We collected 40 previously published or ‘in press’ studies from a variety of sources that reported 72 phylogenetic hypotheses, hereafter referred to as source trees (STs), for genera-level matrix representation with parsimony (MRP) supertree analyses. These studies were screened for data duplication and quality prior to a more rigorous evaluation of their contained STs (see below). Three of these studies were excluded because of complete data duplication with subsequent, more inclusive studies by the same authors. In all instances, we preferred more recent molecular studies using cladistic methodologies (deemed higher quality) to older morphometric studies using either clustering algorithms or no formal analysis; this resulted in the exclusion of three additional studies. With one exception (see below), we limited our selection to studies published after 1966.
We used the ‘garbage in, garbage out’ protocol of Bininda-Emonds et al. (2004) as our criteria for ST selection. We selected the most comprehensive ST presented in each study; the only exceptions to this occurred when multiple STs were reported and there was data duplication involving the most comprehensive ST. Independence among STs was assessed conservatively, STs were excluded on the basis of relatively minor data duplication among studies. We identified 17 independent STs (Holman, 1961; Crowe, 1978; Gutiérrez et al., 1983; Helm-Bychowski & Wilson, 1986; Sibley & Ahlquist, 1990; Randi et al., 1991; Zink & Blackwell, 1998; Armstrong et al., 2001; Birks & Edwards, 2002; Dimcheff et al., 2002; Drovetski, 2002; Pereira et al., 2002; Sorenson et al., 2003; Chubb, 2004; Nishibori et al., 2004; Pereira & Baker, 2004; Crowe et al., 2006). In addition, we also identified two nucleotide sequences, mitochondrial control region and cytochrome b, that were recycled extensively across a further 16 studies; this resulted in two sets of nonindependent STs, one for each of these mitochondrial sequences. For these two sets, we followed the recommendation of Bininda-Emonds et al. (2004), and conducted an interim ‘mini supertree’ analysis on all of the available STs (control region: Fumihito et al., 1995; Kimball et al., 1997, 1999; Lucchini et al., 2001; Drovetski, 2002; and cytochrome b: Kornegay et al., 1993; Ellsworth et al., 1996; Kimball et al., 1997, 1999, 2001; Bloomer & Crowe, 1998; Munechika et al., 1999; Gutiérrez et al., 2000; Armstrong et al., 2001; Bush & Strobeck, 2003; Zhan et al., 2003; Shibusawa et al., 2004; Wen et al., 2005) and included the resulting ‘mini supertrees’ as STs in the main supertree analyses. Because of insufficient overlap among taxa, there were four, rather than two, resulting ‘mini supertrees’. In an attempt to balance the quality of the included STs with taxonomic coverage, we included one osteological taxonomy of the Odontophoridae (Holman, 1961); it was highly congruent with two less complete STs from molecular studies that address relationships among genera (Zink & Blackwell, 1998) and also with other families (Gutiérrez et al., 1983). Prior to coding the STs for MRP, nodes with published bootstrap support values < 50% were collapsed.
Wilkinson et al. (2005) recently compared the properties of 14 supertree methods and demonstrated systematic biases in the way conflicts are resolved among STs: binary coding tends to resolve conflicts in favour of unbalanced STs, whereas additive binary coding tends to resolve conflicts in favour of balanced STs (Wilkinson et al., 2005). We therefore used both coding methods [in Radcon (Thorely & Page, 2000)] to generate matrix representation of STs: binary coding (Baum, 1992; Ragan, 1992) and Purvis’ modification of this method (additive binary coding; Purvis, 1995), which eliminates redundancy inherent to binary coding (Purvis, 1995). Our MRP ‘mini supertree’ and main supertree analyses were conducted using the parsimony ratchet (Nixon, 1999), which increases the efficiency of heuristic searches for candidate trees, as implemented by pauprat (Sikes & Lewis, 2001) in paup* (Swofford, 2002). STs were weighted uniformly, that is to say that the initial weight of all characters was set to 1, which is the default setting of pauprat (Sikes & Lewis, 2001). For each of the MRP matrices we ran 30 independent searches consisting of 200 iterations, with 15% of the characters perturbed at each iteration. After the 30 independent searches we extracted the set of optimal candidate trees (shortest length) and removed duplicates trees. The ratchet searches returned 893 and 338 unique optimal candidate trees for the binary coded and the additive binary coded STs respectively. We generated the 50% majority rule consensus supertrees from these two sets of unique optimal candidate trees in paup*.
The 50% majority rule consensus supertrees from the two coding methods were highly congruent and both contained a large polytomy associated with the most recent radiation, the Phasianidae. Although Numididae was consistently placed as sister to Odontophoridae and Phasianidae in both 50% majority rule consensus supertrees, the support for this node was relatively weak (binary coding: 58% and additive binary coding: 54%). To account for this family-level uncertainty, we resolved both candidate supertrees (binary and additive binary) in each of two ways: with Odontophoridae as sister to Numididae and Phasianidae, and with Numididae as sister to Odontophoridae and Phasianidae. Although highly congruent, minor discrepancies did exist between the supertree topologies. So, we consulted the underlying STs, and, in all instances, the supertree constructed from additive binary coded STs matched the STs better. Because of the lack of redundancy in coding, additive binary coding is arguably a better representation of the STs than binary coding (Purvis, 1995). We performed the comparative analyses on both candidate supertrees and the results did not differ, so we only present results from analyses based on the supertree constructed from additive binary coded STs.
We attempted to provide resolution to the polytomy encompassing the Phasianidae, as follows. First, we assumed monophyly for each unambiguous branch of the polytomy, and resolved discrepancies between the supertree topologies conservatively. As a result, Alectoris, Pternistis, Rollulus, (Xenoperdix with Arborophila), and (Coturnix with Margaroperdix) were each considered as additional branches. Because of inconsistent or ambiguous affinities among the STs, Meleagris, Perdix, Tragopan, (Afropavo with Pavo), and (Rheinardia with Argusianus) were also considered as separate monophyletic branches. This increased the size of the polytomy to 18 branches. We then obtained sequence data from GenBank for six genes or introns: three mitochondrial (cytochrome b, ND2 and 12S rDNA) and three nuclear (ovomucoid intron G, WPG pseudogene and zona pellucida C). For each sequence, we generated a consensus sequence for each of the 18 branches; this facilitated a more thorough search of tree space at the level of the polytomy. We made a global alignment for all of the sequences for each gene in kPrank (Higgins et al., 2005; Loytynoja & Goldman, 2005), using a guide tree imported from ClustalX (Thompson et al., 1997). From this global alignment, we then generated a consensus sequences for each branch of the polytomy. Preliminary work revealed that an 80% threshold for representation in the consensus sequence resulted in > 99% identity to ancestral sequences inferred using Bayesian methods (results not shown). Ambiguous sites in the consensus sequences, i.e. when no single nucleotide met the 80% threshold, were represented by the corresponding IUB DNA symbol (Cornishbowden, 1986). We used mrmodeltest 2.2 (Nylander, 2004) to determine the best model of nucleotide evolution for each partition. Based on a concatenated partitioned alignment of all six sequences and with Numida meleagris as the outgroup, we inferred a phylogeny for these 18 clades using mrbayes 3.1.2 (Huelsenbeck & Ronquist, 2001; Ronquist & Huelsenbeck, 2003). We ran four chains simultaneously in a Metropolis-coupled MCMC search of tree space in two independent iterations of 10 million generations, using default settings. After a burn in of 2.5 million generations, we sampled trees every 1000 generations; this resulted in 7500 candidate trees from which we constructed a 50% majority rule consensus.
We grafted the resulting genera-level phylogeny of the Phasianidae onto the supertree constructed from the additive binary coded STs (Fig. 1c), producing a composite supertree phylogeny (Fig. 1b). From the composite supertree phylogeny, we also generated a fully resolved supertree topology based on our ‘best informed guess’ (BIG) of the remaining unresolved relationships (Fig. 1a). Finally, we allowed for the same family-level uncertainty in the relationships among Numididae, Odontophoridae and Phasianidae, resolving both the composite and BIG phylogenies in each of the two possible ways.
Figure 1. Genera-level supertrees for the Galliform birds based on additive binary coding of source trees: fully resolved ‘best informed guess’ topology (a), composite supertree phylogeny (b), and 50% majority rule consensus supertree (c). Each of the three topologies was also resolved so that the Odontophoridae was sister to the Numididae and the Phasianidae, by switching the placements of the Odontophoridae and the Numididae at the node marked with an asterisk. Figure 1 continued.
Bivariate contrast analyses on the relationships between all variables were performed both using raw data and Model 1 regression through the origin using PICs (Felsenstein, 1985). For the analysis of the relationship between egg size and clutch size we also controlled for female body mass by including all three variables in a multiple regression including the 33 genera for which we had data for all three variables. All branch lengths were set equal to one for the PIC analyses and polytomies were resolved to zero-length branch lengths for the analyses based on the consensus trees (the BIG trees were fully resolved). We then tested for correlations between contrasts and their SD to check whether branch length transformations were needed to avoid type I error (Diaz-Uriarte & Garland, 1998). As we did not detect any relationships between absolute values for the contrasts and their SD for any of the analyses, no transformations were needed. As our supertree analyses generated eight different trees [two fully resolved BIG trees based on the additive binary coded supertree (Fig. 1a), two composite super trees with the additive binary supertree (Fig. 1b), two supertrees based on additive binary coding (Fig. 1c) and two supertrees based on binary coding (not presented here)], we performed all analyses on all eight trees to investigate if our analyses were sensitive to which tree we used. As the results were similar (i.e. no results changed from significance to nonsignificance or vice versa) with only two exceptions (see Results) regardless of which tree was used, we only report the results from the phylogenetic analyses based on the BIG tree. For the analysis of the relationship between SSD and MS, the low sample size did not allow for a phylogenetically independent matched pairs analysis (see Harvey & Pagel, 1991) as it only yielded very few matched pairs. Instead, we used a normal ancova with SSD as the dependent continuous variable and MS (monogamy or polygamy) as a categorical independent variable and mean body mass (sexes pooled) as a covariate for the raw data at the genus level. This allowed us to estimate the relationship between these variables while controlling for body mass. All data were log10 transformed prior to raw data analyses and before calculations of independent contrasts to ensure normality. Independent contrasts were calculated using the pdap: pdtree module within Mesquite (Midford et al., 2002; Maddison & Maddison, 2004).
For the analyses of directional evolution of SSD in relation to female life history traits, we used discrete (4.0) (Pagel, 1994, 1997). This program is based on a Markov model for trait evolution and allows for estimation of ancestral states, investigation of correlated evolution between two traits, investigation of the directionality of changes in traits, and how changes in one trait precedes changes in another trait. A likelihood ratio test is used to compare the maximum likelihood fits of a model that only allows for independent evolution of two traits to a model that allows for dependent evolution between two traits. The likelihood ratio test statistic is χ2 distributed with d.f. = 4 for the comparison between the independent and the dependent model (Pagel, 1997).
One can investigate the pattern of co-evolution for a pair of traits through investigation of the relative magnitudes of their joint transition rates (see Fig. 2). So, for example, one can ask which path (upper vs. lower) from ‘low-small’ to ‘high-large’ is most likely on the data. One approach (see e.g. Cézilly et al., 2000; Kolm et al., 2006) is to ask which of the eight joint rates (represented by the arrows in Fig. 2) are indistinguishable from zero, using a likelihood ratio test of the nested models (focal rate = 0) vs. (focal rate = ML estimate). Rates that are indistinguishable from zero suggest that these paths are unlikely. If one can identify the ancestral states (i.e. which box in Fig. 2 is ancestral), a full description of the likely evolutionary paths through time is possible.
Figure 2. Flow diagram of the possible transitions of a hypothetical model of dependent correlated evolution of two traits (a and b) that can take two stages each (low or high; small or large). Each potential transition is given by qab.
Download figure to PowerPoint
Following transformation of female body mass, egg size and clutch size into discrete characters, body mass and egg size showed a perfect relationship (i.e. for the 34 genera for which we had data on these traits, all genera with larger than average females had larger than average eggs and vice versa). As discrete assumes no simultaneous changes in two traits, we could thus not disentangle the directional evolution of these two traits in relation to each other. Moreover, this perfect correlation also meant that we only performed a discrete analysis on SSD in relation to female body mass. As the bivariate contrast analyses did not show any relationship between female body mass and clutch size, we did not perform any discrete analysis for this combination of traits. Further, although it would be very interesting to perform a discrete analysis on clutch size in relation to egg size to disentangle the evolution of these two traits for Galliformes (as done for cichlid fishes by Kolm et al., 2006), discrete requires larger sample sizes (N. Kolm, personal observation) than we had for robust tests. To investigate whether the relative frequency of trait values might affect our results (Nosil & Mooers, 2005) for our data, we randomized the distribution of female body mass and SSD (from Fig. 5) across one of our BIG trees 100 times and then performed directional discrete analyses to investigate how often chance alone would yield the same result as from our discrete analyses based on the actual transitions in the tree. Only one of our 100 randomized datasets produced the same significant set of transitions, suggesting this was not a problem.
Figure 5. Current combinations of states of Female body mass (= Egg size) and sexual size dimorphism Galliform genera.
Download figure to PowerPoint