User:R. Eric Collins/MBL/PAUP

David Swofford author of PAUP* and uncredited for many of the methods within

Model Selection

• Parsimony
  • long branch attraction because branch lengths not taken into account so similar bases could equally be from equilibration or conservation

• Models are _always_ wrong
  • don't/can't expect them to match reality

• What is a _good_ model?
  • as simple as necessary but no simpler
  • a balance between under- and over-fitting

• heterotachy: differential rates of evolution at different sites on different branches
  • can confuse maximum likelihood and choose long-branch tree
  • new mixture models are being written to address this issue

• Model Selection Criteria
  • Likelihood ratio tests: δ = -2(ln L0 - ln L1)
    • chi-squared (frequentist) based, so always have possibility of Type I error, depending on error tolerance (α)
    • i.e. reject simple model in favor of more complex model even if simple model is true
  • Akaike information criterion (AIC)
    • AIC_i = -2lnL_i + 2K
    • tends to overestimate (be liberal with) number of parameters
  • Bayesian information criterion (BIC)
    • BIC_i = -2lnL_i + K ln n where n is sample size (typically number of sites)
    • converges on correct answer as more data is added

• PAUP
  • tips
    • restrict dataset after loading datafile instead of making multiple copies of data subsets
    • uses -ln L so all things are minimized in PAUP (lower is better)
  • ModelTest
    • any reasonable tree can be used, actual tree topology has little effect on model selection
    • shouldn't have to run ModelTest, better to understand model selection well enough to winnow down model manually

to specify Tamura-Nei (transitions have equal probabilities, each transversion has different probabilities) lscores/nst=6 rclass=(abaaca)

• to only operate on a subset of data
  • first load all the data
  • taxset: set macro of taxa names
  • delete: delete a certain subset of taxa
  • exclude: exclude subset of characters