Part III. A Decision-Theoretic Approach and Bayesian testing

Size: px

Start display at page:

Download "Part III. A Decision-Theoretic Approach and Bayesian testing"

Kelly Robinson
5 years ago
Views:

1 Part III A Decision-Theoretic Approach and Bayesian testing 1

2 Chapter 10 Bayesian Inference as a Decision Problem The decision-theoretic framework starts with the following situation. We would like to choose between various possible actions after observing data. Let Θ denote the set of all possible states of nature values of parameter), and let D denote the set of all possible decisions actions). A loss function is any function L : Θ D [0, ); the idea is that Lθ, d) gives the cost penalty) associated with decision d if the true state of the world is θ Inference as a decision problem Denote by fx, θ) the sampling distribution, for a sample x X. Let Lθ, δ) be our loss function. Often the decision d is to evaluate or estimate a function hθ) as accurately as possible. For point estimation, we want to evaluate hθ) = θ; our set of decisions is D = Θ, the parameter space; and Lθ, d) is the loss in reporting d when θ is true. For hypothesis testing, if we want to test H 0 : θ Θ 0, then our possible decisions are D = {accept H 0, reject H 0 }. 2

3 We would like to evaluate the function hθ) = { 1 if θ Θ0 0 otherwise as accurately as possible. Our loss function is Lθ, accept H 0 ) = Lθ, reject H 0 ) = { l00 if θ Θ 0 l 01 otherwise { l10 if θ Θ 0 l 11 otherwise. Note: l 01 is the Type II-error, accept H 0 although false), and l 10 is the Type I-error reject H 0 although true) Decision rules and Bayes estimators In general we would like to find a decision rule: δ : X D, a function which makes a decision based on the data. We would like to choose δ such that we incur only a small loss. In general there is no δ that uniformly mimimizes Lθ, δx)). In a Bayesian setting, for a prior π and data x X, the posterior expected loss of a decision is a function of the data x, defined as ρπ, d x) = Lθ, d)πθ x)dθ. Θ For a prior π the integrated risk of a decision rule δ is the real number defined as rπ, δ) = Lθ, δx)fx θ)dxπθ)dθ. We prefer δ 1 to δ 2 if and only if rπ, δ 1 ) < rπ, δ 2 ). Θ X Proposition. An estimator minimizing rπ, δ) can be obtained by selecting, for every x X, the value δx) that minimizes ρπ, δ x). 3

4 Proof additional material) rπ, δ) = Lθ, δx))fx θ)dxπθ)dθ Θ X = Lθ, δx))πθ x)px)dθdx = X Θ X ρπ, δ x)px)dx Recall that px) = fx θ)πθ)dθ).) This proves the assertion. Θ A Bayes estimator associated with prior π, loss L, is any estimator δ π which minimizes rπ, δ): For every x X it is δ π = arg min ρπ, d x). d Then rπ) = rπ, δ π ) is called Bayes risk. This is valid for proper priors, and for improper priors if rπ) <. If rπ) = one can define a generalized Bayes estimator as the minimizer, for every x, of ρπ, d x). Fact: For strictly convex loss functions, Bayes estimators are unique Some common loss functions In principle the loss function is part of the problem specification. A very popular choice is squared error loss Lθ, d) = θ d) 2. This loss function is convex, and penalizes large deviations heavily. Proposition. The Bayes estimator δ π associated with prior π under squared error loss is the posterior mean, δ π x) = E π Θ θ x) = θfx θ)πθ)dθ Θ fx θ)πθ)dθ. To see this, recall that for any random variable Y, EY a) 2 ) is minimized by a = EY. Another common choice for the loss function is absolute error loss Lθ, d) = θ d. Proposition: The posterior median is a Bayes estimator under absolute error loss. 4

5 10.4 Bayesian testing Let fx, θ) be our sampling distribution, x X, θ Θ, and suppose that we cant to test H 0 : θ Θ 0. Then the set of possible decisions is D = {accept H 0, reject H 0 } = {1, 0}, where 1 stands for acceptance. We use the loss function Lθ, φ) = 0 if θ Θ 0, φ = 1 a 0 if θ Θ 0, φ = 0 0 if θ Θ 0, φ = 0 a 1 if θ Θ 0, φ = 1. Proposition Under this loss function, the Bayes decision rule associated with a prior distribution π is { 1 if P φ π π θ Θ 0 x) > a 1 x) = a 0 +a 1 0 otherwise Note: In the special case that a 0 = a 1, the rule states that we accept H 0 if P π θ Θ 0 x) > 1 2. Proof additional material) The posterior expected loss is ρπ, φ x) = a 0 P π θ Θ 0 x)1φx) = 0) + a 1 P π θ Θ 0 x)1φx) = 1) = a 0 P π θ Θ 0 x) + 1φx) = 1) a 1 a 0 + a 1 )P π θ Θ 0 x)), and a 1 a 0 + a 1 )P π θ Θ 0 x) < 0 if and only if P π θ Θ 0 x) > a 1 a 0 +a 1. Example: X Binn, θ), Θ 0 = [0, 1/2), πθ) = 1; then P π θ < 1 2 x ) = = θx 1 θ) n x dθ 1 0 θx 1 θ) n x dθ 1 n+1 2) Bx + 1, n x + 1) { } 1 n x)!x! , x + 1 n + 1)! which can be evaluated for particular n and x, and compared with the acceptance level a 1 a 0 +a 1. 5

6 Example: X N θ, σ 2 ), σ 2 is known, θ N µ, τ 2 ). We already calculated To test H 0 : θ < 0: Let z a0,a 1 be the equivalently, if πθ x) N µx), w 2 ) µx) = σ2 µ + τ 2 x σ 2 + τ 2 w 2 σ 2 τ 2 = σ 2 + τ. 2 P π θ < 0 x) = P π θ µx) w a 1 a 0 +a 1 < µx) ) = Φ µx) ) w w quantile; then we accept H 0 if µx) > z a0,a 1 w, or, x < σ2 τ 2 µ 1 + σ2 τ 2 ) z a0,a 1 w. For σ 2 = 1, µ = 0, τ 2, we accept H 0 if x < z a0,a 1. Compare this to the frequentist test: Agian σ 2 = 1. Accept H 0 if x < z 1 α = z α. This corresponds to a 0 a 1 = 1 1. So a 0 α a 1 = 19 for α = 0.05, and = 99 for α = 0.01, for example. a 0 a 1 Note: If the prior probability of H 0 is 0, then so will be posterior probability. Testing H 0 : θ = θ 0 against H 1 : θ > θ 0 often really means testing H 0 : θ θ 0 against H 1 : θ > θ 0, which is natural to test in a Bayesian setting. is Definition: The Bayes factor for testing H 0 : θ Θ 0 against H 1 : θ Θ 1 B π x) = P π θ Θ 0 x)/p π θ Θ 1 x). P π θ Θ 0 )/P π θ Θ 1 ) It measures the extent to which the data x will change the odds of Θ 0 relative to Θ 1. 6

7 If B π x) > 1 the data adds support to H 0 ; if B π x) < 1 the data adds support to H 1 ; if B π x) = 1 the data does not help to distinguish between H 0 and H 1. Note that the Bayes factor still depends on the prior π. Special case: If H 0 : θ = θ 0, H 1 : θ = θ 1, then is the likelihood ratio. B π x) = fx θ 0) fx θ 1 ) More generally, / B π x) = Θ 0 πθ)fx θ)dθ Θ 0 πθ)dθ Θ 1 πθ)fx θ)dθ Θ 1 πθ)dθ = Θ 0 πθ)fx θ)/p π θ Θ 0 )dθ Θ 1 πθ)fx θ))/p π θ Θ 1 )dθ = px θ Θ 0) px θ Θ 1 ) is the ratio of how likely the data is under H 0 and how likely the data is under H 1. Compare this to the frequentist likelihood ratio Λx) = max θ Θ 0 fx θ) max θ Θ1 fx θ). In Bayesian statistics we average instead of taking maxima. Note: with φ π from the Proposition, and ρ 0 = P π θ Θ 0 ), ρ 1 = P π θ Θ 1 ), B π x) = P π θ Θ 0 x)/1 P π θ Θ 0 x)) ρ 0 /ρ 1 and so / φ π x) = 1 B π x) > a 1 ρ 0. a 0 ρ 1 7

8 Also, by inverting the equality it follows that P π θ Θ 0 x) = 1 + ρ ) 1 1 B π x)) 1. ρ 0 Example: Let X Binn, p), H 0 : p = 1/2, H 1 : p 1/2. Choose as prior an atom of size ρ 0 at 1/2, otherwise uniform; then B π x) = px p = 1/2) px p Θ 1 ) = n ) x 2 n n. x) Bx + 1, n x + 1) So P p = 12 ) x = If ρ 0 = 1/2, n = 5, x = 3: B π x) = 15 8 P p = 12 ) x = ρ 0) ρ 0 > 1, then ) 1 x!n x)! 2 n. n 1)! ) = ; so the data adds support to H 0, the posterior probability of H 0 is 15/23 > 1/2. We could have chosen an alternatively prior: atom of size ρ 0 at 1/2, otherwise Beta1/2, 1/2); this prior favours 0 and 1. Then we obtain for n=10: x P p = 1 x) Example: Let X N θ, σ 2 ), σ 2 is known, H 0 : θ = 0. We choose as prior mass ρ 0 at θ = 0, otherwise N 0, τ 2 ). Then B π ) 1 = px θ 0) px θ = 0) = σ2 + τ 2 ) 1/2 exp{ x 2 /2σ 2 + τ 2 ))} σ 1 exp{ x 2 /2σ 2 )} 8

9 and P θ = 0 x) = ρ 0 ρ 0 σ 2 σ 2 + τ exp 2 τ 2 x 2 2σ 2 σ 2 + τ 2 ) ) ) 1. Example: ρ 0 = 1/2, τ = σ, put z = x/σ x P θ = 0 z) For τ = 10σ more diffusive prior) x P θ = 0 z) so x gives stronger support for H 0 than under the tighter prior. Note: For x fixed, τ 2, ρ 0 > 0, we have P θ = 0 x) 1. For the noninformative prior πθ) 1 we have that px πθ) = 2πσ 2 ) 1/2 e x θ)2 2σ 2 dθ = 2πσ 2 ) 1/2 e θ x)2 2σ 2 dθ and so P θ = 0 x) = which is not equal to 1. = ρ ) 1 0 2π expx 2 /2) ρ 0 9

10 Lindley s paradox: Let X N θ, σ 2 x /n), H 0 : θ = 0, n is fixed. If σ/ n) is large enough to reject H 0 in classical test, then for large enough τ 2 the Bayes factor will be larger than 1, indicating support for H 0. In contrast, if σ 2, τ 2 x are fixed, and n such that σ/ = k n) α fixed, which is just significant at level α in classical test, then B π x). Results which are just significant at some fixed level in the classical test will, for large n, actually be much more likely under H 0 than under H 1. A very diffusive prior proclaims great scepticism, which may overwhelm the contrary evidence of the observations Least favourable Bayesian answers The choice of prior affects the result. Suppose that we want to test H 0 : θ = θ 0 against H 1 : θ θ 0, and the prior probability on H 0 is ρ 0 = 1/2. Which prior g in H 1, after observing x, would be least favourable to H 0? Let G be a family of priors on H 1 ; put and P x, G) = Bx, G) = inf g G fx θ 0 ) Θ fx θ)gθ)dθ ) 1 fx θ 0 ) fx θ 0 ) + sup fx θ)gθ)dθ = g G Bx, G) Θ A Bayesian prior g G on H 0 will then have posterior probability at least P x, G) on H 0 for ρ 0 = 1/2). If ˆθ is the m.l.e. of θ, and if G A is the set of all prior distributions, then Bx, G A ) = fx θ 0) fx ˆθx)) and P x, G A ) = 1 + fx ˆθx)) fx θ 0 ) 1 10

11 Other natural families are G S the set of distributions symmetric around θ 0, and G SU the set of unimodal distributions symmetric around θ 0. Example: Let X N θ, 1), H 0 : θ = θ 0, H 1 : θ θ 0, then we can calculate p-value P x, G A ) P x, G SU ) The Bayesian approach will typically reject H 0 less frequently than the frequentist approach Consider the general test problem H 0 : θ = θ 0 against H 1 : θ = θ 1, and consider repetitions in which one uses the most powerful test with level α = In frequentist analysis, only 1% of the true H 0 will be rejected. But this does not say anything about the proportion of errors made when rejecting! Example: Suppose the probability of type II error is 0.99, and θ 0 and θ 1 occur equally often, then about half of the rejections of H 0 will be in error Comparison with frequentist hypothesis testing Frequentist hypothesis testing: Asymmetry between H 0, H 1 : fix type I error, minimize type II error; UMP tests do not always exist general 2-sided tests, e.g.); p-values: - have no intrinsic optimality, space of p-values lacks a decision-theoretic foundation - are routinely misinterpreted - do not take type II error into account; confidence regions: are a pre-data measure, can often have very different post data coverage probabilities. 11

12 Example: Let X N θ, 1/2), H 0 : θ = 1, H 1 : θ = 1, and suppose that x = 0. Then the UMP p-value is 0.072, but the p-value for the test of H 1 against H 0 takes exactly the same value. Example: Let X 1,..., X n be i.i.d. N θ, σ 2 ), both θ, σ 2 are unknown. Then a 1001 α) confidence for θ is ) s s C = x t α/2, x + t α/2. n n For n = 2, α = 0.5, the predata coverage probability is 0.5. However, Brown Ann.Math.Stat. 38, 1967, ) showed that P θ C x /s < 1 + 2) > 2/3. Bayesian testing compares the probability of the actual data under the two hypotheses 12

13 Chapter 11 Hierarchical and empirical Bayesian methods 11.1 Hierarchical Bayes: A hierarchical model consists of modelling a parameter θ through randomness at different levels; for example, θ β π 1 θ β), where β π 2 β); so that then πθ) = π 1 θ β)π 2 β)dβ. When dealing with complicated posterior distributions, rather than evaluating the integrals, we might use simulation to approximate the integrals. For simulation in hierarchical models, we simulate first from β, then, given β, we simulate from θ. We hope that the distribution of β is easy to simulate, and also that the conditional distribution of θ given β is easy to simulate. This approach is particularly useful for MCMC Markov chain Monte Carlo) methods, e.g.: see next term Empirical Bayes: Let x fx θ). The empirical Bayes method chooses a convenient prior family πθ λ) typically conjugate), where λ is a hyperparameter, so px λ) = fx θ)πθ λ)dθ. 13

14 Rather than specifying λ, we estimate λ by ˆλ, for example by frequentist methods, based on px λ), and we substitute ˆλ for λ; πθ x, ˆλ) is called a pseudo-posterior. We plug it into Bayes Theorem for inference. The empirical Bayes approach * is neither fully Bayesian nor fully frequentist; * depends on ˆλ, different ˆλ will lead to different procedures; * if ˆλ is consistent, then asymptotically will lead to coherent Bayesian analysis. * often outperforms classical estimators in empirical terms. Example: James-Stein estimators Let X i N θ i, 1) be independent given θ i, i = 1,..., p,where p 3. In vector notation: X N θ, I p ). Here the vector θ is random; assume that we have realizations θ i, i = 1,..., p. The obvious estimate which is the least-squares estimate) for θ i is ˆθ i = x i, leading to ˆθ = X. Abbreviate X = x. Our decision rule would hence be δx) = x. But this is not the best rule! Assume that θ i N 0, τ 2 ), then px τ 2 ) = N 0, 1 + τ 2 )I p ), and the posterior for θ given the data is ) τ 2 θ x N 1 + τ x, τ I 2 p. Under quadratic loss, the Bayes estimator δx) of θ is the posterior mean τ τ 2 x. In the empirical Bayes approach, we would use the m.l.e. for τ 2, which is ) τˆ x 2 2 = 1 1 x 2 > p), p 14

15 and the empirical Bayes estimator is the estimated posterior mean, δ EB x) = ˆτ ˆτ x = 2 1 p ) + x x 2 is the truncated James-Stein estimator. It can can be shown to outperform the estimator δx) = x. Alternatively, the best unbiased estimator of 1/1 + τ 2 ) is p 2 x 2, giving δ EB x) = 1 p ) x. x 2 This is the James-Stein estimator. It can be shown that under quadratic loss function the James-Stein estimator has smaller integrated risk than δx) = x. Note: both estimators tend to shrink towards 0. It is now known to be a very general phenomenon that when comparing three or more populations, the sample mean is not the best estimator. Shrinkage estimators are an active area of research. Bayesian computation of posterior probabilities can be very computerintensive; see the MCMC and Applied Bayesian Statistics course. 15

16 Chapter 12 Principles of Inference 12.1 The Likelihood Principle The Likelihood Principle: The information brought by an observation x about θ is entirely contained in the likelihood function Lθ x). From this follows: if x 1 and x 2 are two observations with likelihoods L 1 θ x) and L 2 θ x), and if L 1 θ x) = cx 1, x 2 )L 2 θ x) then x 1 and x 2 must lead to identical inferences. Example. We know that πθ x) fx θ)πθ). If f 1 x θ) f 2 x θ) as a function of θ, then they have the same posterior, so they lead to the same Bayesian inference. Example: Binomial versus negative binomial. a) Let X Binn, θ) be the number of successes in n independent trials, with p.m.f. ) n f 1 x θ) = θ x 1 θ) n x x then πθ x) ) n θ x 1 θ) n x πθ) x θ x 1 θ) n x πθ). 16

17 b) Let N NegBinx, θ) be the number of independent trials until x successes, with p.m.f. ) n 1 f 2 n θ) = θ x 1 θ) n x x 1 and πθ x) θ x 1 θ) n x πθ). Bayesian inference about θ does not depend on whether a binomial or a negative binomial sampling scheme was used. M.l.e. s satisfy the likelihood principle, but many frequentist procedures do not! Example: Binn, θ)-sampling. We observe x 1,..., x n ) = 0,..., 0, 1). An unbiased estimate for θ is ˆθ = 1/n. If instead we view n as geometricθ), then the only unbiased estimator for θ is ˆθ = 1n = 1). Unbiasedness typically violates the likelihood principle: it involves integrals over the sample space, so it depends on the value of fx θ) for values of x other than the observed value. Example: a) We observe a Binomial randome variable, n=12; we observe 9 heads, 3 tails. Suppose that we want to test H 0 : θ = 1/2 against H 1 : θ > 1/2. We can calculate that the UMP test has P X 9) = b) If instead we continue tossing until 3 tails recorded, and observe that N = 12 tosses are needed, then the underlying distribution is negative binomial, and P N 12) = The conditionality perspective Example Cox 1958) A scientist wants to measure a physical quantity θ. Machine 1 gives measurements X 1 N θ, 1), but is often busy. Machine 2 gives measurements X 1 N θ, 100). The availability of machine 1 is beyond the scientist s control, independent of object to be measured. Assume that on any given occasion machine 1 is available with probability 1/2; if available, the scientist chooses machine 1. 17

18 A standard 95% confidence interval is about x 16.4, x ) because of the possibility that machine 2 was used. Conditionality Principle: If two experiments on the parameter θ are available, and if one of these two experiments is selected with probability 1/2, then the resulting inference on θ should only depend on the selected experiment. The conditionality principle is satisfied in Bayesian analysis. In the frequentist approach, we could condition on an ancillary statistic, but such statistic is not always available. A related principle is the Stopping rule principle SRP): A it stopping rule is a random variable that tells when to stop the experiment; this random variable depends only on the outcome of the first n experiments does not look into the future). The stopping rule principle states that if a sequence of experiments is directed by a stopping rule, then, given the resulting sample, the inference about θ should not depend on the nature of the stopping rule. The likelihood principle implies the SRP. The SRP is satisfied in Bayesian inference, but it is not always satisfied in frequentist analysis. 18

A Very Brief Summary of Bayesian Inference, and Examples

A Very Brief Summary of Bayesian Inference, and Examples Trinity Term 009 Prof Gesine Reinert Our starting point are data x = x 1, x,, x n, which we view as realisations of random variables X 1, X,, X