STAT 535: Chapter 5:

More Conjugate Priors

David B. Hitchcock
E-Mail: hitchcock@stat.sc.edu

Spring 2022

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Why are Conjugate Priors Nice?

▶ Recall that a conjugate prior is a prior which (along with the

data model) produces a posterior distribution that has the
same functional form as the prior (but with new, updated
parameter values).
▶ In the Beta-binomial setup, the beta prior was conjugate
because the posterior was also a beta distribution.
▶ Conjugate priors are nice because
1. we can typically derive the posterior without needing any
difficult computation;
2. it is typically easy to understand the respective contributions of
the prior information and the data information to the posterior.
▶ We will now examine a couple of other Bayesian models with
conjugate priors.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

The Poisson Distribution

▶ Recall that the Poisson distribution is a common model for

count data: Data whose possible values are the nonnegative
integers 0, 1, 2, . . ..
▶ The Poisson distribution is indexed by a parameter λ > 0, and
(given λ) the pdf of a Poisson random variable Y |λ is:

λy e −λ
f (y |λ) =
y!

▶ If our data consists of a random sample on n such counts,

then the likelihood function is the joint density function
f (y1 |λ)f (y2 |λ) · · · f (yn |λ), since Y1 , Y2 , . . . , Yn are
independent.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Choice of Prior
▶ When our data model is Poisson, what is a good choice for
the prior for the parameter λ?
▶ Since λ > 0, we should use as a prior some distribution whose
support is (0, ∞).
▶ The Gamma distribution is a good choice for the prior, since
its support is (0, ∞).
▶ Note that the parameterization of the Gamma distribution
that we will use in this class is different from the one in the
STAT 511 course.
▶ We will consider a Gamma pdf with a shape parameter s and
a rate parameter r :
r s s−1 −r λ
f (λ) = λ e , λ > 0.
Γ(s)

▶ Note that the rate parameter is the reciprocal of the scale

parameter used in the other parameterization.
David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors
The Gamma/Poisson Bayesian Model
▶ If our data Y1 , . . . , Yn are iid Poisson(λ), then a gamma(s, r )
prior on λ is a conjugate prior.
Likelihood:
n
e −λ λyi e −nλ λ yi
P
Y
L(λ|y ) = = Qn
yi ! i=1 (yi !)
i=1

Prior:
r s s−1 −r λ
f (λ) = λ e , λ > 0.
Γ(s)
⇒ Posterior:
yi +s−1 −(n+r )λ
P
f (λ|y ) ∝ λ e , λ > 0.

⇒ f (λ|y ) is gamma yi + s, n + r . (Conjugate!)
P

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Properties of the Gamma (Mean)

▶ Under this shape/rate parameterization, the mean of the

Gamma(s, r ) prior distribution is
s
E (λ) =
r
▶ Based on our prior beliefs, we would choose appropriate values
of the hyperparameters s and r .


▶ Similarly, the mean of the Gamma Pyi + s, n + r posterior
distribution is P
yi + s
E (λ|y ) =
n+r
▶ This posterior mean could be used as a Bayesian estimator of
λ.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Properties of the Gamma (Variance)

▶ If we have a good guess of the prior mean of λ, how can we

specifically choose which s and r to use in our prior?
▶ Under this shape/rate parameterization, the variance of the
Gamma(s, r ) prior distribution is
s
Var (λ) =
r2
▶ The prior variance (and standard deviation) can guide our
choices of s and r .
▶ Plotting the potential prior using the plot gamma function in
the bayesrules package can also be helpful in choosing the
prior.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

The Posterior Mean in the Gamma/Poisson Bayesian
Model

▶ The posterior mean is:

P
yi + s
λ̂B =
Pn + r
yi s
= +
n + r n P
+r    
n yi r s
= +
n+r n n+r r
▶ Again, the data get weighted more heavily as n → ∞.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example: Fraud Risk Phone Calls

▶ The textbook gives an example using data on fraud risk phone

calls per day, which can be modeled with a Poisson
distribution.
▶ The parameter of interest is λ, the mean number of fraud risk
calls per day.
▶ Prior belief: The mean number of such calls per day is around
5.
▶ So let’s choose s and r so that s/r = 5.
▶ Also, we believe that λ is very likely between 2 and 7.
▶ Let’s try to plot a few possible priors that have s/r = 5 (see R
examples).

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example: Fraud Risk Phone Calls

▶ The choice of s = 10 and r = 2 seems to reflect our prior

beliefs.
▶ We collect n = 4 counts as our data, and the data were:
P
6, 2, 2, 1 ( i yi = 11 and ȳ = 2.75).


▶ So our posterior is Gamma Pyi + s, n + r =
Gamma(11 + 10, 4 + 2) = Gamma(21, 6)
▶ A Bayesian estimate of λ is thus the posterior mean
21/6 = 3.5.
▶ See R plots to see how the data has updated our prior beliefs.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Bayesian Inference: Posterior Intervals

▶ Simple values like the posterior mean E [θ|y ] and posterior

variance var [θ|y ] can be useful in learning about θ.
▶ Quantiles of p(θ|y ) (especially the posterior median) can also
be a useful summary of θ.
▶ The ideal summary of θ is an interval (or region) with a
certain probability of containing θ.
▶ Note that a classical (frequentist) confidence interval does
not exactly have this interpretation.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Bayesian Credible Intervals

▶ A credible interval (or in general, a credible set) is the

Bayesian analogue of a confidence interval.
▶ A 100(1 − α)% credible set C is a subset of Θ such that
Z
p(θ|y ) dθ = 1 − α.
C

▶ If the parameter space Θ is discrete, a sum replaces the

integral.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Quantile-Based Intervals

▶ If θL∗ is the α/2 posterior quantile for θ, and θU∗ is the 1 − α/2
∗ ∗
posterior quantile for θ, then (θL , θU ) is a 100(1 − α)%
credible interval for θ.
Note: P[θ < θL∗ |y ] = α/2 and P[θ > θU
∗ |y ] = α/2.

⇒ P{θ ∈ (θL∗ , θU
∗
)|y }
/ (θL∗ , θU
= 1 − P{θ ∈ ∗
)|y }
 
∗ ∗
= 1 − P[θ < θL |y ] + P[θ > θU |y ]

= 1 − α.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Quantile-Based Intervals
Picture:
Gamma(21,6) posterior

0.5
0.4
0.3
f(λ|y)

0.2

area=0.025 area=0.025
area=0.95
0.1
0.0

0 2 4 6 8

Figure: Between 2.17 and 5.15 is posterior probability 0.95.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 2: Quantile-Based Interval

▶ Consider 10 flips of a coin having P{Heads} = θ.

▶ Suppose we observe 2 “heads”.
▶ We model the count of heads as binomial:
 
10 y
p(y |θ) = θ (1 − θ)10−y , y = 0, 1, . . . , 10.
y

▶ Let’s use a uniform prior for θ:

p(θ) = 1, 0 ≤ θ ≤ 1.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 2: Quantile-Based Interval

▶ Then the posterior is:

p(θ|y ) ∝ p(θ)L(θ|y )
 
10 y
= (1) θ (1 − θ)10−y
y
∝ θy (1 − θ)10−y , 0 ≤ θ ≤ 1.

▶ This is a beta distribution for θ with parameters

y + 1 and 10 − y + 1.
▶ Since y = 2 here, p(θ|y = 2) is beta(3,9).
▶ The 0.025 and 0.975 quantiles of a beta(3,9) are (.0602,
.5178), which is a 95% credible interval for θ.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

HPD Intervals / Regions

▶ The equal-tail credible interval approach is ideal when the

posterior distribution is symmetric.
▶ But what if p(θ|y ) is skewed?
Picture:
A Skewed Posterior Density
0.15
0.10
π(θ)
0.05
0.00

0 5 10 15

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

HPD Intervals / Regions

▶ Note that values of θ around 1 have much higher posterior

probability than values around 7.5.
▶ Yet 7.5 is in the equal-tails interval and 1 is not!
▶ A better approach here is to create our interval of θ-values
having the Highest Posterior Density.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

HPD Intervals / Regions

Defn: A 100(1 − α)% HPD region for θ is a subset C ∈ Θ defined

by
C = {θ : p(θ|y ) ≥ k}
where k is the largest number such that
Z
p(θ|y ) dθ = 1 − α.
θ:p(θ|y )≥k

▶ The value k can be thought of as a horizontal line placed over

the posterior density whose intersection(s) with the posterior
define regions with probability 1 − α.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

HPD Intervals / Regions

Picture: (95% HPD Interval)

Gamma(21,6) posterior

0.5
0.4
0.3
f(λ|y)

0.2
0.1

area=0.90
0.0

0 2 4 6 8

⇒ P{θL∗ < θ < θU∗ } = 0.90.

The values between θL∗ = 2.25 and θU

∗ = 4.72 here have the

highest posterior density.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

HPD Intervals / Regions

▶ The HPD region will be an interval when the posterior is

unimodal.
▶ If the posterior is multimodal, the HPD region might be a
discontiguous set.
Picture:
Bimodal posterior distribution

0.25
0.20
0.15
f(λ|y)

0.10
0.05
0.00

0 2 4 6 8

▶ The set {θ : θ ∈ (2.85, 4.1) ∪ (6.0, 7.25)} is the HPD region

for θ here.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 1 Revisited: HPD Interval

▶ See course web page for finding an HPD interval in R for λ in

the fraud risk call example.
▶ A 90% quantile-based credible interval for λ is (2.345, 4.844).
▶ Also note the hpd function in TeachingDemos package in R.
▶ See code for Example 2 (coin-flipping data) in R.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

The Normal-Normal Model

▶ Why is it so common to model data using a normal

distribution?
▶ Approximately normally distributed quantities appear often in
nature.
▶ CLT tells us any variable that is basically a sum of
independent components should be approximately normal.
▶ Note ȳ and S 2 are independent when sampling from a normal
population — so if beliefs about the mean are independent of
beliefs about the variance, a normal model may be
appropriate.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Why Normal Models?

▶ The normal model is analytically convenient (exponential

family, sufficient statistics ȳ and S 2 )
▶ Inference about the population mean based on a normal
model will be correct as n → ∞ even if the data are truly
non-normal.
▶ When we assume a normal likelihood, we can get a wide class
of posterior distributions by using different priors.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

▶ Simple situation: Assume data Y1 , . . . , Yn are iid N(µ, σ 2 ),

with µ unknown and σ 2 known.
▶ We will make inference about µ.
▶ The likelihood is
n
1 2
(2πσ 2 )− /2 e − 2σ2 (Yi −µ)
Y 1
L(µ|y ) =
i=1

▶ The parameter of interest µ can take values from −∞ to ∞.

▶ A conjugate prior for µ is µ ∼ N(δ, τ 2 ):
1 2
p(µ) = (2πτ 2 )− /2 e − 2τ 2 (µ−δ)
1

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

So the posterior is:

p(µ|y ) ∝ L(µ|y )p(µ)

n
1 2 1 2
e − 2σ2 (Yi −µ) e − 2τ 2 (µ−δ)
Y
∝
i=1
  X n 
1 1 2 1 2
= exp − (Yi − µ) + 2 (µ − δ)
2 σ2 τ
i=1
  X n 
1 1 2 2 1 2 2
= exp − (Yi − 2Yi µ + µ ) + 2 (µ − 2µδ + δ )
2 σ2 τ
i=1

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

So the posterior is:

  X
1 1
p(µ|y ) ∝ exp − 2 2 τ 2 Yi2 − 2τ 2 µnȳ + nµ2 τ 2
2σ τ

2 2 2 2 2
+ σ µ − 2σ µδ + σ δ
 
1 1
= exp − 2 2 µ2 (σ 2 + nτ 2 ) − 2µ(δσ 2 + τ 2 nȳ )
2σ τ
 X 
2 2 2
+ δ σ +τ Yi2
n 1h  1 n δ nȳ  io
∝ exp − µ2 2 + 2 − 2µ 2 + 2 + k
2 τ σ τ σ
(where k is some constant)

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

+ nȳ
n 1 h 1  δ 
n  2 τ2 σ2
io
Hence p(µ|y ) ∝ exp − + µ − 2µ 1
+ k
2 τ 2 σ2 τ2
+ σn2
( " 2 #)
δ
+ nȳ

1 1 n τ2 σ2
∝ exp − + µ− 1
2 τ 2 σ2 τ2
+ σn2

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

▶ Hence the posterior for µ is simply a normal distribution with

mean
δ
τ2
+ nȳ
σ2
1 n
τ2
+ σ2
and variance −1
τ 2σ2

1 n
+ =
τ 2 σ2 σ 2 + nτ 2

▶ The precision is the reciprocal of the variance.

1
▶ Here, is the prior precision . . .
τ2
n
▶ is the data precision . . .
σ2
1 n
▶ . . . and + 2 is the posterior precision.
τ 2 σ

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (variance known)

▶ Note the posterior mean E [µ|y ] is simply

1/τ 2 n/σ 2
1/τ 2
δ+ ȳ ,
+ n/σ 2 1/τ 2 + n/σ2

a combination of the prior mean and the sample mean.

▶ If the prior is highly precise, the weight is large on δ.
▶ If the data are highly precise (e.g., when n is large), the
weight is large on ȳ .
2
▶ Clearly as n → ∞, E [µ|y ] ≈ ȳ , and var [µ|y ] ≈ σn if we
choose a large prior variance τ 2 .
▶ This implies that for τ 2 large and n large, Bayesian and
frequentist inference about µ will be nearly identical.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Conjugate analysis with Normal Data (mean known)
▶ Now suppose Y1 , . . . , Yn are iid N(µ, σ 2 ) with µ known and
σ 2 unknown.
▶ We will make inference about σ 2 .
▶ Our likelihood
n
n 1 P
n − 2σ 2 [ n (Yi −µ)2 ]
2 2 −2
L(σ |y ) ∝ (σ ) e i=1

▶ Let W denote the sufficient statistic n1 (Yi − µ)2 .

P

▶ The conjugate prior for σ 2 is the inverse gamma distribution.

▶ If a r.v. Y ∼ gamma, then 1/Y ∼ inverse gamma (IG).
▶ The prior for σ 2 is
βα 2
p(σ 2 ) = (σ 2 )−(α+1) e −(β/σ ) for σ 2 > 0
Γ(α)
where α > 0, β > 0.
David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors
A Conjugate analysis with Normal Data (mean known)
▶ Note the prior mean and variance are
β
E (σ 2 ) = provided that α > 1
α−1
β2
var (σ 2 ) = provided that α > 2
(α − 1)2 (α − 2)

▶ So the posterior for σ 2 is:

p(σ 2 |y ) ∝ L(σ 2 |y )p(σ 2 )

n n
∝ (σ 2 )− 2 e − 2σ2 w (σ 2 )−(α+1) e −(
β/σ 2 )

n
n β+ w
2 −(α+ 2 +1) − σ22
= (σ ) e
▶ Hence the posterior is clearly an IG(α + n2 , β + n2 w )
distribution, where w = n1 (Yi − µ)2 .
P
Conjugate!
David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors
A Conjugate analysis with Normal Data (mean known)

▶ How to choose the prior parameters α and β?

▶ Note

[E (σ 2 )]2 [E (σ 2 )]2
 
2
α= + 2 and β = E (σ ) +1
var (σ 2 ) var (σ 2 )

so we could make guesses about E (σ 2 ) and var (σ 2 ) and use

these to determine α and β.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

▶ When Y1 , . . . , Yn are iid N(µ, σ 2 ) with both µ, σ 2 unknown,

the conjugate prior for the mean explicitly depends on the
variance:

2
p(σ 2 ) ∝ (σ 2 )−(α+1) e −β/σ
1 − 1
(µ−δ)2
p(µ|σ 2 ) ∝ (σ 2 )− 2 e 2σ 2 /s0

▶ The prior parameter s0 measures the analyst’s confidence in

the prior specification.
▶ When s0 is large, we strongly believe in our prior.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

The joint posterior for (µ, σ 2 ) is:

p(µ, σ 2 |y ) ∝ L(µ, σ 2 |y )p(σ 2 )p(µ|σ 2 )

β n
1 P 1
n 3 − σ 2 − 2σ 2 (Yi −µ)2 − 2 (µ−δ)2
2 −α− 2 − 2 2σ /s0
∝ (σ ) e i=1

n 3 − β − 1 ( Y 2 −2nȳ µ+nµ2 )− 1
(µ2 −2µδ+δ 2 )
P
= (σ 2 )−α− 2 − 2 e σ2 2σ2 i 2σ 2 /s0
 β

n 1 1 P
2 −α− 2 − 2 − σ 2 − 2σ 2 ( Yi2 −nȳ 2 )
= (σ ) e
 
1 2 2 2
× (σ 2 )−1 e − 2σ2 {(n+s0 )µ −2(nȳ +δs0 )µ+(nȳ +s0 δ )}

Note the second part is simply a normal kernel for µ.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

▶ To get the posterior for σ 2 , we integrate out µ:

Z ∞
p(σ 2 |y ) = p(µ, σ 2 |y ) dµ
−∞
n 1 1 1 P
Yi2 −nȳ 2 )]
∝ (σ 2 )−α− 2 − 2 e − σ2 [β+ 2 (

since the second piece (which depends on µ) just integrates to

a normalizing constant.
▶ Hence since −α − n2 − 12 = −(α + n
2 − 12 ) − 1, we see the
posterior for σ 2 is inverse gamma:
X
σ 2 |y ∼ IG α + n
− 12 , β + 1
(Yi − ȳ )2


2 2

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)
▶ Note that
p(µ, σ 2 |y )
p(µ|σ 2 , y ) =
p(σ 2 |y )
▶ After lots of cancellation,

p(µ|σ 2 , y ) ∝ σ −2 exp{− 2σ1 2 [(n + s0 )µ2 − 2(nȳ + δs0 )µ

+ (nȳ 2 + s0 δ 2 )]}
n h io
nȳ 2 +s0 δ 2
= σ −2 exp − 2σ2 /(n+s
1
0)
µ2 − 2 nȳn+s
+δs0
0
µ+ n+s0

▶ Clearly p(µ|σ 2 , y ) is normal:

nȳ + δs0 σ 2
 
2
µ|σ , y ∼ N ,
n + s0 n + s0

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

2

˙ ȳ , σn .
▶ Note as s0 → 0, µ|σ 2 , y ∼N
▶ Note also the conditional posterior mean is
   
n s0
ȳ + δ.
n + s0 n + s0
▶ The relative sizes of n and s0 determine the weighting of the
sample mean ȳ and the prior mean δ.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

The marginal posterior for µ is:

Z ∞
p(µ|y ) = p(µ, σ 2 |y ) dσ 2
0
Z ∞
2β + (s0 + n)(µ − δ)2
 
n 3
2 −α− 2 − 2
= (σ ) exp − dσ 2
0 2σ 2

A A
Letting A = 2β + (s0 + n)(µ − δ)2 , z = 2σ 2
⇒ σ2 = 2z and
dσ 2 = − 2zA2 dz,

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

Z ∞ −α− n − 3
A 2 2 A −z
p(µ|y ) ∝ e dz
0 2z 2z 2
Z ∞ −α− n − 1
A 2 2 1 −z
= e dz
0 2z z
n 1
Z ∞ n 1
−α− 2 − 2
∝A z −α− 2 − 2 −1 e −z dz
0

This integrand is the kernel of a gamma density and thus the

integral is a constant. So

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

A Model for Normal Data (mean and variance both
unknown)

n 1
p(µ|y ) ∝ A−α− 2 − 2
 − 2α+n+1
2
2
= 2β + (s0 + n)(µ − δ)
" #− 2α+n+1
2
(s0 + n)(µ − δ)2
∝ 1+
2β

which is a (scaled) noncentral t kernel having noncentrality

parameter δ and degrees of freedom n + 2α.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 1: Midge Data

▶ Example 1: Y1 , . . . , Y9 are a random sample of midge wing

iid
lengths (in mm). Assume the Yi′ s ∼ N(µ, σ 2 ).
▶ Example 1(a): If we know σ 2 = 0.01, make inference about µ.
(See R example)
▶ A Bayesian point estimate for the population mean midge
wing length is the posterior mean, 1.806 mm.
▶ A 95% credible interval for µ is (1.741, 1.871), so with
posterior probability 0.95, the population mean midge wing
length is between 1.741 and 1.871 mm.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 1: Midge Data

▶ Example 1(b): Make inference about µ and σ 2 , both

unknown (see R example).
▶ This requires choosing the hyperparameters α and β of the
inverse gamma prior on σ 2 .
▶ 95% credible interval for σ 2 : (0.012, 0.028), with posterior
median 0.0188.
▶ To approximate the posterior distribution for µ: We will
randomly generate many values from the posterior distribution
of σ 2 .
▶ Then we will generate many values from the posterior of µ,
given each respective generated value of σ 2 .
▶ 95% credible interval for µ: (1.727, 1.90), with posterior
median 1.81 mm.

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors

Example 2: Brain Data

▶ The textbook has an example of Bayesian inference about the

mean hippocampal volume of the brain in a population of
college football players who have a history of concussions.
▶ Example 2: Y1 , . . . , Y25 are a random sample of hippocampal
volumes (in cm3 ) of such football players. Assume the
iid
Yi′ s ∼ N(µ, σ 2 ).
▶ Example 2(a): If we know σ = 0.5 ⇒ σ 2 = 0.25, make
inference about µ. We assume a N(6.5, 0.42 ) prior on µ.
▶ The posterior mean is 5.78 cm3 . With posterior probability
0.95, the mean hippocampal volume of the brains for the
population of concussed players is between 5.59 and 5.97cm3 .

David B. Hitchcock E-Mail: hitchcock@stat.sc.edu Chapter 5: More Conjugate Priors