Significance

Roofit and RooStats
1

I try to perform the significance using an unbinned and reweigthed dataset taking a look to : root.cern.ch/viewvc/trunk/tutori … ificance.C but it doesn’t work.
I have also some problems to plot the profile-log-likelihood function.
Any suggestion?

2

Hi,

can you please post your macro so I can understand why it does not work for you a code like the one of the rs502 tutorial.

Thank you

Lorenzo

3

Here in attachment there is the macro that I’m using.

Thanks

4

Hi,

can you post please also the TFile so I can try running your program. Looking at the code it looks fine to me. Which ROOT version are you using ?
They were some problem producing the likelihood plot with a version older than 5.26.

Cheers

Lorenzo

5

The TFile seems to be too big to attach it.
How can I do?
Anyway I’m using ROOT v.5.26

6

Can you not put the file in public afs or somewhere else where I can download it ?
Cheers

Lorenzo

7

Done :slight_smile:

8

Hi,
thanks for the file. I will investigate it. It could be a problem due to the fact you are using a weighted data set

Cheers

Lorenzo

9

yes I’m using it!

Thanks a lot

10

Hi,

I have found first a problem in RooFit in creating the NLL. The Profile Likelihood calculator uses the option CloneData ( in RooAbsPdf::createNLL) and it does not work for your specific data set. It is maybe related to the fact that is weighted. I have submitted a bug in RooFit for this. (see savannah.cern.ch/bugs/index.php?63997 )
With this fix (or by removing CloneData at the line 173 of roofit/roostats/src/ProfileLikelihoodCalculator.cxx) you should be able to get the plot and a significance value.

However, I think, since you are having a weighted data sets, the interval and significance returned by the ProfileLikelihood calculator are for the pseudo-likelihood, and not for the real likelihood. They could probably be corrected using the same approximation that is used in correcting the error matrix. But I am not sure of its validity in case of the significance.
Also, the approximation, I think is valid if the weights do not differ by a lot from 1. In your case, some of the wights are negative !

I would try to re-formulate the problem by using an standard unweighted likelihood fit.

Best Regards

Lorenzo

11

You are rigth, some of the weights are negative !
But we are using sPlot and this is the reason why we have negative weigths.
So I’m not sure that significance is correct and how I can do.