Multileptons To-Do-List

Analysis Design

Define the Normalization Regions: Z+jets has a smaller contribution to the SR than e.g. single top. Should we include a single top Normalization Region? Should we let go of the Z Normalization Region? Also ZZ is larger than Z+jets. But keep in mind: not only the contribution in the SR matters but also how much we trust the particular MC with respect to the MC from other processes and how much the contamination in other regions is (apart from SR).
- First idea: study single top and ZZ but likeli don't use them and use Z only for contamination in WZ (WZ is contaminated by Z+LF while SR is contaminated by Z+HF).
How do we make sure the contribution from multijets to the normalization regions is negligible?
How do we get the data driven estimate for the BG from conversions (which are not modeled in MC)?
Define a single top and a ZZ normalization region.

What sort of jets to be used in Normalization Regions (baseline, overlap removal, ...)?
Nice output format to be used as input to the fit machinery.

What to do about the factorization scale, renormalization scale, ..., bla scale uncertainties?

Write the code to perform the fit and to handle the estimates and their uncertainties.
How to get the total background expectation in the Signal Region and its uncertainty?
How to treat BG with large uncertainties? DONE.
How to treat asymmetric uncertainties? DONE. more info: [[Francesco's talk][1]]
How to constrain the poisson parameters to values > 0? 1/2 DONE.
How to deal with the uncorrelated systematics: One common nuisance parameter: but how to make sure it's uncertainty is not reduced too much in the fit? -> As many nuisance parameters. This will weaken the ability of the fit to the nuisance params from the data.
Any need to introduce conditionals into the likelihood? I.e. Po(ns;s,b)P(nb;b) -> Po(ns;s|b)P(nb;b)? -> Alex
pythia vs. herwig type of systematics (the bad in Wouter's text): even the experts aren't sure what to do. Two strategies: Calculate the limits with both and use the looser one. Or use a toy experiment approach. Maybe like this: If it's dominating: Publish both results, this tells the theorists that they need to work more. If it's negligible: just neglect it. If it's of the order the other uncertainties: you're screwed. Use maybe one up the approaches described (toys, looser one, ...).

Publish a plot with the different BG estimates in the different signal grid points? (BG estimate will depend on input signal.)
How to get a goodness-of-fit measure? -> look at Fred James' book.