Fit to data doesn't seem to use data error

[Kreikkk]

Hi! I have a problem with fit to data in RooFit. I have to fit my MC to data in 4 regions. Here is the code I use to fill a DataSet.

RooRealVar x("x", "x", 0, 9);
RooRealVar w("weight", "weight", 1);
RooCategory cat("region", "region");
RooDataSet data("data", "data", RooArgSet(x, w, cat), WeightVar("weight"), StoreError(RooArgSet(x, w)));

for (int i=1; i<=nBins; i++) {
    x = DMBHist[0] -> GetBinCenter(i);
    cat.setLabel("A"); data.add(RooArgSet(x, cat), DMBHist[0] -> GetBinContent(i), DMBHist[0] -> GetBinError(i));
    cat.setLabel("B"); data.add(RooArgSet(x, cat), DMBHist[1] -> GetBinContent(i), DMBHist[1] -> GetBinError(i));
    cat.setLabel("C"); data.add(RooArgSet(x, cat), DMBHist[2] -> GetBinContent(i), DMBHist[2] -> GetBinError(i));
    cat.setLabel("D"); data.add(RooArgSet(x, cat), DMBHist[3] -> GetBinContent(i), DMBHist[3] -> GetBinError(i));
}

Then I perform a binned fit using RooMinimizer, but the fit doesn’t seem to use data error at all beacuse error for data is huge, but the final estimate has very small error.

When I change the last argument in data.add() method that is supposed to be the error of the bin to any number I get the same error for the final estimate. Here is how I perform a fit.

RooArgList models;
for(int j=0; j<4; j++) models.add(*TRooFit::BuildModel(*hist[j], data));

RooSimultaneous model("model", "model", models, cat);

RooAbsReal *nll = model.createNLL(data);
RooMinimizer m(*nll);

m.migrad();
m.hesse();
m.minos();

What can be a problem and how to make RooMinimizer use data error correctly?
Thanks in advance!

[moneta]

Hi,
By default when doing a binned fit, it is assumed that the bin content is distributed according to a Poisson distribution given the observed number of events. If you have a defined bin error for each bin, you should use either a chi2 fit , or a weighted likelihood fit. In this second case you can use the option SumW2Error(true) when calling fitTo. You need to call fitTo and not use createNLL and the RooMinimizer

Cheers

Lorenzo

[Kreikkk]

Thanks a lot! The problem is that the relative error of the estimate that I get is ~100 times less, than the relative error of data and of MC that is used for the fit. That’s why I thaught RooFit doesn’t use error of data. The estimate itself is pretty adequate, but it’s relative error is less than 0.5% which doesn’t seem to be right.
Is it possible for error to fall so significantly during the fit? What could be the problem?

Best regards

[moneta]

Hi,

Yes this can be possible if something is not correct during the minimization, for example the estimated Covariance matrix from the fit could be wring due to numerical errors. I see you also call Minos. Is Minos running correctly ? Are the errors from Minos compatible with the Hessian ones ?

Lorenzo

[Kreikkk]

Hi!

This happened because migrad could not converge. I have changed the upper limit for fit parameters, so it could converge and the error is correct now.

Thank you a lot!