Closed JaneZXJ closed 8 years ago
spond
commented
8 years ago Dear @JaneZXJ,
Values like this should be interpreted as infinite. There is usually a very wide confidence interval on such values, so you should simply treat them as large values, greater than 1 (if the LRT test is significant).
Sergei
murrellb
commented
8 years ago To chime in here:
Imagine you're flipping a coin, and you want to estimate the ratio of heads/tails. Lets say you flip that coin 6 times, and see only heads. You can reject the null hypothesis that the heads/tails ratio is 1 with a bit of confidence, but the point estimate for this would be that the ratio is infinity, which is just a pathology of parameter point estimates (but not p-values!) for small samples in the maximum-likelihood framework.
In the analogy, non-synonymous substitutions are heads, and synonymous substitutions are tails, so all that these sorts of problematic point estimates require is that there are no synonymous substitutions inferred at that site or on that branch or on that bit of the tree (depending on the model you're using).
Cheers, Ben
On Thu, Feb 18, 2016 at 10:13 AM, Sergei Pond notifications@github.com wrote:
Dear @JaneZXJ https://github.com/JaneZXJ,
Values like this should be interpreted as infinite. There is usually a very wide confidence interval on such values, so you should simply treat them as large values, greater than 1 (if the LRT test is significant).
Sergei
— Reply to this email directly or view it on GitHub https://github.com/veg/hyphy/issues/393#issuecomment-185842213.
JaneZXJ
commented
8 years ago Thanks, Ben, It's a good analogy to understand this issue. So if it's the problem of small sample size in ML estimation. Is is a test for suitable sample size in the performed model ? Or it's really not a problem for episodic selection analysis here. Thanks again!
Cheers, Jane
Xiaojia Zhu Ph.D student Ornithological research group Key Laboratory of Zoological Systematics and Evolutionary Center 5 Institute of Zoology, CAS 1 Bei Chen West road Chao Yang District Beijing,100101 xiaojia0402@hotmail.com zhuxiaojia@ioz.ac.cn Office: 010-64807188 Phone: 86-13581827838
From: Ben Murrell Date: 2016-02-19 02:26 To: veg/hyphy CC: JaneZXJ Subject: Re: [hyphy] If ω > 1000, 5000,even 10000, how to explain? (#393) To chime in here:
Imagine you're flipping a coin, and you want to estimate the ratio of heads/tails. Lets say you flip that coin 6 times, and see only heads. You can reject the null hypothesis that the heads/tails ratio is 1 with a bit of confidence, but the point estimate for this would be that the ratio is infinity, which is just a pathology of parameter point estimates (but not p-values!) for small samples in the maximum-likelihood framework.
In the analogy, non-synonymous substitutions are heads, and synonymous substitutions are tails, so all that these sorts of problematic point estimates require is that there are no synonymous substitutions inferred at that site or on that branch or on that bit of the tree (depending on the model you're using).
Cheers, Ben
On Thu, Feb 18, 2016 at 10:13 AM, Sergei Pond notifications@github.com wrote:
Dear @JaneZXJ https://github.com/JaneZXJ,
Values like this should be interpreted as infinite. There is usually a very wide confidence interval on such values, so you should simply treat them as large values, greater than 1 (if the LRT test is significant).
Sergei
— Reply to this email directly or view it on GitHub https://github.com/veg/hyphy/issues/393#issuecomment-185842213.
— Reply to this email directly or view it on GitHub.
Hi, Sergei, In my results of GAB or BSR, the positive selected lineages have extremely high ratios. Is that the case of overfitting of small sample size, as in your BSR paper? My datasets include 11 seqs [http://www.datamonkey.org/spool/upload.590080565620739.1_bsr.php] , 19 seqs [http://www.datamonkey.org/spool/upload.785445843820169.1_bsr.php] or 24 seqs [http://www.datamonkey.org/spool/upload.391348098717127.1_bsr.php].
I'm not sure about that, could you give me more explanation? Thanks so much!
My appreciation for your help. Jane