Closed GoogleCodeExporter closed 9 years ago
Original comment by DiegoDL84
on 17 Nov 2014 at 4:43
There is a bug in jModelTest that seems to make it stop when the criteria
produces a negative score. Actually the data does not contain information
enough for making a reliable selection. In the first case, there are only 60
sites for 30 taxa (i.e., k = 58 branch length parameters + 0 to 10 model
parameters). AICc expects (n - k + 1) to be a positive number, where n is the
sample size (i.e., the sequence length). That means that you should not use
AICc for that datasets.
However, jModelTest should not get stuck in any case, so this is a bug I must
fix.
Thank you for your report!
Original comment by DiegoDL84
on 17 Nov 2014 at 5:00
Hi, thanks!
First of all, please note that it happens not only in the AICc table (e.g.
example2). I would be able to generate more instances if you need it.
Second, can you please indicate in which cases this may occur? I also have
alignments (e.g. 194 taxa, 246 length - attached) whose penalty should be
negative (-140 for 0 additional parameters in this case), but the running
completed without a problem.
I want to get the AICc score for most of my data-sets. What exactly should I
check to predict that jModelTest is going to get stuck?
Thank you!
Original comment by shira...@gmail.com
on 18 Nov 2014 at 11:27
Attachments:
I fixed the problem with AICc (only in the repository so far). The correction
part of the AICc is 2*k*(k+1)/(s - k - 1), where s is the sample size and k the
number of parameters. The problem in those alignments was that for some models,
(s = k - 1), producing a division by zero. I fixed it such that in that case
the AICc score will be zero. However, I will also add a WARNING, because the
results will not be accurate when s is close to k.
Moreover, AICc produces negative values when s <= k and the correction part is
larger than the AIC score.
What I could not see so far is the problem with example3.phy. Could you please
attach the output for this file?
Original comment by DiegoDL84
on 19 Nov 2014 at 4:39
Could you try the new 2.1.7 version?
Original comment by DiegoDL84
on 20 Nov 2014 at 3:18
Hi,
Thank you! It ends as expected for inputs it got stuck on before.
I'll try it on many inputs in the next two days, and let you know if it
finished safely.
Tell me, is there supposed to be any difference in computation between v.2.1.7
and v.2.1.5? Is it possible that I get different scores if I use the same PhyML
version?
Thanks,
Shiran
Original comment by shira...@gmail.com
on 21 Nov 2014 at 8:22
No. The results will be the same.
Original comment by DiegoDL84
on 21 Nov 2014 at 8:27
Original issue reported on code.google.com by
shira...@gmail.com
on 17 Nov 2014 at 7:14Attachments: