Wrong fit errors when fitting weighted data set

[ekpkit]

Hello,

I am performing an unbinned ML fit of weighted MC, where a RooAbsPdf is fitted to a weighted RooDataSet and the average weight is small with a value of about 1/333. The command looks like this:

pdf.fitTo(data, RooFit::Minos(kTRUE));

The resulting errors on the fitted pdf parameters are much too large. When fitting the unweighted data set the parameter errors drop more than an order of magnitude. This problem also occurs when adding the option RooFit::SumW2Error(kTRUE) or without using Minos.
I am using root version 5.34.10.
The same issue has also been reported in this unresolved post

Fit errors with weighted events

Hello, I fill a mass distribution histogram (the left histogram in the attachements, yellow color - fit function). After that I fit it and check the extracted number of events. If I don’t weight my events I obtain: N_D0 = 209 +/ 19 <= both the value and its error are fine! If events were weighted with weight>1: N_D0 = 4444 +/- 86 <= the value is good, the error is too small If events were weighted with weight<1: N_D0 = 6 +/- 3 <= the value is good, the error is too big O…

Is there a way to fit a weighted data set and obtain a correct error estimate on the fitted parameters ?

Thank you,
Andreas

[moneta]

Hi,
Can you please post the code reproducing this problem ? Bare in mind that using weights is just an approximation, which in some case it might not be correct, in particular if the weights are very different from each others (i.e. some very small and some very large). In that case one would need the full weight distribution and/or have the right model.

Best Regards

Lorenzo