Results of combination Higgs toy combination, within and across experiments, with RooStats

16
Results of combination Higgs toy Results of combination Higgs toy combination, within and across combination, within and across experiments, with RooStats experiments, with RooStats Gr Gr égory Schott égory Schott Institute for Experimental Nuclear Physics of the Karlsruhe Institute of Technology on behalf of the ATLAS-CMS Higgs groups and statistics forum ATLAS-CMS statistics meeting, 5 th July 2010

description

Results of combination Higgs toy combination, within and across experiments, with RooStats. Gr égory Schott Institute for Experimental Nuclear Physics of the Karlsruhe Institute of Technology on behalf of the ATLAS-CMS Higgs groups and statistics forum. - PowerPoint PPT Presentation

Transcript of Results of combination Higgs toy combination, within and across experiments, with RooStats

Page 1: Results of combination Higgs toy combination, within and across experiments, with RooStats

Results of combination Higgs toy Results of combination Higgs toy combination, within and across combination, within and across

experiments, with RooStatsexperiments, with RooStats

GrGrégory Schottégory Schott

Institute for Experimental Nuclear Physicsof the Karlsruhe Institute of Technology

on behalf of the ATLAS-CMS Higgs groups and statistics forum

ATLAS-CMS statistics meeting, 5th July 2010

Page 2: Results of combination Higgs toy combination, within and across experiments, with RooStats

2Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Introduction I will present the results of ATLAS, CMS and the combination

summary of the results obtained by Kyle Cranmer and myself thanks to Kyle for the FC, PL and some Hybrid results in case of PL we both get results that of course agree very well

This is on-going work; while a large part of the results are available, not everything has been done yet and not everything has been validated yet

Toy study (made with mockup numbers), H→WW at 160 GeV with 1fb-1

don't draw physics conclusion of the results

Significance and upper limit computed with RooStats (for a given data) different approaches tested as recommanded by the statistics forums

Validation of RooStats (ROOT 5.27.04, June 2009) Conclusion and discussions

Page 3: Results of combination Higgs toy combination, within and across experiments, with RooStats

3Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

The combined model Combination of H→WW→ee, , e analyses Parameter of interest: ratio of cross-sections, r = = / SM

CMS analysis: 3 observables, 37 nuisance parameters Lognormal distribution of nuisance parameters Uncertainty on control region measurement in the systematics

ATLAS analysis: 9 observables, 12 nuisance parameters Truncated gaussian distribution of nuisance parameters Control regions measurements included in the analysis

Combined analysis: common parameter interest (thus, 100% correlated) no other correlation (yet) across the experiments

Concretely, all the information for the combination is shared via 2 ROOT files (for ATLAS & CMS) containing RooFit/RooStats workspaces

graphical representation of the objects in the combined

model

Page 4: Results of combination Higgs toy combination, within and across experiments, with RooStats

4Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Toy-data Data used for significance:

generated from the model in the hypothesis that signal is present

Data used for exclusion: generated from the model in the background-only hypothesis

Check the backups if you're interested by the number of observed events assumed (or the likelihood functions)

Page 5: Results of combination Higgs toy combination, within and across experiments, with RooStats

5Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Profiled likelihood results

significance: reasonably good estimate if large enough

95% CL UL taken for -log (r) = 1.921 assumption of Wilks's asymptotia that's not always valid!

with systematics

r UL

ATLAS 0 0.79CMS 0 0.28

COMBI 0 0.25

with systematics

r signif.

ATLAS 1.13 2.70CMS 0.98 4.89

COMBI 1.02 5.58

^

^

Page 6: Results of combination Higgs toy combination, within and across experiments, with RooStats

6Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Hybrid Frequentist-Bayesian

run for the 3 test statisticsmentionned in the previoustalk

need to study the advantageand inconvenient of the 3approaches

test statistics significance (no syst.)

significance (with syst.)

ATLASQLEP

QTEV

()

3.78-

3.07 ± 0.012.8 ± 0.1

-

CMSQLEP

QTEV

()

6.22 ± 0.02--

4.77 ± 0.02> 4.6

4.3 ± 0.1

COMBIQLEP

QTEV

()

---

> 4.6> 3.5

-

CMS, 477k toys

QLEP()QTEV

B-toysSB-toys

data

Page 7: Results of combination Higgs toy combination, within and across experiments, with RooStats

7Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

CMS model with 'LandS'

Hybrid test statistics distributions

computing the p-value for significancein this approach is challenging:

speed improvements would be useful or use importance sampling techniques

CMS distribution (and results previousslide) made with a RooFit-independenttool

QLEPQLEP

Page 8: Results of combination Higgs toy combination, within and across experiments, with RooStats

8Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Summary of upper limits

technique test stat rule sampling UL ATLAS UL CMS UL COMBIFeldman-Cousins (no

syst.)() CLS+B

toys 0.69 ± 0.05 - -

Profile LR (Wilks) () CLS+Basymptotic 0.79 0.28 0.25

Feldman-Cousins++ () CLS+Btoys 0.78 ± 0.05 0.26 ± 0.02 0.23 ± 0.02

Hybrid QLEP CLStoys ~ 0.68 0.29 ± 0.03

(LandS)-

Hybrid QLEP CLS+Btoys ~ 0.61 - -

Bayesian n/a, flat prior on r MCMC* 0.72 0.31 0.28

ATLAS

r

post

erio

r (r |

 dat

a)

CMS

with BAT [ Caldwell, Kollar, Kroeninger, Comp.Phys.Com. 180, 2197 (2009) ]

95% CL UL

95% CL upper limits: results with systematics (except if indicated otherwise)

Page 9: Results of combination Higgs toy combination, within and across experiments, with RooStats

9Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Validations Some validation of the analysis performed with independent software

(LandS, M. Chen, CMS) Since 'LandS' was developped specifically for this analysis, it is

faster than RooStats which is a general tool showing that RooStats still need to improve

table of compared results: we still need to work on the missing parts

PL and toy-MC results using an earlier version of the ATLAS model were validated with an independent code (Hao Liu, et. al.)

further on-going work on validating Bayesian limits: S. Schmitz (CMS) currently using BAT, RooStats providing another MCMCCalculator

Page 10: Results of combination Higgs toy combination, within and across experiments, with RooStats

10Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Conclusion and outlook A good start, work is on-going on

Lessons learned: we achieved in a short time scale to combine the Atlas and

CMS analyses -> proof of principle that we can do that proof of principle we can compare different statistical methods

to one another still some weaknesses identified that RooStats need to improve we also need to continue the validations of RooStats

still need to understand those results discussion has started on statistical methods we want to use in common

proper treatment of correlations across experiments will require planning and thoughtful parametrization of systematics

activity will continue (other masses, other channels, more thorough correlation studies, ...)

Page 11: Results of combination Higgs toy combination, within and across experiments, with RooStats

11Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

End of talk−

Backup slides

Page 12: Results of combination Higgs toy combination, within and across experiments, with RooStats

12Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Toy-data Data used for significance:

generated from the model in the hypothesis that signal is present

Data used for exclusion: generated from the model in the background-only hypothesis Observables:

1) nobs_bin1 = 4 2) nobs_bin2 = 2 3) nobs_bin3 = 3 4) obs_s_em_0j = 19 5) obs_ww_em_0j = 63 6) obs_tt_em_0j = 3814 7) obs_wj_em_0j = 123 8) obs_s_ee_0j = 6 9) obs_ww_ee_0j = 14 10) obs_wj_ee_0j = 53 11) obs_s_mm_0j = 16 12) obs_ww_mm_0j = 26

Observables: 1) nobs_bin1 = 15 2) nobs_bin2 = 7 3) nobs_bin3 = 13 4) obs_s_em_0j = 36 5) obs_ww_em_0j = 52 6) obs_tt_em_0j = 3650 7) obs_wj_em_0j = 143 8) obs_s_ee_0j = 9 9) obs_ww_ee_0j = 5 10) obs_wj_ee_0j = 49 11) obs_s_mm_0j = 18 12) obs_ww_mm_0j = 28

Page 13: Results of combination Higgs toy combination, within and across experiments, with RooStats

13Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

slide from K. Cranmer, 01.07.2010

ATLAS likelihood model

Page 14: Results of combination Higgs toy combination, within and across experiments, with RooStats

14Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

CMS likelihood model

slide from A. Korytov, 24.06.2010

Page 15: Results of combination Higgs toy combination, within and across experiments, with RooStats

15Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Profile likelihood (zoom)

Page 16: Results of combination Higgs toy combination, within and across experiments, with RooStats

16Grégory Schott - ATLAS-CMS statistics meeting - 01.07.2010

Bayesian upper limits

r

posterior (r | data)ATLAS

posterior (r | data)COMBI

r