Bayesian Analysis (2007) 2, Number 1, pp.

137–166

A Quantitative Study of Quantile Based Direct

Prior Elicitation from Expert Opinion
Dipak K. Dey∗ and Junfeng Liu†

Abstract. Eliciting priors from expert opinion enjoys more efficiency and re-
liability by avoiding the statistician’s potential subjectivity. Since elicitation on
the predictive prior probability space requires too-simple priors and may be bur-
dened with additional uncertainties arising from the response model, quantitative
elicitation of flexible priors on the direct prior probability space deserves much
attention. Motivated by precisely acquiring the shape information for the gen-
eral location-scale-shape family beyond the limited and simple location-scale fam-
ily, we investigate multiple numerical procedures for a broad class of priors, as
well as interactive graphical protocols for more complicated priors. We highlight
the quantile based approaches from several aspects, where Taylor’s expansion is
demonstrated to be an efficient approximate alternative to work on the regions
in which the shape parameter is highly sensitive. By observing inherent associa-
tions between the scale and shape parameters, we put more weight on practical
solutions under a proper sensitivity index (SI) rather than presumability. Our
proposed methodology is demonstrated through skew-normal and Gamma hyper-
parameter elicitation where the shape parameter is numerically solved in a stable
way. The performance comparisons among different elicitation approaches are also
provided.
Keywords: location parameter, prior elicitation, quantile, scale parameter, shape
parameter, skewness, Taylor’s expansion

1 Introduction
Bayesian analysis is markedly recognized by the subjective probability belief, or quan-
titative a priori description of unknown parameter θ. Without external support statis-
ticians can only implement the whole computational process at his/her own will, say
conjugate priors, Jeffreys’ priors and others. On the other hand, expert opinion may
be helpful when we investigate new, rare, complex or poorly understood phenomena.
Multivariate-normal related prior elicitation on the predictive prior space by requesting
response summaries from experts was developed by Kadane, et al (1980), Garthwaite
and Dickey (1988), Al-Awadhi and Garthwaite (1998) among others. However, those
algorithms were limited to simple normal linear or AR(1) time series models. On the
other hand, direct non-informative prior elicitation was discussed through piecewise
conjugate priors (Meeden, 1992), entropy based priors (Jaynes, 1968, 1983), mixture
of natural conjugate priors (Dalal and Hall, 1983) and others. Quantile based univari-
∗ Department
of Statistics, University of Connecticut, Storrs, CT, http://www.stat.uconn.edu/~dey
∗ Department of Statistics, West Virginia University, Morgantown, WV,
mailto:[email protected]

c 2007 International Society for Bayesian Analysis ba0005

138 Quantile based direct prior elicitation

ate prior elicitation for simple cases, say symmetric ones, was studied by Peterson and
Miller (1964), Garthwaite and Dickey (1985), O’Hagan (1998) among others. A recent
comprehensive review on probability elicitation was written by Garthwaite, Kadane and
O’Hagan (2005). Berger (1985, Chapter 3) also discussed subjective prior determination
on the direct prior space by showing that a lack of sufficient tail information in con-
tinuous parameter space causes much difficulty for most of the approaches in practice,
including the “histogram” approach, the “relative likelihood” approach, the entropy
based method, and even the most used “matching a given functional form” method
which often needs prior moments and others. Berger (1985, Chapter 3) envisioned that
a quantile based approach poses as a better method since estimation of probabilities of
regions are more attractive than working on moments. However, two situations deserve
caution in application of quantile approaches: disagreement among multiple quantiles
and incidental matching by multiple functional forms (Berger, 1985, Chapter 3). The
key point to ease these concerns is to efficiently and precisely recover flexible parametric
priors in a quantitative way other than those weak symmetric ones in order to imple-
ment the “sketching” principle (Berger, 1985, Chapter 3) for the downstream graphical
verification. In this paper, we propose efficient quantile based approaches for direct
prior elicitation with quantitative examples, and study a broad class of parametric pri-
ors with a focus on asymmetric cases, which have yet received little attention.

The paper is organized as follows: Section 2 develops approximate and exact quantile
based numerical elicitation algorithms for general symmetric and skewed priors; Section
3 introduces Taylor’s expansion as an efficient tool for joint scale-shape parameter elic-
itation in sensitive regions and makes comparison among multiple expansion options
with skew-normal and Gamma distributions as examples; Section 4 provides interactive
graphical display procedures, especially for more complicated Student’s t related prior
elicitation; and Section 5 concludes the paper with future directions.

2 Symmetric and Asymmetric Prior Elicitation

2.1 Normal Prior
Garthwaite and Dickey (1985) evaluated double- and single-bisection methods for sub-
jective probability assessment in a location-scale family, where single-bisection means
median plus one quantile, double-bisection means 0.25, 0.50 and 0.75 quantiles: q0.25 ,
q0.50 and q0.75 . They showed that, the double-bisection method is more favorable for
most densities. Throughout this paper, we will use the notation IQR to represent the
inter-quartile range q0.75 − q0.25 and additional subscripts to specify certain parametric
families. More specifically, Zq represents the IQR of standard normal distribution. Sup-
pose further q0.75 − q0.50 ≈ q0.50 − q0.25 , then the prior mean and the standard deviation
are estimated by

IQR
µ̂N = q0.50 and σ̂N = respectively. (1)
Zq
Dey and Liu 139

Note that, normal distribution is symmetric and shape parameter free. If both the ex-
perimenter and the expert agree on the symmetry, then only two quantiles are necessary.

2.2 Student’s t Prior

If symmetry is agreed on for Student’s t prior, then three non-redundant quantiles are
necessary to take into account the degrees of freedom ν. The elicitation by Kadane,
et al (1980) requested 0.50, 0.75 and 0.9375 quantiles: q0.50 , q0.75 and q0.9375 . The
median q0.50 can be taken as the location parameter µ and a(x) is defined as (q0.9375 -
q0.50 )/(q0.75 -q0.50 ). Since both the numerator and the denominator are independent of
the center and the ratio is independent of the spread, thus a(x) depends on degrees
of freedom ν only (a monotonic function of ν) and ν is determined by checking the
look-up table of ν vs. a(x). For the spread of Student’s t-distribution, after eliciting
the degrees of freedom ν and obtaining the corresponding standard upper quartile t ν,.75 ,
S(q)=(q.75 −q.50 )2 /t2ν,.75 is used to elicit the scale parameter σ. This idea can be applied
to general location-scale-shape family (Sections 2.4, 2.5 and 2.6).

2.3 Log-normal Prior

Among the scarcity of literature on asymmetric prior elicitation, Garthwaite (1989)
modeled asymmetry by a “split-normal” distribution, and O’Hagan (1998) used 61 , 36
and 56 quantiles to elicit a log-normal prior for positive skew distribution. For simplicity,
we denote q0.25 , q0.50 and q0.75 to be the lower quartile, median and upper quartile of
log-normal distribution, the following proposition provides parameter solution for log-
normal case.

Proposition 1 If X has a log-normal distribution, i.e., lnX ∼ N (µ, σ 2 ), then the

2
variance D(X) = q0.50 r2 (r2 − 1) and the mean E(X) = rq0.50 , where q0.50 = eµ is the
2
median of X, r = exp( ln (q0.75 /q0.25 ) ), Z is the IQR for standard normal distribution.
2Zq2 q

The proof is done by observing E(etlnX ) = E(X t ), followed by simplifications. The

parameter elicitation is straightforward based on 0.25, 0.50 and 0.75 quantiles when
q0.75 − q0.50 > q0.50 − q0.25 . A limitation of log-normal distribution is the strictly
positive domain and the inherent constraint logq0.75 -logq0.50 =logq0.50 -logq0.25 , which is
obviously too restrictive a rule for the expert to follow. Thus the log-normal discussion
is trivial. We now study general asymmetric prior elicitation in the following sections,
where special attention is paid to the shape parameter elicitation.

2.4 Skew-normal Prior

Definition 1: Skew-normal distribution We construct a skew-normal random vari-
able X by cZ + +  (which spans the whole real domain), where c is a constant, Z + is
140 Quantile based direct prior elicitation

2
a folded normal random variable, e.g., the positive part of N (0, σZ ) random variable, 
is a normal random variable with mean µ and variance σ , and Z + is independent of .
2

The pdf of X is

f (x|c, µ, σZ , σ ) = √ 2
2 +σ 2
φ( √ 2x−µ
2 +σ 2
)Φ( cσ
σ
Z √ (x−µ)
2 2
),
c2 σ Z  c σZ  c σZ +σ2

p
where σ (= c2 σZ 2 + σ 2 ) is the scale parameter, λ=( cσZ ) is the skewness (shape) pa-
 σ
rameter, and µ is the location parameter. The derivation is given in Sahu, Dey and
Branco (2003). When c = 0, we have an ordinary normal distribution. Under reparam-
eterization Y = X−µσ , we get the simplified density function of f (y|λ) = 2φ(y)Φ(λy)q
which is the skew-normal density function given by Azzalini (1985), with E(Y ) = π2 δ
q
and D(Y ) = 1 − π2 δ 2 , where δ = √1+λλ
2
. In the general case, E(X) = µ + σ 2
π δ and
D(X) = σ 2 (1 − π2 δ 2 ).

R Qq (λ)
Theorem 1 Assume −∞ 2φ(x)Φ(λx)dx = q; where, q th quantile Qq (λ) is de-
termined by λ, the skewness parameter. Then (i) Qq (−λ) = −Q1−q (λ); (ii) ∀λ ∈
φ(λQq (λ))
(−∞, 0], q ∈ [0, 1], Q0q (λ) = (1+λ2 )Φ(λQ q (λ))
.

The proof is given in the appendix. We can see that, ∀λ ∈ (−∞, 0], Q0.50 (−λ) =
−Q0.50 (λ) and Q00.50 (−λ) = Q00.50 (λ); the median of cZ + +  is not equal to the median
of cZ + plus the median of , since Qq (λ) is not a linear function of λ. The boundary
condition q determines the individual solution from the nonlinear ordinary differential
φ(λQq (λ))
equation Q0q (λ) = (1+λ2 )Φ(λQ q (λ))
. Conditional on the skewness (shape) parameter λ, we
use the notation IQRR(λ) to represent the inter-quartile range ratio of (q0.75,λ − q0.50,λ )
to (q0.50,λ − q0.25,λ ), a monotonic function of λ (Section 2.2). Presumably, the skewness
(shape) parameter λ can be obtained by checking the look-up table of the ratio of
IQRR(λ) to λ. However, the top-right panel of Figure 1 shows that, a small change of
the IQRR(λ) leads to a large change of hyperparameter λ when λ is close to zero, say
|λ| < 1, which is not our expectation. Thus “IQRR(λ) vs. λ” is not a good elicitation
method for λ. To formalize this scenario, we define the sensitivity index (SI) as

∂ hyperparameter/∂ elicitation input,

for which we expect a moderate magnitude (∼1). For example, in the top-right panel
of Figure 1, the elicitation input is IQRR and the hyperparameter is λ. The guideline
is to work on “elicitation input vs. hyperparameter” curves with moderate derivative
or sensitivity index (SI), say around 1. We first consider the regions in which the
shape parameter is not sensitive (|λ| > 1) by the following one-way procedure, and the
sensitive regions will be discussed on in Section 3. The rationale for choosing 1 as the
λ critical value will be described later in Remark 1.
Dey and Liu 141

1. The skewness parameter λ can be obtained by the monotonic relationship between

λ and IQRR(λ) (top-right, Figure 1).
2. The scale parameter σ can be obtained by (q0.75,λ − q0.50,λ ) vs. λ plot (bottom-
left, Figure 1): find out (q0.75,λ − q0.50,λ ) at elicited λ, and the elicited σ is thus
(q0.75 − q0.50 )/(q0.75,λ − q0.50,λ ). Or by (q0.75,λ − q0.25,λ ) vs. λ plot (bottom-
right, Figure 1): find out (q0.75,λ − q0.25,λ ) at elicited λ, and the elicited σ is thus
(q0.75 − q0.25 )/(q0.75,λ − q0.25,λ ).
3. The location parameter µ is simply q0.25 −σq0.25,λ , q0.50 −σq0.50,λ , or q0.75 −σq0.75,λ .

2.5 Skew Student’s t Prior

Definition 2: Skew Student’s t distribution We construct a skew Student’s
t random variable X by cT + +  (which spans the whole real domain), where c is
a constant, T + is a folded Student’s t random variable, that is, the positive part of
T (0, 1, ν) random variable,  is a Student’s t random variable with mean µ and scale η,
both of them have degrees of freedom ν, and T + is independent of . The pdf of X is

(x−µ)2
Γ( ν+1 (x−µ)2 − ν+1 ν+ η2 +c2
2 )
)− 2 ηc √(x−µ)
2 1
f (x|c, µ, η, ν) = √ 1 (1 + ν(η 2 +c2 ) )
2 T
1,ν+1 [( ν+1 ],
c2 +η 2 Γ( ν2 )(νπ) 2 2
η +c 2

p
where σ (= c2 + η 2 ) is the scale parameter, λ (= ηc ) is the skewness (shape) parameter,
µ is the location parameter, and ν is the degrees of freedom. The derivation is given
in Sahu, Dey and Branco (2003). When c = 0, we have an ordinary Student’s t-
distribution. Under reparameterization of Y = X−µ σ , we get the simplified density
function of

Γ( ν+1
2 ) y 2 − ν+1 ν+y 2 − 1
f (y|λ, µ, σ, ν) = 2 1 (1 + ν )
2 T
1,ν+1 [( ν+1 ) 2 λy].
Γ( ν2 )(νπ) 2

A different attribute for skew Student’s t-distribution is that, the nominal location-scale-
shape family is conditional on the degrees of freedom ν by observing that ν cannot be
combined with λ to form a composite shape parameter. From Figure 2, we note that, the
inter-quartile information of skew Student’s t-distribution highly depends on the degrees
of freedom ν; when ν reaches a value to approach the limiting normal distribution, the
sensitivity index (SI) in the top-right panel gets substantially larger than 1, as skew-
normal does (Figure 1). Conditioning on ν, the elicitation procedure is the same as that
in Section 2.4 when λ is not close to 0, say |λ| > 1, and the sensitive regions will be
discussed on in Section 3.

2.6 Normal-exponential Prior

The following normal-exponential distribution is dated back to Aigner et al (1977).
142 Quantile based direct prior elicitation

Definition 3: Normal-exponential distribution A normal-exponential random

variable X is constructed by λE1 +  (which spans the whole real domain), where λ is
a constant, E1 is an exponential random variable with mean 1,  is a normal random
variable with mean µ and variance η 2 , and E1 is independent of . The pdf of X has
closed form

2
α
σ exp(− α2 ) exp(−α x−δ x−δ
σ )Φ( σ ).

The mean is δ + σ( α1 − α), where σ (=η) is the scale parameter, α (= λη ) is the shape
2
parameter and δ (=µ + ηλ ) is the location parameter. The proof is a matter of in-
tegration. The superficial skewness is produced by the normal part, other than the
exponential part. As α approaches zero, the normal-exponential density becomes a flat
degenerate density; as α exceeds two, the density becomes an approximate normal den-
sity (Figure 3), where leftward displacement induced by the combination of shape and
scale parameters is observed. Although the curves are quite steep in the IQR vs. α
plot, reasonable sensitivity index (SI) is reached after taking 5th root (bottom-middle
and bottom-right, Figure 4). The parameter elicitation procedure at α > 0.15, where
the IQRR is substantially away from the limit 1.7107 as α departs from 0, is similar
to skew-normal case (Section 2.4). We observe that, the skewness parameter elicitation
is sensitive for skew-normal and normal-exponential priors when the true value is close
to zero, while it is less sensitive for skew Student’s t case under small degrees of freedom.

A more rigorous mathematical statement to support the above elicitation proce-

dures is that any three distinct quantiles uniquely determines the location-scale-shape
parameter set. For two density functions belonging to the same location-scale-shape
family with two sets of parameters, we have observed that one or two density function
crossovers occurs if these two sets of parameters are not exactly equal to each other. If
two distinct density functions have two identical sets of three quantiles, then at least
four density function crossovers are required, since there is at most one common quantile
between two consecutive crossovers, and there is no common quantile either less than
the first crossover or greater than the last crossover. The open problem is that there are
at most three density function crossovers under any two sets of different parameters.

2.7 Approximate Scale Parameter Elicitation

Now we propose an elicitation scheme based on approximation for Student’s t-distribution.
The expert provides three quantiles: q0.25 , q0.50 and q0.75 , but is not sure about normal-
ity assumptions. Suppose Fλ,µ,σ is the cumulative density function of the distribution of
expert, a member of location-scale-shape (µ-σ-λ) family, and p is the inter-quartile prob-
ability between q0.25 and q0.75 , then by first order Taylor’s expansion at g(?)=µ+σg(λ),
where g(λ) is some characteristic point (median, mean, mode, etc) of the standardized
Dey and Liu 143

density fλ,0,1 with µ = 0 and σ = 1. We have

(1)
Fλ,µ,σ (q0.75 ) ≈ Fλ,µ,σ (g(?)) + fλ,µ,σ (g(?))(q0.75 − g(?)) + 12 fλ,µ,σ (g(?))(q0.75 − g(?))2 · · ·
(1) .
Fλ,µ,σ (q0.25 ) ≈ Fλ,µ,σ (g(?)) + fλ,µ,σ (g(?))(q0.25 − g(?)) + 21 fλ,µ,σ (g(?))(q0.25 − g(?))2 · · ·
(2)

Consequently, if we take 12 = Fλ,µ,σ (q0.75 ) − Fλ,µ,σ (q0.25 ) ≈ fλ,µ,σ (g(?))(q0.75 − q0.25 )

= σ1 fλ,0,1 (g(λ))IQR, then a rough approximation for scale parameter elicitation is
2IQRfλ,0,1 (g(λ)).

Theorem 2 Under Taylor’s expansion approximation (2), denote

P∞ 1 (k)
∆= k=1 k+1 fλ,0,1 (g(λ))[(q0.75,λ − g(λ))k+1 − (q0.25,λ − g(λ))k+1 ].

Then

−2∆
(i) the relative error of the scale parameter elicitation by 2IQRfλ,0,1 (g(λ)) is 1−2∆ ,
a function only associated with the shape parameter λ;

1
(ii) ∆ = 2 − fλ,0,1 (g(λ))IQR(λ), where IQR(λ) is the standardized IQR under Fλ,0,1 .

The proof is given in the appendix. For g(λ) = mode function M (λ), mean function
E(λ) or median function q0.50 (λ), we have the following observations:

1
2 −∆
Corollary 1 IQR(λ) = fλ,0,1 (g(λ)) .

Corollary 2 Under skew-normal distribution (Section 2.4),

P∞ q
1 (k)
limλ→0 k=1 k+1 fλ,0,1 (g(λ))[(q0.75,λ − g(λ))k+1 − (q0.25,λ − g(λ))k+1 ] = 21 (1 − 2
π Zq ).

Corollary 3 Under skew Student’s t-distribution with degrees of freedom ν (Sec-

P∞ 1 (k)
tion 2.5), limλ→0 k=1 k+1 fλ,0,1 (g(λ))[(q0.75,λ − g(λ))k+1 − (q0.25,λ − g(λ))k+1 ]= 21 (1 −
Γ( ν+12 )
√ t ),
where tν,q is the standardized inter-quartile range for Student’s t-distribution
Γ( ν2 ) νπ ν,q
with ν degrees of freedom.

We point out that (2) seems to work for many flexible location-scale-shape classes
(details later, in Propositions 3, 4). Suppose f is the probability density function of
general Student’s t-distribution with ν degrees of freedom (shape parameter) and inter-
144 Quantile based direct prior elicitation

quantile range IQR, then we get the following approximation (after neglecting the error)

p 1 Γ( ν+1
2 ) (q0.50 − µ)2 − ν+1
=√ ν (1 + ) 2 , (3)
IQR νπσ Γ( 2 ) νσ 2

where µ is the location parameter and σ is the scale parameter. Then µ̂ = q0.50 is the
estimate of the prior mean (location parameter), and the approximate estimate of σ can
be obtained as

IQR Γ( ν+1
2 )
σ̂ = √ . (4)
p νπ Γ( ν2 )

Proposition 2 Scale-parameter elicitation is implemented by scaling the requested

IQR by certain value, or in accordance with the underlying distribution assumption.
Denote the inter-quantile range Tq,ν of Student’s t-distribution with degrees of freedom
ν to be T0.75,ν -T0.25,ν , then the ratio of the approximate scale parameter by (4) to the
exact scale parameter from Student’s t version q of (1) is a monotonically decreasing func-
2
tion of the degrees of freedom ν with limit Zq π =1.0763 (Zq =1.349).

We now describe an exploratory scale parameter elicitation for Student’s t-distribution

IQR
with degrees of freedom ν (shape parameter), the exact scale parameter is Tq,ν , which
IQR
goes to Zq as ν increases. By Stirling’s formula, we can show that the approximately
IQR
elicited scale parameter (4) goes to p√2π , where the inter-quartile probability p=0.50.
Although (4) seems to be a rough approximation, from Figure 6 it turns out to be an
efficient estimator retaining proportional high-fidelity (stable relative error) as degrees
of freedom ν increases. Of the two relative error curves, one is between the approximate
and the exact scale parameters both under Student’s t-distribution (solid curve), the
other is between the approximate scale parameter under normal distribution and the
exact scale parameter under Student’s t-distribution (dotted curve), both approach the
limit ratio and the former one approaches the limit more quickly. Consequently, when
the degrees of freedom ν is greater than 5, the approximation (4) under Student’s t is
realistic after scaling by 1.0763. However, the crude approximation under the normal
assumption is worse for most values of degrees of freedom, say less than 20. Taylor’s
expansion based approximate scale parameter elicitation may be quite efficient in the
sense of a proportional high-fidelity when the shape parameter (say degrees of freedom
ν for Student’s t-distribution) approaches certain limit (say positive infinity). This ob-
servation will guide us to elicit scale and shape parameters jointly in certain regions
by means of flexible Taylor’s expansion based iterations (Section 3). From exploratory
point of view, we may potentially elicit the degrees of freedom reversely conditional on
σ̂, although intuitively they are independent parameters without any inherent associ-
pσ̂ pσ̂
ation. The ν ∼ IQR plot derived from (4) is given in Figure 7, where IQR range of
∼(0.2, 0.396) gives sensitive ν (shape parameter) estimates in (0,10). In other words,
Dey and Liu 145

other values will lead to the normal distribution. Theoretically, If we assume symmetry
a priori, then eliciting degrees of freedom ν and the scale parameter σ concurrently
is not possible, since these two parameters are mutually determined within a certain
domain for σ (Figure 7).

For only scale-parameter model, no approximation is needed. For example, if f is

the pdf of an exponential distribution with only scale parameter σ, then σ̂ should be
obtained by solving the equation

IQR
q0.50 σ − lnσ = ln( ). (5)
p

Although approximation (4) is shown to be an efficient scale elicitation for symmetric

Student’s t priors, a similar procedure may not work for some skewed cases. For the
log-normal case in Proposition 1, assume we are given q0.25 , q0.50 and q0.75 for X which
√
forms a symmetric normal distribution for ln(X) requiring q0.50 = q0.25 × q0.75 . We use
first order Taylor’s expansion to make a rough approximation: p = IQRf (q 0.50 ), where
f is the pdf of a log-normal distribution, then µ̂=ln(q0.50 ) and

IQR IQR
p= √ , thus σ̂ = √ . (6)
q0.50 2πσ q0.50 2πp

The exact σ elicitation under log-transformation is Z∆q , where ∆ is ln( qq0.75 ), and (6) be-
q 0.25

comes π2 (e∆ − e−∆ ). The ratio of the approximate and exact scale elicitation is thus
q 2
approximately π2 Zq (1 + ∆6 ) with quadratic increment of ∆. The weakness of Taylor’s
expansion may be due to the fact that, log-normal distribution is not a location-scale
family member.

IQR
For the skew-normal case (Section 2.4), the exact scale elicitation is IQR(λ) , where
IQR(λ) is the IQR for the standard skew-normal distribution with location parameter
0, scale parameter 1 and skewness (shape) parameter λ.

q
2
. If we apply Taylor’s expansion at the location parameter µ, then σ̂ = π IQR,
thus location based expansion makes use of no skewness
q information. The ratio
2
of the approximate to the exact scale elicitation is π IQR(λ).

q
. If we apply Taylor’s expansion at mean Eλ = π2 √1+λ λ
2
, then σ̂ is [2φ(Eλ )Φ(λEλ )]
IQR/p, where p=0.50 and Eλ is the mean of the standardized skew-normal distri-
bution with location parameter 0, scale parameter 1 and skewness parameter λ.
The ratio of the approximate to the exact scale elicitation is IQR(λ)[2φ(E λ )Φ(λEλ )]/p.
146 Quantile based direct prior elicitation

. If we apply Taylor’s expansion at mode Mλ , then σ̂ is IQR[2φ(Mλ )Φ(λMλ )]/p,

where p=0.50 and Mλ is the mode of the standardized skew-normal distribution
with location parameter 0, scale parameter 1 and skewness parameter λ. The ratio
of the approximate to the exact scale elicitation is [2φ(Mλ )Φ(λMλ )] IQR(λ)/p.
. If we apply Taylor’s expansion at median q0.50,λ , then σ̂ is [2φ(q0.50,λ )Φ(λq0.50,λ )]
IQR/p, where p=0.50 and q0.50,λ is the median of the standardized skew-normal
distribution with location parameter 0, scale parameter 1 and skewness parameter
λ. The ratio of the approximate to the exact scale elicitation is [2φ(q0.50,λ )Φ(λq0.50,λ )]
IQR(λ)/p.

Remark 1 The scale parameter σ elicitation based on (2) expanded at the location
parameter µ leads to identical results under naive normal or skew-normal distributions.
If we ignore the skewness completely by assuming naive normal, then the elicited scale
IQR
parameter σ is Zq accordingly. All of these options are compared by a “whiskers” plot
(Figure 8). Except for the native normal based approximation, all skew-normal based
approximations are visually undistinguishable as horizontal segments expanding from
∼-1 to ∼1, which is exactly where λ is highly sensitive to the IQRR(λ) (the top-right
panel of Figure 1), also represented by a horizontal segment expanding from ∼-1 to ∼1.
Thus the sensitive region and the critical value of λ are to be identified visually. Over-
all, the performance ordering may be Taylor’s expansion at: median, mean, mode, and
location. We envision that this observation also works for skew Student’s t conditional
on the degrees of freedom ν. Thus quantile (probability) based prior elicitation could
increase the potential flexibility and efficiency.

Proposition 3 For the skew-normal distribution,

q the ratio of the approximate scale
parameter to the exact scale parameter goes to π2 Zq =1.0763 as λ approaches 0.

IQR
For the normal-exponential case (Section 2.6), the exact scale elicitation is IQR(α) ,
where α is the shape parameter, and IQR(α) is the IQR for the standard normal-
exponential distribution with location parameter 0 and scale parameter 1. Again,
p=0.50.
2
α
. If we apply Taylor’s expansion at the location parameter δ, then σ̂= 2p exp(− α2 )IQR,
in contrast to skew-normal case, location based expansion incorporates the shape
α2
information. The ratio of the approximate to the exact scale elicitation is 2p
α exp( 2 )
IQR(α).
2
. If we apply Taylor’s expansion at mean Eα =σ( α1 − α), then σ̂ = αp exp( α2 −
1)Φ( α1 − α)IQR. The ratio of the approximate to the exact scale elicitation is
α α2 1
p exp( 2 − 1)Φ( α − α)IQR(α).

. If we apply Taylor’s expansion at mode Mα which does not have an explicit for-
2
mula, then σ̂ = αp exp[−( α2 + αMα )]Φ(Mα )IQR. The ratio of the approximate to
Dey and Liu 147
2
α
the exact scale elicitation is p exp[−( α2 + αMα )]Φ(Mα )IQR(α).
2
. If we apply Taylor’s expansion at median q0.50,α , then σ̂ is αp exp[−( α2 + αq0.50,α )]
Φ(q0.50,α )IQR. The ratio of the approximate to the exact scale elicitation is
α α2
p exp[−( 2 + αq0.50,α )]Φ(q0.50,α )IQR(α).

Remark 2 The comparisons are given by a “whiskers” plot (Figure 9), where the
location or median based approximation performs more stably and retains more pro-
portional high-fidelity (≈1.09) than others at small α. Note that the location based
approximation takes into account the shape information, which is a different attribute
from skew-normal and skew Student’s t. The performance ordering may be Taylor’s ex-
pansion at: median, location, mean and mode at small α; at large α, the location based
approximation cannot be used since the shape and scale parameters jointly induces
large location displacement, and all other approximations have visually undistinguish-
able performances.

Proposition 4 For normal-exponential distribution, the ratio of the approximate

scale to the exact scale approaches 1.0763 as α approaches positive infinity.

3 Elicitation on Shape-Parameter-Sensitive Regions

3.1 Skew-normal case
Complementary to the discussion on skew-normal in Section 2.4, we propose several
iterative algorithms for |λ| ≤1 under alternative moderate sensitivity index (∼1). If the
original procedure in Section 2.4 is followed, then λ will be highly sensitive to minor
changes of IQRR(λ); while the approximate σ elicitation based on first order Taylor’s
expansion
q at mean (with a favorable explicit expression, Section 2.4) retains propor-
2
tional ( π Zq ) high-fidelity (the middle panel of Figure 8) for that very λ range.

 Iteration Based on Taylor’s Expansion at Median

1. Start with current λ, apply the first order Taylor’s expansion at median q0.50,λ with
location parameter 0 and scale parameter 1. Then σ̂ is IQR[2φ(q0.50,λ )Φ(λq0.50,λ )]/p,
the elicited scale parameter is √ σ̂2 , and the standardized (q0.75,λ -q0.50,λ ) is (q0.75 -
π Zq

q0.50 )/( √ σ̂2 ).

π Zq

2. The skewness parameter λ can be obtained by (q0.75,λ −q0.50,λ ) vs. λ plot (bottom-
left, Figure 1) under moderate sensitivity index (∼1). Note that, one-to-two
matching around small λ is not an issue, since the sign of λ is correctly kept.

3. Go to 1 until convergence (complete λ and σ).

148 Quantile based direct prior elicitation

4. The location parameter µ is simply q0.25 −σq0.25,λ , q0.50 −σq0.50,λ or q0.75 −σq0.75,λ .

 Iteration Based on Taylor’s Expansion at Mean

q
1. Start with current λ, apply the first order Taylor’s expansion at mean Eλ = π2 √1+λ λ
2

with location parameter 0 and scale parameter 1. Then σ̂ is IQR[2φ(Eλ )Φ(λEλ )]/p,
the elicited scale parameter is √ σ̂2 , and the standardized (q0.75,λ -q0.50,λ ) is (q0.75 -
π Zq

q0.50 )/( √ σ̂2 ).

π Zq

2. The skewness parameter λ can be obtained by (q0.75,λ −q0.50,λ ) vs. λ plot (bottom-
left, Figure 1) under moderate sensitivity index (∼1). Note that, one-to-two
matching around small λ is not an issue, since the sign of λ is correctly kept.

3. Go to 1 until convergence (complete λ and σ).

4. The location parameter µ is simply q0.25 −σq0.25,λ , q0.50 −σq0.50,λ or q0.75 −σq0.75,λ .

 Iteration Based on Taylor’s Expansion at Mode

q
1. Start with current λ, apply the first order Taylor’s expansion at mode Mλ = π2 √1+λ λ
2

with location parameter 0 and scale parameter 1. Then σ̂ is IQR[2φ(Eλ )Φ(λEλ )]/p,
the elicited scale parameter is √ σ̂2 , and the standardized (q0.75,λ -q0.50,λ ) is (q0.75 -
π Zq

q0.50 )/( √ σ̂2 ).

π Zq

2. The skewness parameter λ can be obtained by (q0.75,λ −q0.50,λ ) vs. λ plot (bottom-
left, Figure 1) under moderate sensitivity index (∼1). Note that, one-to-two
matching around small λ is not an issue, since the sign of λ is correctly kept.

3. Go to 1 until convergence (complete λ and σ).

4. The location parameter µ is simply q0.25 −σq0.25,λ , q0.50 −σq0.50,λ or q0.75 −σq0.75,λ .

 Iteration Based on the Inter-quartile Ranges

1. Start with current λ, (q0.75,λ − q0.25,λ ) vs. λ plot (Figure 1) with moderate sensi-
tivity index (∼1). The elicited σ is (q0.75 − q0.25 )/(q0.75,λ − q0.25,λ ).

2. The skewness parameter λ can be obtained by (q0.75,λ −q0.50,λ ) vs. λ plot (bottom-
left, Figure 1) under moderate sensitivity index (∼1).

3. Go to 1 until convergence (complete λ and σ).

4. The location parameter µ is simply q0.25 −σq0.25,λ , q0.50 −σq0.50,λ or q0.75 −σq0.75,λ .
Dey and Liu 149

Remark 3 We create the (q0.25,λ , q0.50,λ , q0.75,λ ) vs. λ look-up table (the top-left
panel of Figure 1), where the λ step is 0.01, the iteration involves obtaining the optimal
λ given (q0.75,λ − q0.50,λ ) and the optimal (q0.75,λ − q0.25,λ ) given λ, till convergence.
The one-way elicitation of λ followed by σ (Section 2.4) is replaced by a two-way joint
λ and σ elicitation. The elicitation comparison is given in Table 1, and Figures 11
through 14, where ratios are calculated at all declared 50% quantiles (from 0.01 to 0.20
with step 0.01). The IQR, mean and median based Taylor’s expansion iterations are
consistent, while mode based Taylor’s expansion iteration is likely to be less satisfactory.
Compared to the IQR based iteration, we recommend the explicit mean based Taylor’s
expansion in view of numerical stability. In the top-right panel of Figure 1, when the
IQRR varies within a very short range around 1.0, the λ ranges among [-1,1], leading
to a sensitivity index with a huge magnitude. Under iterative elicitation procedures,
highly sensitive regions are avoided, and the applied regions in the IQR(λ) vs. λ plots
(the bottom-left and bottom-right panels of Figure
q 1) show a workable sensitivity
q index
3
(∼1). By observing that, the standard mean π2 √1+λ λ
2
has a derivative π2 (1 + λ2 )− 2
(∼1 at small λ) and conclusion (ii) in Theorem 2, we conclude that, iteration by Taylor’s
expansion at mean will work well under moderate sensitivity index. Generally, ∂shape
parameter/∂standardized IQR needs to be moderate in order to apply the iterative
algorithms, similar procedures will apply to the normal-exponential prior and skew
Student’s t prior conditional on the degrees of freedom ν. Because the standard IQRs are
tabulated by numerical integration which is subject to accuracy evaluation for irregular
skewed classes, we prefer an explicit expression involved in the algorithm, such as mean
based Taylor’s expansion. One piece of evidence is seen in Table 1, where there is a
switching between two elicited λ’s at q0.50 =0.07 and q0.50 =0.08.

3.2 Gamma Case

Gamma distribution can be generalized into a location-scale-shape family member as
x−µ
fα,β,µ = β1 Γ(α)
1
( x−µ
β )
α−1 − β
e 1[x≥µ] , whose quantile plots are given in Figure 5. We can
find that the sensitivity index is around 100 by using (q0.75,α − q0.50,α )/(q0.50,α − q0.25,α )
vs. α plot when α ≥1.5, and the sensitivity index is decreased to less than 10 by using
(q0.75,α − q0.50,α ) or (q0.50,α − q0.25,α ) vs. α plot. For α ≥1.5, we propose the following
iteration procedure.

Iteration Based on the Inter-quartile Ranges

1. Start with current α, (q0.75,α − q0.50,α ) vs. α plot (Figure 5) with sensitivity index
∼10, the elicited β is (q0.75 − q0.50 )/(q0.75,α − q0.50,α ).
2. The shape parameter α can be obtained by (q0.50,α − q0.25,α ) vs. α plot, where
(q0.50,α − q0.25,α )=(q0.50 − q0.25 )/β.
3. Repeat 1-2 until convergence (complete α and β).
4. The location parameter µ is simply q0.25 −βq0.25,α , q0.50 −βq0.50,α or q0.75 −βq0.75,α .
150 Quantile based direct prior elicitation

Remark 4 IMSL library is used to call the gamin function to return the standard
quantiles for arbitrary α (Table 2). We create the look-up table of (q0.25,α , q0.50,α ,
q0.75,α ) vs. α, where the α step is 0.001, the iteration alternates between finding the
optimal α given (q0.50,α − q0.25,α ) and finding the optimal (q0.75,α − q0.50,α ) given α
from the look-up table (the lower panels of Figure 5), till convergence. The one-way
elicitation of α followed by β (the top-right panel of Figure 5) is replaced by a two-way
joint elicitation of α and β. The comparison is given in Table 2. where the numerical
difference is substantially greater than 0.001 (α step). One-way elicitation introduced in
Section 2.4 only works well for those regions where the shape parameter is not sensitive,
only two-way iterating elicitation works on regions where the shape parameter is highly
sensitive, including the IQR based iteration and Taylor’s expansion based iterations.
Unfortunately, we also observe those regions where both elicitations encounter high
sensitivity index (SI), such as the tail regions (λ approaches 10) in the lower panels of
Figure 1.

4 Interactive Graphical Display of Prior Probability Dis-

tribution
In this section, we propose visual aids to help elicit skewed priors and behavior of the
tail shape with an emphasis on the degrees of freedom ν of skew Student’s t priors.
This section is a follow-up of preceding sections to streamline the “sketching” principle
recommended by Berger (1985, Chapter 3).

Trial on Degrees of Freedom

A. We consider ν = 1, 5, 10, 15, 30, . . . to apply the general procedure in Section 2.5,
with an exact or an approximate scale elicitation, to select the most agreeable ν
for skewed Student’s t prior by repeatedly drawing distribution shapes.

B. We start with a tentative ν out of (1, 5, 10, 15, 30,. . .) to obtain σ̂ in an exact
way, then immediately update the degrees of freedom for Student’s t-distribution
by reverse of (4) and show the distribution shape to the expert in order to make
an exploratory effort to reach an agreement on ν.

Other Cases

C. The expert provides his/her opinion in the form of a graph, then we estimate
q0.25 , q0.50 and q0.75 directly from the graph. If the graph is close to symmetric
normal we use (1); if it is close to log-normal with a positive domain, we apply
Proposition 1; if it is close to symmetric heavier tail, we go to option [B].

D. For generalized symmetric Student’s t-distribution with two degrees of freedom,

the graphical approach seems more necessary and similar procedures apply.
Dey and Liu 151

E. For skew-normal and normal-exponential priors, the aforementioned general pro-

cedures apply (Sections 2.4, 2.6 and 3), we need to redraw distribution shape for
expert’s evaluation.

Now we briefly explore the shapes of generalized Student’s t-distribution. General-

ized Student’s t-distribution is a special case of a scale mixture of normal distribution. It
is obtained as X|λ ∼ N (µ, λ/σ 2 ) and λ ∼ Gamma(ν1 /2, ν2 /2), the marginal distribution
ν1 +1 2 ν1 +1
Γ( )
produces the generalized t, which has the form f (x) = √πν1
2
2
ν
σΓ( 21 )
(1 + (x−µ)
ν2 σ 2 )
− 2
.
When ν1 = ν2 we get back to the usual Student’s t-distribution. The two degrees of free-
dom parameters make the distribution extremely flexible keeping the symmetry intact.
For example, t(30,1) and t(1,30), t(10,1) and t(1,10) densities are shown in Figure 10.
The tail could be even heavier than Cauchy. For prior elicitation, this distribution can
be used for any symmetric specification, similar to Section 2, the realistic approximate
ν +1
p Γ( 12 ) (q0.50 −µ)2 − ν1 +1
algorithm will be to set IQR as √1
πν2 σΓ( ν21 ) (1 + ν2 σ 2 ) 2 , thus

IQR 1 Γ( ν12+1 )
µ̂ = q0.50 and σ̂ = √ . (7)
p πν2 Γ( ν21 )

Note that the only difference between (4) and (7) is the presence of ν2 in the denomi-
nator, a very minor change in the code.

5 Concluding Remarks
We demonstrate that, for those univariate skewed priors belonging to location-scale-
shape families, the elicitation procedure on the non-sensitive regions is: shape param-
eter followed by the scale and location parameters. Moreover, we take the initiatives
to quantitatively study technical details for flexible direct prior elicitation, which will
streamline the induced interactive graphical procedures. The quantile based Taylor’s ex-
pansion approximation was interestingly found to enjoy sufficient efficiency to pinpoint
the scale parameter on shape-parameter sensitive regions, when the shape information is
incorporated for approximation. The exact procedure is in the sense of numerical com-
putation, and the iterative elicitation is numerically tested to be satisfactorily stable
and efficient. Prior elicitation from multiple expert opinions could be done by averaging
these summaries (Kadane, et al, 1980) or applying linear regression to take into account
multiple experts (Gill and Walker, 2005). Throughout we assumed that experts provide
a unimodal distribution, which is realistic in practice. However, if the experts believe
that the distribution could be multimodal then we have to consider discrete mixtures of
unimodal distributions. If the prior predictive space is necessary and a well-established
response model is available, then expert opinions on the response distribution may be
more practical. But we can envision that the calculation load will be much more in-
tensive under skewed priors. How to develop efficient prior elicitation algorithm under
that circumstance is another area of research.
152 Quantile based direct prior elicitation

Supplemental Material: Proofs, Figures and Results

Proof of Theorem 1

R Qq (λ) R +∞
Proof of part 1. From 2 −∞ φ(x)Φ(λx)dx = q, we have 2 Qq (λ) φ(x)Φ(λx)dx =
R Qq (λ)
1 − q and 2 +∞ −φ(−x)Φ(λx)dx = 1 − q. Substituting y = −x, it follows that
R −Qq (λ)
2 −∞ φ(y)Φ(−λy)dy = 1 − q, which proves part 1.

Proof of part 2.

R Qq (λ+∆λ )
2φ(x)Φ((λ + ∆λ )x)dx
R−∞
Qq (λ) R Q (λ+∆ )
= 2φ(x)Φ((λ + ∆λ )x)dx + Qqq(λ) λ 2φ(x)Φ((λ + ∆λ )x)dx
R−∞
Qq (λ)
= −∞ 2φ(x)Φ((λ + ∆λ )x)dx (8)
+(Qq (λ + ∆λ ) − Qq (λ))2φ(pQq (λ) + (1 − p)Qq (λ + ∆λ ))
×Φ((λ + ∆λ )(pQq (λ) + (1 − p)Qq (λ + ∆λ ))), (∃p ∈ [0, 1])
R Qq (λ)
= −∞
2φ(x)Φ(λx)dx = q.

Thus,

R Qq (λ) Q (λ+∆ )−Qq (λ)

−∞ 2xφ(x)( Φ((λ+∆x∆
λ )x)−Φ(λx)
λ
)dx + q λ
∆λ
×2φ(pQq (λ) + (1 − p)Qq (λ + ∆λ )) (9)
×Φ((λ + ∆λ )(pQq (λ) + (1 − p)Qq (λ + ∆λ )))
= 0.

R Qq (λ)
Letting ∆λ approach 0, we get −∞ 2xφ(x)φ(λx)dx +Q0q (λ)2φ(Qq (λ))Φ(λQq (λ)) =
0. The final result follows after some simplifications. Our result shows that Qq (λ) is
differentiable. The proof ends.

Proof of Theorem 2

Obviously, g(?) retains linearity of the standardized characteristic point by g(?) =

µ + σg(λ), where g(?) is the characteristic point of general location-scale-shape family
and g(λ) is the characteristic point of the standard location-scale-shape family with
location 0 and scale 1. Thus, (2) becomes
(1)
Fλ,µ,σ (q0.75 ) = Fλ,0,1 (g(λ)) + σ1 fλ,0,1 (g(λ))(q0.75 − g(?)) + 12 fλ,0,1 (g(λ))(q0.75,λ − g(λ))2 · · ·
(1) .
Fλ,µ,σ (q0.25 ) = Fλ,0,1 (g(λ)) + σ1 fλ,0,1 (g(λ))(q0.25 − g(?)) + 12 fλ,0,1 (g(λ))(q0.25,λ − g(λ))2 · · ·
(10)

Consequently,
Dey and Liu 153

1 1
P∞ 1 (k)
2 = σ fλ,0,1 (g(λ))IQR + k=1 k+1 fλ,0,1 (g(λ))[(q0.75,λ − g(λ))2 − (q0.25,λ − g(λ))2 ].

P∞ 1 (k) 2 2
Denote ∆= k=1 k+1 fλ,0,1 (g(λ))[(q0.75,λ − g(λ)) − (q0.25,λ − g(λ)) ], the relative error
(k) (k) (k)
f (g(λ)) f (g(λ)) fλ,0,1 (g(λ)) −2∆
for scale parameter elicitation is ( λ,0,1 0.50 − λ,0,1
0.50−∆ )/( 0.50−∆ ) = 1−2∆ .
The
second statement in Theorem 2 comes from the following fact: the true scale parameter
IQR
σ is equal to IQR(λ) by direct numerical calculation from the look-up table IQR(λ) vs.
IQRfλ,0,1 (g(λ))
λ, or 0.50−∆ by keeping the first term from the fully Taylor’s expansion. The
proof ends.
154 Quantile based direct prior elicitation

Table 1: Iteration Comparison: Inter-quartile range (IQR), Taylor’s expansion at me-

dian, mean and mode (skew-normal: q0.25 = −10, q0.75 = 10)

q0.50 λ σ µ
IQR median mean mode IQR median mean mode IQR median mean mode
0.01 -0.23 -0.23 -0.23 -0.23 15.06 15.07 15.07 15.07 2.69 2.69 2.69 2.69
0.02 -0.37 -0.32 -0.32 -0.32 15.42 15.29 15.29 15.29 4.26 3.70 3.70 3.70
0.03 -0.41 -0.41 -0.41 -0.41 15.56 15.56 15.56 15.56 4.69 4.69 4.69 4.69
0.04 -0.53 -0.54 -0.54 -0.54 15.99 16.04 16.04 16.04 5.93 6.04 6.04 6.04
0.05 -0.54 -0.55 -0.55 -0.55 16.03 16.08 16.08 16.08 6.04 6.15 6.15 6.15
0.06 -0.64 -0.56 -0.56 -0.56 16.44 16.12 16.12 16.12 7.00 6.24 6.23 6.24
0.07 -0.70 -0.67 -0.69 -0.67 16.70 16.58 16.66 16.58 7.56 7.30 7.47 7.30
0.08 -0.68 -0.68 -0.68 -0.71 16.62 16.62 16.62 16.76 7.39 7.39 7.39 7.67
0.09 -0.71 -0.72 -0.72 -0.72 16.74 16.80 16.80 16.80 7.65 7.75 7.74 7.75
0.10 -0.77 -0.82 -0.77 -0.85 17.01 17.25 17.02 17.40 8.17 8.62 8.18 8.88
0.11 -0.82 -0.85 -0.85 -0.86 17.23 17.39 17.38 17.44 8.60 8.87 8.86 8.95
0.12 -0.85 -0.86 -0.86 -0.89 17.37 17.43 17.42 17.58 8.85 8.94 8.93 9.21
0.13 -0.88 -0.89 -0.89 -0.92 17.50 17.57 17.56 17.72 9.08 9.19 9.18 9.43
0.14 -0.89 -0.92 -0.91 -0.95 17.55 17.71 17.65 17.86 9.17 9.41 9.33 9.67
0.15 -0.94 -0.95 -0.92 -1.01 17.78 17.85 17.69 18.14 9.55 9.65 9.40 10.12
0.16 -0.95 -1.01 -0.95 -1.03 17.83 18.12 17.83 18.23 9.64 10.10 9.63 10.26
0.17 -1.01 -1.03 -1.01 -1.04 18.09 18.21 18.09 18.28 10.07 10.24 10.07 10.33
0.18 -1.01 -1.04 -1.01 -1.11 18.09 18.25 18.09 18.60 10.07 10.30 10.07 10.82
0.19 -1.04 -1.09 -1.04 -1.24 18.24 18.48 18.22 19.18 10.29 10.65 10.27 11.66
0.20 -1.09 -1.11 -1.04 -1.26 18.46 18.57 18.22 19.27 10.62 10.78 10.27 11.80

Table 2: Shape parameter elicitation comparison: iterative inter-quartile range algo-

rithm (Gamma: q0.25 = −2.0, q0.75 = 3.0)
q0.50 α (iterative) α (IQR ratio) q0.50 α (iterative) α (IQR ratio)
0.02 1.655 1.652 0.22 4.329 4.324
0.04 1.778 1.775 0.24 4.975 4.969
0.06 1.919 1.916 0.26 5.775 5.782
0.08 2.080 2.076 0.28 6.818 6.825
0.10 2.266 2.261 0.30 8.189 8.197
0.12 2.481 2.476 0.32 10.040 10.049
0.14 2.711 2.728 0.34 12.631 12.640
0.16 3.029 3.024 0.36 16.408 16.418
0.18 3.382 3.378 0.38 22.226 22.244
0.20 3.809 3.804 0.40 31.864 31.899
Dey and Liu 155

Figure 1: Relationship between quantiles and skewness parameter λ for skew-normal

q −q0.50,λ
distribution. (Top-left: (q0.75,λ , q0.50,λ , q0.25,λ ) vs. λ; top-right: ( q0.75,λ
0.50,λ −q0.25,λ
) vs. λ;
bottom-left: (q0.75,λ − q0.50,λ ) vs. λ; bottom-right: (q0.75,λ − q0.25,λ ) vs. λ)
156 Quantile based direct prior elicitation

Figure 2: Relationship between quantiles and skewness parameter λ for skew Student’s
q0.75,λ −q0.50,λ
t distribution. (Top-left: (q0.75,λ , q0.50,λ , q0.25,λ ) vs. λ; top-right: ( q0.50,λ −q0.25,λ ) vs. λ;
bottom-left: (q0.75,λ − q0.50,λ ) vs. λ; bottom-right: (q0.75,λ − q0.25,λ ) vs. λ)
Dey and Liu 157

0.010

0.08
α=0.1
0.008

α=0.01

0.06
0.006

0.04
f(x)

f(x)
0.004

0.02
0.002
0.000

0.00

0 100 200 300 400 500 −20 0 20 40 60 80 100

x x
0.4
0.00 0.05 0.10 0.15 0.20 0.25 0.30

α=1 α=10
0.3
0.2
f(x)

f(x)

0.1
0.0

−10 −5 0 5 10 −14 −12 −10 −8 −6

x x

Figure 3: Relationship between density function and shape parameter α for normal-
exponential distribution.
158 Quantile based direct prior elicitation

Figure 4: Relationship between quantiles and shape parameter α for normal-exponential

distribution. (Top-left: (q0.75,α , q0.50,α , q0.25,α ) vs. α; top-middle: (q0.75,α − q0.50,α ) vs.
q0.75,α −q0.50,α
α; top-right: (q0.75,α − q0.25,α ) vs. α; bottom-left: ( q0.50,α −q0.25,α ) vs. α; bottom-middle:
1 1
(q0.75,α − q0.50,α ) 5 vs. α; bottom-right: (q0.75,α − q0.25,α ) 5 vs. α)
Dey and Liu 159

2.0
(q(0.75,α)−q(0.50,α)):((q(0.50,α)−q(0.25,α))
50

1.8
q(0.25,α), q(0.50,α), q(0.75,α)

40

1.6
30

1.4
20

1.2
10

1.0
0

0 10 20 30 40 50 0 10 20 30 40 50

α α
5

5
4

4
q(0.75, α)−q(0.50,α)

q(0.50, α)−q(0.25,α)
3

3
2

2
1

1
0

0 10 20 30 40 50 0 10 20 30 40 50

α α

Figure 5: Relationship between quantiles and shape parameter α for Gamma distribu-
tion. (Top-left: (q0.75,α , q0.50,α , q0.25,α ) vs. α; top-right: (q0.75,α − q0.50,α )/(q0.50,α −
q0.25,α ) vs. α; bottom-left: (q0.75,α − q0.50,α ) vs. α; bottom-right: (q0.50,α − q0.25,α ) vs.
α)
160 Quantile based direct prior elicitation

Figure 6: Scale parameter elicitation comparison vs. degrees of freedom ν in terms of

relative error for Student’s t prior.

Figure 7: Estimated degrees of freedom ν from σ̂.

Dey and Liu 161

Figure 8: Scale parameter elicitation comparison for skew-normal. (Left and middle:
approximate σ/exact σ vs. λ, Right: mean, mode and median vs. λ.)

Figure 9: Scale parameter elicitation comparison for normal-exponential. (Left and

middle: approximate σ/exact σ vs. α, Right: mean, mode and median vs. α.)
162 Quantile based direct prior elicitation

Figure 10: Generalized Student’s t probability density function with different degrees
of freedom.
1.20

1.05

1.10
1.15

1.04
1.10

1.03

1.05
1.05

1.02
ratio

ratio

ratio

1.00
1.01
1.00

0.95
1.00
0.95

0.99

0.90
0.90

0.98

λ σ µ

Figure 11: Skew-normal prior elicitation comparison (relative to IQR based iteration
algorithm) among Taylor’s expansions at median, mean and mode: the left panel is for
λ, the middle panel is for σ, the right panel is for µ; from left to right in each panel,
the ratios are for median, mean and mode
Dey and Liu 163

1.04

1.15
1.15

1.03

1.10
1.10

1.02
1.05

1.01

1.05
ratio

ratio

ratio
1.00
1.00

1.00
0.99
0.95

0.95
0.98

λ σ µ

Figure 12: Skew-normal prior elicitation comparison (relative to median based iteration
algorithm) among IQR based iteration, Taylor’s expansions at mean and mode: the left
panel is for λ, the middle panel is for σ, the right panel is for µ; from left to right in
each panel, the ratios are for IQR, mean and mode
1.06

1.15
1.20

1.05
1.15

1.04

1.10
1.03
1.10
ratio

ratio

ratio

1.05
1.02
1.05

1.01

1.00
1.00

1.00
0.99

λ σ µ

Figure 13: Skew-normal prior elicitation comparison (relative to mean based iteration
algorithm) among IQR based iteration, Taylor’s expansions at median and mode: the
left panel is for λ, the middle panel is for σ, the right panel is for µ; from left to right
in each panel, the ratios are for IQR, median and mode
164 Quantile based direct prior elicitation

1.02

1.15
1.15
1.10

1.10
1.00
1.05

1.05
1.00
ratio

ratio

ratio
0.98

1.00
0.95

0.95
0.90

0.96

0.90
0.85

λ σ µ

Figure 14: Skew-normal prior elicitation comparison (relative to mode based iteration
algorithm) among IQR based iteration, Taylor’s expansions at median and mean: the
left panel is for λ, the middle panel is for σ, the right panel is for µ; from left to right
in each panel, the ratios are for IQR, median and mean
Dey and Liu 165

References
Aigner, D. J., Lovell, C. K., and Schmidt, P. (1977). “Formulation and estimation
of stochastic frontier production function model.” Journal of Econometrics, 6(1):
21–37.
Al-Awadhi, S. A. and Garthwaite, P. H. (1998). “An elicitation method for multivariate
normal distributions.” Communications in Statistics: Theory and Methods, 27(5):
1123–1142.
Azzalini, A. (1985). “A class of distributions which includes the normal ones.” Scandi-
navian Journal of Statistics, 12(2): 171–178.
Berger, J. O. (1985). Statistical Decision Theory and Bayesian Analysis. New York:
Springer-Verlag, 2 edition.
Dalal, S. R. and Hall, W. J. (1983). “Approximating priors by mixtures of natural
conjugate priors.” Journal of the Royal Statistical Society - Series B, 45(2): 278–
286.
Garthwaite, P. H. (1989). “Fractile assessment for a linear regression model: an exper-
imental study.” Organizational Behavior and Human Decision Processes, 43: 188–
206.
Garthwaite, P. H. and Dickey, J. M. (1985). “Double- and single-bisection methods for
subjective probability assessment in a location-scale family.” Special Bayesian issue
of J. Econometrics, 29: 149–163.
— (1988). “Quantifying expert opinion in linear regression problems.” Journal of the
Royal Statistical Society - Series B, 50(3): 462–474.
Garthwaite, P. H., Kadane, J. B., and O’Hagan, A. (2005). “Statistical methods for
elicitating probability distributions.” Journal of the American Statistical Association,
100(470): 680–701.
Gill, J. and Walker, L. D. (2005). “Elicited priors for Bayesian model specifications in
political science research.” The Journal of Politics, 67(3): 841–872.
Jaynes, E. T. (1968). “Prior probabilities.” IEEE Transactions on Systems Science and
Cybernetics, SSC-4: 227–241.
Kadane, J. B., Dickey, J. M., Winkler, R. L., Smith, W. S., and Peters, S. C. (1980).
“Interactive elicitation of opinion for a normal linear model.” Journal of the American
Statistical Association, 75(372): 845–854.
Meeden, G. (1992). “An elicitation procedure using piecewise conjugate priors.” In
Goel, P. K. and Iyengar, N. S. (eds.), Bayesian analysis in statistics and econometrics.
Springer-Verlag.
O’Hagan, A. (1998). “Eliciting expert beliefs in substantial practical applications.” The
Statistician, 47: 21–35.
166 Quantile based direct prior elicitation

Peterson, C. R. and Miller, A. (1964). “Mode, median and mean as optimal strategies.”
Journal of Experimental Psychology, 68: 363–367.
Sahu, S., Dey, D. K., and Branco, M. (2003). “A new class of multivariate skew distri-
butions with application to Bayesian regression models.” The Canadian Journal of
Statistics, 31: 129–150.

Acknowledgments
The authors are grateful to Joseph Kadane for his insightful suggestions which have lead to
significant improvement of the presentation, we also thank the editor and two referees for their
constructive comments, Anthony Billings for his dedicated proof reading of the manuscript and
Jun Tan for helping us with graphical outputs.