Patterns of morphological disparity yield important insight into the causes of diversification and adaptive radiation in East African cichlids. However, comparisons of cichlid disparity have often failed to consider the effects that differing clade ages or stochasticity may have on disparity before making interpretations. Here, a model of branching morphological evolution allows assessment of the relative contributions of differing turnover and morphological change rates, clade ages, and stochastic variation to the observed patterns of disparity in four endemic tribes of Lake Tanganyika cichlids. Simulations compare the likelihood of generating the observed disparity of the four tribes using 200-parameter combinations and four model conditioning variations, which allows inference of evolutionary rate differences among clades. The model is generally robust to model conditioning, the approach to data analysis, and model assumptions. Disparity differences among the first three cichlid tribes, Ectodini, Lamprologini, and Tropheini, can be explained entirely by stochasticity and age, whereas the fourth tribe, Cyprichromini, has likely experienced lower rates of turnover and morphological change. This rate difference is likely related to the low dietary diversity of the Cyprichromini. These results highlight the importance of considering both clade age and stochastic variation when interpreting morphological diversity and evolutionary processes.


Cichlids are a highly diverse fish family that constitutes approximately 4% of extant vertebrate species (Salzburger and Meyer 2004; Hulsey 2009) and are one of the most rapidly diversifying vertebrate lineages (Genner et al. 2007; Moyle et al. 2009). This species richness is largely contained in the endemic species flocks of the East African Great Lakes (Farias et al. 2000). Standing diversity and disparity in cichlids have generally been explained using ecological opportunity and/or key innovations in trophic and reproductive adaptations (Albertson et al. 1999, Seehausen 2006, and references therein). Studies of disparity patterns within and between clades and ecological groups have helped to establish the adaptive nature of the cichlid radiation (Chakrabarty 2005; Clabaut et al. 2007; Hulsey 2009) and the processes involved in the ecological diversification within clades (Albertson et al. 1999; Streelman and Danley 2003; Clabaut et al. 2007).

Several broad patterns have emerged from these studies of cichlid disparity. Lake Tanganyika cichlids are more morphologically disparate than Lake Malawi cichlids (Chakrabarty 2005), and some tribes within Lake Tanganyika, such as Ectodini and, to a lesser extent, Lamprologini, have higher morphological disparity than other Great Lake tribes (Chakrabarty 2005). However, the biological significance of such differences is unclear because the tribes are of different ages (Genner et al. 2007; Day et al. 2008) and disparity increases nearly linearly with time under constant diversification and morphological evolution rates (Foote 1996; Pie and Weitz 2005; Ricklefs 2006a). Thus, older clades should be more disparate than younger clades. Furthermore, stochastic variation in morphological divergence or species turnover can produce different disparity among clades even if age and rates of origination, extinction, and morphological change are the same.

Because differences in disparity can be generated by differences in evolutionary rates, age, or stochastic variation, comparing turnover and morphological change rates rather than standing disparity is likely to yield insight into the most evolutionarily meaningful differences among groups. However, previous analyses of evolutionary rates in East African cichlids have examined rates of diversification and genetic evolution, but not rates of morphological evolution. Although species richness and net diversification rates are very high in the East African lacustrine cichlids, genetic diversification rates in this clade are significantly lower than in Neotropical cichlids (Farias et al. 1999, 2000). Coupled high speciation and extinction rates in the African cichlids could account for their low genetic diversity (Farias et al. 1999). Net diversification rates are also highly variable among the different species flocks of East African cichlids; the diversification rates of the Lake Victoria and Lake Malawi flock are approximately six times greater than those of the cichlid tribes of Lake Tanganyika (Day et al. 2008).

Evolutionary rates have not been studied in detail partially because of the ongoing difficulty in resolving the biogeographic origin of East African cichlids (Farias et al. 2000; Salzburger et al. 2002, 2004; Salzburger and Meyer 2004; Koblmuller et al. 2008). Further, the scarcity of cichlid fossils (Murray 2001a, 2001b; Genner et al. 2007), incomplete lineage sorting of many clades (Seehausen 2006), and considerable morphological convergence (Clabaut et al. 2007; Hulsey 2009) have made the resolution of phylogenetic trees and time calibration of molecular clocks very difficult. Recently published dates for the origin of many of the tribes of East African cichlids (Genner et al. 2007; Day et al. 2008) make analysis of evolutionary rates a possibility. Because calibration of molecular clocks in East African cichlids is still a controversial topic, model designs that are dependent on relative rather than absolute clade ages will provide the most robust conclusions.


Many studies reconstruct taxonomic diversification from modern phylogenies through comparisons of simple null models of evolution with observed data on disparity patterns (Yule 1924; Raup et al. 1973; Hey 1992; Nee 2001, 2004, 2006; Sims and McConway 2003; McConway and Sims 2004; Ricklefs 2006a; Rabosky 2009, 2010). These studies allow the calculation of speciation and extinction rates, inferences about patterns of diversification, comparisons of diversification patterns among clades, and identification of changes in diversification rates over time (Zink and Slowinski 1995; Barraclough and Nee 2001; Ricklefs 2006b; Bininda-Edmonds et al. 2007; Rabosky and Lovette 2008; Alfaro et al. 2009; Crisp and Cook 2009, Quental and Marshall 2009). One important result of these studies is that clade size and age are independent (Magallon and Sanderson 2001; Ricklefs 2006b; Ricklefs et al. 2007; Rabosky 2009), which indicates that speciation and extinction are relatively balanced over time (Ricklefs 2007).

Diversification models can also incorporate morphology to determine rates of morphological evolution (e.g., Foote 1993; Harmon et al. 2003; Agrawal et al. 2009; Hulsey et al. 2010), often by comparing the observed disparity of clades to that produced by a null model of Brownian motion. Brownian motion allows traits to increase or decrease in any given instant with a magnitude and direction that is independent of the current state of the trait (O’Meara et al. 2006). In a Brownian model, older clades or clades with early speciation should have higher disparity (Ricklefs 2006a).

Sidlauskas (2007) developed a simulation-based approach for comparing the effects on extant disparity of morphological change rate, diversification rate, and timing of the onset of diversification without using a detailed phylogeny. However, although his model suggested differences in morphological change rates between two sister clades of South American fish, it revealed a surprising lack of sensitivity to turnover rates (Sidlauskas 2007). Simulation studies suggest that higher turnover rates should result in lower variance in time-dependent models of Brownian Motion due to the decreased average node age (O’Meara et al. 2006; Ricklefs 2006a), so conditioning on the number of extant species actually observed may have biased the results against detection of turnover rates between clades (Sidlauskas 2007).

This study seeks to test the sensitivity of the Sidlauskas (2007) model to its underlying assumptions and to further assess the importance of turnover rate in determining standing disparity. Then, an adaptation of the approach used by Sidlauskas (2007) is used to determine whether differences in evolutionary rates are likely to account for the disparity patterns of four tribes of cichlids from Lake Tanganyika.

The tribes of Lake Tanganyika provide an optimal system for two reasons. First, unlike the Lake Malawi and Lake Victoria cichlid flocks, the Lake Tanganyika cichlid flock is polyphyletic, and contains 16 (Koblmuller et al. 2008) or 17 (Takahashi 2003) tribes of known ages (Genner et al. 2007; Day et al. 2008), many of which have now been conclusively demonstrated to be monophyletic (Salzburger et al. 2002, 2005; Takahashi 2003; Clabaut et al. 2005; Koblmuller et al. 2005). Thus, within Lake Tanganyika, there are well-established monophyletic clades of the same taxonomic rank among which rates may be compared without depending on phylogenetic resolution at the genus level.

Second, feeding preferences strongly predict Tanganyikan cichlid morphology, as do depth preferences, although to a lesser extent (Ruber and Adams 2001; Clabaut et al. 2007). In particular, postcranial morphology tends to correlate with water depth preferences whereas head and mouth shape tend to correlate with feeding preferences (Clabaut et al. 2007). Thus, rates of morphological change would be expected to be higher in tribes that have diversified to have a wide range of trophic strategies and depth preferences, whereas groups that specialize on particular niches might exhibit lower rates of morphological evolution. These factors make rate comparisons within these tribes both biologically informative and feasible using available data.



Many measures of disparity have been proposed, including total variance, total range, mean distance, number of unique pairwise character combinations, principal coordinate analysis volume, average pairwise dissimilarity, and participation ratio (Ciampaglio et al. 2001). Because variance does not depend on the number of species in a clade or the sample size, it is preferable as a measure of disparity over morphospace volume and range (Foote 1997; Ricklefs 2006a). Therefore, total multidimensional variance (Van Valen 1974; Foote 1993) is used as the measure of disparity for this study.

Morphological data were obtained from FishBase for 80 species in four tribes of Lake Tanganyika endemic cichlids (Froese and Pauly 2010, see Table 1). These tribes were selected because they are the monophyletic clades of Tanganyikan cichlids with morphological data available for more than five species. Furthermore, these species include the tribes of Tanganyikan cichlids with the highest disparity (Ectodini, followed by Lamprologini), the lowest disparity (Cyprichromini) (Chakrabarty 2005), and the highest estimated diversification rates (Tropheini followed by Lamprologini) (Day et al. 2008), making these tribes interesting for rate comparisons. The species were divided by tribe using the morphological phylogeny of Takahashi (2003), which agrees well with molecular phylogenies (Koblmuller et al. 2008). The characters include standard length (% total length), preanal length (% total length), predorsal length (% total length), prepelvic length (% total length), prepectoral length (% total length), body depth (% total length), head length (% total length), eye diameter (% head length), preorbital length (% head length), and the aspect ratio (calculated as height squared divided by surface area) of the caudal fin. The data are based on single specimens for all but two species, for which the data are based on two specimens, and are available online via DRYAD (doi:10.5061/dryad.1ps0t).

Table 1.  Tribe and sample data. Tribe data include the estimated ages using both Gondwanaland and fossil calibrations and the total number of species of the tribes discussed in this study (Genner et al., 2007). The number of species for which data were available is also provided. Note that the ratios between ages of tribes are almost identical for the two sets of dates, although the absolute age estimates are very different. Age estimates from a second source are also included (Day et al. 2008).
TribeAge (Gondwanaland)Age (Fossil)Age (Day)Number of spp.Number of samples
Lamprologini16.01 7.296.679–8344
Tropheini 6.76 3.023.419–3611
Cyprichromini10.22 4.604.810 6
Limnochromini10.83 4.884.313 5
Perrisodini 9.24 4.153.1 8 5

To test whether analyses of these data were biased by sampling effects, the traditional morphometric FishBase data were also compared with the geometric morphometric data used by Chakrabarty (2005) using the programs CoordGen6h and DisparityBox6i (Webster and Sheets 2010). The Chakrabarty dataset includes 18 landmarks, 10 of which are cranial landmarks. Cranial morphology is more variable than postcranial morphology in cichlids (Clabaut et al. 2007), and so this dataset ensures that the most variable morphological features are analyzed. The two datasets yield highly consistent results with respect to the relative disparity of the four tribes included in this study. Given that the Chakrabarty dataset is a different kind of data, has larger sample sizes per species (over nine specimens per species on average), and has more information on cranial morphology than the FishBase data, the consistency between the Chakrabarty and FishBase datasets suggests that the results described in this study are not an artifact of sampling. Furthermore, the analyses described in the following sections were repeated using the Chakrabarty data and produced very similar results to analysis of the FishBase data. Further details are provided in Appendix S1.

Principal components analysis of the variance–covariance matrix of the data was conducted using the eigen function in R. The first four principal components represented 46.5%, 21.6%, 11.3%, and 7.8% of the total variance, for a total of 87.2% of the total morphological variance (Table 2). The first principal component correlates most strongly and negatively with prepelvic length, eye diameter, and preorbital length, the second correlates negatively with predorsal length and prepectoral length, the third correlates positively with preanal length and negatively with head length, and the fourth correlates positively with standard length and negatively with body depth. The eigenvalues from this analysis were then used to scale the change on each principal component axis in the simulations described below.

Table 2.  Variance data. Empirical total multidimensional variance, calculated using the Froese and Pauly (2010) data, standard errors (SE) as estimated from 1000 bootstrap repetitions, the variance of the first four PC axes, and the standard errors of the first four PC axes are provided.
TribeTotal VarianceSEPC1–4 VarianceSE PC Variance
Tropheini109.3524.50 96.5323.84
Cyprichromini 33.3910.07 21.36 7.70


The simulations are based on the description of the Matlab script “morphtreegen” (Sidlauskas 2007). The simulations presented in the main text use a speciation-dependent model in which morphological evolution occurs at speciation events. This is most similar to a punctuated equilibrium model of evolution (Ricklefs 2006a; Agrawal et al. 2009; Bokma 2010). The alternative to a speciation-dependent model is a time-dependent model, in which morphological evolution occurs in each extant species during each time step. Time-dependent models are more representative of morphological evolution by anagenesis (Ricklefs 2006a; Agrawal et al. 2009; Bokma 2010). The results from the two types of models were compared to assess the robustness of the model results to assumptions about the processes involved in morphological evolution. The parameters for this model are speciation rate (p), extinction rate (q), time (t, in 100,000 year time steps), rate of morphological change (r), a vector of the eigenvalues of the first four PC axes (λ), and for some model variations, the minimum and maximum number of species. See Figure 1 for a schematic diagram of the simulations and analysis.

Figure 1.

Schematic of simulations and analysis. The schematic represents a complete set of runs of any given model variation, meaning that this approach was repeated using the No Condition, Extant, Min, or Min–Max versions. In each trial, the simulation runs for a number of times steps equivalent to the age of the simulated tribe (each time step is equivalent to 100,000 years using the Gondwanaland calibration of the Lake Tanganyika cichlid ages from Genner et al. 2007). Success is defined as a trial in which the simulated variance equals either the empirical PC1–PC4 variance or total variance (Appendix S4). The “Same Age” analysis compares tribes when the trials are run for the same number of time steps to determine the relative contribution of clade age to standing disparity patterns (Appendix S6). See the main text for additional details on how each step is conducted.

The model is a discrete-time, time-homogeneous birth–death model with a speciation-dependent morphological evolution component. During each time step, each extant species can persist unchanged, give rise to a new species, or go extinct, using the assumption that only one event can occur in any given time step. For each new species, the program selects a random number from a normal distribution with a mean of 0 and a standard deviation of 1 for each of the four PC axes, and then multiplies these random numbers by (rλ).5, where r is the rate of morphological change and λ is the eigenvalue of the corresponding PC axis calculated from the empirical data. Scaling the change to λ therefore causes greater change to occur along axes with greater empirical disparity. If either branching or extinction occurs, the standing diversity count is adjusted accordingly. If all lineages are extinct, the simulation terminates.

After the last time step, or after all lineages are extinct, the final standing diversity is compared with the selected condition for the model. If the model has no condition, then all simulations are counted, including those in which all species went extinct. If the model is conditioned on the survival of the clade, then only trials in which the simulated tribe has at least one extant species after the last time step are counted. Likewise, for simulations conditioned on the minimum or minimum and maximum number of species, then only simulations producing species richness within the specified bounds are recorded and counted toward the number of trials. These bounds are the minimum and maximum species diversities for all four tribes; thus, for this study, the minimum species number is 10 (the species diversity of Cyprichromini) and the maximum species number is 83 (the species diversity of Lamprologini) for all simulations (Table 1). For convenience, models using these four conditions will be referred to as the “No Condition,”“Extant,”“Min,” and “Min–Max” versions for the remainder of the article. An additional version, in which the species richness is confined to the bounds of only the two species compared in any one pairwise analysis, was also included to verify that the results were consistent with those given by other model variations (Appendix S2).

If the simulation meets all model criteria, then the program calculates the variance of the clade. The model output is a table of the final morphological variance for each simulation. Once the specified number of trials (1000 in all cases for this study) has been completed, the simulation moves to the next set of parameter values and begins again. Code for the model and subsequent analyses was written in R and is available online via DRYAD (doi:10.5061/dryad.1ps0t).


Once the simulations for all parameter values have run, the program counts the total number of simulations producing variance within one standard error of the empirical PC1–PC4 variance of the tribe. The probability of generating the empirical disparity using a given set of origination and extinction rates or a given rate of morphological change is assumed to be the number of trials producing the observed disparity divided by the total number of simulations with that parameter value. For each value of r, there were 10,000 total simulations (1000 trials at each of ten p and q values) and for each value of p and q there were 20,000 trials (1000 trials at each of twenty r values). The values used for p, q, and r are discussed in the Parameter Estimation section below.

After calculating the probabilities of achieving the observed disparity of each tribe for each of the values of r, p, and q, matrices were constructed to compare each combination of tribes based on the approach of Sidlauskas (2007). Twenty by twenty cell probability matrices were used to compare values of r, and ten by ten cell matrices were used to compare values of p and q, between each possible pair of tribes. Each cell in the matrix represents the probability of generating the empirical disparity of both tribes, and is therefore the product of the total probability of generating each individual tribe given the relative parameter values. For example, in an r probability matrix, cell (i, j) is equal to


where ML stands for the morphological diversity (PC1–PC4 variance) of the tribe with lower disparity and MH stands for that of the tribe with higher disparity.

Because this study, unlike that of Sidlauskas (2007), uses tribes with very different ages and numbers of species, the prior probabilities for turnover rates are not uniform. Therefore, the value in each cell of each pq probability matrix was also multiplied by the prior probability of achieving the current number of species for each clade given the turnover rate. Thus, in a pq probability matrix, cell (i, j) is given by the equation


where nL and nH are the observed species richness for the clades with low and high morphological diversity, respectively. The prior probabilities are calculated using the equation


where Pn,t is the probability of n species at time t, and α and β are given by the equations of Raup (1985):


For a pure-birth process, the prior probabilities are given by the equation (Aldous 2001)


Thus, each cell in the probability matrix represents the probability of the observations (observed disparity of the two tribes) given the corresponding model and parameter values. Thus, these probabilities are the likelihoods of the model given the disparity data:


where c is a proportionality constant. Because the entire probability matrix is multiplied by the constant c, the constant does not affect the differences in likelihoods and thus these probabilities can be treated as likelihoods. Thus, the probability matrix allows the calculation of the total likelihood of each of the two opposing hypotheses that rates of morphological change or turnover of the more disparate clade were either higher or were less than or equal to the evolutionary rates in the clade with lower disparity. The likelihood of higher evolutionary rates in the more disparate tribe is the sum of all cells above the diagonal of the matrix, whereas the likelihood of equal or lower rates in the high disparity clade is the sum of the cells in and below the diagonal. Sidlauskas (2007) found that the lower turnover rates generated higher empirical disparity. However, initial simulations indicate that in the speciation-dependent model used for this study, all four model conditioning variations yielded statistically higher disparity given higher turnover rates because speciation is the only available way to change morphology (t-test, P < 0.001 in all variations). Results were very similar from a time-dependent model (Appendix S3). Thus, for this study, high turnover rates are considered to result in higher disparity, and the likelihood that the difference in disparity resulted from higher turnover rates in the clade with higher disparity is the sum of the cells above the diagonal. Note that this approach compares the likelihoods for differences in evolutionary rates, but does not provide an estimate of the values of p, q, and r.

Support for each of these hypotheses is then calculated as the natural log of the likelihood. Then, ΔAICC and Akaike weights are used to compare alternative models. Akaike weights compare models on equal footing with a penalty for additional complexity. This approach provides more power and is more biologically reasonable than rejecting a null model with 95% confidence (Hunt et al. 2008).

In situations where rate differences are not well supported, differences in disparity may be due to differences in clade age or to stochastic differences in evolutionary processes. An additional set of analyses in which the tribes are treated as if they were all the same age is designed to test the relative contribution of clade age in determining differences in disparity.


This model requires the specification of the values of p, q, and r to be tested. To determine the range of realistic possibilities for p and q, the diversification rates for all four Lake Tanganyikan tribes in the analysis were estimated. These parameters were estimated using the clade age and species richness of each clade and the method-of-moments estimator of Magallon and Sanderson (2001) for values of ɛ ranging from 0 to 0.9 in 0.1 increments using the function rate.estimate in the R package GEIGER (Harmon et al. 2008), where ɛ= p/q. Although the large number of ages and rates made it impractical to use every combination of p and q provided by this approach, the range of origination and extinction rates used here spans the full range of rates calculated for all clades. For the simulations, the speciation rate, p, ranges from 0.01 to 0.21 in increments of 0.02, and the extinction rate q is set to 90% of the p value, giving a total of 10 different p and q combinations. Magallon and Sanderson (2001) suggest that higher values of ɛ than 0.9 are not realistic, and high coupled speciation and extinction rates have been suggested for East African cichlids (Farias et al. 1999). However, other researchers have indicated that adaptive radiations may have low or even zero extinction rates, especially early in diversification (Purvis 2008), and recent molecular work has found no evidence for nonzero extinction rates (Day et al. 2008). Thus, a second set of simulations was run using pure-birth assumptions (zero extinction) to compare the results with those from a birth–death model. The values of p ranged from 0.01to 0.06 in increments of 0.005, to include the full range of maximum likelihood estimates for p under zero extinction.

Preliminary simulations were used to determine a range of rates of morphological change that produce the observed variance of the four tribes examined in this study. Based on these results, r values were selected to include these values and extend to values that were somewhat too high or too low to ensure that the full range of reasonable possibilities was examined. The values used in this study span r values from 0.005 to 0.15 in increments of 0.075.

The ages and species richness of the four tribes described in this study are provided in Table 1. The ages of the divergence of the cichlids are calibrated on the interpreted Gondwanaland origination of cichlids, which is favored by Genner et al. (2007) over calibration using fossils of distant relatives of the East African cichlids. However, for this study, it is the relative ages of the tribes rather than the absolute ages that are important (See Discussion: Assessment of the model). Ages were also estimated by Day et al. (2008). The estimates are similar to those of Genner et al. However, the differences between the ages of the Ectodini, Lamprologini, Cyprichromini, and Tropheini are slightly reduced in the Day et al. study. Therefore, using the Genner et al. ages will slightly deemphasize the likelihood of differences in rates, making any findings of rate differences using the Genner et al. ages a conservative estimate.



The calculated variance and standard errors for each of the four tribes used in simulations (Ectodini, Lamprologini, Tropheini, and Cyprichromini) and for Limnochromini and Perrisodini are given in Table 2. The principal components analysis was conducted using the morphological data for all six tribes. The inclusion of the morphological data for Limnochromini and Perrisodini did not substantially alter the calculation of eigenvalues or the percent of total variation described by the PC axes due to the small number of species in each clade, but they were included in the eigenvalue calculations because of their use in linear regressions. The model generates simulated data for the first four PC axes, which is then compared with the real data for the first four PC axes only. However, comparing the simulated data with total morphological variance rather than PC1–PC4 variance does not substantially alter the results (Appendix S4). Bootstrap iterations demonstrate that variance does not decrease with small subsamples of species from the more disparate tribes and thus, that the low disparity of Cyprichromini is not due to the number of species (Appendix S5). The consistency in variance regardless of the number of samples selected is expected because variance is not dependent on the number of species or the sample size (Foote 1993, 1997; Ricklefs 2006a).


The results of the rate comparisons are summarized in Tables 3 and 4, and in Appendices S2, S3, S4, S6, and S7. Note that there is little support for higher turnover rates in tribes with higher disparity except for between Cyprichromini and both Lamprologini and Tropheini. Ectodini may also have experienced higher turnover rates than Cyprichromini, although the support for this hypothesis is too low to be conclusive. There is also substantial support for lower rates of morphological change in Cyprichromini than the other three tribes. Support for differences in morphological change rates tends to be most pronounced with the Min–Max model variation and least pronounced with the No Condition model variation. The analyses in which tribes are treated as the same age indicates that age differences, in addition to stochasticity, are important and can account for the all of the differences in disparity among tribes with the exception of the low disparity of Cyprichromini (Appendix S6).

Table 3.  Comparisons of rates of morphological change. These are the results of the analyses for each pairwise comparison of morphological change rates when a success is defined as a trial in which the simulated variance is within one standard error of the observed PC1–PC4 total variance. Tribe names (Ectodini, Lamprologini, Tropheini, and Cyprichromini) are abbreviated by their first letter in the subscripts.
 ModelAICcΔAICcAkaike wt.
 No ConditionrT>rC5.8240.0000.714
 No ConditionrL>rC7.9900.0000.633
 No ConditionrE>rC7.2130.0000.563
 No ConditionrE>rL9.2840.7250.410
 No ConditionrE>rT7.7971.4180.330
 No ConditionrL>rT8.4900.9230.387
Table 4.  Comparisons of turnover rates. The results of the analyses for each pairwise comparison of turnover rates when a success is defined as a trial in which the simulated variance is within one standard error of the observed PC1–PC4 total variance are presented below. Tribe names (Ectodini, Lamprologini, Tropheini, and Cyprichromini) are abbreviated by their first letter in the subscripts.
 ModelAICcΔAICcAkaike wt.
 No ConditionpqTpqC34.5752.9970.183
 No ConditionpqLpqC41.1463.6420.139
 No ConditionpqEpqC36.0110.1930.476
 No ConditionpqEpqL42.1980.0000.903
 No ConditionpqEpqT36.2620.0000.861
 No ConditionpqLpqT41.3310.0000.532

The pure-birth model also yields results that are consistent with the results in Tables 3 and 4 (Appendix S7). The interpretations for morphological change rates are identical to those obtained from the birth–death model, and the Akaike weights are actually higher than those presented in Table 3. There are some differences in interpretations of turnover rates from those made using a birth–death model, but the general patterns are largely consistent.


The effect of using different conditions in the model is also analyzed. The first approach to comparing the models is to determine whether the interpretations of results change when different conditions are used. In the results of this study, there are no cases in which changing the model conditions changes which tribe is interpreted to have the higher rate of morphological evolution, and interpretations of which tribe experiences higher turnover rates do not change as long as ΔAICc>∼0.3 (Table 4 and Appendix S4). Figure 2 provides a visualization of the variations in support for hypotheses of rate differences using different model variations. The second approach to assessing the sensitivity of the model to different types of conditioning uses the number of successes under each parameter set. The model results correlate well with each other, which also suggests that the type of model conditioning does not significantly affect interpretations (Appendix S8). Also note that the simulations conducted under a time-dependent model (Appendix S3), a pure-birth model (Appendix S7), and a Min–Max model in which final species diversity is conditioned on the observed diversity of the two tribes in each pairwise comparison (Appendix S2) all provide results that are very similar to those given in Tables 3 and 4. It is also important that the results presented in this study are not dependent on the number of clades included in the analysis because analyses are all done by pairwise comparisons.

Figure 2.

Akaike weights for higher evolutionary rates in more disparate tribes. Each point represents a pairwise comparison between two tribes under a single model. The dashed lines are at Akaike weights of 0.5 and indicate equal support for higher versus lower or equal rates in the tribe of higher disparity. Points that fall below or to the left of the dashed lines indicate that lower or equal rates in the more disparate tribe are more likely, whereas higher evolutionary rates in the dashed lines are supported for the comparisons above or to the right of the dashed lines. Note that comparisons between tribes using different models are clustered together, indicating that the model is robust to the conditioning scenario chosen.


Regression of the multivariate variance on clade age for the six tribes presented in Tables 1 and 2 was used as a separate assessment of the relationship between clade age and disparity. Figure 3 shows the resulting correlation between age and morphological disparity. Note that, although a significant correlation is expected, the correlation is only moderate and is not statistically significant when all tribes are included (Fig. 3A, r2= 0.39, P = 0.18). However, the correlation is very strong when Cyprichromini is excluded (Fig. 3C, r2= 0.92, P = 0.009). The variance for Cyprichromini falls well below that predicted by its age and outside the 95% confidence interval, as demonstrated by the large negative residual (Fig. 3B, residual =–70.5) in comparison with those of the other tribes (see also Fig. 3D).

Figure 3.

Correlation between age and disparity. (A) The correlation between variance and age for the tribes Ectodini, Lamprologini, Tropheini, Perrisodini, Limnochromini, and Cyprichromini (r2 = 0.39, P = 0.18). (B) Residuals for the regression in (A). Note that Cyprichromini has a very large negative residual. (C) Correlation between variance and age for the tribes in (A), excluding Cyprichromini. Note that the correlation is much stronger (r2 = 0.92, P = 0.009). If rates of morphological change, speciation, and extinction are equal, then the increase in variance over time would be expected to be nearly linear (Foote 1996, Pie and Weitz 2005). (D) Residuals for the correlation in (C).



Before assessing possible reasons for differences in disparity among the tribes, it is necessary to establish that differences in disparity are real and not simply sampling artifacts. In particular, because Cyprichromini has both the smallest sample size and the lowest disparity of the tribes that were simulated for this study, the low disparity could be due to poor sampling.

However, several lines of evidence suggest that the low variance of Cyprichromini is not due to differences in sample size. First, there are only an estimated 10 species in the tribe Cyprichromini. Furthermore, all six confirmed, described species (Brandstatter et al. 2005; Day et al. 2008) are included in this study. There are an estimated three (Day et al. 2008) or four (Genner et al. 2007) possible species that have yet to be described and are currently classified as geographic variants of Cyprichromis leptosoma. Thus, the six species for which data were available represent at least 60% of the total species within the tribe Cyprichromini, and the remaining possible species are unlikely to add significantly to the total disparity due to their similarity to C. leptosoma.

In contrast, less than 60% of the total estimated species diversity is represented by the data for the other three tribes (Table 1). Moreover, variance is independent of the sample size or the number of species in a clade (Foote 1993, 1997; Ricklefs 2006a). This is further supported by the bootstrap analyses, which indicate that sampling only six species with replacement yields the same variance as sampling the full sample size of Ectodini, Lamprologini, and Tropheini (Appendix S5). Moreover, both Limnochromini and Perrisodini have smaller sample sizes than Cyprichromini (both are represented by only five species), but have variance roughly comparable to that of Tropheini, which has a similar age. Finally, there is not a significant correlation between morphological variance and the number of samples for Ectodini, Lamprologini, Tropheini, Cyprichromini, Limnochromini, and Perrisodini (r2= 0.35, P = 0.21). Overall, this suggests that the differences in variance among the tribes are due to factors such as stochasticity, clade age, or differences in evolutionary rates rather than sample size.

There is no indication that the greater disparity of Ectodini is due to differences in either morphological change or turnover rates when compared with Lamprologini and Tropheini. In fact, there is strong support from the speciation-dependent model that Ectodini may have experienced a lower turnover rate than Lamprologini and Tropheini. For a speciation-dependent model, disparity increases with turnover rates. Thus, the high disparity of Ectodini relative to the other tribes is due to its greater age, similar morphological change rate, and equivalent or lower turnover rate. Likewise, there is little support for differences in evolutionary rates between Tropheini and Lamprologini, and thus, the disparity differences between these tribes are also likely due largely to age differences.

However, rates of morphological change are substantially greater in Ectodini, Lamprologini, and Tropheini when compared with Cyprichromini. Under the Min–Max model, the Akaike weights for lower rates in Cyprichromini than the other three tribes were all greater than 0.80 (Table 3). The other model variations provide slightly lower support for rate differences, but the interpretations are consistent with the findings from the Min–Max model. These results are also strongly supported under pure-birth conditions (Appendix S7), suggesting that these results are highly robust to assumptions about the value of ɛ. Furthermore, there is considerable support for higher turnover rates in both Lamprologini and Tropheini in comparison to Cyprichromini with average Akaike weights over 0.80 in both cases. Support for higher turnover rates in Ectodini than Cyprichromini is lower, with an average Akaike weight of only 0.60, which is too low to be conclusive. Under a pure-birth model, higher rates in the more disparate tribe are only supported for Tropheini relative to Cyprichromini (Appendix S7). Thus, interpretations of turnover rate differences may be slightly more susceptible to assumptions regarding the value of ɛ than those for morphological change rates.

The linear regression analysis provides a second approach to examining rate differences among the tribes. According to Foote (1996) and Ricklefs (2006a), for a constant number of species, disparity increases linearly with time if rates of speciation, extinction, and morphological change are constant. Pie and Weitz (2005) found that for the more realistic branching random walks, the relationship between variance and time is nonlinear, but is approximately linear after a short initial period (t = 1/λ). Thus, given maximum likelihood estimates for origination rates for the tribes in this study (ɛ= 0.9), disparity should increase linearly after about 0.5–2 million year (My) if the evolutionary rates are constant. All of the tribes examined in this study, including Perrisodini and Limnochromini, are substantially older than this threshold. Thus, if the rates are the same among the tribes, a linear relationship is expected between clade age and disparity.

While age and morphological variance are strongly positively correlated in the five tribes other than Cyprichromini, the latter falls well below the expected variance. Cyprichromini falls outside the 95% confidence intervals, based on the calculated regression error (Fig. 3A), indicating that Cyprichromini is a statistical outlier. Further, if Cyprichromini is included in the regression, there is not a significant correlation between age and variance even though one is expected under equal rates of evolution. This suggests that the differences between the tribes Ectodini, Tropheini, Lamprologini, Perrisodini, and Limnochromini are all consistent with evolution under the same turnover and morphological change rates, but that Cyprichromini likely had lower evolutionary rates to have the currently low standing disparity. This evidence, combined with the general agreement among all model variations, indicates that the Cyprichromini evolved with lower morphological change and turnover rates.

This approach may also provide a means for comparing rates among large numbers of clades without simulations. However, to use this approach without simulations, further testing is required to determine how much variation from a calculated correlation would be consistent with constant rates. Furthermore, linear regressions do not provide information on whether differences in turnover versus morphological change rates are most likely responsible for disparity patterns. Thus, the regression approach may be most useful as a way to determine whether simulation studies are likely to reveal information on rate differences or, as in this study, as a supplement to simulation-based study, rather than as the only method for assessing rate differences.

The strong linear correlation between disparity and age indicates that age differences are largely responsible for standing disparity patterns among Ectodini, Lamprologini, and Tropheini. Further evidence that these patterns are due at least partially to age differences and cannot be explained simply by the fact that stochastic variation comes from the simulations in which all tribes are treated as if they are of the same age. There would be moderate support for turnover rate differences if the tribes were of the same age, but there is no support for rate differences given the actual ages of the tribes. These results indicate that older tribes may have higher disparity because the three tribes evolved under similar evolutionary rates, but have had different amounts of time to diversify (Appendix S6).

The lower evolutionary rates experienced by Cyprichromini are most interesting within an ecological context. On the basis of previous work on the adaptive character of cichlid morphology, tribes with a low diversity of trophic niches and preferred water depths might be expected to have lower evolutionary rates in an adaptive radiation as these are the ecological traits that correlate most strongly with morphology. Cyprichromini, which has lower rates than in the other tribes described in this study, has become progressively more adapted to a pelagic lifestyle, and perhaps more importantly, to the planktotrophic niche (Brandstatter et al. 2005). In contrast, members of the Ectodini, Lamprologini, Tropheini, Limnochromini, and Perrisodini all consume multiple food types (Konings 1998; Koblmuller et al. 2004, 2007, 2010; Duftner et al. 2005; Day et al. 2007).

Given the ecology of Cyprichromini, there are two possible explanations for standing disparity patterns. The first is that the species of Cyprichromini reached an adaptive optimum as open-water planktotrophs that subsequently resulted in a low rate of morphological change. That the Cyprichromini species have become progressively more specialized (Brandstatter et al. 2005) may provide support for this hypothesis in that it indicates that the tribe has been evolving toward a particular niche, and perhaps thus also toward a particular morphological optimum. This might also indicate that the morphological evolution of Cyprichromini could be better explained by an Ornstein–Uhlenbeck model than a Brownian Motion model (e.g., Hunt et al. 2008). Sidlauskas (2008) found that the approach used in this study may indicate low morphological change rates in clades that repeatedly evolve the same morphology as opposed to those which diffuse through a larger portion of morphospace. A species-level phylogeny would be required to test this hypothesis (Sidlauskas 2008). The second possibility is that the Cyprichromini have only succeeded in specializing for one trophic niche because their low morphological change and turnover rates do not enable them to diversify as rapidly into the range of trophic niches occupied by the other tribes. However, if this second option is the case, then the underlying reason for low rates of morphological change in this group remains to be determined.

Regardless of the ordering of cause and effect, these results suggest congruence between trophic diversity and rate of morphological evolution. Similarly, in comparing the lack of disparity in younger clades (e.g., the species flocks of Lake Malawi and Lake Victoria) with the more morphologically disparate assemblage of Lake Tanganyika, the differences in taxonomic diversity, clade age, and stochastic variation should be examined. The results presented here emphasize the importance of establishing differences in evolutionary rates before making biological interpretations to avoid these confounding factors.


This study develops an approach for comparing rates among clades that are not sister taxa. However, the approach requires the use of ages calculated using fossil or Gondwanaland calibrations of molecular clocks are necessary for comparing clades of different ages. These two calibrations yield very different ages (Genner et al. 2007) and the results are controversial (Stauffer et al. 2006; Koblmuller et al. 2008). However, this study does not seek to calculate diversification or morphological rates, but rather, only to determine whether these rates are greater in some clades relative to others. As should be the case, the ratios between the ages of the tribes are identical (Table 1), regardless of whether the fossil calibrations or the Gondwanaland calibrations are used. Therefore, the assertion that each time step represents 100,000 years may be inaccurate, but the relative proportions of time for diversification of each tribe should represent the actual ages very well. Accordingly, inaccuracies in the absolute age estimates of the tribes should have a minimal effect on the results of this study.

Although it is the ratio of the ages rather than the absolute ages of the tribes that are important for this model, it is also informative to consider how errors in the estimates of these ages are likely to influence the interpretations presented in this article. Note that some phylogenies place Cyprichromini as the sister group to Ectodini (e.g., Takahashi 2003), although many studies differ in their interpretations of the relationships of the Tanganyikan cichlids (Koblmuller et al. 2008). The age used for Cyprichromini in this study is less than half that of Ectodini, and therefore, the expected variance in Cyprichromini would be significantly lower than in Ectodini due to the differences in clade age alone. Furthermore, the Day et al. (2008) ages for the six tribes examined in this study also indicate that the tribe Cyprichromini is older relative to the other tribes than estimated by Genner et al. (2007, See Table 1). If Cyprichromini is in fact relatively older than Genner et al. estimated, then the expected variance of Cyprichromini would be higher if the evolutionary rates were equal, and thus the difference between expected and observed disparity would be even larger for this clade. In this scenario, support for rate differences between Cyprichromini and the other tribes would be higher than estimated by this study. Therefore, this study represents conservative findings with respect to the likelihood of lower evolutionary rates in Cyprichromini.

Either a pure-birth or a birth–death model may be appropriate in representing the radiation of cichlids. However, these two models yield similar results. In particular, morphological change rate comparisons are very similar between the two types of model, although differences in turnover rates among clades are somewhat more dependent on the extinction rate (Appendix S7). This provides further support that the results of this model are robust to its underlying assumptions, and that the interpreted similarities and differences in evolutionary rates among cichlid tribes are not artifacts of model design.

One of the previously unresolved questions regarding this model (Sidlauskas 2007) was that differences in turnover rates were not determined to be a major factor in generating clades of differing levels of disparity, despite simulation studies that suggest that turnover rates should play a major role in determining standing disparity (O’Meara et al. 2006; Ricklefs 2006a). In contrast, differences in turnover rate are strongly supported by this study. Sidlauskas (2007) suggests that his model does not indicate differences in turnover rate because the model is conditioned upon achieving the modern species diversity (equivalent to the Min–Max model of this study). However, the Min–Max model of this study yielded results that are consistent with results from the other three model conditions (the No Condition, Extant, and Min models). All four models generally produce numbers of successes under the same parameter combinations that correlate well with each other (Appendix S8). It may be that including the nonuniform priors for turnover rates results in more support for differences in turnover rates. It is also possible that turnover rates have played a more substantial role in shaping the disparity of Tanganyikan cichlids than that of South American characiform fish. Either way, the results presented here indicate that this model can distinguish differences in turnover rates, regardless of the model conditioning variation that is employed.

Additional research could further our understanding of the effects of turnover rates on standing disparity in cichlids. For example, assumptions of time homogeneous rates or simple random extinction could contribute to the disparity patterns generated by the model. Temporally constant rates are less likely in large clades, and many studies suggest that evolutionary rates are heterogeneous with respect to time (Purvis 2008). Furthermore, simulations indicate that random extinction of 95% of species may reduce total branch length in a phylogeny by as little as 19%, whereas selective extinction may have a much more substantial influence on phylogenetic diversity (Nee and May 1997) and thus also on disparity. Thus, simulations that incorporate these elements could further assess the robustness of the results presented here.

Although future research may provide insight into the nature of evolutionary processes in East African cichlids, the interpretations presented in this study are generally robust to the type of conditioning and to multiple assumptions of the model. This includes time-dependent and speciation-dependent models, five different approaches to conditioning model results, pure-birth versus birth–death models, and two different approaches to comparing model results with empirical disparity. This finding has encouraging implications for future studies. That the results are generally consistent across such a wide range of assumptions implies that the choice of model conditioning does not substantially alter the interpretations of results and the models may be used interchangeably to some extent. However, when possible, using multiple versions of the model and testing assumptions to demonstrate the robustness of the model is advised.


The results presented here provide strong evidence that the tribe Cyprichromini evolved with a low rate of morphological change in comparison with Lamprologini, Tropheini, and Ectodini, and a low turnover rate in comparison to Lamprologini and Tropheini. It is less clear that the turnover rate differed significantly between Cyprichromini and Ectodini. These low evolutionary rates may be associated with their low trophic diversity. However, there is little indication of differences in evolutionary rates among the latter three tribes. Simulation and regression analyses indicate that differences among these tribes are due to differences in age and stochastic variation. These patterns are consistent regardless of model variations or approaches to data analysis.

The exploration of model conditions strongly suggests that these results are not highly sensitive to differences between pure-birth and birth–death models, time-dependent and speciation-dependent models, or five variations of model conditioning. Although the Min–Max variation of the model is somewhat more sensitive to rate differences among groups, the conditioning of the model does not change the direction of interpretations of rate heterogeneity among clades for any significant difference in AICC. This implies that interpretations of differences in rates are unlikely to be substantially altered by choice of model conditions.

Associate Editor: Dr. Gene Hunt


Gene Hunt and two anonymous reviewers provided comments that greatly improved the quality of this manuscript. I am extremely grateful to M. Foote for advice and direction in both the design and implementation of this project. I thank D. Hulsey for providing extremely useful comments on an earlier version of this manuscript. I am also grateful to B. Sidlauskas for providing access to the original morphtreegen Matlab script and for helpful discussion on the direction of this project, and P. Chakrabarty for providing access to his data on cichlid morphology. I thank E. King, D. Bapst, M. Hopkins, M. Brady, K. Jenkins, and J. Adams for helpful conversation on this project and comments on earlier drafts of this manuscript. I also thank E. King and J. Hoerner for the use of their computers to run simulations. This work was supported in part by the NSF Graduate Research Fellowship Program and the University of Chicago Department of Geophysical Sciences.