(Dated: July 1, 2004) Abstract. In Run I, the DØ Collaboration adopted a Bayesian method for calculating confidence limits that

Size: px
Start display at page:

Download "(Dated: July 1, 2004) Abstract. In Run I, the DØ Collaboration adopted a Bayesian method for calculating confidence limits that"

Transcription

1 DRAFT Calculating Confidence Limits DØNote 4491 J. Linnemann 1, M. Paterno 2 and H.B. Prosper 3 1 Michigan State University, East Lansing, Michigan Fermi National Accelerator Laboratory, Batavia, Illinois Florida State University, Tallahassee, Florida (Dated: July 1, 2004) Abstract In Run I, the DØ Collaboration adopted a Bayesian method for calculating confidence limits that was recommended by the Confidence Limits Working Group [1]. This paper describes the method and discusses a few important issues pertaining to it. 1

2 2 I. INTRODUCTION This paper addresses the question of how to present the results of an experiment that observes few or no events. There is broad agreement that the answer is to set an upper limit at some specified confidence level (CL) on the cross section for the process of interest. The aim of this paper is: 1) to provide arguments for the adoption of a fundamentally Bayesian method, and 2) to describe that method in detail. To motivate our choice, and to provide background for the reader, we review in Section II the concepts of confidence and credible intervals, while in Section III we briefly review a few published prescriptions for constructing confidence intervals. We then present in Section IV our reasons for choosing the method adopted by the DØ Collaboration, during Run I, and describe the procedure. Finally, in Section VII, we provide a summary and conclusions of the paper. II. FREQUENTIST AND BAYESIAN INTERVALS A What is Probability? There are two schools of statistical inference (Frequentist and Bayesian) that differ in their interpretation of the concept of probability. The laws of probability theory the mathematical rules that govern the manipulation of abstract symbols called probabilities are used by both. Much of the confusion that pervades the subject of intervals can be traced, ultimately, to a failure either to accept, or appreciate, that there is a difference between the two approaches and that this difference matters. The confusion becomes particularly acute when the numerical results of the two schools agree, from which it is sometimes concluded that the conceptual difference between these schools is merely philosophical, and as such irrelevant. This is not so! To see this, let us begin with the definition of probability according to the (dominant) Frequentist school.

3 3 1 The Frequentist Meaning of Probability Although introduced in the earlier work of J. W. Gibbs, the definition we give of probability is associated with R. Fisher. Two other definitions (those of J. Neyman and R. Mises), also of the Frequentist school, are less suitable. A clear explanation of each can be found in Ref. [2]. See, also, Ref. [3] for an historical perspective. Letting x denote the outcome of some observation, and I some other information of relevance to that observation, we denote the probability of x, given I, as P (x I). The Frequentist school defines this probability P (x I) as the limit (as the sample size tends to infinity) of the relative frequency with which x occurs, out of a sample of observations that satisfy whatever constraints implied by the specification of I. To be precise, let k be the number of observations that satisfy both x and I, i.e., the successes, in n observations that satisfy I, i.e., the trials, the Frequentist definition of probability is k P (x I) = lim n n. (II.1) This definition is extremely important in practice because it provides the connection between frequency data, for example, counts in the bins of a histogram, and the true underlying density function of which the distribution of counts is an estimate. However, the definition is not without difficulty. The problem is that the limit does not exist in a strict mathematical sense! The reason is that there is no deterministic rule that connects the outcome of trial n + 1 with that of the previous trial n, because, by assumption, the outcomes are random. Moreover, the limit cannot be found empirically since no sequence of trials is ever infinite. In practice, probabilities are predicted from a judicious use of combinatoric reasoning of which the derivation of the binomial distribution is a classic example. The relationship between these abstractly derived probabilities and the relative frequencies we measure in experiments is a subtle subject that is well beyond the scope of this paper. We merely note that the pivotal theorem that underpins the relative frequency interpretation establishes that the limit in Eq. (II.1) exists with probability one,

4 4 provided that the order of trials is unimportant and the probability of success at each trial is the same. This theorem is due to de Finetti and is itself based on another important theorem (see Appendix C) by the same author. 2 The Bayesian Meaning of Probability In the Bayesian school, probability is interpreted as a number that measures the degree of belief in or plausibility of a proposition. It therefore makes sense to talk about the probability of an hypothesis, whereas such a statement has no meaning for Frequentists. The Bayesian school derives probabilities using the same combinatoric reasoning as deployed by the Frequentists, and so the basic difference between the two schools lies in the fact that in the Frequentist interpretation probability describes a state of Nature, while in the Bayesian view a probability describes a state of knowledge. It is this difference that leads to the difference in the interpretation of Frequentist and Bayesian intervals and upper limits. In order to highlight these differences, we shall reserve the term confidence intervals for Frequentist intervals and credible intervals for Bayesian intervals. B Confidence Intervals The concept of the confidence interval was introduced by Neyman in a seminal paper published in 1937 [5]. It is easier to explain the concept with an example. Let us consider the measurement of a cross section x. We suppose that the cross section has one fixed value x 0 and that we have acquired data D that can be used to measure, or as statisticians say estimate, the cross section. By a measurement we shall mean a procedure that yields an interval [L(D), U(D)] whose lower and upper limits, L and U, respectively, depend on the data D. Now imagine repeating the experiment an arbitrarily large number of times. Each repetition of the experiment would, in general, yield different data D and a different interval. Some fraction of the intervals in this ensemble of intervals will bracket,

5 5 that is, contain, the true (but unknown) value x 0 of the cross section. That fraction, a relative frequency, is called the coverage probability of the ensemble of intervals. Neyman required that the coverage probability for this ensemble be greater than or equal to the desired confidence level, say β = If this were so, then the probability to draw an interval at random from the ensemble of intervals would be, by construction, β. There is a snag, however, with constructing a single ensemble of intervals satisfying the aforementioned criterion: we do not know the true value of the cross section. It is therefore clearly necessary, as Neyman recognized, that the bound on coverage probability hold true whatever the true value of the parameter to be estimated, here the cross section x. This requires one to construct an ensemble of intervals for every possible value of the cross section. If each ensemble has, by construction, a coverage probability greater than or equal to the desired confidence level then one is assured that an interval will be drawn with a probability not less than the confidence level. Intervals that satisfy this criterion, called the Neyman criterion, are called confidence intervals and are said to cover. A few things are worth emphasizing: 1) the intervals themselves are the random variables and 2) the coverage property refers to repeated applications of the method. Consequently, it applies to repetitions of different experiments as well as to repetitions of the same experiment. 1 Credible Intervals A credible interval [L(D), U(D)] is constructed so that the probability, i.e., the measure of the plausibility of the proposition, that the true value x 0 is in the range [L(D), U(D)] is exactly β. As in the Frequentist prescription, the unknown true value x 0 is assumed fixed. Unlike the Frequentist prescription, there is no statement concerning the relative frequency of the result of a set of measurements, and no reference to an ensemble of intervals. There is, however, the necessity of assigning a probability in the absence of the measurement to the proposition that the value x lies between x and x + dx. This probability is called the

6 6 prior probability (often just the prior) for the quantity of interest, here, the cross section. The probability of x in light of the observed data D is called the posterior probability and is denoted P (x D, I). Since x is a continuous parameter, we can write this probability in terms of its probability density f(x D, I) where P (x D, I) = f(x D, I) dx. The Bayesian upper limit x UP on x at confidence level β is defined by β = x UP 0 f(x D, I) dx, (II.2) where f(x D, I) is called the posterior density. As we decribe later, the Bayesian definition of probability allows one to describe directly P(theory data), while the Frequentist definition does not. The likelihood, used by both Frequentists and Bayesians, only describes P(data theory). III. COMMENTS ON RECENT LITERATURE We have surveyed recent literature and summarize our perceptions in this Section. It is clear from our survey that particle physicists, being pragmatists, use a Bayesian-like approach to deal with systematic uncertainty. Statements such as:...the Poisson distribution is smeared with Gaussian errors... occur often in publications, and usually without comment. Unfortunately, however, mixing the two interpretations of probability sometimes causes confusion or renders the interpretation of the probabilities ambiguous. To illustrate both points, we consider briefly papers by Helene [7] and Cousins and Highland [8]. We also comment on the method of Feldman and Cousins [9] and the CLs method developed at LEP [10]. A Helene Helene [7] notes the difficulty of taking into account cogent prior information in the Frequentist approach (e.g., that a rate must be positive). After performing a proper Bayesian calculation of confidence intervals for the Poisson mean, using a flat prior, Helene purports

7 7 to show the superiority of the Bayesian intervals to those of the Frequentists. He tries to do this using an ensemble of Monte Carlo generated experiments. The first point of confusion arises because of Helene s use of a Frequentist concept, the coverage probability over an ensemble (the fraction of intervals that bracket the true value), in a Bayesian context[26]. For Bayesians, an ensemble is not needed to interpret a probability; in particular, it is not needed to interpret a Bayesian confidence level. The second point of confusion arises because the Monte Carlo method used by Helene, as noted by James [13] (see also, Prosper [14]), is not the correct method to compute coverage probabilities. To compute the latter one should keep the unknown Poisson mean fixed, generate an ensemble of experiments, and compute the fraction of intervals that contain the true fixed value that is, the coverage probability. The procedure should then be repeated with another fixed value for the Poisson mean, and repeated often enough to provide a fair sampling of the entire parameter space. B Cousins and Highland Cousins and Highland [8] consider the very important problem of how to include systematic uncertainty in a Poisson upper limit. They write the Poisson mean, µ, as µ = R S where R might be, for example, a cross section times branching fraction the parameter of interest and S is some sensitivity factor for the measurement, for example, the effective integrated luminosity, which is of no intrinsic interest; that is, the sensitivity is a nuisance parameter. A key assumption is that subsidiary measurements lead to an estimate Ŝ for S, with a known standard deviation δs. Cousins and Highland acknowledge, from the outset, the inability of Frequentist methods to account for systematic uncertainty in a straightforward and conceptually coherent way. They therefore use a Bayesian approach to represent the information about S. Given a likelihood function, f(ŝ S), describing the result of the subsidiary measurement of the sensitivity they compute the posterior probability density W (S Ŝ) assuming a flat prior

8 8 density for S. They then convolve the (cumulative) Poisson distribution with the posterior probability density W (S Ŝ) for S and, having thereby removed the nuisance parameter S, compute upper limits for R using the resulting smeared cumulative Poisson distribution function where C(N R) ( N 0 P oisson(rs, n) n=0 In Ref. [8], 90% upper limits were computed using ) W (S Ŝ) ds, (III.1) P oisson(µ, k) = e µ µ k. (III.2) k! W (S Ŝ) = Gaussian(S, Ŝ, δs), (III.3) for δs/ŝ 0.3, where Gaussian(λ, a, σ) = 1 σ 2π exp[ 1 2 ( ) 2 λ a ]. (III.4) The authors have clearly accepted, and make use of, the notions of Bayesian prior and posterior probabilities as well as the utility of integrating with respect to a parameter, not to mention the practical utility of a flat prior, yet they refrain from performing a consistent Bayesian calculation. The reason offered is that physicists prefer to treat R using Frequentist methods, since they regard them as objective. We concede that using a prior for R comes at a price: the upper limit on R would be more sensitive to the choice of prior for R than to the choice of prior for S since the posterior probability for S is dominated by its likelihood function. This, however, does not alter the fact that choice of prior for S is no more or less subjective than would be a similar choice of prior for R. Moreover, it is not clear how the confidence level is to be interpreted, because of the mixture of Frequentist and Bayesian elements in the calculation. We see no obvious way of proving that the intervals derived by this mixed prescription possess Frequentist coverage in general, though they do possess the property in some specific cases which have been examined. σ

9 9 1 Further Comments In the method of Cousins and Highland the estimate Ŝ is required to be unbiased, that is, < Ŝ >= S. It is clear that we should also require Ŝ > 0, which implies that the likelihood function f(ŝ S) must be restricted to the domain Ŝ [0, ). If the likelihood is a Gaussian it would have to be renormalized, yielding f(ŝ S) = Gaussian(Ŝ, S, δs)/ Gaussian(Ŝ, S, δs) dŝ, 0 [ ( )] S = Gaussian(Ŝ, S, δs)/1 1 + Erf 2 δs, Ŝ > 0, (III.5) 2 where Erf is the error function. One consequence of the truncation is that the estimate Ŝ is no longer unbiased; but it is not clear that this is a problem. What, in principle, is more important is that the likelihood f(ŝ S) be correctly normalized before it is used in Bayes Theorem to compute the posterior density W (S Ŝ), which itself must be normalized with respect to S. In practice, however, we have found that the use of an unrenormalized Gaussian for the posterior probability density, W (S Ŝ), is a good approximation for δs/ŝ < 0.3. (Strictly speaking, since Ŝ > δs/0.3, to be consistent, the likelihood function should be renormalized on the restricted domain Ŝ [δs/0.3, ), instead of [0, ). However, we have not investigated the effect of this.) Although it is unclear that coverage obtains in general for the confidence intervals computed with this method, there is one circumstance in which the limits cover. In Eq. (III.1) if we swap the order of the sum and the integral we shall have a sum over terms each of which is of the form P (n R) = 0 P oisson(rs, n)w (S Ŝ) ds. Since n=0 P (n R) = 1, P (n R) is a non-poisson discrete probability distribution. If the counts were distributed according to P (n R), rather than according to a Poisson distribution, application of the standard Frequentist procedure to compute upper limits namely, setting the cumulative distribution, Nn=0 P (n R), equal to 1 β, where β is the desired confidence level would yield exactly the same upper limits as obtained with the method of Cousins and Highland. But with the new sampling distribution P (n R) coverage is guaranteed!

10 10 C Feldman and Cousins Feldman and Cousins[9] have described a construction of confidence intervals that avoids unphysical intervals, is consistent with Neyman s viewpoint[5] and can be applied when there is a known background. The algorithm proceeds as follows: for each possible value of the parameter of interest, say the cross section x, event counts are ordered in descending order of their likelihood ratio l(x)/l(ˆx), where ˆx is the maximum likelihood estimate of the parameter x; a range of counts is constructed by including counts one at a time, starting with the first of the ordered counts, until the probability content of the range of counts is no less than the desired confidence level. This procedure results in a confidence band, which, for a given count n, yields a confidence interval [L(n), U(n)]. The Feldman and Cousins procedure has great merit, however, it is far from clear how systematic uncertainty can be taken into account in a manner consistent with a fully Frequentist viewpoint and in a way that is computationally feasible. D The CLs Method The CLs method [10], which has been used by some groups within DØ, was developed in the context of the Higgs search at LEP. The method is a generalization of a formula proposed by Zech [11]. (See also the discussion by Bob Cousins [12].) Limits calculated with the CLs method are based on the sampling distribution of the statistic or rather its logarithm Q = = M exp[ (s i + b i )] (s i + b i ) n i i exp( b i ) b n i i ( ) M ni si + b i exp( s i ), i b i, (III.6) q ln Q, M M = s i + n i ln (1 + s i /b i ), i=1 i=1 (III.7)

11 11 under the signal plus background and background only hypotheses, S and B, respectively. The index i = 1 M is over channels, where s i l i σ is the mean signal count with l i the acceptance times integrated luminosity, b i is the mean background count and n i is the observed count in the i th channel. Given the sampling distributions p(q S) and p(q B) one calculates a CLs limit, at level β, by solving the following ratio of p-values (that is, tail probabilities) β = q q 0 p(q S) q q 0 p(q B), (III.8) for the upper limit on the cross-section σ, where q 0 is the observed value of the statistic q. In practice, the densities p(q H), with H = S or B, depend on parameters that are not known precisely such as the mean backgrounds b i and the effective luminosities l i. In order to account for systematic uncertainty in these parameters, the densities p(q H) are integrated with respect to l i and b i, weighted by a known prior density π(l i, b i ), β = q q 0 p(q S) π(li, b i ) dl i db i q q 0 p(q B) π(l i, b i ) dl i db i. (III.9) The principal difficulty with Eq. (III.8) is that it has not been derived from broadly accepted principles. Moreover, since the more general procedure, Eq. (III.9), is neither purely Frequentist nor purely Bayesian, the confidence level β cannot be readily interpreted as either a relative frequency or a degree of belief, as clearly stressed in Ref. [10]. IV. OUR SUGGESTED PROCEDURE A Requirements In choosing a procedure for the setting of upper limits, we felt it necessary to be guided by a few principles. The question of how best to treat uncertainty is under active discussion in our field [16]. We believe that a consensus may be emerging and we are hopeful that the following principles will form part of this consensus:

12 12 Practicality and Power. The method must have sufficient analytical power to answer the questions we want to ask. This includes the handling of searches in which the predicted background is greater than the number of observed events, the handling of unphysical regions, and dealing with uncertainties of all kinds. Generality. The chosen method should apply to a wide variety of problems so that special cases should come up only rarely. Self-consistent. Ideally, the method should be conceptually sound and avoid techniques that violate underlying mathematical theorems. This will assure that conclusions will always be warranted, especially in circumstances where the answers are not readily apparent. B Consideration of the Two Schools Proponents of the Frequentist approach hold that it is objective and, for this reason, is preferred for scientific investigations. Consider a typical Frequentist calculation of limits for a cross section. One devises a procedure, which for a well-specified ensemble, constructs for each possible true value of the cross section x an ensemble of intervals. By construction the coverage probability of each ensemble of intervals within the class of ensembles is greater than or equal to the desired confidence level, say 95%. Moreover, as noted in Sect. II B, if one repeats such a procedure over an arbitrarily large number of different experiments, each measuring a different quantity, the confidence level over this class of classes of intervals is 95%. This is held to be a useful property. Sampling from this class of classes is analogous to the following sampling experiment: pick a ball at random from an urn chosen at random from ten urns, each containing 95% red balls and 5% white ones. Replace the ball in the urn from which it was drawn and repeat the experiment as many times as desired. Each urn is analogous to an ensemble of experiments of a give type for which 95% limits have been calculated. Clearly, the relative frequency with which the statement I have picked a red

13 13 ball is true will 0.95 as the number of repetitions grows without limit. Therefore, the argument goes, one would expect that in the ensemble of 95% CL limits published over the past fifty years at least 95% of statements of the form y < y UP are true, provided that they were all arrived at via a correct Frequentist calculation. A confidence interval is a thoroughly Frequentist concept. Therefore, it is of interest to consider why Fisher, a staunch Frequentist and the inventor of many of the statistical ideas in common use, rejected Neyman s innovation out of hand. Fisher rejected the notion of confidence intervals because of what he regarded as the inherent subjectivity in the definition of their associated confidence level. If an ensemble objectively exists then there is no problem, but Fisher [6] noted, in connection with significance tests, that.. if we possess a unique sample on which significance tests are to be performed, there is always... a multiplicity of populations to each of which we can legitimately regard our sample as belonging; so the phrase repeated sampling from the same population does not enable us to determine which population is to be used to define the probability level, for no one of them has objective reality, all being products of the statistician s imagination. One may say the same of almost all ensembles that define Frequentist confidence levels. In the end, the choice of which ensemble to use is a matter of expert judgement, convention or abstract reasoning rather than the result of an empirically verifiable procedure. It is of course true that the ensemble of published 95% limits has a coverage probability. But in view of the invariably non-objective nature of the ensembles used by each experiment it is less than clear that the coverage probability of the ensemble of published limits is indeed 95%. More to the point, even if it is, it is not clear how that fact is to be used in a scientific investigation. There is, however, a more serious problem, albeit a conceptual one. Because it relies on the relative frequency interpretation of probability, the Frequentist view encounters diffculties addressing systematic uncertainties. The supporting mathematical theory does not

14 14 provide a general, practical, algorithm to eliminate nuisance parameters. Unfortunately, most analyses contain such parameters. The Bayesian approach, on the other hand, provides a general method to handle uncertainties, irrespective of their provenance and nature. In practice, many nominally Frequentist analyses in high energy physics include the idea of integrating over uncertainties. The Cousins and Highland and CLs methods are good examples. But integrating over parameters has no justification within the Frequentist school because it eschews the use of a probability density for a parameter. Such a thing simply makes no sense in this approach. In the Bayesian school, nuisance parameters are eliminated by integrating over their probability densities because that is what probability theory dictates should be done. For practical reasons, as well as for conceptual clarity, the Run I Confidence Limits Working Group [1], recommended the Bayesian approach. This approach is preferred if for no other reason than the fact that most people, whether they are scientists, or not, cannot help but think in a Bayesian way. This is hardly surprising since the Bayesian reasoning mirrors closely the way we reason inductively [18]. One feature of the Bayesian school, recognized by some as a flaw and by others as a virtue, is that the posterior probability depends, necessarily, on the priors used in the analysis. Critics cite this as evidence that the method is subjective. Supporters answer that this is merely a platitude, because the result of an analysis will always depend on the information that goes into it. We are not aware of any useful analysis in high energy physics that does not rely upon the use of judgement (that is, subjective input) on the part of the analyst. It is the rules by which the probabilities are manipulated that should be (and are) objective. The explicit dependence on prior information allows one to study this dependence and to assess the robustness of the conclusions. Moreover, as noted above, the choice of ensemble in Frequentist analysis is invariably subjective; it is whatever a physicist judges to be sensible. A degree of subjectivity seems unavoidable in any realistic analysis, whether the approach is Frequentist or Bayesian. Critics charge that some aspects of the Bayesian approach are arbitrary. Although this charge is a serious one, it cannot be leveled at Bayesians without also being leveled at

15 15 Frequentists. The chief weakness of the Bayesian method is the difficulty of specifying a prior to represent minimal information about a parameter. If all that one is willing to say about a cross section to be measured is that it is a number between zero and forty milli-barns, then the following two-part question arises: can such minimal information be represented by a prior and, if so, how? Several different priors have been advocated [19] for parameters for which minimal information is available or assumed, but it is not clear how one is to choose between them. We believe, however, that this problem is not fatal, but merely serves as a warning that we may have to rely upon a convention, or a formal rule, to make progress. However, given that every serious scientific investigation entails subjective choices it is important to examine the sensitivity of one s conclusions to the choices one makes, including the choice of priors. Should it be found that the conclusions are not robust with respect to reasonable choices the appropriate response it to acquire more and better data. C The Method In our judgement, the benefits of a Bayesian method to construct credible intervals outweigh its principal defect. Its main benefit is a practical one: the method for constructing intervals is conceptually simple, even in the presence of nuisance parameters. Moreover, the method is general and mathematically coherent. The unsatisfactory aspect is the need to accept a convention, or a formal rule, for choosing certain prior probabilities, as described below: 1. Define the model. For the case of new particle searches, there is one accepted model: the expected number of events µ is related to the signal cross section x, the signal efficiency ɛ, the integrated luminosity L, and the expected background b through: µ = b + Lɛx. (IV.1)

16 16 In different analyses, one may replace b by some more complicated expression, containing perhaps a sum of many terms, each of which is a product of background acceptances, background cross sections, branching fractions, and integrated luminosity. Sometimes it is convenient to subsume Lɛ into a single parameter, the effective integrated luminosity, l Lɛ. 2. Given the model, determine the likelihood function for the data. In the case of counting experiments, there is one conventionally accepted likelihood function: it is proportional to the Poisson distribution with expectation value (mean) µ. The probability of observing k events, given an expectation value of µ, is P (k µ, I) = P oisson(µ, k). (IV.2) Here I indicates all the information used to build µ, as well as the assumptions that led us to choose the Poisson distribution. With the above model, the likelihood function is proportional to P (k x, L, ɛ, b, I) = e (b+lɛx) (b + Lɛx) k. (IV.3) k! Note that the above is P(data theory), where theory includes the parameters x, L, ɛ, and b; that is, it is the probability of observing k events, given x, L, ɛ, and b. It is not P(theory data), that is, it is not the probability for any of x, L, ɛ, or b, given k. 3. Assign prior probabilities to all parameters. Next, we use the available information (such as the knowledge of the integrated luminosity, within some bounds of uncertainty) to assign prior probabilities to each of the parameters in the problem. In general, the prior can contain correlations between the parameters. It is useful to think of the prior as P(theory), where theory includes both the cross section x and the nuisance parameters θ. When correlations are negligible, the prior can be factorized into a product of independent priors. Even when correlations are present, ignoring them would not be wrong,

17 17 but would rather correspond to accepting a less precise answer than could otherwise be obtained. In most cases, it will be of interest to factor the prior P (x, θ I) = P (x θ, I)P (θ I), = P (x I)P (θ I) (IV.4) into P (x I) and P (θ I). The choice of prior probabilities for the nuisance parameters is discussed in detail in Section VI B. The choice of prior for the signal cross section is more of a problem. As a matter of convention, we suggest adopting a flat prior of finite range: P (x I) = dx/m if 0 x M, 0 otherwise, (IV.5) where M is chosen sufficiently large so that the likelihood function for x > M is negligible, or more specifically, so that the answer does not depend on M, or remains finite as the limit M is taken at the end of the calculation. We have found that this conventional choice leads to answers that most physicists find intuitively reasonable, and which in the simple case of no background and no uncertainty in sensitivity, give the same upper limits as the Frequentist procedure. This works because in these circumstances, the cross section is tightly coupled to the expected number of signal events. Nonetheless, we recommend checking the robustness of limits by studying their behavior over a plausible class of priors. Appendix B describes an important case where the flat prior can cause problems in limit calculations. It is perhaps worth pointing out that we do not recommend choosing a signal prior flat in a theoretical parameter this can give rather different upper limits from our choice of flat prior in cross section. For example, the signal cross section can be a rather steeply falling function of a SUSY mass parameter, or of a contact term scale parameter. Choosing a prior flat in such a parameter is the same as choosing a strong dependence in cross section, and will thus represent a strong, rather than weak, opinion about the

18 18 likely signal count. As will be seen in Section VI A, such a steeply falling signal prior would result in significantly smaller upper limits than a flat prior. xxxxxxxxxxxx now should put in a specific example xxxxxxxxxxxxx compare DeMortier signal count flat 4. Apply Bayes Theorem to find the posterior probability. Bayes Theorem, when interpreted in a Bayesian way, relates the pre-data knowledge of the parameters (the prior probabilities) to the post-data knowledge of the parameters (the posterior probabilities), with the linkage being the likelihood function and a normalization constant. This process can be thought of as a form of logic [23], in which the prior knowledge and likelihood are premises for an inference whose outcome is the updated knowledge taking into account both the prior knowledge and the data. That is, Bayes theorem can be usefully sketched as P(theory data) P(data theory) P(theory). Stated using abstract notation, P (A BI) = P (B AI)P (A I) P (B I), (IV.6) where, in the context in which this theorem is used, P (A BI), P (B AI) and P (A I) are the posterior, likelihood and prior, respectively. The denominator in Eq. (IV.6) is determined through the normalization condition: P (A BI) = 1. all A (IV.7) To apply Bayes Theorem to our problem, we identify the following propositions: A is the proposition that the cross section lies between x and x+dx, the integrated luminosity is between L and L + dl, the signal efficiency is between ɛ and ɛ + dɛ, and the background count is between b and b + db B is the proposition that we have observed k events in the data I reflects all relevant prior knowledge. This includes the descriptions of the parameters x, L, ɛ, and b, as well as the assumptions that went into our model.

19 19 Thus Bayes Theorem, for our problem, becomes P (x, L, ɛ, b k, I) e (b+lɛx) (b + Lɛx) k P (x I)P (L, ɛ, b I), k! (IV.8) where the constant of proportionality is determined by the condition M 0 1 dx dl dɛ db f(x, L, ɛ, b k, I) = 1, (IV.9) where the probability density f(x, L, ɛ, b k, I) is defined by P (x, L, ɛ, b k, I) f(x, L, ɛ, b k, I)dx dl dɛ db. (IV.10) Our notation was careful to write Eq. (IV.8) in terms of densities times a differential, since, strictly speaking, Bayes Theorem is a theorem about probabilities, not probability densities. 5. Remove nuisance parameters. To remove the nuisance parameters L, ɛ, and b, we merely integrate Eq. (IV.8) over them. The result is the posterior distribution for the cross section x 1 P (x k, I) = dx dl dɛ db f(x, L, ɛ, b k, I), f(x k, I)dx. (IV.11) 6. Use the resulting posterior probability distribution to calculate quantities of interest. Subsequent to the Bayesian analysis, all information about the cross section is contained in the posterior density function for x. We recommend that this function be published along with the specific choices made for the prior probabilities. Nevertheless, it may be of interest to convey the most important information in as few numbers as possible. For example, in a search for some new process, the relevant number is usually the upper limit, calculated by integrating Eq. (IV.11) up to the upper limit as described in Section II B 1. Thus, the 95% CL upper limit U is obtained by solving 0.95 = U 0 dx f(x k, I). (IV.12)

20 20 If it is determined that the posterior distribution for x is peaked away from zero, it may then be more appropriate to quote the mean and the variance of the distribution and, of course, to claim discovery! V. A BAYESIAN EXAMPLE FOR ZERO BACKGROUND To illustrate the method described above we apply it to the problem considered by Cousins and Highland, namely, that of calculating an upper limit U on x for the case of zero background. We assume that the probability to observe n events is given by the Poisson distribution P (n x, l, I) = P oisson(xl, n), Eq. (III.2). The parameter x may be taken to be the cross section times the branching fraction. The sensitivity factor l is then the effective integrated luminosity; that is, l = ɛl. Following Cousins and Highland [8], we assume that we have an estimate ˆl of l with uncertainty δ l. As discussed in Section VI B, there are strong arguments to favor representing this information as a prior probability density that goes to zero at l = 0. Accordingly, we model the sensitivity information as a Gamma prior density Gamma(λ, a, b) (see Section VI B 4). A prior of this form is expected if the sensitivity is ultimately the result of a counting experiment that yields a count k, which is then scaled by some factor γ, that is, ˆl = γk, δ l = γ k, (V.1) (V.2) assuming that the probability to observe count k can be modeled by a Poisson distribution P (k λ, I) = P oisson(λ, k), (V.3) with mean count λ. When this likelihood is combined with a flat prior in λ it yields a Gamma posterior density Gamma(λ, k + 1, 1), which serves as an informative prior P (λ I) = Gamma(λ, k + 1, 1), (V.4)

21 21 for the likelihood function P (n x, l, I) = P (n y, λ) = P oisson(yλ, n), where we have used the fact that l = γλ and where y γx. We have here chosen to write the prior for λ instead of l to simplify the calculations that follow. If we are given ˆl and δl, we can find the effective count k and scale factor γ from k = η 2, γ = ˆlη 2, (V.5) (V.6) where η = δ l /ˆl is the relative uncertainty. In practice, it is often convenient first to integrate out the nuisance parameters P (n y, I) = = and then apply Bayes Theorem 0 P oisson(yλ, n) Gamma(λ, k + 1, 1) dλ, Γ(n + k + 1) y n, (V.7) Γ(n + 1)Γ(k + 1) (1 + y) n+k+1 P (y n, I) = P (n y, I) P (y I) 0 P (n y, I) P (y I), (V.8) where P (y I) = f(y I) dy is the prior for the scaled cross section y = γx. If we adopt the convention suggested above and take P (y I) to be flat we obtain P (y n, I) = Γ(n + k + 1) y n dy, (V.9) Γ(n + 1)Γ(k) (1 + y) n+k+1 as the posterior probability for y, from which upper limits, or credible intervals can be computed as described in Section IV C. In Table I we show the 90% upper limits computed using the posterior probability in Eq. (V.9), where, without loss of generality, we have set the estimated sensitivity ˆl = 1. VI. COVERAGE AND SENSITIVITY TO PRIOR OF BAYESIAN INTERVALS A Dependence on Choice of Prior for Cross Section The choice of a flat prior in cross section is clearly a convention. What motivates such a choice, and how dependent are the resulting intervals on the details of this choice? We have

22 22 n η U(n) a U(n) b TABLE I: (a) 90% Bayesian upper limits for the Cousins and Highland problem, computed using a Gamma prior for the sensitivity, with the estimated sensitivity ˆl = 1, for different relative errors η = δ l /ˆl. (b) For comparison we show the Cousins and Highland limits[8] computed using a Gaussian posterior probability for the sensitivity l. (Note: The translation to the R and S notation of Cousins and Highland is R x, S l, Ŝ ˆl and σ r η.)

23 23 performed the following tests. Our basic problem is finding limits on the cross section in the model µ = b + Lɛx = b + s. (VI.1) To examine the dependence on priors for s, take the expected background b, the luminosity and the efficiency as perfectly known constants, reducing our problem to finding Poisson upper confidence limits. In this discussion, we can substitute s for x, as they differ by a known scale factor (sensitivity) which we take for this discussion as ɛl = 1. We now consider the β = 95% upper limit for several plausible priors. The prior we recommend for s is = 1/M, where M is some maximum value. Jeffreys has offered various suggestions for priors [2, 19]. The constant prior is consistent with Jeffreys recommendations for a finite-range location parameter. For a semi-infinite parameter he suggested 1/s, but his analysis emphasizing invariance under change of variable leads to 1/ s for a Poisson mean. This latter is derived from a procedure usually referred to as Jeffreys prior by statisticians [27]. Together, these ideas suggest consideration of the family 1/s p. A second interesting family is the exponential family e as. Phenomenologically, this family places finite rather than infinite weight at s = 0 and still includes the flat prior as a member. More theoretically, this family results from a natural Bayesian updating procedure (see Appendix A) by taking as the prior the posterior of a previous experiment with a flat prior, with 0 events observed (independent of the original experiment s expected background). The parameter a gives the relative sensitivity of the previous experiment to the present experiment by a = ɛ 0 L 0 /ɛl. A previous experiment with one tenth the present experiment s sensitivity would have a = 0.1, while an equally sensitive previous experiment with which you wished to combine results would have a = 1. It is not difficult to see how the prior affects the limits. The Bayesian tool for limit calculation is the posterior density f(s k, I) = const P (k s + b)f(s I) = e s (s + b) k s p, (VI.2)

24 24 or f(s k, I) = const P (k s + b)f(s I) = e s(1+a) (s + b) k. (VI.3) We will set the right tail probability of f(s k, I) to 1 β. Changing the prior density f(s I) will change the shape of the posterior distribution and thus the upper limit on s. Priors which emphasize the region near s = 0 more heavily than the flat prior will produce smaller (more optimistic ) limits, while priors which emphasize larger s will produce larger ( more conservative ) upper limits. It is impossible to define an upper limit if the posterior is not normalizable, though this requirement does not necessarily imply that the prior itself must be normalizable. Consider values of s in the range from m to M (we can let the limits tend toward 0 and later). The Bayesian credible limits U of size β are defined by β = U m M f(s k, I) ds/ f(s k, I) ds. m In terms of the (un-normalized) cumulative distribution function (VI.4) w(x) = the definition of the interval becomes x m ds e s (s + b) k f(s I), (VI.5) β = w(u)/w(m), (VI.6) which is solved by U = w 1 (β w(m) ). (VI.7) The limits depend weakly on M so long as M is chosen sufficiently larger than k; larger M gives slightly larger limits. This is in accord with the expectation that if one really knows that M is small before the experiment, this knowledge should be incorporated in the estimation of the upper limit. The power family of priors shows considerable dependence on the chosen power. Jeffreys 1/s suggestion leads to an upper limit of 0 regardless of the number of events seen, unless

25 25 noxxxxcheckxxx events are seen and rigorously no background is expected. This choice concentrates too much probability near s = 0. Lest this prior be derided as absurd, it is equivalent to a prior flat in ln(s), and thus is invariant under transformation of variable to any power of s: a prior flat in the logarithm of mass would give the same result as s 1/mass 2. However, the square root prior in s is invariant under even more general transformations. For k = 0, we know that no signal event have been observed, and the limits are independent of b. The 95% limits are 1.9 for the 1/ s prior, 3.0 for the flat prior and 3.9 for a s prior. An e s prior would give 1.5 for the upper limit (half the value for the flat prior) as a result of combining two k = 0 experiments of equal sensitivity. The Figure C shows the dependence of the 95% upper limit for the cases of b = 3, k = 0, 3, 10 for the power prior as a function of the power (with 0 representing the flat prior). When we have substantial data (for example k = 10, b = 3 the assumptions of the prior become less important, but when little data are available, the prior counts significantly; this is the case with many search limit calculations. So we must choose our convention carefully. One of the many criteria which have been proposed for choosing priors to represent lack of relevant knowledge is whether the resulting credible intervals satisfy Frequentist coverage[19]. One attraction of choosing a criterion with good coverage properties is that it satisfies expectations built up by limit definitions in simpler situations. Further, in the simplest case of a Poisson mean with no background, a flat prior for the mean results in a limit of the same size as the Frequentist limit, giving a connection in continuity for the simple cases where the limit values can be calculated in either scheme. Though coverage is clearly a criterion outside the interest of the many Bayesians, it has been studied often in the statistical literature [19]. The coverage of Bayesian intervals can be shown to converge to Frequentist values as k becomes large, possibly as strongly as k 3/2 for carefully selected priors. For a lucid discussion of these issues see Ref. [20]. To summarize, the flat prior has some attractive features, but does not stand out as a special, stationary point in limit calculations. The choice of a prior matters most, in a case

26 26 such as calculation of upper limits, when there is insufficient data to dominate the prior. As a result, in a plausible range of priors (such as s to flat, the Bayesian limits may vary by as much as 30% xxxxxx. The conclusion to be drawn from this is that limits are intrinsically, unavoidably, soft when the data on which they are based are limited. xxxxx should include something from Narsky about coverage. B Dependence on Choice of Informative Prior for Efficiency and Luminosity The prior for the cross section is intended to be minimally-informative, that is, contain little of our opinions about the un-measured cross section. We will in this section study the sensitivity of the upper limit to varying forms of the informative prior for the nuisance parameters, each representing the same estimate and uncertainty. For simplicity, consider a measurement with a perfectly known expected background, b. The efficiency and luminosity ɛ, L are nuisance parameters, parameters we must estimate, but which are not the goal of the experiment. We typically estimate these parameters in side experiments, or by approximate calculation. We represent the results of these subsidiary measurements ˆɛ ± δ ɛ, ˆL±δL with informative priors for the unknown actual values ɛ, L of the nuisance parameters. We may find it convenient to combine these into an effective luminosity, l = ɛl. Then we will integrate out these nuisance parameters, with most weight being placed at the most likely values of the nuisance parameters. This results in an averaged likelihood, known as a marginal likelihood. We define the fractional uncertainty in the sensitivity by ) 2 ( δɛ η = + ˆɛ ( δl ˆL ) 2 = δ l ˆl. (VI.8) For the purposes of this section, it is convenient to define a scaled sensitivity variable φ = l/ˆl where ˆφ = 1 ± η. In this spirit, we will use η to parameterize the informative prior for φ, rather than adjust the posterior mean and rms of this distribution to precisely match the estimates. Without loss of generality, we can further consider unit expected sensitivity

27 27 ˆl = 1, so that s = lx = ˆlφx = φx and we can easily compare numerical values of the upper limits with other results. In the usual fashion, the posterior probability for the cross section will be given by P (x k) P (x) dφp (k φx + b)f(φ η) (VI.9) and the proportionality constant is fixed by dividing by the integral of the expression over x. f(φ η) is the informative prior density for the sensitivity. We now consider a number of possible forms for this informative prior, and their motivation. 1 Truncated Gaussian The most obvious candidate is a Gaussian form, 1 T Gauss(φ η) = η 2πZ exp 1 ( ) 2 φ 1, φ 0, 2 η Z = d φ 1 0 η 2π exp 1 ( ) 2 φ 1. 2 η (VI.10) (VI.11) As noted in Section V, because a Gaussian allows the possibility of negative values of sensitivity, it must be truncated so that φ 0, with the probability normalized to one by dividing by the integral over the allowed range. On its face, this seems somewhat odd, as the need for the truncation goes hand in hand with there being a nonzero probability density at φ = 0, meaning that it is possible that the experiment had no sensitivity. The truncation is a nearly negligible effect for small η, but the distribution is more problematic at larger values of η, culminating with φ = 0 not only possible, but, according to the truncated Gaussian, nearly the most probable value. Note that the effect of the truncation is to give a posterior φ mean sensitivity > 1 and a rms less than the expected value of η. 2 Lognormal Prior If the uncertainty is due to the sum of a number of small additive effects, a Gaussian, as above is an appropriate model. However, if the uncertainty is better modeled by a number

28 28 of small multiplicative errors, then a Gaussian in ln φ, the lognormal distribution [21], is more appropriate. For our case, lnor(φ η) = [ 1 ηφ 2π exp 1 ] 2 (ln φ/η)2, (VI.12) where we have again used ˆφ = 1 for simplicity, so that ln ˆφ = 0. The skewness of the distribution again comes into play: the posterior mean and rms of φ are e η2 /2 and e η2 1, both larger than expected values. 3 Beta prior One can also study some less phenomenological priors. Consider a sensitivity with an efficiency variable ɛ dominating the uncertainty, that is, δ L = 0, and η = δ ɛ /ɛ. A plausible prior density in this case is the Beta distribution, Beta(ɛ, a, b) = Γ(a + b) Γ(a)Γ(b) ɛa 1 (1 ɛ) b 1, 0 ɛ 1, (VI.13) which arises naturally from the following consideration. One performs a Monte Carlo simulation of n events, of which k pass certain cuts. The probability to obtain k counts out of n is given by the binomial distribution n Binomial(ɛ, k, n) = ɛ k (1 ɛ) n k, k (VI.14) which when combined with either a flat prior in ɛ, or a Beta prior, yields a Beta posterior density for the efficiency. This posterior density then serves as a Beta prior in a subsequent application of Bayes Theorem. The Beta distribution is the conjugate prior[22] for a binomial likelihood. This means that the posterior of a binomial variable with a Beta(a, b) prior is again a Beta distribution with a increased by the number of successes and b increased by the number of failures, to give a 1 total successes and b 1 total failures. A flat (a = b = 1) distribution in these terms represents no successes and no failures. (For comparison, the Jeffreys prior has a = b =

Journeys of an Accidental Statistician

Journeys of an Accidental Statistician Journeys of an Accidental Statistician A partially anecdotal account of A Unified Approach to the Classical Statistical Analysis of Small Signals, GJF and Robert D. Cousins, Phys. Rev. D 57, 3873 (1998)

More information

E. Santovetti lesson 4 Maximum likelihood Interval estimation

E. Santovetti lesson 4 Maximum likelihood Interval estimation E. Santovetti lesson 4 Maximum likelihood Interval estimation 1 Extended Maximum Likelihood Sometimes the number of total events measurements of the experiment n is not fixed, but, for example, is a Poisson

More information

Solution: chapter 2, problem 5, part a:

Solution: chapter 2, problem 5, part a: Learning Chap. 4/8/ 5:38 page Solution: chapter, problem 5, part a: Let y be the observed value of a sampling from a normal distribution with mean µ and standard deviation. We ll reserve µ for the estimator

More information

Harrison B. Prosper. CMS Statistics Committee

Harrison B. Prosper. CMS Statistics Committee Harrison B. Prosper Florida State University CMS Statistics Committee 08-08-08 Bayesian Methods: Theory & Practice. Harrison B. Prosper 1 h Lecture 3 Applications h Hypothesis Testing Recap h A Single

More information

Statistics for Particle Physics. Kyle Cranmer. New York University. Kyle Cranmer (NYU) CERN Academic Training, Feb 2-5, 2009

Statistics for Particle Physics. Kyle Cranmer. New York University. Kyle Cranmer (NYU) CERN Academic Training, Feb 2-5, 2009 Statistics for Particle Physics Kyle Cranmer New York University 91 Remaining Lectures Lecture 3:! Compound hypotheses, nuisance parameters, & similar tests! The Neyman-Construction (illustrated)! Inverted

More information

Unified approach to the classical statistical analysis of small signals

Unified approach to the classical statistical analysis of small signals PHYSICAL REVIEW D VOLUME 57, NUMBER 7 1 APRIL 1998 Unified approach to the classical statistical analysis of small signals Gary J. Feldman * Department of Physics, Harvard University, Cambridge, Massachusetts

More information

A Calculator for Confidence Intervals

A Calculator for Confidence Intervals A Calculator for Confidence Intervals Roger Barlow Department of Physics Manchester University England Abstract A calculator program has been written to give confidence intervals on branching ratios for

More information

Constructing Ensembles of Pseudo-Experiments

Constructing Ensembles of Pseudo-Experiments Constructing Ensembles of Pseudo-Experiments Luc Demortier The Rockefeller University, New York, NY 10021, USA The frequentist interpretation of measurement results requires the specification of an ensemble

More information

Statistics of Small Signals

Statistics of Small Signals Statistics of Small Signals Gary Feldman Harvard University NEPPSR August 17, 2005 Statistics of Small Signals In 1998, Bob Cousins and I were working on the NOMAD neutrino oscillation experiment and we

More information

Bayesian inference. Fredrik Ronquist and Peter Beerli. October 3, 2007

Bayesian inference. Fredrik Ronquist and Peter Beerli. October 3, 2007 Bayesian inference Fredrik Ronquist and Peter Beerli October 3, 2007 1 Introduction The last few decades has seen a growing interest in Bayesian inference, an alternative approach to statistical inference.

More information

Primer on statistics:

Primer on statistics: Primer on statistics: MLE, Confidence Intervals, and Hypothesis Testing ryan.reece@gmail.com http://rreece.github.io/ Insight Data Science - AI Fellows Workshop Feb 16, 018 Outline 1. Maximum likelihood

More information

Bayesian vs frequentist techniques for the analysis of binary outcome data

Bayesian vs frequentist techniques for the analysis of binary outcome data 1 Bayesian vs frequentist techniques for the analysis of binary outcome data By M. Stapleton Abstract We compare Bayesian and frequentist techniques for analysing binary outcome data. Such data are commonly

More information

Second Workshop, Third Summary

Second Workshop, Third Summary Statistical Issues Relevant to Significance of Discovery Claims Second Workshop, Third Summary Luc Demortier The Rockefeller University Banff, July 16, 2010 1 / 23 1 In the Beginning There Were Questions...

More information

Parameter estimation and forecasting. Cristiano Porciani AIfA, Uni-Bonn

Parameter estimation and forecasting. Cristiano Porciani AIfA, Uni-Bonn Parameter estimation and forecasting Cristiano Porciani AIfA, Uni-Bonn Questions? C. Porciani Estimation & forecasting 2 Temperature fluctuations Variance at multipole l (angle ~180o/l) C. Porciani Estimation

More information

P Values and Nuisance Parameters

P Values and Nuisance Parameters P Values and Nuisance Parameters Luc Demortier The Rockefeller University PHYSTAT-LHC Workshop on Statistical Issues for LHC Physics CERN, Geneva, June 27 29, 2007 Definition and interpretation of p values;

More information

Confidence Intervals. First ICFA Instrumentation School/Workshop. Harrison B. Prosper Florida State University

Confidence Intervals. First ICFA Instrumentation School/Workshop. Harrison B. Prosper Florida State University Confidence Intervals First ICFA Instrumentation School/Workshop At Morelia,, Mexico, November 18-29, 2002 Harrison B. Prosper Florida State University Outline Lecture 1 Introduction Confidence Intervals

More information

Physics 509: Error Propagation, and the Meaning of Error Bars. Scott Oser Lecture #10

Physics 509: Error Propagation, and the Meaning of Error Bars. Scott Oser Lecture #10 Physics 509: Error Propagation, and the Meaning of Error Bars Scott Oser Lecture #10 1 What is an error bar? Someone hands you a plot like this. What do the error bars indicate? Answer: you can never be

More information

Statistical Methods in Particle Physics

Statistical Methods in Particle Physics Statistical Methods in Particle Physics Lecture 11 January 7, 2013 Silvia Masciocchi, GSI Darmstadt s.masciocchi@gsi.de Winter Semester 2012 / 13 Outline How to communicate the statistical uncertainty

More information

Hypothesis Testing. Part I. James J. Heckman University of Chicago. Econ 312 This draft, April 20, 2006

Hypothesis Testing. Part I. James J. Heckman University of Chicago. Econ 312 This draft, April 20, 2006 Hypothesis Testing Part I James J. Heckman University of Chicago Econ 312 This draft, April 20, 2006 1 1 A Brief Review of Hypothesis Testing and Its Uses values and pure significance tests (R.A. Fisher)

More information

Fourier and Stats / Astro Stats and Measurement : Stats Notes

Fourier and Stats / Astro Stats and Measurement : Stats Notes Fourier and Stats / Astro Stats and Measurement : Stats Notes Andy Lawrence, University of Edinburgh Autumn 2013 1 Probabilities, distributions, and errors Laplace once said Probability theory is nothing

More information

Ridge regression. Patrick Breheny. February 8. Penalized regression Ridge regression Bayesian interpretation

Ridge regression. Patrick Breheny. February 8. Penalized regression Ridge regression Bayesian interpretation Patrick Breheny February 8 Patrick Breheny High-Dimensional Data Analysis (BIOS 7600) 1/27 Introduction Basic idea Standardization Large-scale testing is, of course, a big area and we could keep talking

More information

Lecture 5. G. Cowan Lectures on Statistical Data Analysis Lecture 5 page 1

Lecture 5. G. Cowan Lectures on Statistical Data Analysis Lecture 5 page 1 Lecture 5 1 Probability (90 min.) Definition, Bayes theorem, probability densities and their properties, catalogue of pdfs, Monte Carlo 2 Statistical tests (90 min.) general concepts, test statistics,

More information

Statistical Methods for Discovery and Limits in HEP Experiments Day 3: Exclusion Limits

Statistical Methods for Discovery and Limits in HEP Experiments Day 3: Exclusion Limits Statistical Methods for Discovery and Limits in HEP Experiments Day 3: Exclusion Limits www.pp.rhul.ac.uk/~cowan/stat_freiburg.html Vorlesungen des GK Physik an Hadron-Beschleunigern, Freiburg, 27-29 June,

More information

Harrison B. Prosper. CMS Statistics Committee

Harrison B. Prosper. CMS Statistics Committee Harrison B. Prosper Florida State University CMS Statistics Committee 7 August, 2008 Bayesian Methods: Theory & Practice. Harrison B. Prosper 1 h Lecture 2 Foundations & Applications h The Bayesian Approach

More information

Parameter Estimation. William H. Jefferys University of Texas at Austin Parameter Estimation 7/26/05 1

Parameter Estimation. William H. Jefferys University of Texas at Austin Parameter Estimation 7/26/05 1 Parameter Estimation William H. Jefferys University of Texas at Austin bill@bayesrules.net Parameter Estimation 7/26/05 1 Elements of Inference Inference problems contain two indispensable elements: Data

More information

Physics 403. Segev BenZvi. Credible Intervals, Confidence Intervals, and Limits. Department of Physics and Astronomy University of Rochester

Physics 403. Segev BenZvi. Credible Intervals, Confidence Intervals, and Limits. Department of Physics and Astronomy University of Rochester Physics 403 Credible Intervals, Confidence Intervals, and Limits Segev BenZvi Department of Physics and Astronomy University of Rochester Table of Contents 1 Summarizing Parameters with a Range Bayesian

More information

A Convolution Method for Folding Systematic Uncertainties into Likelihood Functions

A Convolution Method for Folding Systematic Uncertainties into Likelihood Functions CDF/MEMO/STATISTICS/PUBLIC/5305 Version 1.00 June 24, 2005 A Convolution Method for Folding Systematic Uncertainties into Likelihood Functions Luc Demortier Laboratory of Experimental High-Energy Physics

More information

Hypothesis testing. Chapter Formulating a hypothesis. 7.2 Testing if the hypothesis agrees with data

Hypothesis testing. Chapter Formulating a hypothesis. 7.2 Testing if the hypothesis agrees with data Chapter 7 Hypothesis testing 7.1 Formulating a hypothesis Up until now we have discussed how to define a measurement in terms of a central value, uncertainties, and units, as well as how to extend these

More information

Probability and Statistics

Probability and Statistics Probability and Statistics Kristel Van Steen, PhD 2 Montefiore Institute - Systems and Modeling GIGA - Bioinformatics ULg kristel.vansteen@ulg.ac.be CHAPTER 4: IT IS ALL ABOUT DATA 4a - 1 CHAPTER 4: IT

More information

Confidence intervals and the Feldman-Cousins construction. Edoardo Milotti Advanced Statistics for Data Analysis A.Y

Confidence intervals and the Feldman-Cousins construction. Edoardo Milotti Advanced Statistics for Data Analysis A.Y Confidence intervals and the Feldman-Cousins construction Edoardo Milotti Advanced Statistics for Data Analysis A.Y. 2015-16 Review of the Neyman construction of the confidence intervals X-Outline of a

More information

Chapter 9. Non-Parametric Density Function Estimation

Chapter 9. Non-Parametric Density Function Estimation 9-1 Density Estimation Version 1.2 Chapter 9 Non-Parametric Density Function Estimation 9.1. Introduction We have discussed several estimation techniques: method of moments, maximum likelihood, and least

More information

Estimation of reliability parameters from Experimental data (Parte 2) Prof. Enrico Zio

Estimation of reliability parameters from Experimental data (Parte 2) Prof. Enrico Zio Estimation of reliability parameters from Experimental data (Parte 2) This lecture Life test (t 1,t 2,...,t n ) Estimate θ of f T t θ For example: λ of f T (t)= λe - λt Classical approach (frequentist

More information

2. A Basic Statistical Toolbox

2. A Basic Statistical Toolbox . A Basic Statistical Toolbo Statistics is a mathematical science pertaining to the collection, analysis, interpretation, and presentation of data. Wikipedia definition Mathematical statistics: concerned

More information

Chapter Three. Hypothesis Testing

Chapter Three. Hypothesis Testing 3.1 Introduction The final phase of analyzing data is to make a decision concerning a set of choices or options. Should I invest in stocks or bonds? Should a new product be marketed? Are my products being

More information

Objective Bayesian Upper Limits for Poisson Processes

Objective Bayesian Upper Limits for Poisson Processes CDF/MEMO/STATISTICS/PUBLIC/5928 Version 2.1 June 24, 25 Objective Bayesian Upper Limits for Poisson Processes Luc Demortier Laboratory of Experimental High-Energy Physics The Rockefeller University...all

More information

Experiment 2 Random Error and Basic Statistics

Experiment 2 Random Error and Basic Statistics PHY9 Experiment 2: Random Error and Basic Statistics 8/5/2006 Page Experiment 2 Random Error and Basic Statistics Homework 2: Turn in at start of experiment. Readings: Taylor chapter 4: introduction, sections

More information

Part 2: One-parameter models

Part 2: One-parameter models Part 2: One-parameter models 1 Bernoulli/binomial models Return to iid Y 1,...,Y n Bin(1, ). The sampling model/likelihood is p(y 1,...,y n ) = P y i (1 ) n P y i When combined with a prior p( ), Bayes

More information

Chapter 5. Bayesian Statistics

Chapter 5. Bayesian Statistics Chapter 5. Bayesian Statistics Principles of Bayesian Statistics Anything unknown is given a probability distribution, representing degrees of belief [subjective probability]. Degrees of belief [subjective

More information

A General Overview of Parametric Estimation and Inference Techniques.

A General Overview of Parametric Estimation and Inference Techniques. A General Overview of Parametric Estimation and Inference Techniques. Moulinath Banerjee University of Michigan September 11, 2012 The object of statistical inference is to glean information about an underlying

More information

7. Estimation and hypothesis testing. Objective. Recommended reading

7. Estimation and hypothesis testing. Objective. Recommended reading 7. Estimation and hypothesis testing Objective In this chapter, we show how the election of estimators can be represented as a decision problem. Secondly, we consider the problem of hypothesis testing

More information

Chapter 9. Non-Parametric Density Function Estimation

Chapter 9. Non-Parametric Density Function Estimation 9-1 Density Estimation Version 1.1 Chapter 9 Non-Parametric Density Function Estimation 9.1. Introduction We have discussed several estimation techniques: method of moments, maximum likelihood, and least

More information

Bayesian Statistics. State University of New York at Buffalo. From the SelectedWorks of Joseph Lucke. Joseph F. Lucke

Bayesian Statistics. State University of New York at Buffalo. From the SelectedWorks of Joseph Lucke. Joseph F. Lucke State University of New York at Buffalo From the SelectedWorks of Joseph Lucke 2009 Bayesian Statistics Joseph F. Lucke Available at: https://works.bepress.com/joseph_lucke/6/ Bayesian Statistics Joseph

More information

Why Try Bayesian Methods? (Lecture 5)

Why Try Bayesian Methods? (Lecture 5) Why Try Bayesian Methods? (Lecture 5) Tom Loredo Dept. of Astronomy, Cornell University http://www.astro.cornell.edu/staff/loredo/bayes/ p.1/28 Today s Lecture Problems you avoid Ambiguity in what is random

More information

Physics 509: Bootstrap and Robust Parameter Estimation

Physics 509: Bootstrap and Robust Parameter Estimation Physics 509: Bootstrap and Robust Parameter Estimation Scott Oser Lecture #20 Physics 509 1 Nonparametric parameter estimation Question: what error estimate should you assign to the slope and intercept

More information

Treatment of Error in Experimental Measurements

Treatment of Error in Experimental Measurements in Experimental Measurements All measurements contain error. An experiment is truly incomplete without an evaluation of the amount of error in the results. In this course, you will learn to use some common

More information

Bayesian Statistical Methods. Jeff Gill. Department of Political Science, University of Florida

Bayesian Statistical Methods. Jeff Gill. Department of Political Science, University of Florida Bayesian Statistical Methods Jeff Gill Department of Political Science, University of Florida 234 Anderson Hall, PO Box 117325, Gainesville, FL 32611-7325 Voice: 352-392-0262x272, Fax: 352-392-8127, Email:

More information

Divergent Series: why = 1/12. Bryden Cais

Divergent Series: why = 1/12. Bryden Cais Divergent Series: why + + 3 + = /. Bryden Cais Divergent series are the invention of the devil, and it is shameful to base on them any demonstration whatsoever.. H. Abel. Introduction The notion of convergence

More information

Bayesian Inference: What, and Why?

Bayesian Inference: What, and Why? Winter School on Big Data Challenges to Modern Statistics Geilo Jan, 2014 (A Light Appetizer before Dinner) Bayesian Inference: What, and Why? Elja Arjas UH, THL, UiO Understanding the concepts of randomness

More information

A Discussion of the Bayesian Approach

A Discussion of the Bayesian Approach A Discussion of the Bayesian Approach Reference: Chapter 10 of Theoretical Statistics, Cox and Hinkley, 1974 and Sujit Ghosh s lecture notes David Madigan Statistics The subject of statistics concerns

More information

10. Exchangeability and hierarchical models Objective. Recommended reading

10. Exchangeability and hierarchical models Objective. Recommended reading 10. Exchangeability and hierarchical models Objective Introduce exchangeability and its relation to Bayesian hierarchical models. Show how to fit such models using fully and empirical Bayesian methods.

More information

Probability theory basics

Probability theory basics Probability theory basics Michael Franke Basics of probability theory: axiomatic definition, interpretation, joint distributions, marginalization, conditional probability & Bayes rule. Random variables:

More information

STAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01

STAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01 STAT 499/962 Topics in Statistics Bayesian Inference and Decision Theory Jan 2018, Handout 01 Nasser Sadeghkhani a.sadeghkhani@queensu.ca There are two main schools to statistical inference: 1-frequentist

More information

Chapter 4: An Introduction to Probability and Statistics

Chapter 4: An Introduction to Probability and Statistics Chapter 4: An Introduction to Probability and Statistics 4. Probability The simplest kinds of probabilities to understand are reflected in everyday ideas like these: (i) if you toss a coin, the probability

More information

Bayesian Inference for Normal Mean

Bayesian Inference for Normal Mean Al Nosedal. University of Toronto. November 18, 2015 Likelihood of Single Observation The conditional observation distribution of y µ is Normal with mean µ and variance σ 2, which is known. Its density

More information

1 Using standard errors when comparing estimated values

1 Using standard errors when comparing estimated values MLPR Assignment Part : General comments Below are comments on some recurring issues I came across when marking the second part of the assignment, which I thought it would help to explain in more detail

More information

Inference for a Population Proportion

Inference for a Population Proportion Al Nosedal. University of Toronto. November 11, 2015 Statistical inference is drawing conclusions about an entire population based on data in a sample drawn from that population. From both frequentist

More information

Conditional probabilities and graphical models

Conditional probabilities and graphical models Conditional probabilities and graphical models Thomas Mailund Bioinformatics Research Centre (BiRC), Aarhus University Probability theory allows us to describe uncertainty in the processes we model within

More information

Stat260: Bayesian Modeling and Inference Lecture Date: February 10th, Jeffreys priors. exp 1 ) p 2

Stat260: Bayesian Modeling and Inference Lecture Date: February 10th, Jeffreys priors. exp 1 ) p 2 Stat260: Bayesian Modeling and Inference Lecture Date: February 10th, 2010 Jeffreys priors Lecturer: Michael I. Jordan Scribe: Timothy Hunter 1 Priors for the multivariate Gaussian Consider a multivariate

More information

Should all Machine Learning be Bayesian? Should all Bayesian models be non-parametric?

Should all Machine Learning be Bayesian? Should all Bayesian models be non-parametric? Should all Machine Learning be Bayesian? Should all Bayesian models be non-parametric? Zoubin Ghahramani Department of Engineering University of Cambridge, UK zoubin@eng.cam.ac.uk http://learning.eng.cam.ac.uk/zoubin/

More information

Bayesian Methods for Machine Learning

Bayesian Methods for Machine Learning Bayesian Methods for Machine Learning CS 584: Big Data Analytics Material adapted from Radford Neal s tutorial (http://ftp.cs.utoronto.ca/pub/radford/bayes-tut.pdf), Zoubin Ghahramni (http://hunch.net/~coms-4771/zoubin_ghahramani_bayesian_learning.pdf),

More information

Statistical Methods for Particle Physics Lecture 4: discovery, exclusion limits

Statistical Methods for Particle Physics Lecture 4: discovery, exclusion limits Statistical Methods for Particle Physics Lecture 4: discovery, exclusion limits www.pp.rhul.ac.uk/~cowan/stat_aachen.html Graduierten-Kolleg RWTH Aachen 10-14 February 2014 Glen Cowan Physics Department

More information

A Measure of the Goodness of Fit in Unbinned Likelihood Fits; End of Bayesianism?

A Measure of the Goodness of Fit in Unbinned Likelihood Fits; End of Bayesianism? A Measure of the Goodness of Fit in Unbinned Likelihood Fits; End of Bayesianism? Rajendran Raja Fermilab, Batavia, IL 60510, USA PHYSTAT2003, SLAC, Stanford, California, September 8-11, 2003 Maximum likelihood

More information

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008

Gaussian processes. Chuong B. Do (updated by Honglak Lee) November 22, 2008 Gaussian processes Chuong B Do (updated by Honglak Lee) November 22, 2008 Many of the classical machine learning algorithms that we talked about during the first half of this course fit the following pattern:

More information

The dark energ ASTR509 - y 2 puzzl e 2. Probability ASTR509 Jasper Wal Fal term

The dark energ ASTR509 - y 2 puzzl e 2. Probability ASTR509 Jasper Wal Fal term The ASTR509 dark energy - 2 puzzle 2. Probability ASTR509 Jasper Wall Fall term 2013 1 The Review dark of energy lecture puzzle 1 Science is decision - we must work out how to decide Decision is by comparing

More information

An introduction to Bayesian reasoning in particle physics

An introduction to Bayesian reasoning in particle physics An introduction to Bayesian reasoning in particle physics Graduiertenkolleg seminar, May 15th 2013 Overview Overview A concrete example Scientific reasoning Probability and the Bayesian interpretation

More information

September Math Course: First Order Derivative

September Math Course: First Order Derivative September Math Course: First Order Derivative Arina Nikandrova Functions Function y = f (x), where x is either be a scalar or a vector of several variables (x,..., x n ), can be thought of as a rule which

More information

Sample size determination for a binary response in a superiority clinical trial using a hybrid classical and Bayesian procedure

Sample size determination for a binary response in a superiority clinical trial using a hybrid classical and Bayesian procedure Ciarleglio and Arendt Trials (2017) 18:83 DOI 10.1186/s13063-017-1791-0 METHODOLOGY Open Access Sample size determination for a binary response in a superiority clinical trial using a hybrid classical

More information

Error analysis for efficiency

Error analysis for efficiency Glen Cowan RHUL Physics 28 July, 2008 Error analysis for efficiency To estimate a selection efficiency using Monte Carlo one typically takes the number of events selected m divided by the number generated

More information

Fitting a Straight Line to Data

Fitting a Straight Line to Data Fitting a Straight Line to Data Thanks for your patience. Finally we ll take a shot at real data! The data set in question is baryonic Tully-Fisher data from http://astroweb.cwru.edu/sparc/btfr Lelli2016a.mrt,

More information

Statistical Data Analysis Stat 3: p-values, parameter estimation

Statistical Data Analysis Stat 3: p-values, parameter estimation Statistical Data Analysis Stat 3: p-values, parameter estimation London Postgraduate Lectures on Particle Physics; University of London MSci course PH4515 Glen Cowan Physics Department Royal Holloway,

More information

Confidence Intervals Lecture 2

Confidence Intervals Lecture 2 Confidence Intervals Lecture 2 First ICFA Instrumentation School/Workshop At Morelia,, Mexico, November 18-29, 2002 Harrison B. rosper Florida State University Recap of Lecture 1 To interpret a confidence

More information

Stat 542: Item Response Theory Modeling Using The Extended Rank Likelihood

Stat 542: Item Response Theory Modeling Using The Extended Rank Likelihood Stat 542: Item Response Theory Modeling Using The Extended Rank Likelihood Jonathan Gruhl March 18, 2010 1 Introduction Researchers commonly apply item response theory (IRT) models to binary and ordinal

More information

Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics

Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics Wooldridge, Introductory Econometrics, 4th ed. Appendix C: Fundamentals of mathematical statistics A short review of the principles of mathematical statistics (or, what you should have learned in EC 151).

More information

Statistics and Data Analysis

Statistics and Data Analysis Statistics and Data Analysis The Crash Course Physics 226, Fall 2013 "There are three kinds of lies: lies, damned lies, and statistics. Mark Twain, allegedly after Benjamin Disraeli Statistics and Data

More information

Why is the field of statistics still an active one?

Why is the field of statistics still an active one? Why is the field of statistics still an active one? It s obvious that one needs statistics: to describe experimental data in a compact way, to compare datasets, to ask whether data are consistent with

More information

Testing Simple Hypotheses R.L. Wolpert Institute of Statistics and Decision Sciences Duke University, Box Durham, NC 27708, USA

Testing Simple Hypotheses R.L. Wolpert Institute of Statistics and Decision Sciences Duke University, Box Durham, NC 27708, USA Testing Simple Hypotheses R.L. Wolpert Institute of Statistics and Decision Sciences Duke University, Box 90251 Durham, NC 27708, USA Summary: Pre-experimental Frequentist error probabilities do not summarize

More information

Delayed Choice Paradox

Delayed Choice Paradox Chapter 20 Delayed Choice Paradox 20.1 Statement of the Paradox Consider the Mach-Zehnder interferometer shown in Fig. 20.1. The second beam splitter can either be at its regular position B in where the

More information

Introduction to Applied Bayesian Modeling. ICPSR Day 4

Introduction to Applied Bayesian Modeling. ICPSR Day 4 Introduction to Applied Bayesian Modeling ICPSR Day 4 Simple Priors Remember Bayes Law: Where P(A) is the prior probability of A Simple prior Recall the test for disease example where we specified the

More information

Computing Likelihood Functions for High-Energy Physics Experiments when Distributions are Defined by Simulators with Nuisance Parameters

Computing Likelihood Functions for High-Energy Physics Experiments when Distributions are Defined by Simulators with Nuisance Parameters Computing Likelihood Functions for High-Energy Physics Experiments when Distributions are Defined by Simulators with Nuisance Parameters Radford M. Neal Dept. of Statistics, University of Toronto Abstract

More information

Bayesian Estimation An Informal Introduction

Bayesian Estimation An Informal Introduction Mary Parker, Bayesian Estimation An Informal Introduction page 1 of 8 Bayesian Estimation An Informal Introduction Example: I take a coin out of my pocket and I want to estimate the probability of heads

More information

Divergence Based priors for the problem of hypothesis testing

Divergence Based priors for the problem of hypothesis testing Divergence Based priors for the problem of hypothesis testing gonzalo garcía-donato and susie Bayarri May 22, 2009 gonzalo garcía-donato and susie Bayarri () DB priors May 22, 2009 1 / 46 Jeffreys and

More information

2016 SISG Module 17: Bayesian Statistics for Genetics Lecture 3: Binomial Sampling

2016 SISG Module 17: Bayesian Statistics for Genetics Lecture 3: Binomial Sampling 2016 SISG Module 17: Bayesian Statistics for Genetics Lecture 3: Binomial Sampling Jon Wakefield Departments of Statistics and Biostatistics University of Washington Outline Introduction and Motivating

More information

UNIVERSITY OF NOTTINGHAM. Discussion Papers in Economics CONSISTENT FIRM CHOICE AND THE THEORY OF SUPPLY

UNIVERSITY OF NOTTINGHAM. Discussion Papers in Economics CONSISTENT FIRM CHOICE AND THE THEORY OF SUPPLY UNIVERSITY OF NOTTINGHAM Discussion Papers in Economics Discussion Paper No. 0/06 CONSISTENT FIRM CHOICE AND THE THEORY OF SUPPLY by Indraneel Dasgupta July 00 DP 0/06 ISSN 1360-438 UNIVERSITY OF NOTTINGHAM

More information

Incompatibility Paradoxes

Incompatibility Paradoxes Chapter 22 Incompatibility Paradoxes 22.1 Simultaneous Values There is never any difficulty in supposing that a classical mechanical system possesses, at a particular instant of time, precise values of

More information

Lecture Notes 1: Decisions and Data. In these notes, I describe some basic ideas in decision theory. theory is constructed from

Lecture Notes 1: Decisions and Data. In these notes, I describe some basic ideas in decision theory. theory is constructed from Topics in Data Analysis Steven N. Durlauf University of Wisconsin Lecture Notes : Decisions and Data In these notes, I describe some basic ideas in decision theory. theory is constructed from The Data:

More information

2) There should be uncertainty as to which outcome will occur before the procedure takes place.

2) There should be uncertainty as to which outcome will occur before the procedure takes place. robability Numbers For many statisticians the concept of the probability that an event occurs is ultimately rooted in the interpretation of an event as an outcome of an experiment, others would interpret

More information

LECTURE 5 NOTES. n t. t Γ(a)Γ(b) pt+a 1 (1 p) n t+b 1. The marginal density of t is. Γ(t + a)γ(n t + b) Γ(n + a + b)

LECTURE 5 NOTES. n t. t Γ(a)Γ(b) pt+a 1 (1 p) n t+b 1. The marginal density of t is. Γ(t + a)γ(n t + b) Γ(n + a + b) LECTURE 5 NOTES 1. Bayesian point estimators. In the conventional (frequentist) approach to statistical inference, the parameter θ Θ is considered a fixed quantity. In the Bayesian approach, it is considered

More information

Lecture : Probabilistic Machine Learning

Lecture : Probabilistic Machine Learning Lecture : Probabilistic Machine Learning Riashat Islam Reasoning and Learning Lab McGill University September 11, 2018 ML : Many Methods with Many Links Modelling Views of Machine Learning Machine Learning

More information

The Laplace Rule of Succession Under A General Prior

The Laplace Rule of Succession Under A General Prior 1 The Laplace Rule of Succession Under A General Prior Kalyan Raman University of Michigan in Flint School of Management Flint, MI 48502 May 2000 ------------------------------------------------------------------------------------------------

More information

Statistical Methods in Particle Physics Lecture 2: Limits and Discovery

Statistical Methods in Particle Physics Lecture 2: Limits and Discovery Statistical Methods in Particle Physics Lecture 2: Limits and Discovery SUSSP65 St Andrews 16 29 August 2009 Glen Cowan Physics Department Royal Holloway, University of London g.cowan@rhul.ac.uk www.pp.rhul.ac.uk/~cowan

More information

Relationship between Least Squares Approximation and Maximum Likelihood Hypotheses

Relationship between Least Squares Approximation and Maximum Likelihood Hypotheses Relationship between Least Squares Approximation and Maximum Likelihood Hypotheses Steven Bergner, Chris Demwell Lecture notes for Cmpt 882 Machine Learning February 19, 2004 Abstract In these notes, a

More information

One-parameter models

One-parameter models One-parameter models Patrick Breheny January 22 Patrick Breheny BST 701: Bayesian Modeling in Biostatistics 1/17 Introduction Binomial data is not the only example in which Bayesian solutions can be worked

More information

Mathematical Statistics

Mathematical Statistics Mathematical Statistics MAS 713 Chapter 8 Previous lecture: 1 Bayesian Inference 2 Decision theory 3 Bayesian Vs. Frequentist 4 Loss functions 5 Conjugate priors Any questions? Mathematical Statistics

More information

Slope Fields: Graphing Solutions Without the Solutions

Slope Fields: Graphing Solutions Without the Solutions 8 Slope Fields: Graphing Solutions Without the Solutions Up to now, our efforts have been directed mainly towards finding formulas or equations describing solutions to given differential equations. Then,

More information

Just Enough Likelihood

Just Enough Likelihood Just Enough Likelihood Alan R. Rogers September 2, 2013 1. Introduction Statisticians have developed several methods for comparing hypotheses and for estimating parameters from data. Of these, the method

More information

A noninformative Bayesian approach to domain estimation

A noninformative Bayesian approach to domain estimation A noninformative Bayesian approach to domain estimation Glen Meeden School of Statistics University of Minnesota Minneapolis, MN 55455 glen@stat.umn.edu August 2002 Revised July 2003 To appear in Journal

More information

1 Measurement Uncertainties

1 Measurement Uncertainties 1 Measurement Uncertainties (Adapted stolen, really from work by Amin Jaziri) 1.1 Introduction No measurement can be perfectly certain. No measuring device is infinitely sensitive or infinitely precise.

More information

Probability, Entropy, and Inference / More About Inference

Probability, Entropy, and Inference / More About Inference Probability, Entropy, and Inference / More About Inference Mário S. Alvim (msalvim@dcc.ufmg.br) Information Theory DCC-UFMG (2018/02) Mário S. Alvim (msalvim@dcc.ufmg.br) Probability, Entropy, and Inference

More information

I. Induction, Probability and Confirmation: Introduction

I. Induction, Probability and Confirmation: Introduction I. Induction, Probability and Confirmation: Introduction 1. Basic Definitions and Distinctions Singular statements vs. universal statements Observational terms vs. theoretical terms Observational statement

More information

Invariant HPD credible sets and MAP estimators

Invariant HPD credible sets and MAP estimators Bayesian Analysis (007), Number 4, pp. 681 69 Invariant HPD credible sets and MAP estimators Pierre Druilhet and Jean-Michel Marin Abstract. MAP estimators and HPD credible sets are often criticized in

More information