It has been asserted in an answer to a recent question about single gene phylogenetic hypotheses that "the use of parsimony to choose the best tree is methodologically flawed." I wonder what the community thinks these flaws are, and what method is free from such flaws.
I suspect that long-branch attraction is widely regarded as mainly a problem for parsimony. Indeed, this was the most important argument suggested by Felsenstein (1978) and led to a move towards maximum likelihood analyses. However, in his classic paper, Bergsten (Cladistics 21: 163-193, 2005) nicely showed that long-branch attraction is not confined to parsimony but is also problematic for NJ, ML and Bayesian analyses.
As a PhD student, I had the privilege of sharing an office with James S. Farris and I currently have Fredrik Ronquist on my PhD advisory board. There are excellent arguments on both sides of the debate over which method is best. I too would be interested in reading the replies to Andrew's deceptively simple question.
Nice question. I am eager to read the replies, although I can probably predict what the talking points will be.
Dear Andrew
A difficult question.
First of all more you've got homoplasy more parsimony is problematic. It means (in my mind of course) that parsimony is working well as soon as you've got a moderate rate of mutations. A way to partially solve it is to use a Goloboff correction....but it is not optimal (even if I used it a lot).
As you know mitochondrial genes are clearly biased for AT especially in "our" insects. Considering this point, I think that only max lik is able to deal with such problem (of course using the most complete models).
Anyway, when it is possible, I prefer to compare parsimony, phenetic and maxlik trees and discuss the 3 approaches rather than using a single method.
I had many problems with MrBayes......Bayesian statistics are extremely powerful for many things...but I'm not fan for phylogenetic analysis.
To summarize, flaws exist for all the available methods, we've got just to accept that and staying cautious in our conclusions !!!
Cheers from Luc
I suspect that long-branch attraction is widely regarded as mainly a problem for parsimony. Indeed, this was the most important argument suggested by Felsenstein (1978) and led to a move towards maximum likelihood analyses. However, in his classic paper, Bergsten (Cladistics 21: 163-193, 2005) nicely showed that long-branch attraction is not confined to parsimony but is also problematic for NJ, ML and Bayesian analyses.
As a PhD student, I had the privilege of sharing an office with James S. Farris and I currently have Fredrik Ronquist on my PhD advisory board. There are excellent arguments on both sides of the debate over which method is best. I too would be interested in reading the replies to Andrew's deceptively simple question.
Well, it seems that everybody is interested in reading the replies to this question and I am interested in the answers as well.
I understand the fear of long-branch attraction but as George Sangster pointed, the other methods are affected by the long branch attraction as well. I still wonder why parsimony is computed in TNT in about few seconds and almost the same tree under Markov chains in days and still the ESS values are too small when using GTR+I+gamma. Maybe this is the point why the much longer method is regarded as more scientific (I am kidding)
- Badges
- Science topic
- Similar topics
- Biological Science
- Evolutionary Biology
- Phylogenetics