Hi @StephanH ,
Thanks for your reply! I have been working on it, hence the slow reply. Also, while waiting, I found another work-around by Wouter in this link [Simultaneous fits]. I modified my code with this approach and it also worked (I have attached the code). I ran both methods (NLL-method for mine, label-method for yours), I get the exact same answer, which is expected since the only difference is how the data files were handled. Both the NLL-method and label-method converged with no error messages. As I understand it, the NLL-method basically builds the negative log likelihood piece by piece and minimizes that.
Coincidently, I was looking over related forum posts again and I found this post [RooSimultaneous fit and Datasets in RooFit]. In here, Wouter basically seems to give both NLL-method and label-method as solutions, although his implementation of the “label-method” is a bit different (he uses addColumn).
I do have a follow-up question which is related to this issue in a sense that it is about simultaneous/multi-dimension fitting but about a different “how-to” issue. Let me know if I should move it to a different thread but I’ll address it here first.
Now that I have the fit working, I want to make a likelihood ratio plot of one of the quantities (say it is a quantity representing the signal). By likelihood ratio I mean profiled likelihood (found by maximizing the likelihood with respect to all nuisance parameters as a function of the signal) divided by the likelihood at the best fit signal. I would then want to integrate the likelihood ratio plot to 90% to set a 90% upper limit on the quantity using the Feldman-Cousins ranking method. To summarize, the goal now is to produce the likelihood ratio plot and integrate it to 90% using Feldman-Cousins ranking.
My first thought to approach this was the “brute force” method. Since I have my simultaneous PDF, I could feed it the results of the fit to get the denominator of my likelihood ratio by calculating the likelihood for each bin and multiply them. For the numerator, I would do the same but I instead vary my signal quantity (hold the other parameters fixed). I’ll calculate the ratio as I go and since I have the numbers, I can easily store them, make the plots and integrate.
Second thing I found was the FeldmanCousins method in RooStats, in particular this example here: [https://root.cern.ch/root/html/tutorials/roostats/StandardFeldmanCousinsDemo.C.html]. I am even less familiar with RooStats than RooFit so if I could confirm (or not) that this example is what I want, I can give it a go.
I bring this issue back full circle with the two fitting methods, in a sense that if this is the example I want, would it work for the two methods of handling the data?
Thanks,
Ian L.
reduced_2.cpp (8.5 KB)