createChi2 of binned and weighted data set

Roofit and RooStats
1

Hello,

I am using ROOT 5.27/01 (trunk@32187, Feb 02 2010, 16:13:55 on linux).

I would like to do a chi2 fit to a weighted and binned data set. I start with unbinned data from a tree, then create the binned clone.

If I do it this way:[code]

RooDataSet tmp(“tmp”,“tmp”,dataVarSet,Import(tree1));
RooDataSet data = new RooDataSet(“my_data”, “my_data”, dataVarSet, WeightVar(“weight”),Import(tmp), StoreError(dataVarSet) );

RooAbsData data_binned = (RooAbsData)data->binnedClone();
data = (RooDataSet*)data_binned;

//—cut out some code here—

RooAbsReal* chi2 = model->createChi2((RooDataHist&)*data, Extended(true),SumW2Error(true),PrintLevel(1),Range(“cutrange”) );[/code]

Then I get these runtime errors:

[#0] ERROR:Eval -- RooChi2Var::RooChi2Var(chi2_model_my_data_binned) INFINITY ERROR: bin 0 has zero error, but function is not zero (0.0347996) [#0] ERROR:Eval -- RooChi2Var::RooChi2Var(chi2_model_my_data_binned) INFINITY ERROR: bin 0 has zero error, but function is not zero (0.0347996)

And when I do it this way:

RooDataHist *data_binned   = data->binnedClone();
RooDataSet  *data_binned_w = (RooDataSet*) data->emptyClone();
for (int i=0 ; i<data_binned->numEntries() ; i++) {
data_binned_w->add(*data_binned->get(i),data_binned->weight(*data_binned->get(i),1,true,false) ) ;
}
RooDataSet* data = (RooDataSet*)data_binned_w;

//---cut out some code here---

RooAbsReal* chi2 = model->createChi2((RooDataHist&)*data, Extended(true),SumW2Error(true),PrintLevel(1),Range("cutrange") );

I get these runtime errors:

[#0] ERROR:Eval -- RooChi2Var::RooChi2Var(chi2_model_my_data) INFINITY ERROR: bin 3 has zero error, but function is not zero (0) [#0] ERROR:Eval -- RooChi2Var::RooChi2Var(chi2_model_my_data) INFINITY ERROR: bin 3 has zero error, but function is not zero (0)

Can someone explain these errors, or give an example of a chi2 fit to a weighted and binned data set?

Many thanks,
Anthony

2

I’ll ask again.
Can anyone explain these errors, or give an example of a chi2 fit to a weighted and binned data set?

3

The RooFit chi2 fit does not work when the bins have zero entries.
You should either use a binned likelihood fit or use the standard chi2 fit provided by ROOT.
In this case bins with zero entries are excluded from the fit

Best Regards

Lorenzo

4

Ok, good to know. I will use another method. Can someone add that to the documentation? It would have saved me a lot of time.

[quote=“moneta”]
You should either use a binned likelihood fit or use the standard chi2 fit provided by ROOT. [/quote]

Ok, so I’ve tried a binned likelihood fit via a RooAbsPdf::createNLL()
but I have a weighted and binned data set.
There is no option SumW2Error(Bool_t flag) for createNLL(). Doing this:

RooAbsReal* nll;
nll = model->createNLL(*datahist, Extended(true),Range("cutrange"), CloneData(true) ) ;
RooMinuit m(*nll);
m.migrad();
m.hesse();
m.minos();//new
result = m.save();

gives wrong error estimates.
I find it hard to believe there is no solution yet to fit a weighted AND binned dataset and get the correct error calculation.

I found this bug report, which may be relevent. Any progress?
savannah.cern.ch/bugs/?63997

I am using today’s version of ROOT (trunk@33086, Apr 19 2010).

I welcome any suggestions for fitting weighted and binned data.

Many thanks for your help,
Anthony

5

This bug is not yet been fixed. It could be that it affects your case.

You can use plain ROOT, make a TH1 and fit it with your function.
If your model is described as a RooAbsPdf you can get the TF1 simply with the method RooAbsPdf::asTF

TH1 * h1 = datahist->createHistogram(x);
TF1 * f = model->asTF(RooArgList(x) , parameters ); 
h1->Fit(f);

Best Regards

Lorenzo

6

[quote=“moneta”]This bug is not yet been fixed. It could be that it affects your case.
[/quote]

Ok, I’ve added myself to the mailing list for that bug.

Thanks for the suggestion, but this will not work. I have more than 3 observables.
createNLL looks the most promising for me, but it would require a SumW2Error(true) argument that doesn’t exist.

7

It will work but you need to create a THNSparse and fit it or use directly the ROOT::Fit::BinData class to create a ROOT::Fit::Chi2Function to minimize.
If you are interested I can send you an example.
However, due to the curse of dimensionality, I suspect you will have low statistics in many bins. In this case, you need to use a likelihood fit.
Why don’t you transform the binned data set in a weighted unbinned data sets using as first approximation the bin center values ?

Best regards

Lorenzo