I am curious to know if this feature explicitly exists in RooFit or if I can do it with the current functionality.
I have a root file that contains two variables, mass and event_number. I fit mass to a model that is the sum of a Breit-Wigner for the mass peak and a first order polynomial for the background. I want to know the corresponding values of the event_number variable for only those events that fit the BW part of the total model (i.e. not the polynomial part).
In other words, if I fit one variable, I want to know the corresponding values of another variable, for those events corresponding to only a part of the fit.
You cannot determine the signal event numbers from a fit, only the probability that a given event is signal.
For simplicity, consider a binned fit. In a given bin, you have some contributions from signal and some from background. Your fit can estimate what fraction of the events in a given bin are signal or background, but there is no way to disentangle which event is which–the only information in a given bin is the total number of events.
But you can extract from the fit the probability that an event in a given bin is signal or background. This is the principle behind background-subtraction techniques; the probability that a given event is signal in one variable allows you to estimate the signal distribution in another variable by weighting the events by their signal probability*. So if you wanted to see the distribution of the transverse momentum of signal events, as determined from your mass fit, you could do that using some background subtraction technique. At least one of these is built into RooStats: the sPlot method.
To summarize, I believe the answer to your question is no, in principle it cannot be done. The best you can do is determine a probability that a given event is signal or background. However, if what you really want to do is determine the signal distribution of some variable based on a mass fit, that can be done using background subtraction.
I may have overlooked something in this explanation, and @moneta can correct me, but I hope this at least helps.
*Not all background subtraction techniques use the probability, strictly defined. The sPlot method makes use of something called s-weights, for example.