I produced a model in RooStats, and I would like to set a limit on a single observable.
Some indipendent uncertainty sources have been included, and a set of nuisance parameters has been defined.
I would like to compare the results given by different calculators, study the impact of uncorrelated uncertainty sources, possibly turning on/off some of them.
I tested my model using a ProfileLikelihoodCalcuator, and I got some numbers.
For other calculators I just used StandardHypoTestInverterDemo.C out of the box. I tested
- Hybrid Calculator (CLs, 1-Sided PL)
- Asymptotic Calculator (CLs, 2-Sided PL)
- Frequentist Calculator (CLs+b, 2-Sided PL -> FeldmanCousins)
AsymptoticCalculator gave an upper limit very close to those produced by ProfileLikelihood Calculator; Hybrid Calculator limit looks a bit tighter.
However FC strategy fails when the largest systematic uncertainties sources are on (plots attached refer to the same Workspace).
Why (just) a FC approach should fail in a such drmatical way? Have you got an opinion about that?