Which criteria would be better, AIC, BIC or AICc based on real running data

[Runhua Lei]

Hi Rob:
I did read through your paper and the one that you mentioned.
I run six different combinations of AIC, AICc, and BIC, Linked or unlinked.  It seemed to me that unlinkedAIC was the better choice for partition.  However, it is very interested to see that how different criteria make total different schemes (8 partitions to 23 partitions).  I ran Partitionfinder with default.  Could be something wrong with my setting or preference? What do you think?

Thank you very much,

All the best,

Runhua

	AICc	lnL	Number of params	Number of subsets
linkedAICc	285197.6468	-142282.1487	310	23
unlinkedAICc	284895.0972	-141529.743	864	8
	AIC	lnL	Number of params	Number of subsets
linkedAIC	285187.2469	-142282.6235	311	23
unlinkedAIC	284787.4859	-141529.743	864	8
	BIC	lnL	Number of params	Number of subsets
linkedBIC	286823.5273	-142543.0236	181	9
unlinkedBIC	287929.9599	-142409.8873	324	3

[Paul Frandsen]

Hi Runhua,

It is totally reasonable for the different methods to generate different numbers of subsets. The numbers that you list fall in line with what I would expect considering how the different methods work. The runs using unlinked branch lengths require many more parameters to be estimated for each subset, therefore, fewer total subsets would be expected because the much larger number of parameters estimated per subset incur a higher penalty in the calculation of AIC(c)/BIC.

To decide which scheme to use, you should first pick your preferred information criterion (Rob suggested AICc for good reason). Then you should select the scheme with the best score associated with that criterion. If you do decided to choose AICc, then the best scheme is the one generated with unlinked branch lengths. However, keep in mind that if you choose this scheme, and you are planning on estimating a phylogenetic tree, you should use software that allows you to estimate a tree with unlinked branch lengths among subsets. You can usually figure this information out by consulting the software manual/documentation.

My best,

Paul

--
You received this message because you are subscribed to the Google Groups "PartitionFinder" group.
To unsubscribe from this group and stop receiving emails from it, send an email to partitionfind...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

Bioinformatics and Genomics

Office of Research Information Services

Office of the CIO

Smithsonian Institution

paulfrandsen.com

[Runhua Lei]

Hi Paul:
Thanks for your explanation.
I talked quite a lot of people at IPS about mitogenome data and partition.  They strongly suggested that mitogenome was one loci and unlinked model may not fit my data.
Do you think that it is reasonable to choose linked AICc best scheme?
Thanks,

Runhua

[Paul Frandsen]

Hi Runhua,

I think you are confusing what the linked and unlinked branch lengths mean. In a model with unlinked branch lengths, the underlying branching pattern is the same, but each branch is given the 'freedom' to have it's own length for each subset. This is opposed to the linked model in which branches are scaled in tandem with a single rate multiplier. This has nothing to do with whether it is one locus or not--it just assumes that different portions of the dataset are best fit with different model parameters, e.g. 3rd codon positions will most often evolve very differently than 2nd codon positions and should be modeled differently even though they are from the same locus. As Rob pointed out, unlinked branch lengths are just a rather crude method to account for heterotachy. Generally they add so many parameters that they are rarely favored. However, with your dataset, the unlinked branch length model had a better AICc score. This could be *because* it is a mitogenome, not in spite of it. Mitogenomes can exhibit some weird behavior, heterotachy being one of them.

Hope that helps.

Paul