Journal of Multivariate Analysis 128 (2014) 210225

Contents lists available at ScienceDirect

Journal of Multivariate Analysis

journal homepage: www.elsevier.com/locate/jmva

Kendalls tau for hierarchical data

H. Romdhani, L. Lakhal-Chaieb, L.-P. Rivest
Dpartement de Mathmatiques et de Statistique, Universit Laval, Pavillon Alexandre-Vachon 1045, av. de la Mdecine,
Qubec G1V 0A6, Canada

article

info

Article history:
Received 19 February 2013
Available online 4 April 2014
AMS subject classifications:
primary 62H20
secondary 62H15
Keywords:
Exchangeable copula
Hierarchical Archimedean copula
Hierarchical meta-elliptical copula
Intra class independence test
Kendalls tau
Pitman efficiency

abstract
This paper is concerned with hierarchical data having three levels. The level 1 units are
nested in the level 2 units or subclusters which are themselves nested in the level 3 clusters.
The model for this data is assumed to fulfill some symmetry assumptions. The level 1 units
within each subcluster are exchangeable and a permutation of the subclusters belonging
to the same cluster leaves the model unchanged. We are interested in measuring the dependence associated to clusters and subclusters respectively. Two exchangeable Kendalls
tau are proposed as non parametric measures of these two associations and estimators for
these measures are proposed. Their asymptotic properties are then investigated under the
proposed hierarchical model for the data. These statistics are then used to estimate the
intra-class correlation coefficients for data drawn from elliptical hierarchical distributions.
Hypothesis tests for the cluster and subcluster effects based on the proposed estimators
are developed and their performances are assessed using Pitman efficiencies and a Monte
Carlo study.
2014 Published by Elsevier Inc.

1. Introduction
Hierarchical data structures are commonly found in many application areas of statistics especially in social sciences but
also in other fields such as economics, finance and risk management. The hierarchy arises naturally from the organization
of the data: the variable of interest is observed on units that are grouped into subclusters that are themselves grouped into
clusters. Education provides a well known example where scores are observed on students which are clustered in schools
and the schools themselves are grouped in geographical regions. In this paper, students, regions and schools are called units,
clusters and subclusters, respectively.
Statistical models for hierarchical data characterize the association within and between subclusters. Multilevel models,
also known as hierarchical or nested linear models, investigate the variation at different levels of the hierarchy. The standard
reference for these models is [4]; most of the models presented there are based on the normality assumption. In survey
sampling, a multistage sample design is considered in the presence of a hierarchical data structure [13, chapter 4]. In this
context, the dependence levels associated to the subclusters and to the clusters respectively, are typically of prime interest
as the precision of the statistical methods applied to these data depend on the strength of this dependence.
In the bivariate case, Kendalls tau is a measure of association defined as the probability of concordance minus the
probability of discordance. Joe [6] defined an ordering of multivariate concordance and constructed a multivariate Kendalls
tau; see also [2]. For clustered data, Romdhani, Rivest and Lakhal-Chaieb [12] introduced an exchangeable version of
Kendalls tau as a measure of intra cluster association. The empirical counterpart of the latter association measure is

Corresponding author.
E-mail addresses: hela.romdhani.1@ulaval.ca (H. Romdhani), Lajmi.Lakhal@mat.ulaval.ca (L. Lakhal-Chaieb), Louis-Paul.Rivest@mat.ulaval.ca
(L.-P. Rivest).
http://dx.doi.org/10.1016/j.jmva.2014.03.016
0047-259X/ 2014 Published by Elsevier Inc.

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

211

computed by considering all possible pairs of bivariate observations such that the two elements of a pair come from different
clusters. The properties of the resulting estimator were investigated under an exchangeable multivariate distribution model.
This paper considers three level hierarchical data. It proposes two association measures based on the exchangeable
Kendalls tau. The first one, associated to the subclusters, measures the association between two units from the same
subcluster and is defined by considering pairs of bivariate vectors drawn from two subclusters coming from different
clusters. The second one is related to the clusters and measures the association between two units from the same cluster
but different subclusters. It is defined by considering pairs of bivariate vectors coming from two different clusters such that
each observation of a bivariate observation is drawn from two different subclusters. As will be seen in Sections 4 and 5, these
statistics allow one to perform valid inference for hierarchical data when the normality assumption is questionable.
The sampling distributions of the two association measures based on the exchangeable Kendalls tau are investigated
under a general model for hierarchical data. These models are presented in Section 2 along with the different association
measures. In Section 3, we give estimators for the two exchangeable Kendalls tau and investigate their asymptotic
distributions. Section 4 presents two estimators for the intra class correlation coefficients for elliptical distributions; one
is the standard moment estimator and the other is based on Kendalls tau. Section 5 investigates tests for the cluster and
subcluster effects based on Kendalls tau estimators proposed in Section 3. The cluster effect tests performance is assessed
using Pitman efficiencies and a simulation study is conducted for the two tests. Section 6 provides a numerical example.
Proofs and technical details can be found in the appendices.
2. Models for nested data
2.1. A general model for three level data
Let I denote the number of clusters, ni the number of subclusters in cluster i, i = 1, . . . , I, and mij , for j = 1, . . . , ni , the
size of the jth subcluster of cluster i. Let Yij , i = 1, . . . , I, j = 1, . . . , ni and = 1, . . . , mij denotes the random variable
for the th unit of the jth subcluster
cluster i. The random vector representing the data in cluster i is the Ni 1 vector
nof
i
T
Yi = (YTi1 , . . . , YTini )T where Ni =
m
ij is the total number of observations in the ith cluster and Yij = (Yij1 , . . . , Yijmij )
j
denotes the mij 1 random vector for subcluster j of cluster i. Measurements from different clusters are assumed to be
independent.
(n)
The dependence
within a cluster is modeled using a family of cumulative distribution functions (cdf) {Fm1 ,m2 ,...,mn (Y1 , . . . ,
mj
Yn ) : Yj ,
j mj = N } indexed by (n, m1 , . . . , mn ). This family is assumed to satisfy the following permutability
property: let yj represent an mj 1 vector for j = 1, . . . , n and let Pj be a permutation matrix of dimension mj mj for
j = 1, . . . , n. Then
)
Fm(n1),...,mn (y1 , . . . , yn ) = Fm(n(
(P(1) y(1) , . . . , P (n) y (n) ),
1) ,...,m(n)

for any permutation { (1), . . . , (n)} of the integers {1, . . . , n}. The joint cdf is then invariant to permutations within the
subclusters and between the subclusters themselves. We assume in addition that
1)
Fm(n1),...,mn (y1 , . . . , yn1 , (, . . . , )T ) = Fm(n1,...,
mn1 (y1 , . . . , yn1 ).

(1)

(1)

The cdf of the vector Yj is given by Fmj (yj ) and is assumed to be closed under margins that is Fmj (yj1 , . . . , yj,mj 1 , ) =
(n)
Fm1 ,...,mn

Fmj 1 (yj1 , . . . , yj,mj 1 ). In the notation of [8], the cdf

is h-extendible.
These assumptions have several implications. They imply that all the variables Yij have the same marginal cdf F (y) =
(1)

F1 (y). The dependence between two units of the same subcluster is characterized by a common bivariate cdf given by
(1)

Fs (y1 , y2 ) = F2 ((y1 , y2 )T ) while the common bivariate cdf of two units from two different subclusters of a cluster is
(2)

Fc (y1 , y2 ) = F1,1 (y1 , y2 ). The indices c and s refer to the clusters and the subclusters levels respectively. Two special cases
are of interest. If the N random variables of a cluster are exchangeable, then there are no subcluster effect and Fc (y1 , y2 ) =
Fs (y1 , y2 ). It may also happen that the subclusters within a cluster are independent; this implies Fc (y1 , y2 ) = F (y1 )F (y2 ).
(n )
Examples of families of cdf Fmi1i ,...,min satisfying these conditions are presented in the next two subsections.
i

2.2. The standard normal and elliptical distributions for hierarchical data
The standard normal random effect model for hierarchical data writes
Yij = + ai + bj(i) + ij ,

i = 1, . . . , I , j = 1, . . . , ni , = 1, . . . , mij

(1)

where is the overall mean, ai , bj(i) and ij are independent normal random variables with respective variances a2 , b2 and
2 . A positive value of a2 induces dependence between subclusters of the same cluster. A positive value of b2 makes the
dependence between units of the same subcluster stronger than the dependence between those coming from two different
subclusters of the same cluster. The intra cluster correlations associated to clusters and to subclusters are then respectively

212

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

c = a2 /(a2 + b2 + 2 ) and s = (a2 + b2 )/(a2 + b2 + 2 ). Milliken and Johnson [11], chapter 5, discuss inference
procedures for c and s under model (1).
Under model (1), Yi is a normally distributed random vector with covariance matrix (a2 + b2 + 2 ) (c , s , mi ) where
mi = (mi1 , . . . , mini ), (c , s , mi ) is the Ni Ni matrix given by
(c , s , mi ) = (1 c )INi + s 1Ni 1TNi +

ni

(c s )ej(i) eTj(i) ,

(2)

j =1

where INi is the Ni Ni identity matrix, 1Ni is a Ni 1 vector of 1 and ej(i) the Ni 1 vector whose th entry is 1 if unit
belongs to subcluster j of cluster i and 0 otherwise.
The correlation matrix (s , c , mi ) is positive definite if s c . Therefore, it can be employed to generalize the normal
to hierarchical elliptical distributions; the resulting families of distributions fulfill the conditions presented in Section 2.1.
A generalization of the normal model is obtained by considering gN (xT x), the density of a random vector Y = RZN where
ZN is a N 1 vector of independent N (0, 1) random variables and R is a positive random variable independent of ZN . The
density of the hierarchical elliptical distribution with covariance matrix 2 (s , c , m), associated with gN is given by
gN yT (s , c , m)1 y/ 2

fs ,c (y11 , . . . , ynmn , m) =
where N =

j=1

| 2 (s , c , m)|1/2

y N

(3)

mj . The hierarchical normal and t with degrees of freedom distributions are special cases of (3)

corresponding to g (x) exp(x2 /2) and g (x) (1 + x/v)(+N )/2 , respectively.

For model (3), the bivariate densities for the dependence within subclusters and clusters are respectively given by
g (y1 , y2 , s ) and g (y1 , y2 , c ) where
g2 (y21 2 y1 y2 + y22 )/{(1 2 ) 2 }

g (y1 , y2 , ) =

(1 2 )1/2

y1 , y2 .

Observe that for model (3), a null cluster correlation, c = 0, does not imply that the subclusters are independent, except
for the normal model.
2.3. Hierarchical Archimedean copulas
The hierarchical Archimedean copulas (HAC) [1,7] are a generalization of the multivariate Archimedean copulas reviewed
by McNeil and Neslehov [10], that allows for partial exchangeability. Nested Archimedean copulas capture the dependence
within subclusters with Archimedean copulas and combine them using a second Archimedean copula that accounts for
between subcluster association. In this paper we consider the case where:

the dependence structure within each subcluster is generated using the same generator function s,s with dependence
parameter s ,
the dependence structure between the subclusters of each cluster is characterized by a common generator function c ,c
with dependence parameter c ,
the generators s and c come from the same parametric family: Claytons and Franks are considered in this work.
In this case, if s c then Cs ,c , given by

1
1
Cs ,c ui11 , . . . , uini min , mi = c ,c c,1c s,s s
,s (ui11 ) + + s,s (ui1mi1 )

+ + c,1c s,s s,1s (uini 1 ) + + s,1s (uini mini ) ,

(4)

is a copula.
This is deduced from the following theorem that gives general conditions on the generators s,s and c ,c for C to be a
copula, see [1,9] for a proof.
Theorem 1. Let L denote the class of completely monotone functions,

L = : [0, ) [0, 1] : (0) = 1; () = 0; (1)i (i) 0, i = 1, . . . , .

If r ,r L for r = 1, 2 and if

2,12 1,1 L then C1 ,2 given by (4) is a copula.

If the generators 1,1 and 2,2 belong to the same parametric family then the condition 2,12 1,1

1 2 for most Archimedean copulas. This is particularly true for Claytons, Gumbels and Franks copulas.

L is met if

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

213

When c = 0 units from different subclusters of the same cluster are independent, since Cs ,0 ui11 , . . . , uini min

s,s

s,1s (ui11 )

+ +

s,1s (ui1mi1 )

s,s

s,1s (uini 1 )

+ +

s,1s (uini mini )

. On the other hand, if s = c = ,

all pairs of units drawn from

the same cluster
have the same dependence level, regardless of their subclusters. The

corresponding model is C ui11 , . . . , uini min = {1 (ui11 ) + + 1 (uini min )}. This two level hierarchy is investigated

in [12]. Finally, note that s = c = 0 corresponds to the independence copula C0,0 ui11 , . . . , uini min = ui11 uini min .
Hierarchical Archimedean copulas satisfy the conditions laid out in Section 2.1, they have a uniform marginal distribution.
If F is an arbitrary marginal cdf and if {Uij } has cdf given by (4), then Yij = F 1 (Uij ) are random variables with cdf F whose
dependence is characterized by a hierarchical Archimedean copula. Such random variables are considered in Section 5.

2.4. Association measures

We use Kendalls tau to assess the dependence within and between subclusters. To measure the association between two
units from the same subcluster we consider a pair of bivariate observations (Yij , Yijm ) and (Yi j , Yi j m ) drawn from the jth
subcluster of cluster i and the j th subcluster of cluster i with i = i , respectively. Kendalls tau is the probability that this
pair is concordant minus the probability that it is discordant,

s = 2P {(Yij Yi j )(Yijm Yi j m ) > 0} 1.

(5)

One can express s in terms of the bivariate cdf for units drawn from the same subcluster Fs as

s = 4

y1

Fs (y1 , y2 )dFs (y1 , y2 ) 1.

(6)

y2

The second Kendalls tau c measures the association between two units from different subclusters of the same cluster.
It is defined by considering a pair (Yij , Yikm ) and (Yi j , Yi k m ), with i = i , j = k and j = k ,

c = 2P {(Yij Yi j )(Yikm Yi k m ) > 0} 1.

(7)

As in (6), c can be expressed in terms of Fc , the bivariate cdf for two units drawn from different subclusters of the same
cluster.
For the hierarchical copula families of Clayton and Frank,
Kendalls tau r , r = s, c are respectively given by r /(r + 2)
and 1 4 {D1 (r ) 1} /r , where D1 () = (1/) 0 t / {exp(t ) 1} dt is a Debye function of the first kind. For the
hierarchical elliptical models of Section 2.4, the link between the correlations and Kendalls tau, as given in [5], is

r =

arcsin(r ),

r = s, c .

(8)

3. Estimators for Kendalls tau

3.1. Kendalls tau estimators for nested data
The data set has been presented in Section 2.1. We propose to estimate s and c by considering the

ni ni

ni<i i nji

mij (mij 1)mi j (mi j 1) pairs of ordered bivariate vectors coming from the same subcluster and the i<i
mij
j=k
j =k
mik mi j mi k pairs of ordered bivariate vectors coming from two different subclusters of the same cluster, respectively. The
estimators are sample versions of (5) and (7) and are denoted s and c , respectively

s =
=

c =
=

2
I (I 1)
2
I (I 1)
2
I (I 1)
2
I (I 1)

ni
ni
I

i<i

ni
ni
I

i<i

wii jj

mij mi j

sign (Yij Yi j )(Yijm Yi j m )

mij (mij 1)mi j (mi j 1)
=m =m

wii jj h1 (Yij , Yi j )

ni
ni
I

vii jj kk

i<i j=k j =
k
ni
ni
I

mij mik mi j mi k

sign (Yij Yi j )(Yikm Yi k m )

m

mij mik mi j mi k

vii jj kk h2 (Yij , Yik , Yi j , Yi k )

i<i j=k j =k

where sign(x) is 1, 0, or 1 when x is negative, null, and positive respectively. And {wii jj : 1 i < i I , 1 j
ni , 1 j ni } and {vii jj kk : 1 i < i I , 1 j = k ni , 1 j = k ni } are weights such that

i<i

ni ni
j

wii jj = I (I 1)/2 and

i<i

ni ni
j=k

j =k

vii jj kk = I (I 1)/2.

214

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

For each estimator two sets of weights are considered. For s , the first one gives equal weight to all the possible pairs
{(Yij , Yi j ), i = i }; it has w1,ii jj = 1/(ni ni ). The second one gives the same weight to all the pairs of ordered observations
drawn from the two different subclusters: j in cluster i and j in cluster i
I (I 1)mij (mij 1)mi j (mi j 1)/(ni ni )

w2,ii jj =

ni
i
I
1
1

i1 =i1 j1

j1

mi1 j1 (mi1 j1 1)mi j (mi j 1)/(ni1 ni )

1 1

1 1

For c , the first set of weights gives equal weight to all the possible quadruplets {(Yij , Yik , Yi j , Yi k ), i = i , j = k, j = k },
v1,ii jj kk = 1/{ni (ni 1)ni (ni 1)}. The second one gives the same weight to all the pairs of ordered observations drawn
from the four different subclusters: j = k in cluster i and j = k in cluster i
I (I 1)mij mik mi j mi k /{ni (ni 1)ni (ni 1)}

v2,ii jj kk =

ni
i
I
1
1

k1
i1 =i1 j1 =k1 j1 =

mi1 j1 mi1 k1 mi j mi k /{ni1 (ni1 1)ni (ni 1)}

1

1 1

1 1

The two weighting schemes are compared in Section 3.4 where the resulting estimators of Kendalls tau are denoted s , s ,
c and c respectively.
When investigating the sampling distributions of (s , c ) it is convenient to define Ai and Bi as the estimations of s and
c respectively, obtained from Yi ,
ni
ni
I

Ai =

i =i j

wii jj h1 (Yij , Yi j )/(I 1)

ni
ni
I

i =i j
ni

ni

Bi =

(9)

wii jj /(I 1)

vii jj kk h2 (Yij , Yik , Yi j , Yi k )/(I 1)

i =i j=k j =k

ni
ni
I

i =i j=k j =k

.
vii jj kk /(I 1)

Let w
i and v i be the denominator of the fractions for Ai and Bi respectively. Both sequences have an average value of 1 and
one has

s =

I
1

I i =1

w
i Ai

I
1

c =

I i =1

v i Bi .

For the first set of weights, w

i = v i = 1; i = 1, . . . , I and the clusters are given an equal weight; this is not so with the
second set of weights which gives larger weights to big clusters. The asymptotic distribution of (s , c )T is given next; its
proof can be found in Appendix A.
Proposition 1. Let {Yi : i = 1, . . . , I } represent independent hierarchical random vectors whose cdfs fulfill
the conditions ofT
Section 2.1. If as I , ni and mij , i = 1, . . . , I , j = 1, . . . , niare bounded,
then
the
limiting
distribution
of
I (s s ; c c )

is a centered bivariate normal with covariance matrix equal to 1

12

1 = 4 lim

I
1

I i=1

2 = 4 lim

I
1

I i=1

Var

n
i

12
2

where 1 , 2 and 12 are given by

w
ij g1 (Yij ) ,

(10)

Var

n
i

v ijk g2 (Yij , Yik ) ,

(11)

j=k

and

12 = 4 lim

I
1

I i =1

Cov

ni

w
ij g1 (Yij );

ni

j1 =k1

v ij1 k1 g2 (Yij1 , Yik1 ) ,

(12)

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

215

where

w
ij =

ni
I

wii jj /(I 1),

ni
I

v ijk =

i ,i =i j =1

vii jj kk /(I 1),

i ,i =i j =k

mij

2 Fs (Yij , Yijm ) + Fs (Yijl , Yijm ) 1

g1 (Yij ) =

l=m

1 ,

mij (mij 1)
mij mik

g2 (Yij , Yik ) =

2 Fc (Yij , Yikm ) + Fc (Yij , Yikm ) 1

2 ,

mij mik

with Fs (y1 , y2 ) = 1 F (y1 ) F (y2 ) + Fs (y1 , y2 ) and Fc (y1 , y2 )

= 1 F (y1 ) F (y2 ) + Fc (y1 , y2 ).
A consistent estimator of the asymptotic covariance matrix of I (s s ; c c )T is given by

T
w
i (Ai s )
w
i (Ai s )
=
v i (Bi c )
v i (Bi c )
I i

where Ai , Bi , w
i = j w
ij , v i = j=k v ijk have been defined in Eq. (9).

1
12

12
2

I
4

(13)

The variances (10) and (11) and the covariance (12) are evaluated in Appendix B by expanding the variances and the
covariance of sums as sums of variances and covariances. This leads to the following expressions for the limiting covariance
matrix of Proposition 1,

1 = 16 lim

ni
I
1

2 = 16 lim

2
2v ijk

mij mik

j=k

ni

2V1,c + 4(mij 2)V1,c + (mij 2)(mij 3)V3,s + w

ij

w
ik V4,s ,

(14)

k,k=j

V1,c + (mij + mik 2)V2,c + (mij 1)(mik 1)V3,c

ni

4v ijk
mij

mij (mij 1)

ni
I
1

w
ij2

v ijk1 V4,c + (mij 1)V5,c + v ijk

k1 ,k1 =k,j

ni

ni

v ij1 k1 V6,c

(15)

j1 ,j1 =j,k k1 ,k1 =j,k,j1

and

12 = 16 lim

ni
I
1

w
ij

ni

v ijk1
k1 ,k1 =j

mij

4V1,s,c + 2(mij 2)V2,s,c +

ni

ni

v ij1 k1 V3,s,c

(16)

j1 ,j1 =j k1 ,k1 =j,j1

where
V1,s = Var [Qs (Y111 , Y112 )] ,

V2,s = Cov [Qs (Y111 , Y112 ); Qs (Y111 , Y113 )] ,

V3,s = Cov [Qs (Y111 , Y112 ); Qs (Y113 , Y114 )] ,

V1,c = Var {Qc (Y111 , Y121 )} ,

V4,s = Cov [Qs (Y111 , Y112 ); Qs (Y121 , Y122 )] ,

V2,c = Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y122 )} ,

V3,c = Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y122 )} ,

V4,c = Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )} ,

V5,c = Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )} ,

V6,c = Cov {Qc (Y111 , Y121 ); Qc (Y131 , Y141 )}

V1,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y111 , Y121 )} ,

V2,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y113 , Y121 )} ,

V3,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y131 )} ,

with Qs (Y1 , Y2 ) = Fs (Y1 , Y2 ) + Fs (Y1 , Y2 ) and Qc (Y1 , Y2 ) = Fc (Y1 , Y2 ) + Fc (Y1 , Y2 ).
Note that if all the subclusters contain a single
observation i.e. mij = 1 for all i and j, under the two weighting schemes,
v ijk = 1/{ni (ni 1)} and 2 is the limit of (16/I ) i {2V1,c + 4(ni 2)V4,c + (ni 2)(ni 3)V6,c }/{ni (ni 1)}. This is equal
to the asymptotic variance of the estimator of the exchangeable Kendalls tau for two level cluster data given in [12] with I
clusters of sizes ni , i = 1, . . . , I.
3.2. Special cases
This section considers special cases: the independence between subclusters and the lack of a subcluster effect. If the
subclusters are independent, then Fc (y1 , y2 ) = F (y1 )F (y2 ), Qc (Y1 , Y2 ) = F (Y1 )F (Y2 ) + {1 F (Y1 )}{1 F (Y2 )} and F (Y1 )
and F (Y2 ) are independent random variables uniformly distributed on (0, 1). Then one has V1,c = 1/36, V2,c = sp /36,
2
V3,c = sp
/36 and V4,c = V5,c = V6,c = 0, where sp = 12Cov[F (Y111 ), F (Y112 )] is the subcluster Spearmans rho. Then (15)

216

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

simplifies to

2,0 =

2
v ijk

ni
I
1

lim

9 I I

mij mik

j=k

{1 + (mij 1)sp }{1 + (mik 1)sp }.

(17)

On the other hand when there is no subcluster effect, Fs = Fc and the model involves I clusters of sizes
j mij , i =
1, . . . , I. Then Qs = Qc = Q , V1,s = V1,c = V1 , V2,s = V2,c = V4,c = V1,s,c = V2 , V3,s = V4,s = V3,c =
V5,c = V6,c = V2,s,c = V3,s,c = V3 where V1 = Var [Q (Y111 , Y112 )], V2 = Cov [Q (Y111 , Y112 ); Q (Y111 , Y113 )] and
V3 = Cov [Q (Y111 , Y112 ); Q (Y113 , Y114 )]. Thus (14)(16) simplify to

2w
ij2

ni

(mij 2)(mij 3)
w
ik V3 ,
1 = 16 lim
{V1 + 2(mij 2)V2 } + w
ij
w
ij +
I I
mij (mij 1)
mij (mij 1)
i
j
k,k=j

ni
ni
2
I

2v ijk
1
mij + mik 2
2
2 = 16 lim
V1 + 2v ijk
v ijk +
v ijk1
I I
mij mik
mij mik
mij k ,k =j,k
i
1 1
j=k

ni
ni
ni

(mij 1)(mik 1)
mij 1
V2 + v ijk 2
v ij1 k1 V3 ,
v ijk1 +
v ijk + 4
ni
I
1

mij mik

ni

12 = 16 lim

1
I

w
ij

4
mij

mij

ni

v ijk1 V2 + 2

k1 ,k1 =j

k1 ,k1 =j,k

j1 ,j1 =j,k k1 ,k1 =j,k,j1

ni

mij 2
mij

v ijk1 +

k1 ,k1 =j

ni

ni

v ij1 k1 V3 .

j1 ,j1 =j k1 ,k1 =j,j1

3.3. The balanced case

Consider now the case corresponding to ni = n, i = 1, . . . , I and mij = m, j = 1, . . . , n. The two weightings w1,ii jj
ij = 1/n and v ijk = 1/{n(n 1)} for i = 1, . . . , I and
and w2,ii jj (respectively v1,ii jj kk and v2,ii jj kk ) are equivalent and w
j, k = 1, . . . , n. The asymptotic variances (14) and (15) and the covariance (16) have then the convenient forms

1 =
2 =

12 =

16
nm(m 1)
16

2V1,s + 4(m 2)V2,s + (m 2)(m 3)V3,s + (n 1)m(m 1)V4,s ,

2V1,c + 4(m 1)V2,c + 2(m 1)2 V3,c + 4(n 2)

2
n(n
1)m

m V4,c + (m 1)V5,c + (n 2)(n 3)m2 V6,c ,

16
4V1,s,c + 2(m 2)V2,s,c + (n 2)mV3,s,c .
nm

If the subclusters are independent, e.g. Fc (y1 , y2 ) = F (y1 )F (y2 ), then (17) reduces to

2 =

8
9n(n 1)m2

1 + (m 1)sp

(18)

In addition, we have V4,s = 0 and 1 simplifies to the asymptotic variance of Kendalls tau estimator in the clustered data
case with In clusters of size m, computed in [12].
3.4. Monte Carlo evaluation of the finite sample properties
A first set of simulations was conducted to assess the finite sample performances of s and c and of their variance
estimators, and to compare the two weighting schemes described in Section 2.4. The number of clusters was set to
I = 20 and 1000 unbalanced samples were generated using nested normal and Student multivariate distributions such
that (s , c ) {(0.7, 0.2) , (0.7, 0.5) , (0.5, 0.2)}. The cluster size ni and the mij were either 2 or 3 with means of n = 2.5
= 2.36. The same size vectors mi , i = 1, . . . , 20, were used in all the simulations. The expectations and the variances
and m
of the estimators r , r = s, c were evaluated together with the expectation of the variance estimators r /I , r = 1, 2 and
the coverage of the 95% confidence intervals, r 1.96(r /I )1/2 , r = s, c.
The results, presented in Table 1, show that the four estimators are virtually unbiased and that the variance estimators
have a small negative bias. This shows up in the coverage of the confidence intervals that are slightly below their nominal
value of 95%. Table 1 also shows that the first weighting scheme leads to uniformly smaller variances. Thus the equal weight
strategy leads to better estimators of the parameters. It also gives slightly better coverage rates for s than the unequal
weights.

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

217

Table 1
Empirical averages and variances (103 ) of s , s , c and c , along with the empirical averages of the their estimated variances (103 ) and the coverage
rate of the 95% confidence interval, for I = 20 based on 1000 simulated samples.

Normal

Student(5)

0.7

0.2

0.698
(2.40)
(2.41)
92.1

0.201
(15.0)
(13.9)
90.6

0.698
(2.67)
(2.60)
92.2

0.201
(15.5)
(14.2)
90.7

0.698
(3.17)
(2.89)
90.7

0.198
(15.7)
(15.8)
92.7

0.698
(3.35)
(3.13)
91

0.197
(16.2)
(16.0)
93.2

0.7

0.5

0.705
(3.20)
(3.04)
92

0.506
(9.62)
(8.75)
89.1

0.705
(3.44)
(3.21)
90.3

0.506
(9.76)
(8.97)
88.8

0.696
(4.40)
(3.85)
91.1

0.498
(11.0)
(10.3)
90.6

0.698
(4.52)
(4.01)
89.9

0.499
(11.4)
(10.5)
91.1

0.5

0.2

0.499
(5.08)
(5.03)
92.5

0.203
(11.6)
(10.5)
91

0.499
(5.56)
(5.17)
91.1

0.203
(12.1)
(10.5)
91.2

0.497
(6.39)
(5.83)
91.8

0.197
(12.4)
(11.9)
91.6

0.498
(6.80)
(5.85)
90.5

0.197
(12.7)
(11.9)
91.4

4. Estimators for the intra cluster correlations of elliptical distributions

In this section we use s and c to estimate the correlations s and c for elliptical distributions with correlation matrix
given by (2); they are compared with moment estimators derived using the standard ANOVA table for hierarchical data. Our
model for Yi is an elliptical distribution with density (3) defined in Section 2.2. Only the balanced case is considered with m
and n standing for the sizes of the subclusters and of the clusters respectively.
The moment estimators of s and c are presented on page 633 of [11]. They are expressed in terms of the mean
I

2
squares for the clusters, MSC =
i=1 nm(Yi.. Y... ) /{(I 1)}, for the subclusters nested in the clusters, MSS (C ) =

i =1

j =1

s =
c =

m(Yij. Yi.. )2 /{I (n 1)} and for the errors, MSE =

i=1

n m
j =1

k=1

(Yijk Yij. )2 /{In(m 1)}. They are given by

MSC + (n 1)MSS (C ) nMSE

MSC + (n 1)MSS (C ) + n(m 1)MSE
MSC MSS (C )
MSC + (n 1)MSS (C ) + n(m 1)MSE

(19)

The alternative estimators of the two correlation coefficients are based on the exchangeable Kendalls tau. From (8) they
are

r , = sin

r ,

r = s, c .

These two sets of estimators were compared by Monte Carlo simulation. The vectors Yi were simulated using the
normal and Student elliptical distributions with correlation matrix (2) with (s , c ) {(0.9, 0.7) , (0.9, 0.3) , (0.7, 0.3)}.
The simulations used I = 20, n = 3 and m {2, 5}. Four estimates were calculated s, , s , c , , c . The Monte Carlo
expectations and the Monte Carlo variances of the four estimators are reported in Table 2; they were evaluated using 1000
repetitions.
Table 2 shows that the estimators are nearly unbiased. As expected r outperforms r , , r = s, c, for the multivariate
normal distribution. The loss of efficiency can be as large as 17% for s, and 12% for c , . While for the multivariate Student
distribution r , uniformly outperforms r , r = s, c with a gain of efficiency as large as 40% for s, and 28% for c , .
5. Hypothesis tests
5.1. Description of the test statistics and Monte Carlo comparisons
This section investigates tests for the two null hypotheses that the cluster, respectively the subcluster, has no impact
on the variability of the data. The null hypothesis are H0s : s = c and H0c : c = 0. The standard test statistics for
these hypotheses are F -tests constructed with the mean squared MSC , MSS (C ), and MSE introduced in Section 4. They are
compared with test statistics based on Kendalls tau estimators proposed in Section 3. They reject the null hypothesis H0s if
s c is large while H0c is rejected for large values of c .
Rather than using the asymptotic covariance matrix of Proposition 1 to construct test statistics we propose using
permutation tests. To test H0s , permutations of the observations are carried out within the clusters. This creates new
subclusters and the permutation distribution is obtained by calculating s c for each new set of subclusters. The
permutation p-value is then estimated by the proportion of permuted samples for which s c is larger than its observed

218

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

Table 2
Empirical averages and variances (103 ) of s, , s , c , and c , for I = 20, based on 1000 simulated samples with n = 3.

0.9

0.3

Normal

2
5

0.7

2
5

0.7

0.3

2
5

Student(5)

s,

c ,

s,

c ,

0.896
(0.87)
0.897
(0.58)
0.897
(1.45)
0.896
(0.93)

0.895
(0.72)
0.895
(0.49)
0.893
(1.23)
0.891
(0.80)

0.299
(23.6)
0.284
(20.1)
0.696
(9.76)
0.687
(9.43)

0.289
(20.6)
0.276
(17.8)
0.684
(8.83)
0.675
(8.24)

0.898
(1.14)
0.896
(0.67)
0.896
(1.55)
0.898
(1.06)

0.895
(1.26)
0.893
(0.75)
0.890
(1.94)
0.888
(1.46)

0.295
(24.6)
0.298
(23.7)
0.689
(12.71)
0.694
(10.3)

0.289
(29.8)
0.280
(29.8)
0.669
(15.68)
0.672
(13.2)

0.697
(5.62)
0.697
(3.16)

0.693
(4.79)
0.692
(2.84)

0.292
(16.5)
0.296
(14.5)

0.283
(14.6)
0.288
(12.7)

0.695
(7.20)
0.698
(3.63)

0.689
(8.51)
0.687
(5.08)

0.294
(19.1)
0.301
(16.0)

0.28
(23.1)
0.289
(19.9)

value. To test H0c , the subclusters are

the permutation units; the null distribution is obtained by calculating c on samples
obtained by assigning randomly the
ni subclusters to the I clusters.
We conducted simulations to compare the power of the standard F tests and of the permutation tests using Kendalls
tau of Section 3. The simulations used balanced data with I = 20, n = 2 and m = 3; all tests were carried out at the
5% level. The random vectors Yi were generated using Claytons and Franks copulas and two marginal distributions, the
standard normal and the t with 5 degrees of freedom. To estimate the power of the test for H0c , we fixed s = 0.7 and let r
increase from 0 to 0.7. For the test for H0s , we fixed c = 0.3 and the power was evaluated by letting s vary between 0.3 and
1. Each permutation test used 500 permutations and the power was evaluated by calculating the percentage of rejection
among 1000 Monte Carlo trials. Smoothed power curves for the permutation tests based on Kendalls tau and the F tests are
reported in Figs. 1 and 2.
For H0s : s = c , Fig. 1 reveals that the two F tests and Kendalls tau tests are comparable, except possibly for t (5)
margins and Franks copula where the test based on Kendalls tau is more powerful. In Fig. 2 the two tests for H0c : c = 0
are comparable for Franks copula while the F test is better for Claytons.
5.2. Pitman relative efficiency
To pursue our investigation of the tests for H0c , we calculated the Pitman efficiency of the test based on c with respect
to the ANOVA F -test for balanced samples, when the alternative hypothesis is expressed in terms of Claytons and Franks
copulas. In large samples, the F -test is equivalent to a test that rejects H0c if c is large, where c is defined in (19). Using
Theorem 14.19 in [14], the Pitman efficiency of the test based on Kendalls tau with respect to the F test is

e ,m,n =

1/2

1/2

c (0)/2

2
,

where 2 = 8{1 + (m 1)sp }2 /{9n(n 1)m2 } is the asymptotic variance of I 1/2 c under H0c as given in (18), 22 =
{1 + (m 1)1 }2 /{n(n 1)m2 } is the corresponding asymptotic variance for I 1/2 c derived in the appendix with 1 equal
to the within subcluster correlation, and c (0) is the derivative of c , expressed as a function of c , at c = 0. This derivative
does not depend on neither m nor n therefore
e ,m,n =

2
{1 + (m 1)1 }2
1 + (m 1)1
=
e
,
,1,n
8c (0)2 {1 + (m 1)sp }2
1 + (m 1)sp
9

(20)

where e ,1,n is the Pitman efficiency of a test of independence based on Kendalls tau in the two level hierarchical model
studied in [12]. In (20), the within subcluster correlation 1 is equal to

1 =

R2

(y1 F ) (y2 F )dFs (y1 , y2 )/F2 ,

(21)

where F and F2 are the mean and the variance of F . In a similar way, sp is the correlation between F (Y111 ) and F (Y112 ).
Values of 1 are given in Table 3 for the copulas and the marginal distributions used in the simulations. In most cases sp > 1
and the efficiency (20) decreases with m.
Pitman efficiencies are reported in Table 4. The m = 1 efficiencies are taken from [12]. The low efficiencies for Claytons
copula agree with the findings of Genest & Verret (2005) [3] who showed that the test of independence based on Kendalls
tau performs poorly for Claytons alternatives. Fig. 2 has 1 = 0.7 and m = 3. In the first row, the F test is better; this agrees
with the Clayton efficiencies in Table 4. The same is true of the second row, for instance Kendalls tau is slightly better for
Frank copula and a t (5) margin; the corresponding Pitman efficiency is 1.14.

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

219

Fig. 1. Empirical power comparison of the tests for H0s using Kendalls tau (solid line) and a F statistic (dashed line) for data generated using Claytons
(row 1) and Franks (row 2) copulas with standard normal (column 1) and t (5) (column 2) margins; the power is given in terms of the difference
s c .

Table 3
Values of 1 for various copulas and marginal distributions; the row for uniform margins gives Spearmans rho, sp .
Clayton

Frank

0.2

0.5

0.7

0.2

0.5

0.7

t (3 )
t (5 )
N (0, 1)
Uniform

0.313
0.3228
0.316
0.294

0.626
0.667
0.684
0.682

0.706
0.808
0.850
0.872

0.202
0.255
0.282
0.296

0.476
0.596
0.659
0.694

0.629
0.774
0.846
0.887

Table 4
Pitman efficiencies of the test based on 2 with respect to the F test for H0s .
Clayton
m=1

s
t(3)
t(5)
N (0 , 1 )

1.07
0.85
0.84

Frank
m=3

m = 10

m=1

0.2

0.5

0.7

0.2

0.5

0.7

1.13
0.91
0.89

0.98
0.83
0.85

0.83
0.77
0.82

1.18
0.97
0.94

0.93
0.82
0.85

0.74
0.74
0.81

2.14
1.35
1.10

m=3

m = 10

0.2

0.5

0.7

0.2

0.5

0.7

1.66
1.21
1.06

1.43
1.13
1.03

1.42
1.14
1.03

1.26
1.09
1.02

1.14
1.04
1.00

1.18
1.06
1.01

220

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

Fig. 2. Empirical power comparison of the tests for H0c using Kendalls tau (solid line) and a F statistic (dashed line) for data generated using Claytons
(row 1) and Franks (row 2) copulas with standard normal (column 1) and t (5) (column 2) margins.

Table 5
Point estimates of s and c with the two weightings and their 95% confidence intervals.

RMT85
REV84

0.299
[0.200, 0.397]
0.246
[0.209, 0.282]

0.225
[0.118,0.331]
0.184
[0.084, 0.283]

c
0.082

[0.083, 0.247]
0.098

[0.005, 0.201]

c
0.031

[0.119, 0.181]
0.088

[0.026, 0.202]

6. Illustration

Consider the data set MU284 presented in [13], Appendix B. It consists of

mij = 284 Swedish municipalities clustered

in
ni = 50 neighborhoods (subclusters) and in I = 8 geographical regions (clusters). The average number of subclusters
by cluster is 6.25 and the average subclusters size is 5.68. The variables of interest are RMT85 and REV84 the revenues from
the 1985 municipal taxation and the real estate values according to 1984 assessment, respectively.
Table 5 provides the estimates of s and c with the two weighting schemes along with the respective 95% confidence
intervals. Table 6 presents the p-values of the c and s c permutation based tests computed with 500 permutations.
These variables have skewed non normal distributions and the assumptions for the validity of the ANOVA tests are not met;
thus the 11.9% and 10.3% levels of the F -tests for both REV84 and RMT85 variables are misleading. Only the tests based on c
and s c are reliable. We conclude that both RMT85 and REV84 present a significant association between municipalities
of the same neighborhood. However the region has a small effect for REV84 and is not significant at the 5% level for RMT85.
This is contradictory with the 95% confidence interval obtained for REV84 which contains 0. A possible explanation is that
I = 8 may not be large enough for the asymptotic variances to be reliable. The permutation test based on the statistic c is
more trustworthy.

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

221

Table 6
p-values of the c and s c based tests and of the ANOVA F -tests.
H0C

Null hypothesis

H0s

Test statistic
RMT85
REV84

0.062
0.028

s c

0.228
0.023

<10
<103

0.119
0.103

From a survey sampling perspective, the small c means that 8 geographical regions are not very different; stratifying by
regions would not improve the precision of survey estimates much. It is the municipalities that differ from one neighborhood
to the next. Thus a cluster sample of neighborhoods would not give estimates as precise as a random sample of municipalities
of the same size.
7. Conclusion
This paper introduces non parametric association measures for hierarchical data based on an exchangeable Kendalls
tau. The sampling properties of the proposed statistics have been investigated under a general semi-parametric model for
hierarchical data. They provide flexible modeling tools for nested data when the normality assumption fails. These statistics
have been applied to estimate the intra-cluster correlation coefficients of nested elliptical distributions. Two intra class
independence tests based on the proposed measures have been shown to be better than the standard methods based on
ANOVA tables when the assumption of normality is not met.
Appendix A. Proof of Proposition 1
Let a1 and a2 be two constants. The proof consists in showing that, asymptotically, any linear combination a1 s + a2 c of

s and c has a normal distribution with variance equal to a21 1 + a22 2 + 2a1 a2 12 . Recall that

mij mi j

2 1 (Yij Yi j )(Yijm Yi j m ) > 0 1

h1 (Yij , Yi j ) =
mij (mij 1)mi j (mi j 1)
l=m =m
and
h2 (Yij , Yik , Yi j , Yi k ) =

mij mik mi j mi k

2 1 (Yij Yi j )(Yikm Yi k m ) > 0 1

mij mik mi j mi k

Thus,
a1 s + a2 c =

ni
ni
I

2
I (I 1)

i=i

ha1 a2 ,ii jj (Yij , Yi j ),

where
ha1 a2 ,ii jj (Yij , Yij ) = a1 wii jj h1 (Yij , Yi j ) + a2

ni
ni

vii jj kk h2 (Yij , Yik , Yi j , Yi k ).

k,k=j k ,k =j

Since a1 s + a2 c is bounded, it has a finite variance. The Hjek projection of a1 s + a2 c {a1 s + a2 c } is given by

a1 s + a2 c {a1 s + a2 c } =

a1 s + a2 c {a1 s + a2 c } |Yi .

i=1

Now since
E {1 (Yij Yi j )(Yijm Yi j m ) > 0 |Yi } = Fs (Yij , Yijm ) + Fs (Yij , Yijm ) and

E {1 (Yij Yi j )(Yikm Yi k m ) > 0 |Yi } = Fc (Yij , Yikm ) + Fc (Yij , Yikm ),

E ha1 a2 ,ii jj (Yij , Yi j )|Yi is given by

mij

a1 wii jj

=m

2 Fs (Yij , Yijm ) + Fs (Yij , Yijm ) 1

mij (mij 1)

= a1 wii jj g1 (Yij ) + a2

ni
ni

ni

+ a2

k,k=j k ,k =j

vii jj kk g2 (Yij , Yik ),

k,k=j k ,k =j

where g1 (Yij ) and g2 (Yij , Yik ) are defined in Proposition 1.

mij mik

ni

vii jj kk

2 Fc (Yij , Yikm ) + Fc (Yij , Yikm ) 1

mij mik

222

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

Thus the Hajek projection of a1 1 + a2 2 {a1 1 + a2 2 } is

ni
I
2

a1 w
ij g1 (Yij ) + a2

ni

I
2

v ijk g2 (Yij , Yik ) =

{ a1 w
i A0i + a2 v i B0i },

k=j

(A.1)

i
i
ijk g2 (Yij , Yik )/vi . Since (A0i , B0i ) are for i = 1, . . . , I bounded
ij g1 (Yij )/w
i and B0i =
where A0i =
j w
k=j v
independent
random
vectors
with
a
null
expectation
the
central limit theorem implies that the asymptotic distribution

of IH a1 1 + a2 2 {a1 1 + a2 2 } is normal. The variance of the Hajek projection is given by the expectation of

4( a21 w
i2 A20i + 2a1 a2 w
i v i A0i B0i + a22 v i2 B20i )/I 2 . The estimator proposed in Proposition 1 is obtained by replacing, in this
random variable, A0i and B0i by their convergent predictors (Ai 1 )2 and (Bi 2 )2 .
Applying theorem 11.2 in [14], to complete the proof we need to show that the variance of the Hajek projection divided
by that of a1 1 + a2 2 converges to 1. From (A.1) the variance of the Hajek projection is

4
I2

a21

Var

n
i

w
ij g1 (Yij ) +

a22

j =1

i=1

+ 2a1 a2

Cov

Var

i =1

n
i

i=1

w
ij g1 (Yij ),

ni

n
i

v ijk g2 (Yij , Yik )

j=k

v ij1 k1 g2 (Yij1 , Yik1 )

j1 =k1

As 1 is a weighted sum of random variables, its variance is a sum of variances plus a sum of covariances involving
Cov{h1 (Yij , Yi j ), h1 (Yi1 k , Yi k1 )}. The sum of the variances is O(1/I 2 ) and is asymptotically negligible. A covariance is non
1

null if one of (i, i ) is equal to one of (i1 , i1 ); there are 4 ways in which this can happen thus
Var [1 ] = 4 {I (I 1)}

ni ni1
ni
I
I

i=i i1 =i j,k

wii jj wii1 kk1 Cov[h1 (Yij , Yi j ), h1 (Yik , Yi1 k1 )] + O(1/I 2 ).

(A.2)

k1

In a similar way,
Var [2 ] = 4 {I (I 1)}2

ni
ni
ni
ni
I
I
1

i=i i1 =i j=k j =
k j1 =k1 j1 =
k1

vii jj kk vii1 j1 j1 k1 k1

Cov[h2 (Yij , Yik , Yi j , Yi k ), h2 (Yij1 , Yik1 , Yi1 j1 , Yi1 k1 )] + O(1/I 2 )

and
n

Cov[1 , 2 ] = 4 {I (I 1)}

ni
i1
ni
ni
I
I
I

i ,i =i i ,i =
i
1 1

j1 =k1

j1 =k1

wii jj vii1 j1 j1 k1 k1

Cov[h1 (Yij , Yi j ), h2 (Yij1 , Yik1 , Yi1 j1 , Yi1 k1 )] + O(1/I 2 ).

Now, by conditioning on Yi one can prove that
Cov[h1 (Yij , Yi j ), h1 (Yik , Yi1 k1 )] = E E h1 (Yij , Yi j ) 1 |Yi E h1 (Yik , Yi1 k1 ) 1 |Yi

= Cov{g1 (Yij ), g1 (Yik )}.

In a similar way, one shows that
Cov[h2 (Yij , Yik , Yi j , Yi k ), h2 (Yij1 , Yik1 , Yi1 j , Yi1 k )] = Cov{g2 (Yij , Yik ), g2 (Yij1 , Yik1 )}
1

Cov[h1 (Yij , Yi j ), h2 (Yij1 , Yik1 , Yi j , Yi k )] = Cov{g1 (Yij ), g2 (Yij1 , Yik1 )}.

1 1

1 1

Now from (A.2), neglecting O(1/I 2 ) terms,

Var [1 ] = 4 {I (I 1)}2

ni ni1
ni
I
I

i=i i1 =i j,k

= 4I 2

ni
I

= 4I

wii jj wii1 kk1 Cov{g1 (Yij ), g1 (Yik )}

k1

w
ij w
ik Cov{g1 (Yij ), g1 (Yik )}

j ,k

Var

n
i

w
ij g1 (Yij ) .

j=1

This shows that 1 and the Hajek projection of 1 1 have the same asymptotic variance. The proofs for 2 and for the
covariance between 1 and 2 use similar arguments that are omitted. This ends the proof of Proposition 1.

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

223

Table B.7
Covariances needed to compute the asymptotic variance of s with fixed i, j and = m.
k

1 , m1

Covariance

Frequency

k=j

{1 , m1 } = {, m}
1 = or m, m1 = m,
1 = , m, m1 = or m
1 = , m, m1 = , m, 1

Var {Qs (Y111 , Y112 )}

Cov {Qs (Y111 , Y112 ); Qs (Y111 , Y113 )}
Cov {Qs (Y111 , Y112 ); Qs (Y113 , Y112 )}
Cov {Qs (Y111 , Y112 ); Qs (Y113 , Y114 )}

2
2(mij 2)
2(mij 2)
(mij 2)(mij 3)

k = j

1 = m1

Cov {Qs (Y111 , Y112 ); Qs (Y121 , Y122 )}

mik (mik 1)

Table B.8
Covariances needed to compute the asymptotic variance of c with fixed i, j = k, j1 = k1 and m.
j1 , k1

1 , m1

Covariance

Frequency

j1 = j k1 = k

1
1
1
1

Var {Qc (Y111 , Y121 )}

Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y122 )}
Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y122 )}
Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y122 )}

2
2(mik 1)
2(mij 1)
2(mik 1)(mij 1)

j1 = j, k1 = k, j

1 = , any m1
1 = , any m1

Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )}

Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )}

mik1
(mij 1)mik1

j1 = k, k1 = j, k

1 = m, any m1
1 = m, any m1

Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )}

Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )}

mik1
(mik 1)mik1

j1 = j, k, k1 = j

any 1 , m1 =
any 1 , m1 =

Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )}

Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )}

mij1
(mij 1)mij1

j1 = j, k, k1 = k

any 1 , m1 = m
any 1 , m1 = m

Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )}

Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )}

mij1
(mik 1)mij1

j1 = j, k, k1 = k, j, j1

any 1 , any m1

Cov {Qc (Y111 , Y121 ); Qc (Y131 , Y141 )}

mij1 mik1

= , m1 = m
= , m1 = m
= , m1 = m
= , m1 =

Appendix B. Explicit expressions of 1 , 2 and 12

The asymptotic variance 1 is the limit of
ni
I
4

I i=1 j,k=1

w
ij w
ik Cov[g1 (Yij ), g1 (Yik )] =

mij
mik

Cov Qs (Yij , Yijm ); Qs (Yik1 , Yikm1 )

,
16w
ij w
ik
mij (mij 1)mik (mik 1)
j,k=1
=m =m

ni
I
1

I i=1

where Qs (, ) is defined in Section 3.1. To evaluate this expression one has to consider the cases enumerated in Table B.7
where the indices i, j and = m are fixed. The feasible covariances and their frequencies are given in the table for the
possible configurations of the indices. For instance, if k = j, there are two ways in which {1 , m1 } = {, m}, namely
1 = , m1 = m and 1 = m, m1 = . Now if 1 is equal to one of (, m), there are (mij 2) possible values of m1 for
a total of2(mij 2) cases. By symmetry, one gets the same result if m1 is equal to one of (, m) while 1 is not. When
{1 , m1 } {, m} = , there are (mij 2)(mij 3) ways to select 1 = m1 . When k = j equalities between {1 , m1 } and
{, m} is not an issue anymore since these indices represent units in different subclusters and there are mik (mik 1) possible
values for {1 , m1 }. In (14) these frequencies are multiplied by mij (mij 1) to account for the possible values of the indices
= m.
In Section 3.1, Var [Qs (Y111 , Y112 )] = V1,s , Cov [Qs (Y111 , Y112 ); Qs (Y111 , Y113 )] = V2,s , Cov [Qs (Y111 , Y112 ); Qs (Y113 , Y114 )] =
V3,s , Cov [Qs (Y111 , Y112 ); Qs (Y121 , Y122 )] = V4,s .
Consider now the evaluation of 2 which is the limit of

ni
ni
I
4

I i =1 j =
k

ni
ni
I
1

I i =1 j =
k

v ijk v ij1 k1 Cov[g2 (Yij , Yik ); g2 (Yij1 , Yik1 )]

j1 =k1

j1 =k1

16v ijk v ij1 k1

mij mik mij1 mik1

mij mik mij1 mik1

Cov[Qc (Yij , Yikm ); Qc (Yij1 1 , Yik1 m1 )].

m1

The possible values of Cov[Qc (Yijl , Yikm ); Qc (Yij1 l1 , Yik1 m1 )] and their frequencies are given in Table B.8 for fixed values of i,
j = k, j1 = k1 , , and m. To get (15), the frequencies in Table B.8 are multiplied by mij mik to account for the possible values
of the indices and m.
In Table B.8 the frequencies for the cases corresponding to j1 = j, k1 = k include those for j1 = k, k1 = j; thus
the 2 possible variances in the first row of Table B.8 correspond to the cases j1 = j, k1 = k, 1 = , m1 = m and
j1 = k, k1 = j, 1 = m, m1 = . In Section 3, V1,c = Var {Qc (Y111 , Y121 )}, V2,c = Cov {Qc (U111 , U121 ); Qc (Y111 , Y122 )}, V3,c =
Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y122 )}, V4,c = Cov {Qc (Y111 , Y121 ); Qc (Y111 , Y131 )}, V5,c = Cov {Qc (Y111 , Y121 ); Qc (Y112 , Y131 )},
V6,c = Cov {Qc (Y111 , Y121 ); Qc (Y131 , U141 )}.

224

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

Table B.9
Covariances needed to compute the asymptotic covariance of s and c , the indices i, j, j1 = k1 , = m are
fixed.
j1 = k1

1 , m1

Term

Frequency

j1 = j, k1 = j

1 = , any m1
1 = m, any m1
1 = , m, any m1

Cov {Qs (Y111 , Y112 ); Qc (Y111 , Y121 )}

Cov {Qs (Y111 , Y112 ); Qc (Y112 , Y121 )}
Cov {Qs (Y111 , Y112 ); Qc (Y113 , Y121 )}

mik1
mik1
mik1 (mij 2)

j1 = j, k1 = j

any 1 , m1 =
any 1 , m1 = m
any 1 , m1 = , m

Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y111 )}

Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y112 )}
Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y113 )}

mij1
mij1
mij1 (mij 2)

j1 = j, k1 = j, k,

any 1 , any m1

Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y131 )}

mij1 mik1

The asymptotic covariance 12 is the limit of

ni
ni
I
4

I i =1

w
ij v ij1 k1 Cov[g1 (Yij ); g2 (Yij1 , Yik1 )]

j1 =k1

ni
ni
I
1

I i=1

mij mij1 mik1

16w
ij v ij1 k1

j1 =k1

mij (mij 1)mij1 mik1 =m

1

Cov[Qs (Yij , Yijm ); Qc (Yij1 1 , Yik1 m1 )].

m1

The possible values of the covariance and their frequencies are given in Table B.9. The indices i, j, j1 = k1 , , and m are fixed
and, to get (16), the frequencies are multiplied by mij (mij 1), the possible values of (1 , m1 ). This yields the announced
formula.
In Section 3.1,
V1,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y111 , Y121 )} ,
V2,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y113 , Y121 )} ,
V3,s,c = Cov {Qs (Y111 , Y112 ); Qc (Y121 , Y131 )} .

Appendix C. Large sample distribution of c under the assumption c = 0

We want to find an expression for 22 , the limiting variance of I 1/2 2 under H0c : c = 0. From (19), the numerator of
c is
MSC MSS (C )

m
I

n(Yi.. Y... )2

I
n

(Yij. Yi.. )2

i=1

I (n 1)
m

i =1

I
n

(Yij. Y... )2
i

n

I

I (n 1) i=1

n1

n (Yi.. Y... )2
2

(Yij. Y... )(Yik. Y... )

(Yij. Y... )2

j ,k

m (Yij. Y... )(Yik. Y... )

I

1
j=k

2m (Yij. F )(Yik. F )
I

1
j <k

I i =1

n1

I i =1

n1

where F is the marginal mean of F . Under H0c , the random variable Yij. F s are independent with variance {1 + (m
1)1 }2 F2 /m. Thus MSC MSS (C ) is approximately equal to the mean of I identically distributed random variables with mean
0 and variance 2n {1 + (m 1)1 }2 F4 /(n 1). The denominator of c , MSC + (n 1)MSS (C ) + n(m 1)MSE converges in

probability to nmF2 . According to the Central Limit Theorem and Slutskys theorem, the limiting distribution of

is N 0, 2 {1 + (m 1)1 }2 /{n(n 1)m2 } as stated.

I ( c c )

H. Romdhani et al. / Journal of Multivariate Analysis 128 (2014) 210225

225

References
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]

K.J. Bandeen-Roche, K.Y. Liang, Modelling failure-time associations in data with multiple levels of clustering, Biometrika 83 (1996) 2939.
C. Genest, J. Neslehov, N. Ben Ghorbal, Estimators based on Kendalls tau in multivariate copula model, Aust. N. Z. J. Stat. 53 (2) (2011) 157177.
C. Genest, F. Verret, Locally most powerful rank tests of independence for copula models, J. Nonparametr. Stat. 17 (2005) 521535.
H. Goldstein, Multilevel Statistical Models, Wiley Series in Probability and Statistics, 2010.
H. Hult, F. Lindskog, Multivariate extremes, aggregation and dependence in elliptical distributions, Adv. Appl. Probab. 34 (2002) 587608.
H. Joe, Multivariate concordance, J. Multivariate Anal. 35 (1990) 1230.
H. Joe, Multivariate Models and Dependence Concepts, Chapman & Hall, London, 1997.
J.F. Mai, M. Scherer, H-extendible copulas, J. Multivariate Anal. 110 (2012) 151160.
A.J. McNeil, Sampling nested Archimedean copulas, J. Statist. Comput. Simul. 78 (2008) 567581.
A.J. McNeil, J. Neslehov, Multivariate Archimedean copulas, d-monotone functions and l1-norm symmetric distributions, Ann. Stat. 37 (2009)
30593097.
G.A. Milliken, D.E. Johnson, Analysis of Messy Data, in: Designed Experiments, vol. 1, 2009.
H. Romdhani, L. Lakhal-Chaieb, L.-P. Rivest, An exchangeable Kendalls tau for clustered data, Canad. J. Statist. (2014) Under revision.
C.E. Srndal, B. Swensson, J. Wretman, Model Assisted Survey Sampling, Springer Verlag, New York, 1992.
A.W. Van Der Vaart, Asymptotic Statistics, in: Cambridge Series in Statistical and Probabilistic Mathematics, 1998.