so, how to conceptualize a species in a phylogenetic system (Baum, 1992). Some cladists (Nelson, 1989a, 1989b; Vrana and Wheeler, 1992) contend that there is no “species problem” because con­trary to the “modern” synthetists (e.g., Mayr and Simpson) there is no fundamen­tal unit of evolution. In this view, all taxa are equivalent and monophyletic and are recognized only by synapomorphy. What­ever their hierarchical level, taxa evolve only in the sense that they differentiate by developing new apomorphies (Nelson, 1989a, 1989b). A consequence of this view is that taxa can only be recognized by cla­distic analysis, and taken to its extreme, the only terminal units suitable for anal­ysis can be individuals (Vrana and Whee-ler, 1992).

This regress need not stop here, for in some sense, even individuals may not be monophyletic (de Queiroz and Don­oghue, 1988). This view seems to give no starting point for cladistics that is inde­pendent of the method itself. Vrana and Wheeler (1992) argued that all observation is done at the level of the individual any­way, but they begged the question of what is a representative sample of individuals for analysis. Moreover, their approach would lead to the “discovery” of spurious incongruence in the analysis of sexually dimorphic taxa. At least some sexually polymorphic characters would be apparent synapomorphies for gender groups rather than clades and would thus conflict with sexually monomorphic characters.

Many cladists, including ourselves, take the view that species are different from higher level taxa. They recognize a point at which populations of biparental organ­isms that show basically reticulating (to- kogenetic) relationships split into inde­pendent lineages that show primarily ancestor-descendant (phylogenetic) rela­tionships. Hennig's diagram (1966: fig. 6) is a much-cited example of this model. Of course, the “changeover point” of rela­tionships is often blurred, because di­verged lineages, especially plants, retain the plesiomorphic ability to interbreed (Rosen, 1979); however, this does not alter the generality of the model. Controversy has arisen over definitions (reviewed by Baum, 1992). Some authors (Nelson and Platnick, 1981:11; Wiley, 1981:31; Cracraft, 1989:35; Wheeler and Nixon, 1990) have emphasized “diagnosability” as the crite­rion for a phylogenetic species. In a phy­logenetic system, a synapomorphy is the only diagnostic character (Hennig, 1966); however, it is widely agreed that a newly separated lineage does not instantly (or in some cases ever) acquire an apomorphy shared by all individuals of that group (Hennig, 1966; Wiley, 1981; de Queiroz and Donoghue, 1990a), and limitations on sam­pling can lead to difficulty in recognizing apomorphies (Nelson and Platnick, 1981). Given these difficulties, a group of organ­isms that has a separate historical fate may

nevertheless be diagnosable only by sym- plesiomorphies or, worse, by a unique combination of characters (Nelson and Platnick, 1981; Wheeler and Nixon, 1990; Baum, 1992), none of which are shared by all individuals in the population (which is then “polythetic”). Proponents of the diag­nosability criterion have not come up with a clear solution to this problem (e.g., Cra­craft, 1989, 1992). Moreover, de Queiroz and Donoghue (1990a:68) accused these authors of treating species and higher taxa by different standards. We shall return to this point below.

De Queiroz and Donoghue (1988,1990a, 1990b) have provided an extended discus­sion of the tension between two contrast­ing species concepts, each having its lim­itations and neither able to classify all organisms. Species can be defined as the most inclusive populations, using repro­ductive cohesiveness as the criterion. However, this concept fails to deal with uniparental organisms. A species can also be defined as the least inclusive mono­phyletic group. These authors provide an important refinement of the criterion of monophyly by recognizing two implicit components: (1) descent from a common ancestor and (2) exclusivity, which rules out reticulating groups such as parts of populations. In their view, a monophyletic species concept is inapplicable to ancestral populations because these populations do not include all the descendants of their common ancestor. This conundrum ap­pears to have led Nelson (1989a, 1989b) to effectively abandon any species concept: paraphyletic taxa are anathema and can only be dealt with by splitting them into monophyletic subtaxa. This process will regress down to the level where relation­ships cease to be hierarchical (ancestor­descendant), and monophyly is no longer demonstrable by synapomorphy. De Quei­roz and Donoghue defined monophyly ac­cording to ancestor-descendant relation­ships rather than by synapomorphy (cf. Hennig, 1966). As they pointed out (1990a: 70-71), new apomorphies do not spread instantaneously through a population, and so a new phylogenetic species may show

no synapomorphy shared by all individ­uals until some time after the point at which its lineage split from that of its ancestor. This delay can apply to either, neither, or both of two daughters of an ancestral lin­eage. Donoghue (1985) and de Queiroz and Donoghue (1988) proposed that a group lacking a synapomorphy, rather than be­ing treated as “paraphyletic” (Nelson, 1989a, 1989b), be treated as unresolved and labeled accordingly (as a “metataxon”). Thus, they explicitly recognized that such groups may as likely turn out to be mono­phyletic as paraphyletic, either after fur­ther investigation (discovering hitherto unobserved apomorphies) or time (to evolve a synapomorphy).

The “diagnosability” criterion proposed by Cracraft and others is in effect a com­promise between the views of species as populations and as monophyletic groups (a disjunctive species concept: de Queiroz and Donoghue, 1988:334). De Queiroz and Donoghue (1988:334, 1990a:67) rejected dualism as imprecise and confusing, being based upon a combination of properties (both monophyly and reproductive cohe­sion). However, they pointed out (1990b: 89) that “in biparental organisms, repro­duction and common descent are inti­mately tied to interbreeding.” In other words, interbreeding (predominant in populations) and common descent (pre­dominant in monophyletic groups) are manifestations of a single process, repro­duction. Thus, a dualistic concept could be viewed as unifying rather than confusing.

A distinction must be drawn between theoretical concepts and operational defi­nitions of species. A phylogenetic species concept is needed in a research program directed at deriving estimates of phylog­eny among taxa, but a working unit also is needed as a way of getting started in phy­logenetic reconstruction. This paper fo­cuses on the means of recognizing these operational units. We have adopted Nel­son and Platnick's (1981:11) operational definition. We propose that species can be recognized as clusters in phenetic space because their distinctness must be due to some underlying factor, such as infraspe­

cific polymorphism (e.g., sex differentia­tion or ontogenetic stages), pleiotropic ef­fects of a single gene difference (an unlikely explanation), or phylogenetic divergence. Correlated geographic patterning can help to distinguish divergence from intrapopu- lational effects (cf. Cracraft, 1992). Phenetic differentiation, therefore, is the observable manifestation of underlying fixed quali­tative differences (expected of monophy­letic taxa; Nelson and Platnick, 1981) that are either observed as synapomorphies or predicted following speciation (de Queiroz and Donoghue, 1990a: fig. 4). Such units are suitable as terminals in cladistic anal­ysis. They have the methodological advan­tage of being independent of that proce­dure as well as precluding any need to begin at absurdly atomistic levels (cf. Vrana and Wheeler, 1992).

Operational units detected as phenetic clusters might be compatible with more than one phylogenetic species concept, but is there a conceptual definition of a species that corresponds to our operational defi­nition? We return to the tension recog­nized by de Queiroz and Donoghue (1989, 1990a, 1990b). Sometimes species appear to be taxa (they possess autapomorphies and can be said to be monophyletic), and some­times they appear to be populations (they lack autapomorphies and can be said to be metataxa). Perhaps this duality should be accepted, and species should be treated as an ambivalent concept.

Models of Pattern among Populations

A variety of techniques have been used for analyzing geographic variation (see Gould and Johnston, 1972; Thorpe, 1976, 1983b; Whiffin, 1982, for review); however, the question of how to choose an appro­priate algorithm, given a set of assump­tions about underlying biological (evolu­tionary) processes, has frequently been neglected. The choice of technique for quantitative analysis should depend upon the type of pattern postulated to exist in the study group. Many patterns of geo­graphic variation within species, both real and hypothetical, have been recognized (Mayr, 1962; Gould and Johnston, 1972;

Endler, 1977). Levels of gene flow, selec­tion, mutation, dispersal, and extinction all affect geographic pattern; all these pro­cesses, but especially gene flow, seem to vary widely among species (Slatkin, 1985). Historically, the role attributed to gene flow has changed. Mayr (1962) argued for its importance as a homogenizing force among individuals of “biological species.” Sub­sequently, this view was rejected on the basis that limitation of dispersal and strong local selection pressures should outweigh the homogenizing effects of gene flow, leading to divergence between popula­tions (Ehrlich and Raven, 1969; Endler, 1977; Barton, 1989). Slatkin (1985), al­though observing that most species are morphologically uniform, gave examples showing wide variation among species in levels of gene flow. He postulated that sto­chastic rare but significant events of gene flow had almost certainly been underem­phasized but concluded that gene flow did not necessarily account for morphological uniformity within species. Other authors have explained genetic and phenotypic continuity between populations within species by arguing that selection favors in­termediate rather than extreme genotypes (Eldredge and Gould, 1972) or that macro­mutation is the primary cause of diver­gence but is a rare event (Lovtrup, 1987).

Given these various models, phenotypic characters may be distributed among pop­ulations in a variety of patterns, e.g., retic­ulate, clinal, or whatever, but they are un­likely to be primarily hierarchical (Thorpe, 1976:432). Hierarchical patterns arise from diverging, noninterbreeding lineages as a result Of speciation and, as such, call for cladistic analysis, which is explicitly based upon such a model (Hennig, 1966; Nelson and Platnick, 1981; Wiley, 1981). Within species, the pattern of relationships among populations is likely to conflict with the basic assumptions of the cladistic model, and cladistic analysis could yield spurious conclusions (Swofford and Berlocher, 1987: 321). At best, the level of homoplasy caused by gene flow would be so high as to render no more than minimal progress likely (Ackery and Vane-Wright, 1984). By con­

trast, de Queiroz and Donoghue (1988, 1990a) recognized monophyletic entities, defined as complete ancestor-descendant lineages, comprised of populations, indi­vidual organisms, cells, organelles, chro­mosomes, or genes. However, their views have been fiercely contested (Nelson, 1989a; Nixon and Wheeler, 1990; Wheeler and Nixon, 1990), and even de Queiroz and Donoghue acknowledged that within spe­cies, hierarchical patterns would be re­stricted in time, to parts of lineages, or to uniparental populations. Genetic data, when analyzed cladistically for popula­tions within species, can show strong hi­erarchical structure, but this structure may (Wallis and Arntzen, 1989) or may not (Lamb et al., 1989) be congruent with re­lationships established from morphologi­cal data (see also Crozier, 1990). Organelle genes in particular have a hierarchical pat­tern of descent because, by contrast with nuclear genes or phenotypic characters, they are inherited uniparentally and tend to become homogeneous within popula­tions (Dowling et al., 1990:260); however, exceptions even to these rules are emerg­ing in plants (Harrison and Doyle, 1990; Harris and Ingram, 1991). Theoretical models have shown that individual genes that exhibit ancestral allelic polymorphism do not necessarily track the phylogenies of organisms when time between successive speciation events is short (Farris, 1977; Nei, 1987; Pamilio and Nei, 1988).

Phenetic cluster analysis also has a hi­erarchical model (although not a phylo­genetic one) and, in this respect, is subject to the same limitation as cladistics for an­alyzing relationships among populations (Thorpe, 1976:432). Ecologists have long been aware of analogous dangers in cluster analysis of vegetation, which usually var­ies continuously (Dale, 1975). Some phe­netic methods, particularly ordination, are designed to reveal multiple, continuous, and overlapping patterns of variation (Sneath and Sokal, 1973; Thorpe, 1976, 1983b). These methods seem to be the most appropriate under a nonhierarchical mod­el of infraspecific variation (Swofford and Berlocher, 1987:321).

Hopper and Burgman (1983) applied cla- distic and phenetic cluster analysis, but not ordination, to populations within a species (Eucalyptus caesia Benth.) postulated to comprise two subspecies. They did not di­rectly address the question of whether re­lationships among populations were ex­pected to be hierarchical or otherwise. Their results showed very little congru­ence between cladograms derived from dif­ferent data sets (morphometries and allo- zymes) for the same set of populations. Congruence in their data was also poor be­tween phenograms and between pheno­grams and cladograms. Nonhierarchical relationships among populations probably were at least partly responsible for this lack of congruence. Colless (pers. comm.) car­ried out a principal component analysis of their data, which provided stronger sup­port for recognition of the two subspecies than did any of the cladograms or pheno­grams.

A study group comprised of several pop­ulations may include more than one spe­cies or a single species in the process of speciation. The purpose of the exercise (as in the present study) may be to determine whether more than one species (or sub­species) can be recognized. Where more than two taxa are involved, the study group should show a mixture of nonhierarchical relationships (at the population level) and hierarchical relationships (at the subspe- cies/species level), and a combination of ordination, clustering, and/or cladistic techniques may be justified to fully eluci­date these relationships. However, in such a study, cladograms or dendrograms should be interpreted cautiously because appar­ently hierarchical relationships shown at the lower levels would probably be false.

Study Group

Telopea is a small genus that includes four named species restricted to southeastern Australia. It belongs to the Proteaceae, tribe Embothrieae, subtribe Embothriinae and is distinguished from the other genera in the subtribe (Embothrium, Oreocallis, and Allox­ylon) by characters of the inflorescence, no­tably the crowding of the flowers and the

large, colored involucral bracts (Johnson and Briggs, 1983; Weston and Crisp, 1991). Weston and Crisp (1987) analyzed the phy­logeny of the Embothriinae, including the species of Telopea, and Crisp and Weston (1987b) outlined the taxonomic history of Telopea and the current state of knowledge of the species. The problems in delimiting three of the four species were discussed, and a proposal to investigate these species using multivariate analysis of morpho­metric data was outlined. This information is briefly restated as follows.

Telopea speciosissima is restricted to New South Wales, Australia, where it extends more or less continuously from Ulladulla northwards to the northern Blue Moun­tains and the Watagan Mountains (near Cessnock). In the Gibraltar Range, be­tween Grafton and Glen Innes in northern New South Wales, there is a restricted oc­currence of populations that have always been referred to T. speciosissima. From an initial examination of specimens in the Australian National Herbarium (CANB) and the National Herbarium of New South Wales (NSW), we observed subtle differ­ences in morphology between the Gibral­tar Range plants and those from popula­tions farther south. However, no single character showed a clear discontinuity be­tween north and south; considerable vari­ation was apparent among the southern populations.

Similar problems have been encoun­tered in distinguishing between T. mon­gaensis and T. oreades. When Cheel (1947) first described T. mongaensis, he gave no satisfactory diagnosis (Crisp and Weston, 1987b). Subsequent publications have questioned the status of the species (Crisp and Weston, 1987b, and references there­in). Telopea mongaensis may differ from T. oreades in habit, presence/absence of a lig­notuber, leaf width, leaf lobing, leaf apex, and/or leaf margin curvature (Cheel, 1947; Crisp and Weston, 1987b; L. Johnson, pers. comm.), but not one of these characters alone affords a clear-cut circumscription of T. mongaensis. After sampling local popu­lations of T. mongaensis and T. oreades, we observed that the leaf tissue of northern

(“typical” T. mongaensis) plants was com­pletely free of sclereids, in contrast to the leaves of southern (“typical” T. oreades) plants, which have abundant sclereids (Fig. 1). This observation was made indepen­dently by T. Ananda Rao (pers. comm.). We suggested that this character may provide a simple diagnosis (Crisp and Weston, 1987b). If it were to correlate with other independently sampled morphological variables, then the existence of two taxa would be strongly corroborated.

(a)

Figure 1. Light micrographs of leaf-lamina transverse sections of (a) Telopea oreades and (b) T. mongaensis. Sclereids (S) are present only in T. oreades. Adaxial surface is uppermost. Hand-cut sections of fresh material stained in lactophenol/aniline blue/fucsin/hemotoxylin.

In this paper, we report upon multivar­iate analyses of the morphological varia­tion in Telopea, undertaken with the pur­pose of resolving these taxonomic problems.

Materials and Methods
Population Samples

In the spring of 1984, three species of Telopea were sampled over their known range (Fig. 2; Appendix). Six individuals were sampled from each of nine southern and five Gibraltar Range T. speciosissima populations, two T. mongaensis popula­tions, and two T. oreades populations. A single stem bearing at least the previous

(b)

season's vegetative growth plus the cur­rent season's inflorescence(s) was taken from each individual. Specimens were taken at midflowering stage (half the flow­ers open, half still in bud) whenever pos­sible. Characters of the inflorescence were measured immediately, then the stems were pressed and dried for later evaluation of vegetative characters. The specimens have been lodged at CANB and NSW. One of the “populations” of T. oreades, from south of Monga, was mixed with respect to sclereid presence and therefore was sub­divided for analysis.

Morphometries

Characters sampled from each individ­ual plant (=stem) are listed in Table 1. Most are self-explanatory, but the following re­quire some additional comments. We treat­ed as ordinal variables some leaf shape characters (apex, basal curvature, and po­sition of widest portion) that could have been quantified more accurately, e.g., by image analysis techniques; however, these techniques were unavailable. Leaf apex shape (character 7) was given three states, which differed slightly among groups. In

Figure 2. Map of Telopea populations sampled. For population abbreviations and locality data, see Ap­

pendix.

Character	T. speciosissima3	T. mongaensis and T. oreades
1. Length of longest involucral bract	+	+
2. Max. width of broadest involucral bract	+	+
3. No. flowers per inflorescence	+	+
4. No. involucral bracts	+	+
5. Length of longest leaf	+	+
6. Max. width of broadest leaf	+	+
7. Leaf apex shapeb	+	+
8. Petiole length (longest leaf)	+
9. Broadest 16 of broadest leap	+
10. Basal curvature of longest leaP	+
11. Diameter of stem 15 cm below inflorescence	+
12. Hair density on leaf undersurfacebc	+
13. Width of leaf lamina 16-way above base0	+
14. No. marginal teeth, lower 16 of leaf laminac	+
15. Total no. marginal teeth on leaf lamina'	+
16. Length of longest mature flower	+
17. No. lobes per leaP	+
18. Depth of deepest leaf-lobe sinusb	+
19. Presence of leaf sclereids'1		+
20. Leaf venation prominence		+

Table 1. Characters evaluated for individuals of populations of three species of Telopea.

Study group

^a Southern and Gibraltar Range populations.

^b See Morphometries section in text for explanation.

^c Tenth leaf below inflorescence.

^d Two-state (binary) character.

T. mongaensis and T. oreades, the states were obtuse or rounded (score = 0), acute (1), and acuminate (2). For T. speciosissima, scores were truncate (0), rounded (1), and acute (2). Scores were averaged over 10 leaves in each individual of T. mongaensis and T. ore­ades and over 5 leaves in each individual of T. speciosissima. We considered measur­ing the angle subtended at the leaf apex, but in T. speciosissima the presence of teeth and lobes made this impractical, whereas in T. mongaensis and T. oreades, the occa­sional presence of acuminate tips made a simple measure of apical angle meaning­less. Hair density (character 8) was sam­pled at the center of the leaf lamina using a dissecting microscope fitted with a grat­icule. Broadest third (character 10) was a measure of leaf shape: whether the lamina was broadest in the lower (score = 0), mid­dle (1), or upper (2) third. Basal curvature (character 11) also was a measure of leaf shape: whether the margins of the lower third of the lamina were concave (score = 0), straight (1), or convex (2). Two appar­

ently independent variables associated with leaf lobing were recorded. Number of lobes per leaf (character 17) was aver­aged over all leaves on the seasonal growth unit being sampled. Small lobes were dif­ficult to distinguish from large teeth, even though these structures were inferred not to be homologous (large lobes have toothed margins). To differentiate lobes and teeth, the relative prominence of the vein lead­ing to the lobe or tooth was used as a cri­terion: in a lobe, this vein was as or more prominent than the vein that branched and looped around to meet the next more distal lateral vein, whereas in a tooth, this vein was less prominent than the looping vein. The second leaf-lobe variable, depth of deepest leaf-lobe sinus (character 18), was the depth of the sinus immediately below (proximal to) the longest lobe on the sea­sonal growth unit as measured along a line parallel to the central axis of the lobe. Al­though presence of leaf lobes might be di­agnostic for T. mongaensis, this character was not recorded for this species or T. ore-

ades because lobed leaves are ontogeneti- cally intermediate (Crisp and Weston, 1987b) and are not normally seen in adult plants. Finding and identifying interme­diate-aged plants was not feasible in the present study. Leaves of plants of T. spe- ciosissima from the Gibraltar Range ap­peared to be harsher in texture—almost prickly—than those from populations far­ther south; however, we could not readily quantify this feature and did not include it in the analysis.

Quantitative Analysis

Two main classes of analytical tech­niques were used in this study: ordination and cluster analysis. Numerous alterna­tives are available within each class, and because they are likely to yield substan­tially different results, we employed sev­eral. Different techniques of ordination employ different indices of similarity or dissimilarity and different criteria for con­structing principal axes and embody dif­ferent assumptions about the data. The na­ture of the data being analyzed affects the appropriateness of a given technique.

Faith et al. (1987) investigated the rela­tive robustness of a variety of dissimilarity measures in ordination, using simulated ecological data. Robustness was measured as the strength, over a range of models, of the linear or monotonic relationship be­tween the input dissimilarities and the cor­responding Euclidean distances between samples in the ordination space. They found that Kulczynski's dissimilarity coef­ficient correlated best with the original dis­tances in ecological space. However, this coefficient is most appropriate when there is reason to assume an intrinsic polarity in the data, e.g., presence versus absence in sites-by-taxa (ecological) data or derived versus primitive in phylogenetic data (Faith, 1989). By contrast, phenetic data are not expected to show polarity, i.e., shared high values and shared low values are ex­pected to be equally informative, and a dis­tance measure that is sensitive only to the separation between objects, such as Man­hattan distance or its range-standardized

form, Gower's distance, is more suitable (Faith, 1984, 1989; Belbin, 1989). Gower's coefficient has the additional advantage of being suitable for mixtures of qualitative and quantitative variables (Gower, 1971), such as our data.

Pimentel (1981) empirically tested sev­eral ordination techniques using a system­atic data set that was known on the basis of biosystematic studies to have a partic­ular structure. The methods that repro­duced this structure most faithfully were considered the best. Pimentel used two metric and two nonmetric ordination tech­niques: principal component analysis (PCA), principal coordinate analysis (PCoA), nonmetric multidimensional scal­ing (NMDS), and nonlinear mapping (NLMAP), respectively. The results from these different analyses were surprisingly variable. The PCoA using Gower's simi­larity coefficient (the complement of Gow­er's distance) performed better than PCA, NMDS was superior to both of the metric techniques, and NLMAP was consistently inferior to all of the others.

Most metric ordination methods can be seen as variants of PCoA. They differ in the way the variables are standardized and in their (often implicit) distance indices. For example, PCA, when based upon a cor­relation matrix among the attributes, is for­mally equivalent to PCoA of a matrix of squared Euclidean distances (Gower, 1966). An important assumption inherent in PCoA is that the relationship between the dissimilarity coefficient and the ordination space is linear, an assumption that often is unrealistic (Pimentel, 1981; Faith et al., 1987). Departure from linearity, as well as interactions between the variables, may re­sult in second and sometimes higher axes that have a systematic relationship to the first axis in PCoA ordinations, the “arch” effect (Gauch, 1982), i.e., variability asso­ciated with the same pattern may dominate several axes, confounding other important underlying patterns. The arch effect is par­ticularly evident in community-ecology data, where there is a pattern of successive replacement of species through space. This

phenomenon has stimulated the develop­ment of metric and nonmetric techniques for which nonlinearity does not have such adverse consequences (Gauch, 1982; James and McCulloch, 1990; see below).

We used seven different ordination al­gorithms: PCA, PCoA, canonical variate analysis (CVA), reciprocal averaging or correspondence analysis (RA), detrended correspondence analysis (DCA), metric multidimensional scaling (MMDS), and NMDS. We briefly describe each of these techniques and note their strengths and weaknesses. For a more detailed discussion of these methods, see Thorpe (1976,1983b), Gauch (1982), Gittins (1985), Minchin (1987), James and McCulloch (1990), and references therein.

PCA is one of the methods most com­monly used in multivariate systematic studies (Thorpe, 1983b; James and Mc­Culloch, 1990: table 1). However, the dis­tance index implicit in PCA, squared Eu­clidean distance, is inferior to others, such as Gower's coefficient for systematic data (Pimentel, 1981) and Kulczynski's coeffi­cient for simulated ecological data (Faith et al., 1987). PCA exaggerates the distinct­ness of outlying individuals, and the coef­ficients of individual components are highly subject to sampling variability (James and McCulloch, 1990). Neverthe­less, PCA has the advantage of being readi­ly understood in geometric terms, unlike more sophisticated methods (see Campbell and Atchley, 1981:269; Gauch, 1982:138­141). Moreover, it may yield results supe­rior to those of other more complex algo­rithms if its assumptions are approximate­ly met (Gauch, 1982; James and McCulloch, 1990). We used PCA based on covariance matrices of variables standardized to unit range (PCA-R) and unit standard deviation (PCA-S).

We also ran PCoA analyses of a Man­hattan distance matrix with variables stan­dardized to unit range (PCoA-RM) and a Euclidean distance matrix with variables standardized to unit standard deviation (PCoA-SE) to see whether the different dis­similarity indices yielded markedly differ­ent results for our data (cf. Pimentel, 1981).

CVA or multiple discriminant analysis is equivalent to PCoA of a matrix of Ma- halanobis's generalized distances (Sneath and Sokal, 1973). It is also mathematically and geometrically closely related to PCA; it is equivalent to a two-stage application of PCA (Campbell and Atchley, 1981; Git­tins, 1985). CVA ordinates predefined groups rather than individuals. It maxi­mizes the between-group variance relative to the within-group variance, thus accen­tuating differences between groups (Campbell and Atchley, 1981). By taking into account the within-group correlation, CVA has the desirable property of negat­ing the effect of information redundancy in the characters (Thorpe, 1976,1983b). We followed Thorpe (1983b) in defining local populations as groups for CVA to test whether they aggregated into higher geo­graphic groupings. We also used CVA to derive discriminant functions for the taxa that we resolved in our study. CVA in­volves the same assumptions as PCoA and also requires that the group covariance ma­trices be homogeneous (Gittins, 1985:75­76). Consequently, CVA results are suscep­tible to the effects of heteroscedasticity and misclassification of predefined groups as well as to nonlinearity.

Reciprocal averaging is so named be­cause the ordination scores of the variables are averages of the ordination scores of the samples, and reciprocally, the ordination scores of the samples are averages of the ordination scores of the variables. This re­sult may be obtained by an iterative pro­cedure in which scores for samples and then variables are estimated by successive approximations until a stable result is achieved (Hill, 1973). RA is equivalent to a weighted PCoA based on chi-squared dis­tances (Gauch, 1982; Faith et al., 1987). Strictly, these are applicable to categorical data consisting of counts (such as the num­ber of species in a site or bracts in an in­florescence), whereas continuous data are best handled by other ordination methods (James and McCulloch, 1990). Only 5 of our 19 variables are of this type (Table 1); however, the results of RA are usually similar to those of PCA (Hill, 1973). RA

differentially weights high shared-attrib­ute values. It was developed to analyze community-ecology data, which show an asymmetry (hence nonlinearity) in the de­gree to which presence or absence of spe­cies reflect underlying environmental gra­dients (Faith, 1989). In morphometric data, relationships between variables are ex­pected to be linear, although this is not always the case (Pimentel, 1981). We used this analysis in conjunction with DCA to check for the effects of nonlinearity in our data.

Detrended correspondence analysis is based on RA but corrects for systematic relationships between pairs of ordination axes (Hill, 1979; Gauch, 1982). These cor­rections are achieved by detrending by segments at each iteration. For example, scores for axis 2 are derived by dividing axis 1 into a number of segments and ad­justing the scores for axis 2 in each segment to have an average of zero. This process results not only in orthogonal axes but also in the elimination of a systematic relation­ship of any kind between any two axes (Hill, 1979; Gauch, 1982). The major dis­advantage is that it is an ad hoc technique that sometimes fails and can even intro­duce distortions of its own (Belbin, 1989; James and McCulloch, 1990). We ran DCA and RA to see whether this form of de­trending had any significant effect on re­sults; that is, to test for the effect of non­linear relationships in the data. The number of detrending segments used was 10 for T. speciosissima and 5 for T. mongaensis IT. ore- ades.

Recent empirical and simulation studies have shown that MDS is the most robust ordination technique available (Pimentel, 1981; Minchin, 1987). Robustness is de­fined here as the ability of the technique to recover an underlying Euclidean ordi­nation space from data that do not fit a simple linear model of responses but may be highly skewed or noisy or show uneven responses in different parts of the space (Minchin, 1987; Faith, 1989). Adding to the appeal of MDS is its criterion of fit, the minimization of “stress,” although this brings with it the danger of the analysis

terminating in a local, rather than a global, optimum (Belbin, 1989). To minimize the latter possibility, we ran each analysis 10 times, starting from different random con­figurations, and checked for convergence of the final stress values. Multidimensional scaling can be either metric or nonmetric, depending upon expected linearity or oth­erwise of the relationship between the dis­similarity coefficient and the underlying ordination space. Whereas metric ordina­tion techniques assume a linear relation­ship, NMDS assumes only monotonicity.

Faith et al. (1987) and Belbin (1989) have developed a “hybrid” MDS technique to use with the Kulczynski dissimilarity co­efficient for data showing intrinsic polar­ity. This technique is called hybrid because it is metric for low dissimilarities and non­metric above a threshold value. The ra­tionale derives from an observation that the relationship between the dissimilarity coefficient and ecological distance is usu­ally linear at small ecological distances but monotonic at best when distances are high (Faith et al., 1987; Faith, 1989). However, morphometric data such as ours are likely to vary continuously and show no inherent polarity, and hybrid scaling seems inap­propriate. We ran both metric and non­metric MDS analyses, using Gower's dis­similarity coefficient. NMDS is more robust than MMDS, but its weaker assumption of monotonicity can result in the loss of use­ful information (Faith et al., 1987).

MMDS and NMDS often give results similar to those of PCA, and their main disadvantage relative to PCA is that the axes are not functions of the original vari­ables, so direct interpretations must be qualitative and subjective (James and McCulloch, 1990). However, Belbin (1989) developed an algorithm, dubbed principal axis correlation (PCC), that fits a set of at­tributes to an ordination space using mul­tiple linear regression. These attributes may be extrinsic or intrinsic (such as the vari­ables used in the original ordination). Principal axis correlation takes each vari­able and finds the location of its best-fit vector in the ordination space, resulting in two pieces of information: the direction of

the vector and its correlation with that di­rection. There is no requirement for the fitted variables to correspond with the ac­tual ordination axes; in any case, the ori­entation of MDS ordination axes is arbi­trary (Belbin, 1989; Faith and Norris, 1989). These vectors can be plotted on a scatter diagram of the ordination and interpreted by inspection, and the correlation coeffi­cients can be taken to indicate their sig­nificance. We used PCC to interpret the contribution of our original variables to the pattern observed in the MMDS and NMDS ordinations.

To complement the ordinations, mini­mum spanning trees (MST) were con­structed. These connect all points (individ­uals) under study with single links to form the tree of shortest length. When com­bined with ordination techniques, MST can identify apparent misplacements of indi­viduals that have occurred in the ordina­tion plot due to the restriction to two or three axes (Sneath and Sokal, 1973:256; Whiffin, 1982).

Several statistics were used to estimate the relative contribution of the original variables (characters) to the patterns vi­sualized in the ordination space. In PCA, the loadings on the principal components were employed, although these are sensi­tive only to patterns parallel with the prin­cipal axes. Their value depends upon the assumption that each influence in the data is confined to a single dimension, which is not necessarily true. With NMDS, we used the PCC algorithm in the PATN pack­age (Belbin, 1989) to find the best linear relationship in any direction between each character and the ordination space, be­cause in this type of ordination, the ori­entation of the axes is arbitrary. In CVA, the coefficients of the latent vectors, weighted by the within-group standard deviations, were used to indicate those variables that contributed most to the dis­crimination between the predefined groups. When postulated taxa are defined as a priori groups, these values estimate the relative diagnostic value of the char­acters. However, Sneath and Sokal (1973: 406) suggest that CVA does not seem to be

a useful guide to diagnostic characters; moreover, a character that is the perfect discriminator (has a unique state for each taxon) cannot be used in a normal CVA analysis because the resulting zero vari­ances would violate the assumption of ho- moscedasticity.

Cluster analysis, like ordination, in­cludes a plethora of algorithms combining different distance coefficients with a vari­ety of rules for forming clusters: whether agglomerative or divisive, hierarchical or otherwise, or overlapping or exclusive and how objects are added to or split from groups (Sneath and Sokal, 1973; Belbin, 1989). Cluster analysis is most appropriate for categorical rather than continuous data, but nevertheless it is very widely used in systematics (James and McCulloch, 1990). Its most serious deficiency is that it forces the data into groups, irrespective of wheth­er they really exist in nature, and it is not efficient when the data are vectors of cor­related measurements (James and Mc­Culloch, 1990:147-148). We carried out some phenetic cluster analyses of the se­quential, agglomerative, hierarchical, and nonoverlapping type. The unweighted pair-group method using averages (UPGMA) forms relatively weak and often unequally sized clusters (Sneath and Sokal, 1973). It is widely favored because it is space conserving; that is, the difference between the input and output distances is mini­mized (Belbin, 1989:88). By contrast, incre­ment in sums of squares (ISS) (Burr, 1970) is very strongly space dilating and pro­duces tight, equally sized clusters, irre­spective of the nature of the data, but car­ries a risk of misclassification of some objects (Belbin, 1989:94). On the CSIRO Di­vision of Entomology VAX computer, we used an algorithm developed by M. Dall­witz (pers. comm.) that varies clustering intensity between zero (equivalent to UPGMA) and one (equivalent to ISS). We ran Dallwitz's algorithm with the intensity factor set at 0.5 to give a moderately space­dilating analysis. In this study, we were mainly interested in the highest order clusters, and a divisive clustering algo­rithm was theoretically more appropriate

(Gauch, 1982; Belbin, 1989:98). Therefore we also ran PDIV, the polythetic-divisive equivalent of UPGMA from the PATN package (Belbin, 1989). All these cluster analyses were generated from matrices of Gower's distance.

Computations were carried out on (1) the VAX-11/750 VMS computer system at CSI- RO Division of Entomology, Canberra, us­ing the Rothamsted GENSTAT statistical package (for CVA and MST) with programs developed by D. Colless (for PCA and PCoA) and M. Dallwitz (ISS) and (2) on MS-DOS personal computers using the PATN pattern analysis package developed by Belbin (1989) for several ordinations and cluster analyses.

Results

Telopea speciosissima

Ordinations.—We have illustrated only four results: those for PCA-R, DCA, NMDS, and CVA (Figs. 3a-d, respectively), be­cause these represent the variation among all the different ordinations. PCA-S and PCoA-SE are almost identical and differ from PCA-R (Fig. 3a) in that axis 3 in PCA-S and PCoA-SE resembles axis 2 in PCA-R, in showing a secondary, ontogenetic influ­ence in the data (see below). A very close resemblance between the results of MMDS and NMDS and between those of RA and DCA implies that nonlinear relationships among variables are not a problem inher­ent in this data set. Because RA is one of the ordination techniques most susceptible to nonlinearity, this conclusion probably holds for the other methods, too. The CVA (Fig. 3d) ordination resembles none of the others in detail, mainly because the algo­rithm ordinates populations rather than individuals. Overall, there are some dif­ferences among results of different tech­niques, including some techniques that theoretically should be similar (e.g., PCA-R and PCA-S). Some segregate different mor­phological trends more clearly than do others, which confirms the wisdom of us­ing several different ordination techniques in the exploratory phase of data analysis (Reyment et al., 1984: chapter 16).

The results of all ordinations have one

significant feature in common: the first di­mension shows geographic separation of the Gibraltar Range populations from the rest of T. speciosissima. In none of the PCA and PCoA ordinations are the individuals fully sorted into discrete geographic groups. Although the Gibraltar Range and southern groups scarcely overlap on axis 1 (e.g., Fig. 3a), there is no obvious morpho­metric disjunction between them. A simi­lar pattern was obtained by MMDS and NMDS (Fig. 3c), although there is no fixed criterion for orienting the axes relative to the scatter of the objects in this form of ordination (Belbin, 1989:108; Faith and Norris, 1989). However, the results for RA, DCA (Fig. 3b), and CVA (Fig. 3d) show distinct clustering into geographic groups; only the Watagans population and to a lesser extent the Bells Line population oc­cupy an intermediate position; these pop­ulations are the most northerly of the southern group. A clearer resolution of geographic races is expected from CVA be­cause the distance coefficient used (Ma- halanobis's) is designed to discriminate between populations by emphasizing be- tween-group variance relative to within- group variance. However, chi-squared dis­tance, the index implicit in RA and DCA, is not designed to achieve this discrimi­nation, and we cannot explain its appar­ently greater discriminatory power over Manhattan distance, Euclidean distance, and squared Euclidean distance in this in­stance.

Most ordinations reveal a secondary pat­tern, which appears to reflect ontogenetic variation within populations. Several in­dividuals, particularly those marked with the letter “j” in Figures 3a-d, are dispersed away from the main geographic groups, usually along axis 2 (although in PCA-S and PCoA-SE, this occurs along axis 3). These individuals belong to four different populations and are well separated from the remaining members of their popula­tions, which fall within the main geo­graphic groups. They come from three southern populations (OK, JB, and TR) and one Gibraltar Range population (TN). On closer examination, these individuals share

extreme values for variables 17 (number of leaf lobes) and 18 (depth of deepest leaf­lobe sinus), both of which are character­istic of developmentally intermediate leaves, as discussed below. Individuals from

(C)

Figure 3. Scatter diagrams in two dimensions from ordinations of all 84 individuals of Telopea speciosissima sensu lato (• = Gibraltar Range, O = Watagan Mountains, 0 = Bells Line, □ = other southern; “j” denotes juvenile/intermediate leaf morphology), (a) Principal component analysis using range-standardized data (PCA-R). Component 1 separates the Gibraltar Range individuals (right) from southern individuals (left). Component 2 separates specimens with ontogenetically intermediate leaves (below) from specimens with entirely adult leaves (above), (b) Detrended correspondence analysis, (c) Nonmetric multidimensional scaling using Gower's distance, (d) The first two canonical variates from CVA. The original fourteen populations were defined as a priori groups for the analysis. The five Gibraltar Range populations have clustered together (right), as have the nine southern populations (left), except that the Watagans population occupies an inter­mediate position.

the same populations that do not separate from the main geographic groups have av­erage to low scores for these characters. In the CVA (Fig. 3d), these individuals are well separated from one another and in-

stead cluster closely with other individuals from the same populations. If CVA is in­sensitive to this pattern, then this separa­tion is occurring within rather than be­tween populations, which is consistent with an ontogenetic interpretation. Axis 2 in the CVA has no obvious interpretation.

Character	PCAa		CVAb		NMDSe
	Axis 1	Axis 2	Populations0	Taxad
1. Bract length	-1.04	-0.57	-0.48	0.21	0.59
2. Bract width	-0.42	-0.52	-0.02	-0.24	0.34
3. No. flowers	-0.23	-0.56	0.02	-0.27	0.23
4. No. bracts	-0.61	-0.45	-0.11	0.29	0.51
5. Leaf length	-0.80	0.21	-0.10	0.01	0.67
6. Leaf width	0.09	-1.02	0.49	0.20	0.11
7. Leaf apex	2.12	0.48	0.68	-0.29	0.61
8. Petiole length	-1.13	0.58	0.09	0.02	0.78
9. Widest 16 of leaf	-1.52	1.19	0.12	-0.03	0.75
10. Leaf base curvature	2.31	-0.52	-0.05	-0.18	0.68
11. Stem diameter	0.39	-0.38	0.06	-0.10	0.32
12. Density of hairs	2.32	0.99	0.93	-1.00	0.83
13. Leaf base width	1.25	-1.07	0.45	-0.32	0.65
14. No. teeth, lower half	1.63	-0.01	0.70	-0.72	0.76
15. Total no. teeth	1.02	-0.07	-0.54	0.53	0.63
16. Flower length	0.38	-0.02	0.20	0.15	0.34
17. No. leaf lobes	-0.07	-1.37	-0.39	-0.13	0.62
18. Depth of leaf sinus	-0.60	-1.46	-0.26	0.27	0.72

Table 2. Contribution of characters of Telopea speciosissima to observed patterns in ordination analyses. Characters and their numbers correspond with Table 1. Extreme values are highlighted in bold and discussed in the text. See also Figures 3a-c.

^a Principal component analysis; loadings on principal components.

^b Canonical variate analysis; coefficients of latent vectors weighted by within-group standard deviations.

^c Populations as a priori groups (first axis).

^d Gibraltar Range versus southern populations as a priori groups (first and only axis).

^e Nonmetric multidimensional scaling; principal axis correlation (PCC) coefficients.

To investigate whether the ontogeneti- cally intermediate individuals were dis­torting the geographic pattern in the data, we reran the ordinations without them. The results were essentially similar; all showed a phenetic separation of Gibraltar Range populations from southern populations, and this pattern was again most pro­nounced in DCA and CVA and to a lesser extent in MMDS and NMDS. Again, the Watagans population occupied an inter­mediate position in the ordination space.

Contribution of characters.—Characters with the highest loadings on the first com­ponent of the PCA were (in order) hair density, leaf base curvature, leaf apex, number of teeth on the lower half of the leaf, and widest third of the leaf (Table 2). This component accounts for 32% of the

total variance, and insofar as it separates Gibraltar Range populations from south­ern populations, these characters contrib­ute most to that distinction. On the second principal component, which accounts for another 13% of the total variance, depth and number of leaf lobes have the highest loadings (Table 2). These characters diag­nose ontogenetically intermediate plants, which are separated by the second com­ponent. Principal axis correlation found strong linear relationships between the NMDS space and hair density, petiole length, number of teeth on the lower half of the leaf, and the widest third of the leaf in a direction corresponding to the geo­graphic separation of Gibraltar Range and southern populations (Table 2), roughly agreeing with the PCA loadings. In the direction corresponding with the separa­tion of juveniles, PCC gave a high corre­lation with depth of leaf sinus, also agree­ing with PCA. In the CVA ordination of populations, the values of the coefficients of the latent vectors, weighted by the with­in-group standard deviations, were high-

1 o 60

5o

Figure 4. UPGMA phenogram, using Gower's distances, of Telopea speciosissima sensu lato. Symbols as in Figure 3. See text for further description.

est for hair density, number of teeth, and leaf apex shape on the first axis, which dis­tinguished the two geographic groups (Ta­ble 2). When the Gibraltar Range and southern populations were defined as two a priori groups for CVA, so that the anal­ysis discriminated between these groups rather than between populations, the weighted coefficients were highest for hair density and number of teeth on the lower half of the leaf, but leaf apex no longer figured strongly (Table 2). Cramer values, which are the ratio of between-group vari­ance to total variance, were calculated for the two groups (Gibraltar Range and southern; Table 3). Cramer values indicate the diagnostic value of characters and vary from 0 (no difference between groups) to 1 (perfect discrimination between groups). Highest values were for hair density (0.89), number of teeth on the lower half of the leaf (0.71), and leaf apex shape (0.59). These results agree well with those of the weight­ed coefficients from CVA.

Minimum spanning tree.—The MST is too complex to show on an ordination scatter diagram, but it is consistent with our geo­graphic interpretation of the pattern in the ordinations; all of the Gibraltar Range in­dividuals form a cluster separate from the southern populations, which themselves form a cluster. However, all but one of the Watagans individuals cluster with the

southern populations, suggesting that this population is not as intermediate as ap­pears from the CVA.

Cluster analysis.—Results from the clas­sification analyses were similar to those from the ordinations but were less infor­mative. UPGMA (Fig. 4) placed all the adult Gibraltar Range individuals into one of the five highest level clusters and put nearly all the southern individuals into another. All individuals from the apparently inter­mediate populations (Watagans and Bells Line) fell into the southern group, except one individual from the Watagans, which fell into the Gibraltar Range cluster. Ju­venile (j) individuals (identified from the ordinations) showed no clear pattern in the phenograms. Nearly all fell into the remaining three high-level clusters; one cluster included the only juvenile from the Gibraltar Range, another included two southern juveniles, and a third cluster comprised three southern juveniles plus one southern adult. One southern juvenile fell into the cluster of southern adults. A similar result was obtained using the space­dilating algorithm of Dallwitz (Fig. 5). The highest and third highest level clusters in­cluded all adult individuals from the Gi­braltar Range and southern populations, respectively, except that again one indi­vidual from the Watagans fell into the Gi­braltar Range cluster. The juveniles showed

no clear pattern, and neither did the in­termediates. The divisive algorithm (PDIV; not illustrated) gave a similar result. In all the cluster analyses, the five highest level clusters included two that were broadly consistent with the first axes of the ordi­nations, but the intermediate populations were mostly classified with the southern populations, and the juveniles were not clearly classified at all. Removal of the ju­venile individuals did not clarify the clus­ter analyses. For example, in the PDIV analysis, two of the five highest level clus­ters corresponded with the Gibraltar Range and southern groups, but the Watagans and Bells Line populations were divided among the major and some minor clusters.

60 4o 1o

1 o

Figure 5. Phenogram (Dallwitz's modified ISS, intensity 0.5) of Telopea speciosissima sensu lato. Symbols as in Figure 3. See text for further description.

Telopea mongaensis and T. oreades

Ordinations.—Most ordinations showed a clear separation of individuals into two groups in the first dimension, e.g., NMDS (Fig. 6a). The other results are similar, ex­cept for CVA. In Figure 6a, the two ordi­nation clusters correspond to the tra­ditional concepts of the two species; all individuals from the Gunrock Creek (GC) and north of Monga (MN) populations, which have been included in T. mongaensis, fell into the left cluster, and all individuals from the Errinundra Plateau population (EP), which have been included in T. oread­es, fell into the right cluster. However, the

fourth population, from south of Monga (MS) was split by the analyses; two indi­viduals were grouped with T. mongaensis, and four were grouped with T. oreades. The MS population is apparently heteroge­neous, including individuals of both T. ore­ades and T. mongaensis.

Canonical variate analysis in which pop­ulations were defined a priori as groups did not yield two discrete clusters (Fig. 6b). The typical T. mongaensis populations (MN and GC) were grouped together and sep­arated from the typical T. oreades popula­tion (EP), and the apparently heteroge­neous MS population was placed in an intermediate position. Unlike the other or­dinations, CVA did not split this popula­tion because the analysis is constrained to minimize between-population variance relative to within-population variance. The scatter diagram (Fig. 6b) appears to show an arch effect (Gauch, 1982).

Because we scored the sclereid character in binary mode (sclereids were either pres­ent or absent), this character was funda­mentally different from the others, which were all morphometric. Moreover, it ap­peared to discriminate perfectly between the two species. To test whether the purely morphometric characters (including all those that had been used traditionally) could distinguish T. mongaensis from T. ore­ades, we reran the ordinations omitting

sclereids. The results were essentially the same as before (Fig. 6c), although the gap between the species is not as evident. Again, the south of Monga population ap­pears heterogeneous, and one individual (MS6), which previously clustered with T. oreades, now occupies a position interme­diate between the species. This individual has abundant sclereids, as do all specimens of T. oreades, but with the sclereid character removed, the remaining morphometric characters show that it also has affinity with T. mongaensis. By contrast, rerunning the

2

(C)

Figure 6. Scatter diagrams in two dimensions from ordinations of all 24 individuals of Telopea mongaensis and T. oreades. Population symbols: ▲ = Errinundra Plateau, • = south of Monga, O = north of Monga, A = Gunrock Creek, (a) Nonmetric multidimensional scaling (NMDS) using Gower's distance. A minimum spanning tree connects all individuals. Telopea oreades (right) is well separated from T. mongaensis (left). The south of Monga (MS) population is heterogeneous: some individuals cluster with each species, (b) First two canonical variates from CVA. Groups defined a priori were the four populations. No separation of species is evident, (c) NMDS except that the sclereid character is omitted from the data. Separation of the species {T. oreades at right, T. mongaensis at left) is still evident but less marked.

CVA without sclereids did not change the results (cf. Fig. 6b).

Contribution of characters.—The first prin­cipal component of the PCA accounted for 62% of the total variance and distinguished strongly between T. oreades and T. mon­gaensis. The loadings on this axis (Table 4) indicate that sclereids, leaf width, leaf length, bract width, leaf apex shape, and, in the opposite direction, venation prom­inence contributed most to that distinc­tion. The PCC analysis (Table 4) on the NMDS ordination (Fig. 6a), which was al-

most identical to the PCA, gave high val­ues for the same set of characters, and the direction of maximum correlation was along or close to the axis of separation of the species. For sclereids, the correlation coefficient was very high (0.98). The CVA ordination of populations gave an arching pattern, which did not distinguish the spe­cies. Correspondingly, a different set of characters contributed to this pattern. On the first axis, leaf width, leaf apex, and, opposite these, leaf length had the largest weighted coefficients, and on the second axis, number of flowers, bract width, and, opposite these, sclereids had the highest weighted coefficients (Table 4). These axes had no obvious biological interpretation. Another CVA analysis was done on the two species, defined as the two primary clusters in the UPGMA classification (Fig. 7), thereby splitting the MS population. The sclereid character was omitted from this analysis because it was invariant and opposite in value in each group, violating the assumption of homoscedasticity. Of the remaining characters, leaf apex, bract width, and, opposite these, venation prom­inence showed high weighted coefficients (Table 4). Cramer values (Table 3) were highest for sclereids (a perfect value of 1.0), leaf width (0.87), leaf length (0.79), bract

	Character	PCAa	Populationsc		Taxad	NMDSe
			Axis 1	Axis 2
1.	Bract length	-0.70	0.13	-0.31	-0.47	0.66
2.	Bract width	-0.99	0.33	-0.86	0.79	0.87
3.	No. flowers	0.11	-0.20	-1.27	0.66	0.82
4.	No. bracts	0.34	-0.30	0.80	-0.66	0.81
5.	Leaf length	-1.17	-0.72	0.58	-0.13	0.88
6.	Leaf width	-1.34	1.00	-0.56	0.65	0.90
7.	Leaf apex	-0.99	0.78	0.08	0.97	0.77
19.	Presence of sclereids	-2.36	0.31	1.18		0.98
20.	Venation prominence	1.33	0.03	0.03	-0.93	0.81

Table 4. Contribution of characters of Telopea oreades and T. mongaensis to observed patterns in ordination analyses. Characters and their numbers correspond with Table 1. Extreme values are highlighted in bold and discussed in the text. See also Figures 6a and 6b.

CVA^b

^a Principal component analysis; loadings on first principal component.

^b Canonical variate analysis; coefficients of latent vectors weighted by within-group standard deviations.

^c Populations as a priori groups (first axis).

^d The two primary clusters in Figure 8 as a priori groups (first and only axis), sclereids omitted.

^e Nonmetric multidimensional scaling; principal axis correlation (PCC) coefficients.

width (0.78), and venation prominence (0.74).

Minimum spanning tree.—The MST (Fig. 6a) confirms the distinction between T. mongaensis and T. oreades made by the or­dinations. Moreover, it supports the ob­servation that the south of Monga popu­lation is heterogeneous by placing individuals on widely separated branches. One individual (MS6) was placed on the branch linking the two species, supporting the suggestion from the ordinations that this population is intermediate.

Cluster analysis.—All of the classification analyses (UPGMA, PDIV, and ISS) agreed in producing two primary clusters corre­sponding to T. oreades and T. mongaensis. Moreover, most analyses reconstructed the populations themselves, e.g., GC, MN, and most of EP in Figure 7. Omission of the sclereid character blurred the distinction between populations but did not alter the deepest level clustering of the two species (Fig. 8). As in the ordinations, the MS pop­ulation was divided between the species. With sclereids included, all algorithms placed individuals MSI and MS4-6 with T. oreades and put individuals MS2 and MS3 with T. mongaensis (Fig. 7). However, when sclereids were omitted, individual MS6 clustered with T. mongaensis, confirming the

MN1 MN4

Distance

Distance

j • MS2

* • MS3

MN6

MN5

MN2

MN3

GC2

GC3

GC5

GC6

GC1

GC4

MS5 MSI

MS4

EP3

EP5

EP1

EP2

EP4

Figure 7. UPGMA phenogram of Telopea mon­gaensis and T. oreades. The two primary clusters cor­respond to T. oreades (below) and T. mongaensis (above). As in the ordinations, the south of Monga population has been split; individuals MS2 and MS3 have been classified with T. mongaensis, whereas MSI, MS4, MS5, and MS6 have been classified with T. oread­es.

-• MS6

A EP6

suggestion that this population is inter­mediate between the species.

Discussion

Telopea speciosissima

The initial hypothesis that the Gibraltar Range waratah populations may constitute a phenotype distinguishable from T. spe­ciosissima sensu strict© is supported by the morphometric analyses. Ordination anal­ysis (Figs. 3a-d) separated the Gibraltar Range populations and the southern pop­ulations into two clusters. There is some morphological overlap among individuals, and the Watagan Mountains population in particular appears to be intermediate (Fig. 3d). This is the northernmost of the south­ern populations sampled (Fig. 2); however, there is no suggestion of a cline connecting

the northern and southern populations be­cause there is a 400-km gap between the Watagans and the Gibraltar Range. The CVA indicates that the southern popula­tions of T. speciosissima are more divergent from one another than are the Gibraltar Range populations (Fig. 3d). Despite this, there is no evidence for recognition of more than two taxa within T. speciosissima. Clus­ter analysis also tends to separate the Gi­braltar Range waratahs from the southern waratahs, but the geographic pattern is not as clear as in the ordination analysis. Only Dallwitz's algorithm (Fig. 5) made this dis­tinction in the deepest level of the pheno­gram. By contrast, the ordinations always showed geographic separation in the first dimension. Cluster analysis is inherently unable to accommodate secondary pattern in the data, thus causing perturbation of the primary pattern.

Figure 8. UPGMA phenogram of Telopea mon­gaensis and T. oreades, with sclereids omitted from the data. The result differs mainly in that individual MS6, previously classified with T. oreades, is now classified with T. mongaensis.

Leaf texture, a character not included in the morphometric analyses because of dif-

Acuity in quantifying it, may show a clear discontinuity between the two taxa. The Watagans waratahs have the softer texture of the southern populations and may not be as intermediate as they appear in the CVA. We suggest that they should be in­cluded in typical T. speciosissima, with which they are geographically contiguous.

Once phenotypic clusters are circum­scribed, the application of taxon status is not straightforward because there are no universally recognized criteria. The geo­graphical-morphological method, where­by scarcely differentiated allopatric taxa are treated as subspecies, has been used wide­ly, but not in a formal or consistent manner (Davis and Heywood, 1963). In practice, this method does not give a clear criterion of rank because closely related allopatric taxa are just as frequently treated as species (e.g., Cracraft, 1992). We follow the view of Stebbins (1950) that there should be dis­continuities between species, and where none are evident, subspecific rank is pre­ferred. Because there is a clear geographic disjunction, corresponding to an almost complete discontinuity in multivariate morphometric space, between the Gibral­tar Range populations and the southern populations of T. speciosissima, we conclude that these should be treated as two species. This treatment is consistent with our op­erational definition of a species.

Ontogenetic variation was revealed by the ordinations as a separate pattern with­in (rather than between) populations of T. speciosissima. In most analyses, the second dimension separated individuals into two groups (Figs. 3a-c): those with lobed leaves in one group and those with simple leaves in the other. Because CVA shows no evi­dence of this pattern (Fig. 3c), it appears to reflect variation between individuals within populations and appears to be in­dependent of the between-population geo­graphic pattern. These groups represent intermediate and adult developmental stages, respectively. Johnson and Briggs (1975:121-123, table 2) recognized in pro- teaceous leaf development a sequence from a pinnately lobed or compound interme­diate (/₂) stage to a simple adult (/₃) stage.

Most Telopea species, including T. speciosis­sima, develop both the /₂ and the /₃ stages. The results of the ordinations suggest that in T. speciosissima, these stages differ not only in the presence or absence of lobing but also in the size of the lobes, as mea­sured by the sinus depth. Although each leaf of T. speciosissima can be assigned to a particular developmental stage, there is not a strict sequence of ontogenetic stages on the plant. Frequently, intermediate and adult leaves are both present on sexually mature plants, often on the same seasonal shoot. Similar and apparently disordered “heteroblastic” sequences occur in other plants, such as Acacia, and have been dis­cussed in more detail by Crisp and Weston (1987a:70) and Weston (1988).

In all of the diagnostic analyses (Tables 2, 3), hair density contributes most to the distinction between southern T. speciosis­sima and the Gibraltar Range populations. In most analyses, number of teeth on the margins of the lower half of the leaf lamina ranked second, and leaf apex shape and widest third of the leaf also were ranked highly in more than one analysis. The dif­ferences between the diagnoses probably reflect the different properties of the sta­tistics, as well as the imprecise relation­ships between ordination axes (or PCC vectors) and the clusters whose distinc­tions are being tested. Those analyses with a precise definition of the groups should be most useful for diagnosis, namely CVA and Cramer values. These analyses all rank (in order) hair density, number of teeth in the lower half of the leaf, and leaf apex shape as the best discriminators.

With respect to the ontogenetic pattern in the data, CVA and Cramer values are uninformative because this pattern is seen within populations, not between groups. Here, the second-axis PCA loadings and the PCC are most useful because these can be aligned with the observed pattern. Both of these analyses indicate that depth of leaf sinuses and, to a lesser extent, number of leaf lobes are most strongly related to this pattern (Table 2). Both characters are in­dicative of ontogenetically intermediate leaves.

Telopea mongaensis and T. oreades

Telopea oreades and T. mongaensis should continue to be treated as separate species. The lack of overlap between the phenetic clusters (Figs. 6a, 7) and the perfect dis­crimination by the sclereid character sat­isfy our operational definition of species. Furthermore, T. oreades and T. mongaensis are sympatric in the Monga area of New South Wales, growing together in one stand while apparently maintaining their integ­rity. Sympatry is a long-standing criterion of the species rank (Ehrendorfer, 1968).

Only the CVA using populations as pre­defined groups (Fig. 6b) failed to clearly distinguish between the taxa, and this was simply the result of heterogeneity in the MS population. The same heterogeneity probably caused the arching pattern in this ordination, because CVA assumes homo­geneity of variances within groups. An arch effect can be caused by a nonlinear rela­tionship between the input distances and the ordination space (Gauch, 1982), but if nonlinearity were a problem in our data, some difference should have been ob­served between PCA and NMDS or be­tween MMDS and NMDS.

The absence of sclereids in T. mongaensis (Fig. 1) distinguishes this species qualita­tively from T. oreades and is an autapo- morphy for the former, given that all other species of Telopea have sclereids (cf. Wes­ton and Crisp, 1987: fig. 3). When the scler­eid character was excluded and the data were reanalyzed, nearly identical results were obtained (Figs. 6c, 8), showing that these species differ morphometrically as well as cladistically. The diagnostic statis­tics confirmed sclereids as best character for distinguishing the species (Tables 3,4). According to the Cramer values, the best morphometric characters for discrimina­tion were (in order) leaf width, leaf length, bract width, and venation prominence. Re­sults of PCA and PCC agreed well with this result, but CVA (of the two species, not the populations) differed in rating leaf apex highest after sclereids. To the extent that leaf width and prominence of veins were used previously (Crisp and Weston, 1987b), the conventional taxonomy of T.

oreades and T. mongaensis would seem to be vindicated. However, in the original treat­ment (Cheel, 1947), prominent veins were incorrectly attributed to T. oreades, whereas the present study indicates that T. mon­gaensis has the more prominent veins. Moreover, all previous workers over­looked the excellent sclereid character (Crisp and Weston, 1987b). Although we have examined many specimens from the Monga area besides those used in the pres­ent study, we have seen only one that ap­pears intermediate in the sclereid charac­ter. All other specimens either have abundant sclereids or none at all; the sin­gle exception has just a few and could be a hybrid.

All analyses revealed heterogeneity within the south of Monga population. The consistent grouping of individuals MS2 and MS3 with T. mongaensis and MSI, MS4, MS5, and MS6 with T. oreades by both ordination and cluster analyses when the full data set was used suggests that this population is mixed rather than hybrid. Nonetheless, comparison of the cluster analyses includ­ing (Fig. 7) and excluding (Fig. 8) sclereids indicates possible breakdown between the species. Individual MS6 was classified with T. oreades when sclereids were included and with T. mongaensis when they were exclud­ed. This individual is closer to T. oreades according to the sclereid character but is closer to T. mongaensis according to leaf shape and size; therefore, it may be a hy­brid.

The geographic situation of the south of Monga population of T. oreades is curious. Although it is coextensive with the south­ernmost population of T. mongaensis, it is disjunct by 150 km from the nearest known population of T. oreades (Fig. 2). Under such circumstances, divergence from southern populations of T. oreades, or introgression with T. mongaensis, or both might be ex­pected. Despite the minor exceptions, this population has maintained its identity with T. oreades to a remarkable degree, suggest­ing that there are both strong reproductive barriers between it and T. mongaensis and little pressure towards divergence. At least part of the reproductive isolation is due to

flowering times: in the Monga area and elsewhere, T. oreades flowers about a month earlier than T. mongaensis. Additional evi­dence is needed to investigate the degree of intergradation between these species, especially in the south of Monga locality, and we have initiated studies on polymor­phism in allozymes and DNA restriction sites.

Comparison of Multivariate Techniques

A retrospective examination of the rel­ative utility of the algorithms used in the present study indicates that ordination analysis was particularly useful. In the first instance, we used it as a heuristic proce­dure, displaying patterns in the data un­burdened by a priori groupings. The on­togenetic influence revealed in the T. speciosissima data was completely unex­pected. The great power of ordination is its capacity to elucidate multiple patterns in the data and to separate them into dif­ferent dimensions, allowing independent underlying variables to be identified, viz. the geographic and ontogenetic factors in T. speciosissima. All ordination algorithms used (MMDS, NMDS, PCA, PCoA, RA, and DCA) revealed the same basic patterns in the data, even though each embodied dif­ferent distance measures and assumptions; thus, we are confident that the patterns are real. Nevertheless, some inconsistencies were seen, especially between data stan­dardized by range and those standardized by standard deviation, e.g., PCA-R and PCoA-RM versus PCA-S and PCoA-SE. These differences emphasize the need to consider carefully the form of standard­ization and distance measure in relation to the kind of data being used (Faith et al., 1987; Belbin, 1989). The diversity in our results probably would have been greater had our data shown nonlinear relation­ships.

Canonical variate analysis is a form of multiple discriminant analysis designed for diagnosis (Sneath and Sokal, 1973). It pro­vides a variety of tests of significance but is laden with such strict assumptions that such tests are invalid for most real data sets (Thorpe, 1983b). In the present study, we

followed the suggestion by Thorpe (1983b) of using CVA to test whether the popu­lation samples tended to aggregate into higher level groupings that might be in­terpreted as taxa. This approach exploits the capacity of CVA to minimize within- group variation relative to between-group variation, thereby suppressing the effects of such within-population phenomena as ontogeny or size and shape variation while accentuating any geographic pattern pres­ent. The outcome was highly successful in the case of T. speciosissima because the on­togenetic influence, so apparent in the PCA (Fig. 3a), was eliminated, and the natural taxa were highlighted (Fig. 3c).

Thorpe (1983a, 1983b) has analyzed and discussed in detail the disturbing effect on ordination of ontogenetic characters and has recommended negating the effect of such data rather than eliminating charac­ters that may be very useful in elucidating geographic pattern. As one alternative, he suggested using multiple group PCA as a way of segregating the ontogenetic influ­ence, which operates within populations. Multiple group PCA and CVA are very similar algorithms in this respect and give virtually identical results (Thorpe, 1983a: 23). In the present study of T. speciosissima, CVA was effective in removing the onto­genetic influence that showed so strongly in the PCA (cf. Figs. 3a, 3c).

Nevertheless, our analysis of the T. ore­ades and T. mongaensis populations pro­vides a caveat about CVA. The unexpected heterogeneity of the south of Monga pop­ulation distorted the result, giving an arch effect (Fig. 6b), because the assumption of homogeneity of variances was violated. If we had used ordination algorithms with more rigid assumptions, we may not have identified this heterogeneity.

Cluster analysis was not as useful as or­dination in the present study. Its hierar­chical model is inappropriate for studies of one or two species. In the analyses of T. speciosissima, individuals from populations identified in the ordinations as morpho­logically intermediate (Watagans and Bells Line), as well as ontogenetically juvenile individuals, were scattered among several

clusters, even at the highest levels of clas­sification. Thus, their relationships to each other and to the main geographic clusters were unclear. Using cluster analysis alone, the juveniles, and thus the ontogenetic in­fluence in the data, would not have been identified because the individuals could not have been identified as having anything in common. It is doubtful whether the Wa­tagans and Bells Line populations would have been identified as intermediate ei­ther, although the classification of some individuals from each into different major clusters may have pointed to this conclu­sion. By its hierarchical, categorical struc­ture, cluster analysis is inherently inca­pable of describing gradients or multiple patterns in data, and therefore ordination is clearly superior in dealing with such data. However, in the case of T. oreades and T. mongaensis, two species that appear to be cladistically divergent (at least according to the sclereid character), cluster analysis was more informative and revealed the possibly hybrid nature of individual MS6. In this case, no secondary influence in the data was strong enough to obfuscate the strong primary, categorical pattern.

Conclusions

Our initial hypothesis that the Gibraltar Range waratahs might constitute a distin­guishable taxon was supported by this study. Because this group is geographically disjunct from typical T. speciosissima and is distinguishable morphologically from all but a handful of individuals of that species, we propose that it be recognized as a spe­cies. A formal taxonomic account will ap­pear in the Flora of Australia (Crisp and Weston, in press).

Conventional recognition of T. oreades and T. mongaensis as distinct species has been supported by this study. However, we have shown that they are better diag­nosed by a somewhat different set of char­acters, especially leaf sclereids, which dis­tinguishes them qualitatively. Moreover, these species are distinguishable mor­phometrically, even when the sclereid character is not used. We will investigate further the suggestion of hybridization in

the south of Monga locality by using mo­lecular techniques.

Ordination techniques were more useful than cluster analysis in elucidating the variation patterns of Telopea. They parti­tioned independent influences in the data (ontogenetic and geographic) and were an excellent heuristic aid. They also assisted in the identification of diagnostic charac­ters. Canonical variate analysis, when based upon population samples, was the best tool for segregating morphological-geograph­ical groups that might be interpreted as taxa through its capacity to negate within- group variation, but only as long as its strict assumptions were met. Cluster analysis was confounded by the multiple patterns in T. speciosissima but was more useful in eluci­dating the simpler categorical pattern in T. oreades and T. speciosissima. Our results em­phasize the diverse nature of variation pat­terns below the species level and the im­portance of using a variety of multivariate techniques in exploratory analysis at this taxonomic level.

Acknowledgments

We are indebted to Dan Faith and Don Colless for discussions on methods and also to Don for assistance with running the analyses. David Cannatella, Jenny Chappill, Peter Cranston, Dan Faith, Toby Kellogg, Pauline Ladiges, Peter Linder, Chris Meacham, and David Morrison commented on various drafts of the manuscript. The project was partly supported by the Australian National Botanic Gardens. Helen Thomp­son, Peter Schumack, and Garry Richards gave tech­nical assistance. Bob Coveny, Ian Telford, Joan Taylor, and Geoff Butler assisted with the field work. Neda Plovanic provided the leaf tissue micrographs.

 

Appendix. Telopea populations sampled. Each population is represented by six individuals.
Study group	Abbrev.	Lat. (S)	Long. (E)	Locality and altitude
T. speciosissima	OK	34°03′	150°29′	New South Wales (NSW), ca. 5 km NNW of Oakdale, 550 m
	RN	34°04′	151°01′	NSW, Royal National Park, near Enga- dine Railway Station, 180 m
	WA	33°04′	151°21′	NSW, Watagan Mts., 10 km W of Cooran- bong, 410 m
	BW	33°32′	151°17′	NSW, Brisbane Water National Park, 1 km NNW of Warrah Trig, 150 m
	TR	35°05′	149°23′	NSW, Turpentine Range, 8 km from Tianjara Falls along road to Nowra, 400 m
	JB	35°09′	150°40′	NSW, Jervis Bay Territory, between Aus­tralian National Botanic Gardens annex and Lake Windemere, 45 m
	CF	34°37′	150°39′	NSW, Carrington Falls reserve, 7 km ESE of Robertson, 560 m
	KT	33°46′	150°23′	NSW, Blue Mts., Kings Tableland, 5 km S of Great Western Hwy., 820 m
	BL	33°33′	150°16′	NSW, Blue Mts., Bells Line, 5 km N of Mt. Victoria, 1,030 m
	NW	29°29′	152°15′	NSW, Gibraltar Range, 6.6 km along NW Fire Trail from Gwydir Hwy. (western end), 920 m
	BT	29°33′	152°16′	NSW, Gibraltar Range, 1 km S of Bound­ary Trig, 1,060 m
	TG	29°29′	152°19′	NSW, Gibraltar Range, The Granites, 1,060 m
	MH	29°32′	152°19′	NSW, Gibraltar Range, 4 km from ranger station along road to Mulligans Hut, 980 m
	TN	29°31′	152°23′	NSW, Gibraltar Range, The Needles lookout, 880 m
T. mongaensis/ T. oread es	GC	34°38′	150°25′	NSW, Gunrock Creek, 8 km along Mery- la Road from Moss Vale-Fitzroy Falls Road, 540 m
	MN	35°34′	149°55′	NSW, Mongarlowe River, 1.5 km N of Monga, 680 m
	MS	35°38′	149°55′	NSW, 6.7 km S of Monga along River Road, 740 m
	EP	37°21′	148°52′	Victoria, ascent to Errinundra Plateau, 1 km above Kanuka Creek, 1,000 m