report BF or log10bf?

[TimL]

Hi all:

Okay - probably a silly question, but one nonetheless that my limited
physiologist's mind is having trouble discerning from the literature.
When using a Bayes Factor (BF) as a test of association, do you report
the BF or the log10bf, which is what BimBam outputs? I've found other
literature to determine the significance thresholds to use to
determine the magnitude of linkage for BF (e.g. the Jeffrey scale of
strength and think I have a good handle on that), but those decisions
require me to know whether I should be dealing with the straight BF
for those decisions or the log10bf.

Thanks,

Tim L.

[Yongtao Guan]

Jeffrey's interpretation is more suitable for single test.
Usually in GWAS, one performs hundreds of thousands tests, so that one
has a much smaller prior odds because one doesn't expect much
interesting findings among those hundreds of thousands tests a priori.
Recall: posterior odds = BF x prior odds.

To report BF in log10 scale or not doesn't matter because one doesn't
change the actually value.
In the following paper, log10(BFs) were reported for type-setting
convenience, I guess.
http://quartus.uchicago.edu/~yguan/pdfs/ReinerEtal2008.pdf

Grant

--
Yongtao Guan, PhD
Postdoctoral Fellow
Dept of Human Genetics
University of Chicago
http://home.uchicago.edu/~ytguan/

[Mathew Barber]

Hi Tim,

In the context of GWAS I don't think that using Jeffrey's scale of strength is a good idea.
In our recent paper "Polymorphisms of the HNF1A Gene Encoding Hepatocyte Nuclear Factor-1α are Associated with C-Reactive Protein" in ASHG, (pubmed link)
we report log10BFs and give the following justification.

"To interpret a BF, the relationship “Posterior Odds = Prior Odds × BF” is helpful. For example, if 1 in 10,000 SNPs are genuinely associated with CRP (Prior Odds = 1:10,001), then a single-SNP log10(BF) of 5, 4, or 3 will result, respectively, in Posterior Odds of approximately 10:1, 1:1 and 1:10, which correspond to a 91%, 50%, and 9% chance of being a genuine association. (Naturally, different prior odds yield different probabilities of a genuine association, and the reader is free to substitute alternative prior odds. Note that the proportion of associations with a given p value that are genuine will also depend on the prior odds, but not in an easily specified way.)The top nine SNPs [log10(BF) > +3] associated with plasma CRP concentration in the combined PARC phase 1 + phase 2 analysis are shown in Table 2."

I hope this helps,

Mat

ps It may be useful to add that assuming that 1 in 10,000 SNPs are genuinely associated is the same as assuming that there are 100 SNPs that are genuinely associated if you presume that there are 10 million SNPs in total.

On Wed, Mar 4, 2009 at 2:47 PM, TimL <jtli...@uncc.edu> wrote:

[jtli...@uncc.edu]

Thanks to both Grant and Mat...both were helpful. Now I've got more
to rattle around in my brain!

Cheers!

Tim

On Mar 4, 4:17 pm, Mathew Barber <mathar...@googlemail.com> wrote:
> Hi Tim,
>
> In the context of GWAS I don't think that using Jeffrey's scale of strength
> is a good idea.

> In our recent paper "Polymorphisms of the *HNF1A* Gene Encoding Hepatocyte

> Nuclear Factor-1α are Associated with C-Reactive Protein" in ASHG, (pubmed

> link<http://www.ncbi.nlm.nih.gov/pubmed/18439552?ordinalpos=1&itool=Entrez...>
> )

> we report log10BFs and give the following justification.

> *
> *"*To interpret a BF, the relationship "Posterior Odds = Prior Odds × BF" is

> helpful. For example, if 1 in 10,000 SNPs are genuinely associated with CRP
> (Prior Odds = 1:10,001), then a single-SNP log10(BF) of 5, 4, or 3 will
> result, respectively, in Posterior Odds of approximately 10:1, 1:1 and 1:10,
> which correspond to a 91%, 50%, and 9% chance of being a genuine
> association. (Naturally, different prior odds yield different probabilities
> of a genuine association, and the reader is free to substitute alternative
> prior odds. Note that the proportion of associations with a given p value
> that are genuine will also depend on the prior odds, but not in an easily
> specified way.)The top nine SNPs [log10(BF) > +3] associated with plasma CRP
> concentration in the combined PARC phase 1 + phase 2 analysis are
> shown in Table

> 2<http://www.sciencedirect.com/science?_ob=ArticleURL&_udi=B8JDD-4SBYX2...>
> .*"

>
> I hope this helps,
>
> Mat
>
> ps It may be useful to add that assuming that 1 in 10,000 SNPs are genuinely
> associated is the same as assuming that there are 100 SNPs that are
> genuinely associated if you presume that there are 10 million SNPs in total.
>