Upload
gerry
View
13
Download
0
Embed Size (px)
DESCRIPTION
Some Topics in Statistical Data Analysis. Invisibles School IPPP Durham July 2013. Glen Cowan Physics Department Royal Holloway, University of London [email protected] www.pp.rhul.ac.uk/~cowan. TexPoint fonts used in EMF. Read the TexPoint manual before you delete this box.: A A A A. - PowerPoint PPT Presentation
Citation preview
G. Cowan Invisibles 2013 / Statistical Data Analysis 1
Some Topics in Statistical Data Analysis
Invisibles SchoolIPPP DurhamJuly 2013
Glen CowanPhysics DepartmentRoyal Holloway, University of [email protected]/~cowan
G. Cowan Invisibles 2013 / Statistical Data Analysis 2
Outline
Lecture 1: Introduction and basic formalismProbability, statistical tests, parameter estimation.
Lecture 2: Discovery and LimitsAsymptotic formulae for discovery/limitsExclusion without experimental sensitivity, CLs, etc.Bayesian limitsThe Look-Elsewhere Effect
G. Cowan Invisibles 2013 / Statistical Data Analysis Lecture 1 page 3
Some statistics books, papers, etc. J. Beringer et al. (Particle Data Group), Review of Particle Physics, Phys. Rev. D86, 010001 (2012); see also pdg.lbl.gov sections on probability statistics, Monte Carlo
G. Cowan, Statistical Data Analysis, Clarendon, Oxford, 1998see also www.pp.rhul.ac.uk/~cowan/sda
R.J. Barlow, Statistics: A Guide to the Use of Statistical Methodsin the Physical Sciences, Wiley, 1989
see also hepwww.ph.man.ac.uk/~roger/book.html
L. Lyons, Statistics for Nuclear and Particle Physics, CUP, 1986
F. James., Statistical and Computational Methods in Experimental Physics, 2nd ed., World Scientific, 2006
S. Brandt, Statistical and Computational Methods in Data Analysis, Springer, New York, 1998
G. Cowan Invisibles 2013 / Statistical Data Analysis 4
A definition of probability Consider a set S with subsets A, B, ...
Kolmogorovaxioms (1933)
Also define conditional probability:
G. Cowan Invisibles 2013 / Statistical Data Analysis 5
Interpretation of probabilityI. Relative frequency
A, B, ... are outcomes of a repeatable experiment
cf. quantum mechanics, particle scattering, radioactive decay...
II. Subjective probabilityA, B, ... are hypotheses (statements that are true or false)
• Both interpretations consistent with Kolmogorov axioms.• In particle physics frequency interpretation often most useful, but subjective probability can provide more natural treatment of non-repeatable phenomena: systematic uncertainties, probability that Higgs boson exists,...
G. Cowan Invisibles 2013 / Statistical Data Analysis 6
Bayes’ theoremFrom the definition of conditional probability we have
and
but , so
Bayes’ theorem
First published (posthumously) by theReverend Thomas Bayes (1702−1761)
An essay towards solving a problem in thedoctrine of chances, Philos. Trans. R. Soc. 53(1763) 370; reprinted in Biometrika, 45 (1958) 293.
G. Cowan Invisibles 2013 / Statistical Data Analysis 7
Frequentist Statistics − general philosophy In frequentist statistics, probabilities are associated only withthe data, i.e., outcomes of repeatable observations.
Probability = limiting frequency
Probabilities such as
P (WIMPs exist), P (0.298 < Ωm < 0.332),
etc. are either 0 or 1, but we don’t know which.The tools of frequentist statistics tell us what to expect, underthe assumption of certain probabilities, about hypotheticalrepeated observations.
The preferred theories (models, hypotheses, ...) are those for which our observations would be considered ‘usual’.
G. Cowan Invisibles 2013 / Statistical Data Analysis 8
Bayesian Statistics − general philosophy In Bayesian statistics, interpretation of probability extended todegree of belief (subjective probability). Use this for hypotheses:
posterior probability, i.e., after seeing the data
prior probability, i.e.,before seeing the data
probability of the data assuming hypothesis H (the likelihood)
normalization involves sum over all possible hypotheses
Bayesian methods can provide more natural treatment of non-repeatable phenomena: systematic uncertainties, probability that Higgs boson exists,...
No golden rule for priors (“if-then” character of Bayes’ thm.)
G. Cowan Invisibles 2013 / Statistical Data Analysis 9
Hypothesis testingA hypothesis H specifies the probability for the data, i.e., the outcome of the observation, here symbolically: x.
x could be uni-/multivariate, continuous or discrete.
E.g. write x ~ f (x|H).
x could represent e.g. observation of a single particle, a single event, or an entire “experiment”.
Possible values of x form the sample space S (or “data space”).
Simple (or “point”) hypothesis: f (x|H) completely specified.
Composite hypothesis: H contains unspecified parameter(s).
The probability for x given H is also called the likelihood ofthe hypothesis, written L(x|H).
G. Cowan Invisibles 2013 / Statistical Data Analysis 10
Definition of a (frequentist) hypothesis testConsider e.g. a simple hypothesis H0 and alternative H1.
A test of H0 is defined by specifying a critical region w of thedata space such that there is no more than some (small) probability, assuming H0 is correct, to observe the data there, i.e.,
P(x w | H0 ) ≤
Need inequality if data arediscrete.
α is called the size or significance level of the test.
If x is observed in the critical region, reject H0.
data space Ω
critical region w
G. Cowan Invisibles 2013 / Statistical Data Analysis 11
Definition of a test (2)But in general there are an infinite number of possible critical regions that give the same significance level .
Roughly speaking, place the critical region where there is a low probability (α) to be found if H0 is true, but high if the alternativeH1 is true:
G. Cowan Invisibles 2013 / Statistical Data Analysis 12
Type-I, Type-II errors Rejecting the hypothesis H0 when it is true is a Type-I error.
The maximum probability for this is the size of the test:
P(x w | H0 ) ≤
But we might also accept H0 when it is false, and an alternative H1 is true.
This is called a Type-II error, and occurs with probability
P(x S w | H1 ) =
One minus this is called the power of the test with respect tothe alternative H1:
Power =
G. Cowan Invisibles 2013 / Statistical Data Analysis 13
Rejecting a hypothesisNote that rejecting H0 is not necessarily equivalent to thestatement that we believe it is false and H1 true. In frequentiststatistics only associate probability with outcomes of repeatableobservations (the data).
In Bayesian statistics, probability of the hypothesis (degreeof belief) would be found using Bayes’ theorem:
which depends on the prior probability (H).
What makes a frequentist test useful is that we can computethe probability to accept/reject a hypothesis assuming that itis true, or assuming some alternative is true.
G. Cowan Invisibles 2013 / Statistical Data Analysis 14
Defining a multivariate critical region
Each event is a point in x-space; critical region is now definedby a ‘decision boundary’ in this space.
What is best way to determine the decision boundary?
WH1
H0
Perhaps with ‘cuts’:
G. Cowan Invisibles 2013 / Statistical Data Analysis 15
Other multivariate decision boundaries
Or maybe use some other sort of decision boundary:
WH1
H0
WH1
H0
linear or nonlinear
G. Cowan Invisibles 2013 / Statistical Data Analysis 16
Test statisticsThe boundary of the critical region for an n-dimensional dataspace x = (x1,..., xn) can be defined by an equation of the form
We can work out the pdfs
Decision boundary is now a single ‘cut’ on t, defining the critical region.
So for an n-dimensional problem we have a corresponding 1-d problem.
where t(x1,…, xn) is a scalar test statistic.
G. Cowan Invisibles 2013 / Statistical Data Analysis 17
Constructing a test statisticHow can we choose a test’s critical region in an ‘optimal way’?
Neyman-Pearson lemma states:
For a test of size α of the simple hypothesis H0, to obtainthe highest power with respect to the simple alternative H1,choose the critical region w such that the likelihood ratio satisfies
everywhere in w and is less than k elsewhere, where k is a constantchosen such that the test has size α.
Equivalently, optimal scalar test statistic is
N.B. any monotonic function of this is leads to the same test.
G. Cowan Invisibles 2013 / Statistical Data Analysis 18
Testing significance / goodness-of-fitSuppose hypothesis H predicts pdf observations
for a set of
We observe a single point in this space:
What can we say about the validity of H in light of the data?
Decide what part of the data space represents less compatibility with H than does the point less
compatiblewith H
more compatiblewith H
(Not unique!)
G. Cowan Invisibles 2013 / Statistical Data Analysis 19
p-values
where (H) is the prior probability for H.
Express level of agreement between data and H with p-value:
p = probability, under assumption of H, to observe data with equal or lesser compatibility with H relative to the data we got.
This is not the probability that H is true!
In frequentist statistics we don’t talk about P(H) (unless H represents a repeatable observation). In Bayesian statistics we do; use Bayes’ theorem to obtain
For now stick with the frequentist approach; result is p-value, regrettably easy to misinterpret as P(H).
G. Cowan 20
Significance from p-valueOften define significance Z as the number of standard deviationsthat a Gaussian variable would fluctuate in one directionto give the same p-value.
1 - TMath::Freq
TMath::NormQuantile
Invisibles 2013 / Statistical Data Analysis
E.g. Z = 5 (a “5 sigma effect”) corresponds to p = 2.9 × 10.
G. Cowan Invisibles 2013 / Statistical Data Analysis 21
The significance of an observed signalSuppose we observe n events; these can consist of:
nb events from known processes (background)ns events from a new process (signal)
If ns, nb are Poisson r.v.s with means s, b, then n = ns + nb
is also Poisson, mean = s + b:
Suppose b = 0.5, and we observe nobs = 5. Should we claimevidence for a new discovery?
Give p-value for hypothesis s = 0:
G. Cowan page 22
Distribution of the p-valueThe p-value is a function of the data, and is thus itself a randomvariable with a given distribution. Suppose the p-value of H is found from a test statistic t(x) as
Invisibles 2013 / Statistical Data Analysis
The pdf of pH under assumption of H is
In general for continuous data, under assumption of H, pH ~ Uniform[0,1]and is concentrated toward zero for the relevant alternatives. pH
g(pH|H)
0 1
g(pH|H′)
G. Cowan page 23
Using a p-value to define test of H0
So the probability to find the p-value of H0, p0, less than is
Invisibles 2013 / Statistical Data Analysis
We started by defining critical region in the original dataspace (x), then reformulated this in terms of a scalar test statistic t(x).
We can take this one step further and define the critical region of a test of H0 with size as the set of data space where p0 ≤ .
Formally the p-value relates only to H0, but the resulting test willhave a given power with respect to a given alternative H1.
G. Cowan Invisibles 2013 / Statistical Data Analysis 24
Quick review of parameter estimationThe parameters of a pdf are constants that characterize its shape, e.g.
random variable
Suppose we have a sample of observed values:
parameter
We want to find some function of the data to estimate the parameter(s):
← estimator written with a hat
Sometimes we say ‘estimator’ for the function of x1, ..., xn;‘estimate’ for the value of the estimator with a particular data set.
G. Cowan Invisibles 2013 / Statistical Data Analysis 25
Properties of estimatorsIf we were to repeat the entire measurement, the estimatesfrom each would follow a pdf:
biasedlargevariance
best
We want small (or zero) bias (systematic error):→ average of repeated measurements should tend to true value.
And we want a small variance (statistical error):→ small bias & variance are in general conflicting criteria
G. Cowan Invisibles 2013 / Statistical Data Analysis 26
The likelihood functionSuppose the entire result of an experiment (set of measurements)is a collection of numbers x, and suppose the joint pdf forthe data x is a function that depends on a set of parameters :
Now evaluate this function with the data obtained andregard it as a function of the parameter(s). This is the likelihood function:
(x constant)
G. Cowan Invisibles 2013 / Statistical Data Analysis 27
The likelihood function for i.i.d.*. data
Consider n independent observations of x: x1, ..., xn, where x follows f (x; ). The joint pdf for the whole data sample is:
In this case the likelihood function is
(xi constant)
* i.i.d. = independent and identically distributed
G. Cowan Invisibles 2013 / Statistical Data Analysis 28
Maximum likelihood estimatorsIf the hypothesized is close to the true value, then we expect a high probability to get data like that which we actually found.
So we define the maximum likelihood (ML) estimator(s) to be the parameter value(s) for which the likelihood is maximum.
ML estimators not guaranteed to have any ‘optimal’properties, (but in practice they’re very good).
G. Cowan Invisibles 2013 / Statistical Data Analysis 29
Example: fitting a straight line
Data:
Model: yi independent and all follow yi ~ Gauss(μ(xi ), σi )
assume xi and i known.
Goal: estimate 0
Here suppose we don’t care about 1 (example of a “nuisance parameter”)
G. Cowan Invisibles 2013 / Statistical Data Analysis 30
Maximum likelihood fit with Gaussian data
In this example, the yi are assumed independent, so thelikelihood function is a product of Gaussians:
Maximizing the likelihood is here equivalent to minimizing
i.e., for Gaussian data, ML same as Method of Least Squares (LS)
G. Cowan Invisibles 2013 / Statistical Data Analysis 31
1 known a priori
For Gaussian yi, ML same as LS
Minimize χ2 → estimator
Come up one unit from
to find
G. Cowan Invisibles 2013 / Statistical Data Analysis 32
Correlation between
causes errors
to increase.
Standard deviations from
tangent lines to contour
ML (or LS) fit of 0 and 1
G. Cowan Invisibles 2013 / Statistical Data Analysis 33
The information on 1
improves accuracy of
If we have a measurement t1 ~ Gauss (1, σt1)
G. Cowan Invisibles 2013 / Statistical Data Analysis 34
Bayesian methodWe need to associate prior probabilities with 0 and 1, e.g.,
Putting this into Bayes’ theorem gives:
posterior likelihood prior
← based on previous measurement
‘non-informative’, in anycase much broader than
G. Cowan Invisibles 2013 / Statistical Data Analysis 35
Bayesian method (continued)
Usually need numerical methods (e.g. Markov Chain MonteCarlo) to do integral.
We then integrate (marginalize) p(0, 1 | x) to find p(0 | x):
In this example we can do the integral (rare). We find
G. Cowan Invisibles 2013 / Statistical Data Analysis 36
Digression: marginalization with MCMCBayesian computations involve integrals like
often high dimensionality and impossible in closed form,also impossible with ‘normal’ acceptance-rejection Monte Carlo.
Markov Chain Monte Carlo (MCMC) has revolutionizedBayesian computation.
MCMC (e.g., Metropolis-Hastings algorithm) generates correlated sequence of random numbers:
cannot use for many applications, e.g., detector MC;effective stat. error greater than if all values independent .
Basic idea: sample multidimensional look, e.g., only at distribution of parameters of interest.
G. Cowan Invisibles 2013 / Statistical Data Analysis 37
Although numerical values of answer here same as in frequentistcase, interpretation is different (sometimes unimportant?)
Example: posterior pdf from MCMCSample the posterior pdf from previous example with MCMC:
Summarize pdf of parameter ofinterest with, e.g., mean, median,standard deviation, etc.
G. Cowan Invisibles 2013 / Statistical Data Analysis 38
Bayesian method with alternative priorsSuppose we don’t have a previous measurement of 1 but rather, e.g., a theorist says it should be positive and not too much greaterthan 0.1 "or so", i.e., something like
From this we obtain (numerically) the posterior pdf for 0:
This summarizes all knowledge about 0.
Look also at result from variety of priors.
G. Cowan Invisibles 2013 / Statistical Data Analysis 39
Interval estimation: confidence interval from inversion of a test
Suppose a model contains a parameter μ; we want to know whichvalues are consistent with the data and which are disfavoured.
Carry out a test of size α for all values of μ.
The values that are not rejected constitute a confidence intervalfor μ at confidence level CL = 1 – α.
The probability that the true value of μ will be rejected isnot greater than α, so by construction the confidence interval will contain the true value of μ with probability ≥ 1 – α.
The interval depends on the choice of the test (critical region).
If the test is formulated in terms of a p-value, pμ, then the confidence interval represents those values of μ for which pμ > α.
To find the end points of the interval, set pμ = α and solve for μ.
G. Cowan Invisibles 2013 / Statistical Data Analysis 40
Frequentist upper limit on Poisson parameterConsider again the case of observing n ~ Poisson(s + b).
Suppose b = 4.5, nobs = 5. Find upper limit on s at 95% CL.
Relevant alternative is s = 0 (critical region at low n)
p-value of hypothesized s is P(n ≤ nobs; s, b)
Upper limit sup at CL = 1 – α found from
G. Cowan Invisibles 2013 / Statistical Data Analysis 41
n ~ Poisson(s+b): frequentist upper limit on sFor low fluctuation of n formula can give negative result for sup; i.e. confidence interval is empty.
G. Cowan Invisibles 2013 / Statistical Data Analysis 42
Limits near a physical boundarySuppose e.g. b = 2.5 and we observe n = 0.
If we choose CL = 0.9, we find from the formula for sup
Physicist: We already knew s ≥ 0 before we started; can’t use negative upper limit to report result of expensive experiment!
Statistician:The interval is designed to cover the true value only 90%of the time — this was clearly not one of those times.
Not uncommon dilemma when limit of parameter is close to a physical boundary.
G. Cowan Invisibles 2013 / Statistical Data Analysis 43
Expected limit for s = 0
Physicist: I should have used CL = 0.95 — then sup = 0.496
Even better: for CL = 0.917923 we get sup = 10!
Reality check: with b = 2.5, typical Poisson fluctuation in n isat least √2.5 = 1.6. How can the limit be so low?
Look at the mean limit for the no-signal hypothesis (s = 0)(sensitivity).
Distribution of 95% CL limitswith b = 2.5, s = 0.Mean upper limit = 4.44
G. Cowan Invisibles 2013 / Statistical Data Analysis 44
The Bayesian approach to limitsIn Bayesian statistics need to start with ‘prior pdf’ (), this reflects degree of belief about before doing the experiment.
Bayes’ theorem tells how our beliefs should be updated inlight of the data x:
Integrate posterior pdf p(| x) to give interval with any desiredprobability content.
For e.g. n ~ Poisson(s+b), 95% CL upper limit on s from
G. Cowan Invisibles 2013 / Statistical Data Analysis 45
Bayesian prior for Poisson parameterInclude knowledge that s ≥ 0 by setting prior (s) = 0 for s < 0.
Could try to reflect ‘prior ignorance’ with e.g.
Not normalized but this is OK as long as L(s) dies off for large s.
Not invariant under change of parameter — if we had used insteada flat prior for, say, the mass of the Higgs boson, this would imply a non-flat prior for the expected number of Higgs events.
Doesn’t really reflect a reasonable degree of belief, but often usedas a point of reference;
or viewed as a recipe for producing an interval whose frequentistproperties can be studied (coverage will depend on true s).
G. Cowan Invisibles 2013 / Statistical Data Analysis 46
Bayesian interval with flat prior for sSolve to find limit sup:
For special case b = 0, Bayesian upper limit with flat priornumerically same as one-sided frequentist case (‘coincidence’).
where
G. Cowan Invisibles 2013 / Statistical Data Analysis 47
Bayesian interval with flat prior for sFor b > 0 Bayesian limit is everywhere greater than the (one sided) frequentist upper limit.
Never goes negative. Doesn’t depend on b if n = 0.
G. Cowan Invisibles 2013 / Statistical Data Analysis 48
Priors from formal rules Because of difficulties in encoding a vague degree of beliefin a prior, one often attempts to derive the prior from formal rules,e.g., to satisfy certain invariance principles or to provide maximuminformation gain for a certain set of measurements.
Often called “objective priors” Form basis of Objective Bayesian Statistics
The priors do not reflect a degree of belief (but might representpossible extreme cases).
In Objective Bayesian analysis, can use the intervals in afrequentist way, i.e., regard Bayes’ theorem as a recipe to producean interval with certain coverage properties.
G. Cowan Invisibles 2013 / Statistical Data Analysis 49
Priors from formal rules (cont.) For a review of priors obtained by formal rules see, e.g.,
Formal priors have not been widely used in HEP, but there isrecent interest in this direction, especially the reference priorsof Bernardo and Berger; see e.g.
L. Demortier, S. Jain and H. Prosper, Reference priors for highenergy physics, Phys. Rev. D 82 (2010) 034002, arXiv:1002.1111.
D. Casadei, Reference analysis of the signal + background model in counting experiments, JINST 7 (2012) 01012; arXiv:1108.4270.
G. Cowan Invisibles 2013 / Statistical Data Analysis 50
Systematic uncertainties and nuisance parametersIn general our model of the data is not perfect:
x
L (x
|θ) model:
truth:
Can improve model by including additional adjustable parameters.
Nuisance parameter ↔ systematic uncertainty. Some point in theparameter space of the enlarged model should be “true”.
Presence of nuisance parameter decreases sensitivity of analysisto the parameter of interest (e.g., increases variance of estimate).
G. Cowan Invisibles 2013 / Statistical Data Analysis 51
Prototype search analysis Search for signal in a region of phase space; result is histogramof some variable x giving numbers:
Assume the ni are Poisson distributed with expectation values
signal
where
background
strength parameter
G. Cowan Invisibles 2013 / Statistical Data Analysis 52
Prototype analysis (II)Often also have a subsidiary measurement that constrains someof the background and/or shape parameters:
Assume the mi are Poisson distributed with expectation values
nuisance parameters (s, b,btot)Likelihood function is
G. Cowan Invisibles 2013 / Statistical Data Analysis 53
The profile likelihood ratioBase significance test on the profile likelihood ratio:
maximizes L forSpecified μ
maximize L
The likelihood ratio of point hypotheses gives optimum test (Neyman-Pearson lemma).
The profile LR hould be near-optimal in present analysis with variable μ and nuisance parameters .
G. Cowan Invisibles 2013 / Statistical Data Analysis 54
Test statistic for discoveryTry to reject background-only (μ= 0) hypothesis using
i.e. here only regard upward fluctuation of data as evidence against the background-only hypothesis.
Note that even though here physically μ ≥ 0, we allow to be negative. In large sample limit its distribution becomesGaussian, and this will allow us to write down simple expressions for distributions of our test statistics.
55
p-value for discovery
G. Cowan CERN Academic Training 2012 / Statistics for HEP / Lecture 2
Large q0 means increasing incompatibility between the dataand hypothesis, therefore p-value for an observed q0,obs is
will get formula for this later
From p-value get equivalent significance,
G. Cowan Invisibles 2013 / Statistical Data Analysis 56
Example of a p-valueATLAS, Phys. Lett. B 716 (2012) 1-29
57
Expected (or median) significance / sensitivity
When planning the experiment, we want to quantify how sensitivewe are to a potential discovery, e.g., by given median significanceassuming some nonzero strength parameter μ′.
G. Cowan CERN Academic Training 2012 / Statistics for HEP / Lecture 2
So for p-value, need f(q0|0), for sensitivity, will need f(q0|μ′),
G. Cowan Invisibles 2013 / Statistical Data Analysis 58
Distribution of q0 in large-sample limitAssuming approximations valid in the large sample (asymptotic)limit, we can write down the full distribution of q as
The special case μ′ = 0 is a “half chi-square” distribution:
In large sample limit, f(q0|0) independent of nuisance parameters;f(q0|μ′) depends on nuisance parameters through σ.
Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554
G. Cowan Invisibles 2013 / Statistical Data Analysis 59
Cumulative distribution of q0, significanceFrom the pdf, the cumulative distribution of q is found to be
The special case μ′ = 0 is
The p-value of the μ = 0 hypothesis is
Therefore the discovery significance Z is simply
Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554
60
Monte Carlo test of asymptotic formula
G. Cowan Invisibles 2013 / Statistical Data Analysis
Here take τ = 1.
Asymptotic formula is good approximation to 5level (q0 = 25) already forb ~ 20.
Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554
I.e. when setting an upper limit, an upwards fluctuation of the data is not taken to mean incompatibility with the hypothesized μ:
From observed qμ find p-value:
Large sample approximation:
95% CL upper limit on μ is highest value for which p-value is not less than 0.05.
G. Cowan Invisibles 2013 / Statistical Data Analysis 61
Test statistic for upper limits
For purposes of setting an upper limit on μ use
where
cf. Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554.
G. Cowan Invisibles 2013 / Statistical Data Analysis 62
Example: the on/off problem
Measure two Poisson distributed values:
n ~ Poisson(s+b) (primary or “search” measurement)
m ~ Poisson(τb) (control measurement, τ known)
The likelihood function is
Use this to construct profile likelihood ratio (b is nuisanceparmeter):
(Cranmer 2005; Cousins, Linnemann, and Tucker 2008; Li and Ma 1983,...)
G. Cowan Invisibles 2013 / Statistical Data Analysis 63
Ingredients for profile likelihood ratio
To construct profile likelihood ratio from this need estimators:
and in particular to test for discovery (s = 0),
64
Monte Carlo test of asymptotic formula
G. Cowan CERN Academic Training 2012 / Statistics for HEP / Lecture 2
Here take τ = 1.
Asymptotic formula is good approximation to 5level (q0 = 25) already forb ~ 20.
Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554
65
Monte Carlo test of asymptotic formulae
G. Cowan CERN Academic Training 2012 / Statistics for HEP / Lecture 2
Consider again n ~ Poisson (μs + b), m ~ Poisson(τb)Use qμ to find p-value of hypothesized μ values.
E.g. f (q1|1) for p-value of μ=1.
Typically interested in 95% CL, i.e., p-value threshold = 0.05, i.e.,q1 = 2.69 or Z1 = √q1 = 1.64.
Median[q1 |0] gives “exclusion sensitivity”.
Here asymptotic formulae goodfor s = 6, b = 9.
Cowan, Cranmer, Gross, Vitells, arXiv:1007.1727, EPJC 71 (2011) 1554
G. Cowan Invisibles 2013 / Statistical Data Analysis 66
I. Discovery sensitivity for counting experiment with b known:
(a)
(b) Profile likelihood ratio test & Asimov:
II. Discovery sensitivity with uncertainty in b, σb:
(a)
(b) Profile likelihood ratio test & Asimov:
Expected discovery significance for counting experiment with background uncertainty
G. Cowan Invisibles 2013 / Statistical Data Analysis 67
Counting experiment with known backgroundCount a number of events n ~ Poisson(s+b), where
s = expected number of events from signal,
b = expected number of background events.
Usually convert to equivalent significance:
To test for discovery of signal compute p-value of s = 0 hypothesis,
where Φ is the standard Gaussian cumulative distribution, e.g.,Z > 5 (a 5 sigma effect) means p < 2.9 ×107.
To characterize sensitivity to discovery, give expected (meanor median) Z under assumption of a given s.
G. Cowan Invisibles 2013 / Statistical Data Analysis 68
s/√b for expected discovery significanceFor large s + b, n → x ~ Gaussian(μ,) , μ = s + b, = √(s + b).
For observed value xobs, p-value of s = 0 is Prob(x > xobs | s = 0),:
Significance for rejecting s = 0 is therefore
Expected (median) significance assuming signal rate s is
G. Cowan Invisibles 2013 / Statistical Data Analysis 69
Better approximation for significancePoisson likelihood for parameter s is
So the likelihood ratio statistic for testing s = 0 is
To test for discovery use profile likelihood ratio:
For now no nuisance params.
G. Cowan Invisibles 2013 / Statistical Data Analysis 70
Approximate Poisson significance (continued)
For sufficiently large s + b, (use Wilks’ theorem),
To find median[Z|s], let n → s + b (i.e., the Asimov data set):
This reduces to s/√b for s << b.
G. Cowan Invisibles 2013 / Statistical Data Analysis 71
n ~ Poisson(s+b), median significance,assuming s, of the hypothesis s = 0
“Exact” values from MC,jumps due to discrete data.
Asimov √q0,A good approx.for broad range of s, b.
s/√b only good for s « b.
CCGV, EPJC 71 (2011) 1554, arXiv:1007.1727
G. Cowan Invisibles 2013 / Statistical Data Analysis 72
Extending s/√b to case where b uncertainThe intuitive explanation of s/√b is that it compares the signal, s, to the standard deviation of n assuming no signal, √b.
Now suppose the value of b is uncertain, characterized by a standard deviation σb.
A reasonable guess is to replace √b by the quadratic sum of√b and σb, i.e.,
This has been used to optimize some analyses e.g. where σb cannot be neglected.
G. Cowan Invisibles 2013 / Statistical Data Analysis 73
Asymptotic significanceUse profile likelihood ratio for q0, and then from this get discoverysignificance using asymptotic approximation (Wilks’ theorem):
Essentially same as in:
Or use the variance of b = m/τ,
G. Cowan Invisibles 2013 / Statistical Data Analysis 74
Asimov approximation for median significanceTo get median discovery significance, replace n, m by theirexpectation values assuming background-plus-signal model:
n → s + bm → τb
, to eliminate τ:ˆ
G. Cowan Invisibles 2013 / Statistical Data Analysis 75
Limiting cases
Expanding the Asimov formula in powers of s/b andσb
2/b (= 1/τ) gives
So the “intuitive” formula can be justified as a limiting caseof the significance from the profile likelihood ratio test evaluated with the Asimov data set.
G. Cowan Invisibles 2013 / Statistical Data Analysis 76
Testing the formulae: s = 5
G. Cowan Invisibles 2013 / Statistical Data Analysis 77
Using sensitivity to optimize a cut
G. Cowan Invisibles 2013 / Statistical Data Analysis 78
Summary on discovery sensitivity
For large b, all formulae OK.
For small b, s/√b and s/√(b+σb2) overestimate the significance.
Could be important in optimization of searches withlow background.
Formula maybe also OK if model is not simple on/off experiment, e.g., several background control measurements (checking this).
Simple formula for expected discovery significance based onprofile likelihood ratio test and Asimov approximation:
G. Cowan Invisibles 2013 / Statistical Data Analysis 79
Bayesian limits on s with uncertainty on bConsider n ~ Poisson(s+b) and take e.g. as prior probabilities
Put this into Bayes’ theorem,
Marginalize over the nuisance parameter b,
Then use p(s|n) to find intervals for s with any desired probability content.
G. Cowan Invisibles 2013 / Statistical Data Analysis 80
A toy exampleFor each event we measure two variables, x = (x1, x2).
Suppose that for background events (hypothesis H0),
and for a certain signal model (hypothesis H1) they follow
where x1, x2 ≥ 0 and C is a normalization constant.
G. Cowan Invisibles 2013 / Statistical Data Analysis 81
Likelihood ratio as test statisticIn a real-world problem we usually wouldn’t have the pdfs f(x|H0) and f(x|H1), so we wouldn’t be able to evaluate thelikelihood ratio
for a given observed x, hencethe need for multivariate methods to approximate this with some other function.
But in this example we can find contours of constant likelihood ratio such as:
G. Cowan Invisibles 2013 / Statistical Data Analysis 82
Event selection using the LRUsing Monte Carlo, we can find the distribution of the likelihoodratio or equivalently of
signal (H1)
background (H0)
From the Neyman-Pearson lemmawe know that by cutting on thisvariable we would select a signalsample with the highest signalefficiency (test power) for a givenbackground efficiency.
G. Cowan Invisibles 2013 / Statistical Data Analysis 83
Search for the signal processBut what if the signal process is not known to exist and we wantto search for it. The relevant hypotheses are therefore
H0: all events are of the background typeH1: the events are a mixture of signal and background
Rejecting H0 with Z > 5 constitutes “discovering” new physics.
Suppose that for a given integrated luminosity, the expected numberof signal events is s, and for background b.
The observed number of events n will follow a Poisson distribution:
G. Cowan Invisibles 2013 / Statistical Data Analysis 84
Likelihoods for full experimentWe observe n events, and thus measure n instances of x = (x1, x2).
The likelihood function for the entire experiment assumingthe background-only hypothesis (H0) is
and for the “signal plus background” hypothesis (H1) it is
where s and b are the (prior) probabilities for an event tobe signal or background, respectively.
G. Cowan Invisibles 2013 / Statistical Data Analysis 85
Likelihood ratio for full experimentWe can define a test statistic Q monotonic in the likelihood ratio as
To compute p-values for the b and s+b hypotheses given an observed value of Q we need the distributions f(Q|b) and f(Q|s+b).
Note that the term –s in front is a constant and can be dropped.
The rest is a sum of contributions for each event, and each termin the sum has the same distribution.
Can exploit this to relate distribution of Q to that of singleevent terms using (Fast) Fourier Transforms (Hu and Nielsen, physics/9906010).
G. Cowan Invisibles 2013 / Statistical Data Analysis 86
Distribution of Q
Take e.g. b = 100, s = 20.
f (Q|b)f (Q|s+b)
p-value of b only
p-value of s+b
Suppose in real experimentQ is observed here.
If ps+b < α, reject signal model s at confidence level 1 – α.
If pb < 2.9 × 107, reject background-only model (signif. Z = 5).
G. Cowan Invisibles 2013 / Statistical Data Analysis 87
Including systematic uncertainties
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 88
Low sensitivity to μIt can be that the effect of a given hypothesized μ is very smallrelative to the background-only (μ = 0) prediction.
This means that the distributions f(qμ|μ) and f(qμ|0) will bealmost the same:
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 89
Having sufficient sensitivityIn contrast, having sensitivity to μ means that the distributionsf(qμ|μ) and f(qμ|0) are more separated:
That is, the power (probability to reject μ if μ = 0) is substantially higher than α. Use this power as a measure of the sensitivity.
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 90
Spurious exclusionConsider again the case of low sensitivity. By construction the probability to reject μ if μ is true is α (e.g., 5%).
And the probability to reject μ if μ = 0 (the power) is only slightly greater than α.
This means that with probability of around α = 5% (slightly higher), one excludes hypotheses to which one has essentially no sensitivity (e.g., mH = 1000 TeV).
“Spurious exclusion”
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 91
Ways of addressing spurious exclusion
The problem of excluding parameter values to which one hasno sensitivity known for a long time; see e.g.,
In the 1990s this was re-examined for the LEP Higgs search byAlex Read and others
and led to the “CLs” procedure for upper limits.
Unified intervals also effectively reduce spurious exclusion bythe particular choice of critical region.
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 92
The CLs procedure
f (Q|b)
f (Q| s+b)
ps+bpb
In the usual formulation of CLs, one tests both the μ = 0 (b) andμ > 0 (μs+b) hypotheses with the same statistic Q = 2ln Ls+b/Lb:
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 93
The CLs procedure (2)
As before, “low sensitivity” means the distributions of Q under b and s+b are very close:
f (Q|b)
f (Q|s+b)
ps+bpb
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 94
The CLs solution (A. Read et al.) is to base the test not onthe usual p-value (CLs+b), but rather to divide this by CLb (~ one minus the p-value of the b-only hypothesis), i.e.,
Define:
Reject s+b hypothesis if: Reduces “effective” p-value when the two
distributions become close (prevents exclusion if sensitivity is low).
f (Q|b) f (Q|s+b)
CLs+b = ps+b
1CLb
= pb
The CLs procedure (3)
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 95
Setting upper limits on μ = σ/σSM
Carry out the CLs procedure for the parameter μ = σ/σSM, resulting in an upper limit μup.
In, e.g., a Higgs search, this is done for each value of mH.
At a given value of mH, we have an observed value of μup, andwe can also find the distribution f(μup|0):
±1 (green) and ±2 (yellow) bands from toy MC;
Vertical lines from asymptotic formulae.
G. Cowan Cargese 2012 / Statistics for HEP / Lecture 2 96
How to read the green and yellow limit plots
ATLAS, Phys. Lett. B 710 (2012) 49-66
For every value of mH, find the CLs upper limit on μ.
Also for each mH, determine the distribution of upper limits μup one would obtain under the hypothesis of μ = 0.
The dashed curve is the median μup, and the green (yellow) bands give the ± 1σ (2σ) regions of this distribution.
G. Cowan Invisibles 2013 / Statistical Data Analysis 97
Summary
General idea of a statistical test:Divide data spaced into two regions; depending onwhere data are then observed, accept or reject hypothesis.
Significance tests (also for goodness-of-fit):p-value = probability to see level of incompatibilitybetween data and hypothesis equal to or greater thanlevel found with the actual data.
Parameter estimationMaximize likelihood function → ML estimator.Bayesian estimator based on posterior pdf.Confidence interval: set of parameter values not rejected in a test of size α = 1 – CL.
G. Cowan Invisibles 2013 / Statistical Data Analysis 98
Extra slides
G. Cowan Invisibles 2013 / Statistical Data Analysis 99
Jeffreys’ priorAccording to Jeffreys’ rule, take prior according to
where
is the Fisher information matrix.
One can show that this leads to inference that is invariant undera transformation of parameters.
For a Gaussian mean, the Jeffreys’ prior is constant; for a Poisson mean μ it is proportional to 1/√μ.
G. Cowan Invisibles 2013 / Statistical Data Analysis 100
Jeffreys’ prior for Poisson mean
Suppose n ~ Poisson(μ). To find the Jeffreys’ prior for μ,
So e.g. for μ = s + b, this means the prior (s) ~ 1/√(s + b), which depends on b. Note this is not designed as a degree of belief about s.
G. Cowan Invisibles 2013 / Statistical Data Analysis 101
Proof of Neyman-Pearson lemmaWe want to determine the critical region W that maximizes the power
subject to the constraint
First, include in W all points where P(x|H0) = 0, as they contributenothing to the size, but potentially increase the power.
G. Cowan Invisibles 2013 / Statistical Data Analysis 102
Proof of Neyman-Pearson lemma (2)
For P(x|H0) ≠ 0 we can write the power as
The ratio of 1 – to is therefore
which is the average of the likelihood ratio P(x|H1) / P(x|H0) overthe critical region W, assuming H0.
(1 – ) / is thus maximized if W contains the part of the samplespace with the largest values of the likelihood ratio.
G. Cowan Invisibles 2013 / Statistical Data Analysis 103
Choosing a critical regionTo construct a test of a hypothesis H0, we can ask what are the relevant alternatives for which one would like to have a high power.
Maximize power wrt H1 = maximize probability to reject H0 if H1 is true.
Often such a test has a high power not only with respect to a specific point alternative but for a class of alternatives. E.g., using a measurement x ~ Gauss (μ, σ) we may test
H0 : μ = μ0 versus the composite alternative H1 : μ > μ0
We get the highest power with respect to any μ > μ0 by taking the critical region x ≥ xc where the cut-off xc is determined by the significance level such that
α = P(x ≥xc|μ0).
G. Cowan Invisibles 2013 / Statistical Data Analysis 104
Τest of μ = μ0 vs. μ > μ0 with x ~ Gauss(μ,σ)
Standard Gaussian quantile
Standard Gaussiancumulative distribution
G. Cowan Invisibles 2013 / Statistical Data Analysis 105
Choice of critical region based on power (3)
But we might consider μ < μ0 as well as μ > μ0 to be viable alternatives, and choose the critical region to contain both high and low x (a two-sided test).
New critical region now gives reasonable power for μ < μ0, but less power for μ > μ0 than the original one-sided test.
G. Cowan Invisibles 2013 / Statistical Data Analysis 106
No such thing as a model-independent testIn general we cannot find a single critical region that gives themaximum power for all possible alternatives (no “UniformlyMost Powerful” test).
In HEP we often try to construct a test of
H0 : Standard Model (or “background only”, etc.)
such that we have a well specified “false discovery rate”,
α = Probability to reject H0 if it is true,
and high power with respect to some interesting alternative,
H1 : SUSY, Z′, etc.
But there is no such thing as a “model independent” test. Anystatistical test will inevitably have high power with respect tosome alternatives and less power with respect to others.
G. Cowan Invisibles 2013 / Statistical Data Analysis 107
Choice of test for discoveryIf μ represents the signal rate, then discovering the signal processrequires rejecting H0 : μ = 0.
Often our evidence for the signal process comes in the form ofan excess of events above the level predicted from backgroundalone, i.e., μ > 0 for physical signal models.
So the relevant alternative hypothesis is H0 : μ > 0.
In other cases the relevant alternative may also include μ < 0 (e.g., neutrino oscillations).
The critical region giving the highest power for the test of μ = 0 relative to the alternative of μ > 0 thus contains high values of theestimated signal rate.
G. Cowan Invisibles 2013 / Statistical Data Analysis 108
Choice of test for limitsSuppose the existence of the signal process (μ > 0) is not yet established.
The interesting alternative in this context is μ = 0.
That is, we want to ask what values of μ can be excluded on the grounds that the implied rate is too high relative to what isobserved in the data.
The critical region giving the highest power for the test of μ relativeto the alternative of μ = 0 thus contains low values of the estimatedrate, .
Test based on one-sided alternative → upper limit.
G. Cowan Invisibles 2013 / Statistical Data Analysis 109
More on choice of test for limitsIn other cases we want to exclude μ on the grounds that some othermeasure of incompatibility between it and the data exceeds somethreshold.
For example, the process may be known to exist, and thus μ = 0is no longer an interesting alternative.
If the measure of incompatibility is taken to be the likelihood ratiowith respect to a two-sided alternative, then the critical region can contain data values corresponding to both high and low signal rate.
→ unified intervals, G. Feldman, R. Cousins, Phys. Rev. D 57, 3873–3889 (1998)
A Big Debate is whether to focus on small (or zero) valuesof the parameter as the relevant alternative when the existence of a signal has not yet been established. Professional statisticians have voiced support on both sides of the debate.
G. Cowan Invisibles 2013 / Statistical Data Analysis 110
p-value example: testing whether a coin is ‘fair’
i.e. p = 0.0026 is the probability of obtaining such a bizarreresult (or more so) ‘by chance’, under the assumption of H.
Probability to observe n heads in N coin tosses is binomial:
Hypothesis H: the coin is fair (p = 0.5).
Suppose we toss the coin N = 20 times and get n = 17 heads.
Region of data space with equal or lesser compatibility with H relative to n = 17 is: n = 17, 18, 19, 20, 0, 1, 2, 3. Addingup the probabilities for these values gives:
G. Cowan Invisibles 2013 / Statistical Data Analysis 111
Variance of estimators from information inequalityThe information inequality (RCF) sets a lower bound on the variance of any estimator (not only ML):
Often the bias b is small, and equality either holds exactly oris a good approximation (e.g. large data sample limit). Then,
Estimate this using the 2nd derivative of ln L at its maximum:
G. Cowan Invisibles 2013 / Statistical Data Analysis 112
Information inequality for n parametersSuppose we have estimated n parameters
The (inverse) minimum variance bound is given by the Fisher information matrix:
The information inequality then states that V I is a positivesemi-definite matrix, where Therefore
Often use I as an approximation for covariance matrix, estimate using e.g. matrix of 2nd derivatives at maximum of L.
G. Cowan Invisibles 2013 / Statistical Data Analysis 113
ML example: parameter of exponential pdf
Consider exponential pdf,
and suppose we have i.i.d. data,
The likelihood function is
The value of τ for which L(τ) is maximum also gives the maximum value of its logarithm (the log-likelihood function):
G. Cowan Invisibles 2013 / Statistical Data Analysis 114
ML example: parameter of exponential pdf (2)
Find its maximum by setting
→
Monte Carlo test: generate 50 valuesusing τ = 1:
We find the ML estimate:
G. Cowan Invisibles 2013 / Statistical Data Analysis 115
MCMC basics: Metropolis-Hastings algorithmGoal: given an n-dimensional pdf generate a sequence of points
1) Start at some point
2) Generate
Proposal densitye.g. Gaussian centredabout
3) Form Hastings test ratio
4) Generate
5) If
else
move to proposed point
old point repeated
6) Iterate
G. Cowan Invisibles 2013 / Statistical Data Analysis 116
Metropolis-Hastings (continued)This rule produces a correlated sequence of points (note how each new point depends on the previous one).
For our purposes this correlation is not fatal, but statisticalerrors larger than naive
The proposal density can be (almost) anything, but chooseso as to minimize autocorrelation. Often take proposaldensity symmetric:
Test ratio is (Metropolis-Hastings):
I.e. if the proposed step is to a point of higher , take it; if not, only take the step with probability If proposed step rejected, hop in place.