The performance of the above-mentioned measures was compared on 1000 simulated replicates, each with 100 taxa and 600 nucleotides and based on a random tree (data from Desper and Gascuel 2004 (link)). Trees were simulated using the beta-splitting model (Aldous 1996 ), which generalizes the uniform distribution on phylogenies and the standard Yule–Harding branching process (Yule 1925 ; Harding 1971 ), both of which are typically used to generate a distribution of biologically relevant trees. Deviations from molecular clock were introduced to each tree (Desper and Gascuel 2004 (link)). Sequence data were generated using the K2P + covarion model, similar to Galtier (2001) (link), where evolutionary rates vary among sites and over time. Analyses were performed under the incorrect models HKY+ Gamma;4 (moderate model violation) and JC+Γ4 (serious violation). For comparison with the Bayesian approach and with the results from our previous study, we used 1500 smaller simulated data sets, each generated under HKY+Γ4 with 12 taxa and 1000 nucleotides, and based on a distribution of phylogenies generated using the standard speciation process with deviations from the molecular clock (data from Anisimova and Gascuel 2006 (link)). The data were analyzed under both the correct model HKY + Γ4 and the incorrect model JC + Γ4. The Bayesian MCMC analyses were conducted with MrBayes v3.1.2 (Huelsenbeck and Ronquist 2001 (link)) as described in Anisimova and Gascuel (2006) (link). To address concerns that generations used in our previous study may not have been sufficient (despite good convergence diagnostics), we also run longer chains ( generations) under each model.
It has previously been noticed that bootstrap proportions as well as PP can be too high not only for incorrect but also for nonexisting (i.e., zero-length) branches (e.g., Lewis et al. 2005 (link); Yang 2007 (link); Guindon et al. 2010 ). Thus, we tested how often branch partitions were inferred with high supports on star-like data, as can be the case for viral data or samples of deep divergence confounded by selection (adaptive radiation). We simulated 100- and 12-taxa star trees with branches drawn from the exponential distribution with a mean of 0.1 expected substitutions per branch per site. All star trees were simulated under HKY + Γ4.
It has previously been noticed that bootstrap proportions as well as PP can be too high not only for incorrect but also for nonexisting (i.e., zero-length) branches (e.g., Lewis et al. 2005 (link); Yang 2007 (link); Guindon et al. 2010 ). Thus, we tested how often branch partitions were inferred with high supports on star-like data, as can be the case for viral data or samples of deep divergence confounded by selection (adaptive radiation). We simulated 100- and 12-taxa star trees with branches drawn from the exponential distribution with a mean of 0.1 expected substitutions per branch per site. All star trees were simulated under HKY + Γ4.
