Are constant terms neglected in a binned pdf in Roofit?

Dear all,

I would like to know whether all constant terms which do not depend on the fit parameters are considered in the expression of a binned likelihood in Roofit (using RooAbsPdf::createNLL() or RooAbsPdf::fitTo) ?

Let me give you a bit of context. I have a Time-Of-Flight distribution whose generating process depends on many experimental parameters. The distribution presents tails on both side and is Gaussian-like in the center of the distribution. In other words the generating process is unknown and I am faced with the problem of having to choose an adequate fitting pdf, which I also want to be parsimonious. Additionnally, the experimental data are recorded using an MultiChannelAanlyzer whose output is an histogram. Thus, I model my data using a binned likelihood in RooFit using different pdf (Gaussian, Gaussian-Exp convolutions…). I would like to use the nll from RooFit to compute AIC differences between the different models, all within a set of candidate models that I defined a priori, in order to study which one would be closer to the true process generating my data. I would like to avoid having unnecessary fitting parameters in my approximating model used for the analysis. It seems that the AIC approach is well suited for this kind of problems. If the data suggest many models are equally good approximations to the true unknown model I could use AIC weights to average over different models. If someone was already faced with similar situations maybe he/she can also comment on the aforementionned approach?

Thank you very much in advance.

Regards,

MM

Hi,

Constant terms are often omitted from the likelihood and if you are using this technique (AIC) based on the ML value, this could be certainly a problem.
Normally we always use likelihood ratio values and therefore the cancel terms cancel out.
The here might be some options in createNLL, but I presume some terms depending on n could be in some case omitted. I would reccomend you to verify correctly the log-likelihood obtained in a simple case
to be sure it is correct.
I have to admit I have no experience with AIC to compare different models. If you have a publication showing this for high energy physics I would be interested to look at it.

Best Regards

Lorenzo

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.