Effect size similar to Cohen's d in limma
3
1
Entering edit mode
@january-weiner-4252
Last seen 5.4 years ago
European Union

What is the recommended way of calculating a standardized effect size for each gene in limma, preferably based on the fit object returned by lmFit? Possibly, would that be something like

f$coefficients / f$stdev.unscaled
limma effect size • 4.6k views
ADD COMMENT
5
Entering edit mode
@gordon-smyth
Last seen 1 hour ago
WEHI, Melbourne, Australia

Cohen's d is just logFC divided by the estimate of sigma, so in limma it is:

fit$coefficients / sqrt(fit$s2.post)

If there are two groups, and the coefficient holds the logFC between them, then this an empirical Bayes version of Cohen's d. The ordinary (not EB) Cohen's d would have fit$sigma in place of sqrt(fit$s2.post).

ADD COMMENT
0
Entering edit mode

Thanks, I'll go with that.

ADD REPLY
0
Entering edit mode

Thanks, I'll go with that.

ADD REPLY
0
Entering edit mode

Thank you Gordon, a follow-up question is how to compute the the confidence interval for the Cohen'd in limma?

ADD REPLY
1
Entering edit mode
@ryan-c-thompson-5618
Last seen 28 days ago
Icahn School of Medicine at Mount Sinai…

Well, Cohen's d statistic is defined as the difference between two groups over the standard deviation. If you want to be consistent with the way other statistics are calculated in limma, you should use the posterior estimate of the standard deviation as calculated by eBayes(), which would be accessed by sqrt(f$s2.post). For the ordinary sample standard deviations, use f$sigma. As for the numerator, you should use the logFC value for whatever coefficient or contrast you are testing, which you can get from topTable(). However, be careful if you are trying to compute this for a coefficient or contrast that does not represent a simple difference between two means, such as an interaction term.

ADD COMMENT
0
Entering edit mode
svlachavas ▴ 830
@svlachavas-7225
Last seen 13 months ago
Germany/Heidelberg/German Cancer Resear…

Dear January,

im not an expert about these details but I think this post has some useful information 

C: Standard error and effect size from Limma

ADD COMMENT
0
Entering edit mode

Thanks. I have seen this question. However, I am looking for a standardized effect size -- one that takes into account the variability. For example, a log fold change of 4 with CI ranging from 0.5 to 8 has a lower standardized effect size than a log fold change of 3 with CI from 2.5 to 3.5. 

ADD REPLY
0
Entering edit mode

I see. So the more experts on the field will have an appropriate explanation regarding your specific notification 

ADD REPLY

Login before adding your answer.

Traffic: 546 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6