Approximate Bayesian computation for spatial extremes via...

35
Approximate Bayesian computation for spatial extremes via open-faced sandwich adjustment Ben Shaby SAMSI August 3, 2010 Ben Shaby (SAMSI) OFS adjustment August 3, 2010 1 / 29

Transcript of Approximate Bayesian computation for spatial extremes via...

Page 1: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Approximate Bayesian computation for spatial extremesvia open-faced sandwich adjustment

Ben Shaby

SAMSI

August 3, 2010

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 1 / 29

Page 2: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Outline

1 Introduction

2 Spatial Extremes

3 The OFS adjustment

4 Simulation

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 2 / 29

Page 3: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Outline

1 Introduction

2 Spatial Extremes

3 The OFS adjustment

4 Simulation

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 3 / 29

Page 4: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

“Posterior” distributions

I will describe how to draw from a “posterior” distribution.

Note the finger quotes.

What do we want out of a posterior distribution?

For our purposes, we will want a distribution that1 Describes our state of knowledge (uncertainty) about a parameter.2 Produces equi-tailed credible intervals that have nominal frequentist

coverage rates.

This is not a very Bayesian view!

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 4 / 29

Page 5: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

“Posterior” distributions

I will describe how to draw from a “posterior” distribution.

Note the finger quotes.

What do we want out of a posterior distribution?

For our purposes, we will want a distribution that1 Describes our state of knowledge (uncertainty) about a parameter.2 Produces equi-tailed credible intervals that have nominal frequentist

coverage rates.

This is not a very Bayesian view!

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 4 / 29

Page 6: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

A good “posterior”’

An ideal ''posterior''

empirical density of θπ(θ|x)

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 5 / 29

Page 7: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Outline

1 Introduction

2 Spatial Extremes

3 The OFS adjustment

4 Simulation

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 6 / 29

Page 8: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Extreme values

Of the environmental variables we care about, usually what we really careabout are the extremes.

heat waves

storms

sea levels

Why do we care?

manage risk (insurance, etc.)

emergency preparedness

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 7 / 29

Page 9: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Floods

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 8 / 29

Page 10: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Heat waves

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 9 / 29

Page 11: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Extreme values

“Extreme values” can mean many things

We consider only “block maxima” (block minima).

Asymptotically follow generalized extreme value (GEV) distribution

G(x) = exp

−[1− ξ

(x− ητ

)]−1/ξ

+

where z+ = max(z, 0).

η is a location and τ a scale parameter.

ξ is a shape parameter, and determines the tail behavior.

Then any maximal process should have GEV marginal distributions!

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 10 / 29

Page 12: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Spatial extremes

It is possible to construct processes with spatial structure and GEVmarginals. This leads us to max stable processes.

The “Smith model” is one example.

0 2 4 6 8 10

0.0

0.1

0.2

0.3

0.4

x

Z(x

)

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 11 / 29

Page 13: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Smith process in 2 dimensions

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 12 / 29

Page 14: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

GP margins

Lest you fear that this process is unrealistic, the margins don’t have to bethe same everywhere.

Unit Frechet margins Gaussian process GEV parameters

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 13 / 29

Page 15: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Pairwise likelihoods

Unfortunately, joint likelihoods for the Smith process are not knownfor n ≥ 2.

But we can write the pairwise likelihood, a form of compositelikelihood.

Lp(θ;y) =∏i 6=j

f(yi, yj ;θ)

It turns out that Lp(θ;y) that behaves “similarly” to the likelihood.

Can we trick MCMC into doing something useful with Lp(θ;y)?

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 14 / 29

Page 16: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Pairwise likelihoods

Unfortunately, joint likelihoods for the Smith process are not knownfor n ≥ 2.

But we can write the pairwise likelihood, a form of compositelikelihood.

Lp(θ;y) =∏i 6=j

f(yi, yj ;θ)

It turns out that Lp(θ;y) that behaves “similarly” to the likelihood.

Can we trick MCMC into doing something useful with Lp(θ;y)?

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 14 / 29

Page 17: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The quasi-posterior

Yes!

We define the quasi-posterior distribution as

πp,n(θ|yn) =Lp,n(θ;yn)π(θ)∫

Θ Lp,n(θ;yn)π(θ) dθ,

We will assume, for convenience, that π(θ) proper.

Lp,n is not necessarily a density, so πp,n(θ|yn) is not a true posterior.

Lp,n is integrable, so as long as the prior π(θ) is proper, thenπp,n(θ|Zn) will be a proper density.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 15 / 29

Page 18: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

More definitions

Now we can write down a quasi-Bayes estimator

Define loss in the usual way.

Define quasi-posterior risk Rn(θ) as the quasi-posterior expectationof loss.

The pairwise quasi-Bayes estimator is then

θQB = argminθ∈Θ

Rn(θ).

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 16 / 29

Page 19: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Outline

1 Introduction

2 Spatial Extremes

3 The OFS adjustment

4 Simulation

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 17 / 29

Page 20: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The sandwich matrix

Pn = E0[∇0`p,n∇0`′p,n]

Bn = −E0[∇20`p,n]

Sn = Bn P−1n Bn

Bread

Peanut butter

Bread

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 18 / 29

Page 21: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The sandwich matrix

Pn = E0[∇0`p,n∇0`′p,n]

Bn = −E0[∇20`p,n]

Sn = Bn P−1n Bn

Bread

Peanut butter

Bread

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 18 / 29

Page 22: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The sandwich matrix

Pn = E0[∇0`p,n∇0`′p,n]

Bn = −E0[∇20`p,n]

Sn = Bn P−1n Bn

Bread

Peanut butter

Bread

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 18 / 29

Page 23: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Asymptotic normality of quasi-Bayes estimators

Then as long as we don’t use a crazy prior, Chernozhukov and Hong(2003) says that:

Theorem

S1/2n (θQB − θ0)

D−→ N(0, I)

When we use pairwise likelihoods for MCMC, the sandwich matrixdescribes the (asymptotic) sampling variability of the estimator.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 19 / 29

Page 24: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Convergence of the quasi-posterior

Furthermore, (also from Chernozhukov and Hong, 2003)

Theorem

Asymptotically, πp,n(θ|yn) ∼ N(θ0,B−1n ).

This has important consequences for inference from the MCMC sample!

B−1 6= B−1PB−1!

⇒ Equi-tailed credible intervals based on MCMC quantiles will NOThave the correct frequentist coverage probabilities

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 20 / 29

Page 25: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

“Distortion” of the posterior

The two curves are very different!

0.5 1.0 1.5 2.0

02

46

8

θ

Den

sity

empirical density of θπp (θ|x)

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 21 / 29

Page 26: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The OFS adjustment

The main idea:

Whereas θQB is distributed like a sandwich normal (S−1n ), the

quasi-posterior looks like a single slice of bread normal (B−1n ).

We want to complete the sandwich by joining the slice of bread B−1n

to the open-faced sandwich BnP−1n to get S−1

n .

−→

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 22 / 29

Page 27: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The OFS adjustment

The main idea:

Whereas θQB is distributed like a sandwich normal (S−1n ), the

quasi-posterior looks like a single slice of bread normal (B−1n ).

We want to complete the sandwich by joining the slice of bread B−1n

to the open-faced sandwich BnP−1n to get S−1

n .

−→

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 22 / 29

Page 28: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

The OFS adjustment

The trick:

Let Ω = B−1P1/2B1/2, the (OFS) adjustment matrix.

Take samples from πp(θ|y) obtained via MCMC and pre-multiply

them (after centering) by an estimator Ω of Ω

If everything goes according to plan, if you squint a bit, each(centered) sample Z ∼ N(0,B−1), making the transformed sampleZ∗ = ΩZ ∼ N(0,S−1).

So we should end up with a sample that has the right frequentistproperties.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29

Page 29: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Outline

1 Introduction

2 Spatial Extremes

3 The OFS adjustment

4 Simulation

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 24 / 29

Page 30: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Simulated data

I simulated 1000 datasets, eachwith

y ∼ Smith process(Σ)

Unit Frechet margins

Σ =

[0.75 −0.5−0.5 1.25

]100 spatial locations

100 blocks

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 25 / 29

Page 31: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

MCMC with OFS

For each realization of y, we run MCMC using the pairwise likelihood.

The OFS matrix is constructed via the four combinations of:1 P a Monte Carlo estimate of the expected information at θ02 P a moment estimate of the expected information at θ3 B the sample covariance of the MCMC sample4 B the observed information at θ

Intervals are constructed as equi-tailed quantiles of the adjustedMCMC sample, and coverage rates computed.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 26 / 29

Page 32: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Coverage rates

σ11

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Σ11

nominal coverage

cove

rage

σ12

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Σ12

nominal coverage

cove

rage

σ22

0.0 0.2 0.4 0.6 0.8 1.0

0.0

0.2

0.4

0.6

0.8

1.0

Σ22

nominal coverage

cove

rage

Dashed lines are OFS-adjusted samples, solid line is un-adjusted.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 27 / 29

Page 33: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Summary

In summary:

Max stable processes are useful for modeling spatial extremes, buttheir corresponding joint densities are unavailable.

One can construct a quasi-posterior using pairwise likelihoods, but

The quasi posterior does not reflect parameter uncertainty.

Using OFS, we can adjust MCMC samples of the quasi posterior tohave the properties we want.

A few caveats:

The OFS matrix can be difficult to estimate (in particular, the“peanut butter” center).

This approach would really shine in hierarchical models, which I havenot shown you.

It’s not really Bayesian.

Ribatet et al. (2010) have a different approach to the same problem.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 28 / 29

Page 34: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

Summary

In summary:

Max stable processes are useful for modeling spatial extremes, buttheir corresponding joint densities are unavailable.

One can construct a quasi-posterior using pairwise likelihoods, but

The quasi posterior does not reflect parameter uncertainty.

Using OFS, we can adjust MCMC samples of the quasi posterior tohave the properties we want.

A few caveats:

The OFS matrix can be difficult to estimate (in particular, the“peanut butter” center).

This approach would really shine in hierarchical models, which I havenot shown you.

It’s not really Bayesian.

Ribatet et al. (2010) have a different approach to the same problem.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 28 / 29

Page 35: Approximate Bayesian computation for spatial extremes via ...stat.duke.edu/~bs128/papers/shaby_jsm_2010.pdfBen Shaby (SAMSI) OFS adjustment August 3, 2010 23 / 29. Outline 1 Introduction

References

Victor Chernozhukov and Han Hong. An MCMC approach to classicalestimation. J. Econometrics, 115(2):293–346, 2003. ISSN 0304-4076.

Mathieu Ribatet, Daniel Cooley, and Anthony Davison. Bayesian inferencefrom composite likelihoods, with an application to spatial extremes.Extremes, 2010. to appear.

Ben Shaby (SAMSI) OFS adjustment August 3, 2010 29 / 29