Chapter 4

Measuring Dependence
Given the wide selection of copula models, how should one model be chosen over the others in
empirical work? One of the key considerations is the nature of the dependence captured by
dierent copulas. The nature of the dependence captured by the dependence parameter(s)
varies from one copula to another. Moreover, commonly used measures of dependence are
related to the parameter(s) in copula functions. This chapter starts with a brief review of
the widely used measures of dependence linear correlation and rank correlation, followed
by the application of measures of rank correlation to the models with common random
eects. Measures of tail dependence will also be discussed.
4.1 Review of Dependence Measures
This section reviews the commonly used measures of dependence in statistics literature. We
focus on the dependence measures that have appeared more often in empirical work, instead
of a complete list of all the measures of dependence.
Two random variables X
1
and X
2
are said to dependent or associated if they are not
independent in the sense that F(x
1
, x
2
) = F
1
(x
1
)F
2
(x
2
), or S(x
1
, x
2
) = S
1
(x
1
)S
2
(x
2
). Let
(X
1
, X
2
) denote a scalar measure of dependence. Embrechts et al. (2002) listed four desir-
able properties of dependence measure:
(i) Symmetry: (X
1
, X
2
) = (X
2
, X
1
);
(ii) Normalization: 1 (X
1
, X
2
) +1;
30
CHAPTER 4. MEASURING DEPENDENCE 31
(iii) (X
1
, X
2
) = +1 if and only if (X
1
, X
2
) are comonotonic; (X
1
, X
2
) = 1 if and only
if (X
1
, X
2
) are countermonotonic;
(iv) For a strictly monotonic transformation T : R R of X
1
:
(T(X
1
), X
2
) =
(
(X
1
, X
2
) if T is increasing,
(X
1
, X
2
) if T is decreasing.
4.1.1 Correlation coecient
The most commonly used measure of dependence (or association) between two random
variables X
1
and X
2
is Pearsons correlation coecient, which is dened as

X
1
X
2
=
Cov (X
1
, X
2
)
[Var(X
1
)]
1
2
[Var(X
2
)]
1
2
, (4.1)
where Cov(X
1
, X
2
) = E[X
1
X
2
] E[X
1
]E[X
2
], Var(X
1
) and Var(X
2
) are the variances of
X
1
and X
2
, respectively.
It is well known that:
(i)
X
1
X
2
is a measure of linear dependence,
(ii)
X
1
X
2
is symmetric,
(iii) the lower and upper bounds on the inequality 1
X
1
X
2
+1 measure perfect
negative and positive linear dependence, and
(iv) it is invariant with respect to strictly increasing linear transformations of the variables.
The weakness of using the correlation coecient as a measure of dependence includes:
(i) in general, zero correlation does not imply independence,
(ii) it is not dened for heavy-tail distributions whose second moments do not exist,
(iii) it is not invariant under strictly increasing nonlinear transformations, and
(iv) attainable values of the correlation coecients within interval [1, +1] between two
variables depend upon their respective marginal distributions.
These limitations have motivated alternative measures of dependence based on ranks.
CHAPTER 4. MEASURING DEPENDENCE 32
4.1.2 Rank correlation
Consider two random variables X
1
and X
2
with continuous distribution functions F
1
and
F
2
, respectively, and joint distribution function F. Two well-established measures of rank
correlation are Spearmans rho and Kendalls tau.
Spearmans rho is the linear correlation between the distribution functions, dened as

S
(X
1
, X
2
) = (F
1
(X
1
), F
2
(X
2
)),
where =
X
1
X
2
is dened in (4.1).
Kendalls tau is dened as

K
(X
1
, X
2
) = P{(X
1
X
0
1
)(X
2
X
0
2
) > 0} P{(X
1
X
0
1
)(X
2
X
0
2
) < 0}, (4.2)
where (X
1
, X
2
) and (X
0
1
, X
0
2
) are two independent pairs of random variables from F. The
rst term on the right hand side of equation (4.2) is referred to as the probability of con-
cordance, and the second term as the probability of discordance, and hence

K
(X
1
, X
2
) = P{concordance} P{discordance}.
The similarity between Spearmans rho and Kendalls tau is that both of them measure
monotonic dependence between random variables, and both are based on the concept of
concordance, which refers to the property that large values of one random variable are
associated with large values of another, whereas discordance refers to large values of one
being associated with small values of the other.
These two well-established measures of rank correlation have properties of symmetry,
normalization, comonotonicity and countermonotonicity, and both assume the value of zero
under independence. Further,

S
(X
1
, X
2
) =
K
(X
1
, X
2
) = 1 if and only if C = C
L
= Max(u
1
+u
2
1, 0),

S
(X
1
, X
2
) =
K
(X
1
, X
2
) = +1 if and only if C = C
U
= Min(u
1
, u
2
),

S
(X
1
, X
2
) =
K
(X
1
, X
2
) = 0 if and only if C = u
1
u
2
.
4.2 Measures of Rank Correlation for Models of Common
Random Eects
Spearmans rho and Kendalls tau can be expressed in terms of copulas as follows:

S
(X
1
, X
2
) =
S
(C) = 12
Z
1
0
Z
1
0
C(u
1
, u
2
)du
1
du
2
3; (4.3)
CHAPTER 4. MEASURING DEPENDENCE 33

K
(X
1
, X
2
) =
K
(C) = 4
Z
1
0
Z
1
0
C(u
1
, u
2
)

2
C(u
1
, u
2
)
u
1
u
2
du
1
du
2
1 (4.4a)
= 1 4
Z
1
0
Z
1
0
C(u
1
, u
2
)
u
1
C(u
1
, u
2
)
u
2
du
1
du
2
. (4.4b)
More details about (4.3) and (4.4) can be found in Joe (1997) or Nelsen (2006).
For the two approaches of modeling the joint behavior of random variables in Chapter
3, the following proposition shows that the regular copula and its associated survival copula
have same rank correlations.
Proposition 4.1 The rank correlation of the survival copula
b
C is equal to that of the
regular copula C, that is,

S
(C) =
S
(
b
C), (4.5)

K
(C) =
K
(
b
C), (4.6)
where the survival copula
b
C is dened by equation (2.8).
Proof. From (2.8) and the expression of Spearmans rho in terms of copulas in (4.3), we
have

S
(
b
C) = 12
Z
1
0
Z
1
0
b
C(u
1
, u
2
)du
1
du
2
3
= 12
Z
1
0
Z
1
0
(u
1
+u
2
1 +C(1 u
1
, 1 u
2
))du
1
du
2
3
= 12
Z
1
0
Z
1
0
(1 u
1
u
2
+C(u
1
, u
2
))du
1
du
2
3
= 12
Z
1
0
Z
1
0
(1 u
1
u
2
)du
1
du
2
+ 12
Z
1
0
Z
1
0
C(u
1
, u
2
)du
1
du
2
3
= 12
Z
1
0
Z
1
0
C(u
1
, u
2
)du
1
du
2
3
=
S
(C),
because in the third last line,
R
1
0
R
1
0
(1 u
1
u
2
)du
1
du
2
= 0.
CHAPTER 4. MEASURING DEPENDENCE 34
Similarly, using (2.8), (4.4b), and the properties of copulas in Section 2.2.1, we obtain

K
(
b
C) = 1 4
Z
1
0
Z
1
0

b
C(u
1
, u
2
)
u
1

b
C(u
1
, u
2
)
u
2
du
1
du
2
= 1 4
Z
1
0
Z
1
0

C(1 u
1
, 1 u
2
)
(1 u
1
)
1

C(1 u
1
, 1 u
2
)
(1 u
2
)
1

du
1
du
2
= 1 4
Z
1
0
Z
1
0

C(u
1
, u
2
)
u
1
1

C(u
1
, u
2
)
u
2
1

du
1
du
2
= 1 4
Z
1
0
Z
1
0
C(u
1
, u
2
)
u
1
C(u
1
, u
2
)
u
2
du
1
du
2
+ 4
Z
1
0
Z
1
0
C(u
1
, u
2
)
u
1
du
1
du
2
+ 4
Z
1
0
Z
1
0
C(u
1
, u
2
)
u
2
du
1
du
2
4
Z
1
0
Z
1
0
du
1
du
2
= 1 4
Z
1
0
Z
1
0
C(u
1
, u
2
)
u
1
C(u
1
, u
2
)
u
2
du
1
du
2
=
K
(C).
2
Unlike Pearsons correlation coecient, rank correlations depend on the copula of a
bivariate distribution and not on the functional forms of the marginal distributions. In
other words, each copula species a unique dependence structure and the rank correlation is
a function of the dependence parameter(s) embedded in the copula. Because of the limited
dependence parameter space, the Clayton, Pareto, Gumbel, and Hougaard copulas permit
only non-negative association, while the Frank copula allows positive as well as negative
association.
Furthermore, Kendalls tau can be evaluated directly from the Laplace transform of the
common random eect , as shown in the following theorem.
Theorem 4.1 Let X
1
and X
2
be random variables with copulas generated by the models
of common random eects (3.4) or (3.11). Then Kendalls tau is given by

K
(X
1
, X
2
) = 1 + 4
Z
1
0

0
(
1
(s))
1
(s)ds,
where (s) is the Laplace transform of the common random eect.
Proof. Genest and MacKay (1986) gave the following expression for Kendalls tau,

K
(X
1
, X
2
) = 1 + 4
Z
1
0
(s)

0
(s)
ds,
CHAPTER 4. MEASURING DEPENDENCE 35
where (s) is the Archimedean generator.
Since the Archimedean generator, (s), is the inverse of the Laplace transform of the
common random eect, (s), that is, (s) =
1
(s), then using the formula for the derivative
of an inverse function, we have
(s)

0
(s)
=
0
(
1
(s))
1
(s).
The desired result follows immediately. 2
Table 4.1 illustrates Spearmans rho and Kendalls tau for the copulas specied in Chap-
ter 3.
Table 4.1: Copulas and Their Rank Correlations
Copula type Copula Function C(u
1
, u
2
) Spearmans rho Kendalls tau
Product u
1
u
2
0 0
Clayton (u

1
+u

2
1)
1/
Complicated form

2+
Frank
1

ln

1 +
(e
u
11)(e
u
21)
e

1 +
12

{D
2
() D
1
()} 1 +
4

{D
1
() 1}
Gumbel e

(lnu1)

+(lnu2)

1/
No closed form 1
1
Hougaard u
1
+u
2
1 +e

(ln(1u1))

+(ln(1u2))

1/
No closed form 1
1
Pareto u
1
+u
2
1 +

(1 u
1
)

+ (1 u
2
)

1/
Complicated form

2+
The dependence measures of Frank copula depend on Debye faction, dened as D
k
(x) =
k
x
k
R
x
0
t
k
e
t
1
dt, for k = 1, 2. D
k
(x) = D
k
(x) +
kx
k+1
Figure 4.1 shows the scatter plots for bivariate distributions with identical marginal
exponential distributions (with mean of 1) and identical rank correlation but dierent de-
pendence structures. Perspective plots of the corresponding copula densities are given in
Figure 4.2. If these random variables represent the insurance losses, then the Gumbel and
Pareto copulas would be preferable models for insurers since extreme losses have tendency
to occur together. Measures of tail dependency discussed in next section can be used to
capture the extremal dependence.
4.3 Measures of Tail Dependency
As we have seen from Figures 4.1 and 4.2, copulas with same rank correlation may have
dramatically dierent tail behavior. Measures of tail dependence may help to distinguish
CHAPTER 4. MEASURING DEPENDENCE 36
Figure 4.1: Simulated Samples from Five Copulas with Same Marginal Distri-
butions (Exponential with Mean of 1) and Same Rank Rorrelation (Kendalls
tau = 0.50)
0 2 4 6 8 10
0
2
4
6
8
Clayton Copula
x1
x
2
0 2 4 6 8 10
0
2
4
6
8
1
0
Frank Copula
x1
x
2
0 2 4 6 8 10
0
2
4
6
8
1
0
Gumbel Copula
x1
x
2
0 2 4 6 8
0
2
4
6
8
Hougaard Copula
x1
x
2
0 2 4 6 8 10
0
2
4
6
8
1
0
Pareto Copula
x1
x
2
CHAPTER 4. MEASURING DEPENDENCE 37
Figure 4.2: Five Copulas with the Same Rank Correlation (Kendalls tau =
0.50)

copulas. In some empirical applications, the joint behavior of tail values of random variables
is of particular interest. For example, investors may be more concerned about the probability
that the rates of returns of all securities in a portfolio fall below given levels. This requires
measures of tail dependency. The tail dependency measure can be dened in terms of
conditional probability that one random variable exceeds some value given that another
exceeds some value. Various measures of tail dependency can be found in Joe (1997),
Nelsen (2006), Venter (2002), and Frahm (2006).
4.3.1 Tail concentration functions
Let X
1
and X
2
be random variables with continuous distribution functions F
1
and F
2
, and
copula C. Then U
1
= F
1
(X
1
) and U
2
= F
2
(X
2
) are standard uniform random variables.
The right and left tail concentration functions can be dened with reference to how much
probability is in regions near (1, 1) and (0, 0). For any z in (0, 1), the left tail concentration
CHAPTER 4. MEASURING DEPENDENCE 38
function is dened as
L(z) = P(U
1
< z | U
2
< z)
=
P(U
1
< z, U
2
< z)
P(U
2
< z)
=
C(z, z)
z
, (4.7)
and the right tail concentration function is
R(z) = P(U
1
> z | U
2
> z)
=
P(U
1
> z, U
2
> z)
P(U
2
> z)
=
1 P(U
1
< z) P(U
2
< z) +P(U
1
< z, U
2
< z)
1 P(U
2
< z)
=
1 2z +C(z, z)
1 z
. (4.8)
The relationship between the tail concentration functions of regular copula and its as-
sociated survival copula is detailed in the following proposition.
Proposition 4.2 Let
b
C be the survival copula associated with the regular copula C. Then
the left (right) tail concentration function of
b
C is equal to the right (left) tail concentration
function of C, that is
L
b
C
(z) = R
C
(1 z), R
b
C
(z) = L
C
(1 z).
Proof. By the denition of the left tail concentration function in (4.7),
L
b
C
(z) =
b
C(z, z)
z
=
2z 1 +C(1 z, 1 z)
1 (1 z)
= R
C
(1 z);
Similarly,
R
b
C
(z) =
1 2z +
b
C(z, z)
1 z
=
C(1 z, 1 z)
1 z
= L
C
(1 z).
2
CHAPTER 4. MEASURING DEPENDENCE 39
4.3.2 Upper (lower) tail dependence coecients
The degree of extreme co-movements of random variables can be dened by taking limits
of equations (4.7) and (4.8). The upper (lower) tail dependence coecients capture the
probability that one event is extermal conditional on another extreme event, which are
given by

L
= lim
z0
P(U
1
< z|U
2
< z)
= lim
z0
C(z, z)
z
, (4.9)
and

R
= lim
z1
P(U
1
> z|U
2
> z)
= lim
z1
1 2z +C(z, z)
1 z
. (4.10)
If
R
(
L
) is positive, then the two variables are said to be right (left) tail dependent, with
larger values indicating stronger dependence.
4.3.3 Extremal dependence coecients
Let U
Min
= Min{U
1
, U
2
}, and U
Max
= Max{U
1
, U
2
}. The lower extremal dependence coe-
cient is dened as

L
= lim
z0
P(U
Max
< z | U
Min
< z)
= lim
z0
P(U
Max
< z, U
Min
< z)
P(U
Min
< z)
= lim
z0
P(U
1
< z, U
2
< z)
1 P(U
1
> z, U
2
> z)
= lim
z0
P(U
1
< z, U
2
< z)
P(U
1
< z) +P(U
2
< z) P(U
1
< z, U
2
< z)
= lim
z0
C(z, z)
2z C(z, z)
, (4.11)
CHAPTER 4. MEASURING DEPENDENCE 40
whereas the upper extremal dependence coecient is dened as

R
= lim
z1
P(U
Max
> z | U
Min
> z)
= lim
z1
P(U
Max
> z, U
Min
> z)
P(U
Min
> z)
= lim
z1
P(U
1
> z, U
2
> z)
1 P(U
1
< z, U
2
< z)
= lim
z1
1 P(U
1
< z) P(U
2
< z) +P(U
1
< z, U
2
< z)
1 P(U
1
< z, U
2
< z)
= lim
z1
1 2z +C(z, z)
1 C(z, z)
. (4.12)
Thus the lower extremal dependence coecient can be interpreted as the probability that
the best performer is aected by the worst one provided that the latter has an extremely
bad performance, while the upper extremal dependence coecient measures the probability
that the worst performer is aected by the best given that the latter has an extremely good
performance.
The following proposition relates the tail dependence coecients and the extremal de-
pendence coecients.
Proposition 4.3 Let
L
and
R
be the tail dependence coecients dened by equations
(4.9) and (4.10), and
L
and
R
be the corresponding extremal dependence coecients
dened by equations (4.11) and (4.12). Then

L
=

L
2
L
,
R
=

R
2
R
.
Proof. From the denition of lower extremal dependence coecient given in (4.11), and
using lim
z0
C(z,z)
z
=
L
by equation (4.9), we have

L
= lim
z0
C(z, z)
2z C(z, z)
= lim
z0
C(z,z)
z
2
C(z,z)
z
=

L
2
L
.
CHAPTER 4. MEASURING DEPENDENCE 41
Similarly, using (4.12) and (4.10),

R
= lim
z1
1 2z +C(z, z)
1 C(z, z)
= lim
z1
12z+C(z,z)
1z
2
12z+C(z,z)
1z
=

R
2
R
.
2
Table 4.2 summarizes the measures of tail dependency for the copula functions specied
in Chapter 3. If the dependency over the right tail is of particular interest to practition-
ers, then the Gumbel and Pareto copulas should be considered. Mis-specication of the
dependence structure, especially the dependency over the tails, may result in devastating
consequences, which will be shown in insurance applications in Chapter 6.