Errors from a model


when doing a extended likelihood fit using a given model (for example using a given Pdf of signal and of background, RooGaussian for example or any other), one obtain a given number of signal and background (where only Nsignal and Nbackground are floatted, all the others are fixed)

Are the errors of the fit including the errors of the modelisation of the Pdf ?

-If “yes”, how to separate the statistical error with these “systematic” modelisation of the Pdf ?

-If “no”, how to deal with computing the errors of the modelisation, apart from doing it “by hand” by making a loop to vary each parameter of the Pdf by its uncertainty ?

Would the RooKeysPdf (which is non parametrised) be a special case ?

thank you for your time


Given an extended pdf of the form NsigFsig(X;A) +NbkgFbkg(X;B)
you have a variety of options

If you fit with parameters A,B constant, the errors in Nsig, Nbkg will not reflect any modeling uncertainty in their error.

If, on the other hand you fit with parameters A,B floating, your errors in Nsig,Nbkg will reflect the propagated errors on the fitted parameters A and B as well

So you can choose which scenario you want.