Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian...

43
Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics & ISR University of Maryland College Park, MD (Victor De Oliveira, David Bindel, Boris and Sandra Kozintsev) 1

Transcript of Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian...

Page 1: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Bayesian Transformed GaussianRandom Field: A Review

Benjamin Kedem

Department of Mathematics & ISR

University of Maryland

College Park, MD

(Victor De Oliveira, David Bindel, Boris and

Sandra Kozintsev)

1

Page 2: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Berger, De Oliveira, Sanso (2001). Objective

Bayesian Analysis of Spatially Correlated Data.

JASA, 96, 1361-1374.

(•) Bindel, De Oliveira, Kedem (1997). An

Implementation of the Bayesian Transformed

Gaussian Spatial Prediction Model.

http://www.math.umd.edu/ bnk/btg page.html.

Box, Cox (1964). An Analysis of Transforma-

tions (with discussion). JRSS, B 26, 211-252.

De Oliveira (2000). Bayesian Prediction of

Clipped Gaussian Random Fields. Comp. Data

Analysis, 34, 299-314.

(•) De Oliveira, Kedem, Short (1997). Bayesian

Prediction of Transformed Gaussian Random

Fields. JASA, 92, 1422-1433.

2

Page 3: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Handcock, Stein (1993). A Bayesian Analysis

of Kriging. Technometrics, 35, 403-410.

(•) De Oliveira, Ecker (2002). Bayesian Hot

Spot Detection in the Presence of a Spatial

Trend: Application to Total Nitrogen Con-

centration in the Chesapeake Bay. Environ-

metrics, 13, 85-101.

(•) Kedem, Fokianos (2002). Regression Mod-

els for Time Series Analysis, New York: Wiley.

Kozintsev, Kedem (2000). Generation of ”Sim-

ilar” Images From a Given Discrete Image. J.

Comp. Graphical Stat., 2000, Vol. 9, No. 2,

286-302.

Kozintseva (1999). Comparison of Three Meth-

ods of Spatial Prediction. M.A. Thesis, De-

partment of Mathematics, University of Mary-

land, College Park.

3

Page 4: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Spatial/temporal geostatistical data often dis-

play:

• Non-Gaussian skewed sampling distributions.

• Positive continuous data.

• Heavy right tails.

• Bounded support.

• Small data sets observed irregularly (gaps).

4

Page 5: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Possible remedies:

• Bayesian Transformed Gaussian (BTG): A

Bayesian approach combined with paramet-

ric families of nonlinear transformations to

Gaussian data.

• BTG provides a unified framework for in-

ference and prediction/interpolation in a

wide variety of models, Gaussian and non-

Gaussian.

• Will describe BTG and illustrate it using

spatial and temporal data.

5

Page 6: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Stationary isotropic Gaussian random field.

Let Z(s), s ∈ D ⊂ Rd, be a spatial process or

a random field.

A random field Z(s) is Gaussian if for all

s1, ..., sn ∈ D, the vector (Z(s1), ..., Z(sn)) has

a multivariate normal distribution.

Z(s) is (second order) stationary when for

s, s + h ∈ D we have

(•) E(Z(s)) = µ,

(•) Cov(Z(s + h), Z(s)) ≡ C(h).

The function C(·) is called the covariogram or

covariance function.

6

Page 7: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

We shall assume that C(h) depends only on

the distance ‖h‖ between the locations s + h

and s but not on the direction of h.

In this case the covariance function as well as

the process are called isotropic.

The corresponding isotropic correlation func-

tion is given by K(l) = C(l)/C(0), where l is

the distance between points.

Useful special case: Matern correlation

Kθ(l) =

12θ2−1Γ(θ2)

(

lθ1

)θ2κθ2

(

lθ1

)

if l 6= 0

1 if l = 0

where θ1 > 0, θ2 > 0, and κθ2 is a modified

Bessel function of the third kind of order θ2.

7

Page 8: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Kθ(l) =

12θ2−1Γ(θ2)

(

lθ1

)θ2κθ2

(

lθ1

)

if l 6= 0

1 if l = 0

Matern (θ1 = 8, θ2 = 3).

2040

6080

100120

X20

40

60

80

100

120

Y

-4-3

-2-1

01

23

Z

8

Page 9: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Spherical correlation:

Kθ(l) =

1 − 32

(

)

+ 12

(

)3if l ≤ θ

0 if l > θ

where θ > 0 controls the correlation range.

Spherical (θ = 120).

2040

6080

100120

X20

40

60

80

100

120

Y

-2-1

01

2Z

9

Page 10: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Exponential correlation:

Kθ(l) = exp(lθ2 log θ1)

θ1 ∈ (0,1), θ2 ∈ (0,2]

Exponential (θ1 = 0.5, θ2 = 1).

2040

6080

100120

X20

40

60

80

100

120

Y

-4-2

02

46

Z

10

Page 11: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Rational quadratic:

Kθ(l) =

(

1 + l2

θ21

)−θ2

θ1 > 0, θ2 > 0.

Rational quadratic (θ1 = 12, θ2 = 8).

2040

6080

100120

X20

40

60

80

100

120

Y

-4-2

02

4Z

11

Page 12: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Clipped, at 3 levels, realizations from Gaussian

random fields. Top left: Matern (8,3). Top

right: spherical (120). Bottom left: exponen-

tial (0.5,1). Bottom right: rational quadratic

(12,8).

www.math.umd.edu/~bnk/bak/ generate.cgi?4

Kozintsev(1999), Kozintsev and Kedem(2000).

12

Page 13: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Ordinary Kriging.

Given the data

Z ≡ (Z(s1), . . . , Z(sn))′

observed at locations s1, . . . , sn in D, the prob-

lem is to predict (or estimate) Z(s0) at loca-

tion s0 using the best linear unbiased predictor

(BLUP) obtained by minimizing

E(Z(s0) −n∑

i=1

λiZ(si))2 subject to

n∑

i=1

λi = 1

13

Page 14: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Define

1 = (1,1, . . . ,1)′, 1 × n vector

c = (C(s0 − s1), . . . , C(s0 − sn))′

C = (C(si − sj)), i, j = 1, ..., n

λ = (λ1, λ2, . . . , λn)′

Then

λ = C−1

(

c +1 − 1′C−1c

1′C−111

)

.

The ordinary kriging predictor is then

Z(s0) = λ′Z.

14

Page 15: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Define,

m =1 − 1′C−1c

1′C−11

and the kriging variance

σ2k(s0) = E(Z(s0) − Z(s0))

2 = C(0) − λ′c + m.

Under the Gaussian assumption,

Z(s0) ± 1.96σk(s0)

is a 95% prediction interval for Z(s0). For non-

Gaussian fields this may not hold.

15

Page 16: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Bayesian Spatial Prediction: The BTG Model.

RF Z(s), s ∈ D observed at

s1, . . . , sn ∈ D

Parametric family of monotone transformations

G = gλ(.) : λ ∈ Λ.

⋆ Assumption: Z(.) can be transformed into

a Gaussian random field by a member of G.

A useful parametric family of transformations

often used in applications to ‘normalize’ posi-

tive data is the Box-Cox (1964) family of power

transformations,

gλ(x) =

xλ−1λ if λ 6= 0

log(x) if λ = 0.

16

Page 17: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

For some unknown ‘transformation parameter’

λ ∈ Λ, gλ(Z(s)), s ∈ D is a Gaussian random

field with

Egλ(Z(s)) =p∑

j=1

βjfj(s),

covgλ(Z(s)), gλ(Z(u)) = τ−1Kθ(s,u),

Regression parameters: β = (β1, . . . , βp)′

Covariates: f(s) = (f1(s), . . . , fp(s))

Variance: τ−1 = vargλ(Z(s))

Simplifying assumption: Isotropy,

Kθ(s,u) = Kθ(||s− u||)

θ = (θ1, . . . , θq) ∈ Θ ⊂ Rq.

17

Page 18: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Data: Zobs = (Z1,obs, . . . , Zn,obs)

gλ(Zi,obs) = gλ(Z(si)) + ǫi ; i = 1, . . . , n,

ǫ1, . . . , ǫn are i.i.d. N(0, ξτ ).

Parameters: η = (β, τ, ξ, θ, λ).

Prediction problem:

Predict Z0 = (Z(s01), . . . , Z(s0k)) from the pre-

dictive density function, defined by

p(zo|zobs) =

Ωp(zo, η|zobs)dη

=

Ωp(zo|η, zobs)p(η|zobs)dη,

where Ω = Rp × (0,∞)2 × Θ × Λ.

18

Page 19: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Notation: For a = (a1, . . . , an), we write

gλ(a) ≡ (gλ(a1), . . . , gλ(an)).

The Likelihood:

L(η; zobs) =

(

τ

)n2∣

∣Ψξ,θ

−12 exp

−τ

2Q

Jλ,

Q =(

gλ(zobs) − Xβ)′

Ψ−1ξ,θ

(

gλ(zobs) − Xβ)

.

X n × p design matrix, Xij = fj(si).

Ψξ,θ = Σθ + ξI, n × n matrix.

Σθ;ij = Kθ(si, sj).

I identity matrix.

Jλ =∏n

i=1 |g′(zi,obs)|, the Jacobian.

19

Page 20: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

The Prior.

Insightful arguments in Box and Cox(1964),

De Oliveira, Kedem, Short (1997), as well as

practical experience lead us to the prior

p(η) ∝p(ξ)p(θ)p(λ)

τJpnλ

,

where p(ξ), p(θ) and p(λ) are the prior marginals

of ξ, θ and λ, respectively, which are assumed

to be proper.

Unusual prior: it depends on the data through

the Jacobian.

For more on prior selection see Berger, De

Oliveira, Sanso (2001).

20

Page 21: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Simplifying assumption: No measurement noise

(ξ = 0).

gλ(Zi,obs) = gλ(Z(si)), i = 1, . . . , n,

p(β, τ, θ, λ) ∝p(θ)p(λ)

τJp/nλ

(1)

η = (β, τ, θ, λ)′.

Also, write

z = zobs

21

Page 22: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

The Posterior.

p(η|z) = p(β, τ, θ, λ|z) = p(β, τ |θ, λ, z)p(θ, λ|z).

To get the first factor:

(β|τ, θ, λ, z) ∼ Np(βθ,λ,1

τ(X′

Σ−1θ

X)−1)

(τ |θ, λ, z) ∼ Ga(n − p

2,

2

qθ,λ

)

where

βθ,λ = (X′Σ

−1θ

X)−1X

′Σ

−1θ

gλ(z)

qθ,λ = (gλ(z) − Xβθ,λ)′Σ

−1θ

(gλ(z) − Xβθ,λ).

and

p(β, τ |θ, λ, z) = p(β|τ, θ, λ, z)p(τ |θ, λ, z)

is Normal-Gamma.

22

Page 23: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

To get the second factor:

p(θ, λ|z) ∝

|Σθ|−1/2|X′

Σ−1θ

X|−1/2q−n−p

2

θ,λJ1−p

nλ p(θ)p(λ)

In addition to the joint posterior distribution

p(η|z) derived above, the predictive density p(zo|z)

also requires p(zo|η, z). We have

p(zo|η, z) = (τ

2π)k/2|Dθ|

−1/2k∏

j=1

|g′λ(zoj)|

× exp

−τ

2(gλ(zo) − Mβ,θ,λ)

′D

−1θ

(gλ(z) − Mβ,θ,λ)

where Mβ,θ,λ,Dθ are known.

23

Page 24: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

We now have the integrand p(zo|η, z)p(η|z) needed

for p(zo|z). By integrating out β and τ we ob-

tain the simplified form of the predictive den-

sity:

p(zo|z) =

Λ

Θp(zo|θ, λ, z)p(θ, λ|z)dθdλ

=

Λ

Θ p(zo|θ, λ, z)p(z|θ, λ)p(θ)p(λ)dθdλ∫

Λ

Θ p(z|θ, λ)p(θ)p(λ)dθdλ

where

p(zo|θ, λ, z) =Γ(n−p+k

2 )∏k

j=1 |g′λ(zoj)|

Γ(n−p2 )πk/2|qθ,λCθ|

1/2

×[1 + (gλ(zo) − mθ,λ)′(qθ,λCθ)−1

×(gλ(zo) − mθ,λ)]−n−p+k

2

and from Bayes theorem,

p(z|θ, λ) ∝ |Σθ|−1/2|X′

Σ−1θ

X|−1/2q−n−p

2

θ,λJ1−p

where mθ,λ,Cθ are known.

24

Page 25: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

BTG Algorithm:

Predictive Density Approximation

1. Let S = z(j)o : j = 1, . . . , r be the set

of values obtained by discretizing the effective

range of Z0.

2. Generate independently θ1, . . . , θm i.i.d. ∼

p(θ) and λ1, . . . , λm i.i.d. ∼ p(λ).

3. For zo ∈ S, the approximation to p(zo|z) is

given by

pm(zo|z) =

∑mi=1 p(zo|θi, λi, z)p(z|θi, λi)

∑mi=1 p(z|θi, λi)

p(zo|θ, λ, z) and p(z|θ, λ) given above.

(⋆) Z0 = Median of (Z0|Z)

25

Page 26: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Software: tkbtg application. Hybrid of C++,

Tcl/Tk, and FORTRAN 77 (Bindel et al (1997)).

http://www.math.umd.edu/~bnk/btg_page.html

The tkbtg Interface Layout

26

Page 27: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Example: Spatial Rainfall Prediction

Rain gauge positions and weekly rainfall to-

tals in mm, Darwin, Australia, 1991.

x

y

0 2 4 6 8 10 12

02

46

810

12

29.55 41.85

26.7633.98

31.27

54.35 30.57

22.8266.76 33.68 19.31

23.6963.1 46.06

24.47

28.9868.2526.83 18.24

31.14

21.717.72

29.92 23.6

27

Page 28: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

1. Use the Box-Cox transformation family.

2. λ ∼ Unif(−3,3).

3. m = 500.

4. Correlation: Matern and exponential.

5. No covariate information. Assume constant

regression: Egλ(Z(s)) = β1.

6. Data apparently not normal.

10 20 30 40 50 60 70

0.0

0.01

0.02

0.03

0.04

Total Rainfall in mm

Mean=33.94

Median=29.73

SD=15.06

28

Page 29: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

No. z z 95% PI

1 29.55 33.74 (10.70, 56.78)2 41.85 34.32 (15.29, 53.35)3 26.76 36.71 (20.93, 52.49)4 33.98 33.39 (18.49, 48.29)5 31.27 32.99 (9.53, 56.45)6 54.35 39.13 (16.46, 61.79)7 30.57 24.45 (15.40, 33.59)8 22.82 23.09 (11.52, 34.66)9 66.76 64.12 (28.25, 100)

10 33.68 35.16 (18.01, 52.30)11 19.31 24.51 (15.62, 33.40)12 23.69 26.45 (15.35, 37.54)13 63.10 72.07 (44.14, 100)14 46.06 40.60 (18.58, 62.63)15 24.47 22.32 (14.00, 30.63)16 28.98 21.62 (13.43, 29.81)17 68.25 46.54 (19.25, 73.84)18 26.83 29.52 (16.57, 42.46)19 18.24 19.00 (10.79, 27.21)20 31.14 37.36 (20.33, 54.39)21 21.70 22.97 (14.71, 31.22)22 17.72 22.69 (11.64, 33.74)23 29.92 26.56 (12.21, 40.91)24 23.60 21.83 (10.85, 32.81)

29

Page 30: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Spatial prediction and contour maps from theDarwin data using Matern correlation.

0

2

4

6

8

10

X

0

2

4

6

8

10

Y

1020

3040

5060

7080

Z

x

y

0 2 4 6 8 10 12

02

46

810

12

20

25

30

30

30

30

30

35 35

40

40

45

50

55

60

65

70

30

Page 31: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Spatial prediction and contour maps from theDarwin data using exponential correlation.

0

2

4

6

8

10

X

0

2

4

6

8

10

Y

1020

3040

5060

7080

Z

x

y

0 2 4 6 8 10 12

02

46

810

12

20

25

25

25

30

30

35 3540

40

40

45

45

50

50

55

55

6060

65

70

31

Page 32: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Predictive densities, 95% prediction intervals,

and cross-validation: Predicting a true value

from the remaining 23 observations using Matern

correlation. The vertical line marks the loca-

tion of the true value.

x

p(x)

0 20 40 60 80 100

0.0

0.04

0.08

0.12

True: 41.85Median: 34.31(15.60, 53.02)

x

p(x)

0 20 40 60 80 100

0.0

0.04

0.08

0.12

True: 63.10Median: 70.76(41.52, 100)

x

p(x)

0 20 40 60 80 100

0.0

0.04

0.08

0.12

True: 28.98Median: 22.1(13.26, 30.94)

x

p(x)

0 20 40 60 80 100

0.0

0.04

0.08

0.12

True: 21.70Median: 23.05(14.84, 31.26)

32

Page 33: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Comparison of BTG With Kriging and Trans-

Gaussian kriging (Kozintseva (1999)).

Cross validation results using artificial data on

50 × 50 grid.

Data obtained by transforming a Gaussian (0,1)

RF using inverse Box-Cox transformation.

In Kriging and TG kriging λ, θ, were known.

Not in BTG (!)

λ = 0: Log-Normal.

λ = 1: Normal.

λ = 0.5: Between Normal and Log-Normal.

In most cases BTG has more reliable but larger

prediction intervals.

BTG predicts at the original scale. TG kriging

does not.

33

Page 34: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Matern(1,10)

λ 0 0.5 1

KRG MSE 68397.48 7.15 0.58TGK MSE 55260.90 7.08 0.58BTG MSE 64134.30 7.31 0.56

KRG AvePI 2.42 2.51 2.42TGK AvePI 291.80 8.21 2.42BTG AvePI 330.68 10.23 2.87

KRG % out 100% 48% 6%TGK % out 18% 8% 6%BTG % out 12% 6% 6%

Exponential(e−0.03,1)λ 0 0.5 1

KRG MSE 12212.32 1.83 0.13TGK MSE 11974.73 1.84 0.13BTG MSE 12520.70 1.89 0.14

KRG AvePI 1.45 1.43 1.45TGK AvePI 267.92 5.24 1.45BTG AvePI 466.69 6.10 1.63

KRG % out 98% 64% 2%TGK % out 20% 4% 2%BTG % out 6% 2% 2%

34

Page 35: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Application of BTG to Time Series Prediction.

Short time series observed irregularly.

Set: s = (x, y) = (t,0).

Can predict/interpolate as in state space pre-

diction: k–step prediction forward, backward,

and “in the middle”.

Example 1: Monthly data of unemployed women

20 years of age and older, 1997–2000. Data

source: Bureau of Labor Statistics. N = 48.

Example 2: Monthly airline passenger data,

1949–1960. Data source: Box-Jenkins (1976).

Use only N = 36 out of 144 observations,

t = 51, ...,86.

35

Page 36: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Example: Prediction of Monthly number of un-

employed women (Age ≥ 20), 1997–2000. Data

in hundreds of thousands.

Cross validation and 95% prediction intervals.

Observations at times t = 12,13,36 are out-

side their 95% PI’s. N = 48 − 1 = 47.

t

No.

Une

mpl

oyed

Wom

en

0 10 20 30 40

1015

2025

3035

TruePredictedLowerUpper

Unemployed Women 1997--2000

36

Page 37: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Forward and backward one and two step pre-

diction in the unemployed women series.

In 2–step higher dispersion.

x

p(x)

10 20 30 40 50

0.0

0.05

0.10

0.15

0.20

True: 18.34Median: 21.78(17.24, 26.33)

One step forward

x

p(x)

10 20 30 40 50

0.0

0.05

0.10

0.15

0.20

True: 18.34Median: 22.20(16.36, 28.05)

Two step forward

x

p(x)

10 20 30 40 50

0.0

0.05

0.10

0.15

0.20

True: 28.98Median: 26.17(21.94, 30.40)

One step backward

x

p(x)

10 20 30 40 50

0.0

0.05

0.10

0.15

0.20

True: 28.98Median: 24.69(19.77, 29.61)

Two step backward

37

Page 38: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Example: Time series of monthly international

airline passengers in thousands, January 1949-

December 1960. N=144. Seasonal time se-

ries.

t

No.

Pas

seng

ers

0 20 40 60 80 100 120 140

100

200

300

400

500

600

38

Page 39: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

BTG cross validation and prediction intervals

for the monthly airline passengers series, t =

51, ...,86, using Matern correlation. Observa-

tions at t = 62,63 are outside the PI. N =

36 − 1 = 35.

t

No.

Pas

seng

ers

50 60 70 80

150

200

250

300

350

400

TruePredictedLowerUpper

39

Page 40: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Application to Rainfall: Heuristic Argument.

Let Xn represent the area average rain rate

over a region such that

Xn = mXn−1 + λ + ǫn, n = 1,2,3, · · · ,

where the noise ǫn is a martingale difference.

It can be argued that necessary conditions for

Xn → LogNormal, n → ∞,

are that m → 1− and λ → 0+.

40

Page 41: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

The monotone increase in m (Curve a) and the

monotone decrease in λ (Curve b) as a func-

tion of the square root of the area. Source:

Kedem and Chiu(1987).

This suggests that the lognormal distribution

as a model for averages or rainfall amounts

over large areas or long periods.

41

Page 42: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

It is interesting to obtain the posterior p(λ | z)

of λ, the transformation parameter in the Box-

Cox family, given the data, where the data are

weekly rainfall totals from Darwin, Australia.

With a uniform prior for λ, the medians of

p(λ | z) in 5 different weeks are:

Week 1 median = − 0.45 (to the left of 0).

Week 2 median = 0.45 (to the right of 0).

Week 3 median = 0.15 (Not far from 0).

Week 4 median = 0.20 (Not far from 0).

Week 5 median = 0.95 (Close to 1).

42

Page 43: Bayesian Transformed Gaussian Random Field: A Reviebnk/Ch_Pred1sld.pdf · 2010-11-02 · Bayesian Transformed Gaussian Random Field: A Review Benjamin Kedem Department of Mathematics

Weekly posterior p(λ | z) of λ given rainfall

totals from Darwin, Australia, for 5 different

weeks.

(a)

lambda

dens

ity

-3 -2 -1 0 1 2 3

0.0

0.5

1.0

1.5

2.0

MAE = -0.45

(b)

lambda

dens

ity

-3 -2 -1 0 1 2 3

0.0

0.5

1.0

1.5

2.0

MAE = 0.45

(c)

lambda

dens

ity

-3 -2 -1 0 1 2 3

0.0

0.5

1.0

1.5

2.0

MAE = 0.15

(d)

lambda

dens

ity

-3 -2 -1 0 1 2 3

0.0

0.5

1.0

1.5

2.0

MAE = 0.2

(e)

lambda

dens

ity

-3 -2 -1 0 1 2 3

0.0

0.5

1.0

1.5

2.0

MAE = 0.95

43