See Gadagkar and Kumar (2005) for an example of datasets where maximum likelihood is more appropriate than maximum parsimony.
Thanks for the example. Using computer simulations is something that should have donned on me from the start.
In the paper, they show that ML outperforms MP when there is a certain amount of heterotachy. Is there a way of determining the amount of heterotachy in a data set independent of the phylogenetic method?
Well, different phylogenetic methods are more appropriate depending on the data set. I hear you, Taq, but when it comes to molecular phylogenetics it's pretty hard to have an absolute standard given the diversity of sequences involved. So, for example, I've published molecular phylogenies of certain gram-positive protein systems, but opted to use more than one phylogenetic method (ML and Bayesian analysis) to see the extent to which the phylogenies were congruent.
I totally get what you are saying. Going back to my protein assay analogy, the appropriate method can also depend on the protein mixture. If you have an nearly pure sample with just one protein in it, then you can very precisely and reliably measure the concentration of the protein by using UV absorbance since specific amino acids have specific absorbances. If you have an unknown mix of proteins you can't say how many UV absorbing amino acids there are per protein molecule, so the UV method may not be as appropriate.
So, for example, I've published molecular phylogenies of certain gram-positive protein systems, but opted to use more than one phylogenetic method (ML and Bayesian analysis) to see the extent to which the phylogenies were congruent.
By congruence, do you mean congruence to the accepted species tree? If so, are you assuming a lack of horizontal transfer?