Tom Junk - SFU Department of Statistics and Actuarial...
Transcript of Tom Junk - SFU Department of Statistics and Actuarial...
BanffChallenge2
TomJunkFermilab
BIRSSta6s6csinHEPWorkshopJuly2010
1TomJunkBanffChallenge2
TomJunkBanffChallenge2 2
CommonStandardsofEvidence
Physicistsliketotalkabouthowmany“sigma”aresultcorrespondstoandgenerallyhavelessfeelforp‐values.
Thenumberof“sigma”iscalleda“z‐value”andisjustatransla6onofap‐valueusingtheintegralofonetailofaGaussian
Double_tzvalue=‐TMath::NormQuan6le(Double_tpvalue)
1σ⇒15.9%
Tip:mostphysiciststalkaboutp‐valuesnowbuthardlyusethetermz‐value
Folklore:95%CL‐‐goodforexclusion3σ:“evidence”5σ:“observa6on”Someargueforamoresubjec6vescale.
€
pvalue =1− erf zvalue / 2( )( )
2z-value (σ) p-value
1.0 0.159
2.0 0.0228
3.0 0.00135
4.0 3.17E-5
5.0 2.87E-7
BanffChallenge2Problem#1–StackedplotshownHEP‐style
3TomJunkBanffChallenge2
• Observeddatashownaspointswithsqrt(n)errorbars(yes,theconven6on’scrazybutthat’sthewaywedoit.)• Signalpredic6onshownstackedontopofthebackgroundpredic6on.UsefulbecausewecancomparethethedatawithH0andH1withjustoneplot.
DiscriminantVariable
Even
ts
nbackground=10000nsignal=210ndata=9815
TomJunkBanffChallenge2 4
€
−2lnQ ≡ LLR ≡ −2ln L(data | s+ b, ˆ θ )
L(data |b, ˆ ˆ θ )
Problem1,nosystema6cuncertainty1MillionsimulatedexperimentsforH0and1MillionsimulatedexperimentsforH1
Nuisanceparametersalwaysattheirnominalvalues
p‐value=5.95x10‐4z‐value=3.24‐2lnQobs=1.98
hatsdon’tmaierheresincethere’snofit.
TomJunkBanffChallenge2 5
€
−2lnQ ≡ LLR ≡ −2ln L(data | s+ b, ˆ θ )
L(data |b, ˆ ˆ θ )
Problem1,withsystema6cuncertainty1MillionsimulatedexperimentsforH0and1MillionsimulatedexperimentsforH1
p‐value=1.91x10‐5z‐value=4.11‐2lnQobs=‐15.43
nowdotwofitspersimulatedexperiment‐‐fitforallnuisanceparameters,rateandshape
Eachpseudoexperimentgetsrandomlyfluctuatednuisanceparameters(“prior‐predic6veensemble”)
TomJunkBanffChallenge2 6
ATricktoUseonly1MSimulatedH0Experiments
• Fitthedistribu6onof‐2lnQ(H0)toasumoftwoGaussians–canintegratethatanaly6callywitherf’s.• Needtocheckfitquality.Arealjobwouldbetoes6matetheuncertainty(extrapola6onuncertaintyifneedbe).• Forarealdiscoveryofapar6cle,we’djustusetheneededCPU.Maybethefiiergetsstuckoncein1x107experiments–needtoknowthat.
ThesumoftwoGaussiansisagoodapproxima6onherebutapooroneiftheproblemismorediscrete–onebin,forexample,orlotsoflows/bbinsandoneveryhighs/bbinwithjustafewexpectedeventsinit.
TomJunkBanffChallenge2 7
AProblemwithProblem#2
DiscriminantVariable
Even
ts
nbackground=10000nsignal=200ndata=98439209dataeventsinthevisiblepartofthehistogram
634observedeventsoutof9843areintheupperoverflowbin(!)(~6.44%ofthem)Backgroundtemplatemodelsthis.
IdiscouragetheuseofROOT’sover‐andunderflowbinsforseveralreasons:
1)Theyarenot(usually)ploied.Hardtovalidatethemifyoucannotseethem2)TheyarenotincludedinTH1::Integral()orinfSumwwhendumped.Soscalingbydividingbytheintegralandmul6plyingbythedesiredyieldwon’tgetitright.
rootaccumulatesentriesbeyondthehistogramedgesinunderflowandoverflowbins,andtreatsthemasspecialbins(why?)Sugges6ontoallstudents:constrainallselecteddatatobeinvisiblebins(maxandmin). Problems1and3havenoentriesinthe
underfloworoverflowbins.
SoIsolvedaproblemthatisslightlydifferentandpossiblymoreinstruc6ve.
TomJunkBanffChallenge2 8
Problem2
€
−2lnQ ≡ LLR ≡ −2ln L(data | s+ b, ˆ θ )
L(data |b, ˆ ˆ θ )
Nosystema6cs:‐2lnQ=17.46z‐value=0.20p‐value=0.42
Withsystema6cs:‐2lnQ=‐17.33z‐value=4.1p‐value=1.75x10‐5
DiscriminantVariable
Even
ts
GOFnotevaluatedwithoutsystema6cs–preiypoorthough.Showsthattheno‐systema6csinterpreta6onisincorrect.
nbackground=10000nsignal=200ndata=9843
TomJunkBanffChallenge2 9
Problem2’sFit
Notperfectonthetail,probablyjustneedtorunmorepseudoexperiments
TomJunkBanffChallenge2 10TomJunkBanffChallenge2 10
Problem3
€
−2lnQ ≡ LLR ≡ −2ln L(data | s+ b, ˆ θ )
L(data |b, ˆ ˆ θ )
Nosystema6cs:‐2lnQ=‐43.1z‐value=7.3p‐value=1.4x10‐13
Withsystema6cs:‐2lnQ=‐21.2z‐value=4.44p‐value=4.5x10‐6
DiscriminantVariable
Even
ts
nbackground=80nsignal=72ndata=134
TomJunkBanffChallenge2 11
CombiningProblems1+2+3Jointfits–correlatedsystema6cuncertain6esinarealproblem.Wearetoldtodecorrelatethenuisanceparametersbetweenchannels.
Nosystema6cs:
‐2lnQcomb=‐2lnQ1‐2lnQ2‐2lnQ3
Withsystema6cs–spoiledabitbythedifferentfits,ifnuisanceparametersarecorrelated.Inthiscasethesumrules6llworksbecausealldataandallnuisanceparametersareindependent.
GOFispoorforbothhypotheses–seeprob.2.Largesensi6vity.Nosystema6cscanruleoutbothH0andH1.
Nosystema6cs:‐2lnQ=‐23.7z‐value=7.0p‐value=1.2x10‐12
Withsystema6cs:‐2lnQ=‐53.96z‐value=7.5p‐value=3.6x10‐14
TomJunkBanffChallenge2 12
Combina6onSignificancesareaBitofanExtrapola6onwithjust1MSimulatedOutcomes
TomJunkBanffChallenge2 13
Es6matesofSensi6vity• Well,1Millionsimulatedexperimentsisn’tenough–canget1Millionmoreincombina6onbyaddingthe‐2lnQ’sfrom1,2,and3’stogether.
• Wilks’sTheoremprobablyisagoodapproxima6onheretoo.
• Importancesamplingcouldbeusedtoimproveprecisionintails
• Fordiscovery,we’duserealCPUasthesysteam6cswillbecorrelatedandtheremaybeasinglebinaddingadiscretecomponenttoit.
• Ourfavoritesensi6vityes6mate:pmed,signalisthemedianexpectedp‐valueassumingasignalispresent.1Mpseudoexperimentsnotquiteenough.
• Astand‐in:the“o‐value”(namedbytheCDFKarlsruhesingletopteam,butwe’duseditbefore.
• Medianscanbeusedinsteadofmeans,andtheσ’sareRMS’softhe‐2lnQdistribu6on.
€
o − value =−2lnQ bkg − −2lnQ s+b( )
σ bkg2 +σ s+b
2
TomJunkBanffChallenge2 14
€
o − value =−2lnQ bkg − −2lnQ s+b( )
σ bkg2 +σ s+b
2
Problem <‐2lnQ>b RMSb <‐2lnQ>s+b RMSs+b o‐value
1nosyst 41.9 12.3 ‐46.3 14.3 4.7
2nosyst 19.1 8.6 ‐20.0 9.1 3.1
3nosyst 49.5 11.9 ‐68.9 19.5 5.2
123nosyst 110.6 19.1 ‐135.2 25.8 7.6
1syst 21.2 8.9 ‐28.1 13.5 3.0
2syst 12.8 6.6 ‐16.7 9.3 2.6
3syst 21.6 9.5 ‐25.6 12.2 3.0
123syst 55.5 14.6 ‐70.3 20.5 5.0
Toagoodapproxima6on,o‐valuesaddinquadratureforthecombina6on.Trueforthisproblem,butnottrueingeneral.
TomJunkBanffChallenge2 15
BackupMaterial
TomJunkBanffChallenge2 16
Fitting Nuisance Parameters to Reduce Sensitivity to Mismodeling
Means of PDF’s of -2lnQ very sensitive to background rate estimation.
Still some sensitivity in PDF’s residual due to prob. of each outcome varies with bg estimate.