The Effect of Combining Molecular and Morphological Data in Published Phylogenetic Analyses

Understanding the natural world is not possible without an extensive knowledge of organismal morphology. The encyclopedia of life (Wilson, 2003) would be an empty, sterile list without detailed information about the full range of morphological diversity. Morphology plays a pivotal role in our understanding of life cycles, geographical distributions, identification, conservation status, evolution, development, and species delimitation (Buzgo et al., 2004; Endress, 2000; Kaplan, 2001; Scotland et al., 2003a). Nonetheless, there remains considerable debate about the precise role of morphology in one particular area of biology—computer-based phylogenetic inference (Baker and Gatesy, 2002; Hillis and Wiens, 2000; Jenner, 2004; Scotland et al., 2003b; Smith and Turner, 2005; Wortley and Scotland, 2006; Wiens, 2004). One widespread contemporary use of morphological data in phylogeny reconstruction is in combined analysis with molecular sequence data (recent examples include Aagesen and Sanso, 2003; Cameron and Williams, 2003; Cohen et al., 2004; Dorchin et al., 2004; Hebsgaard et al., 2004; Lundberg and Bremer, 2003; Michelangeli et al., 2003; Near et al., 2003; Pedersen et al., 2003; Zrzavý, 2003). Despite the methodological, conceptual, and philosophical issues surrounding combined analyses of independent data partitions (Ballard, 1996; Ballard et al., 1998; Brower et al., 1996; Bull et al., 1993; De Queiroz et al., 1995; Farris, 2000; Huelsenbeck et al., 1996; Kluge, 1998; Levasseur and Lapointe, 2001; Nixon and Carpenter, 1996; Page, 1996; Pupko et al., 2002), the number of combined analyses conducted recently demonstrates that this is a popular way of integrating morphological data into phylogenetic analyses. Recent empirical studies have both questioned (e.g., Gaubert et al., 2005) and defended (e.g., Wahlberg et al., 2005) the role of morphology in phylogenetic analysis, with equal vigor. Although individual studies such as these can be highly informative and compelling, it is doubtful whether they can effectively reflect general trends. In 1998, Baker et al. investigated the contribution of morphological data sets to combined analyses with molecular data across a range of studies and found that the support contributed by morphological data was substantial, and often greater than that provided by molecular data. Since then, molecular sequence data sets have dramatically increased in size (from an average of 156 parsimony-informative characters across the 15 molecular data sets studied in Baker et al. (1998) to an average of 391 parsimony-informative characters across the 26 data sets analyzed in this paper), whereas morphological data sets have remained approximately the same (from an average of 40 parsimony-informative characters in Baker et al., 1998, to 47 parsimony-informative characters here). Since the first phylogenetic analyses using molecular data, more than 20 years ago (e.g., Curtis and Clegg, 1984; Hillis, 1987; Rogers and Bendick, 1985), the composition of molecular data sets has also changed. Molecular data sets can include amplified fragment length polymorphisms (AFLPs), restriction fragment length polymorphisms (RFLPs), protein and DNA sequences, coded gaps, and secondary chemicals. At the time of Baker et al.'s (1998) study, only two-thirds of “molecular” analyses considered comprised DNA sequence data. Today, the vast majority of analyses of molecular data comprise DNA sequences (all of the molecular data sets studied in this paper included DNA sequence, 24 out of 26 exclusively). Changes in the size and composition of molecular data sets used in phylogenetic analysis imply that the relative contribution of morphological data to combined analyses with molecular data may also have changed significantly in recent years. In the context of the ongoing debate surrounding molecular and morphological data in phylogenetic studies, there is a need for a broad survey to assess this development, and in particular to determine the effect of adding morphological data to molecular sequence data in combined analysis today. This has the potential to provide important general insights about current practice that individual, taxon-specific analyses do not. Two important parameters to consider are the degree of resolution provided by a data set and the level of support for the groups resolved, because improvements in resolution and support are frequently cited as the motivation for conducting combined analyses. In this paper we conduct a survey specifically to test the hypothesis that phylogenetic analyses of combined morphological and molecular sequence data sets yield significantly increased levels of resolution and support compared to analyses of molecular data alone. The Web of Knowledge citation database (http://wok.mimas.ac.uk/) is the most comprehensive database of peer-reviewed scientific literature. It was searched for all combined analyses of molecular sequence and morphological data published between 1998 and 2004. This period was chosen because studies published during recent years are most likely to reflect the current situation. The sample was then narrowed down according to a set of criteria designed to ensure the valid comparison of phylogenies within and between studies. Because support values obtained using different search methods (such as parsimony bootstrap and Bayesian posterior probability) are not directly comparable (Alfaro et al., 2003), comparisons were restricted to studies using the same method of analysis. Bayesian methods are becoming increasingly prevalent and are now being used to analyse combined data in a likelihood framework (e.g., Nylander et al., 2004); however, parsimony is still by far the most commonly used method for combined analyses of molecular and morphological data, and therefore parsimony analyses provide the largest possible set of studies for this investigation. Studies were excluded if they reported only the results of a combined phylogenetic analysis, because this did not allow for a direct comparison of the combined analysis with separate analyses of molecular or morphological data for the same taxa. Similarly, those studies that used different sets of taxa in separate and combined analyses (due to the availability of certain data sets for some taxa and not others) were excluded, since the phylogenies could not directly be compared. Finally, studies that used different methods, such as maximum likelihood for molecular data and parsimony for morphological and combined analyses, were excluded for the same reason. A wide range of metrics has been used to compare separate and combined analyses, both in single studies (e.g., Wahlberg et al., 2005) and broader surveys (Baker et al., 1998). These include the number of most parsimonious trees found, degree of resolution in consensus trees, tree statistics such as consistency and retention indices (Farris, 1989; Kluge and Farris, 1969) and, most often, measures of support such as jack-knife, bootstrap, and decay values (Bremer, 1994; Felsenstein, 1985). The most meaningful measures of the relative success of any phylogenetic analysis are accuracy (representation of true evolutionary relationships), resolution (number of branching events recovered), and support (confidence) for the relationships recovered (Bremer et al., 1999; Hillis, 1995; Huelsenbeck, 1995; Lecointre et al., 1993). The first of these, accuracy, cannot be determined without knowledge of the underlying phylogeny. This paper therefore concentrates on resolution and support. Resolution was measured using the consensus fork index (CFI), calculated by dividing the number of nodes found in a strict consensus tree by the total number of possible nodes (Colless, 1980, 1981). The support measure most frequently presented in the studies used here was the nonparametric bootstrap, which provides statistical support for individual clades. This was translated into support for each tree as a whole by counting the number of clades found to attain a value of at least 50% (the usual minimum reported in phylogenetic studies) and the number attaining at least 95% (often taken as indicating strong support for a clade) in bootstrap analyses. To conform, and allow comparison, with previous studies, the following statistics were also extracted from the studies: number of taxa, total number of characters, number of parsimony-informative characters, consistency index, and retention index. The percentage contribution of molecular and morphological data sets to the total number of parsimony-informative characters was also calculated. A two-tailed, paired t-test was used to assess whether there was a significant difference in resolution or support upon adding molecular or morphological data. The distribution of data points was near enough normal to justify the use of this parametric test. A paired version of the t-test was used because the sets of samples to be tested were not independent, comprising analyses of the same taxa with or without an additional data set. Because we could not be certain that the effect of adding data would be positive and wished also to detect any negative effect, we tested the nondirectional null hypothesis “adding a data set makes no difference to resolution and support.” Nondirectional hypotheses are most appropriately tested using a two-tailed t-test, which also has the benefit of being a more conservative test than the one-tailed and therefore providing more convincing evidence should a significant difference be found. Twenty-six data sets fulfilled the criteria described above (Table 1, Appendix 1). The data sets ranged in size from 10 to 103 terminal taxa and covered mosses and flowering plants, vertebrates and invertebrates. Molecular data sets ranged from 246 to 6972 aligned characters, of which 57 to 2093 were parsimony informative. Morphological data sets ranged from 12 to 119 characters in size, of which 12 to 119 were parsimony informative. On average, morphological data sets contributed 11% (ranging from 2% to 35%) and molecular data sets 89% (65% to 98%) of the parsimony-informative characters in the combined analysis. Details of phylogenies reported in source studies or calculated from phylogenies presented therein. PI, parsimony-informative;—, data not available. Details of phylogenies reported in source studies or calculated from phylogenies presented therein. PI, parsimony-informative;—, data not available. When morphological and molecular data were analyzed separately, morphological data almost always gave lower support and resolution than molecular data for the same taxa: fewer nodes were resolved and fewer clades found to be supported at both 50% and 95% bootstrap levels (Table 2). In one instance (Zrzavý, 2003), morphological data provided more resolution than molecular data, and in one other (Marvaldi et al., 2002), greater support. Resolution and support indices for molecular, morphological, and combined phylogenetic analysis. Resolution: absolute CFI value, where available, and effect of adding morphological or molecular data on CFI value of combined analysis. Support: absolute and percentage change in support on adding molecular or morphological data in combined analysis at >50% and >95% BS levels. BS, bootstrap; CFI, consensus fork index; Mol, molecular data; Mor, morphological data; Com, combined data;—, data not available; *, jack-knife reported instead of bootstrap value. Resolution and support indices for molecular, morphological, and combined phylogenetic analysis. Resolution: absolute CFI value, where available, and effect of adding morphological or molecular data on CFI value of combined analysis. Support: absolute and percentage change in support on adding molecular or morphological data in combined analysis at >50% and >95% BS levels. BS, bootstrap; CFI, consensus fork index; Mol, molecular data; Mor, morphological data; Com, combined data;—, data not available; *, jack-knife reported instead of bootstrap value. When morphology was added to molecular data in combined analysis, there was most often an increase in resolution compared to molecular data alone (Fig. 1A). Fourteen studies showed an increase in CFI, three a decrease and seven remained unchanged (these tended to be studies where complete resolution was attained both with and without morphological data). When molecular data was added to morphological data, the change in resolution was always positive (Fig. 1B). Using a two-tailed t-test, the addition of morphology to a molecular data set was found to make a significant difference to resolution (P = 0.01). The effect of adding molecular data to morphological data was also statistically significant (P < 0.001). Effect on resolution of including additional data sets in phylogenetic analysis, measured by CFI. A, Effect of adding morphological data, i.e., difference between molecular and combined analysis, 0.08 ± 0.15 (mean ± SE, t-test, t = 2.7, df 24, P = 0.01). B, Effect of adding molecular data, i.e., difference between morphological and combined analysis, 0.31 ± 0.19 (mean ± SE, t-test, t = 6.71, df 17, P < 0.001). Each point represents a single study, the vertical line within the shaded area is the mean and shaded area the 95% confidence interval. In terms of support, with the addition of morphological to molecular data, 10 studies showed an increase in the number of clades supported at the 50% level, three showed a decrease and four remained the same (Fig. 2A). At the 95% level, nine studies showed an increase, four a decrease, and four were unchanged (Fig. 3A). With the addition of molecular to morphological data, eight studies showed an increase in the number of clades supported at the 50% level and one showed a decrease (Fig. 2B). At the 95% level, all nine studies where bootstrap values were reported showed an increase (Fig. 3B). Using a two-tailed t-test, no significant difference was found between the level of support for molecular and combined phylogenies (P = 0.61 for the 50% level; P = 0.50 for 95%). By contrast, the difference between morphology alone and combined analysis (i.e., the effect of adding molecular data) was significant (P = 0.05 for the 50% level, P = 0.03 for 95%). When support values were scaled by the total number of nodes in the phylogeny, the same patterns were seen (graphs not shown). No relationship was found between patterns of support or resolution and the type of taxon (plant or animal, vertebrate or invertebrate) or level of analysis (intergeneric, interfamilial, interphylum, interkingdom). Effect on support, in terms of the number of nodes receiving a bootstrap value of at least 50%. A, Effect of adding morphological data, i.e., difference between molecular and combined analysis, 1.24 ± 2.36 (mean ± SE, t-test, t = 0.52, df 17, P = 0.61). B, Effect of adding molecular data, i.e., difference between morphological and combined analysis, 11.67 ± 5.06 (mean ± SE, t-test, t = 2.30, df 9, P = 0.05). Each point represents a single study, the vertical line within the shaded area is the mean and shaded area the 95% confidence interval. Effect on support, in terms of the number of nodes receiving a bootstrap value of at least 95%. A, Effect of adding morphological data, i.e., difference between molecular and combined analysis, 0.75 ± 1.08 (mean ± SE, t-test, t = 0.7, df 16, P = 0.5). B, Effect of adding molecular data, i.e., difference between morphological and combined analysis, 8.56 ± 3.36 (mean ± SE, t-test, t = 2.55, df 9, P = 0.03). Each point represents a single study, the vertical line within the shaded area is the mean and shaded area the 95% confidence interval. The aim of this survey was to analyze current phylogenetic practice in order to determine whether combined analyses of morphological and molecular data, as recommended in, and implemented by, recent papers, have yielded significantly increased levels of resolution and support compared to analyses of molecular data alone (e.g., Aagesen and Sanso, 2003; Cameron and Williams, 2003; Edgecombe et al., 2002; Giribet et al., 2002; Jenner, 2004; Lundberg and Bremer, 2003; Meier and Wiegmann, 2002; Schulmeister et al., 2002; Smith and Turner, 2005; Stach and Turbeville, 2002; Virtanen, 2003; Wahlberg and Nylin, 2003; Wiens, 2004). In the sample of studies examined, the inclusion of morphological data was found to have had a significant, positive effect upon resolution in combined analyses. For the same sample, the inclusion of morphological data did not have a statistically significant effect upon support as measured by the number of nodes attaining a given bootstrap value. The molecular data sets studied usually contained an order of magnitude more characters than the morphological ones with which they were combined (Table 1). However, morphological data sets contributed an appreciable proportion (on average 11%) of the parsimony-informative characters in each combined analysis. They therefore might be expected to have a significant effect on the results of phylogenetic analysis, even if not as large as the effect of molecular data. In terms of resolution, the results clearly indicate this positive effect of adding morphology in combined analyses. Resolution is an important aspect of phylogenetic analysis. However, in the absence of support resolution is of limited value and difficult to evaluate (Sanderson, 1995). In terms of support, although the result for this sample masks a great deal of variability, the effect of adding morphological data in combined analysis was not significant. Baker et al. (1998) conducted a similar survey of 15 phylogenetic analyses from 1989 to 1997, focusing on the relative contribution of morphological and molecular data within a combined analysis using partitioned Bremer support (PBS). There is a stark difference between their conclusion—that “the inclusion of morphology has a substantial influence on the support in a simultaneous [combined] analysis” (Baker et al., 1998:434)—and ours. One possible reason for this is the dramatic increase in size of molecular data sets since 1997, during which time they have more than doubled in terms of their average number of parsimony-informative characters, whereas the size of morphological data sets has remained approximately the same (from the data sets presented in these two studies). This means that, although morphology may have had the same absolute effect on analyses conducted between 1998 and 2004 (this study) as it did between 1989 and 1997 (Baker et al., 1998), the relative additional impact it has had when combined with molecular data is now smaller, because molecular data provide a larger amount of support and resolution. More recently, Wahlberg et al. (2005) suggested that combining molecular and morphological data sets can sometimes result in improvements in support as decay that are greater than the total support expected from the of the individual found that morphology did not have a statistically significant effect upon support, the number of supported clades in a majority of studies. This may be in some of the studies the added morphological data sets with the molecular ones to decrease support at some nodes and Sanso, 2003). within individual studies, a conservative change in resolution or support could different positive or negative on individual clades. It may also be for each individual study, the change in support values was and therefore could be to to However, this is not to that in any particular the impact of morphological data upon combined analysis be For in and of combined analyses recovered nodes in strict compared to only using molecular data alone. et al.'s combined molecular and morphological analysis of found dramatically more supported nodes at 50% bootstrap than molecular data alone In contrast, in et al.'s analysis of adding morphological data in a decrease from to nodes supported at 50% bootstrap, and in Meier and of it a decrease from to nodes at 95% These examples that of individual studies can be used to support almost any point of on the impact of morphological data in combined analysis. To a more complete of the role of morphological data, average impact across an sample of studies be compared. The number of studies that fulfilled our criteria the of in the However, the studies a range of taxa from et al., 2003) to flowering and Sanso, 2003; and 2001; and 2002; et al., 2001; 2001; et al., 2004; et al., 2002; and 2001; Michelangeli et al., 2003; and et al., to vertebrates et al., and a of from to there is any in the sample, it is most likely to the potential role of morphological data in phylogenetic analysis, because in groups where morphological data are likely to have to (such as morphological data sets not to be published at way to this would be to and the data sets using a search for between the results of individual studies. The here was by the of whether the inclusion of morphological data in combined phylogenetic analyses has in a significant difference to support and resolution in recent published this was investigated by published CFI and bootstrap One potential of any such survey in the between molecular and morphological data in of the studies the molecular data set comprised sequences from a number of different and was a data within which different phylogenetic may have been However, the morphological data sets were similar in comprising of separate data sets from and life and although the between molecular and morphological data sets is it did not result in a one type of data or have for focusing phylogenetic When and a phylogenetic study, there are a number of different to be considered and one These include which data sets provide the most resolved, and phylogeny, and in terms of time and are to Although this cannot the contribution of a particular data set for a given set of taxa, it can to the that adding a data set have a positive In it is likely that adding a morphological data set to an molecular one have a significant impact on resolution not on support, whereas a molecular data set is likely to both parameters relative to analysis of morphological data alone. It is important to that this upon resolution and support, as two of the of contemporary phylogenetic analysis, cannot reflect the of phylogenetic accuracy, which would need to be determined to provide a full of the of phylogenetic methods and data Resolution and support have if they to clades. The debate surrounding the role of morphological data still no of 2004; 2004; and Scotland, 2005; Scotland et al., 2003b; et al., 2003; Wiens, 2004; Wortley and Scotland, and of of phylogenetic analysis to to be used with morphological and combined data that there is still a for such 2001; Nylander et al., 2004). In most widespread in with molecular sequence data, recent studies have that morphology may be and et al., 2005) or provide support for clades et al., set out to test the hypothesis that phylogenetic analyses of combined morphological and molecular sequence data sets yield significantly increased levels of resolution and support compared to analyses of molecular data alone. results that the hypothesis has been for resolution not with to support. The would to Page, and an for on an of this and provided with statistical analysis. and retention indices reported in source studies index calculated from parsimony-informative characters not available. and retention indices reported in source studies index calculated from parsimony-informative characters not available.