I just get very different fit results, i.e. the number of signal events could be 30 VS. 60, when RooNDKeysPdf or Roo2DKeysPdf are used, respectively. Here all the data sets, options, fit methods, etc. are set exactly same except the different PDF methods are used.
I read from the document that the Roo2DKeysPdf is superseded by the RooNDKeysPdf, so I assume the results should be same. Anyone know what happens? Is there any possible mistake I made?
I can only say that those two are completely different implementations. I would not be surprised to see differences, but a factor 2 sounds too much, unless your background is large and the signal is something like
30 ± 30 and 60 ± 35.
Yes, the background is high, but the signals are like 30 +/- 12 and 60 +/- 13. So the difference is about 1.7 sigma, that seems not so terrible as I think before. Thanks for this reminding!
But the difference is still too large to be considered as systematic uncertainty, I guess. Which one shall I use? Or any other suggestion?
It seems RooNDKeysPdf has considered the correlation between different dimensions, that has been ignored in the Roo2DKeysPdf. So my current conclusion is RooNDKeysPdf is more reliable.