Notes BMDA PDF

.
Bayesian Methods for

Data Analysis
ENAR Annual Meeting
Tampa, Florida March 26, 2006
ENAR - March 2006
Course contents
Introduction of Bayesian concepts using single-parameter models.
Multiple-parameter models and hyerarchical models.
Computation:
approximations to the posterior,
importance sampling and MCMC.
rejection and
Model checking, diagnostics, model fit.

Linear hierarchical models: random effects and mixed models.
Generalized linear models: logistic, multinomial, Poisson regression.
Hierarchical models for spatially correlated data.
ENAR - March 2006
Course emphasis
Notes draw heavily on the book by Gelman et al., Bayesian Data
Analysis 2nd. ed., and many of the figures are borrowed directly from
that book.
We focus on the implementation of Bayesian methods and interpretation
of results.
Little theory, but some is needed to understand methods.
Lots of examples. Some are not directly drawn from biological problems,
but still serve to illustrate methodology.
Biggest idea to get across: inference by simulation.
ENAR - March 2006
Software: R (or SPlus) early on, and WinBUGS for most of the
examples after we discuss computational methods.
All of the programs used to construct examples can be downloaded
from www.public.iastate.edu/ alicia.
On my website, go to Teaching and then to ENAR Short Course.
ENAR - March 2006
Single parameter models

There is no real advantage to being a Bayesian in these simple models.
We discuss them to introduce important concepts:
Priors: non-informative, conjugate, other informative
Computations
Summary of posterior, presentation of results
Binomial, Poisson, Normal models as examples
Some real examples
ENAR - March 2006
Binomial model
Of historical importance: Bayes derived his theorem and Laplace
provided computations for the Binomial model.
Before Bayes, question was: given , what are the probabilities of the
possible outcomes of y?
Bayes asked: what is P r(1 < < 2|y)?
Using a uniform prior for , Bayes showed that
R 2
P r(1 < < 2|y)
y (1 )ny d
p(y)
with p(y) = 1/(n + 1) uniform a priori.

ENAR - March 2006
Laplace devised an approximation to the integral expression in

numerator.
First application by Laplace: estimate the probability that there were
more female births in Paris from 1745 to 1770.
Consider n exchangeable Bernoulli trials y1, ..., yn.
yi = 1 a success, yi = 0 a failure.
Exchangeability:summary is # of successes in n trials y.
For the probability of success, Y B(n, ):

n y
p(y|) =
(1 )ny .
y
ENAR - March 2006
MLE is sample proportion: = y/n.

Prior on : uniform on [0,1] (for now)
p() 1.
Posterior:
p(|y) y (1 )ny .
As a function of , posterior is proportional to a Beta(y + 1, n y + 1)
density.
We could also do the calculations to derive the posterior:

n y
ny
(1
)
y
p(|y) = R n y
ny d
y (1 )
ENAR - March 2006
n!
= (n + 1)
y (1 )ny
y!(n y)!
=
(n + 1)! y
(1 )ny
y!(n y)!
(n + 2)
y+11(1 )ny+11
(y + 1)(n y + 1)
= Beta(y + 1, n y + 1)
Point estimation
y+1
Posterior mean E(|y) = n+2
Note posterior mean is compromise between prior mean (1/2) and
sample proportion y/n.
Posterior mode y/n
Posterior median such that P r( |y) = 0.5.
Best point estimator minimizes the expected loss (more later).
ENAR - March 2006
Posterior variance
(y + 1)(n y + 1)
V ar(|y) =
.
(n + 2)2(n + 3)
Interval estimation
95% credibe set (or central posterior interval) is (a, b) if:
Z
p(|y)d = 0.975
p(|y)d = 0.025 and

0
A 100(1 )% highest posterior density credible set is subset C of

such that
C = { : p(|y) k()}
where k() is largest constant such that P r(C|y) 1 .
For symmetric unimodal posteriors, credible sets and highest
posterior density credible sets coincide.
ENAR - March 2006
10
In other cases, HPD sets have smallest size.

Interpretation (for either): probability that is in set is equal to
1 . Which is what we want to say!.
Credible set and HPD set
ENAR - March 2006
11
Inference by simulation:
Draw values from posterior: easy for closed form posteriors, can also
do in other cases (later)
Monte Carlo estimates of point and interval estimates
Added MC error (due to sampling)
Easy to get interval estimates and estimators for functions of
parameters.
Prediction:
Prior predictive distribution
Z
p(y) =
0
1

n y
1
ny
(1 )
d =
.
y
n+1
A priori, all values of y are equally likely

ENAR - March 2006
12
Posterior predictive distribution, to predict outcome of a new trial y

given y successes in previous n trials
Z
P r(
y = 1|y, )p(|y)d
P r(
y = 1|y) =
0
P r(
y = 1|)p(|y)d
=
0
Z
=
0
ENAR - March 2006
y+1
p(|y)d = E(|y) =
n+2
13
Binomial model: different priors

How do we choose priors?
In a purely subjective manner (orthodox)
Using actual information (e.g., from literature or scientific
knowledge)
Eliciting from experts
For mathematical convenience
To express ignorance
Unless we have reliable prior information about , we prefer to let the
data speak for themselves.
Asymptotic argument: as sample size increases, likelihood should
dominate posterior
ENAR - March 2006
14
Conjugate prior
Suppose that we choose a Beta prior for :
p(|, ) 1(1 )1
Posterior is now
p(|y) y+1(1 )ny+1
Posterior is again proportional to a Beta:
p(|y) = Beta(y + , n y + ).
For now, , considered fixed and known, but they can also get their
own prior distribution (hierarchical model).
ENAR - March 2006
15
Beta is the conjugate prior for binomial model: posterior is in the

same form as the prior.
To choose prior parameters, think as follows: observe successes in
+ prior trials. Prior guess for is /( + ).
ENAR - March 2006
16
Conjugate priors
Formal definition: F a class of sampling distributions and P a class
of prior distributions. Then P is conjugate for F if p() P and
p(y|) F implies p(|y) P .
If F is exponential family then distributions in F have natural conjugate
priors.
A distribution in F has form
p(yi|) = f (yi)g() exp(()T u(yi)).
For an iid sequence, likelihood function is
p(y|) g()n exp(()T t(y)),
ENAR - March 2006
17
for () the natural parameter and t(y) a sufficient statistic.

Consider prior density
p() g() exp(()T )
Posterior also in exponential form
p(|y) g()n+ exp(()T (t(y) + ))
ENAR - March 2006
18
Conjugate prior for binomial proportion

y is # of successes in n exchangeable trials, so sampling distribution is
binomial with prob of success :
p(y|) y (1 )ny
Note that
p(y|) y (1 )n(1 )y
(1 )n exp{y[log log(1 )]}
Written in exponential family form
}
p(y|) (1 ) exp{y log
1
n
ENAR - March 2006
19
where g()n = (1 )n, y is the sufficient statistic, and the logit

log /(1 ) is the natural parameter.
Consider prior for : Beta(, ):

p() 1(1 )1
If we let = 1 and = + 2, can write

p() (1 )
or
p() (1 ) exp{ log

}
1
ENAR - March 2006
20
Then posterior is in same form as prior:

p(|y) (1 )n+ exp{(y + ) log
}
1
Since p(y|) u(1 )ny then prior Beta(, ) suggests that a

priori we believe in approximately successes in + trials.
Our prior guess for the probability of success is /( + )
By varying ( + ) (with /( + ) fixed), we can incorporate more
or less information into the prior: prior sample size
Also note:
+y
++n
is always between the prior mean /( + ) and the MLE y/n.
E(|y) =
ENAR - March 2006
21
Posterior variance
V ar(|y) =
=
( + y)( + n y)
( + + n)2( + + n + 1)
E(|y)[1 E(|y)]
++n+1
As y and n y get large:

E(|y) y/n
var(|y) (y/n)[1 (y/n)]/n which approaches zero at rate 1/n.
Prior parameters have diminishing influence in the posterior as n
.
ENAR - March 2006
22
Placenta previa example

Condition in which placenta is implanted low in uterus, preventing
normal delivery.
In study in Germany, found that 437 out of 980 births with placenta
previa were females so observed ratio is 0.446
Suppose that in general population, proportion of female births is 0.485.
Can we say that the proportion of female babies among placenta previa
births is lower than in the general population?
If y = number of female births, is probability of a female birth, and
p() is uniform, then
p(|y) y (1 )n1
ENAR - March 2006
23
and therefore
|y Beta(y + 1, n y + 1).
Thus a posteriori, is Beta(438, 544) and

438
E(|y) =
438 + 544
V ar(|y) =
438 544
(438 + 544)2(438 + 544 + 1)
Check sensitivity to choice of priors: try several Beta priors with

increasingly more information about .
Fix prior mean at 0.485 and increase prior sample size + .

ENAR - March 2006
24
Prior
mean
0.5
0.485
0.485
0.485
0.485
0.485
0.485
+
2
2
5
10
20
100
200
Post.
median
0.446
0.446
0.446
0.446
0.447
0.450
0.453
Post 2.5th
pctile
0.415
0.415
0.415
0.415
0.416
0.420
0.424
Post 97.5th
pctile
0.477
0.477
0.477
0.477
0.478
0.479
0.481
Results robust to choice of prior, even very informative prior. Prior mean
not in posterior 95% credible set.
ENAR - March 2006
25
Conjugate prior for Poisson rate

y (0, 1, ...) is counts, with rate > 0. Sampling distribution is
Poisson
yi exp()
p(yi|) =
yi !
For exchangeable (y1, ..., yn)
p(y|) =
n
Y
yi exp()
i=1
yi !
ny exp(n)
= Qn
.
y
!
i=1 i
Consider Gamma(, ) as prior for : p() 1 exp()

Then posterior is also Gamma:
p(|y) ny exp(n)1 exp() ny+1 exp(n )
ENAR - March 2006
26
For = n
y + and = n + , posterior is Gamma(, )
Note:
Prior mean of is /
Posterior mean of is
n
y+
E(|y) =
n+
If sample size n then E(|y) approaches MLE of .
If sample size goes to zero, then E(|y) approaches prior mean.
ENAR - March 2006
27
Mean of a normal distribution

y N (, 2) with 2 known
For {y1, ..., yn} an iid sample, the likelihood is:
p(y|) = i
1
1
exp( 2 (yi )2)
2
2 2
Viewed as function of , likelihood is exponential of a quadratic in :

1 X 2
( 2yi + yi2))
p(y|) exp( 2
2 i
Conjugate prior for must belong to family of form
p() = exp(A2 + B + C)
ENAR - March 2006
28
that can be parameterized as

1
p() exp( 2 ( 0)2)
20
and then p() is N (0, 02)
(0, 02) are hyperparameters. For now, consider known.
If p() is conjugate, then posterior p(|y) must also be normal with
parameters (n, n2).
Recall that p(|y) p()p(y|)
Then
ENAR - March 2006
1
p(|y) exp( [
2
i (yi
)2
( 0)2
+
])
2
0
29
Expand squares, collect terms in 2 and in :

1
p(|y) exp( [
2
2
i yi
i yi
2
2 20 + 20
+
])
2
0
1 (02 + 2)n2 2(n

y 02 + 0 2)
exp( [
])
2
202
1 (( 2/n) + 02) 2
y02 + 0( 2/n)
exp(
[ 2(
)])
2
2
2
2
2 ( /n)0
( /n) + 0
Then p(|y) is normal with
Mean: n = (
y 02 + 0( 2/n))/(( 2/n) + 02)
Variance: n2 = (( 2/n)02)/(( 2/n) + 02)
Posterior mean
ENAR - March 2006
30
Note that
n =
y02
+ 0( /n)
=
( 2/n) + 02
n
1
y
02 0
1
n
+
2
02
(by dividing numerator and denominator into ( 2/n)02)

Posterior mean is weighted average of prior mean and sample mean.
Weights are given by precisions n/ 2 and 1/02.
As data precision increases (( 2/n) 0) because 2 0 or because
n , n y.
Also, note that
n =
ENAR - March 2006
y02 + 0( 2/n)
( 2/n) + 02
31
02
2/n
) + y( 2
)
= 0( 2
( /n) + 02
( /n) + 02
Add and subtract 002/(( 2/n) + 02) to see that
02
n = 0 + (
y 0)( 2
)
2
( /n) + 0
Posterior mean is prior mean shrunken towards observed value. Amount
of shrinkage depends on relative size of precisions
Posterior variance:
Recall that
1 (( 2/n) + 02) 2
y02 + 0( 2/n)
p(|y) exp(
[ 2(
)])
2
2
2
2
2 ( /n)0
( /n) + 0
ENAR - March 2006
32
Then
1
n2
=
=
(( 2/n) + 02)
( 2/n)02
n
1
+
2 02
Posterior precision = sum of prior and data precisions.

Posterior predictive distribution
Want to predict next observation y.
Recall p(
y |y)
p(
y |)p(|y)d
We know that
ENAR - March 2006
33
Given , y N (, 2)
|y N (n, n2)
Then
Z
1 (
y )2
1 ( n)2
exp{
} exp{
}d
2
2
2
2
n
1 (
y )2 ( n)2
exp{ [
+
]}d
2
2
2
p(
y |y)
Integrand is kernel of bivariate normal, so (

y , ) have bivariate normal
joint posterior.
Marginal p(
y |y) must be normal.
ENAR - March 2006
34
Need E(
y |y) and var(
y |y):
E(
y |y) = E[E(
y |y, )|y] = E(|y) = n,
because E(
y |y, ) = E(
y |) =
var(
y |y) = E[var(
y |y, )|y] + var[E(
y |y, )|y]
= E( 2|y) + var(|y)
= 2 + n2
because var(
y |y, ) = var(
y |) = 2
Variance includes additional term n as a penalty for us not knowing
the true value of .
Recall n =
ENAR - March 2006
1 + n y
0 2
02
1 + n
02 2
35
In n, prior precision 1/02 and data precision n/ 2 are equivalent.

Then:
For n large, (
y , 2) determine posterior
For 02 = 2/n, prior has the same weight as adding one more
observation with value 0.
When 02 with n fixed, or when n with 02 fixed:
p(|
y ) N (
y , 2/n)
Good approximation in practice when prior beliefs about are vague
or when sample size is large.
ENAR - March 2006
36
Normal variance
Example of a scale model
Assume that y N (, 2), known.
For iid y1, ..., yn:
n
1 X
2
2 n/2
(yi )2)
p(y| ) ( )
exp( 2
2 i=1
2
2 n/2
p(y| ) ( )
for suffient statistic
nv
exp( 2 )
2
1X
(yi )2
v=
n i=1
ENAR - March 2006
37
Likelihood is in exponential family form with natural parameter ( 2) =

2. Then, natural conjugate prior must be of form
p( 2) ( 2) exp(( 2))
Consider an inverse Gamma prior:

p( 2) ( 2)(+1) exp(/ 2)
For ease of interpretation, reparameterize as a scaled-inverted 2

distribution, with 0 d.f. and 02 scale.
Scale 02 corresponds to prior guess for 2.
Large 0: lots of confidence in 02 as a good value for 2.
ENAR - March 2006
38
If 2 inv 2(0, 02) then

2
2
0
0
0
p( 2) ( 2 )( 2 +1) exp( 2 )
0
2
Corresponds to Inv
2
0 0 0
Gamma( 2 , 2 ).
Prior mean is 020/(0 2)

Prior variance behaves like 04/0: large 0, large prior precision.
Posterior
p( 2|y) p( 2)p(y| 2)
2 n/2
( )
ENAR - March 2006
nv 2 ( 0 +1)
002
exp( 2 )( 2 ) 2
exp( 2 )
2 0
2
39
2 ( 21 +1)
( )
exp(
with 1 = 0 + n and
12
1
2
)
1
1
2 2
nv + 002
=
n + 0
p( 2|y) is also a scaled inverted 2

Posterior scale is weighted average of prior guess and data estimate
Weights given by prior and sample degrees of freedom
Prior provides information equivalent to 0 observations with average
squared deviation equal to 02.
As n , 12 v.
ENAR - March 2006
40
Poisson model
Appropriate for count data such as number of cases, number of
accidents, etc. For a vector of n iid observations:
t(y)en
yi e
= p(y|) = Qn
p(y|) = i
,
yi !
y!
i=1
where is the rate, y = 0, 1, ... and t(y) =
statistic for .
Pn
i=1 yi
the sufficient
We can write the model in the exponential family form:

p(y|) en et(y) log
where () = log is natural parameter.
ENAR - March 2006
41
Natural conjugate prior must have form

p() e e log
or p() e 1 that looks like a Gamma(, ).
Posterior is Gamma( + n
y , + n)
ENAR - March 2006
42
Poisson model - Digression

In the conjugate case, can often derive p(y) using
p(y) =
p()p(y|)
p(|y)
In case of Gamma-Poisson model for a single observation:

( + y)
Poisson() Gamma(, )
p(y) =
=
Gamma( + y, + 1)
()y!(1 + )(+y)
1 y
= ( + y 1, y)(
) (
)
+1 +1
the density of a negative binomial with parameters (, ).
ENAR - March 2006
43
Thus we can interpret negative binomial distributions as arising from a

mixture of Poissons with rate , where follows a Gamma distribution:
Z
p(y) =
Poisson (y|) Gamma (|, )d
Negative binomial is robust alternative to Poisson.
ENAR - March 2006
44
Poisson model (contd)
Given rate, observations assumed to be exchangeable.

Add an exposure to model: observations are exchangeable within small
exposure intervals.
Examples:
In a city of size 500,000, define rate of death from cancer per million
people, then exposure is 0.5
Intersection in Ames with traffic of one million vehicles per year,
consider number of traffic accidents per million vehicles per year.
Exposure for intersection is 1.
Exposure is typically known and reflects the fact that in each unit,
the number of persons or cars or plants or animals that are at risk
is different.
ENAR - March 2006
45
Now
p(y|)
Py
exp(
xi)
for xi the exposure of unit i.

With prior p() = Gamma (, ), posterior is
p(|y) = Gamma ( +
X
i
ENAR - March 2006
yi , +
xi)
46
Non-informative prior distns

A non-informative prior distribution
Has minimal impact on posterior
Lets the data speak for themselves
Also called vague, flat, diffuse
So far, we have concentrated on conjugate family.
Conjugate priors can also be almost non-informative.
If y N (, 1), natural conjugate for is N (0, 02), and posterior is
N (1, 12), where
1
0/02 + n
y / 2 2
1 =
,
=
1
1/02 + n/ 2
1/02 + n/ 2
ENAR - March 2006
47
For 0 :
1 y
12 2/n
Same result could have been obtained using p() 1.
The uniform prior is a natural non-informative prior for location
parameters (see later).
It is improper:
Z
Z
p()d =
d =
yet leads to proper posterior for . This is not always the case.
Poisson example: y1, ..., yn iid Poisson(), consider p() 1/2.
R 1/2
0
d = improper
ENAR - March 2006
48
Yet
p(|y)
=
P y n 1/2
P y 1/2 n
e
P y +1/21 n
i i
i i
=
proportional to a Gamma( 21 +
i i
i yi , n),
proper
Uniform priors arise when we assign equal density to each : no

preferences for one value over the other.
But they are not invariant under one-to-one transformations.
Example:
= exp{} so that = log{} is inverse transformation.
ENAR - March 2006
49
d
d
= 1 is Jacobian
Then, if p() is prior for , p() = 1p(log ) is corresponding
prior for transformation.
For p() c, p() 1, informative.
Informative prior is needed to arrive at same answer in both
parameterizations.
ENAR - March 2006
50
Jeffreys prior
In 1961, Jeffreys proposed a method for finding non-informative priors
that are invariant to one-to-one transformations
Jeffreys proposed
p() [I()]1/2
where I() is the expected Fisher information:
d2
I() = E [ 2 log p(y|)]
d
If is a vector, then I() is a matrix with (i, j) element equal to
d2
E [
log p(y|)]
didj
ENAR - March 2006
51
and
p() |I()|1/2
Theorem: Jeffreys prior is locally uniform and therefore non-informative.
ENAR - March 2006
52
Jeffreys prior for binomial proportion

If y B(n, ) then log p(y|) y log + (n y) log(1 ) and
d2
2
2
log
p(y|)
=
y
(n
y)(1
)
d2
Taking expectations:
d2
I() = E [ 2 log p(y|)] = E(y)2 + (n E(y))(1 )2
d
n
2
2
= n (n n)(1 ) =
.
(1 )
Then
1 1
p() [I()]1/2 1/2(1 )1/2 Beta( , ).
2 2
ENAR - March 2006
53
Jeffreys prior for normal mean

y1, ..., yn iid N (, 2), 2 known.
1 X
p(y|) exp( 2
(y )2)
2
1 X
log p(y|) 2
(y )2
2
d2
n
log
p(y|)
=
d2
2
constant with respect to .
Then I() constant and p() constant.
ENAR - March 2006
54
Jeffreys prior for normal variance

y1, ..., yn iid N (, 2), known.
1 X
p(y|) exp( 2
(y )2)
2
1 X
log p(y|) n log 2
(y )2
2
d2
n
3 X
2
log
p(y|)
=
(y
)
i
d 2
2 2 4
n
Take negative of expectation so that:

I() =
n
3
n
2
+
n
=
2 2 4
2 2
and therefore the Jeffreys prior is p() 1.

ENAR - March 2006
55
Invariance property of Jeffreys

The Jeffreys prior is invariant to one-to-one transformations = h()
Then:
p() = p()|
d
dh() 1
| = p()|
|
d
d
Under invariance, choose p() such that p() constructed as above

would match what would be obtained directly.
Consider p() = [I()]1/2 and evaluate I() at = h1():
d2
I() = E[ 2 log p(y|)]
d
d2
d
= E[ 2 log p(y| = h1())[ ]2]
d
d
ENAR - March 2006
56
d 2
= I()[ ]
d
d
Then [I()]1/2 = [I()]1/2 d
as required.
ENAR - March 2006
57
Multiparameter models - Intro

Most realistic problems require models with more than one parameter
Typically, we are interested in one or a few of those parameters
Classical approach for estimation in multiparameter models:
1. Maximize a joint likelihood: can get nasty when there are many
parameters
2. Proceed in steps
Bayesian approach: base inference on the marginal posterior
distributions of the parameters of interest.
Parameters that are not of interest are called nuisance parameters.
ENAR - March 2006
Nuisance parameters
Consider a model with two parameters (1, 2) (e.g., a normal
distribution with unknown mean and variance)
We are interested in 1 so 2 is a nuisance parameter
The marginal posterior distribution of interest is p(1|y)
Can be obtained directly from the joint posterior density
p(1, 2|y) p(1, 2)p(y|1, 2)
by integrating with respect to 2:
Z
p(1|y) = p(1, 2|y)d2
ENAR - March 2006
Nuisance parameters (contd)

Note too that
Z
p(1|y) =
p(1, |2, y)p(2|y)d2
The marginal of 1 is a mixture of conditionals on 2, or a weighted

average of the conditional evaluated at different values of 2. Weights
are given by marginal p(2|y)
ENAR - March 2006

Important difference with frequentists!
By averaging conditional p(1, |2, y) over possible values of 2, we
explicitly recognize our uncertainty about 2.
Two extreme cases:
1. Almost certainty about the value of 2: If prior and sample are very
informative about 2, marginal p(2|y) will be concentrated around
some value 2. In that case,
p(1|y) p(1|2, y)
2. Lots of uncertainty about 2: Marginal p(2|y) will assign relatively
high probability to wide range of values of 2. Point estimate 2 no
longer reliable. Important to average over range of values of 2.
ENAR - March 2006

In most cases, integral not computed explicitly
Instead, use a two-step simulation approach
(k)
1. Marginal simulation step: Draw value 2 of 2 from p(2|y) for

k = 1, 2, ...
(k)
2. Conditional simulation step: For each 2 , draw a value of 1 from
(k)
the conditional density p(1|2 , y)
Effective approach when marginal and conditional are of standard form.
More sophisticated simulation approaches later.
ENAR - March 2006
Example: Normal model

yi iid from N (, 2), both unknown
Non-informative prior for (, 2) assuming prior independence:
p(, 2) 1 2
Joint posterior:
p(, 2|y) p(, 2)p(y|, 2)
n
X
1
n2 exp( 2
(yi )2)
2 i=1
ENAR - March 2006
Note that
n
X
(yi )
(yi2 2yi + 2)
i=1
yi2 2n
y + n2
(yi y)2 + n(
y )2
by adding and subtracting 2n

y 2.
ENAR - March 2006
Example: Normal model (contd)

Let
1 X
(yi y)2
s =
n1 i
2
Then can write posterior for (, 2) as

2
p(, |y)
n2
1
exp( 2 [(n 1)s2 + n(
y )2])
2
Sufficient statistics are (

y , s2)
ENAR - March 2006
Example: Conditional posterior p(| 2, y)

Conditional on 2:
p(| 2, y) = N (
y , 2/n)
We know this from earlier chapter (posterior of normal mean when
variance is known)
We can also see this by noting that, viewed as a function of only:
p(| 2, y) exp(
n
2
)
(
y
)
2
2
that we recognize as the kernel of a N (

y , 2/n)
ENAR - March 2006
Example: Marginal posterior p( 2|y)

To get p( 2|y) we need to integrate p(, 2|y) over :
Z
1
2
2
[(n
1)s
+
n(
y
)
])d
2
2
2 Z
n
(n
1)s
2
n2
)
exp(
(
y
)
)d

exp(
2
2
2
2
2 p
(n
1)s
2 /n
n2 exp(
)
2
2 2
p( 2|y)
n2 exp(
Then
(n 1)s2
p( |y) ( )
exp(
)
2
2
which is proportional to a scaled-inverse 2 distribution with degrees
of freedom (n 1) and scale s2.
2
ENAR - March 2006
2 (n+1)/2
10
Recall classical result: conditional on 2, the distribution of the scaled

sufficient statistic (n 1)s2/ 2 is 2n1.
ENAR - March 2006
11
Normal model: analytical derivation

For the normal model, we can derive the marginal p(|y) analytically:
Z
p(|y) =
Z
p(, 2|y)d 2
1
1 n/2+1
( 2)
exp( 2 [(n 1)s2 + n(
y )2])d 2
2
2
Use the transformation
A
2 2
where A = (n 1)s2 + n(
y )2. Then
z=
d 2
A
= 2
dz
2z
ENAR - March 2006
12
and
Z
z n A
( ) 2 +1 2 exp(z)dz
A
z
0
Z
n 1
n/2
A
z 2 exp(z)dz
p(|y)
ENAR - March 2006
13

n/2
p(|y) A
n 1
2
exp (z)dz
Integrand is unnormalized Gamma(n/2, 1), so integral is constant w.r.t.
Recall that A = (n 1)s2 + n(

y )2. Then
p(|y) An/2
[(n 1)s2 + n(
y )2]n/2
n( y)2 n/2
[1 +
]
2
(n 1)s
the kernel of a tdistribution with n 1 degrees of freedom, centered
at y and with scale parameter s2/n
ENAR - March 2006
14
For the non-informative prior p(, 2) 2, the posterior distribution

of is a non-standard t. Then,
p(
y
|y) = tn1
s/ n
the standard t distribution.
ENAR - March 2006
15

We saw that
y
p( |y) = tn1
s/ n
Notice similarity to classical result: for iid normal observations from

N (, 2), given (, 2), the pivotal quantity
y
|, 2 tn1
s/ n
A pivot is a non-trivial function of the data and the parameter(s)
whose distribution, given , is independent of . Property deduced
from sampling distribution as above.
Baby example of pivot property: y N (, 1). Pivot is x = y .
Given , x N (0, 1), independent of .
ENAR - March 2006
16
Posterior predictive for future obs

Posterior predictive distribution for future observation y is a mixture:
Z Z
p(
y |y) =
p(
y |y, 2, )p(, 2|y)dd 2
First factor in integrand is just normal model, and it does not depend
on y at all.
To simulate y from posterior predictive distributions, do the following:
1. Draw 2 from Inv-2(n 1, s2)
2. Draw from N (
y , 2/n)
3. Draw y from N (, 2)
ENAR - March 2006
17
Can derive the posterior predictive distribution in analytic form. Note

that
Z
p(
y | 2, y) =
p(
y | 2, )p(| 2, y)d
Z
1
n
2
exp( 2 (
y ) ) exp( 2 ( y)2)d
2
2
After some algebra:
1 2
p(
y | , y) = N (
y , (1 + ) )
n
2
Using same approach as in deriving posterior distribution of , we find

that
1
p(
y |y) tn1(
y , (1 + )1/2s)
n
ENAR - March 2006
18
Normal data and conjugate prior

Recall that using a non-informative prior, we found that
p(| 2, y) N (
y , 2/n)
p( 2|y) Inv 2(n 1, s2)
Then, factoring p(, 2) = p(| 2)p( 2) the conjugate prior for 2
would also be scaled inverse 2 and for (conditional on 2) would be
normal. Consider
| 2
2
N(0, 2/0)
Inv-2(0, 02)
Jointly:
p(, 2) 1( 2)(0/2+1) exp(
ENAR - March 2006
1
2
2
[
)
])
0
0
0
0
2
2
19

Note that and 2 are not independent a priori.
Posterior density for (, 2):
Multiply likelihood by N-Inv-2(0, 2/0; 0, 02) prior
Expand the two squares in
Complete the square by adding and subtracting term depending on
y and 0
Then p(, 2|y) N-Inv-2(n, n2 /n; n, n2 ) where
0
n
0 +
y
0 + n
0 + n
= 0 + n, n = 0 + n
0 n
2
2
= 00 + (n 1)s +
(
y )2
0 + n
n =
n
nn2
ENAR - March 2006
20

Interpretation of posterior parameters:
n a weighted average as earlier
nn2 : sum of the sample sum of squares, prior sum of squares, and
additional uncertainty due to difference between sample mean and
prior mean.
ENAR - March 2006
21

Conditional posterior of : As before | 2, y N(n, 2/n).
Marginal posterior of 2: As before 2|y Inv-2(n, n2 ).
Marginal posterior of : As before |y tn (|n, n2 /n).
Two ways to sample from joint posterior distribution:
1. Sample from t and 2 from Inv-2
2. Sample 2 from Inv-2 and given 2, sample from N
ENAR - March 2006
22
Semi-conjugate prior for normal model

Consider setting independent priors for and 2:

2
N(0, 02)
Inv-2(0, 02)
Example: mean weight of students, visual inspection shows weights

between 100 and 200 pounds.
p(, 2) = p()p( 2) is not conjugate and does not lead to posterior
of known form
Can factor as earlier:
| 2, y N(n, n2)
ENAR - March 2006
23
with
n =
1
n
+
y
0
2
2
0
,
n
1
+
2
22
n2
1
02
1
+ n2
NOTE: Even though and 2 are independent a priori, they are not
independent in the posterior.
ENAR - March 2006
24
Semi-conjugate prior and p( 2|y)

The marginal posterior p( 2|y) can be obtained by integrating the joint
p(, 2|y) w.r.t. :
p( 2|y)
N(|0, 02) Inv2( 2|0, 02) N(yi|, 2)d
Integration can be performed by noting that integrand as function of

is proportional to normal density.
Keeping track of normalizing constants that depend on 2 is messy.
Easier to note that:
p(, 2|y)
2
p( |y) =
p(| 2, y)
ENAR - March 2006
25
so that
2
2 2
2
2
)
N(y
|,
)
)
Inv
(
|
,
N(|
,
i
0
0
0
0
2
p( |y)
N(|n, n2)
which is still a mess.
ENAR - March 2006
26
Semi-conjugate prior and p( 2|y)

From earlier page:
N(|0, 02) Inv 2( 2|0, 02) N(yi|, 2)
p( |y)
N(|n, n2)
2
Factors that depend on must cancel, and therefore we know that

p( 2|y) does not depend on in the sense that we can evaluate p( 2|y)
for a grid of values of 2 and any arbitrary value of .
Choose = n and then denominator simplifies to something
proportional to n1. Then
p( 2|y) n N(|0, 02) Inv 2( 2|0, 02) N(yi|, 2)
that can be evaluated for a grid of values of 2.
ENAR - March 2006
27
Implementing inverse CDF method

We often need to draw values from distributions that do not have a
nice standard form. Example is expression 3.14 in text, for p( 2|y)
in the normal model with semi-conjugate priors for and 2.
One approach is the inverse cdf method (see earlier notes), implemented
numerically.
We assume that we can evaluate the pdf (even in unnormalized form)
for a grid of values of in the appropriate range
ENAR - March 2006
28
Implementing inverse CDF method

To generate 1000 draws from a distribution p(|y) with parameter ,
do
1. Evaluate p(|y) for a grid of m values of . Let the evaluations be
denoted by (p1, p2, ..., pm)
Pm1
2. Compute the CDF over the grid: (p1, p1 + p2, ...., i=1 pi, 1) and
denote those (f1, f2, ..., fm)
3. Generate M uniform random variables u in [0, 1].
4. If u [fi1, fi], draw i.
ENAR - March 2006
29
Inverse CDF method - Example
1
2
3
4
5
6
7
8
9
10
Prob()
0.03
0.04
0.08
0.15
0.20
0.30
0.10
0.05
0.03
0.02
CDF (F)
0.03
0.07
0.15
0.30
0.50
0.80
0.90
0.95
0.98
1.00
Consider a parameter with values in (1, 10) with probability

distribution and cumulative distribution functions as above. Under
distribution, values of between 4 and 7 are more likely.
ENAR - March 2006
30
Inverse CDF method - Example

To implement method do:
1. Draw u U (0, 1). For example, in first draw u = 0.45
2. For u (fi1, fi), draw i.
3. For u = 0.45 (0.3, 0.5), draw = 5.
4. Alternative approach:
(a) Flip another coin v U (0, 1).
(b) Pick i1 if v 0.5 and pick i if v > 0.5
5. In example, for u = 0.45, would either choose = 4 or would choose
= 4 with probability 1/2.
6. Repeat many times M .
If M very large, we expect that about 50% of our draws will be
= 4, 5, or 6, about 2% will be 10, etc.
ENAR - March 2006
31
Example: Football point spreads

Data di, i = 1, ..., 672 are differences between predicted outcome of
football game and actual score.
Normal model:
di N (, 2)
Priors:
N (0, 22)
p( 2) 2
To draw values of (, 2) from posterior, do
First draw 2 from p( 2|y) using inverse cdf method
Then draw from p(| 2, y), normal.
To draw 2, evaluate p( 2|y) on the grid [150, 250].
ENAR - March 2006
32
Example: Rounded measurements (prob. 3.5)

Sometimes, measurements are rounded, and we do not observe true
values.
yi are observed rounded values, and zi are unobserved true
measurements.
If zi N (, 2), then
y|, 2 i
yi + 0.5
yi 0.5
Prior: p(, 2) 2
We are interested in posterior inference about (, 2) and in differences
between rounded and exact analysis.
ENAR - March 2006
33
Joint posterior contours
2
0
sigma2
exact normal
10
11
12
13
12
13
mu
2
0
sigma2
rounded
10
11
mu
KSU - April 2005
Multinomial model - Intro

Generalization of binomial model, for the case where observations can
have more than two possible values.
Sampling distribution: multinomial with parameters (1, ..., k ), the
probabilities associated to each of the k possible outcomes.
Example: In a survey, respondents may: Strongly Agree, Agree,
Disagree, Strongly Disagree, or have No Opinion when presented with
a statement such as Instructor for Stat 544 is spectacular.
ENAR - March 2006
34
Multinomial model - Sampling distn

Formally:
y = k 1 vector of counts of #s of observations in each outcome
j : probability of jth outcome
Pk
Pk
j=1 j = 1 and
j=1 yj = n
Sampling distribution:
y
p(y|) kj=1j j
ENAR - March 2006
35
Multinomial model - Prior

Conjugate prior for (1, ..., k ) is Dirichlet distribution, a multivariate
generalization of the Beta:
1
p(|) kj=1j j
with
j > 0j,
and
0 =
k
X
j=1
j > 0j,
and
k
X
j = 1
j=1
ENAR - March 2006
36
The j can be thought of as prior counts associated to jth outcome,

so that 0 would then be a prior sample size.
For the Dirichlet:
j (0 j )
V ar(j ) = 2
0(0 + 1)
ij
Cov(i, j ) = 2
0(0 + 1)
j
E(j ) = ,
0
ENAR - March 2006
37
Dirichlet distribution
The Dirichlet distribution is the conjugate prior for the parameters of
the multinomial model.
If (1, 2, ...K ) D(1, 2, ..., K ) then
(0)
1
p(1, ...k ) =
j j j ,
j (j )
where j 0, j j = 1, j 0 and 0 = j j .
Some properties:
E(j ) =
ENAR - March 2006
j
0
38
j (0 j )
02(0 + 1)
ij
Cov(i, j ) = 2
0(0 + 1)
V ar(j ) =
Note that the j are negatively correlated.

Because of the sum to one restriction, the pdf of the K-dimensional
random vector can be written in terms of K 1 random variables.
The marginal distributions can be shown to be Beta.
Proof for K = 3:
(1, 2, 3) 11 21
p(1, 2) =
1
2
(1 1 2)31
j (0)
ENAR - March 2006
39
To get marginal for 1, integrate p(1, 2) with respect to 2 with

limits of integration 0 and 1 1. Call normalizing constant Q and use
change of variable:
2
v=
1 1
so that 2 = v(1 1) and d2 = (1 1)dv.
The marginal is then:
Z
p(1) = Q
111(v(1 1))21
(1 1 v(1 1))31(1 1)dv

Z 1
= Q111(1 1)2+31
v 21(1 v)31dv
0
= Q111(1 1)2+31
ENAR - March 2006
(2)(3)
(2 + 3)
40
Then,
1 Beta(1, 2 + 3).
This generalizes to what is called the clumping property of the Dirichlet.
In general, if (1, ..., k ) D(1, ..., K ):
p(i) = Beta(1, 0 i)
ENAR - March 2006
41
Multinomial model - Posterior

Posterior must have Dirichlet form also:
1 yj
j
p(|y) kj=1j j
y +j 1
kj=1j j
n =
j (j
+ yj ) = 0 + n is total number of observations.
Posterior mean (a point estimate) of j is:

E(j |y) =
=
ENAR - March 2006
j + yj
0 + n
# of obs. of jth outcome
total # of obs
42
For P
j = 1 j: uniform non-informative prior on all vectors of j such
that j j = 1
For j = 0
restriction.
j: the uniform prior is on the log(j ) with same
In either case, posterior is proper if yj 1 j.
ENAR - March 2006
43
Multinomial model - samples from p(|y)

Gamma method: Two steps for each j :
1. Draw x1, x2, ..., xk from independent
Gamma(, (j + yj )) for any common
Pk
2. Set j = xj / i=1 xi
Beta method: Relies on properties of Dirichlet:
Marginal p(j |y) = Beta(j + yj , n (j + yj ))
Conditional p(j |j , y) = Dirichlet
ENAR - March 2006
44
Multinomial model - Example

Pre-election polling in 1988:
n = 1,447 adults in the US

y1 = 727 supported G. Bush (the elder)
y2 = 583 supported M. Dukakis
y3 = 137 supported other or had no opinion
If no other information available, can assume that observations are

exchangeable given . (However, if information on party affiliation is
available, unreasonable to assume exchangeability)
Polling is done under complex survey design. Ignore for now and
assume simple random sampling. Then, (y1, y2, y3) M ult(1, 2, 3)
Of interest: 1 2, the difference in the population in support of the
two major candidates.
ENAR - March 2006
45
Multinomial model - Example

Non-informative prior, for now: set 1 = 2 = 3 = 1
Posterior distribution is Dirichlet(728, 584, 138). Then:
E(1|y) = 0.502, E(2|y) = 0.403
Other quantities obtain by simulation (see next)
To derive p(1 2|y) do:
1. Draw m values of (1, 2, 3) from posterior
2. For each draw, compute 1 2
Results and program: see quantiles of posterior distributions of 1 and
2, credible set for 1 2 and Prob(1 > 2|y).
ENAR - March 2006
46
Proportion voting for Bush (the elder)
0.46
0.48
0.50
0.52
0.54
theta_1
KSU - April 2005
Difference between two candidates
0.0
0.05
0.10
0.15
0.20
theta_1 - theta_2
KSU - April 2005
Proportion voting for Dukakis
0.36
0.38
0.40
0.42
0.44
theta_2
KSU - April 2005
Example: Bioassay experiment

Does mortality in lab animals increase with increased dose of some
drug?
Experiment:
20 animals randomly allocated to four doses
(treatments), and number of dead animals within each dose recorded.
Dose xi
ni
No. of deaths yi
-0.863
-0.296
-0.053
0.727
5
5
5
5
0
1
3
5
Animals exchangeable within dose.

ENAR - March 2006
47
Bioassay example (contd)

Model
yi Bin (ni, i)
i are not exchangeable because probability of death depends on dose.
One way to model is with linear relationship:
i = + xi.
Not a good idea because i (0, 1).
Transform i:

logit (i) = log
ENAR - March 2006
i
1 i
48
Since logit (i) (, +), can use linear model.

regression)
E[logit (i)] = + xi.
(Logistic
Likelihood: if yi Bin (ni, i), then

p(yi|ni, i) (i)yi (1 i)niyi .
But recall that
so that
ENAR - March 2006
i
log
1 i

= + xi,
exp( + xi)
.
i =
1 + exp( + xi)
49
Then
p(yi|
, , ni, xi)

yi
niyi
exp( + xi)
exp( + xi)
1
1 + exp( + xi)
1 + exp( + xi)
Prior: p(, ) 1.
Posterior:
p(, |y, n, x) ki=1p(yi|, , ni, xi).
We first evaluate p(, |y, n, x) on the grid
(, ) [5, 10] [10, 40]
and use inverse cdf method to sample from posterior.
ENAR - March 2006
50

Posterior evaluated on 200 200 grid of values of (, ).
1
...
...
200
p(1|y)
p(2|y)
..
..
p(200|y)
1
2
..
..
200
p(1|y)
p(2|y)
...
...
p(200|y)
Entry (j, i) in grid is p(, | = i, = j , y).

Sum of column j entries is p(j |y) because:
Z
p(j |y) =
p(j , |y)d
ENAR - March 2006
51
Z
p(j |, y)p(|y)d
X
p(j |, y)
=
Sum of row i entries is p(i|y).
ENAR - March 2006
52

We sample from p(|y), and then from p(|, y) (or the other
way around).
To sample from p(|y):
1. Obtain empirical p(| = j , y), j = 1, ...200 by summing over the
.
2. Use inverse cdf method to sample from p(|y).
To sample from p(|, y):
1. Given a draw , choose appropriate column in grid
2. Use inverse cdf method on p(| = , y).
ENAR - March 2006
53

LD50 is dose at which probability of death is 0.5, or

E
yi
ni

= i = 0.5
so that
0.5 = logit1( + xi)

logit(0.5) = + xi
0 = + xi
Then LD50 is computed as xi = /.

ENAR - March 2006
54
Posterior distribution of the probability of death for each dose
box plot: theta
[4]
1.0
[3]
0.75
0.5
[2]
0.25
[1]
0.0
Posterior distributions of the regression parameters

Alpha: posterior mean = 1.297, with 95% credible set [-0.365, 3.755]
Beta: posterior mean = 11.52, with 95% credible set [3.572, 25.79]
alpha sample: 2000
beta sample: 2000
0.6
0.1
0.075
0.05
0.025
0.0
0.4
0.2
0.0
-2.5
0.0
2.5
5.0
-20.0
0.0
20.0
40.0
Posterior distribution of the LD50 in the log scale

Posterior mean: -0.1064
95% credible set: [-0.268, 0.1141]
LD50 sample: 2000

6.0
4.0
2.0
0.0
-1.0
-0.5
0.0
0.5
Multivariate normal model - Intro

Generalization of the univariate normal model, where now observations
are d 1 vectors of measurements
For the ith sample unit: yi = (yi1 , yi2 , ..., yid ), and yi N (, ), with
(, ) d 1 and d d respectively
For a sample of n iid observations:
n/2
p(y|, ) ||
1X
exp[
(yi )01(yi )]
2 i
1
||n/2 exp[ tr(yi )01(yi )]
2
1
||n/2 exp[ tr1S]
2
ENAR - March 2006
55
with
S=
(yi )(yi )0
S is the d d matrix of sample squared deviations and cross-deviations

from .
ENAR - March 2006
56
Multivariate normal model - Known

We want the posterior distribution of the mean p(|y) under the
assumption that the variance matrix is known.
Conjugate prior for is p(|0, 0) = N (0, 0), as in the univariate
case
Posterior for :
(|y, )
(
"
#)
X
1
0 1
( 0) ( 0) +
exp
(yi )01(yi )
2
i
p
that is a quadratic function of . Completing the square in the

exponent:
p(|y, ) = N (n, n)
ENAR - March 2006
57
with
n = (10 + n1y)(1 + n1)1
n = (1 + n1)1
ENAR - March 2006
58
Posterior precision is equal to the sum of prior and sample precisions:

1
1
1
n = 0 + n
Posterior mean is weighted average of prior and sample means:

1
1
1 1
n = (1
+
n
y
)(
0
0
0 + n )
From usual properties of multivariate normal distribution, can derive

marginal posterior distribution of any subvector of or conditional
posterior distribution of (1) given (2).
ENAR - March 2006
59
Let 0 = ((1) , (2) )0 and correspondingly, let

"
n =
and
"
n =
ENAR - March 2006
(1)
n
(2)
n
(1,1) (1,2)
n n
(2,1) (2,2)
n n
60

(1)
(1,1)
Marginal of subvector (1) is p((1)|, y) = N (n , n
).
Conditional of subvector (1) given (2) is

1|2 (2)
(2)
1|2
p((1)|(2), , y) = N ((1)
+
),
)
n
n
where

1|2 = (1,2)
(2,2)
n
n
1

(2,2)
1|2 = (1,1)
(1,2)
n
n
n
1
(2,1)
n
We recognize 1|2 as the regression coefficient in a regression of (1)

on (2)
ENAR - March 2006
61
Multivariate normal - Posterior predictive

We seek p(
y |y), and note that
p(
y , |y) = N (
y ; , )N (; n, n)
Exponential in p(
y , |y) is a quadratic form in (
y , ), so (
y , ) have a
normal joint posterior distribution.
Marginal p(
y |y) is also normal with mean and variance:
E(
y |y) = E(E(
y |, y)|y) = E(|y) = n
var(
y |y) = E(var(
y |, y)|y) + var(E(
y |, y)|y)
= E(|y) + var(|y)
= + n .
ENAR - March 2006
62
Multivariate normal - Sampling from p(

y |y)
With known, to draw a value y from p(
y |y) note that
Z
p(
y |y) =
p(
y |, y)p(|y)d
Then:
1. Draw from p(|y) = N (n, n)
2. Draw y from p(
y |, y) = N (, )
Alternatively (better), draw y directly from
p(
y |y) = N (n, + n)
ENAR - March 2006
63
Sampling from a multivariate normal
Want to sample y d 1 from N (, ).

Two common approaches
Using Cholesky decomposition of :
1. Get A, a lower triangular matrix such that AA0 =
2. Draw (z1, z2, ..., zd) iid N (0, 1) and let z = (z1, ..., zd)
3. Compute y = + Az
ENAR - March 2006
64
Sampling from a multivariate normal

Using sequential conditional sampling:
Use fact that all conditionals in a multivariate normal are also normal.
1. Draw y1 from N (1, (11))
2. Then draw y2|y1 from N (2|1, 2|1)
3. Etc.
For example, if d = 2,
1. Draw y1 from
N ((1), 2(1))
2. Draw y2 from
(2)
N (
ENAR - March 2006
(12)
( (12))2
(1)
2(2)
+ 2(2) (y1 ),
2(1) )
65
Non-informative prior for

A non-informative prior for is obtained by letting the prior precision
0 go to zero.
With the uniform prior, the posterior for is proportional to the
likelihood.
Posterior is proper only if n > d; otherwise, S is not of full column
rank.
If n > d,
p(|, y) = N (
y , /n)
ENAR - March 2006
66
Multivariate normal - unknown ,

The conjugate family of priors for (, ) is the normal-inverse Wishart
family.
The inverse Wishart is the multivariate generalization of the inverse 2
distribution
If |0, 0 Inv-Wishart0 (1
0 ) then

1
(0 +d+1)/2
p(|0, 0) ||
exp tr(01)
2
for positive definite and symmetric.
For the Inv-Wishart with 0 degrees of freedom and scale 0: E() =
(0 d 1)10
ENAR - March 2006
67

Then conjuage prior is
Inv-Wishart0 (01)
| N(0, /0)
which corresponds to
p(, )
ENAR - March 2006
||(0+d)/2+1

1
0
exp
tr(01) ( 0)01( 0)
2
2
68

Posterior must also be in the Normal-Inv-Wishart form
Results from univariate normal generalize directly:
|y Inv-Wishartn (1
n )
|, y N (n, /n)
|y Mult-tnd+1(n, n/(n(n d + 1)))
y|y Mult-t
Here:
n
0
0 +
y
0 + n
0 + n
= 0 + n
n =
n
n = 0 + n
ENAR - March 2006
69
n
S
n0
= 0 + S +
(
y 0)(
y 0)0
0 + n
X
=
(yi y)(yi y)0
i
ENAR - March 2006
70
Multivariate normal - sampling ,

To sample from p(, |y) do: (1) Sample from p(|y) (2) Sample
from p(|, y)
To sample from p(
y |y), use drawn (, ) and obtain draw y from
N (, )
Sampling from Wishart (S):
1. Draw independent
d 1 vectors 1, 2, ..., from a N (0, S)
P
2. Let Q = i=1 ii0
Method works when > d.
Inv-Wishart.
If Q Wishart then Q1 =
We have already seen how to sample from a multivariate normal given

mean vector and covariance matrix.
ENAR - March 2006
71
Multivariate normal - Jeffreys prior

If we start from the conjugate normal-inv-Wishart prior and let :
0 0
0 1
|0| 0
then resulting prior is Jeffreys prior for (, ):
p(, ) ||(d+1)/2
ENAR - March 2006
72
Example: bullet lead concentrations

In the US, four major manufacturers produce all bullets fired. One of
them is Cascade.
A sample of 200 round-nosed .38 caliber bullets with lead tips were
obtained from Cascade.
Concentration of five elements (antimony, copper, arsenic, bismuth,
and silver) in the lead alloy was obtained. Data for for Cascade are
stored in federal.data in the courses web site.
Assuming that the 200 5 1 observation vectors for Cascade can
be represented by a multivariate normal distribution (perhaps after
transformation), we are interested in the posterior distribution of the
mean vector and of functions of the covariance parameters.
ENAR - March 2006
Particular quantities of interest for inference include:

The mean trace element concentrations (1, ..., 5)
The correlations between trace element concentrations
(11, 12, ..., 45).
The largest eigenvalue of the covariance matrix
In the data file, columns correspond to antimony, copper, arsenic,
bismuth, and silver, in that order.
For this example, the antimony concentration was divided by 100
ENAR - March 2006
Bullet lead example - Model and prior

We concentrate on the correlations that involve copper (four of them).
y|, N (, )
We use the conjugate Normal-Inv-Wishart family of priors, and choose

parameters for p(|0, 0) first:
0,ii = (100, 5000, 15000, 1000, 100)
0,ij
= 0 i, j
0 = 7
ENAR - March 2006
For p(|, mu0, 0):

0 = (200, 200, 200, 100, 50)
0 = 10
Low values for 0 and 0 suggest little confidence in prior guesses
0, 0
We set 0 to be diagonal apriori: we have some information about the
variances of the five element concentrations, but no information about
their correlation.
Sampling from posterior distribution: We follow the usual approach:
1. Draw from Inv-Wishartn (1
n )
2. Draw from N (n, /n)
ENAR - March 2006
3. For each draw, compute:

(a) the ratio between 1 (largest eigenvalue of ) and 2 (second
largest eigenvalue of )
(b) the four correlations copper,j = copper,j /(copper j ), for j
{antimony, arsenic, bismuth, silver}.
We are interested in the posterior distributions of the five means, the
eigenvalues ratio, and the four correlation coefficients
ENAR - March 2006
Scatter plot of the posterior mean concentrations
ENAR - March 2006
Results from R program

# antimony
c(mean(sample.mu[,1]),sqrt(var(sample.mu[,1])))
[1] 265.237302
1.218594
quantile(sample.mu[,1],probs=c(0.025,0.05,0.5,0.95,0.975))
2.5%
5.0%
50.0%
95.0%
97.5%
262.875 263.3408 265.1798 267.403 267.7005
# copper
[1] 259.36335
5.20864
2.5%
5.0%
50.0%
95.0%
97.5%
249.6674 251.1407 259.5157 268.0606 269.3144
# arsenic
[1] 231.196891
8.390751
ENAR - March 2006
2.5%
5.0%
50.0%
95.0%
97.5%
213.5079 216.7256 231.6443 243.5567 247.1686
# bismuth
[1] 127.248741
1.553327
2.5%
5.0%
50.0%
95.0%
97.5%
124.3257 124.6953 127.2468 129.7041 130.759
# silver
[1] 38.2072916 0.7918199
2.5%
5.0%
50.0%
95.0%
97.5%
36.70916 36.93737 38.1971 39.54205 39.73169
ENAR - March 2006
Posterior of 1/2 of
ENAR - March 2006
Summary statistics of posterior dist. of ratio

c(mean(ratio.l),sqrt(var(ratio.l)))
[1] 5.7183875 0.8522507
quantile(ratio.l,probs=c(0.025,0.05,0.5,0.95,0.975))
2.5%
5.0%
50.0%
95.0%
97.5%
4.336424 4.455404 5.606359 7.203735 7.619699
ENAR - March 2006
10
Correlation of copper with other elements
ENAR - March 2006
11
Summary statistics for correlations

# j = antimony
c(mean(sample.rho[,1]),sqrt(var(sample.rho[,1])))
[1] 0.50752063 0.05340247
quantile(sample.rho[,1],probs=c(0.025,0.05,0.5,0.95,0.975))
2.5%
5.0%
50.0%
95.0%
97.5%
0.4014274 0.4191201 0.5076544 0.5901489 0.6047564
# j = arsenic
[1] -0.56623609 0.04896403
2.5%
5.0%
50.0%
95.0%
97.5%
-0.6537461 -0.6448817 -0.570833 -0.4857224 -0.465808
# j = bismuth
[1] 0.63909311 0.04087149
ENAR - March 2006
12
2.5%
5.0%
50.0%
95.0%
97.5%
0.5560137 0.5685715 0.6429459 0.7013519 0.7135556
# j = silver
[1] 0.03642082 0.07232010
2.5%
5.0%
50.0%
95.0%
97.5%
-0.09464575 -0.0831816 0.03370379 0.164939 0.1765523
ENAR - March 2006
13
S-Plus code
#
#
#
#
#
#
#
#
Cascade example
We need to define two functions, one to generate random vectors from
a multivariate normal distribution and the other to generate random
matrices from a Wishart distribution.
Note, that a function that generates random matrices from a
inverse Wishart distribution is not necessary to be defined
since if W ~ Wishart(S) then W^(-1) ~ Inv-Wishart(S^(-1))
# A function that generates random observations from a

# Multivariate Normal distribution: rmnorm(a,b)
# the parameters of the function are
#
a = a column vector of kx1
#
b = a definite positive kxk matrix
#
rmnorm_function(a,b) {
k_nrow(b)
zz_t(chol(b))%*%matrix(rnorm(k),nrow=k)+a
zz
}
# A function that generates random observations from a
# Wishart distribution: rwishart(a,b)
# the parameters of the function are
#
a = df
#
b = a definite positive kxk matrix
# Note a must be > k
ENAR - March 2006
14
rwishart_function(a,b) {
k_ncol(b)
m_matrix(0,nrow=k,ncol=1)
cc_matrix(0,nrow=a,ncol=k)
for (i in 1:a) { cc[i,]_rmnorm(m,b) }
w_t(cc)%*%cc
w
}
#
#Read the data
#
y_scan(file="cascade.data")
y_matrix(y,ncol=5,byrow=T)
y[,1]_y[,1]/100
means_rep(0,5)
for (i in 1:5){ means[i]_mean(y[,i]) }
means_matrix(means,ncol=1,byrow=T)
n_nrow(y)
#
#Assign values to the prior parameters
#
v0_7
Delta0_diag(c(100,5000,15000,1000,100))
k0_10
mu0_matrix(c(200,200,200,100,50),ncol=1,byrow=T)
# Calculate the values of the parameters of the posterior
#
mu.n_(k0*mu0+n*means)/(k0+n)
v.n_v0+n
k.n_k0+n
ENAR - March 2006
15
ones_matrix(1,nrow=n,ncol=1)
S = t(y-ones%*%t(means))%*%(y-ones%*%t(means))
Delta.n_Delta0+S+(k0*n/(k0+n))*(means-mu0)%*%t(means-mu0)
#
# Draw Sigma and mu from their posteriors
#
samplesize_200
lambda.samp_matrix(0,nrow=samplesize,ncol=5) #this matrix will store
#eigenvalues of Sigma
rho.samp_matrix(0,nrow=samplesize,ncol=5)
mu.samp_matrix(0,nrow=samplesize,ncol=5)
for (j in 1:samplesize) {
# Sigma
SS = solve(Delta.n)
# The following makes sure that SS is symmetric
for (pp in 1:5) { for (jj in 1:5) {
if(pp<jj){SS[pp,jj]=SS[jj,pp]} } }
Sigma_solve(rwishart(v.n,SS))
# The following makes sure that Sigma is symmetric
for (pp in 1:5) { for (jj in 1:5) {
if(pp<jj){Sigma[pp,jj]=Sigma[jj,pp]} } }
# Eigenvalue of Sigma
lambda.samp[j,]_eigen(Sigma)$values
# Correlation coefficients
for (pp in 1:5){
rho.samp[j,pp]_Sigma[pp,2]/sqrt(Sigma[pp,pp]*Sigma[2,2]) }
# mu
mu.samp[j,]_rmnorm(mu.n,Sigma/k.n)
ENAR - March 2006
16
}
# Graphics and summary statistics
sink("cascade.output")
# Calculate the ratio between max(eigenvalues of Sigma)
# and max({eigenvalues of Sigma}\{max(eigenvalues of Sigma)})
#
ratio.l_sample.l[,1]/sample.l[,2]
# "Histogram of the draws of the ratio"
postscript("cascade_lambda.eps",height=5,width=6)
hist(ratio.l,nclass=30,axes = F,
xlab="eigenvalue1/eigenvalue2",xlim=c(3,9))
axis(1)
dev.off()
# Sumary statistics of the ratio of the eigenvalues
c(mean(ratio.l),sqrt(var(ratio.l)))
quantile(ratio.l,probs=c(0.025,0.05,0.5,0.95,0.975))
#
# correlations with copper
#
postscript("cascade_corr.eps",height=8,width=8)
par(mfrow=c(2,2))
# "Histogram of the draws of corr(antimony,copper)"
hist(sample.rho[,1],nclass=30,axes = F,xlab="corr(antimony,copper)",
xlim=c(0.3,0.7))
axis(1)
# "Histogram of the draws of corr(arsenic,copper)"
hist(sample.rho[,3],nclass=30,axes = F,xlab="corr(arsenic,copper)")
axis(1)
ENAR - March 2006
17
# "Histogram of the draws of corr(bismuth,copper)"

hist(sample.rho[,4],nclass=30,axes = F,xlab="corr(bismuth,copper)",
xlim=c(0.5,.8))
axis(1)
# "Histogram of the draws of corr(silver,copper)"
hist(sample.rho[,5],nclass=25,axes = F,xlab="corr(silver,copper)",
xlim=c(-.2,.3))
axis(1)
dev.off()
# Summary statistics of the dsn. of correlation of copper with
# antimony
# arsenic
# bismuth
# silver
#
# Means
#
mulabels=c("Antimony","Copper","Arsenic","Bismuth","Silver")
postscript("cascade_means.eps",height=8,width=8)
pairs(sample.mu,labels=mulabels)
dev.off()
ENAR - March 2006
18
# Summary statistics of the dsn. of mean of

# antimony
# copper
# arsenic
# bismuth
# silver
q()
ENAR - March 2006
19
Advanced Computation
Approximations based on posterior modes
Simulation from posterior distributions
Markov chain simulation
Why do we need advanced computational methods?
Except for simpler cases, computation is not possible with available
methods:
Logistic regression with random effects
Normal-normal model with unknown sampling variances j2
Poisson-lognormal hierarchical model for counts.
ENAR - March 26, 2006
Strategy for computation

If possible, work in the log-posterior scale.
Factor posterior distribution: p(, |y) = p(|, y)p(|y)
Reduces to lower-dimensional problem
May be able to sample on a grid
Helps to identify parameters most influenced by prior
Re-parametrizing sometimes helps:
Create parameters with easier interpretation
Permit normal approximation (e.g., log of variance or log of Poisson
rate or logit of probability)
Normal approximation to the posterior

It is often reasonable to approximate a complicated posterior
distribution using a normal (or a mixture of normals) approximation.
Approximation may be a good starting point for more sophisticated
methods.
Computational strategy:
1. Find joint posterior mode or mode of marginal posterior distributions
(better strategy if possible)
2. Fit normal approximations at the mode (or use mixture of normals if
posterior is multimodal)
Notation:
p(|y): joint posterior of interest (target distribution)
q(|y): un-normalized density, typically p()p(y|).

= (, ), typically lower dimensional than .
Practical advice: computations are often easier with log posteriors than
with posteriors.
Finding posterior modes

To find the mode of a density, we maximize the function with respect
to the parameter(s). Optimization problem.
Modes are not interesting per se, but provide the first step for
analytically approximating a density.
Modes are easier to find than means: no integration, and can work
with un-normalized density
Multi-modal posteriors pose a problem: only way to find multiple modes
is to run mode-finding algorithms several times, starting from different
locations in parameter space
Whenever possible, shoot for finding the mode of marginal and
conditional posteriors. If = (, ), and

p(, y) = p(|, y)p(|y)
then
1. Find mode
y)
2. Then find by maximizing p(| = ,
Many algorithms to find modes. Most popular include Newton-Raphson
and Expectation-Maximization (EM).
Taylor approximation to p(|y)

Second order expansion of log p(|y) around the mode is
2

+ 1 ( )
0 d log p(|y)
log p(|y) log p(|y)

( ).
2
2
d
=
Linear term in expansion vanishes because log posterior has a zero
derivative at the mode.
Considering log p(|y) approximation as a function of , we see that:
is constant and
1. log p(|y)
1
0
log p(|y) ( )
2

d
log p(|y)
( )
2
d
=
is proportional to a log-normal density

Then, for large n and in interior of parameter space:

[I()]
1)
p(|y) N(,
with
d2
I() = 2 log p(|y),
d
the observed information matrix.
Normal and normal-mixture approximation

For one mode, we saw that
V )
pnapprox(|y) = N(,
with
1
V = [L00()]
1
or [I()]
is called the curvature of the log posterior at the mode.

L00()
Suppose now that p(|y) has K modes.
Approximation to p(|y) is now a mixture of normals:
pnapprox(|y)
k N(k , Vk )
k
The k reflect the relative mass associated to each mode

It can be shown that k = q(k |y)|Vk |1/2.
The mixture-normal approximation is then
pnapprox(|y)
X
k
1
k )}.
q(k |y) exp{ ( k )0V1
(
k
2
A more robust approximation can be obtained by using a tkernel

instead of the normal kernel:
X
k )](d+)/2,
ptapprox(|y)
q(k |y)[ + ( k )0V1
(
k
k
with d the dimension of and relatively low. For most problems,

= 4 works well.
10
Sampling from a mixture approximation

To sample from the normal-mixture approximation:
1. First choose one of the K normal components, using relative
probability masses k as multinomial probabilities.
2. Given a component, draw a value from either the normal or the t
density.
Reasons not to use normal approximation:
When mode of parameter is near edge of parameter space (e.g., in
SAT example)
When even transformation of parameter makes normal approximation
crazy.
Can do better than normal approximation using more advanced
methods
11
Simulation from posterior - Rejection sampling

Idea is to draw values of p(|y), perhaps by making use of an
instrumental or auxiliary distribution g(|y) from which it is easier to
sample.
The target density p(|y) need only be known up to the normalizing
constant.
Let q(|y) = p()p(y; ) be the un-normalized posterior distribution so
that
q(|y)
.
p(|y) = R
p()p(y; )d
Simpler notation:
1. Target density: f (x)
12
2. Instrumental density: g(x)

3. Constant M such that f (x) M g(x)
The following algorithm produces a variable Y that is distributed
according to f :
1. Generate X g and U U[0,1]
2. Set Y = X if U f (X)/M g(X)
3. Reject draw otherwise.
Proof: The distribution function of Y is given by:

P (Y y) = P
X y|U

f (X)
M g(X)

(X)
P X y, U Mf g(X)

f (X)
P U M g(X)
13
Then
Ry
P (Y y) =
R f (x)/M g(x)
dug(x)dx
0
R R f (x)/M g(x)
dug(x)dx
0
R
1 y
M
f (x)dx
R
.
1
f (x)dx
M
Since the last expression equals
Ry
f (x)dx, we have proven the result.
In the Bayesian context, q(|y) (the un-normalized posterior) plays the

role of f (x) above.
When both f and g are normalized densities:
1. The probability of accepting a draw is 1/M , and expected number
of draws until one is accepted is M .
14
2. For each f , there will be many instrumental densities g1, g2, ....
Choose the g that requires smallest bound M
3. M is necessarily larger than 1, and will approach minimum value 1
when g closely imitates f .
In general, g needs to have thicker tails than f for f /g to remain
bounded for all x.
Cannot use a normal g to generate values from a Cauchy f .
Can do the opposite, however.
Rejection sampling can be used within other algorithms, such as the
Gibbs sampler (see later).
15
Rejection sampling - Getting M

Finding the bound M can be a problem, but consider the following
implementation approach recalling that
q(|y) = p()p(y; ).
Let g() = p() and draw a value from the prior.

Draw U U (0, 1).
where is the mode of p(y; ).
Let M = p(y; ),
Accept the draw if
u
p()p(y; ) p(y; )
q(|y)
=
.
=
M p() p(y; )p() p(y; )
Those from the prior that are likely under the likelihood are kept in
the posterior sample.
16
Importance sampling
Also known as SIR (Sampling Importance Resampling), the method is
no more than a weighted bootstrap.
Suppose we have a sample of draws (1, 2, ..., n) from the proposal
distribution g(). Can we convert it to a sample from q(|y)?
For each i, compute
i = q(i|y)/g(i)
X
wi = i/
j .
j
Draw from the discrete distribution over {1, ..., n} with weight wi
on i, without replacement.
17
The sample of re-sampled s is approximately a sample from q(|y).

Proof: suppose that is univariate (for convenience). Then:
Pr( a)
n
X
wi1,a(i)
i=1
i 1,a (i )
i P
n1 i i
Eg q(|y)
g() 1,a (i )
Eg q(|y)
g()
R
q(|y)d
R a
q(|y)d
Z a
p(|y)d.
18
The size of the re-sample can be as large as desired.

The more g resembles q, the smaller the sample size that is needed for
the re-sample to approximate well the target distribution p(|y).
A consistent estimator of the normalizing constant is
n
i.
19
Importance sampling in a different context
Suppose that we wish to estimate E[h()] for p(|y) (e.g., a

posterior mean).
If N values i can be drawn from p(|y), can compute Monte Carlo
estimate:
X
1
E[h()]
=
h(i).
N
Sampling from p(|y) may be difficult. But note:
Z
E[h()|y] =
h()p(|y)
g()d.
g()
20
Generate values from g(), and estimate

X
p(i|y)
1
h(i)
E[h()]
=
N
g(i)
The w(i) = p(i|y)/g(i) are the same importance weights from
earlier.
Will not work if tails of g are short relative to tails of p.
21
Importance sampling (contd)

Importance sampling is most often used to improve normal
approximation to posterior.
Example: suppose that you have a normal (or t) approximation to the
posterior and use it to draw a sample that you hope approximates a
sample from p(|y).
Importance sampling can be used to improve sample:
1. Obtain large sample of size L: (1, ..., L) from the approximation
g().
2. Construct importance weights w(l) = q(l|y)/g(l)
3. Sample k < L values from (1, ..., L) with probability proportional
to the weights, without replacement.
22
Why without replacement? If weights are approximately constant,

can re-sample with replacement. If some weights are large, re-sample
would favor values of with large weights repeteadly unless we sample
without replacement.
Difficult to determine whether importance sampling draws approximate
draws from posterior
Diagnostic: monitor weights and look for outliers
Note: draws from g() can be reused for other p(|y)!
Given draws (1, ..., m) from g(), can investigate sensitivity to prior by
re-computing importance weights. For posteriors p1(|y) and p2(|y),
compute (for the same draws of )
pj (i|y)
w (i) =
,
g(i)
j
j = 1, 2.
23
Markov chain Monte Carlo

Methods based on stochastic process theory, useful for approximating
target posterior distributions.
Iterative: must decide when convergence has happened
Quite general, can be implemented where other methods fail
Can be used even in high dimensional examples
Three methods: Metropolis-Hastings, Metropolis and Gibbs sampler.
M and GS special cases of M-H.
24
Markov chains
A process Xt in discrete time t = 0, 1, 2, ..., T where
E(Xt|X0, X1, ..., Xt1) = E(Xt|Xt1)
is called a Markov chain.
A Markov chain is irreducible if it is possible to reach all states from
any other state:
pn(j|i) > 0,
pm(i|j) > 0,
m, n > 0
where pn(j|i) denotes probability of getting to state j from state i in

n steps.
25
A Markov chain is periodic with period d if pn(i|i) = 0 unless n = kd

for some integer k. If d = 2, the chain returns to i in cycles of 2k
steps.
If d = 1, the chain is aperiodic.
Theorem: Finite-state, irreducible, aperiodic Markov chains have a
limiting distribution: limn pn(j|i) = , with pn(j|i) the probability
that we reach j from i after n steps or transitions.
26
Properties of Markov chains (contd)

Ergodicity: If a MC is irreducible and aperiodic and has stationary
distribution then we have ergodicity:
a
n
1X
a(Xt)
n t
E{a(X)} as n .
a
n is an ergodic average.
Also, rate of convergence can be calculated and is geometric.
27
Numerical standard error

Sequence {X1, X2, ..., Xn} is not iid.
The asymptotic standard error of an is approximately
s
X
a2
{1 + 2
i(a)}
n
i
where i is the ith lag autocorrelation in the function a{Xt}.
First term a2/n is the usual sampling variance under iid sampling.
Second term is a penalty for the fact that sample is not iid and is
usually bigger than 1.
Often, the standard error is computed assuming a finite number of lags.
28
Markov chain simulation

Idea: Suppose that sampling from p(|y) is hard, but that we can
generate (somehow) a Markov chain {(t), t T } with stationary
distribution p(|y).
Situation is different from the usual stochastic process case:
Here we know the stationary distribution.
We seek an algorithm to transition from (t) to (t+1)) and that will
take us to the stationary distribution.
Idea: start from some initial guess 0 and let the chain run for n steps
(n large), so that it reaches its stationary distribution.
After convergence, all additional steps in the chain are draws from the
stationary distribution p(|y).
29
MCMC methods all based on the same idea; difference is just in how
the transitions in the MC are created.
In MCMC simulation, we generate at least one MC for each parameter
in the model. Often, more than one (independent) chain for each
parameter
30
The Gibbs Sampler

An iterative algorithm that produces Markov chains with joint stationary
distribution p(|y) by cycling through all possible conditional posterior
distributions.
Example: suppose that = (1, 2, 3), and that the target distribution
is p(|y). Steps in the Gibbs sampler are:
1.
2.
3.
4.
Start with a guess (10, 20, 30)

Draw 11 from p(1|2 = 20, 3 = 30, y)
Draw 21 from p(2|1 = 11, 3 = 30, y)
Draw 31 from p(3|1 = 11, 2 = 21, y)
Steps above complete one iteration of the GS

Repeat the steps above n times, and after convergence (see later),
draws (n+1, n+2, ...) are sample from stationary distribution p(|y).
31
The Gibbs Sampler (contd)

Baby example: = (1, 2) are bivariate normal with mean y = (y1, y2),
variances 12 = 22 = 1 and covariance . Then:
p(1|2, y) N (y1 + (2 y2), 1 2)
and
p(2|1, y) N (y2 + (1 y1), 1 2)
In this case, would not need GS, just for illustration
See figure: y1 = 0, y2 = 0, and = 0.8.
32
Example:
normal
Gibbs sampling in the bivariate
Just as illustration, we consider the trivial problem of sampling from

the posterior distribution of a bivariate normal mean vector.
Suppose that we have a single observation vector (y1, y2) where

y1
y2

N

1
1
,
.
2
1
(1)
With a uniform prior on (1, 2), the joint posterior distribution is

KSU - April 2005
1
|y
2

N

y1
1
,
.
y2
1
(2)
1
The conditional distributions are

1|2, y
N(y1 + (2 y2), 1 2)
2|1, y
N(y2 + (1 y1), 1 2)
We generate four independent chains, starting from four different

corners of the 2-dimensional parameter space.
KSU - April 2005
After 50 iterations..
After 1000 iterations, and eliminating the path lines.
The non-conjugate normal example

Let y N (, 2), with (, 2) unknown.
Consider the semi-conjugate prior:
N (0, 02)
2 Inv 2(0, 02).
Joint posterior distribution is
( 1 2
2 n
2
2
p(, 2) ( ) e
P(y ) ) (
i
1 ( )2 )
0
202
2
0 0
0
(
)
2 ( 2 +1))
2 2
( )
33
Non-conjugate example (contd)
Derive full conditional distributions p(| 2, y) and p( 2|, y).
For :
1 1 X
1
p(| , y) exp{ [ 2
(yi )2 + 2 ( 0)2]}
2
0
1
exp{ [(n02 + 2)2 2(n02y + 20)]}
2
2

2
2
2
1 n0 + 2
n0 y + 0
exp{
[ 2
]}
2 202
n02 + 2
2
N (n, n2),
34
where
n =
n
1
y
02 0
1
n
+
2
02
and
n2
n
2
1
.
+ 12
0
35
Non-conjugate normal example (contd)

The full conditional for 2 is
2
p( |, y)
n+
1X
2 ( 2 0 +1)
( )
exp{ [ (yi
)2 + 002]}
Inv 2(n, n2 ),
where
n = 0 + n
02 = nS 2 + 002,
and
S=n
(yi )2.
36
Non-conjugate normal example (contd)

For the non-conjugate normal model, Gibbs sampling consists of
following steps:
1. Start with a guess for 2, 2(0).
2. Draw (1) from a normal
(n( 2(0)), n2( 2(0))), where
with
n( 2(0)) =
2(0)
y + 12 0
n
2(0)
and
n2( 2(0))
mean
+ 12
and
variance
1
= n
1 .
+
2
2(0)
0
37
3. Next draw 2(1) from Inv 2(n, n2 ((1))), where

n2 ((1)) = nS 2((1)) + 002,
and
2
(1)
S (
)=n
(yi (1))2.
4. Repeat many times.
38
A more interesting example

Poisson counts, with a change point: consider a sample of size n of
counts (y1, ..., yn) distributed as Poisson with some rate, and suppose
that the rate changes, so that
For i = 1, ..., m ,
yi Poi()
For i = m + 1, ..., n ,
yi Poi()
and m is unknown.
Priors on (, , m):

Ga(, )
Ga(, )
m U[1,n]
39
Joint posterior distribution:
p(, , m|y)
m
Y
eyi
i=1
n
Y
eyi 1e 1e n1
i=m+1
y1 +1e(m+)y2 +1e(nm+)
with
y1
Pm
i=1 yi
and
y2
Pn
i=m+1 yi .
Note: if we knew m, problem would be trivial to solve. Thus Gibbs,

where we condition on all other parameters to determine Markov chains,
appears to be the right approach.
To implement Gibbs sampling, need full conditional distributions. Pick

and choose pieces that depend on each parameter
40
Conditional of :
p(|m, , y)
m
i=1 yi +1
Ga( +
m
X
exp((m + ))
yi, m + )
i=1
Conditional of :
p(|, m, y)
n
i=m+1 yi +1
Ga( +
n
X
exp((n m + ))
yi, n m + )
i=m+1
Conditional of m = 1, 2, ..., n:
p(m|, , y) = c1q(m|, )
41
1 y1 +1
c
exp((m
y2 +1
+ ))
exp((n m + )).
Note that all terms in joint posterior depend on m, so conditional of

m is proportional to joint posterior.
Distribution does not look like any standard form so cannot sample
directly. Need to obtain normalizing constant, easy to do for relatively
small n and for this discrete case:
c=
n
X
q(k|, , y).
k=1
To sample from p(m|, , y) can use inverse cdf on a grid, or other

methods.
42
Non-standard distributions
It may happen that one or more of the full conditionals is not a standard
distribution
What to do then?
Try direct simulation: grid approximation, rejection sampling
Try approximation: normal or t approximation, need mode at each
iteration (see later)
Try more general Markov chain algorithms: Metropolis or MetropolisHastings.
43
Metropolis-Hastings algorithm
More flexible transition kernel: rather than requiring sampling from
conditional distributions, M-H permits using many other proposal
densities
Idea: instead of drawing sequentially from conditionals as in Gibbs,
M-H jumps around the parameter space
The algorithm is the following:
1. Given a draw t in iteration t, sample a candidate draw from a
proposal distribution J(|)
2. Accept the draw with probability
p(|y)/J(|)
r=
.
p(|y)/J(| )
44
3. Stay in place (do not accept the draw) with probability 1 r, i.e.,
(t+1) = (t).
Remarkably, the proposal distribution (text calls it jump distribution)
can have just about any form.
When proposal distribution is symmetric, i.e.
J(|) = J(|),
Metropolis-Hastings acceptance probability is
p(|y)
r=
.
p(|y)
This is the Metropolis algorithm
45
Proposal distributions
Convergence does not depend on J, but rate of convergence does.
Optimal J is p(|y) in which case r = 1.
Else, how do we choose J?
1. It is easy to get samples from J
2. It is easy to compute r
3. It leads to rapid convergence and mixing: jumps should be large
enough to take us everywhere in the parameter space but not too
large so that draw is accepted (see figure from Gilks et al. (1995)).
Three main approaches:
random walk M-H (most popular),
independence sampler, and approximation M-H
46
Independence sampler
Proposal distribution Jt does not depend on t.
Just find a distribution g() and generate values from it
Can work very well if g(.) is a good approximation to p(|y) and g has
heavier tails than p.
Can work awfully bad otherwise
Acceptance probability is
p(|y)/g()
r=
p(|y)/g()
47
With large samples (central limit theorem operating) proposal could be

normal distribution centered at mode of p(|y) and with variance larger
than inverse of Fisher information evaluated at mode.
48
Random walk M-H

Most popular, easy to use.
Idea: generate candidate using random walk.
Proposal is often normal centered at current draw:
J(|(t1)) = N (|(t1), V )
Symmetric: think of drawing values (t1) from a N (0, V ). Thus,
r simplifies.
Difficult to choose V :
1. V too small: takes long to explore parameter space
49
2. V too large: jumps to extremes are less likely to be accepted. Stay

in the same place too long
3. Ideal V : posterior variance. Unknown, so might do some trial and
error runs
Optimal acceptance rate (from some theory results) are between 25%
and 50% for this type of proposal distribution. Gets lower with higher
dimensional problem.
50
Approximation M-H
Idea is to improve approximation of J to p(|y) as we know more about
.
E.g., in random walk M-H, can perhaps increase acceptance rate by
considering
J(|(t1)) = N (|(t1), V(t1) )
variance also depends on current draw
Proposals here typically not symmetric, requires full r expression.
51
Starting values
If chain is irreducible, choice of 0 will not affect convergence.
With multiple chains (see later), can choose over-dispersed starting
values for each chain. Possible algorithm:
1. Find posterior modes
2. Create over-dispersed approximation to posterior at mode (e.g., t4)
3. Sample values from that distribution
Not much research on this topic.
52
How many chains?

Even after convergence, draws from stationary distribution are
correlated. Sample is not i.i.d.
An iid sample of size n can be obtained by keeping only last draw from
each of n independent chains. Too inefficient.
Compromise: If autocorrelation at lag k and larger is negligible, then
can generate an almost i.i.d. sample by keeping only every kth draw in
the chain after convergence. Need a very long chain.
To check convergence (see later) multiple chains can be generated
independently (in parallel).
Burn-in:
iterations required to reach
(approximately). Not used for inference.
stationary
distribution
53
Convergence
Impossible to decide whether chain has converged, can only monitor
behavior.
Easiest approach: graphical displays (trace plots in WinBUGS).
p
of Gelman
A bit more formal (and most popular in terms of use): (R)
and Rubin (1992).
To use the G-R diagnostic, must generate multiple independent chains
for each parameter.
The G-R diagnostic compares the within-chain sd to the between-chain
sd.
54
Before convergence, the within-chain sd is an underestimate of the

sd(|y), and between-chain sd is overestimate.
If chains converge, all draws are from stationary distribution, so within
and between chain variances should be similar.
Consider scalar parameter and let ij be jth draw in ith chain, with
i = 1, ..., n and j = 1, ..., J
Between-chain variance:
n X
1X
2
(
j ..) , j =
ij ,
B=
J 1
n
.. =
1
j
J
Within-chain variance:
X
1
W =
(ij j )2.
J(n 1)
55
An unbiased estimate of var(|y) is weighted average:

1
n1
W+ B
var(|y)
=
n
n
Early in iterations, var(|y)
overestimates true posterior variance

Diagnostic measures potential reduction in the scale if iterations are
continued:
p
p var(|y)
)
(R) = (
W
which goes to 1 as n .
56
Hierarchical models - Introduction

Consider the following study:
J counties are sampled from the population of 99 counties in the
state of Iowa.
Within each county, nj farms are selected to participate in a fertilizer
trial
Outcome yij corresponds to ith farm in jth county, and is modeled
as p(yij |j ).
We are interested in estimating the j .
Two obvious approaches:
1. Conduct J separate analyses and obtain p(j |yj ): all s are
independent
2. Get one posterior p(|y1, y2, ..., yJ ): point estimate is same for all
s.
Neither approach appealing.

Hierarchical approach: view the s as a sample from a common
population distribution indexed by a parameter .
Observed data yij can be used to draw inferences about s even though
the j are not observed.
A simple hierarchical model:
p(y, , ) = p(y|)p(|)p().
p(y|) =
Usual sampling distribution
p(|) =
Population dist. for or prior
p() =
Hyperprior
Hierarchical models - Rat tumor example

Imagine a single toxicity experiment performed on rats.
is probability that a rat receiving no treatment develops a tumor.
Data: n = 14, and y = 4 develop a tumor.
From the sample: tumor rate is 4/14 = 0.286.
Simple analysis:
y|
|,
Bin(n, )
Beta(, )
3
Posterior for is
p(|y) = Beta( + 4, + 10)
Where can we get good guesses for and ?
One possibility: from the literature, look at data from similar
experiments. In this example, information from 70 other experiments
is available.
In jth study, yj is the number of rats with tumors and nj is the sample
size, j = 1, .., 70.
See Table 5.1 and Figure 5.1 in Gelman et al..
Model the yj as independent binomial data given the study-specific j
and sample size nj
From Gelman et al.: hierarchical representation of tumor experiments
Representation of hierarchical model in Figure 5.1 assumes that the

Beta(, ) is a good population distribution for the j .
Empirical Bayes as a first step to estimating mean tumor rate in
experiment number 71:
1. Get point estimates of , from
earlier 70 experiments as follows:
P
70
2. Mean tumor rate is r = (70)1 j=1 yj /nj and standard deviation is
P
1
[(69)
)2]1/2. Values are 0.136 and 0.103 respectively.
j (yj /nj r
Using method of moments:
= 0.136, + = /0.136
+
2
=
0.103
( + )2( + + 1)
Resulting estimate for (, ) is (1.4, 8.6).
Then posterior is
p(|y) = Beta(5.4, 18.6)
with posterior mean 0.223, lower than the sample mean, and posterior
standard deviation 0.083.
Posterior point estimate is lower than crude estimate; indicates that in
current experiment, number of tumors was unusually high.
Why not go back now and use the prior to obtain better estimates for
tumor rates in earlier 70 experiments?
Should not do that because:
1. Cant use the data twice: we use the historical data to get the
prior, and cannot now combine that prior with the data from the
experiments for inference
2. Using a point estimate for (, ) suggests that there is no uncertainty
about (, ): not true
3. If the prior Beta(, ) is the appropriate prior, shouldnt we know

the values of the parameters prior to observing any data?
Approach: place a hyperprior on the tumor rates j with parameters
(, ).
Can still use all the data to estimate the hyperparameters.
Idea: Bayesian analysis on the joint distribution of all parameters
(1, ..., 71, , |y)
Hierarchical models - Exchangeability

J experiments. In experiment j, yj p(yj |j ).
To create joint probability model, use idea of exchangeability
Recall: a set of random variables (z1, ..., zK ) are exchangeable if their
joint distribution p(z1, ..., zK ) is invariant to permutations of their
labels.
In our set-up, we have two opportunities for assuming exchangeability:
1. At the level of data: conditional on j , the yij s are exchangeable if
we cannot distinguish between them
2. At the level of the parameters: unless something (other than the
data) distinguishes the s, we assume that they are exchangeable as
well.
In rat tumor example, we have no information to distinguish between

the 71 experiments, except sample size. Since nj is presumably not
related to j , we choose an exchangeable model for the j .
Simplest exchangeable distribution for the j :
p(|) = Jj=1p(j |).
The hyperparameter is typically unknown, so marginal for must be
obtained by averaging:
Z
p() =
Jj=1p(j |)p()d.
Exchangeable distribution for (1, ..., J ) is written in the form of a

mixture distribution
Mixture model characterizes parameters (1, ..., J ) as independent

draws from a superpopulation that is determined by unknown
parameters
Exchangeability does not hold when we have additional information on
covariates xj to distinguish between the j .
Can still model exchangeability with covariates:
Z
p(1, ..., J |x1, ..., xJ ) = Jj=1p(j |xj , )p(|x)d,
with x = (x1, ..., xJ ). In Iowa example, we might know that different
counties have very different soil quality.
Exchangeability does not imply that j are all the same; just that
they can be assumed to be draws from some common superpopulation
distribution.
10
Hierarchical models - Bayesian treatment

Since is unknown, posterior is now p(, |y).
Model formulation:
p(, ) = p()p(|) = joint prior
p(, |y) p(, )p(y|, ) = p(, )p(y|)
The hyperparameter gets its own prior distribution.
Important to check whether posterior is proper when using improper
priors in hierarchical models!
11
Posterior predictive distributions

There are two posterior predictive distributions potentially of interest:
1. Distribution of future observations y corresponding to an existing j .
Draw y from the posterior predictive given existing draws for j .
2. Distribution of new observations y corresponding to new j s drawn
from the same superpopulation. First draw from its posterior, then
draw for a new experiment, and then draw y from the posterior
predictive given the simulated .

In rat tumor example:
1. More rats from experiment #71, for example
2. Experiment #72 and then rats from experiment #72
12
Hierarchical models - Computation

Harder than before because we have more parameters
Easiest when population distribution p(|) is conjugate to the
likelihood p(|y).
In non-conjugate models, must use more advanced computation.
Usual steps:
1. Write p(, |y) p()p(|)p(y|)
2. Analytically determine p(|y, ). Easy for conjugate models.
3. Derive the marginal p(|y) by
Z
p(|y) =
p(, |y)d,
13
or, if convenient, by p(|y) =
p(,|y)
p(|y,) .
Normalizing constant in denominator may depend on as well as y.

Steps for computation in hierarchical models are:
1. Draw vector from p(|y). If is low-dimensional, can use inverse
cdf method as before. Else, need more advanced methods
2. Draw from conditional p(|, y). If the j are conditionally
independent, p(|, y) factors as
p(|, y) = j p(j |, y)
so components of can be drawn one at a time.
3. Draw y from appropriate posterior predictive distribution.
14
Rat tumor example

Sampling distribution for data from experiments j = 1, ..., 71:
yj Bin(nj , j )
Tumor rates j assumed to be independent draws from Beta:
j Beta(, )
Choose a non-informative prior for (, ) to indicate prior ignorance.
Since hyperprior will be non-informative, and perhaps improper, must
check integrability of posterior.
15
Defer choice of p(, ) until a bit later.

p(, , |y)
p(, ) j
p(, )p(|, )p(y|, , )

( + ) 1
y
j (1 j )1j j j (1 j )nj yj .
()()
Conditional of : notice that given (, ), the j are independent, with

Beta distributions:
( + + nj )
+y 1
p(j |, , y) =
j j (1 j )+nj yj 1.
( + yj )( + nj yj )
16
Marginal posterior distribution of hyperparameters is obtained using

p(, |y) =
p(, , |y)
p(|y, , )
Substituting into expression above, we get

p(, |y) p(, )j
( + ) ( + yj )( + nj yj )
()()
( + + nj )
Not a standard distribution, but easy to evaluate and only in two

dimensions.
Can evaluate p(, |y) over a grid of values of (, ), and then use
inverse cdf method to draw from marginal and from conditional.
17
But what is p(, )? As it turns out, many of the obvious choices for
the prior lead to improper posterior. (See solution to Exercise 5.7, on
the course web site.)
Some obvious choices such as p(, |y) 1 lead to non-integrable
posterior.
Integrability can be checked analytically by evaluating the behavior of
the posterior as , (or functions) go to .
An empirical assessment can be made by looking at the contour plot
of p(, |y) over a grid. Significant mass extending towards infinity
suggests that the posterior will not integrate to a constant.
Other choices leading to non-integrability of p(, |y) include:
18
A flat prior on prior guess and degrees of freedom
p(
, + ) 1.
+
A flat prior on the log(mean) and log(degrees of freedom)
p(log(/), log( + )) 1.
A reasonable choice for the prior is a flat distribution on the prior mean
and square root of inverse of the degrees of freedom:
p(
, ( + )1/2) 1.
+
This is equivalent to
p(, ) ( + )5/2.
19
Also equivalent to
p(log(/), log( + ) ( + )5/2.
We use the parameterization (log(/), log( + )).
Idea for drawing values of and from their posterior is the usual:
evaluate p(log(/), log( + )|y) over grid of values of (, ), and
then use inverse cdf method.
For this problem, easier to evaluate the log posterior and then
exponentiate.
Grid: From earlier estimates, potential centers for the grid are =
1.4, = 8.6. In the new parameterization, this translates into u =
log(/) = 1.8, v = log( + ) = 2.3.
20
Grid too narrow leaves mass outside. Try [2.3, 1.3] [1, 5].
Steps for computation:
1. Given draws (u, v), transform back to (, ).
2. Finally, sample j from Beta( + yj , + nj yj )
21
From Gelman et al.:
Contours of joint posterior distribution of alpha and beta in reparametrized scale
Posterior means and 95% credible sets for i
Normal hierarchical model

What we know as one-way random effects models
Set-up: J independent experiments, in each want to estimate a mean
j .
Sampling model:
yij |j , 2 N(j , 2)
for i = 1, ..., nj , and j = 1, ..., J.
Assume for now that 2 is known.
If y.j =
1 P
nj
i yij ,
then
y.j |j N(j , j2)
22
with j2 = 2/nj .
Sampling model for y.j is quite general. For nj large, y.j is normal
even if yij is not.
Need now to think of priors for the j .
What type of posterior estimates for j might be reasonable?
1.
2.
j = y.j : reasonable if nj large

P 2 1 P 2
j = y.. = ( j )
.j : pooled estimate, reasonable if
j j y
we believe that all means are the same.
To decide between those two choices, do an F-test for differences

between groups.
ANOVA approach (for nj = n):
23
Source
Between groups
Within groups
df
J-1
J(n-1)
SS
SSB
SSE
MS
SSB / (J-1)
SSE / J(n-1)
E(MS)
n 2 + 2
2
where 2, the variance of the group means can be estimated as

2 =
M SB M SE
n
If M SB >>> M SE then 2 > 0 and F-statistic is significant: do not

pool and use j = y.j .
Else, F-test cannot reject H0 : 2 = 0, and must pool.
Alternative: why not a more general estimator:
j = j y.j + (1 j )
y..
24
for j [0, 1]. Factor (1 j ) is a shrinkage factor.

All three estimates have a Bayesian justification:
1. j = y.j is posterior mean of j if sample means are normal and
p(j ) 1
2. j = y.. is posterior mean if 1 = ... = J and p() 1.
3. j = j y.j + (1 j )
y.. is posterior mean if j N(, 2)
independent of other s and sampling distribution for y.j is normal
Latter is called the Normal-Normal model.
25
Normal-normal model
Set-up (for 2 known):
y.j |j
1, ..., J |,
N(j , j2)
Y
N(, 2)
p(, 2)
It follows that
p(1, ..., J ) =
Z Y
N(j ; , 2)p(, 2)dd 2
The joint prior can be written as p(, ) = p(| )p( ). For , we will
26
consider a contidional flat prior, so that

p(, ) p( )
Joint posterior
p(, , |y) p(, )p(|, )p(y|)
Y
Y
p(, )
N(j ; , )
N(
y.j ; j , j2)
j
Conditional distribution of group means

For p(| ) 1, the conditional distributions of the j given , , y.j
are independent, and
p(j |, , y.j ) = N(j , Vj )
27
with
j2 + 2y.j
j =
,
2
2
j +
Vj1
1
1
= 2+ 2
j
Marginal distribution of hyperparameters

To get p(, |y) we would typically either integrate p(, , |y) with
respect to 1, ..., J , or would use the algebraic approach
p(, |y) =
p(, , |y)
p(|, , y)
In the normal-normal model, data provide information about , , as

shown below.
Consider the marginal posterior:
p(, |y) p(, )p(y|, )
28
where p(y|, ) is the marginal likelihood:

Z
p(y|, ) =
=
p(y|, , )p(|, )d
Z Y
j
N(
y.j ; j , j2)
N(j ; , )d1, ..., dJ
Integrand above is the product of quadratic functions in y.j and j , so

they are jointly normal.
Then the y.j |, are also normal, with mean and variance:
E(
y.j |, ) = E[E(
y.j |j , , )|, ]
= E(j |, ) =
var(
y.j |, ) = E[var(
y.j |j , , )|, ]
29
+var[E(
y.j |j , , )|, ]
= E(j2|, ) + var(j |, )
= j2 + 2
Therefore
p(, |y) p( )
N(
y.j ; , j2 + 2)
We know that
p(, |y) p( )p(| )
"
Y
(j2 + 2)1/2 exp

j
1
2
(
y
)
.j
2(j2 + 2)
Now fix , and think of p(

y.j |) as the likelihood in a problem in
30
which is the only parameter.

Recall that p(| ) 1.
From earlier, we know that p(|y, ) = N(
, V) where
P
= P
y.j /(j2 + 2)
2
2
j 1/(j + )
V =
X
j
1
j2 + 2
To get p( |y) we use the old trick:

p( |y) =
p(, |y)
p(|, y)
Q
p( ) j N(
y.j ; , j2 + 2)
N(;
, V )
31
Expression holds for any , so set =

. Denominator is then just
1/2
V
.
Then
p( |y)
p( )V1/2
N(
y.j ;
, j2 + 2).
What do we use for p( )?

Safe choice is always a proper prior. For example, consider
p( ) = Inv 2(0, 02),
where
02 can be best guess for
0 can be small to reflect prior uncertainty.
32
Non-informative (and improper) choice:

Beware. Improper prior in hierarchical model can easily lead to
non-integrable posterior.
In normal-normal model, natural non-informative p( ) 1 or
equivalently p(log( )) 1 results in improper posterior.
p( ) 1 leads to proper posterior p( |y).
33
Normal-normal model - Computation

1. Evaluate the one-dimensional p( |y) on a grid.
2. Sample from p( |y) using inverse cdf method.
3. Sample from N(
, V)
4. Sample j from N(j , Vj )
34
Normal-normal model - Prediction

Predicting future data y from the current experiments with means
= (1, ..., J ):
1. Obtain draws of (, , 1, ..., J )
2. Draw y from N(j , j2)
Predicting future data y from a future experiment with mean and
sample size n
:
1. Draw , from their posterior
2. Draw from the population distribution p(|, ) (also known as
the prior for )
3. Draw y from N(,

2), where
2 = 2/
n
35
Example: effect of diet on coagulation times

Normal hierarchical model for coagulation times of 24 animals
randomized to four diets. Model:
Observations: yij with i = 1, ..., nj , j = 1, ..., J.
Given , coagulation times yij are exchangeable
Treatment means are normal N (, 2), and variance 2 is constant
across treatments
Prior for (, log , log ) . A uniform prior on log leads to
improper posterior.
p(, , log , log |y) j N(j |, 2)
j i N(yij |j , 2)
P
1 P
1
Crude initial estimates: j = nj

yij = y.j ,
=J
y.j = y..,
36
P 2
P
2
2
1
j , and 2 = (J
(yij y.j ) ,
= J
j2 = (nj 1)1
P
1
1)
(
y.j y..)2.
Data
The following table contains data that represents coagulation time
in seconds for blood drawn from 24 animals randomly allocated to four
different diets. Different treatments have different numbers of observations
because the randomization was unrestricted.
Diet
Measurements
62,60,63,59
63,67,71,64,65,66
68,66,71,67,68,68
56,62,60,61,63,64,63,59
37
Conditional maximization for joint mode

Conjugacy makes conditional maximization easy.
Conditional modes of treatment means are
j |, , , y N(j , Vj )
with
2
2
/
+
n
y
/
j
.j
j =
,
2
2
1/ + nj /
and Vj1 = 1/ 2 + nj / 2.
For j = 1, ..., J, maximize conditional posteriors by using j in place of
current estimate j .
38
Conditional mode of :
2
|, , , y N(
, /J),
=J
j .
Conditional maximization: replace current estimate with

.
39
Conditional maximization
Conditional mode of log : first derive conditional posterior for 2:
2|, , , y Inv 2(n,
2),
PP
with
2 = n1
(yij j )2.
The mode of the Inv 2 is
2n/(n + 2). To get mode for log ,
use transformation. Term n/(n + 2) disappears with Jacobian, so
conditional mode of log is log
.
Conditional mode of log : same reasoning. Note that
2|, , 2, y Inv 2(J 1, 2),
P
2
1
with = (J 1)
(j )2. After accounting for Jacobian of
transformation, conditional mode of log is log .
40
Conditional maximization
Starting from crude estimates, conditional maximization required only
three iterations to converge approximately (see table)
Log posterior increased at each step
Values in final iteration is approximate joint mode.
When J is large relative to the nj , joint mode may not provide good
summary of posterior. Try to get marginal modes.
In this problem, factor:
p(, , log , log |y) = p(|, log , log , y)p(, log , log |y).
Marginal of , log , log is three-dimensional regardless of J and nj .
41
Conditional maximization results

Table shows iterations for conditional maximization.
Parameter
1
2
3
4
log p(params.|y)
Crude
estimate
61.000
66.000
68.000
61.000
64.000
2.291
3.559
-61.604
First
iteration
61.282
65.871
67.742
61.148
64.010
2.160
3.318
-61.420
Stepwise ascent
Second
Third
iteration iteration
61.288
61.290
65.869
65.868
67.737
67.736
61.152
61.152
64.011
64.011
2.160
2.160
3.318
3.312
-61.420
-61.420
Fourth
iteration
61.290
65.868
67.736
61.152
64.011
2.160
3.312
-61.420
42
Marginal maximization
Recall algebraic trick:
p(, log , log |y) =
p(, , log , log |y)

p(|, log , log , y)
j N(j |, 2)j i N(yij |j , 2)
j N(j |j , Vj )
Using in place of :
1/2
p(, log , log |y) j N(j |, 2)j iN(yij |j , 2)j Vj

with and Vj as earlier.
43
Can maximize p(, log , log |y) using EM. Here, (j ) are the missing
data.
Steps in EM:
1. Average over missing data in E-step
2. Maximize over (, log , log ) in M-step
44
EM algorithm for marginal mode

Log joint posterior:
log p(, , log , log |y) n log (J 1) log
1 X
1 XX
2
2
2
(j )
(y
)
ij
j
2 j
2 2 j i
E-step: Average over using conditioning trick (and p(| rest). Need
two expectations:
Eold[(j )2] = E[(j )2|old, old, old, y]
= [Eold(j )]2 + varold(j )
= (j )2 + Vj
Similarly:
Eold[(yij j )2] = (yij j )2 + Vj .
45
EM algorithm for marginal mode (contd)

M-step: Maximize the expected log posterior (expectations just taken
in E-step) with respect to (, log , log ). Differentiate and equate to
zero to get maximizing values (new , log new , log new ). Expressions
are:
1X
new
=
j .
J j
1/2
XX
1
new
=
[(yij j )2 + Vj ]
n j i
and
new
1/2
X
1
[(j new )2 + Vj ]
=
J 1 j
46
EM algorithm for marginal mode (contd)

Beginning from joint mode, algorithm converged in three iterations.
Important to check that log posterior increases at each step. Else,
programming or formulation error!
Results:
Parameter
Value at
joint mode
64.01
2.17
3.31
First
iteration
64.01
2.33
3.46
Second
iteration
64.01
2.36
3.47
Third
iteration
64.01
2.36
3.47
47
Using EM results in simulation

In a problem with standard form such as normal-normal, can do the
following:
1. Use EM to find marginal and conditional modes
2. Approximate posterior at the mode using N or t approximation
3. Draw values of parameters from papprox, and act as if they were
from p.
Importance re-sampling can be used to improve accuracy of draws. For
each draw, compute importance weights, and re-sample draws (without
replacement) using probability proportional to weight.
In diet and coagulation example, approximate approach as above likely
to produce quite reasonable results.
But must be careful, specially with scale parameters.
48
Gibbs sampling in normal-normal case

The Gibbs sampler can be easily implemented in the normal-normal
example because all posterior conditionals are of standard form.
Refer back to Gibbs sampler discussion, and see derivation of
conditionals in hierarchical normal example.
Full conditionals are:
j |all N (J of them)
|all N
2|all Inv 2
2|all Inv 2.
Starting values for the Gibbs sampler can be drawn from, e.g., a t4
approximation to the marginal and conditional mode.
49
Multiple parallel chains for each parameter permit monitoring

convergence using potential scale reduction statistic.
50
Results from Gibbs sampler

Summary of posterior distributions in coagulation example.
Posterior quantiles and estimated potential scale reductions computed
from the second halves of then Gibbs sampler sequences, each of length
1000.
Potential scale reductions for and were computed on the log scale.
The hierarchical variance 2, is estimated less precisely than the unitlevel variance, 2, as is typical in hierarchical models with a small
number of batches.
51
Posterior quantiles from Gibbs sampler

Burn-in = 50% of chain length.
Parameter
1
2
3
4
log p(, log , log |y)

log p(, , log , log |y)
2.5%
58.92
63.96
65.72
59.39
55.64
1.83
1.98
-70.79
-71.07
Posterior quantiles
25.0% 50.0% 75.0%
60.44 61.23 62.08
65.26 65.91 66.57
67.11 67.77 68.44
60.56 61.14 61.71
62.32 63.99 65.69
2.17
2.41
2.7
3.45
4.97
7.76
-66.87 -65.36 -64.20
-66.88 -65.25 -64.00
97.5%
63.69
67.94
69.75
62.89
73.00
3.47
24.60
-62.71
-62.42
52
Checking convergence
Parameter
1
2
3
4
log p(, log , log |y)

log p(, , log , log |y)
R
1.001
1.001
1.000
1.000
1.005
1.000
1.012
1.000
1.000
upper
1.003
1.003
1.002
1.001
1.005
1.001
1.013
1.000
1.001
53
62
58
theta1
66
Chains for diet means
500
600
700
800
900
1000
800
900
1000
800
900
1000
800
900
1000
65
63
theta2
67
iteration
500
600
700
69
67
65
theta3
iteration
500
600
700
63
61
59
theta3
65
iteration
500
600
700
iteration
54
Posteriors for diet means
56
58
60
62
64
66
62
64
theta1
64
66
68
theta3
66
68
70
theta2
70
72
58
60
62
64
theta4
55
80
50
60
70
mu
90 100
Chains for , ,
500
600
700
800
900
1000
800
900
1000
800
900
1000
10
5
sigma2
15
iteration
500
600
700
6000
0
2000
tau2
10000
iteration
500
600
700
iteration
56
Posteriors for , ,
50
55
60
65
70
75
80
mu
1.5
2.0
2.5
3.0
3.5
4.0
4.5
sigma
10
15
20
25
tau
57
Hierarchical modeling for meta-analysis

Idea: summarize and integrate the results of research studies in a
specific area.
Example 5.6 in Gelman et al.: 22 clinical trials conducted to study the
effect of beta-blockers on reducing mortality after cardiac infarction.
Considering studies separately, no obvious effect of beta-blockers.
Data are 22 22 tables: in jth study, n0j and n1j are numbers of
individuals assigned to control and treatment groups, respectively, and
y0j and y1j are number of deaths in each group.
Sampling model for jth experiment: two independent Binomials, with
probabilities of death p0j and p1j .
58
Possible quantities of interest:

1. difference p1j p0j
2. probability ratio p1j /p0j
3. odds ratio
j = [p1j /(1 p1j )]/[p0j /(1 p0j )]
We parametrize in terms of the log odds ratios: j = log j because
the posterior is almost normal even for small samples.
59
Normal approximation to the likelihood

Consider estimating j by the empirical logits:

yj = log
y1j
n1j y1j

log
y0j
n0j y0j
Approximate sampling variance (e.g., using a Taylor expansion):

j2 =
1
1
1
1
+
+
+
y1j n1j y1j y0j n0j y0j
See Table 5.4 for the estimated log-odds ratios and their estimated
standard errors.
60
Raw data
Study
j
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
Control
deaths
total
3
39
14
116
11
93
127
1520
27
365
6
52
152
939
48
471
37
282
188
1921
52
583
47
266
16
293
45
883
31
147
38
213
12
122
6
154
3
134
40
218
43
364
39
674
Treated
deaths
total
3
38
7
114
5
69
102
1533
28
355
4
59
98
945
60
632
25
278
138
1916
64
873
45
263
9
291
57
858
25
154
33
207
28
251
8
151
6
174
32
209
27
391
22
680
Logodds,
yj
0.0282
-0.7410
-0.5406
-0.2461
0.0695
-0.5842
-0.5124
-0.0786
-0.4242
-0.3348
-0.2134
-0.0389
-0.5933
0.2815
-0.3213
-0.1353
0.1406
0.3220
0.4444
-0.2175
-0.5911
-0.6081
sd,
j
0.8503
0.4832
0.5646
0.1382
0.2807
0.6757
0.1387
0.2040
0.2740
0.1171
0.1949
0.2295
0.4252
0.2054
0.2977
0.2609
0.3642
0.5526
0.7166
0.2598
0.2572
0.2724
61
Goals of analysis
Goal 1: if studies can be assumed to be exchangeable, we wish to
estimate the mean of the distribution of effect sizes, or overall average
effect.
Goal 2: the average effect size in each of the exchangeable studies
Goal 3: the effect size that could be expected if a new, exchangeable
study, were to be conducted.
62
Normal-normal model for

meta-analysis
First level: sampling distribution yj |j , j2 N(j , j2)
Second level: population distribution j |, N(, 2)
Third level: prior for hyperparameters , p(, ) = p(| )p( ) 1
mostly for convenience. Can incorporate information if available.
63
Study
j
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
2.5%
-0.58
-0.63
-0.64
-0.44
-0.44
-0.62
-0.62
-0.43
-0.56
-0.48
-0.47
-0.42
-0.65
-0.34
-0.54
-0.50
-0.46
-0.53
-0.51
-0.51
-0.67
-0.69
Posterior quantiles of effect j

normal approx. (on log-odds scale)
25.0%
50.0%
75.0%
-0.32
-0.24
-0.15
-0.36
-0.28
-0.20
-0.34
-0.26
-0.18
-0.31
-0.25
-0.18
-0.28
-0.21
-0.12
-0.36
-0.27
-0.19
-0.44
-0.36
-0.27
-0.28
-0.20
-0.13
-0.36
-0.27
-0.20
-0.35
-0.29
-0.23
-0.31
-0.24
-0.17
-0.28
-0.21
-0.12
-0.37
-0.27
-0.20
-0.22
-0.12
0.00
-0.32
-0.26
-0.17
-0.30
-0.23
-0.15
-0.28
-0.21
-0.11
-0.30
-0.22
-0.13
-0.31
-0.22
-0.13
-0.32
-0.24
-0.17
-0.40
-0.30
-0.23
-0.40
-0.30
-0.22
97.5%
0.13
-0.03
0.06
-0.04
0.13
0.04
-0.16
0.08
-0.05
-0.12
-0.01
0.09
0.02
0.30
0.00
0.06
0.15
0.15
0.17
0.04
-0.09
-0.07
64
Estimand
Mean,
Standard deviation,
Predicted effect, j
2.5%
-0.38
0.01
-0.57
Posterior quantiles
25.0%
50.0%
75.0%
-0.29
-0.25
-0.21
0.08
0.13
0.18
-0.33
-0.25
-0.17
97.5%
-0.12
0.32
0.08
0.0
0.02
0.04
Marginal posterior density p( |y)
0.0
0.5
1.0
1.5
2.0
tau
65
0.4
Conditional posterior means E(j |, y)
19
0.2
18
14
0.0
-0.4
-0.2
1
12
8
16
20
11
4
10
15
9
13
22
-0.6
Estimated Treatment Effects
17
7
3
6
21
2
0.0
0.5
1.0
1.5
2.0
tau
66
0.8
Conditional posterior standard deviations sd(j |, y)
0.6
3
18
2
0.4
13
17
15
5
9 16
22 20
21
12
8 14
11
0.2
Estimated standard deviations
19
6
10
0.0
0.5
1.0
1.5
4 7
2.0
tau
67
Histogram of 1000 simulations of 2, 18, 7, and 14
-1.0
-0.6
-0.2
0.2
-0.8
Study 2
-0.8
-0.4
Study 7
-0.4
0.0
0.4
Study 18
0.0
-0.6
-0.2
0.2
0.6
Study 14
68
-0.24
-0.26
-0.25
E(mu|tau,y)
-0.23
Overall mean effect, E(|, y)
0.0
0.5
1.0
1.5
2.0
tau
69
Example: Mixed model analysis

Sheffield Food Company produces dairy products.
Government cited company because the actual content of fat in yogurt
sold by Sheffield appers to be higher than the label amount.
Sheffield believes that the discrepancy is due to the method employed
by the goverment to measure fat content, and conducted a multilaboratory study to investigate.
Four laboratories in the United States were randomly chosen.
Each laboratory received 12 carefully mixed samples of yogurt, with
instructions to analyze 6 using the government method and 6 using
Sheffields method.
KSU - April 2005
Fat content in all samples was known to be very close to 3%.

Because of technical difficulties, none of the labs managed to analyze
all six samples using the government method within the alloted time.
KSU - April 2005
Method
Government
Lab 1
5.19
5.09
Sheffield
3.26
3.38
3.24
3.41
3.35
3.04
Lab 2
4.09
3.0
3.75
4.04
4.06
3.02
3.32
2.83
2.96
3.23
3.07
Lab 3
4.62
4.32
4.35
4.59
Lab 4
3.71
3.86
3.79
3.63
3.08
2.95
2.98
2.74
3.07
2.70
2.98
2.89
2.75
3.04
2.88
3.20
2
We fitted a mixed linear model to the data, where

Method is a fixed effect with two levels
Laboratory is a random effect with four levels
Method by laboratory interaction is random with six levels
yijk = i + j + ()ij + eijk ,
and i = + i and eijk N(0, 2).
Priors:
p(i) 1
p(j |2 |2 )
2
p(()ij |
)
N(0, 2 )
2
N(0,
)
The three variance components were assigned diffuse inverted gamma

priors.
KSU - April 2005
alpha[1] chains 1:3

5.5
5.0
4.5
4.0
3.5
3.0
5001
6000
8000
10000
iteration
alpha[2] chains 1:3

4.5
4.0
3.5
3.0
2.5
2.0
5001
8000
6000
iteration
10000
beta[1] chains 1:3 sample: 15000

4.0
3.0
2.0
1.0
0.0
3.0
2.0
1.0
0.0
-1.0
0.0
-1.0
1.0
-0.5
0.0
0.5
4.0
3.0
2.0
1.0
0.0
4.0
3.0
2.0
1.0
0.0
-2.0
-1.0
0.0
1.0
-2.0
-1.0
0.0
beta[3] chains 1:3
beta[1] chains 1:3
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
0
20
40
20
40
lag
lag
beta[2] chains 1:3
beta[4] chains 1:3
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
40
20
lag
40
20
lag
beta[1] chains 1:3
beta[2] chains 1:3
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
40
20
20
lag
40
lag
beta[4] chains 1:3
beta[3] chains 1:3
1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
40
20
lag
20
40
lag
beta[1] chains 1:3
beta[2] chains 1:3
1.5
1.5
1.0
1.0
0.5
0.5
0.0
0.0
5001
8000
6000
5001
8000
6000
iteration
iteration
beta[3] chains 1:3

beta[4] chains 1:3
1.5
1.0
1.0
0.5
0.5
0.0
0.0
5001
6000
8000
iteration
5001
6000
8000
iteration
box plot: alpha
5.0
box plot: beta
1.0
[1]
[1]
[3]
0.5
[2]
4.0
[4]
[2]
0.0
3.0
-0.5
2.0
-1.0
Model checking
What do we need to check?
Model fit: does the model fit the data?

Sensitivity to prior and other assumptions
Model selection: is this the best model?
Robustness: do conclusions change if we change data?
Remember: models are never true; they may just fit the data well and
allow for useful inference.
Model checking strategies must address various parts of models:
priors
sampling distribution
hierarchical structure
other model characteristics such as covariates, form of dependence

between response variable and covariates, etc.
Classical approaches to model checking:
Do parameter estimates make sense

Does the model generate data like observed sample
Are predictions reasonable
Is model best in some sense (e.g., AIC, likelihood ratio, etc.)
Bayesian approach to model checking

Does the posterior distribution of parameters correspond with what
we know from subject-matter knowledge?
Predictive distribution for future data must also be consistent with
substantive knowledge
Future data generated by predictive distribution compared to current
sample
Sensitivity to prior and other model components

Most popular model checking approach has a frequentist flavor: we
generate replications of the sample from posterior and observe the
behavior of sample summaries over repeated sampling.
Posterior predictive model checking

Basic idea is that data generated from the model must look like
observed data.
Posterior predictive model checks based on replicated data y rep
generated from the posterior predictive distribution:
Z
p(y rep|y) = p(y rep|)p(|y)d.
y rep is not the same as y. Predictive outcomes y can be anything
(e.g., a regression prediction using different covariates x
). But y rep is
a replication of y.
y rep are data that could be observed if we repeated the exact same
experiment again tomorrow (if in fact the s in our analysis gave rise
to the data y).
Definition of a replicate in a hierarchical model:

p(|y) p(|, y) p(y rep|) : rep from same units
p(|y) p(|) p(y rep|) : rep from new units
Test quantity: T (y, ) is a discrepancy statistic used as a standard.
We use T (y, ) to determine discrepancy between model and data on
some specific aspect we wish to check.
Example of T (y, ): proportion of standardized residuals outside of
(3, 3) in a regression model to check for outliers
In classical statistics, use T (y) a test-statistic that depends only on the
data. Special case of Bayesian T (y, ).
For model checking:
Determine appropriate T (y, )

Compare posterior predictive distribution of T (y rep, ) to posterior
distribution of T (y, )
Bayes p-values
p-values attempt to measure tail-area probabilities.
Classical definition:
class p value = Pr(T (y rep) T (y)|)
Probability is taken over distribution of y rep with fixed
Point estimate typically used to compute the pvalue.
Posterior predictive p-values:
Bayes p value = Pr(T (y rep, ) T (y, )|y)
Probability taken over joint posterior distribution of (, y rep):

Z Z
Bayes p value =
IT (yrep,)T (y,)p(|y)p(y rep|)ddy rep
Relation between p-values

Small example:
Sample y1, ..., yn N (, 2)
Fit N (0, 2) and check whether = 0 is good fit
In real life, would fit more general N (, 2) and would decide whether
= 0 is plausible.
Classical approach:
Test statistic is sample mean T (y) = y
p value = P r(T (y rep) T (y)| 2)
= P r(
y rep y| 2)

= Pr
n
y
S

= P
This is special case:
parameters.
tn1
rep
n
y
n
y 2
|
S

not always possible to get rid of nuisance
Bayes approach:
p value = P r(T (y rep) T (y)|y)
Z Z
=
IT (yrep)T (y)p(y rep| 2)p( 2|y)dy repd 2
Note that
IT (yrep)T (y)p(y rep| 2) = P (T (y rep) T (y)| 2)
10
= classical p-value
Then:
Bayes p value = E{p valueclass|y}
where the expectation is taken with respect to p( 2|y).
In this example, classical pvalue and Bayes pvalue are the same.
In general, Bayes can handle easily nuisance parameters.
11
Interpreting posterior predictive pvalues

We look for tail-area probabilities that are not too small or too large.
The ideal posterior predictive p-value is 0.5: the test quantity falls right
in the middle of its posterior predictive.
Posterior predictive p-values are actual posterior probabilities
Wrong interpretation: P r(model is true |data).
12
Example: Independence of Bernoulli trials

Sequence of binary outcomes y1, ..., yn modeled as iid Bernoulli trials
with probability of success .
s
ns
Uniform prior on leads to p(|y) (1 )
with s =
yi .
Sample:
1, 1, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0
Is the assumption of independence warranted?
Consider discrepancy statistic
T (y, ) = T (y) = number of switches between 0 and 1.
13
In sample, T (y) = 3.
Posterior is Beta(8,14).
To test assumption of independence, do:
1. For j = 1, ..., M draw j from Beta(8,14).
rep,j
rep,j
, ..., y20 } independent Bernoulli variables with
2. Draw {y1
probability j .
3. In each of the M replicate samples, compute T (y), the number of
switches between 0 and 1.
p = Prob(T (y rep) T (y)|y) = 0.98.
B
If observed trials were independent, expected number of switches in 20

trials is approximately 8 and 3 is very unlikely.
14
Posterior predictive dist. of number of switches
15
Example: Newcombs speed of light

Newcomb obtained 66 measurements of the speed of light. (Chapter
3). One measurement of -44 is a potential outlier under a normal
model.
From data: y = 26.21 and s = 10.75.
Model: yi|, 2 N (, 2), (, 2) 2.
Posteriors
p( 2|y) = Inv 2(65, s2)
p(| 2, y) = N(
y , 2/66)
or
p(|y) = t65(
y , s2/66).
16
For posterior predictive checks, do for i = 1, ..., M

1. Generate ((i), 2(i)) from p(, 2|y)
rep(i)
rep(i)
from N ((i), 2(i))
, ..., y66
2. Generate y1
17
Example: Newcombs data
18
Example: Replicated datasets
19
Is observed minimum value -44 consistent with model?

Define T (y) = min{yi}
get distribution of T (y rep).
Large negative value of -44 inconsistent with distribution of
T (y rep) = min{yirep} because observed T (y) very unlikely under
distribution of T (y rep).
Model inadequate, must account for long tail to the left.
20
Posterior predictive of smallest observation
21
Is sampling variance captured by model?

Define
T (y) =
s2y
1
= i(yi y)2
65
obtain distribution of T (y rep).

Observed variance of 115 very consistent with distribution (across
reps) of sample variances
Probability of observing a larger sample variance is 0.48: observed
variance in middle of distribution of T (y rep)
But this is meaningless: T (y) is sufficient statistic for 2, so model
MUST have fit it well.
Symmetry in center of distribution
Look at difference in the distance of 10th and 90th percentile from
the mean.
Approx 10th percentile is 6th order statistic y(6), and approx 90th
percentile is y(61).
22
Define T (y, ) = |y(61) | |y(6) |

Need joint distribution of T (y, ) and of T (y rep, ).
Model accounts for symmetry in center of distribution
Joint distribution of T (y, ) and T (y rep, ) about evenly distributed
along line T (y, ) = T (y rep, ).
Probability that T (y rep, ) > T (y, ) about 0.26, plausible under
sampling variability.
23
Posterior predictive of variance and symmetry
24
Choice of discrepancy measure

Often, choose more than one measure, to investigate different attributes
of the model.
Can smoking behavior in adolescents be predicted using information
on covariates? (Example on page 172 of Gelman et al.)
Data: six observations of smoking habits in 2,000 adolescents every
six months
Two models:
Logistic regression model
Latent-class model: propensity to smoke modeled as nonlinear
function of predictors. Conditional on smoking, fit same logistic
regression model as above.
Three different test statistics chosen:
25
% who never smoked

% who always smoked
% of incident smokers: began not smoking but then switched to
smoking and continued doing so.
Generate replicate datasets from both models.
Compute the three test statistics in each of the replicated datasets
Compare the posterior predictive distributions of each of the three
statistics with the value of the statistic in the observed dataset
Results: Model 2 slightly better than 1 at predicting % of always
smokers, but both models fail to predict % of incident smokers.
Test variable
% never smokers
% always-smokers
% incident smokers
T (y)
77.3
5.1
8.4
95% CI
for T (y rep )
[75.5, 78.2]
[5.0, 6.5]
[5.3, 7.9]
p-value
0.27
0.95
0.005
95% CI
for T (y rep )
[74.8, 79.9]
[3.8, 6.3]
[4.9, 7.8]
p-value
0.53
0.44
0.004
26
Omnibus tests
It is often useful to also consider summary statistics:
2
-discrepancy: T (y, ) =
P (yiE(yi|))2
var(yi |)
Deviance: T (y, ) = 2 log p(y|)

Deviance is proportional to mean squared error if model is normal and
with constant variance.
In classical approach, insert null or mle in place of in T (y, ) and
compare test statistic to reference distribution derived under asymptotic
arguments.
In Bayesian approach, reference distribution for test statistic is
automatically calculated from posterior predictive simulations.
27
A Bayesian 2 test is carried out as follows:

Compute the distribution of T (y, ) for many draws of from
posterior and for observed data.
Compute the distribution of T (y rep, ) for replicated datasets and
posterior draws of
Compute Bayesian p-value: probability of observing a more extreme
value of T (y rep, ).
Calculate p-value empirically from simulations over and y rep.
28
Criticisms of posterior predictive checks

Too conservative because data get used twice
Posterior predictive p-values are not really p-values: distribution under
null hypothesis is not uniform, as it should be
What is high/low if pp p-values not uniform?
Some Bayesians object to frequentist slant of using unobserved data
for anything
Lots of work recently on improving posterior predictive p-values: Bayarri
and Berger, JASA, 2000, is good reference
In spite of criticisms, pp checks are easy to use in very general cases,
and intuitively appealing.
29
Pps can be conservative

One criticism for pp model checks is that they tend to reject models
only in the face of extreme evidence.
Example (from Stern):
y N (, 1) and N (0, 9).

Observation: yobs = 10
Posterior p(|y) = N (0.9yobs, 0.9) = N (9, 0.9)
Posterior predictive dist: N (9, 1.9)
Posterior predictive p-value is 0.23, we would not reject the model
Effect of prior is minimized because 9 > 1 and posterior predictive

mean is close to observed datapoint.
To reject, we would need to observe y 23.
30
Graphical posterior predictive checks

Idea is to display the data together with simulated data.
Three kinds of checks:
Direct data display
Display of data summaries or parameter estimates
Graphs of residuals or other measures of discrepancy
31
Model comparison
Which model fits the data best?
Often, models are nested: a model with parameters is nested within
a model with parameters (, ).
Comparison involves deciding whether adding to the model improves
its fit. Improvement in fit may not justify additional complexity.
It is also possible to compare non-nested models.
We focus on predictive performance and on model posterior probabilities
to compare models.
32
Expected deviance
We compare the observed data to several models to see which predicts
more accurately.
We summarize model fit using the deviance
D(y, ) = 2 log p(y|).
It can be shown that the model with the lowest expected deviance is
best in the sense of minimizing the (Kullback-Leibler) distance between
the model p(y|) and the true distribution of y, f (y).
We compute the expected deviance by simulation:
Davg(y) = E|y (D(y, )|y).
33
An estimate is
1 X
Davg(y) =
D(y, j ).
M j
34
Deviance information criterion - DIC

Idea is to estimate the error that would be expected when applying the
model to future data.
Expected mean square predictive error:
#
"
X
1
rep
rep
pred
(yi E(yi |y))2 .
Davg = E
n i
A model that minimizes the expected predictive deviance is best in the
sense of out-of-sample predictive power.
pred
An approximation to Davg is the Deviance Information Criterion or
DIC
pred = 2D
avg(y) D (y),
DIC = D
avg
35
where D(y) = D(y, (y))

and (y)
is a point estimator of such as
the posterior mean.
36
Example: SAT study

We compare three models: no pooling, complete pooling and
hierarchical model (partial pooling).
Deviance is
D(y, ) = 2
log N(yj |j , j2)
j
2
= log(2J ) +
((yj j )2/j2)
The DIC for the three models were: 70.3 (no pooling), 61.5 (complete
pooling), 63.4 (hierarchical model).
Based on DIC, we would pick the complete pooling model.
37
We still prefer the hierarchical model because assumption that all

schools have identical mean is strong.
38
Bayes Factors
Suppose that we wish to decide between two models M1 and M2
(different prior, sampling distribution, parameters).
Priors on models are p(M1) and p(M2) = 1 p(M1).
The posterior odds favoring M1 over M2 are
p(M1|y) p(y|M1) p(M1)
=
.
p(M2|y) p(y|M2) p(M2)
The ratio p(y|M1)/p(y|M2) is called a Bayes factor.
It tells us how much the data support one model over the other.
39
The BF12 is computed as

BF12
R
p(y|1, M1)p(1|M1)d1
=R
.
p(y|2, M2)p(2|M2)d2
Note that we need p(y) under each model to compute p(i|Mi).

Therefore, the BF is only defined when the marginal distribution of the
data p(y) is proper, and therefore, when the prior distribution in each
model is proper.
For example, if y N (, 1) with p() 1, we get
Z
1
1/2
p(y) (2)
exp{ (y )2}d = 1,
2
which is constant for any y and thus is not a proper distribution. This
creates an indeterminacy because we can increase or decrease the BF
simply by using a different constant in the numerator and denominator.
40
Computation of Bayes Factors

Difficulty lies in the computation of p(y)
Z
p(y) =
p(y|)p()d.
The simplest approach is to draw many values of from p() and get
a Monte Carlo approximation to the integral.
May not work well because the s from the prior may not come from
the parameter space region where the sampling distribution has most
mass.
A better Monte Carlo approximation is similar to importance sampling.
41
Note that
1
p(y)
h()
d
p(y)
h()
p(|y)d.
p(y|)p()
=
=
h() can be anything (e.g., a normal approximation to the posterior).

Draw values of from the posterior and evaluate the integral
numerically.
Computations can get tricky because denominator can be small.
As n ,
1
log(BF ) log(p(y|2, M2) log(p(y|1, M1) (d1 d2) log(n),
2
42
where i is the posterior mode under model Mi and di is the dimension

of i.
Notice that the criterion penalizes a model for additional parameters.
Using log(BF ) is equivalent to ranking models using the BIC criterion,
given by
1
BIC = log(p(y|, M ) + d log(n).

2
43
Ordinary Linear Regression - Introduction

Question of interest: how does an outcome y vary as a function of a
vector of covariates X.
We want the conditional distribution p(y|, x), where observations
(y, x)i are assumed exchangeable.
Covariates (x1, x2, ..., xk ) can be discrete or continuous, and typically
x1 = 1 for all n units. The matrix X, n k is the model matrix
Most common version of the model is the normal linear model
E(yi|, X) =
k
X
j xij
j=1
ENAR - March 2006
Ordinary Linear Regression - Model

In ordinary linear regression models, the conditional variance is equal
across observations: var(yi|, X) = 2. Thus, = (1, 2, ..., k , 2).
Likelihood: For the ordinary normal linear regression model, we have
p(y|X, , 2) = N (X, 2I)
with In and n n identity matrix.
Priors: A non-informative prior distribution for (, 2) is
p(, 2) 2
(more on informative priors later)
ENAR - March 2006
Joint posterior:

1 2
2
2 n
1
2
p(, |y) ( )
exp (y X)0(y X)
2
Joint posterior
2 n
2 1
p(, 2|y) ( )

1 2
exp (y X)0(y X)
2
Consider
p(, 2|y) = p(| 2, y)p( 2|y)
Conditional posterior for :
ENAR - March 2006
Expand and complete the squares in

3
p(| 2, y) (viewed as function of ). We get:

h
i
1
2
p(| , y) exp 2 0X 0X 20X 0X
2

1
0X 0X( )
exp 2 ( )
2
for
= (X 0X)1X 0y
Then:
2(X 0X)1)
| 2, y N (,
ENAR - March 2006
Ordinary Linear Regression - p( 2|y)

The marginal posterior distribution for 2 is obtained by integrating
the joint posterior with respect to :
p( 2|y) =
p(, 2|y)d

Z
n
1
( 2) 2 1 exp 2 (y X)0(y X) d
2
By expanding square in integrand, and adding and subtracting

2y 0X(X 0X)1X 0y, we can write:
arge
2
p( |y) =
ENAR - March 2006
2 n
( ) 2 1

i
h
1
+ ( )
0X 0X( )
0(y X )
d
exp 2 (y X )
2
Z

n
1
1
0X 0X( )
d
( 2) 2 1 exp 2 (n k)S 2
exp 2 ( )
2
2
Z
Integrand is kernel of k dimensional normal, so result of integration

is proportional to ( 2)k/2. Then
2
p( |y)
nk
( 2) 2 +1 exp
2

2
(n k)S
proportional to an Inv-2(n k, S 2).
ENAR - March 2006
Ordinary Linear Regression - Notes

Note that and S 2 are the MLEs of and 2
When is the posterior proper? p(, 2|y) is proper if
1. n > k
2. rank(X) = k: columns of X are linearly independent or |X 0X| =
6 0
To sample from the joint posterior:
1. Draw 2 from Inv-2(n k, S 2)
2(X 0X)1)
2. Given 2, draw vector from N(,
S 2, (X 0X)1 once, before starting repeated
For efficiency, compute ,
drawing.
ENAR - March 2006
Regression - Posterior predictive distribution

In regression, we often want to predict the outcome y for a new set of
covariates x
. Thus, we wish to draw values from p(
y |y, X)
By simulation:
1. Draw 2 from Inv-2(n k, S 2)
2(X 0X)1)
2. Draw from N(,
3. Draw yi for i = 1, ..., m from N(
x0i, 2)
ENAR - March 2006
Regression - Posterior predictive distribution

We can derive p(
y |y) in steps, first considering p(
y |y, 2).
Note that
p(
y |y, 2) =
p(
y |, 2)p(| 2, y)d
is normal because exponential in integrand is quadratic function in

(, y).
To get mean and variance use conditioning trick:
2

2
2
E(
y | , y) = E E(
y |, , y)| , y
2
= E(X|
, y)

= X
ENAR - March 2006
where inner expectation averages over y conditional on and outer

averages over .
2

2
2
2
2
var(
y | , y) = E var(
y |, , y)| , y + var E(
y |, , y)| , y
2
= E[ 2I| 2, y] + var[X|
, y]
0X)1X
0)
= 2(I + X(X
0X)1X
0 is
Var has two terms: 2I is sampling variation, and 2X(X
due to uncertainty about
To complete specification of p(
y |y) must integrate p(
y |y, 2) with
respect to marginal posterior distribution of 2.
Result is:
S 2[I + X(X
,
0X)1X])
p(
y |y) = tnk (X
ENAR - March 2006
10
Regression example: radon measurements in

Minnesota
Radon measurements yi were taken in three counties in Minnesota:
Blue Earth, Clay and Goodhue.
14 houses in each of Blue Earth and Clay county and 13 houses in
Goodhue were sampled.
Measurements were taken in the basement and in the first floor.
We fit an ordinary regression model to the log radon measurements,
without an intercept.
We define dummy variables as follows: X1 = 1 if county is Blue Earth
KSU - April 2005
and is 0 otherwise. Similarly, X2 and X3 are dummies for Clay and

Goodhue counties, respectively.
X4 = 1 if measurement was taken in the first floor.
The model is
log(yi) = 1x1i + 2x2i + 3x3i + 4x4i + ei,
with e N(0, 2).
Thus
E(y|Blue Earth, basement) = exp(1)
E(y|Blue Earth, first floor) = exp(1 + 4)
E(y|Clay, basement) = exp(2)
KSU - April 2005
E(y|Clay, first floor) = exp(2 + 4)

E(y|Goodhue, basement) = exp(3)
E(y|Goodhue, first floor) = exp(3 + 4)
We used noninformative priors N(0, 1000) for the regression coefficients
and a noninformative prior Gamma(0.01, 0.01) for the error variance.
KSU - April 2005
beta[1]
3.0
2.5
2.0
1.5
1.0
0.5
3001
3500
4000
4500
5000
4500
5000
iteration
beta[2]
3.0
2.5
2.0
1.5
1.0
0.5
3001
3500
4000
iteration
beta[3]
3.0
2.5
2.0
1.5
1.0
3001
3500
4000
4500
5000
4500
5000
iteration
beta[4]
1.0
0.0
-1.0
-2.0
3001
3500
4000
iteration
box plot: beta
3.0
[1]
[2]
[3]
2.0
1.0
[4]
0.0
-1.0
becb sample: 2000
becff sample: 2000
0.3
0.3
0.2
0.2
0.1
0.1
0.0
0.0
0.0
5.0
10.0
15.0
0.0
ccb sample: 2000
5.0
10.0
15.0
ccff sample: 2000
0.4
0.3
0.2
0.1
0.0
0.4
0.3
0.2
0.1
0.0
0.0
10.0
5.0
0.0
gcb sample: 2000

0.3
0.2
0.2
0.1
0.1
0.0
0.0
5.0
10.0
10.0
gcff sample: 2000
0.3
0.0
5.0
15.0
0.0
5.0
10.0
15.0
Regression - Posterior predictive checks

For regression models, there are well-known methods for checking
model and assumptions using estimated residuals. Residuals:
i = yi x0i
Two useful test statistics are:
Proportion of outliers among residuals
Correlation between squared residuals and fitted values y
Posterior predictive distributions of both statistics can be obtained via
simulation
Define a standardized residual as
ei = (yi x0i)/
ENAR - March 2006
13
If normal model is correct, standardized residuals should be about

N (0, 1) and therefore, |e|i > 3 suggests that ith observation may be
an outlier.
To derive the posterior predictive distribution for the proportion of
outliers q and for , the correlation between (e2, y), do:
1.
2.
3.
4.
5.
6.
Draw ( 2, ) from the joint posterior distribution

Draw y rep from N (X, 2I) given the existing X
Run the regression of y rep on X, and save residuals
Compute
Compute proportion of large standardized residuals q
Repeat for another y rep
Approach is frequentist in nature: we act as if we could repeat the

experiment many times.
Inspection of posterior predictive distribution of and of q provides
ENAR - March 2006
14
information about model adequacy.
ENAR - March 2006
15
Results of posterior predictive checks suggest that there are no outliers

and that the correlation between the squared residuals and the fitted
values is negligible.
The 95% credible set for the proportion of absolute residuals (in the
41 observations) above 3 was (0, 4.88) with a mean of 0.59% and a
median of 0.
The 95% credible set for was (0.31, 0.32) with a mean of 0.011.
Notice that the posterior distribution of the proportion of outliers
is skewed. This sometimes happens when the quantity of interest
is bounded and most of the mass of its distribution is close to the
boundary.
KSU - April 2005
Regression with unequal variances

Consider now the case where y N (X, y ), with y 6= 2I.
Covariance matrix y is n n and has n(n + 1)/2 distinct parameters,
and cannot be estimated from n observations. Must either specify y
or it must be assigned an informative prior distribution.
Typically, some structure is imposed on y to reduce the number of
free parameters.
ENAR - March 2006
16
Regression - known y
As before, let p() 1 be the non-informative prior
Since y is positive definite and symmetric, it has an upper-triangular
1/2
1/2 1/20
square root matrix (Cholesky factor) y such that y y = y
so that if:
y = X + e, e N (0, y )
then
1/2
y = 1/2
X + 1/2
e,
y
y
y
1/2
e N (0, I)
y
With y known, just proceed as in ordinary linear regression, but use

the transformed y and X as above and fix 2 = 1. Algebraically, this
ENAR - March 2006
17
is equivalent to computing
1
1
= (X 01
y X) Xy y
1
= (X 01
X)
y
Note: By using the Cholesky factor you avoid computing the n n

inverse 1
y .
ENAR - March 2006
18
Prediction of new y with known y

Even if we know y , prediction of new observations is more complicated:
must know covariance matrix of old and new data.
Example: heights of children from same family are correlated. To
predict height of a new child when a brother is in the old dataset, we
must include that correlation in the prediction.
y are n
new observations given a n
k matrix of regressors X.
Joint distribution of y, y is:
y
N
|X, X,
y
y|y, , y
ENAR - March 2006

X
y yy
,
yy yy
X
N (y, Vy)
19
where
+ yy1
y = X
yy (y X)
Vy = yy yy1
yy yy
ENAR - March 2006
20
Regression - unknown y
To draw inferences about (, y ), we proceed in steps:
p(|y , y) and then p(y |y).
derive
Let p() 1 as before

We know that
p(y |y) =
p(, y |y) p(y )p(y|, y )
p(|y , y)
p(|y , y)
p(y ) N(y|, y )
,
N(|, V )
V ) depend on y .
where (,
ENAR - March 2006
21

Expression for p(y |y) must hold for any , so we set = .
Note that
V ) |V |1/2
p(|y , y) = N(,
Then
for = .
1
01(y X )]
p(y |y) p(y )|V |1/2|y |1/2 exp[ (y X )

y
2
In principle, p(y |y) could be evaluated for a range of values of y .
However:
1. It is difficult to determine a prior for an n n unstructured matrix.
V depend on y and it is very difficult to draw values of y from
2. ,
p(y |y).
Need to put some structure on y
ENAR - March 2006
22
Regression - y = 2Qy
Suppose that we know y upto a scalar factor 2.
Non-informative prior on , 2 is p(, 2) 2
Results follow directly from ordinary linear regression, by using
1/2
1/2
transformation Qy y and Qy X, which is equivalent to computing
1 0 1
= (X 0Q1
X)
X Qy y
y
1
= (X 0Q1
y X)
0Q1(y X )
s2 = (n k)1(y X )
y
To estimate the joint posterior distribution p(, 2|y) do
ENAR - March 2006
23
1. Draw 2 from Inv-2(n k, s2)

V 2 )
2. Draw from n(,
In large datasets, use Cholesky factor transformation and unweighted
regression to avoid computation of Q1
y .
ENAR - March 2006
24
Regression - other covariance structures

Weighted regression
In some applications, y = diag( 2/wi), for known weights and 2
unknown.
Inference is the same as before, but now Q1
y = diag(wi )
Parametric model for unequal variances:
Variances may depend on weights in non-linear fashion: ii =
2v(wi, ) for unknown parameter (0, 1) and known function
v such as v = wi
Note that for that function v:
= 0 ii = 2 i
ENAR - March 2006
25
= 1 ii = 2/wi i
In practice, to uncouple from the scale of the weights, we multiply
weights by a factor so that their product equals 1.
ENAR - March 2006
26
Parametric model for unequal variances

Three parameters to estimate: , 2, .
Possible prior for is uniform in [0, 1]
Non-informative prior for , 2 is 2
p(, 2, |y) p()p(, 2)ni=1 N(yi; (X)i, 2v(wi, ))
For a given , we are back in the earlier weighted regression case with
Q1
y = diag(v(w1 , ), ..., v(wn , ))
ENAR - March 2006
27
This suggests the following scheme

1.
2.
3.
4.
Draw from marginal posterior p(|y)

1/2
1/2
Compute Qy y and Qy X
Draw 2 from p( 2|, y) as in ordinary regression
Draw from p(| 2, , y) as in ordinary regression
Marginal posterior distribution of :

p(|y) =
p(, 2, |y)
p(, 2|, y)
p(, 2, |y)
p(| 2, , y)p( 2|, y)
p() 2i N(yi|(X)i, 2v(wi, ))

V )
Inv-2(n k, s2) N(,
Expression must hold for any (, 2), so we set = and 2 = s2.

ENAR - March 2006
28
Recall that and s2 depend on .

Also recall that weights are scaled to have product equal to 1.
Then
p(|y) p()|V |1/2(s2)(nk)/2
To carry out computation
First sample from p(|y):
1. Evaluate p(|y) for a range of values of [0, 1]
2. Use inverse cdf method to sample
1/2
1/2
Given , compute y = Qy y and X = Qy X
Compute
0 1 0
= (X X ) X y
V
ENAR - March 2006
= (X X )1
29
0(y X )
s2 = (n k)1(y X )
Draw 2 from Inv-2(n k, s2)
2 V )
Draw from N(,
ENAR - March 2006
30
Including prior information about

Suppose we wish to add prior information about a single regression
coefficient j of the form:
j N(j0, 2j ),
with j0, 2j known.
Prior information can be added in the form of an additional data
point.
An ordinary observation y is normal with mean x and variance 2.
As a function of j , the prior can be viewed as an observation with
0 on all xs except xj
ENAR - March 2006
31
variance 2j .
To include prior information, do the following:
1. Append one more data point to vector y with value j0.
2. Add one row to X with zeroes except in jth column.
3. Add a diagonal element with value 2j to y .
Now apply computational methods for non-informative prior.
Given y , posterior for obtained by weighted linear regression.
ENAR - March 2006
32
Adding prior information for several s

Suppose that for the entire vector ,
N(0, ).
Proceed as before: add k data points and draw posterior inference
by weighted linear regression applied to observations y, explanatory
variables X and variance matrix :

y =
y
0

,
X =
X
Ik

, =
y 0
0
Computation can be carried out conditional on first and then inverse

cdf for or using the Gibbs sampling.
ENAR - March 2006
33
Prior information about 2

Typically we do not wish to include prior information about 2.
If we do, we can use the conjugate prior
2 Inv-2(n0, 02).
The marginal posterior of 2 is
2
2
+
nS
n
0
0
).
2|y Inv-2(n0 + n,
n0 + n
If prior information on is also incorporated, S 2 is replaced by

corresponding value from regression of y on X and , and n is
replaced by length of y.
ENAR - March 2006
34
Inequality constraints on
Sometimes we wish to impose inequality constraints such as
1 > 0
or
2 < 3 < 4.
The easiest way is to ignore the constraint until the end:
Simulate (, 2) from posterior
discard all the draws that do not satisfy the constraint.
Typically a reasonably efficient way to proceed unless constraint
eliminates a large portion of unconstrained posterior distribution.
If so, data tend to contradict the model.
ENAR - March 2006
35
Generalized Linear Models

Generalized linear models are an extension of linear models to the case
where relationship between E(y|X) and X is not linear or normal
assumption is not appropriate.
Sometimes a transformation suffices to return to the linear setup.
Consider the multiplicative model
b2 b3
yi = xb1
i1 xi2 xi3 i
A simple log transformation leads to

log(yi) = b1 log(xi1) + b2 log(xi2) + b3 log(xi3) + ei
When simple approaches do not work, we use GLIMs.
There are three main components in the model:

1. Linear predictor = X
2. Link function g(.) relating linear predictor to mean of outcome
variable: E(y|X) = = g 1() = g 1(X)
3. Distribution of outcome variable y with mean = E(y|X).
Distribution can also depend on a dispersion parameter :
p(y|X, , ) = ni=1p(yi|(X)i, )
In standard GLIMs for Poisson and binomial data, = 1.
In many applications, however, excess dispersion is present.
Some standard GLIMs

Linear model:
Simplest GLIM, with identity link function g() = .
Poisson model:
Mean and variance and link function log() = X, so that
= exp(X) = exp()
For y = (y1, ..., yn):
p(y|) = ni=1
1
exp( exp(i))(exp(i))yi
y!
with i = (X)i.
Binomial model: Suppose that yi Bin(ni, i), ni known. Standard

link function is logit of probability of success :

i
g(i) = log
1 i

= (X)i = i
For a vector of data y:

y i
niyi
ni
1
exp(i)
p(y|) = ni=1
1 + exp(i)
1 + exp(i)
yi
Another link used in econometrics is the probit link:
1(i) = i
with (.) the normal cdf.
In practice, inference from logit and probit models is almost the same,
except in extremes of the tails of the distribution.
Overdispersion
In many applications, the model can be formulated to allow for extra
variability or overdispersion.
E.g. in Poisson model, the variance is constrained to be equal to the
mean.
As an example, suppose that data are the number of fatal car accidents
at K intersections over T years. Covariates might include intersection
characteristics and traffic control devices (stop lights, etc).
To accommodate overdispersion we model the (log)rate as a linear
combination of covariates and add a random effect for intersection with
its own population distribution.
Setting up GLIMs
Canonical link functions: Canonical link is function of mean that
appears in exponent of exponential family form of sampling distribution.
All links discussed so far are canonical except for the probit.
Offset: Arises when counts are obtained from different population sizes
or volumes or time periods and we need to use an exposure. Offset is
a covariate with a known coefficient.
Example: Number of incidents in a given exposure time T are Poisson
with rate per unit of time. Mean number of incidents is T .
Link function would be log() = i, but here mean of y is not but
T .
To apply the Poisson GLIM, add a column to X with values log(T )

and fix the coefficient to 1. This is an offset.
Interpreting GLIMs
In linear models, j represents the change in the outcome when xj is
changed by one unit.
Here, j reflects changes in g() when xj is changed.
The effect of changing xj depends of current value of x.
To translate effects into the scale of y, measure changes relative to a
baseline
y0 = g 1(x0).
A change in x of x takes outcome from y0 to y where
g(y0) = x0 y0 = g 1(x0)
and
y = g 1(g(y0) + (x))
10
Priors in GLIM
We focus on priors for although sometimes is present and has its
own prior.
Non-informative prior for :
With p() 1, posterior mode = MLE for
Approximate posterior inference can be
approximation to posterior at mode.
based
on
normal
Conjugate prior for :

As in regression, express prior information about in terms of
hypothetical data obtained under same model.
Augment data vector and model matrix with y0 hypothetical
observations and X0n0k hypothetical predictors.
11
Non-informative prior for in augmented model.

Non-conjugate priors:
Often more natural to model p(|0, 0) = N (0, 0) with (0, 0)
known.
Approximate computation based on normal approximation (see next)
particularly suitable.
Hierarchical GLIM:
Same approach as in linear models.
Model some of the as exchangeable with common population
distribution with unknown parameters. Hyperpriors for parameters.
12
Computation
Posterior distributions of parameters can be estimated using MCMC
methods in WinBUGS or other software.
Metropolis within Gibbs will often be necessary: in GLIM, most often
full conditionals do not have standard form.
An alternative is to approximate the sampling distribution with a
cleverly chosen approximation.
Idea:
)
Find mode of likelihood (,

perhaps conditional
hyperparameters
Create pseudo-data with their pseudo-variances (see later)
Model pseudo-data as normal with known (pseudo-)variances.
on
13
Normal approximation to likelihood

Objective: find zi and i2 such that normal likelihood
N (zi|(X)i, i2)
is good approximation to GLIM likelihood p(yi|(X)i, ).
)
be mode of (, ) so that i is the mode of i.
Let (,
For L the loglikelihood, write
p(y1, ..., yn) = ip(yi|i, )
= i exp(L(yi|i, ))
14
Approximate factor in exponent by normal density in i:

1
L(yi|i, ) 2 (zi i)2,
2i
where (zi, i2) depend on (yi, i, ).
Now need to find expressions for (zi, i2).
To get (zi, i2), match first and second order terms in Taylor approx
around i to (i, i2) and solve for zi and for i2.
Let L0 = L/i:
1
L = 2 (zi i)
i
0
15
Let L00 = 2L/i2:

L00 =
1
i2
Then
zi
i2
L0(yi|
i, )
= i
L00(yi|
i, )
=
L00(yi|
i, )
Example: binomial model with logit link:

exp(i)
L(yi, |i) = yi log
1 + exp(i)

1
+(ni yi) log
1 + exp(i)
16
= yii ni log(1 + exp(i))

Then
exp(i)
1 + exp(i)
L0 = yi ni
L00 = ni
exp(i)
(1 + exp(i))2
Pseudo-data and pseudo-variances:

2
zi = i +
i2 =
(1 + exp(
i))
exp(
i)
yi
exp(
i)
ni 1 + exp(
i)
1 (1 + exp(
i))2
ni
exp(
i)
17
Models for multinomial responses

Multinomial data: outcomes y = (yi, ..., yK ) are counts in K categories.
Examples:
Number of students receiving grades A, B, C, D or F
Number of alligators that prefer to eat reptiles, birds, fish,
invertebrate animals, or other (see example later)
Number of survey respondents who prefer Coke, Pepsi or tap water.
In Chapter 3, we saw non-hierarchical multinomial models:
y
p(y|) kj=1j j
with j : probability of jth outcome and
n.
Pk
j=1 j
= 1 and
Pk
j=1 yj
18
Here: we model j as a function of covariates (or predictors) X with

corresponding regression coefficients j .
For full hierarchical structure, the j are modeled as exchangeable with
some common population distribution p(|, ).
Model can be developed as extension of either binomial or Poisson
models.
19
Logit model for multinomial data

Here i = 1, ..., I is number of covariate patters. E.g., in alligator
example, 2 sizes four lakes = 8 covariate categories.
Let yi be a multinomial random variable with sample size ni and k
possible outcomes. Then
yi Mult (ni; i1, ..., ik )
with
i yi = ni , and
Pk
j
ij = 1.
ij is the probability of jth outcome for ith covariate combination.

Standard parametrization: log of the probability of jth outcome relative
20
to baseline category j = 1:

log
ij
i1

= ij = (Xj )i,
with j a vector of regression coefficients for jth category.

p(y|) Ii=1kj=1
exp(ij )
!yij
Pk
l=1 exp(il )
For identifiability, 1 = 0 and thus i1 = 0 for all i.

j is effect of changing X on probability of category j relative to
category 1.
21
Typically, indicators for each outcome category are added to predictors

to indicate relative frequency of each category when X = 0. Then
ij = j + (Xj )i
with 1 = 1 = 0 typically.
22
Example from WinBUGS - Alligators

Agresti (1990) analyzes feeding choices of 221 alligators.
Response is one of five categories: fish, invertebrate, reptile, bird,
other.
Two covariates: length of alligator (less than 2.3 meters or larger than
2.3 meters) and lake (Hancock, Oklawaha, Trafford, George).
2 4 = 8 covariate combinations (see data)
For i, j a combination of size and lake, we have counts in five possible
categories yij = (yij1, ..., yij5).
Model
p(yij |ij , nij ) = Mult (yij |nij , ij1, ..., ij5)
23
with
exp(ijk )
ijk = Pk
,
l=1 exp(ijl )
and
ijk = k + ik + jk .
Here,
k is baseline indicator for category k
ik is coefficient for indicator for lake
jk is coefficient for indicator for size
24
box plot: alpha
[5]
0.0
[2]
[4]
[3]
-2.0
-4.0
-6.0
alpha[2]
alpha[3]
alpha[4]
alpha[5]
Mean
Std
2.5%
Median
97.5%
-1.838
-2.655
-2.188
-0.7844
0.5278
0.706
0.58
0.3691
-2.935
-4.261
-3.382
-1.531
-1.823
-2.593
-2.153
-0.7687
-0.8267
-1.419
-1.172
-0.1168
box plot: beta
5.0
[3,2]
[2,2]
[3,3]
[2,3]
[4,2]
2.5
[3,4] [3,5]
[4,3] [4,4]
[2,4] [2,5]
[4,5]
0.0
-2.5
-5.0
beta[2,2]
beta[2,3]
beta[2,4]
beta[2,5]
beta[3,2]
beta[3,3]
beta[3,4]
beta[3,5]
beta[4,2]
beta[4,3]
beta[4,4]
beta[4,5]
Mean
Std
2.5%
Median 97.5%
2.706
1.398
-1.799
-0.9353
2.932
1.935
0.3768
0.7328
1.75
-1.595
-0.7617
-0.843
0.6431
0.8571
1.413
0.7692
0.6864
0.8461
0.7936
0.5679
0.6116
1.447
0.8026
0.5717
1.51
-0.2491
-5.1
-2.555
1.638
0.37
-1.139
-0.3256
0.636
-4.847
-2.348
-1.97
2.659
1.367
-1.693
-0.867
2.922
1.886
0.398
0.7069
1.73
-1.433
-0.7536
-0.8492
4.008
3.171
0.5832
0.4413
4.297
3.842
1.931
1.849
3.023
0.9117
0.746
0.281
box plot: gamma
4.0
[2,4]
2.0
[2,3]
[2,5]
0.0
[2,2]
-2.0
-4.0
gamma[2,2]
gamma[2,3]
gamma[2,4]
gamma[2,5]
Mean
Std
2.5%
Median
97.5%
-1.523
0.342
0.7098
-0.353
0.4101
0.5842
0.6808
0.4679
-2.378
-0.788
-0.651
-1.216
-1.523
0.3351
0.7028
-0.3461
-0.7253
1.476
2.035
0.518
box plot: p[2,,]
box plot: p[1,,]
0.8
0.8
[2,1,2]
[1,2,1]
[1,1,1]
[2,2,1]
0.6
0.6
[2,1,1]
[1,1,5]
0.4
[2,2,2]
0.4
[2,2,3]
[1,2,5]
[1,2,4]
[1,2,3]
[1,1,2]
0.2
[2,1,3]
0.2
[1,1,4]
[2,2,5]
[2,1,5]
[1,1,3]
[2,2,4]
[1,2,2]
[2,1,4]
0.0
0.0
box plot: p[3,,]
box plot: p[4,,]
0.8
[4,2,1]
0.8
[3,1,2]
0.6
0.6
[4,1,1]
[4,1,2]
[3,2,1]
0.4
[3,2,3]
[3,1,5]
[3,1,1]
[3,2,5]
[3,2,2]
0.4
[4,2,2]
[3,2,4]
0.2
[3,1,3]
0.2
[3,1,4]
[4,1,4]
[4,1,3]
0.0
0.0
[4,2,4]
[4,1,5]
[4,2,3]
[4,2,5]
p[1,1,1]
p[1,1,2]
p[1,1,3]
p[1,1,4]
p[1,1,5]
p[1,2,1]
p[1,2,2]
p[1,2,3]
p[1,2,4]
p[1,2,5]
p[2,1,1]
p[2,1,2]
p[2,1,3]
p[2,1,4]
p[2,1,5]
p[2,2,1]
p[2,2,2]
p[2,2,3]
p[2,2,4]
p[2,2,5]
p[3,1,1]
p[3,1,2]
p[3,1,3]
p[3,1,4]
p[3,1,5]
p[3,2,1]
0.5392
0.09435
0.04585
0.06837
0.2523
0.5679
0.02322
0.06937
0.1469
0.1927
0.2578
0.5968
0.08137
0.0095
0.05445
0.4612
0.2455
0.1947
0.0302
0.06833
0.1794
0.5178
0.09403
0.03554
0.1732
0.2937
0.07161
0.04288
0.02844
0.03418
0.06411
0.09029
0.01428
0.04539
0.07437
0.07171
0.07217
0.08652
0.0427
0.01186
0.03517
0.08211
0.06879
0.07042
0.02921
0.03984
0.05581
0.09136
0.04716
0.02504
0.06253
0.07225
0.4046
0.03016
0.007941
0.01923
0.1389
0.3959
0.005301
0.01191
0.03647
0.0793
0.1367
0.4159
0.01939
0.00533
0.009716
0.3055
0.1263
0.07776
7.653E-4
0.01357
0.08555
0.3334
0.02652
0.006063
0.07053
0.1618
0.5384
0.08735
0.04043
0.06237
0.2496
0.5684
0.02013
0.05918
0.1344
0.1852
0.2496
0.6
0.0728
0.04105
0.0476
0.4632
0.2426
0.189
0.02009
0.06097
0.1732
0.5185
0.08517
0.02996
0.167
0.2901
0.676
0.2067
0.1213
0.1457
0.3808
0.7342
0.06186
0.1768
0.3182
0.349
0.4265
0.7551
0.1841
0.143
0.1398
0.6175
0.3867
0.3502
0.1122
0.1643
0.296
0.6933
0.2219
0.1055
0.3261
0.4469
Interpretation of results
Because we set to zero several model parameters, interpreting results is
tricky. For example:
Beta[2,2] ha posterior mean 2.714. This is the effect of lake Oklawaha
(relative to Hancock) on the alligators preference for invertebrates
relative to fish.
Since beta[2,2] > 0, we conclude that alligators in Oklawaha eat more
invertebrates than do alligators in Hancock (even though both may
prefer fish!).
Gamma[2,2] is the effect of size 2 relative to size on the relative
preference for invertebrates. Since gamma[2,2] < 0, we conclude that
large alligators prefer fish more than do small alligators.
The alpha are baseline counts for each type of food relative to fish.
Hierarchical Poisson model

Count data are often modeled using a Poisson model.
If y Poisson() then E(y) = var(y) = .
When counts are assumed exchangeable given and the rates can
also be assumed to be exchangeable, a Gamma population model for
the rates is often chosen.
The hierarchical model is then
yi
Poisson(i)
Gamma(, ).
25
Priors for the hyperparameters are often taken to be Gamma (or

exponential):

Gamma(a, b)
Gamma(c, d),
with (a, b, c, d) known.
The joint posterior distribution is

p(, , |y) iyi i exp{i}i1 exp{i}
a1 exp{b} c1 exp{d}
To carry out Gibbs sampling we need to find the full conditional
distributions.
26
Conditional for i is
p(i| all) yi i+1 exp{i( + 1)},
which is proportional to a Gamma with parameters (yi + , + 1).
The full conditional for is
p(| all) i1
a1 exp{b}.
i
The conditional for does not have a standard form.
For :
p(| all) i exp{i} c1 exp{d}
X
c1

exp{(
i + d)},
i
27
which is proportional to a Gamma with parameters (c,
i i
+ d).
Computation:
Given , , draw each i from the corresponding Gamma conditional.
Draw using a Metropolis step or rejection sampling or inverse cdf
method.
Draw from the Gamma conditional.
See Italian marriages example.
28
Italian marriages Example

Gill (2002) collected data on the number of marriages per 1,000 people in Italy
during 1936-1951.
Question: did the number of marriages decrease during WWII years? (1939
1945).
Model:
Number of marriages yi are Poisson with year-specific means i.
Assuming that rates of marriages are exchangeable across years, we model the
i as Gamma(, ).
To complete model specification, place independent Gamma priors on (, ), with
known hyper-parameter values.
WinBUGS code:
model {
for (i in 1:16) {
y[i] ~ dpois(l[i])
l[i] ~ dgamma(alpha, beta)
}
alpha ~ dgamma(1,1)
beta ~ dgamma(1,1)
warave <- (l[4]+l[5]+ l[6]+l[7]+l[8]+l[9]+l[10]) / 7
nonwarave<- (l[1]+l[2]+l[3]+l[11]+l[12]+l[13]+l[14]+l[15]+l[16]) / 9
diff <- nonwarave - warave
}
list(y = c(7,9,8,7,7,6,6,5,5,7,9,10,8,8,8,7))
Results
lambda: 95% credible sets
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
0.0
5.0
10.0
lambda: box plot
15.0
[12]
[2]
[11]
[13]
[3]
[1]
[4]
[5]
[10]
[6]
[7]
[8]
10.0
5.0
0.0
[9]
[14]
[15]
[16]
Difference between non-war and war years marriage rate

Nonwar - war average
0.4
0.3
0.2
0.1
0.0
-5.0
0.0
5.0
Overall marriage rate:

If i ~ Gamma(, ), then E(i | y) = / .
overallrate
overallstd
mean sd
7.362 1.068
2.281 0.394
2.5% median
5.508
7.285
1.59
2.266
97.5%
9.665
3.125
overallrate chains 1:3 sample: 3000

0.6
0.4
0.2
0.0
4.0
6.0
8.0
10.0
overallstd chains 1:3 sample: 3000

1.5
1.0
0.5
0.0
1.0
2.0
3.0
4.0
Poisson regression
When rates are not exchangeable, we need to incorporate covariates
into the model. Often we are interested in the association between one
or more covariate and the outcome.
It is possible (but not easy) to incorporate covariates into the PoissonGamma model.
Christiansen and Morris (1997, JASA) propose the following model:
Sampling distribution, where ei is a known exposure:
yi|i Poisson(iei).
Under model, E(yi/ei) = i.
29
Population distribution for the rates:

i| Gamma(, /i),
with log(i) = x0i, and = (0, 1, ..., k1, ).
is thought of as an unboserved prior count.
Under population model,
E(i) =
/i
= i
CV2(i) =
=
2i 1
2i
1
30
For k = 0, i is known. For k = 1, i are exchangeable. For k 2, i

are (unconditionally) nonexchangeable.
In all cases, standardized rates i/i are
exchangeable, and have expectation 1.
Gamma(, ),
are
The covariates can include random effects.

To complete specification of model, we need priors on .
Christensen and Morris (1997) suggest:
and independent a priori.
Non-informative prior on s associated to fixed effects.
For a proper prior of the form:
p(|y0)
y0
,
2
( + y0)
31
where y0 is the prior guess for the median of .

Small values of y0 (for example, y0 < and the MLE of ) provide
less information.
When the rates cannot be assumed to be exchangeable, it is common
to choose a generalized linear model of the form:
p(y|) i exp{i}yi i
i exp{ exp(i)}[exp(i)]yi ,
for i = x0i and log(i) = i.
The vector of covariates can include one or more random effects to
accommodate additional dispersion (see epylepsy example).
32
The second-level distribution for the s will typically be flat (if covariate
is a fixed effect) or normal
j Normal(j0, 2j )
if jth covariate is a random effect. The variance 2j represents the
between batch variability.
33
Epilepsy example
From Breslow and Clayton, 1993, JASA.
Fifty nine epilectic patients in a clinical trial were randomized to a new
drug: T = 1 is the drug and T = 0 is the placebo.
Covariates included:
Baseline data: number of seizures during eight weeks preceding trial
Age in years.
Outcomes: number of seizures during the two weeks preceding each of
four clinical visits.
Data suggest that number of seizures was significantly lower prior to
fourth visit, so an indicator was used for V4 versus the others.
34
Two random effects in the model:

A patient-level effect to introduce between patient variability.
A patients by visit effect to introduce between visit within patient
dispersion.
35
Epilepsy study Program and results

model {
for(j in 1 : N) {
for(k in 1 : T) {
log(mu[j, k]) <- a0 + alpha.Base * (log.Base4[j] - log.Base4.bar)
+ alpha.Trt * (Trt[j] - Trt.bar)
+ alpha.BT * (BT[j] - BT.bar)
+ alpha.Age * (log.Age[j] - log.Age.bar)
+ alpha.V4 * (V4[k] - V4.bar)
+ b1[j] + b[j, k]
y[j, k] ~ dpois(mu[j, k])
b[j, k] ~ dnorm(0.0, tau.b);
# subject*visit random effects
}
b1[j] ~ dnorm(0.0, tau.b1)
# subject random effects
BT[j] <- Trt[j] * log.Base4[j] # interaction
log.Base4[j] <- log(Base[j] / 4) log.Age[j] <- log(Age[j])
diff[j] <- mu[j,4] mu[j,1]
}
# covariate means:
log.Age.bar <- mean(log.Age[])
Trt.bar <- mean(Trt[])
BT.bar <- mean(BT[])
log.Base4.bar <- mean(log.Base4[])
V4.bar <- mean(V4[])
# priors:
a0 ~ dnorm(0.0,1.0E-4)
alpha.Base ~ dnorm(0.0,1.0E-4)
alpha.Trt ~ dnorm(0.0,1.0E-4);
alpha.BT ~ dnorm(0.0,1.0E-4)
alpha.Age ~ dnorm(0.0,1.0E-4)
alpha.V4 ~ dnorm(0.0,1.0E-4)
tau.b1 ~ dgamma(1.0E-3,1.0E-3); sigma.b1 <- 1.0 / sqrt(tau.b1)
tau.b
~ dgamma(1.0E-3,1.0E-3); sigma.b <- 1.0/ sqrt(tau.b)
# re-calculate intercept on original scale:
alpha0 <- a0 - alpha.Base * log.Base4.bar - alpha.Trt * Trt.bar
- alpha.BT * BT.bar - alpha.Age * log.Age.bar - alpha.V4 * V4.bar
}
Results
Parameter
alpha.Age
alpha.Base
alpha.Trt
alpha.V4
alpha.BT
sigma.b1
sigma.b
Mean
0.4677
0.8815
-0.9587
-0.1013
0.3778
0.4983
0.3641
Std
0.3557
0.1459
0.4557
0.08818
0.2427
0.07189
0.04349
2.5th
-0.2407
0.5908
-1.794
-0.273
-0.1478
0.3704
0.2871
Median
0.4744
0.8849
-0.9637
-0.09978
0.3904
0.4931
0.362
97.5th
1.172
1.165
-0.06769
0.07268
0.7886
0.6579
0.4552
Individual random effects
box plot: b1
2.0
[4]
[1] [2]
[32]
[33]
[8]
[3]
1.0
[18]
[22]
[11]
[5]
[6]
[7]
[9]
[12] [14]
[13] [15]
[21]
[20]
[19]
[28]
[24]
[23]
[56]
[35]
[25]
[10]
[37]
[40]
[43]
[42]
[27]
[26]
[45]
[44]
[30][31]
[29]
[49]
[46]
[36]
[39]
[53]
[48]
[34]
[50]
[51]
[41]
0.0
-1.0
-2.0
[17]
[16]
[38]
[55]
[59]
[47]
[54]
[57]
[52]
[58]
Diff in seizure counts between V4 and V1
[1]
[2]
[3]
[4]
[8]
[49]
-100.0
-50.0
[7]
[6]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[26]
[27]
[28]
[29]
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
[38]
[39]
[40]
[41]
[42]
[43]
[44]
[45]
[46]
[47]
[48]
[50]
[51]
[52]
[53]
[54]
[55]
[56]
[57]
[58]
[59]
0.0
[5]
[25]
Patients 5 and 49 are different. For #5, the number of events

increases from visits 1 through 4. For #49, there is a significant
decrease.
box plot: b[5, ]
box plot: b[49, ]
1.0
[5,4]
1.0
[5,2]
[49,1]
0.5
[5,3]
0.5
[5,1]
0.0
0.0
-0.5
-1.0
[49,3]
[49,2]
-0.5
[49,4]
Convergence and autocorrelation

We ran three parallel chains for 10,000 iterations
Discarded the first 5,000 iterations from each chain as burn-in
Thinned by 10, so there are 1,500 draws for inference.
sigma.b1 chains 1:3
sigma.b chains 1:3

1.0
0.5
0.0
-0.5
-1.0
1.0
0.5
0.0
-0.5
-1.0
40
20
lag
lag
sigma.b1 chains 1:3
sigma.b chains 1:3

1.5
1.5
1.0
1.0
0.5
0.5
0.0
0.0
5001 6000
40
20
8000
iteration
5001 6000
8000
iteration
Example: hierarchical Poisson model

The USDA collects data on the number of farmers who adopt
conservation tillage practices.
In a recent survey, 10 counties in a midwestern state were randomly
sampled, and within county, a random number of farmers were
interviewed and asked whether they had adopted conservation practices.
The number of farmers interviewed in each county varied from a low
of 2 to a high of 100.
Of interest to USDA is the estimation of the overall adoption rate in
the state.
KSU - April 2005
We fitted the following model:

yi
Poisson(i)
i = i Ei
i
Gamma(, ),
where yi is the number of adopters in the ith county, Ei is the number of

farmers interviewed, i is the expected adoption rate, and the expected
number of adopters in a county can be obtained by multiplying i by
the number of farms in the county.
The hierarchical structure establishes that even though counties may
vary significantly in terms of the rate of adoption, they are still
exchangeable, so the rates are generated from a common population
distribution.
Information about the overall rate of adoption across the state is
KSU - April 2005
contained in the posterior distribution of (, ).

We fitted the model using WinBUGS and chose non-informative priors
for the hyperparameters.
Observed data:
County
1
2
3
4
5
KSU - April 2005
Ei
94
15
62
126
5
yi
5
1
5
14
3
County
6
7
8
9
10
Ei
31
2
2
9
100
yi
19
1
1
4
22
Posterior distribution of rate of adoption in each county

box plot: theta
1.5
[7]
[8]
[5]
1.0
[6]
[9]
0.5
[10]
[2]
[3]
[1]
0.0
[4]
node
mean
theta[1] 0.06033
theta[2] 0.1074
theta[3] 0.09093
theta[4] 0.1158
theta[5] 0.5482
theta[6] 0.6028
theta[7] 0.4586
theta[8] 0.4658
theta[9] 0.4378
theta[10] 0.2238
sd
2.5%
median 97.5%
0.02414
0.08086
0.03734
0.03085
0.2859
0.1342
0.3549
0.3689
0.2055
0.0486
0.02127
0.009868
0.03248
0.06335
0.131
0.3651
0.0482
0.04501
0.1352
0.1386
0.05765
0.08742
0.08618
0.1132
0.4954
0.5962
0.3646
0.3805
0.3988
0.2211
0.1147
0.306
0.1767
0.1831
1.242
0.8932
1.343
1.41
0.9255
0.33
Posterior distribution of number of adopters in each county, given

exposures
box plot: lambda
40.0
[10]
30.0
[6]
[4]
20.0
[1]
[3]
10.0
[9]
[5]
[2]
[7]
0.0
[8]
node
mean
sd
2.5%
median 97.5%
lambda[1]
lambda[2]
lambda[3]
lambda[4]
lambda[5]
lambda[6]
lambda[7]
lambda[8]
lambda[9]
lambda[10]
5.671
1.611
5.638
14.59
2.741
18.69
0.9171
0.9317
3.94
22.38
2.269
1.213
2.315
3.887
1.43
4.16
0.7097
0.7377
1.85
0.1028
2.0
0.148
2.014
7.982
0.6551
11.32
0.09641
0.09002
1.217
13.86
5.419
1.311
5.343
14.26
2.477
18.48
0.7292
0.761
3.589
22.11
node
mean
sd
2.5%
median
alpha
beta
mean
0.8336 0.3181 0.3342 0.7989 1.582

2.094 1.123 0.4235 1.897 4.835
0.4772 0.2536 0.1988 0.4227 1.14
10.78
4.59
10.95
23.07
6.21
27.69
2.687
2.819
8.33
33.0
97.5%
Poisson model for small area deaths

Taken from Bayesian Statistical Modeling, Peter Congdon, 2001
Congdon considers the incidence of heart disease mortality in 758 electoral wards in
the Greater London area over three years (1990-92). These small areas are grouped
administratively into 33 boroughs.
Regressors:
at ward level: xij , index of socio-economic deprivation.
at borough level: wj , where

wi =
1
0
for inner London boroughs

for outer suburban boroughs
We assume borough level variation in the intercepts and in the impacts of deprivation;
this variation is linked to the category of borough (inner vs outer).
KSU - April 2005
Model
First level of the hierarchy
Oij |ij Poisson(ij )
log(ij ) = log(Eij ) + 1j + 2j (xij x

) + ij
ij N (0, )
Death counts Oij are Poisson with means ij .
log(Eij ) is an offset, i.e. an explanatory variable with known coefficient (equal
to 1).
KSU - April 2005
j = (1j , 2j )0 are random coefficients for the intercepts and the impacts of
deprivation at the borough level.
ij is a random error for Poisson over-dispersion. We fitted two models: one
without and another with this random error term.
KSU - April 2005
Model (contd)
Second level of the hierarchy

j =
where
1j
2j

N2(j , )
11 12
=
21 22
1j = 11 + 12wj
2j = 21 + 22wj
11, 12, 21, and 22 are the population coefficients for the intercepts, the
impact of borough category, the impact of deprivations, and, respectively, the
interaction impact of the level-2 regressors and level-1 regressors.
Hyperparameters

1
1
Wishart (R) ,
ij N (0, 0.1) i, j {1, 2}
2
Inv-Gamma(a, b)
KSU - April 2005
Computation of Eij
Suppose that p is an overall disease rate. Then, Eij = nij p and ij = pij /p.
The s are said:
1. externally standardized if p is obtained from another data source (such as a standard
reference table);
2. internally standardized if p is obtained from the given dataset, e.g.
p =
Oij
ij
nij
ij
In our example we rely on the latter approach.

Under (a), the joint distribution of the Oij is a product Poisson; under (b) is
multinomial.
P
However, since likelihood inference is unaffected by whether we condition on ij Oij ,
the product Poisson likelihood is commonly retained.
KSU - April 2005
Model 1: without overdispersion term

(i.e. without ij )
node
11
12
21
22
1
2
Deviance
KSU - April 2005
mean
-0.075
0.078
0.621
0.104
0.294
0.445
std.
dev.
0.074
0.106
0.138
0.197
0.038
0.077
945.800
11.320
Posterior quantiles
2.5%
median
97.5%
-0.224
-0.075
0.070
-0.128
0.078
0.290
0.354
0.620
0.896
-0.282
0.103
0.486
0.231
0.290
0.380
0.318
0.437
0.618
925.000
945.300
969.400
Model 2: with overdispersion term

node
11
12
21
22
1
2
Deviance
mean
-0.069
0.068
0.616
0.105
0.292
0.431
std.
dev.
0.074
0.106
0.141
0.198
0.037
0.075
802.400
40.290
Posterior quantiles
2.5%
median
97.5%
-0.218
-0.069
0.076
-0.141
0.068
0.278
0.335
0.615
0.892
-0.276
0.104
0.499
0.228
0.289
0.376
0.306
0.423
0.600
726.600
800.900
883.500
Including ward-level random variability ij reduces the average Poisson GLM deviance
to 802, with a 95% credible interval from 726 to 883. This is in line with the expected
value of the GLM deviance for N = 758 areas if the Poisson Model is appropriate.
KSU - April 2005
Hierarchical models for spatial data

Based on the book by Banerjee, Carlin and Gelfand Hierarchical Modeling
and Analysis for Spatial Data, 2004. We focus on Chapters 1, 2 and 5.
Geo-referenced data arise in agriculture, climatology, economics,
epidemiology, transportation and many other areas.
What does geo-referenced mean? In a nutshell, we know the geographic
location at which an observation was collected.
Why does it matter? Sometimes, relative location can provide
information about an outcome beyond that provided by covariates.
Example: infant mortality is typically higher in high poverty areas.
Even after incorporating poverty as a covariate, residuals may still be
spatially correlated due to other factors such as nearness to pollution
sites, distance to pre and post-natal care centers, etc.
Often data are also collected over time, so models that include spatial
and temporal correlations of outcomes are needed.
We focus on spatial, rather than spatio-temporal models.
We also focus on models for univariate rather than multivariate
outcomes.
Types of spatial data

Point-referenced data: Y (s) a random outcome (perhaps vectorvalued) at location s, where s varies continuously over some region D.
The location s is typically two-dimensional (latitude and longitude) but
may also include altitude. Known as geostatistical data.
Areal data: outcome Yi is an aggregate value over an areal unit with
well-defined boundaries. Here, D is divided into a finite collection of
areal units. Known as lattice data even though lattices can be irregular.
Point-pattern data: Outcome Y (s) is the occurrence or not of an event
and locations s are random. Example: locations of trees of a species
in a forest or addresseses of persons with a particular disease. Interest
is often in deciding whether points occur independently in space or
whether there is clustering.
Marked point process data: If covariate information is available we talk

about a marked point process. Covariate value at each site marks the
site as belonging to a certain covariate batch or group.
Combinations: e.g. ozone daily levels collected in monitoring stations
with precise location, and number of children in a zip code reporting to
the ER on that day. Require data re-alignment to combine outcomes
and covariates.
Models for point-level data

The basics
Location index s varies continuously over region D.
We often assume that the covariance between two observations at
locations si and sj depends only on the distance dij between the
points.
The spatial covariance is often modeled as exponential:
Cov (Y (si), Y (si0 )) = C(dii0 ) = 2edii0 ,
where ( 2, ) > 0 are the partial sill and decay parameters, respectively.
Covariogram: a plot of C(dii0 ) against dii0 .
For i = i0, dii0 = 0 and C(dii0 ) = var(Y (si)).

Sometimes, var(Y (si)) = 2 + 2, for 2 the nugget effect and 2 + 2
the sill.
Models for point-level data (contd)

Covariance structure
Suppose that outcomes are normally distributed and that we choose an
exponential model for the covariance matrix. Then:
Y |, N(, ()),
with
Y
()ii0
= {Y (s1), Y (s2), ..., Y (sn)}

=
cov(Y (si), Y (si0 ))
= ( 2, 2, ).
Then
()ii0 = 2 exp(dii0 ) + 2Ii=i0 ,
with ( 2, 2, ) > 0.
This is an example of an isotropic covariance function: the spatial
correlation is only a function of d.
Models for point-level data, details

Basic model:
Y (s) = (s) + w(s) + e(s),
where (s) = x0(s) and the residual is divided into two components:
w(s) is a realization of a zero-centered stationary Gaussian process and
e(s) is uncorrelated pure error.
The w(s) are functions of the partial sill 2 and decay parameters.
The e(s) introduces the nugget effect 2.
2 interpreted as pure sampling variability or as microscale variability,
i.e., spatial variability at distances smaller than the distance between
two outcomes: the e(s) are sometimes viewed as spatial processes with
rapid decay.
The variogram and semivariogram

A spatial process is said to be:
Strictly stationary if distributions of Y (s) and Y (s + h) are equal,
for h the distance.
Weakly stationary if (s) = and Cov(Y (s), Y (s + h)) = C(h).
Instrinsically stationary if
E[Y (s + h) Y (s)] = 0, and
E[Y (s + h) Y (s)]2 = V ar[Y (s + h) Y (s)]
= 2(h),
defined for differences and depending only on distance.
2(h) is the variogram and (h) is the semivariogram
10
Examples of semi-variograms
Semi-variograms for the linear, spherical and exponential models.
11
Stationarity
Strict stationarity implies weak stationarity but the converse is not true
except in Gaussian processes.
Weak stationarity implies intrinsec stationarity, but the converse is not
true in general.
Notice that intrinsec stationarity is defined on the differences between
outcomes at two locations and thus says nothing about the joint
distribution of outcomes.
12
Semivariogram (contd)
If (h) depends on h only through its length ||h||, then the spatial
process is isotropic. Else it is anisotropic.
There are many choices for isotropic models. The exponential model
is popular and has good properties. For t = ||h||:
(t) = 2 + 2(1 exp(t)) if t > 0,
= 0 otherwise.
See figures, page 24.
The powered exponential model has an extra parameter for smoothness:
(t) = 2 + 2(1 exp(t)) if t > 0
13
Another popular choice is the Gaussian variogram model, equal to the

exponential except for the exponent term, that is exp(2t2)).
Fitting of the variogram has been traditionally done by eye:
Plot an empirical estimate of the variogram akin to the sample
variance estimate or the autocorrelation function in time series
Choose a theoretical functional form to fit the empirical
Choose values for ( 2, 2, ) that fit the data well.
If a distribution for the outcomes is assumed and a functional form
for the variogram is chosen, parameter estimates can be estimated via
some likelihood-based method.
Of course, we can also be Bayesians.
14
Point-level data (contd)

For point-referenced data, frequentists focus on spatial prediction using
kriging.
Problem: given observations {Y (s1), ..., Y (sn)}, how do we predict
Y (so) at a new site so?
Consider the model
Y = X + , where N(0, ),
and where
= 2H() + 2I.
Here, H()ii0 = (, dii0 ).
15
Kriging consists in finding a function f (y) of the observations that

minimizes the MSE of prediction
Q = E[(Y (so) f (y))2|y].
16
Classical kriging (contd)

(Not a surprising!) Result: f (y) that minimizes Q is the conditional
mean of Y (s0) given observations y (see pages 50-52 for proof):
1(y X )
E[Y (so)|y] = x0o + 0
1 ,
V ar[Y (so)|y] =
2 + 2 0
where
do1), ...,
don))
= (
2(,
2(,
1X)1X 0
1y
= (X 0
2H().
Solution assumes that we have observed the covariates xo at the new
site.
17
If not,in the classical framework Y (so), xo are jointly estimated using

an EM-type iterative algorithm.
18
Bayesian methods for estimation

The Gaussian isotropic kriging model is just a general linear model
similar to those in Chapter 15 of textbook.
Just need to define the appropriate covariance structure.
For an exponential covariance structure with a nugget effect, parameters
to be estimated are = (, 2, 2, ).
Steps:
Choose priors and define sampling distribution
Obtain posterior for all parameters p(|y)
Bayesian kriging: get posterior predictive distribution for outcome at
new location p(yo|y, X, xo).
19
Sampling distribution (marginal data model)

y| N(X, 2H() + 2I)
Priors: typically chosen so that parameters are independent a priori.
As in the linear model:
Non-informative prior for is uniform or can use a normal prior too.
Conjugate priors for variances 2, 2 are inverse gamma priors.
For , appropriate prior depends on covariance model. For simple
exponential where
(si sj ; ) = exp(||si sj ||),
a Gamma prior can be a good choice.
20
Be cautious with improper priors for anything but .
21
Hierarchical representation of model

Hierarchical model representation: first condition on the spatial random
effects W = {w(s1), ..., w(sn)}:
y|, W
W |, 2
N(X + W, 2I)
N(0, 2H()).
Model specification is then completed by choosing priors for , 2 and

for , 2 (hyperparameters).
Note that hierarchical model has n more parameters (the w(si)) than
the marginal model.
Computation with the marginal model preferable because 2H()+ 2I
tends to be better behaved than 2H() at small distances.
22
Estimation of spatial surface W |y

Interest is sometimes on estimating the spatial surface using p(W |y).
If marginal model is fitted, we can still get marginal posterior for W as
Z
p(W |y) = p(W | 2, )p( 2, |y)d 2d.
Given draws ( 2(g), (g)) from the Gibbs sampler on the marginal
model, we can generate W from
p(W | 2(g), (g)) = N(0, 2(g)H((g))).
Analytical marginalization over W is possible only if model has Gaussian
form.
23
Bayesian kriging
Let Yo = Y (so) and xo = x(so). Kriging is accomplished by obtaining
the posterior predictive distribution
Z
p(yo|xo, X, y) =
p(yo, |y, X, xo)d
Z
=
p(yo|, y, xo)p(|y, X)d.
Since (Yo, Y ) are jointly multivariate normal (see expressions 2.18 and
2.19 on page 51), then p(yo|, y, xo) is a conditional normal distribution.
Given MCMC draws of the parameters ((1), ..., (G)) from the posterior
(g)
distribution p(|y, X), we draw values yo for each (g) as
yo(g) p(yo|(g), y, xo).
24
(1)
(2)
(G)
Draws {yo , yo , ..., yo , } are a sample from the posterior predictive

distribution of the outcome at the new location so.
To predict Y at a set of m new locations so1, ..., som, it is best to do
joint prediction to be able to estimate the posterior association among
m predictions.
Beware of joint prediction at many new locations with WinBUGS. It
can take forever.
25
Kriging example from WinBugs

Data were first published by Davis (1973) and consist of heights at 52
locations in a 310-foot square area.
We have 52 s = (x, y) coordinates and outcomes (heights).
Unit of distance is 50 feet and unit of elevation is 10 feet.
The model is
height = + , where N(0, ),
and where
= 2H().
26
Here, H()ij = (si sj ; ) = exp(||si sj ||).

Priors on (, , ).
We predict elevations at 225 new locations.
27
model {
# Spatially structured multivariate normal likelihood
height[1:N] ~ spatial.exp(mu[], x[], y[], tau, phi, kappa)
# exponential correlation function
for(i in 1:N) {
mu[i] <- beta
}
# Priors
beta ~ dflat()
tau ~ dgamma(0.001, 0.001)
sigma2 <- 1/tau
# priors for spatial.exp parameters
phi ~ dunif(0.05, 20)
# prior decay for correlation at min distance (0.2 x 50 ft) is 0.02 to 0.99
# prior range for correlation at max distance (8.3 x 50 ft) is 0 to 0.66
kappa ~ dunif(0.05,1.95)
# Spatial prediction
# Single site prediction
for(j in 1:M) {
height.pred[j] ~ spatial.unipred(beta, x.pred[j], y.pred[j], height[])
}
# Only use joint prediction for small subset of points, due to length of time it takes to run
for(j in 1:10) { mu.pred[j] <- beta }
height.pred.multi[1:10] ~ spatial.pred(mu.pred[], x.pred[1:10], y.pred[1:10], height[])
}
Data Click on one of the arrows for the data

Initial values
Click on one of the arrows for inital values for spatial.exp model
beta
1200.0
1.00E+3
800.0
600.0
400.0
1
500
1000
iteration
1500
2000
phi
0.8
0.6
0.4
0.2
0.0
1
500
1000
iteration
1500
2000
kappa
2.0
1.5
1.0
0.5
1
500
1000
iteration
1500
2000
(samples)means for height.pred
(1) < 700.0

(16) 700.0 - 750.0
N
(40) 750.0 - 800.0
(80) 800.0 - 850.0
(61) 850.0 - 900.0
(26) 900.0 - 950.0
(1) >= 950.0
0.002km
Hierarchical models for areal data

Areal data: often aggregate outcomes over a well-defined area.
Example: number of cancer cases per county or proportion of people
living in poverty in a set of census tracks.
What are the inferential issues?
1. Identifying a spatial pattern and its strength. If data are spatially
correlated, measurements from areas that are close will be more
alike.
2. Smoothing and to what degree. Observed measurements often
present extreme values due to small samples in small areas. Maximal
smoothing: substitute observed measurements by the overall mean
in the region. Something less extreme is what we discuss later.
3. Prediction: for a new area, how would we predict Y given
measurements in other areas?
28
Defining neighbors
A proximity matrix W with entries wij spatially connects areas i and
j in some fashion.
Typically, wii = 0.
There are many choices for the wij :
Binary: wij = 1 if areas i, j share a common boundary, and is 0
otherwise.
Continuous: decreasing function of intercentroidal distance.
Combo: wij = 1 if areas are within a certain distance.
W need not be symmetric.
P
Entries are often standardized by dividing into j wij = wi+. If entries

are standardized, the W will often be asymmetric.
29
The wij can be thought of as weights and provide a means to introduce

spatial structure into the model.
Areas that are closer by in some sense are more alike.
For any problem, we can define first, second, third, etc order neighbors.
For distance bins (0, d1], (d1, d2], (d2, d3], ... we can define
(1)
1. W (1), the first-order proximity matrix with wij = 1 if distance

between i abd j is less than d1.
(2)
2. W (2), the second-order proximity matrix with wij = 1 if distance
between i abd j is more than d1 but less than d2.
30
Areal data models

Because these models are used mostly in epidemiology, we begin with
an application in disease mapping to introduce concepts.
Typical data:
Yi observed number of cases in area i, i = 1, ..., I
Ei expected number of cases in area i.
The Y s are assumed to be random and the Es are assumed to be
known and to depend on the number of persons ni at risk.
An internal standardized estimate of Ei is
P
yi
i
Ei = nir = ni P
,
n
i i
31
corresponding to a constant disease rate across areas.

An external standardized estimate is
Ei =
nij rj ,
where rj is the risk for persons of age group j (from some existing
table of risks by age) and nij is the number of persons of age j in area
i.
32
Standard frequentist approach

For small Ei,
Yi|i Poisson(Eii),
with i the true relative risk in area i.
The MLE is the standard mortality ratio
i = SM Ri =
Yi
.
Ei
The variance of the SM Ri is

var(Yi)
i
var(SM Ri) =
= ,
2
Ei
Ei
33
estimated by plugging i to obtain

Yi
var(SM Ri) = 2 .
Ei
To get a confidence interval for i, first assume that log(SM Ri) is

approximately normal.
From a Taylor expansion:

V ar[log(SM Ri)]
=
1
V ar(SM Ri)
SM Ri2
Yi
1
Ei2
2= .
2
Yi
Ei
Yi
34
An approximate 95% CI for log(i) is

SM Ri 1.96/(Yi)1/2.
Transforming back, an approximate 95% CI for i is
(SM Ri exp(1.96/(Yi)1/2), SM Ri exp(1.96/(Yi)1/2)).
Suppose we wish to test whether risk in area i is high relative to other
areas. Then test
H0 : i = 1 versus Ha : i > 1.
This is a one-sided test.
35
Under H0, Yi Poisson(Ei) so the pvalue for the test is

p =
Prob(X Yi|Ei)
= 1 Prob(X Yi|Ei)
Yi 1
X exp(Ei)E x
i
.
= 1
x!
x=0
If p < 0.05 we reject H0.
36
Hierarchical models for areal data

To estimate and map underlying relative risks, might wish to fit a
random effects model.
Assumption: true risks come from a common underlying distribution.
Random effects models permit borrowing strength across areas to
obtain better area-level estimates.
Alas, models may be complex:
High-dimensional: one random effect for each area.
Non-normal if data are counts or binomial proportions.
We have already discussed hierarchical Poisson models, so material in
the next few transparencies is a review.
37
Poisson-Gamma model
Consider
Yi|i
Poisson(Eii)
i|a, b
Gamma(a, b).
Since E(i) = a/b and V ar(i) = a/b2, we can fix a, b as follows:

A priori, let E(i) = 1, the null value.
Let V ar(i) = (0.5)2, large on that scale.
Resulting prior is Gamma(4, 4).
Posterior is also Gamma:
p(i|yi) = Gamma(yi + a, Ei + b).
38
A point estimate of i is
E(i|y) = E(i|yi) =
=
yi + a
Ei + b
[E(i )]2
yi + V ar(i)
i)
Ei + VE(
ar(i )
Ei( Eyii )
Ei +
E(i )
V ar(i )
E(i )
V ar(i ) E(i )
i)
Ei + VE(
ar(i )
= wiSM Ri + (1 wi)E(i),
where wi = Ei/[Ei + (E(i)/V ar(i))].
Bayesian point estimate is a weighted average of the data-based SM Ri
and the prior mean E(i).
39
Poisson-lognormal models with spatial errors

The Poisson-Gamma model does not allow (easily) for spatial correlation
among the i.
Instead, consider the Poisson-lognormal model, where in the second
stage we model the log-relative risks log(i) = i:
Yi|i
Poisson(Ei exp(i))
i = x0i + i + i,
where xi are area-level covariates.
The i are assumed to be exchangeable and model between-area
variability:
i N(0, 1/h).
40
The i incorporate global extra-Poisson variability in the log-relative

risks (across the entire region).
The i are the spatial parameters; they capture regional clustering.
They model extra-Poisson variability in the log-relative risks at the local
level so that neighboring areas have similar risks.
One way to model the i is to proceed as in the point-referenced data
case. For = (1, ..., I ), consider
|, NI (, H()),
and H()ii0 = cov(i, i0 ) with hyperparameters .
Possible models for H() include the exponential, the powered
exponential, etc.
41
While sensible, this model is difficult to fit because

Lots of matrix inversions required
Distance between i and i0 may not be obvious.
42
CAR model
More reasonable to think of a neighbor-based proximity measure and
consider a conditionally autoregressive model for :
i N(i, 1/(cmi)),
where
i =
wij (i j ),
i6=j
and mi is the number of neighbors of area i. Earlier we called this wi+.

The weights wij are (typically) 0 if areas i and j are not neighbors and
1 if they are.
CAR models lend themselves to the Gibbs sampler. Each i can
be sampled from its conditional distribution so no matrix inversion is
43
needed:
xi +i +i
p(i| all) Poi(yi|Eie
) N(i|i,
).
mic
44
Difficulties with CAR model

The CAR prior is improper. Prior is a pairwise difference prior identified
only up to a constant.
The posterior will still be proper, but to identify anP
intercept 0 for the
log-relative risks, we need to impose a constraint: i i = 0.
In simulation, constraint is imposed numerically by recentering each
vector around its own mean.
h and c cannot be too large because i and i become unidentifiable.
We observe only one Yi in each area yet we try to fit two random effects.
Very little data information.
Hyperpriors for h, c need to be chosen carefully.
45
Consider
h Gamma(ah, bh),
c Gamma(ac, bc).
To place equal emphasis on heterogeneity and spatial clustering, it is

tempting to make ah = ac and bh = bc. This is not correct because
1. The h prior is defined marginally, where the c prior is conditional.
2. The conditional prior precision is cmi. Thus, a scale that satisfies
1
1
sd(i) = p
sd(i)
0.7
m
h )
c
with m
the average number of neighbors is more fair (Bernardinelli
et al. 1995, Statistics in Medicine).
46
Example: Incidence of lip cancer in 56 areas in Scotland

Data on the number of lip cancer cases in 56 counties in Scotland were obtained. Expected
lip cancer counts Ei were available.
Covariate was the proportion of the population in each county working in agriculture.
Model included only one random effect bi to introduce spatial association between counties.
Two counties are neighbors if they have a common border (i.e., they are adjacent).
Three counties that are islands have no neighbors and for those, WinBUGS sets the
random spatial effect to be 0. The relative risk for the islands is thus based on the baseline
rate 0 and on the value of the covariate xi.
We wish to smooth and map the relative risks RR.
Model
model {
# Likelihood
for (i in 1 : N) {
O[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha0 + alpha1 * X[i]/10 + b[i]
RR[i] <- exp(alpha0 + alpha1 * X[i]/10 + b[i])
relative risk (for maps)
}
# CAR prior distribution for random effects:
b[1:N] ~ car.normal(adj[], weights[], num[], tau)
for(k in 1:sumNumNeigh) {
weights[k] <- 1
}
# Other priors:
alpha0 ~ dflat()
alpha1 ~ dnorm(0.0, 1.0E-5)
tau ~ dgamma(0.5, 0.0005)
sigma <- sqrt(1 / tau)
}
# prior on precision
# standard deviation
# Area-specific
Data
list(N = 56,
O = c( 9, 39, 11, 9, 15, 8, 26, 7, 6, 20,
13, 5, 3, 8, 17, 9, 2, 7, 9, 7,
16, 31, 11, 7, 19, 15, 7, 10, 16, 11,
5, 3, 7, 8, 11, 9, 11, 8, 6, 4,
10, 8, 2, 6, 19, 3, 2, 3, 28, 6,
1, 1, 1, 1, 0, 0),
E = c( 1.4, 8.7, 3.0, 2.5, 4.3, 2.4, 8.1, 2.3, 2.0, 6.6,
4.4, 1.8, 1.1, 3.3, 7.8, 4.6, 1.1, 4.2, 5.5, 4.4,
10.5,22.7, 8.8, 5.6,15.5,12.5, 6.0, 9.0,14.4,10.2,
4.8, 2.9, 7.0, 8.5,12.3,10.1,12.7, 9.4, 7.2, 5.3,
18.8,15.8, 4.3,14.6,50.7, 8.2, 5.6, 9.3,88.7,19.6,
3.4, 3.6, 5.7, 7.0, 4.2, 1.8),
X = c(16,16,10,24,10,24,10, 7, 7,16,
7,16,10,24, 7,16,10, 7, 7,10,
7,16,10, 7, 1, 1, 7, 7,10,10,
7,24,10, 7, 7, 0,10, 1,16, 0,
1,16,16, 0, 1, 7, 1, 1, 0, 1,
1, 0, 1, 1,16,10),
num = c(3, 2, 1, 3, 3, 0, 5, 0, 5, 4,
0, 2, 3, 3, 2, 6, 6, 6, 5, 3,
3, 2, 4, 8, 3, 3, 4, 4, 11, 6,
7, 3, 4, 9, 4, 2, 4, 6, 3, 4,
5, 5, 4, 5, 4, 6, 6, 4, 9, 2,
4, 4, 4, 5, 6, 5),
adj = c(
19, 9, 5,
10, 7,
12,
28, 20, 18,
19, 12, 1,
17, 16, 13, 10, 2,
29, 23, 19, 17, 1,
22, 16, 7, 2,
5, 3,
19, 17, 7,
35, 32, 31,
29, 25,
29, 22, 21, 17, 10, 7,
29, 19, 16, 13, 9, 7,
56, 55, 33, 28, 20, 4,
17, 13, 9, 5, 1,
56, 18, 4,
50, 29, 16,
16, 10,
39, 34, 29, 9,
56, 55, 48, 47, 44, 31, 30, 27,
29, 26, 15,
43, 29, 25,
56, 32, 31, 24,
45, 33, 18, 4,

50, 43, 34, 26, 25, 23, 21, 17, 16, 15, 9,
55, 45, 44, 42, 38, 24,
47, 46, 35, 32, 27, 24, 14,
31, 27, 14,
55, 45, 28, 18,
54, 52, 51, 43, 42, 40, 39, 29, 23,
46, 37, 31, 14,
41, 37,
46, 41, 36, 35,
54, 51, 49, 44, 42, 30,
40, 34, 23,
52, 49, 39, 34,
53, 49, 46, 37, 36,
51, 43, 38, 34, 30,
42, 34, 29, 26,
49, 48, 38, 30, 24,
55, 33, 30, 28,
53, 47, 41, 37, 35, 31,
53, 49, 48, 46, 31, 24,
49, 47, 44, 24,
54, 53, 52, 48, 47, 44, 41, 40, 38,
29, 21,
54, 42, 38, 34,
54, 49, 40, 34,
49, 47, 46, 41,
52, 51, 49, 38, 34,

56, 45, 33, 30, 24, 18,
55, 27, 24, 20, 18),
sumNumNeigh = 234)
Results
The proportion of individuals working in agriculture appears to be
associated to the incidence of cancer
alpha1 sample: 10000

4.0
3.0
2.0
1.0
0.0
-0.25
0.0
0.25
0.5
0.75
There appears to be spatial correlation among counties:
box plot: b
2.0
[1]
[3]
[5]
[2]
[12][13]
[7]
1.0
[9]
[10]
[15] [17] [19]
[4]
[16]
[14]
0.0
-1.0
-2.0
[18]
[20][21]
[25]
[26]
[23]
[22]
[24]
[27][28][29]
[36]
[56]
[31][32][33] [35]
[40]
[37][38][39]
[30]
[43]
[34]
[51]
[52]
[41]
[44] [46][47][48] [50]
[53][54][55]
[45]
[42]
[49]
(samples)means for RR
(26) <
(15)
1.0 -
2.0
(9)
2.0 -
3.0
(4)
3.0 -
4.0
(2) >=
200.0km
1.0
4.0
values for O
(38) < 10.0

(14)
(2)
10.0 20.0 -
(2) >= 30.0
200.0km
20.0
30.0
Example: Lung cancer in London

Data were obtained from an annual report by the London Health Authority in which the
association between health and poverty was investigated.
Population under study are men 65 years of age and older.
Available were:
Observed and expected lung cancer counts in 44 census ward districts
Ward-level index of socio-economic deprivation.
A model with two random effects was fitted to the data. One random effect introduces
region-wide heterogeneity (between ward variability) and the other one introduces regional
clustering.
The priors for the two components of random variability, are sometimes known as
convolution priors.
Interest is in:
Determining association between health outcomes and poverty.
Smoothing and mapping relative risks.
Assessing the degree of spatial clustering in these data.
Model
model {
for (i in 1 : N) {
# Likelihood
O[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha + beta * depriv[i] + b[i] + h[i]
RR[i] <- exp(alpha + beta * depriv[i] + b[i] + h[i])
relative risk (for maps)
# Exchangeable prior on unstructured random effects
h[i] ~ dnorm(0, tau.h)
}
# CAR prior distribution for spatial random effects:
b[1:N] ~ car.normal(adj[], weights[], num[], tau.b)
for(k in 1:sumNumNeigh) {
weights[k] <- 1
}
# Other priors:
alpha ~ dflat()
beta ~ dnorm(0.0, 1.0E-5)
tau.b ~ dgamma(0.5, 0.0005)
# Area-specific
sigma.b <- sqrt(1 / tau.b)

tau.h ~ dgamma(0.5, 0.0005)
sigma.h <- sqrt(1 / tau.h)
propstd <- sigma.b / (sigma.b + sigma.h)
}
Data click on one of the arrows to open data
Inits
click on one of the arrows to open initial values
Note that the priors on the precisions for the exchangeable and the spatial
random effects are Gamma(0.5, 0.0005).
That means that a priori, the expected value of the standard deviations is
approximately 0.03 with a relatively large prior standard deviation.
This is not a fair prior as discussed in class. The average number of neighbors
is 4.8 and this is not taken into account in the choice of priors.
Results
Parameter
Mean
Std
2.5th
perc.
97.5th
perc.
Alpha
Beta
Relative size of
spatial std
-0.208
0.0474
0.358
0.1
0.0179
0.243
-0.408
0.0133
0.052
-0.024
0.0838
0.874
sigma.b sample: 1000
sigma.h sample: 1000
20.0
15.0
10.0
5.0
0.0
10.0
7.5
5.0
2.5
0.0
-0.2
0.0
0.2
0.4
0.6
-0.2
0.0
0.2
0.4
0.6
Posterior distribution of RR for first 15 wards

node
mean
sd
2.5%
median 97.5%
RR[1]
RR[2]
RR[3]
RR[4]
RR[5]
RR[6]
RR[7]
RR[8]
RR[9]
RR[10]
RR[11]
RR[12]
RR[13]
RR[14]
RR[15]
0.828
1.18
0.8641
0.8024
0.7116
1.05
1.122
0.821
1.112
1.546
0.7697
0.865
1.237
0.8359
0.7876
0.1539
0.2061
0.1695
0.1522
0.1519
0.2171
0.1955
0.1581
0.2167
0.2823
0.1425
0.16
0.3539
0.1807
0.1563
0.51
0.7593
0.5621
0.5239
0.379
0.7149
0.7556
0.4911
0.7951
1.072
0.4859
0.6027
0.8743
0.5279
0.489
0.8286
1.188
0.8508
0.7873
0.7206
1.015
1.116
0.8284
1.07
1.505
0.7788
0.8464
1.117
0.8084
0.7869
1.148
1.6
1.247
1.154
0.9914
1.621
1.589
1.132
1.702
2.146
1.066
1.26
2.172
1.3
1.141
values for O
(17) <
(19)
5.0
5.0 -
10.0
N
(3)
10.0 -
15.0
(3)
15.0 -
20.0
(2) >= 20.0
2.5km
values for b
(2) < -0.1

(2)
-0.1 - -0.05
N
(18) -0.05 - 1.38778E-17
(19) 1.38778E-17 (2)
0.05 -
(1) >=
2.5km
0.1
0.1
0.05
values for RR
(8) <
0.8
(20)
0.8 -
1.0
(11)
1.0 -
1.2
(5) >=
2.5km
1.2

Notes BMDA PDF

Uploaded by

Copyright:

Available Formats

Notes BMDA PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Notes BMDA PDF

Uploaded by

Copyright:

Available Formats

.

Bayesian Methods for

Tampa, Florida March 26, 2006

ENAR - March 2006

Model checking, diagnostics, model fit.

ENAR - March 2006

Single parameter models

ENAR - March 2006

with p(y) = 1/(n + 1) uniform a priori.

Laplace devised an approximation to the integral expression in

MLE is sample proportion: = y/n.

p(|y)d = 0.025 and

A 100(1 )% highest posterior density credible set is subset C of

In other cases, HPD sets have smallest size.

Credible set and HPD set

ENAR - March 2006

A priori, all values of y are equally likely

Posterior predictive distribution, to predict outcome of a new trial y

ENAR - March 2006

Binomial model: different priors

ENAR - March 2006

Beta is the conjugate prior for binomial model: posterior is in the

ENAR - March 2006

for () the natural parameter and t(y) a sufficient statistic.

ENAR - March 2006

Conjugate prior for binomial proportion

ENAR - March 2006

where g()n = (1 )n, y is the sufficient statistic, and the logit

Consider prior for : Beta(, ):

If we let = 1 and = + 2, can write

p() (1 ) exp{ log

ENAR - March 2006

Then posterior is in same form as prior:

Since p(y|) u(1 )ny then prior Beta(, ) suggests that a

ENAR - March 2006

As y and n y get large:

ENAR - March 2006

Placenta previa example

Thus a posteriori, is Beta(438, 544) and

Check sensitivity to choice of priors: try several Beta priors with

Fix prior mean at 0.485 and increase prior sample size + .

ENAR - March 2006

Conjugate prior for Poisson rate

Consider Gamma(, ) as prior for : p() 1 exp()

ENAR - March 2006

Mean of a normal distribution

Viewed as function of , likelihood is exponential of a quadratic in :

that can be parameterized as

ENAR - March 2006

Expand squares, collect terms in 2 and in :

1 (02 + 2)n2 2(n

(by dividing numerator and denominator into ( 2/n)02)

Posterior precision = sum of prior and data precisions.

Integrand is kernel of bivariate normal, so (

In n, prior precision 1/02 and data precision n/ 2 are equivalent.

ENAR - March 2006

Likelihood is in exponential family form with natural parameter ( 2) =

Consider an inverse Gamma prior:

For ease of interpretation, reparameterize as a scaled-inverted 2