Strange fit behaviour

[cmcnicol]

Hi,

I’m trying to fit a distribution using RooFit and the resulting fit looks strange. However, if I fit the same function using ROOT, I get a reasonably good fit.

This distribution comes from combining several samples with different normalisation weights and I was told that the default fit method, fitTo() can give poor results but I should try the chi2FitTo() method. (If this is true and is known, could someone explain why?)

If I use the values given by the ROOT fit, (“fixed”) then I get a seemingly good fit, but allow those parameters to float and the result is a very poor fit.

I’m also confused because if I generate an Asimov dataset from the created PDF, and then try to fit this I still see the strange effect. I would have thought that the PDF would fit the Asimov dataset perfectly (within statistical errors).

I’ve attached a script which should replicate my problem.

To explain the attached plot. I create a PDF and generate a dataset from this. These are the datapoints and red line which makes sense. The blue line is the result of the fit . I notice a similar but less severe problem when using the chi2FitTo() method.

Can someone tell me if I’m doing something wrong or inappropriate and how to fix it?

Thanks,
Chris

strangeFittingScript.py (3.3 KB)

SlicedZ_uniformBins.root (8.3 KB)

[moneta]

Hi

I would not use chi2fit when fitting tails like you because it is biased due to the not correct handling of empty bins or bins with poor statistics. However a likelihood fit (weighted) should give you the correct result.

I cannot run your code because I miss the input data file

Best Regards

Lorenzo

[cmcnicol]

Hi Lorenzo,

Thanks for your reply, sorry for my delayed response. I’ve attached the file to my original post. (Sorry for not attaching it before, I wasn’t sure who would have access to it on /afs .

Yes, I’d heard about the problem with empty bins. However, I still see a problem with the likelihood fit. (Unless there’s something special I need to do regarding the weights?)

Thanks for your help,
Chris