I made 5 separate seed trees, and I'm testing out different parameters for the -i and -c values for each of the trees. I was wondering how I should be comparing the resulting likelihood values?
1) what exactly do these values represent? Are they likelihood values or log likelihood values?
2) what criteria do I use to compare two trees based on these likelihood values? (do I look at the 2*(log likelihood difference), and how do I assess statistical significance? or does it not matter and I only need to choose the smaller value?)
When comparing trees made from the same alignment and by the same program, the tree with highest likelihood is considered the best. That's the whole point of maximum likelihood methods in general - to find the highest value of a ML function. Should be easy to find more information about it by Googling.
I haven't read RAxML manual in a long time so it could be that something has changed, but I do not consider those two parameters (-i and -c) to be most important in finding best trees. As a general rule, more categories will yield a better tree, but that will incur a heavy time cost and usually with tiny improvement in likelihood values. In my experience, testing various substitution matrices is usually more productive. That can be done automatically by ProtTest3, but you'll need to have a working PhyML on your system. I think the fastest program for model parameterization is IQ-TREE. Its ModelFinder will take an alignment and quickly go through many parameter combinations and rank them. Afterwards, you can use those parameters in any ML or Bayesian tree estimator.
By the way, if you don't have your heart set on RAxML, I suggest IQ-TREE for subsequent ML tree estimations as well. Besides being able to find model parameters, it is considerably faster than RAxML.
