RE: TGraphAssymm BAYESDivide

From: Amnon Harel <amnon.harel_at_cern.ch>
Date: Mon, 31 Jan 2011 17:36:26 +0000


Dear H.,

I hope my answer clarified the tools available in TGraphAsymmErrors (and TEfficiency).

For some types of error intervals (you still did not say which you wanted), the one-to-one correspondence between a/(a+b) and a/b yields a one-to-one translation of error intervals.

FYI: in the CMS paper we had Clopper-Pearson intervals in the graphical display of the dijet centrality ratio. One of many reasonable choices.

 cheers,
 Amnon

-----Original Message-----
From: HSing [mailto:harinder.singh.bawa_at_gmail.com] Sent: Mon 31-Jan-11 6:57 PM
To: Amnon Harel
Cc: roottalk (Mailing list discussing all aspects of the ROOT system.) Subject: Re: [ROOT] TGraphAssymm BAYESDivide

Hello Amnon,

Thanks for the reply but I am not calculating efficiency. Infact I am calculating centrality ratio which is number of events in innner region/ number of events in outer region within each mass bin. I am getting one value per mass bin but what is the error on that value. Thats I am trying to figure out. BayesDivide as wellas TEfficiency both are meant for calc error on efficiency where the value cannot exceed 1.But not in my case. My Numerator can be higher than denominator making value >1.

H.

On Mon, Jan 31, 2011 at 5:42 PM, Amnon Harel <amnon.harel_at_cern.ch> wrote:

> Dear H.
>
> That depends on whether you're willing to condition your statistics
> (i.e. the vertical error bars) on the total number of events observed
> in each bin (the sum of the inner and outer counts).
>
> If you are, your problem reduces to the Binomial, and TGraphAsymmError's
> "Divide" method (BayesDivide just calls it, nothing Bayesian there) is
> one of many that handle that. Personally, I'd never use their "shortest
> interval" approach, but obviously, opinions differ on that :-)
> TEfficiency is a newer, cleaner, and more flexible tool that handles
> Binomial errors. You may want to look at it.
>
> BTW: once it's a Binomial problem, just sum the inner and outer counts
> to get the total.
>
> If you're not willing to apply this conditioning, you'll need to program
> the errors in yourself. Also may want to think about which errors:
> Gaussian approximation? Frequentist intervals (with "exact" or on-average)
> coverage)? etc. Depends on what you want to use them for...
>
> cheers,
> Amnon Harel,
> University of Rochester
>
>
>
> -----Original Message-----
> From: owner-roottalk_at_root.cern.ch on behalf of HSing
> Sent: Mon 31-Jan-11 6:20 PM
> To: roottalk (Mailing list discussing all aspects of the ROOT system.)
> Subject: [ROOT] TGraphAssymm BAYESDivide
>
> Hello,
>
> I am wondering if I can calculate Assymetric error on my datapoints usng
> BayesDivide fcn. The only problem is that I am doing ratio of 2 histograms
> ,
> Number of evts in inner rapidity region/number of events in outer rapidity
> region.
>
>
> But BayesDivide says that TAssymError::BayesDivide(h1,h2)
>
> h1 should be subset of h2. But in my case, my h1 is not a subset of h2.
>
> Could anybody suggest?
>
> H.
>
>
Received on Mon Jan 31 2011 - 18:41:16 CET

This archive was generated by hypermail 2.2.0 : Tue Feb 01 2011 - 05:50:01 CET