I would like to compute Efficiency using TEfficiency, the problem is my input histograms didn’t pass TEfficiency::CheckConsistency() criteria turns out that in some bins pass.GetBinContent(i) <= total.GetBinContent(i).
The only thought I have to overcome this, is by taking the two histograms and rebin them and force the pass histograms to have a bin content in each bin smaller than the total one.
I was wondering if it’s the appropriate (clever) way to do it. Otherwise, I’d be definitely happy to hear your suggestion.
If you have two histogram with independent Poisson counts and you want to make the ratio, i.e. in the case that it is not binomial, I remind you that you can use TGraphAsymmErrors::Divide with the option pots. See ROOT: TGraphAsymmErrors Class Reference.
Otherwise you cannot use TEfficiency if ntotal < npass .
Thanks a lot for your suggestion, It worked, but efficiency plots have values > 100% (It makes sense since h_pass > h_tot).
But from a physics point of view, I think it doesn’t make sense to have this. Right ?
I would be grateful if you could propose an alternative solution ?
I don’t know what the 2 histogram represents, so I cannot say anything about what make sense or not. For sure if one histogram represents a subset of the count of the other (i.e. this is in the efficiency case) then h_pass is < h_tot.
How come? The number of reconstructed jets should always be smaller ot equal to the number of generated jets. If you generate two jets in event and reconstruct three, what is this third one? Pile-up? If so, I’d suggest to tighten your selection criteria and not to account for such jets. Otherwise h_pass does not hold a subset of h_tot, so you can’t use the TEfficiency class.
This is possible if there is the number of reconstructed jets contained a contribution from fake jets.
The process is then more complicated is not a binomial one and neither two independent Poisson process, because there are correlations between and reco and truth jets, I think. You would need some MC simulation or some other techniques to estimate the uncertainty in their ratio.
Or as suggested above tight the selection to make the contribution from additional jets negligible).
However I am not sure how this quantity is useful for you. Normally one looks at the b-tagging efficiency defined as: (number of b tagged jets) / (number of true b jets ) , which seems not what you are doing
The total number of entries in reco hist is less than the total entries in the true hist.
the problem as I said is in some bin the content (h_reco->GetBinContent(1) > h_true->GetBinContent(1)).
if you believe that the in all bins ( h_reco->BinContent() should be less than h_true->GetBinContent()) then the problem is in my code and I believe that the reason is that the reco is not a subset of true.
I mean I had to do something like that
for (int i=0; i < truth; i++){
if (truth_isbjet) h_true->Fill()
for (int j=0; j<reco; j++){
if(reco_isbjet) h_reco->Fill()
}
}
The code above is not what I am using, it was only a thought to overcome this.
what I was doing is looping separately on the truth and reco jets separately (am interested on V->qq (V=WorZ) )
First I retrieve truth information from the dedicated container and choose the leading 2 jets (with max sum pT). then I constructed Dijet system with TLorentzVector (PtEtaPhiM), in order to have b-truth jets I used a variable in our CxAODs called HadronConeExclTruthLabelID which should equal to 5 for b-tagged jet.
I fill information in histograms pT eta, phi and mass of dijet.
like that
h_truth->Fill(dijet.Pt())
For the reco I open a separate loop where I took the reconstructed jets from my reconstructed object (which coontain a lepton, neutrino and 2 jets) and require them to be b-tagged. Then I fill the information as the truth ones.
Please let me know if you need further info.
Thanks
Hi,
Thanks for the clarification. It seems to me than you are filling the histogram with jet results (for truth and reco jets) with many events. As first step then you could assume that the truth and the reco histogram bins are un-correlated, and therefore you could use the Poisson ratio case instead of TEfficiency. At least try this as first approximation, still better than using TEfficiency for this case.
Thank you so much for your suggestion.
If you are refereeing TGraphAsymmErrors::Divide I already tried it with option pois
as you proposed earlier but seems to give some strange results eff = N(reco b-jets)/ N(truth b-jets)
I attached a plot so you can have a look (the efficiency here is computed using TGraphAsymmErrors::Divide )
Hi,
The large error you are getting in the first and last bins are due to a very low statistics, so I think there is noting wrong with this. I think there is nothing wrong with the plot, but maybe, as I mentioned before, I am not sure about the utility of the ratio you are computing
