prior sensitivity and model choice · dic1 = 271 dic2 = 98 dic1 - dic2 = 172 conclude: overwhelming...

33
Sensitivity DIC Prior sensitivity and model choice Alex Cook Week 9 Alex Cook, ST5219, Bayesian Hierarchical Modelling 1/30

Upload: others

Post on 23-Aug-2020

0 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Prior sensitivity and model choice

Alex Cook

Week 9

Alex Cook, ST5219, Bayesian Hierarchical Modelling 1/30

Page 2: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(1, 1)

E(p) = 0.5

V(p) ≈ 0.292

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 2/30

Page 3: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(1, 1)

E(p) = 0.5

V(p) ≈ 0.292

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 3/30

Page 4: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(1, 97)

E(p) = 0.01

V(p) = 0.012

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 4/30

Page 5: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(1, 97)

E(p) = 0.01

V(p) = 0.012

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 5/30

Page 6: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(90, 809)

E(p) = 0.1

V(p) = 0.012

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 6/30

Page 7: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(90, 809)

E(p) = 0.1

V(p) = 0.012

p

dens

ity

0% 1% 2%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 7/30

Page 8: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

HIV trial

x = 51

n = 8197

x ∼ Bin(n, p)

p ∼ Be(90, 809)

E(p) = 0.1

V(p) = 0.012

p

dens

ity

0% 10% 20%

Alex Cook, ST5219, Bayesian Hierarchical Modelling 8/30

Page 9: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Take home

Lesson learnt

Informative priors can have big effects on posteriors

Alex Cook, ST5219, Bayesian Hierarchical Modelling 9/30

Page 10: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Simulated Poisson data

set.seed(54321)

n=100

lambda=10

x=rpois(n,lambda)

dlambda=0.01

lambda vector

=seq(1,20,dlambda)

x

Fre

quen

cy0 10 20

010

20

Alex Cook, ST5219, Bayesian Hierarchical Modelling 10/30

Page 11: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Simulated Poisson data

> prior1=dunif(lambda vector,0,100)

> logprior1=log(prior1)

> a=logprior1

> for(i in 1:n)a=a+

dpois(x[i],lambda vector,log=TRUE)

> a=exp(a-max(a));posterior1=a(sum(a)*dlambda)

> plot(lambda vector,posterior1,type=’l’,

xlab=expression(lambda),ylab=’density’)

> lines(lambda vector,prior1,lty=2)

Alex Cook, ST5219, Bayesian Hierarchical Modelling 11/30

Page 12: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Simulated Poisson data

prior1=dunif(lambda vector,0,100)

0 5 10 15 20

λ

dens

ity

0.0

0.5

1.0

Alex Cook, ST5219, Bayesian Hierarchical Modelling 12/30

Page 13: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Simulated Poisson data

prior2=dexp(lambda vector,0.01)

0 5 10 15 20

λ

dens

ity

0.0

0.5

1.0

Alex Cook, ST5219, Bayesian Hierarchical Modelling 13/30

Page 14: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Simulated Poisson data

prior3=dnorm(lambda vector,0,100)

0 5 10 15 20

λ

dens

ity

0.0

0.5

1.0

Alex Cook, ST5219, Bayesian Hierarchical Modelling 14/30

Page 15: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Take home

Lesson learnt

Informative priors can have big effects on posteriors

Non-informative priors: form doesn’t matter much

Alex Cook, ST5219, Bayesian Hierarchical Modelling 15/30

Page 16: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Classical model selection

yi ∼ N(µi , σ2)

M1 : µi = a + bxi

M2 : µi = a + bxi + cx2i

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

2 3 4 5 6

100

110

120

130

140

x

y

You were probably taught. . .

NHST of H0 : c = 0, if p < α choose M2

Alex Cook, ST5219, Bayesian Hierarchical Modelling 16/30

Page 17: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Classical model selection

yi ∼ N(µi , σ2)

M1 : µi = a + bxi

M2 : µi = a + bxi + cx2i

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

2 3 4 5 6

100

110

120

130

140

x

y

You were probably taught. . .

NHST of H0 : c = 0, if p < α choose M2

Alex Cook, ST5219, Bayesian Hierarchical Modelling 16/30

Page 18: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Classical model selection

yi ∼ N(µi , σ2)

M1 : µi = a + bxi

M2 : µi = a + c exp(xi)

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

2 3 4 5 6

100

110

120

130

140

x

y

What to do now?

NHST of H0 :?

Alex Cook, ST5219, Bayesian Hierarchical Modelling 17/30

Page 19: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Classical model selection

yi ∼ N(µi , σ2)

M1 : µi = a + bxi

M2 : µi = a + c exp(xi)

●●

●●

●● ●

●●

● ●

●●

●●

●●

●●

●●

●●

●●

● ●●

●●

●●

● ●

2 3 4 5 6

100

110

120

130

140

x

y

What to do now?

NHST of H0 :?

Alex Cook, ST5219, Bayesian Hierarchical Modelling 17/30

Page 20: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Classical model selection

Alex Cook, ST5219, Bayesian Hierarchical Modelling 18/30

Page 21: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Bayesian model selection

DIC:

Deviance: D(θ) = −2 log f (data|θ)

Posterior mean deviance: D̄ =∫

D(θ)f (θ|data) dθ

Effective # parameters: pD = D̄ − D(θ̄)

DIC: DIC = D̄ + pD

Spiegelhalter et al (2002, J Roy Stat Soc ser B64:583–639) recommend same scale as AIC

Alex Cook, ST5219, Bayesian Hierarchical Modelling 19/30

Page 22: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Bayesian model selection

DIC:

Deviance: D(θ) = −2 log f (data|θ)

Posterior mean deviance: D̄ =∫

D(θ)f (θ|data) dθ

Effective # parameters: pD = D̄ − D(θ̄)

DIC: DIC = D̄ + pD

Spiegelhalter et al (2002, J Roy Stat Soc ser B64:583–639) recommend same scale as AIC

Alex Cook, ST5219, Bayesian Hierarchical Modelling 19/30

Page 23: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Radish data

days post sowing

% in

fect

ed

0 7 14 21

0

20

40

60

80

100

Alex Cook, ST5219, Bayesian Hierarchical Modelling 20/30

Page 24: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Radish data

Model 1: no difference model

Xi ∼ Bin(ni , p)

p ∼ Be(1, 1)

Alex Cook, ST5219, Bayesian Hierarchical Modelling 21/30

Page 25: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Radish data

Model 2: hierarchical model

Xi ∼ Bin(ni , pi)

pi ∼ Be(a, b)

a ∼ exp(0.0001)

b ∼ exp(0.0001)

Alex Cook, ST5219, Bayesian Hierarchical Modelling 22/30

Page 26: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Calculating the DIC

logposterior1=function(current,data){

current$loglikelihood =sum(dbinom(data$I,data$N,current$p,log=TRUE))

current$D = -2*current$loglikelihoodcurrent$logprior = dbeta(current$p,1,1,log=TRUE)current$logposterior = current$logprior +

current$loglikelihoodcurrent

}

Alex Cook, ST5219, Bayesian Hierarchical Modelling 23/30

Page 27: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Calculating the DIC

phat=mean(dump$p)current$p=phatcurrent=logposterior1(current,data)Dbar1=mean(dump$D)pD1=Dbar1-current$D

DIC1=Dbar1+pD1

Alex Cook, ST5219, Bayesian Hierarchical Modelling 24/30

Page 28: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Calculating the DIC

logposterior2=function(current,data){

current$loglikelihood =sum(dbinom(data$I,data$N,current$p,log=TRUE))

current$D = -2*current$loglikelihoodcurrent$logprior =

sum(dbeta(current$p,current$a,current$b,log=TRUE))+dexp(current$a,0.001,log=TRUE)dexp(current$b,0.001,log=TRUE)

current$logposterior = current$logprior +current$loglikelihood

current

}

Alex Cook, ST5219, Bayesian Hierarchical Modelling 25/30

Page 29: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Calculating the DIC

phat=c();for(i in 1:13)phat[i]=mean(dump$p[,i])current$p=phatcurrent=logposterior2(current,data)Dbar2=mean(dump$D)pD2=Dbar2-current$D

DIC2=Dbar2+pD2

Alex Cook, ST5219, Bayesian Hierarchical Modelling 26/30

Page 30: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Fitted models

0.70 0.75 0.80 0.85 0.90 0.95 1.00

020

4060

8010

0

probability of disease

dens

ity

Alex Cook, ST5219, Bayesian Hierarchical Modelling 27/30

Page 31: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Fitted models

100 150 200 250

0.0

0.2

0.4

0.6

0.8

deviance

dens

ity

Alex Cook, ST5219, Bayesian Hierarchical Modelling 28/30

Page 32: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Comparing the DIC

DIC1 = 271

DIC2 = 98

DIC1 - DIC2 = 172

Conclude:

Overwhelming evidence against homogeneous diseaserates

Alex Cook, ST5219, Bayesian Hierarchical Modelling 29/30

Page 33: Prior sensitivity and model choice · DIC1 = 271 DIC2 = 98 DIC1 - DIC2 = 172 Conclude: Overwhelming evidence against homogeneous disease rates Alex Cook, ST5219, Bayesian Hierarchical

Sensitivity DIC

Better Bayesian model selection

Alex Cook, ST5219, Bayesian Hierarchical Modelling 30/30