 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Published by Blackwell Publishing Ltd, 9600 Garsington

Road, Oxford OX4 2DQ, UK and 350 Main Street, Malden, MA 02148, USA Vol 33: 121–138, 2006

Residuals in the Extended Growth Curve

Model
JEMILA SEID HAMID and DIETRICH VON ROSEN
Department of Biometry and Engineering, Swedish University of Agricultural Sciences

ABSTRACT. The Extended Growth Curve model is considered. It turns out that the estimated
mean of the model is the projection of the observations on the space generated by the design
matrices which turns out to be the sum of two tensor product spaces. The orthogonal complement
of this space is decomposed into four orthogonal spaces and residuals are defined by projecting the
observation matrix on the resulting components. The residuals are interpreted and some remarks
are given as to why we should not use ordinary residuals, what kind of information our residuals
give and how this information might be used to validate model assumptions and detect outliers and
influential observations. It is shown that the residuals are symmetrically distributed around zero and
are uncorrelated with each other. The covariance between the residuals and the estimated model as
well as the dispersion matrices for the residuals are also given.

Key words: decomposition of linear spaces, Extended Growth Curve model, Growth Curve
model, residuals, tensor product of linear spaces

1. Introduction
The study of residuals is one of the most important steps in any model fitting procedure as
residuals represent what is left unexplained after subtracting the estimated model from the
observations. In this paper, residuals in general multivariate linear models will be considered.
In particular, we are going to consider the Extended Growth Curve model, i.e. the multi-


variate linear normal model (MLNM)( m i = 1 Ai Bi Ci ) (cf. definition 2), when m
= 2 which
among others arises when we have a Growth Curve model, i.e. an MLNM(ABC) (cf. definition
1) with certain linear restrictions on the parameters. Many diagnostic tools, both graphical
and formal tests, for validating regression or any univariate linear model are based on resi-
duals. For instance, the normal probability plot and the Shapiro–Wilk test for investigating
normality, the Durbin–Watson test for serial correlation and White’s test for heteroscedas-
ticity (see Sen & Srivastava, 1990). Residuals in multivariate linear models of type multivariate
analysis of variance (manova), have also been studied to some extent. Both in the univariate
and manova models residuals are defined by the difference between the observations and the
estimated mean. However, in the Growth Curve model the mean follows a bilinear structure
and therefore the residuals are also more structured. Indeed, in this model there exist three
types of residuals (see von Rosen, 1995).
The Extended Growth Curve model has many applications and may arise in many differ-
ent situations. One of its principal applications under the condition of equally observed ‘time
points’, is in the analysis of growth curves which is applied extensively in biostatistics, medi-
cal research and epidemiology. Verbyla & Venables (1988) considered models of which the
Extended Growth Curve model is a special case. They gave several examples to illustrate how
the model may arise. They also gave some indications of the applications of the model.
Therefore, if using the Extended Growth Curve model in practice, there is a need to develop
some diagnostic tools for validating the model. To our knowledge there has not yet been any
studies regarding residuals in this model. We hope that this paper will lay a ground for further
122 J. Seid Hamid and D. von Rosen Scand J Statist 33

studies regarding diagnostic tools for validating the model. The residuals defined in this
paper could be used to check the between and within individual assumptions in the model, i.e.
the bilinear structure. The residuals may also give information about outliers (which vaguely
speaking are observations that seem to differ from the rest) as they tend to have unexpectedly
large residuals. For instance, plotting the residuals in a systematic way gives a better insight
for identifying outliers, especially if there are a few of them.
This paper will be focused upon the special case m = 2, in the Extended Growth Curve
model. This paper consists of seven sections. The section 2 gives necessary background infor-
mation. In section 3 we consider the estimated model, A1 B̂1 C1 + A2 B̂2 C2 , and study the lin-
ear space orthogonal to the space generated by the design matrices. We decompose this space
into four orthogonal spaces and define residuals by projecting the observation matrix on the
resulting spaces. In section 4, the interpretation of the residuals is studied via the spaces they
are defined on. In section 5, we study some properties of the residuals. We concentrate on
moment relations as the distributions of the residuals are difficult to obtain. It turns out that
the residuals are symmetrically distributed around zero and are uncorrelated with each other.
The covariance matrices between the residuals and the estimated model as well as the disper-
sion matrices for the residuals are also given in this section. In section 6, we give a numerical
example to illustrate the results of the preceding sections and finally some conclusions and
remarks are stated in section 7.

2. Background
The growth curve model was introduced by Potthoff & Roy (1964) and subsequently studied
among others by Rao (1965) and Khatri (1966). For a review of the model, see von Rosen
(1991). There is a book by Kshirsagar & Smith (1995) about the model. A related paper on
the discrimination between growth curves was presented by Burnaby (1966).

Definition 1
Let X : p × n and B : q × k be the observation matrix and parameter matrix, respectively, and let
A : p × q and C : k × n be the within and between individual design matrices, respectively. Suppose
that q ≤ p and
(C) + p ≤ n, where
(•) denotes the rank of a matrix. The Growth Curve model
is given by
X = ABC + E, (1)
where the columns of E are assumed to be independently p-variate normally distributed with
mean zero and an unknown positive definite dispersion matrix
.

We denote the above model by MLNM(ABC). It is important to note that if A = I the

MLNM(ABC) reduces to the MLNM(BC) which is the classical manova model. Note also
that C is a design matrix similar to the one in univariate linear models and classical multi-
variate linear models.
As mentioned earlier the MLNM(ABC) has many applications and it arises in many differ-
ent situations. Suppose, for example, we have repeated observations on a given experimental
unit, where all the observations are supposed to be measured at the same time points. If these
observations can be associated with some continuous variable, such as time and temperature,
then they may form a response curve. The following example shows how the MLNM(ABC)
may arise and illustrates the observation matrix X, the parameter matrix B and the two de-
sign matrices A and C in definition 1. For additional examples, see e.g. Srivastava & Carter
(1983) and Kshirsagar & Smith (1995).

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 123

Example 1. Dental measurements on 11 girls and 16 boys at four different ages (8, 10, 12
and 14) were taken (Potthoff & Roy, 1964). Each measurement is the distance, in milli-
metres, from the centre of pituitary to pteryo-maxillary fissure. A linear growth curve is to
be fitted to both the girls and the boys. Then we may use the Growth Curve model where
the observation, parameter and design matrices are given by

 
21 21 20.5 23.5 21.5 20 21.5 23 20 16.5
 24.5 26 21.5 23 20 25.5 24.5 22 24 23 
 
 27.5 
 23 21.5 17 22.5 23 22, 
 
 20 21.5 24 24.5 23 21 22.5 23 21 19 
 
 25 25 22.5 22.5 23.5 27.5 25.5 22 21.5 20.5 
 
 28 23 23.5 24.5 25.5 24.5 21.5, 
X4 × 27 = 
 21.5
,
 24 24.5 25 22.5 21 23 23.5 22 19 
 28 24.5 31 
 29 23 24 22.5 26.5 27 24.5 
 
 31 23.5 24 26 25.5 26 23.5, 
 
 23 25.5 26 26.5 23.5 22.5 25 24 21.5 19.5 
 
 28 31 26.5 27.5 26 27 28.5 26.5 25.5 26 
31.5 25 28 29.5 26 30 25
 
1 8
 
b01 b02 1 10  111 016
B= , A =
1
 and C2 × 27 = ,
b11 b12 12  011 116
1 14

where the 111 and 016 in the first row indicate that there are 11 1s and 16 0s. Note also how
we have presented X, i.e. we have used commas to separate the rows.
Khatri (1966) found the maximum likelihood estimator for the parameter matrix B which
is given by,

B̂ = (A S−1 A)−1 A S−1 XC (CC )−1 ,

−1
where S = X(I − C (CC ) C)X and it was assumed that A and C are of full rank. The max-
imum likelihood estimator under the general situation, i.e. without assuming full rank con-
ditions, is given by

B̂ = (A S−1 A)− A S−1 XC (CC )− + (A )o Z1 + A Z2 Co ,

where, Z1 and Z2 are arbitrary matrices, Co is a matrix of full rank spanning the orthogonal
complement of the linear space generated by the columns of C, and G− denotes an arbitrary
generalized inverse in the sense that GG− G = G. For different methods of maximizing the
likelihood function, see e.g. Srivastava & Khatri (1979) and von Rosen (1989). Due to the
arbitrariness of the vectors Z1 and Z2 it is evident that the maximum likelihood estimator is
not unique. However, it is important to note that

AB̂C = A(A S−1 A)− A S−1 XC (CC )− C

is always unique.
Now consider the classical multivariate linear model [MLNM(BC)] which is given by

X = BC + E,

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

124 J. Seid Hamid and D. von Rosen Scand J Statist 33

where X : p × n, B : p × k and C : k × n are the observation, parameter and design matrices re-
spectively. In the univariate and MLNM(BC) cases residuals are obtained by projecting X on
the space orthogonal to C(C ), the column space of C , which is the space generated by the
design matrix, i.e. X(I − C (CC )− C). However, in the MLNM(ABC) the space has a bilinear
structure which generates a tensor product C(C ) ⊗ C S (A), where the S in C S indicates that
the inner product is defined with respect to S−1 , i.e. < x, y > = x S−1 y. If there is no subscript
as in C(C ), the standard inner product is assumed. Therefore, in MLNM(ABC) residuals are
obtained by projecting the observation matrix on the orthogonal complement to C(C ) ⊗
CS (A). von Rosen (1995) decomposed the space into three orthogonal spaces and defined
residuals by projecting X on the resulting spaces.
The Extended Growth Curve model which is given in definition 2 below was introduced by
von Rosen (1989). A canonical form of the model was considered by Banken (1984). A
special case of the model was considered by Srivastava & Khatri (1979). The Extended
Growth Curve is a special case of a more general model which is known in the econometrics
literatures as a multivariate seemingly unrelated regression model. The model without the
nested subspace condition was also considered by Verbyla & Venables (1988) under a different
name; sum of profiles model.

Definition 2
Let X : p × n, Ai : p × qi , Bi : qi × ki and Ci : ki × n. Suppose that qi ≤ p,
(C1 ) + p ≤ n and C(Cm ) ⊆
C(Cm−1 ) ⊆ · · · ⊆C(C1 ), where
(•) and C(•) represent the rank and column space of a matrix
respectively. Then the Extended Growth Curve model is given by,

m
X= Ai Bi Ci + E,
i =1

and denoted by MLNM( m i =1 Ai Bi Ci ), where the columns of E are assumed to be indepen-
dently distributed as a p-variate normal distribution with mean zero and an unknown positive
definite dispersion matrix
.

Verbyla & Venables (1988) in their paper gave an algorithm for obtaining maximum like-
lihood estimators. They also gave examples to illustrate how the model may arise and gave
some remarks as to the applications of the model. von Rosen (1989) gave the explicit form
of the maximum likelihood estimators of the model parameters under the nested subspace
condition between the design matrices.
In this paper, we are going to consider the model in definition 2 when m = 2, i.e. the
MLNM(A1 B1 C1 + A2 B2 C2 ). The following example gives an illustration of the matrices
involved in the model.

Example 2. Consider the data in example 1. Suppose that a linear and a quadratic growth
are to be fitted for the girls and boys respectively. Suppose, also that the growth curve for the
boys has a linear component, then the observation matrix will be the same as the matrix X
given in example 1, C1 = C, A1 = A, B1 = B, where C, A, B are the matrices given in example
1, C2 is a row vector with 27 elements where the first 11 elements are zero and the rest equal
1 and A2 = (64 100 144 196).
Now we give an important theorem which will be used in the sequel. The general case of
the theorem was established and proved by von Rosen (1989).

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 125

Theorem 1
Suppose the conditions in definition 2 are satisfied and let B̂1 and B̂2 be the maximum likelihood
estimators of B1 and B2 respectively. Then, the estimated mean structure for the model given in
definition 2 is given by

A1 B̂1 C1 + A2 B̂2 C2 = (I − T1 )XC1 (C1 C1 )− C1 + (I − T2 )XC2 (C2 C2 )− C2 , (2)

where
T1 = I − A1 (A1 S−1 −  −1
1 A1 ) A1 S1 ,

T2 = I − T1 A2 (A2 T1 S−1 −   −1

2 T1 A2 ) A2 T1 S2 ,

S1 = X(I − C1 (C1 C1 )− C1 )X ,

S2 = S1 + T1 XC1 (C1 C1 )− C1 (I − C2 (C2 C2 )− C2 )C1 (C1 C1 )− C1 X  T1 .
Finally, we give the definition of the dispersion matrix and covariance between two
matrices. Let, A = (a1 , a2 , . . ., aq ) be a p × q matrix, where ai , i = 1, 2, . . ., q, is the ith column
vector. The vec-operator vec(•) is an operator from Rp×q to Rpq defined by

vecA = (a1 , a2 , . . ., aq ) . (3)

Let U and V be two matrices. The covariance matrix between them is defined by

cov(U, V) = E[vec U vec V] − E[vecU]E[vec V]

and the dispersion matrix,

D[U] = cov[U, U].

3. Residuals
In this section, we will define residuals in the Extended Growth Curve model. First, we would
like to give some notations which will be used throughout this paper.
PC1 = C1 (C1 C1 )− C1 ,
PC2 = C2 (C2 C2 )− C2 ,
PA1 = A1 (A1 S−1 −  −1
1 A1 ) A1 S1 ,

PA2 = T1 A2 (A2 T1 S−1 −   −1

2 T1 A2 ) A2 T1 S2 .

Now consider the estimated model given in theorem 1 and replace T1 and T2 by the cor-
responding expressions in the theorem. We get

A1 B̂1 C1 + A2 B̂2 C2 = PA1 XPC1 + PA2 XPC2 . (4)

Apply the vec-operator on both sides of (4). It gives

(C1 ⊗ A1 ) vecB̂1 + (C2 ⊗ A2 ) vecB̂2 ={PC1 ⊗ PA1 } vecX +{PC2 ⊗ PA2 } vecX
= P vecX,
where ⊗ denotes the Kronecker product and

P = (PC1 ) ⊗ (PA1 ) + (PC2 ) ⊗ (PA2 ).

P is an idempotent matrix and hence a projection matrix. Moreover (for details we refer to
Seid Hamid, 2001).

C(P) = C(C1 ) ⊗ CS1 (A1 ) + C(C2 ) ⊗ CS2 (T1 A2 ), (5)

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

126 J. Seid Hamid and D. von Rosen Scand J Statist 33

where now ⊗ on the right-hand side represents a tensor product of linear spaces. Conse-
quently, the estimated model is obtained by projecting X on (5) which is the space generated
by the design matrices. Observe that the subscripts S1 and S2 in C S1 (A1 ) and C S2 (T1 A2 ) indi-
cate that the inner products are defined with respect to the two positive definite matrices S1
and S2 . However, it is possible to show that A1 S−1  −1
1 = A1 S2 , see Seid Hamid (2001). There-
fore, it is straightforward to replace S1 in CS1 (A1 ) by S2 so that we have one inner product,
i.e. < x, y > = x S−1
2 y.
Consider now the ordinary residuals which are obtained by subtracting the estimated model
from the observations, i.e.

X − (A1 B̂1 C1 + A2 B̂2 C2 ).

Applying the vec-operator yields

vec{X − (A1 B̂1 C1 + A2 B̂2 C2 )} = (I − P) vecX.

Moreover, I − P is a projection matrix and C(I − P) = C S2 (P)⊥ , where ⊥ denotes the ortho-
gonal complement. Therefore, residuals may be defined by projecting X on the space ortho-
gonal to the space generated by the design matrices. However, residuals defined in that way
do not give any information about the bilinear structure in the mean and hence there is a
need to define other residuals. Therefore, we will exploit

(C(C1 ) ⊗ CS1 (A1 ) + C(C2 ) ⊗ CS2 (T1 A2 ))⊥ . (6)

Unfortunately, this space is difficult to interpret. However, it is much easier when the above
space is decomposed into different subspaces. In addition to the technical simplification, we
are going to show in the next section that the resulting spaces are interpretable.
Before going into the decomposition, let us observe the following two points that will
simplify the treatment. The first one is that C(C2 ) ⊆C(C1 ) by assumption and that

CS2 (T1 A2 ) ⊆ CS2 (A1 )⊥ . (7)

The second one is that

C(A1 )
C(T1 A2 ) = C(A1 ) + C(A2 ), (8)

where
denotes the orthogonal sum. See Seid Hamid (2001) for the proofs of (7) and (8).
Using the information we have in (7) and (8) about the spaces involved, (6) can be decom-
posed into the following four orthogonal spaces (see Seid Hamid, 2001),

(C(C1 ) ⊗ CS2 (A1 ) + C(C2 ) ⊗ CS2 (T1 A2 ))⊥ = I
II
III
IV, (9)

where

I = C(C1 )⊥ ⊗ CS2 (A1 ),

II = C(C1 )⊥ ⊗ CS2 (A1 )⊥ ,
III = (C(C1 ) ∩ C(C2 )⊥ ) ⊗ CS2 (A1 )⊥ ,
IV = C(C2 ) ⊗ (CS2 (A1 ) + CS2 (A2 ))⊥ .

Now we are ready to give the definition of the residuals for the Extended Growth Curve
model when m = 2. We define them through the above four spaces, i.e. R1 , R2 , R3 and R4 are
obtained by projection X on the spaces given by I, II, III and IV respectively.

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 127

Definition 3
Let T1 and T2 be as defined in theorem 1 and suppose that the conditions in definition 2 are
satisfied. Then residuals are defined by,
R1 = PA1 X(I − PC1 ) = (I − T1 )X(I − PC1 ),
R2 = (I − PA1 )X(I − PC1 ) = T1 X(I − PC1 ),
R3 = (I − PA1 )X(PC1 − PC2 ) = T1 X(PC1 − PC2 ),
R4 = (I − PA1 − PA2 )XPC2 = (T1 + T2 − I)XPC2 .

Observe that A1 S−1  −1

2 in the residuals can be replaced by A1 S1 . Therefore, we are going to
use this fact to replace S2 by S1 whenever necessary as S1 has a much simpler structure and
is easier to handle. Observe also that the above underlying decomposition is by no means
unique and hence residuals could be defined in many other ways. For instance, we may want
to combine the residuals in some way depending on what information we are looking for. We
could also decompose them further and get additional information by looking at the differ-
ent components. In section 4 we are going to decompose R3 into two components each of
which provides important information.
What we have performed so far can equivalently be formulated as follows,

Theorem 2
Suppose the conditions in definition 2 are satisfied. Let R1 , R2 , R3 and R4 be given in definition
4 and let B̂1 and B̂2 be the maximum likelihood estimators of B1 and B2 , respectively. Then

X − (A1 B̂1 C1 + A2 B̂2 C2 ) = R1 + R2 + R3 + R4 .

As the last part of this section we shall consider the space generated by the design matrices.
We can decompose it as follows,

C(C1 ) ⊗ CS2 (A1 ) + C(C2 ) ⊗ CS2 (T1 A2 ) = I
II
III,

where
I = (C(C1 ) ∩ C(C2 )⊥ ) ⊗ CS2 (A1 ),
II = C(C2 ) ⊗ CS2 (A1 ),
III = C(C2 ) ⊗ CS2 (T1 A1 ).
Now, if we project X on the above three spaces, respectively, we get the following:
M1 = PA1 X(PC1 − PC2 ), (10)
M2 = PA1 XPC2 , (11)
M3 = PA2 XPC2 . (12)
Note that the estimated mean equals M1 + M2 + M3 . We are going to show later that this
decomposition is important in calculating the covariance between the residuals and the esti-
mated model. Moreover, each Mi contains important information which we shall exploit in
section 4.

4. Interpretation of the residuals

If we are going to use the Extended Growth Curve model in practice one needs to develop
diagnostic tools for validating the model. As in most statistical models one may think of
what kind of information is possible to obtain from the residuals. We believe that each of the
residuals defined in the previous section bears important information, which may be used to

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

128 J. Seid Hamid and D. von Rosen Scand J Statist 33

validate the model assumptions in some way. In this section, we try to interpret the residuals
given in definition 4 via the spaces on which they are defined. We also give some indication
as to how these residuals might be used to assess the model assumptions.
Suppose without loss of generality that we have two groups of experimental units. Repeated
measurements are taken on each experimental unit. Suppose, also that the observations in
group 1 follow a linear mean structure whereas the observations in group 2 follow a quadratic
mean structure. Assume also that all individuals are measured at the same time points. Note
that example 2 illustrates the presented situation.
Let us now start by looking at R1 + R2 , which equals

R1 + R2 = X(I − PC1 ). (13)

The expression in (13) represents the observations minus the mean value for each group. Thus
R1 + R2 tells us how far the observations are from their group mean. In other words, it tells
us about the between individual assumption in a given group.
If we consider R3 + R4 , it follows that

R3 + R4 = XPC1 − (A1 B̂1 C1 + A2 B̂2 C2 ),

which is the observed mean minus the estimated mean structure. Thus, R3 + R4 tells us how
well the estimated mean fits the observed mean. It tells us about the within individual assump-
tions. However, if we look at them separately,

R3 = X(PC1 − PC2 ) − M1 ,

where M1 is given in (10). Both terms on the right-hand side are zero for the observations
in group 2 and hence R3 is zero for those observations. For the observations in group 1, the
two terms are the observed and the estimated mean structures respectively. This means that
R3 is obtained by subtracting the estimated mean from the observed mean. Hence, R3 tells
us how well the assumed linear growth curve fits the observed mean for the observations in
the first group. It tells us about the within individual assumptions for the group.
Let us go one step further and write R3 as the sum of two expressions, denoted by R13
and R23 :

R3 = PA2 X(PC1 − PC2 ) + (I − PA1 − PA2 )X(PC1 − PC2 ). (14)

The first term, R13 , is obtained by projecting the observed mean for the second group
X(PC1 − PC2 ) on C(T1 A2 ). This means R13 is the projection of the observed mean on C(A1 ) +
C(A1 ) after the contribution because of the linear term has been subtracted. This indi-
cates that R13 is the reduction in R3 if we add a quadratic term in the growth curves of the
observations in group one (those with a linear mean structure). Equivalently, we may use R3
and R23 to assess whether there is a need to add a quadratic term in the growth curves.
Now, let us consider R4 , which could be rewritten as,

R4 = XPC2 − (M2 + M3 ), (15)

where M2 and M3 are defined in (11) and (12) respectively. For the observations in group 1,
the three terms on the right-hand side of equation (15) are zero and therefore R4 for obser-
vations in group 1. However, for the observations in group 2 XPC1 and M2 + M3 are the
observed and estimated mean respectively. Hence, R4 is the difference between the observed
and estimated mean structure. Therefore, we may use R4 to assess how well the assumed qua-
dratic mean structure fits the observed mean for the experimental units in the second group.

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 129

Here, it is worth mentioning that when considering group one it is enough to look at the
residuals R1 , R2 and R3 and the estimated mean M1 , whereas when considering the second
group it is enough to look at the residuals R1 , R2 and R4 and the estimated mean M2 + M3 .
Finally, consider M3 . As mentioned earlier M3 is zero for observations in group 1. For the
observations in the second group, it is the projection of the observed mean on C(T1 A2 ). This
indicates that M3 is the projection of the observed mean on a space where the contribution
of the linear component has been subtracted. It is the reduction we have got because of the
quadratic term in the assumed growth curve. Therefore, one may use R4 and M34 , where

M34 = M3 + R4 (16)

to check how significant the contribution of the quadratic term is.

The observations in the groups mentioned in the beginning of this section do not have to
follow a linear and/or a quadratic mean structure. The interpretation of the residuals given
above is applicable, with few modifications, in much more general situations.

5. Properties of the residuals

Residuals in the univariate linear models have been studied extensively producing a huge
amount of the literature. Moreover, different types of residuals have been introduced such
as ordinary residuals, studentized residuals and others. There have also been studies regard-
ing the distributions and properties of the residuals. For instance, it is a well-known fact that
ordinary residuals in the univariate linear model are symmetrically distributed around zero
and are uncorrelated with the estimated model. This is also true for the classical multivariate
linear model, i.e. the MLNM(BC).
It was shown in von Rosen (1995) that residuals in the Growth Curve model are sym-
metrically distributed around zero and obtained a couple of moment relations for three
types of residuals. These results will now be extended to the MLNM(A1 B1 C1 + A2 B2 C2 ).

Theorem 3
Let R1 , R2 , R3 and R4 be the residuals given in definition 4. Then,

E[Ri ] = 0, i = 1, 2, 3, 4.

Proof. Let us consider R1 + R2 first;

E[R1 + R2 ] = E[X(I − PC1 )
= (A1 B1 C1 + A2 B2 C2 )(I − PC1 ) = 0.

Here, we have used the fact that (I − PC1 ) is the projection matrix on C(C1 )⊥ and C(C2 ) ⊆
C(C1 ). Now, observe that as R1 is an odd function of X(I − PC1 ) which is distributed as a
p-variate normal distribution with mean zero, we have E[R1 ] = 0. Thus E[R2 ] = 0 follows from
the first part of the proof.
We can rewrite R3 as

R3 = PA1 X(PC1 − PC2 ).

This follows from the fact that

 
I − G(G S−1 G)− G S−1 = SGo (Go SG0 )− Go (17)

for any matrix G and S is supposed to be a positive definite matrix. Now,

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

130 J. Seid Hamid and D. von Rosen Scand J Statist 33

 
E[R3 ] = E[S1 Ao1 (Ao1 S1 Ao1 )− Ao1 X(PC1 − PC2 )]
 
= E[S1 Ao1 (Ao1 S1 Ao1 )− Ao1 E[X(PC1 − PC2 )]] = 0,
where we have used C2 (PC1 − PC2 ) = 0, the independence between S1 and X(PC1 − PC2 ), and

the fact that Ao1 A1 = 0. Next, let A be the partitioned matrix (A1 : A2 ). Using (8) and (17) as
well as the fact that projection operators are unique we can rewrite R4 as
 
R4 = S2 Ao (Ao S2 Ao )− Ao XPC2 . (18)
Therefore,

 
E[R4 ] = E[S2 Ao (Ao S2 Ao )− Ao XPC2 ] = 0,
which follows due to the independence between S2 and XPC2 , and the fact that Ao is ortho-
gonal to both Ao1 and Ao2 .
The above theorem together with the assumption of normality tells us that the residuals
are symmetrically distributed around zero. In the next theorem we are going to see that they
are uncorrelated with each other. Let us first give a lemma which will be used in the proof
(for the proof of the lemma we refer to Seid Hamid 2001).

Lemma 1
Let R1 and R2 be the residuals given in definition 4 and let

Ao1 = T(I : 0)
−1/2 and Y = 
−1/2 X(I − C1 (C1 C1 )− C1 ),
where T and  are non-singular and orthogonal matrices of proper sizes, respectively, and
1/2
is a symmetric square root. Then,
R1 =
1/2 2 Y2 (I − Y1 (Y1 Y1 )−1 Y1 )
and
R2 =
1/2 1 Y1 +
1/2 2 Y2 Y1 (Y1 Y1 )−1 Y1 ),
where Y  = (Y1 : Y2 ) and  = (1 : 2 ).

Theorem 4
Let R1 , R2 , R3 and R4 be the residuals given in definition 4. Then,
cov[Ri , Rj ] = 0 i, j = 1, 2, 3, 4, i  = j.

Proof. Let us divide the proof into four parts.

Part 1: here, it is convenient to use the canonical forms of R1 and R2 which is given in lemma
1. If using this representations we get
cov[R1 , R2 ] = cov[R1 ,
1/2 1 Y1 ] + cov[R1 ,
1/2 2 Y2 Y1 (Y1 Y1 )−1 Y1 ]. (19)
Using the fact that Y1 and Y2 are independent and have mean zero, the first term on the
right-hand side of (19) equals zero. Now use the definition of the covariance between two
matrices and the fact that (I − PC1)Y1 = Y1 and E[vecY1 vec Y1 ] = (I − PC1). As a result the
second term also equals to zero.
Part 2: consider R1 and R3 and replace them by their corresponding expressions and use the
fact that X(I − PC1) and X(PC1 − PC2) are independent. Then,

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 131

cov[R1 , R3 ] = E[vecR1 vec R3 ]

= E[vec{(PA1 X(I − PC1 ))}E[vec X(PC1 − PC2 )]{I ⊗ PAo }],
1


which equals zero as A1 Ao1 = 0
and C2 (PC1 − PC2 ) = 0. Note that PA2o represents Ao1 (Ao1 S1 Ao1 )−
o
A1 S1 . The proof of cov[R1 , R3 ] = 0 is analogous.
Part 3: let, A be as in (18) and use expression (18) for R4 . Note that XPC2 is independent
of X(I − PC1 ), S1 and S2 . Furthermore, Ao is orthogonal to both A1 and A2 . As a result
cov[R1 , R4 ] = 0. Analogously, one can show that cov[R2 , R4 ] = 0.
Part 4:

cov[R3 , R4 ] = E[vec{PA1o X(PC1 − PC2 )} vec{PA1o XPC2 }],

which can be shown to be zero after using the fact that X(PC1 − PC2 ) is independent of S1 ,
S2 and XPC2 .
The next theorem gives the covariance between the residuals and the estimated model. Be-
fore stating the theorem let us give the following preparatory lemma (see Seid Hamid, 2001
for the proof).

Lemma 2
Let M1 , M2 and M3 be as defined in (10), (11) and (12), respectively, and let

L1 = A1 (A1
−1 A1 )− A1

     
L2 =
Ao1 (Ao1
Ao1 )−1 Ao1 A2 (A2 Ao1 (Ao1
Ao1 )−1 Ao1 A2 )− A2 Ao1 (Ao1
Ao1 )−1 Ao1
.

Then,
(i) D[M1 ] = 1 (PC1 − PC2 ) ⊗ L1 ,
(ii) D[M2 ] = 2 PC2 ⊗ L1 ,
(iii) D[M3 ] = PC2 ⊗ (3 L1 + 2 L2 ),
(iv) D[M2 + M3 ] = PC2 ⊗ {(1 + 33 )L1 + (2 + 2)L2 },

where,
n −
(C1 ) − 1 n −
(C2 ) − 1
1 = , 2 =
n −
(C1 ) − p +
(A1 ) − 1 n −
(C2 ) − p +
(A1 : A2 ) −
(A1 ) − 1
(n −
(C1 ) − 1)(
(A1 : A2 ) −
(A1 ))
3 = .
(n −
(C1 ) − p +
(A1 ) − 1)(n −
(C2 ) − p +
(A1 : A2 ) − 1)

Theorem 5
Let R1 , R2 , R3 and R4 be the residuals given in definition 4 and L1 , L2 , 1 , 2 and 3 be as
defined in lemma 2. Suppose that the assumptions in definition 2 are satisfied. Let  = 1 − 1 ,
 =  − 33 and  = 1 + 2 , then

(i) cov[Ri , A1 B̂1 C1 + A2 B̂2 C2 ] = 0, i = 1, 2,

(ii) cov[R3 , A1 B̂1 C1 + A2 B̂2 C2 ] = (PC2 − PC2 ) ⊗ L1 ,
(iii) cov[R4 , A1 B̂1 C1 + A2 B̂2 C2 ] = PC2 ⊗ (L1 − L2 ).

Proof. (i) First note that,

XPC1 = A1 B̂1 C1 + A2 B̂2 C2 + R34 ,

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

132 J. Seid Hamid and D. von Rosen Scand J Statist 33

where R34 = R3 + R4 . On the contrary, because of the fact that XPC1 is independent of
X(I − PC1 ) and S1 , we have
cov[R1 , XPC1 ] = 0,
which together with the fact that the residuals are uncorrelated implies
cov[R1 , A1 B̂1 C1 + A2 B̂2 C2 ] = 0.
Using similar arguments it is possible to show that
cov[R2 , A1 B̂1 C1 + A2 B̂2 C2 ] = 0.
(ii) Here, observe that we can rewrite the estimated model as a sum of M1 , M2 and M3 .
Consequently,
cov[R3 , A1 B̂1 C1 + A2 B̂2 C2 ] = cov[R3 , M1 ] + cov[R3 , M2 ] + cov[R3 , M3 ]. (20)
Now consider the last two expressions on the left hand side of (20) separately,
cov[R3 , M2 ] = E[{I ⊗ (PA1o )}E[vecX(PC1 − PC2 )]E[vec {XPC2 }]{I ⊗ (PA1)}] = 0,
and equality holds as X(PC1 − PC2 ) is independent of S1 , S2 and XPC2 . Similarly, it can be
shown that cov[R3 , M3 ] = 0. As a result,
cov[R3 , A1 B̂1 C1 + A2 B̂2 C2 ] = cov[X(PC1 − PC2 ), M1 ] − D[M1 ]. (21)

However, it is possible to show that

E[I − T1 ] =
−1 A1 (A1
−1 A1 )− A1 . (22)
As a result the first term on the right-hand side of (21) equals
(PC1 − PC2 ) ⊗ (A1 (A1
−1 A1 )− A1 ).
The desired result is established using lemma 2 (i).
(iii) Now, consider R4 and we have the following
cov[R4 , A1 B̂1 C1 + A2 B̂2 C2 ] = cov[R4 , M1 ] + cov[R4 , M2 + M3 ]. (23)
Recall that XPC2 is independent of X(PC1 − PC2 ), S1 and S2 . Consequently, the first term on
the right-hand side of (23) equals to zero. The second term equals

cov[XPC2 , M2 + M3 ] − D[M2 + M3 ],
however,

cov[XPC2 , M2 + M3 ] = (PC2 ) ⊗ (L1 + L2 ).

Here, we have used (22) and the fact that (see Seid Hamid, 2001 for the proof)
E[S−1   −1 −   −1
2 T1 A2 (A2 T1 S2 T1 A2 ) A2 T1 ] =
L2 .

The desired result follows after using lemma 2 (iv).

We have established that the residuals are symmetric around zero (theorem 3) and uncor-
related (theorem 4). In theorem 5 we have seen how they are related to the estimated model.
It is of interest to study the variation in the residuals. Moreover, the mean and variance are
useful when approximating the distribution of the residuals. In the next theorem we shall give
the dispersion matrices.

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 133

Theorem 6
Let R1 , R2 , R3 and R4 be the residuals given in definition 4 and let the matrices L1 and L2 and
the constants 1 , 2 and 3 be as defined in lemma 2. Suppose that the assumptions in definition
2 are satisfied. Then,

(i) D[R1 ] = 1 (I − PC1 ) ⊗ (A1 (A1
−1 A1 )− A1 ),

(ii) D[R2 ] = (I − PC1 ) ⊗ (
− 1 A1 (A1
−1 A1 )− A1 ),
(iii) D[R3 ] = (PC1 − PC2 ) ⊗ (
− 2 A1 (A1
−1 A1 )− A1 ),
(iv) D[R3 ] = PC2 ⊗ (
+ 3 L1 + 4 L2 ),

where
n −
(C1 ) − p +
(A1 ) n −
(C1 ) − 2(p −
(A1 )) − 1
1 = , 2 = ,
n −
(C1 ) n −
(C1 ) − p +
(A1 ) − 1
3 = 1 + 33 − 2 and 4 = 2 .

Proof. Using the canonical representation of R2 which is given in lemma 1, we can rewrite
D[R2 ] as follows

D[R2 ] = D[
1/2 1 Y1 ] + D[
1/2 2 Y2 PY1 ] + cov[
1/2 1 Y1 ,
1/2 2 Y2 PY1 ], (24)

where PY1 = Y1 (Y1 Y )− Y1 . Using the fact that Y1 and Y2 are independent and that E[Y2 ] = 0,
the third term on the right-hand side of (23) equals zero. Furthermore, it is possible to show
that
 

1/2 1 1
1/2 =
− A1 (A1
−1 A1 )− A1 .

Consequently, we have

D[
1/2 1 Y1 ] = (I − PC1) ⊗ (
− A1 (A1
−1 A1 )− A1 ).

For the second term, we condition with respect to Y1 and use the fact that Y1 (I − PC1) = Y1 ,
which leads to
p −
(A1 )
D[
1/2 2 Y2 PY1 ] = (I − PC1) ⊗ (A1 (A1
−1 A1 )− A1 ).
n −
(C1 )
Therefore,

D[R2 ] = (I − PC1) ⊗ (
− 1 A1 (A1
−1 A1 )− A1 ).

Now, note that R1 and R2 are uncorrelated. As a result

D[R1 ] = D[X(I − PC1)] − D[R2 ] = 1 (I − PC1) ⊗ (A1 (A1
−1 A1 )− A1 ).

For R3 , it follows by utilizing theorem 5 (ii) and lemma 2 (i) that

D[R3 ] = D[X(PC1 − PC2)] − D[M1 ] − 2cov[R3 , M1 ]

= (PC1 − PC2) ⊗ (
− 2 A1 (A1
−1 A1 )− A1 ).

Finally, we have

D[R4 ] = D[XPC2 ] − D[M2 + M3 ] − 2 cov[R4 , M2 + M3 ],

and the desired result follows from lemma 2 (iv) and theorem 5 (iii).

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

134 J. Seid Hamid and D. von Rosen Scand J Statist 33

Corollary 1
Suppose the assumptions in definition 2 are satisfied with n1 and n2 as the number of indivi-
duals in group I and group II, respectively. The dispersion matrices given in theorem 6 can
be simplified as,
(i) D[R1 ] = 1 1 A1 (A1
−1 A1 )− A1 ,
(ii) D[R2 ] = 1 (
− 1 A1 (A1
−1 A1 )− A1 ),
(iii) D[R3 ] = 1 2 (
− 2 A1 (A1
−1 A1 )− A1 ),
(iv) D[R4 ] = (1 − 1 )3 (
+ 3 L1 + 4 L2 ),
where, 1 = (1 − 1/n1) if an individual belongs to group I and 1 = (1 − 1/n2) if it belongs to group
II, and 2 is an indicator which equals to one if an individual is in group I, and 3 is an indicator
which equals to one if an individual is in group II.

In practice, it may be necessary to estimate the dispersion matrices D[R1 ], D[R2 ] D[R3 ]
and D[R4 ] as well as the covariance matrices cov[R3 , A1 B̂1 C1 + A2 B̂2 C2 ] and cov[R4 , A1 B̂1 C1 +
A2 B̂2 C2 ]. One possible estimator is based on the maximum likelihood estimators which is
obtained by replacing
in the corresponding expressions by its maximum likelihood estima-
tor which is given by
1
ˆ=

(S2 + R4 R4 ).
n
However, the maximum likelihood estimators are not unbiased. If one is interested in un-
biased estimation one may then modify the estimators to get unbiased estimators (see von
Rosen, 1995).

6. Numerical illustration
In this section, we give a numerical example to illustrate what we have discussed in the pre-
vious sections. We are going to consider the Potthoff & Roy (1964) data. This data have also
been considered by von Rosen (1995) where the girls and the boys were assumed to follow a
linear mean growth. Here, we are going to fit the MLNM(A1 B1 C1 + A2 B2 C2 ) where the girls
and the boys are assumed to follow a linear and quadratic mean growth respectively. We are
not going to give any formal diagnostic tools for validating the model. Rather, our aim is to
give a numerical example to illustrate what we have discussed in the preceding sections and
try to see if the residuals are close enough to zero by looking at the standard errors for the
residuals. The values of the residuals together with their standard errors are presented in the
Tables 1–4. Values that are above two standard errors are marked with ∗.
Consider the Potthof & Roy (1964) data. The MLNM(A1 B1 C1 + A2 B2 C2 ) is fitted to the
data where the matrices X, A1 , B1 , C1 , A2 , B2 and C2 are as in example 2 and the residuals
given in definition 4 together with their standard errors obtained from the dispersion matrices
given in corollary 1 are calculated.
We will examine R1 and R2 given in Table 1. As we can see, the values of R1 at all ages
fall within three standard errors from the mean although individuals 10 and 21 seem to have
values larger than two standard errors. Individual 11 seems to have large R1 values; however,
the values are within two standard errors except at ages 12 and 14. Individual 15 has an R1
value larger than two standard errors at age 8. On the contrary, the values of R2 are within
two standard errors for all the observations except for individual 20 at age 12, where the value
is much larger than three standard errors. By looking at data for this individual it actually
seems that the value of the dental measurement at age 12 does not fit the data at all. Next

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 135

Table 1. The values of R1 and R2 and the corresponding standard errors for the girls (seg ) and the boys
(seb ), where individuals 1–11 are girls and the rest are boys
R1 R2
Observation 8 10 12 14 8 10 12 14
1 −1.41 −1.35 −1.30 −1.25 1.22 −0.87 −0.29 0.16
2 −0.95 −0.19 0.58 1.34 0.77 −0.54 0.33 0.07
2 −0.95 −0.19 0.58 1.34 0.77 −0.54 0.33 0.07
3 0.44 0.97 1.50 2.03 −1.12 0.80 −0.10 −0.13
4 2.27 2.31 2.34 2.38 0.05 −0.04 −0.44 0.03
5 0.86 0.39 −0.08 −0.56 −0.54 0.38 −0.51 −0.04
6 −1.11 −1.28 −1.45 −1.62 −0.08 0.05 −0.64 0.03
7 0.16 0.39 −0.62 0.85 0.16 −0.12 −0.71 0.06
8 1.53 0.98 0.43 −0.13 0.29 −0.21 −0.02 0.03
9 −0.90 −1.44 −1.97 −2.51 −0.28 0.21 0.88 −0.08
10 −3.68∗ −3.94∗ −4.20∗ −4.47∗ −1.00 0.71 0.11 −0.12
11 2.77 3.15 3.54∗ 3.92∗ 0.55 −0.38 1.37 −0.01
12 1.65 2.23 2.81 3.39 1.47 −1.04 0.47 0.15
13 −1.40 −1.28 −1.67 −1.06 0.02 −0.03 −1.55 0.09
14 −0.84 −0.61 −0.38 −0.15 0.97 −0.70 −1.34 0.19
15 4.11∗ 2.63 1.16 −0.32 −1.49 1.05 −0.38 −0.15
16 −1.36 −1.37 −1.38 −1.39 −1.51 1.06 −1.84 −0.08
17 1.81 1.56 1.30 1.05 −0.18 0.13 −0.02 −0.02
18 −1.51 −1.36 −1.20 −1.04 0.64 −0.45 −0.02 0.08
19 −0.58 −1.10 −1.62 −2.15 1.71 −1.21 0.41 0.18
20 −2.06 −1.82 −1.59 −1.35 2.18 −1.49 6.87∗ −0.12
21 4.43∗ 4.31∗ 4.19∗ 4.07∗ 0.19 −0.13 1.09 −0.04
22 0.08 −0.78 −1.64 −2.51 0.04 −0.03 −0.58 0.04
23 −1.04 −0.53 −0.03 0.48 −0.33 0.22 −1.69 0.05
24 −3.10 −1.28 0.53 2.34 −2.78 1.97 −0.25 −0.31
25 1.33 0.47 −0.39 −1.26 −1.71 1.22 0.18 −0.21
26 −0.18 0.80 1.62 2.44 0.14 −0.11 −1.34 0.09
27 −1.51 −1.86 −2.20 −2.54 0.64 −0.45 −0.02 0.08
seg 1.75 1.62 1.69 1.92 1.22 0.94 1.61 0.71
seb 1.78 1.65 1.71 1.95 1.22 0.95 1.63 0.71
Values greater than two standard errors are marked with ∗.

Table 2. The calculated of R3 and R4 together with their standard errors, where seg
and seb represent the standard errors for the girls and boys respectively
R3 R4
8 10 12 14 8 10 12 14
Girls −0.054 0.038 −0.051 −0.004 0.000 0.000 0.000 0.000
Boys 0.000 0.000 0.000 0.000 0.141 −0.098 0.232 0.004
seg 1.2 0.91 1.61 0.63
seb 0.4 0.32 0.49 0.26

consider R12 = R1 + R2 which is given in Table 3. This residual is normally distributed with
mean zero, and it actually tells us about the between individual variations. We can therefore
use these residuals to see if there are individuals who seem to deviate from the rest. In addi-
tion to the individuals mentioned above, individual 24 seems to have a large value at age 8
although it is still within three standard errors from the mean. It is also important to note
that some of the individuals with large R1 values have smaller R12 values, see e.g. individual
21 at time 14. This shows why it is good to examine the residuals R1 and R2 separately.
Let us now consider R3 given in Table 2. Note that R3 is zero for the boys. For the girls it
is the difference between the observed and estimated mean M1 . Recall that this residual tells
us how well the estimated linear growth curve fits the data for the girls. It seems that R3 for
the girls is very close to zero at all ages, the values fall within one standard deviation from

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

136 J. Seid Hamid and D. von Rosen Scand J Statist 33

Table 3. The values of R12 , where R12 = R1 + R2 and the correspond-

ing standard errors for girls (observations 1–11) and boys (observations
12–27), values greater than two standard errors are marked with ∗
Observation 8 10 12 14
1 −0.18 −2.23 −1.59 −1.09
2 −0.18 −0.73 0.91 1.41
3 −0.68 1.77 1.41 1.91
4 2.32 2.27 1.91 2.41
5 0.32 0.77 −0.59 −0.59
6 −1.18 −1.23 −2.09 −1.59
7 0.32 0.27 −0.09 0.91
8 1.82 0.77 0.41 −0.09
9 −1.18 −1.23 −1.09 −2.59
10 −4.68∗ −3.23 −4.09 −4.59∗
11 3.32 2.77 4.91∗ 3.91
12 3.13 1.19 3.28 3.53
13 −1.38 −1.31 −2.72 −0.97
14 0.13 −1.31 −1.72 0.03
15 2.63 3.69 0.78 −0.47
16 −2.88 −0.31 −3.22 −1.47
17 1.63 1.69 1.28 1.03
18 −0.88 −1.81 −1.22 −0.97
19 1.13 −2.31 −1.22 −1.97
20 0.13 −3.31 5.28∗ −1.47
21 4.63∗ 4.18∗ 5.28∗ 4.03
22 0.13 −0.81 −2.22 −2.47
23 −1.38 −0.31 −1.72 0.53
24 −5.88∗ 0.69 0.28 2.03
25 −0.38 1.69 −0.22 −1.47
26 0.13 0.60 0.28 2.53
27 −0.88 −2.31 −2.22 −2.47
seg 2.14 1.88 2.33 2.05
seb 2.17 1.91 2.37 2.08

Table 4. The values of M1 and M34 = M3 + R4 , where M1 and M3 are as given in

(10) and (12) respectively
M1 M34
8 10 12 14 8 10 12 14
Girls 21.236 22.189 23.142 24.094 0.000 0.000 0.000 0.000
Boys 0.000 0.000 0.000 0.000 0.418 −0.298 −0.045 0.051

the mean. This indicates that the observed mean and the estimated mean are close to each
other. In other words, it indicates that the data fits the estimated linear curve very well.
If we look at the values of R4 in Table 2, they are zero for the girls. For the boys they are
the difference between the observed and estimated mean (M23 = M2 + M3 ). Recall that R4 tells
us how well the estimated quadratic growth curve fits the data for the boys. The values of R4
for the boys seem to be very close to zero at all ages, the values fall within one standard devi-
ation from the mean. This indicates that the observed and estimated mean values are close
and hence the data fits the quadratic growth curve very well. We can go one step further
to assess the necessity of the quadratic term via the values of R4 and M34 = M3 + R4 . Now,
compare the values of R4 and M34 given in Tables 2 and 4 respectively. It seems that there
is a considerable improvement at the ages of 8 and 10. Furthermore, a small improvement
has been obtained at the age of 14. However, at age 12 the estimated value obtained from
the linear growth curve seems closer to the observed one. Therefore, although the data seem
to fit the quadratic growth curve there is a need to check the significance of the quadratic
term, which will be discussed in a forthcoming paper.

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

Scand J Statist 33 Extended growth curve model 137

7. Concluding remarks
It is a well-known fact that, in any regression model and univariate linear models in general,
the ordinary residuals obtained by subtracting the estimated value from the observed one are
used to investigate how the data fits the estimated model and to validate model assumptions.
This is also true in multivariate linear models. However, because of the bilinear structure
in the model the ordinary residuals consist of two parts which give information about the
within and between assumptions. As a result, the two parts may happen to have large mag-
nitude but with opposite signs. Consequently, the ordinary residuals may be close to zero
and hence lead to the wrong conclusion that data fit the estimated model very well. There-
fore, there is a need to define and analyse other residuals which give a possibility to show
the different parts of the ordinary residuals. It is important that we examine each Ri in order
to have a better understanding of the structure of the model. Our methods can also be used
to define residuals in the MLNM(ABC). von Rosen (1995) has considered this model and
defined three residuals. Although the linear space he considered is the same space which was
obtained is utilized in this paper, our approach is better in the sense that we do not need
to estimate any inner product. The inner products involved in our method are direct con-
sequences of the maximum likelihood approach. Although things get a bit messy, we could
also use our method to define and discuss residuals in the general Extended Growth Curve
model with the nested subspace condition.
Each of the residuals defined in this paper bears important information about the bilinear
structure in the model and we may use this information to validate model assumptions and
detect outliers in the model (see section 4). However, it is not clear how close to zero the
residuals should be in order to conclude that the underlying assumptions are valid. If using
the residuals defined in this paper for detecting outliers, one may use two or three standard
errors as a cut off point. One may also calculate approximate cut off points using approxi-
mate densities based on the first two moments which are already given in here. One can find
this approach for the Growth curve model in von Rosen (1995). However, one could con-
struct test statistics which are functions of the residuals to see how the assumed model fits
the data and/or check the model assumptions. This will be shown in forthcoming research.

Acknowledgements
The first author’s studies have been financially supported by the Swedish International Devel-
opment Agency through the International Science Program, Uppsala University.

References
Banken, L. (1984). Eine Verallgemeinerung des GMANOVA modells. Dissertation, University of Trier,
Trier, Germany [in German].
Burnaby, T. P. (1966). Growth-invariant discriminant functions and generalized distances. Biometrics 22,
96–110.
Khatri, C. G. (1966). A note on a MANOVA model applied to the problems in growth curve. Ann. Inst.
Statist. Math. 18, 75–86.
Kshirsagar, A. M. & Smith, W. B. (1995). Growth curves. Marcel Dekker, New York, NY.
Potthoff, R. F. & Roy, S. N. (1964). A generalized multivariate analysis of variance model useful
especially for growth curves. Biometrika 51, 313–326.
Rao, C. R. (1965). The theory of least squares when the parameters are stochastic and its application
to the analysis of growth curves. Biometrika 52, 447–458.
von Rosen, D. (1989). Maximum likelihood estimators in multivariate linear normal models. J. Multi-
variate Anal. 31, 187–200.

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.

138 J. Seid Hamid and D. von Rosen Scand J Statist 33

von Rosen, D. (1990). Moments of a multivariate linear model with an application to the growth curve
model. J. Multivariate Anal. 35, 243–259.
von Rosen, D. (1991). The growth curve model: a review. Comm. Statist. Theory Methods 20, 2791–2822.
von Rosen, D. (1995). Residuals in the growth curve model. Ann. Inst. Statist. Math. 47, 129–136.
Seid Hamid, J. (2001). Residuals in the extended growth curve model. UUDM Report 2001:29, ISSN 1101–
3591. Department of Mathematics, Uppsala university, Uppsala, Sweden.
Sen, A. & Srivastava, M. S. (1990). Regression analysis, theory and applications. Springer, New York,
NY.
Srivastava, M. S. & Carter, E. M. (1983). An introduction to applied multivariate statistics. North
Holland, New York, NY.
Srivastava, M. S. & Khatri, C. G. (1979). An introduction to multivariate statistics. North Holland,
New York, NY.
Takemura, A. (1983). Tensor analysis of ANOVA decomposition. J. Am. Statist. Assoc. 78, 894–900.
Verbyla, A. P. & Venables, W. V. (1988). An extension of the growth curve model. Biometrika 75,
129–138.

Received April 2002, in final form April 2005

Jemila Seid Hamid, Department of Biometry and Engineering, Swedish University of Agricultural
Sciences, Uppsala, Sweden.
E-mail: [email protected]

 Board of the Foundation of the Scandinavian Journal of Statistics 2006.