NON-HIERARCHICAL MULTILEVEL MODELS

Chapter 1
NON-HIERARCHICAL MULTILEVEL MODELS
Jon Rasbash and William Browne
1.
INTRODUCTION
2.
CROSS-CLASSIFIED MODELS
2.1
TWO WAY CROSS-CLASSIFICATIONS :
A BASIC MODEL
In the models discussed in this book so far we have assumed that
the structures of the populations from which the data have been drawn
are hierarchical. This assumption is sometimes not justied. In this
chapter two main types of non-hierarchical model are considered. Firstly,
cross-classied models. The notion of cross-classication is probably
reasonably familiar to most readers. Secondly, we consider multiple
membership models, where lower level units are inuenced by more than
one higher level unit from the same classication. For example, some
pupils may attend more than one school. We also consider situations
that contain a mixture of hierarchical, crossed and multiple membership
relationships.
This section is divided into three parts. In the rst part we look at situations that can give rise to a two way cross-classication and introduce
some diagrams to describe the population structure, and discuss notation
for constructing a statistical model. In the second part we discuss some
of the possible estimation methods for estimating cross-classied models
and give an example analysis of an educational data set. In the third
part we then describe some more complex cross-classied structures and
give an example analyses of a medical data set.
Suppose, we have data on a large number of patients, attending many
hospitals and we also know the neighbourhood in which the patient lives
and that we regard patient, neighbourhood and hospital all as important
1
2
Table 1.1
Patients cross-classied by hospital and neighbourhood.
Neighbourhood 1 Neighbourhood 2 Neighbourhood 3
Hospital 1 XX
X
Hospital 2 X
X
Hospital 3
XX
X
Hospital 4
X
XXX
Table 1.2
Patients nested within hospitals within neighbourhoods.
Neighbourhood 1 Neighbourhood 2 Neighbourhood 3
Hospital 1 XXX
Hospital 2
XX
Hospital 3
XXX
Hospital 4 XXXX
sources of variation for the patient level outcome measure we wish to
study. Now, typically hospitals will draw patients from many dierent
neighbourhoods and the inhabitants of a neighbourhood will go to many
hospitals. No pure hierarchy can be found and patients are said to be
contained within a cross-classication of hospitals by neighbourhoods.
This can be represented schematically , for the case of twelve patients
contained within a cross-classication of three neighbourhoods by four
hospitals as in table 1.1.
In this example we have patients at level 1 and neighbourhood and
hospital are cross-classied at level 2. The characteristic pattern of a
cross-classication is shown, some rows contains multiple entries and
some columns contain multiple entries. In a nested relationship, if the
row classication is nested within the column classication then all the
entries across a row will fall under a single column and vice versa if
the column classication is nested within the row classication. For
example, if hospitals are nested within neighbourhoods we might observe
the pattern in table 1.2.
3
Many studies follow this simple two-way crossed structure, here are a
few examples :
Education: students cross-classied by primary school and secondary school.
Health: patients cross-classied by general practice and hospital.
Survey data: individuals cross-classied by interviewer and area of
residence.
2.1.1 Diagrams for representing the relationship between
classications. We nd two types of diagrams useful in expressing
the nature of relationships between classications. Firstly, unit diagrams
where we draw every unit (patient, hospital and neighbourhood, in the
case of our rst example) and connect each lowest level unit(patient) to
its parent units (hospital, neighbourhood). Such a representation of the
data in table 1.1 is shown in gure 1.1.
Hospital
Patient
H1
P1
Neighbourhood
H2
P2
P3
N1
Figure 1.1
P4
H3
P5
P6
N2
P7
H4
P8
P9
P10
P11
P12
N3
Diagrams for crossed structure given in table 1.1.
Note that we have two hierarchies present, patients within hospitals
and patients within neighbourhoods, we have organised the topology of
the diagram such that patients are nested within hospitals. However,
4
when we come to add neighbourhoods to the diagram we see that the
connecting lines cross, indicating we have a cross classication. Drawing
the hierarchical structure shown in table 1.2 gives the representation
shown in gure 1.2.
Neighbourhood
Hospital
Patient
N1
H1
P1
Figure 1.2
P2
H4
P3
P4
P5
P6
P7
P8
N2
N3
H2
H3
P9
P10
P11
P12
Diagrams for completely nested structure given in table 1.2.
Clearly, to draw such diagrams that include all units with large data
sets is not practical as there will be far to many nodes on the diagram
to t into a reasonable area. However, they can be used in schematic
form to convey the structure of the relationship between classications.
However, when we have four or ve random classications present (as
commonly occur with social data) even schematic forms of these diagrams can become hard to read. There is a more minimal diagram,
the classication diagram, which has one node for each classication.
Nodes connected by an arrow indicate a nested relationship, nodes connected by a double arrow indicate a multiple-membership relationship
(examples are given later) and unconnected nodes indicate a crossed relationship. Thus the crossed structure in gure 1.1 and the completely
nested structure of gure 1.2 are drawn as
5
Hospital
Neighbourhood
Neighbourhood
Hospital
Patient
Patient
(i) crossed structure
Figure 1.3
2.1.2
(ii) nested structure
Classication Diagrams for nesting and crossing.
Some notation for constructing a statistical model.
The matrix notation used in this book for describing hierarchical models,
that is,
yj = Xj + Zj j + ej
does not readily extend to the case of cross-classications. This is because this notation assumes a unique hierarchy where we write down the
generic equation for the j th level two unit. In a simple cross-classication
we have two sets of level two units, for example, hospitals and neighbourhoods, so which classication is j indexing?
We can extend the basic scalar notation to handle cross-classied
structures. Assume we have patients nested within a cross-classication
of neighbourhoods by hospital, that is the case illustrated in gure 1.3(i).
Suppose we want to estimate a simple variance components model giving
estimates of the mean response and patient, hospital and neighbourhood
level variation. In this case we can write the model in scalar notation as
yi(j ;j ) = 0 + j + j + ei(j ;j )
where 0 estimates the mean response, j1 indexes the neighbourhood
classication, j2 indexes the hospital classication, j is the random
1
2
1
2
1
2
1
6
Table 1.3
ure 1.1
Indexing table for neighbourhoods and hospitals for patients given in g-
i
1
2
3
4
5
6
7
8
9
10
11
12
nhbd(i) hosp(i)
1
2
1
2
1
2
2
3
3
2
3
3
1
1
1
2
2
3
3
3
4
4
4
4
eect for neighbourhood j1 , j is the random eect for hospital j2 ,
yi(j ;j ) is the response for the ith patient from the cell in the crossclassication dened by neighbourhood j1 and hospital j2 and nally
ei(j ;j ) is the patient level residual for the i'th patient from cell in the
cross-classication dened by neighbourhood j1 and hospital j2 .
2
1
2
1
2
Details of how this notation extends to represent more complex models and patterns of cross-classications are given in Rasbash and Browne,
2001. One problem with this notation is that as we t models with more
classications and more complex patterns of crossing, the subscript notation that describes the patterns becomes very cumbersome and dicult to read. We therefore prefer an alternative notation introduced in
Browne et al., 2000.
We can write the same model as
(2)
(3)
yi = 0 + nbhd
+ hosp
+ ei
(i)
(i)
where i indexes the observation level in this case patients, and nbhd(i)
and hosp(i) are functions that return the unit number of the neighbourhood and hospital, respectively, that patient i belongs to. Thus for the
data structure drawn in gure 1.1 the values of nbhd(i) and hosp(i) are
given in table 1.3.
Therefore the model for patient 3 would be
7
y3 = 0 + 1(2) + 1(3) + e3
and for patient 5 would be
y5 = 0 + 1(2) + 2(3) + e5
We number the classications from 2 upwards as we use classication
number 1 to represent the identity classication that applies to the observation level (like level 1 in a hierarchical model). This classication
simply returns the row numbers in the data matrix. As can be seen
random eects require bracketed superscripting with their classication
number to avoid ambiguity.
This simplied notation has the advantage that the subscripting notation does not increase in complexity as we add more classications.
This simplication is achieved because the notation makes no attempt
to describe the patterns of crossing and nesting present. This is useful
information and we therefore advocate the use of this notation in conjunction with the classication diagrams, as shown in gure 1.3, which
displays these patterns explicitily.
2.2
ESTIMATION ALGORITHMS
We will describe three estimation algorithms for tting cross-classication
models in detail and mention other alternatives. Each of these three
methods has advantages and disadvantages in terms of speed, memory
usage and bias and these will be discussed later. All three methods have
been implemented in versions of the MLwiN software package (Rasbash
et al., 2000) and all results in this paper are produced by this package.
2.2.1 An IGLS algorithm for estimating cross-classied models. The iterative generalized least squares estimates for a multilevel
model are those estimates which simultaneously satisfy both of the following equations:
^ = (XT V;1 X);1(XT V;1 y)
^ = (ZT (V );1 Z);1 ZT (V );1 y
where ^ are the estimated xed coecients and ^ is a vector containing
the estimated variances and covariances of the sets of random eects in
the model. V = Cov(y j X; ) and an estimate of V is constructed from
the elements of ^. y is the vector of elements of (y ; X )(y ; X )T
8
and therefore has length n2 (n is the length of the data set). V is the
covariance matrix of y and Z is the design matrix
linking y to V in the
N
regression of y on Z . V has the form V = V V. See Goldstein, 1986
for more details. Some of these matrices are massive for example (V );1
is dimensioned n2 by n2 , making a direct software implementation of
these estimating equations extremely resource intensive both in terms
of CPU time and memory consumed. However, in hierarchical models
V and V have a block diagonal structure which can be exploited by
customised algorithms (see Goldstein and Rasbash, 1996) which allow
ecient computation.
The problem presented by cross-classied models is that V (and therefore V ) no longer has the block diagonal structure which the ecient
algorithm requires.
2.2.2
Structure of V for cross classied models. Lets take a
look at the structure of V, the covariance matrix of y, for cross-classied
models and see how we can adapt the standard IGLS algorithm to handle
cross-classications.
The basic two level cross-classied model (with hospitals + neighbourhoods) can be written as :
(2)
(3)
yi = X + hosp
+ nhbd
+ ei
(i)
(i)
(2)
(3)
hosp
N (0; 2(2) ); nhbd
N (0; 2(3) ); ei N (0; e2 )
(i)
(i)
The variance of our response is now
(2)
(3)
var(yi ) = var(hosp
+ nhbd
+ ei ) = 2(2) + 2(3) + e2 :
(i)
(i)
The covariance between individuals a and b is
(2)
(3)
(2)
(3)
cov(ya ; yb ) = cov(hosp
+ nhbd
+ ea ; hosp
+ nhbd
+ eb )
(a)
( a)
(b)
(b)
which simplies to 2(2) for two individuals from the same hospital
but dierent neighbourhoods, 2(3) for two individuals from the same
neighbourhood but dierent hospitals, 2(2) + 2(3) for two individuals
from the same neighbourhood and the same hospital and zero for two
individuals who are from both dierent neighbourhoods and hospitals.
If we take a toy example of ve patients in two hospitals and introduce
a cross-classication with two neighbourhoods, as shown in table 1.4.
This generates a 5 by 5 covariance matrix for the responses of the ve
patients with the following structure :
9
Table 1.4
Indexing table for hospitals and neighbourhoods for 5 patients
i hosp(i) nhbd(i)
1
2
3
4
5
1
1
1
2
2
1
2
1
2
1
0 h+n+p
h
h+n
0
BB h
h+n+p
h
n
h
h+n+p
0
V=B
B@ h + n
0
n
0
h+n+p
n
0
n
2
2
where n = (2) ; h = (3) and p = e2 .
h
n
0
n
h
h+n+p
1
CC
CC
A
Here the data is sorted patient within hospital, this allows us to split
the covariance matrix into two components. A component for patients
within hospitals which has a block diagonal structure (P) and a component for neighbourhoods which is not block diagonal (Q) :
V=P+Q
where0
h+p h
h
0
0 1
BB h h + p h
0
0 C
C
B
h h+p 0
0 C
P=B h
C
@ 0
0
0 h+p h A
0
0
0
h h+p
and 0
n 0 n 0 n1
BB 0 n 0 n 0 CC
Q=B
B@ n 0 n 0 n CCA
0 n 0 n 0
n 0 n 0 n
Splitting the structure of V into a hierarchical, block-diagonal part
that the IGLS algorithm can handle in an ecient way and a nonhierarchical, non-block diagonal part forms the basis of a relatively ecient algorithm for handling cross-classied models.
If we take the dummy variable indicator matrix of neighbourhoods
(Z), then we have Q = ZZT n :
10
0
BB
Z=B
B@
01 0 1 0 11
1 01
0 1C
C , ZZT n = BBB 01 10 01 10 01 CCC n
1 0C
C
B@
C
0 1A
0 1 0 1 0A
1 0
1 0 1 0 1
We can dene a `pseudo-unit' that spans the entire data set, in our toy
example, all ve points, and declare this pseudo-unit to be level three in
the model (removing the neighbourhood level from the model). We can
now form the three level hierarchical model
(2)
(3)
(3)
yi = 0 + hosp
+ punit
Z + punit
Z + ei
(i)
(i);1 1
(i);2 2
2
3
"
#
4 punit i ; 5 N (0; ); = 0 ; 0
punit i ;
;
(3)
(3)
( )1
(3)
(3)
2
(3) 1
( )2
2
(3) 2
(2)
hosp
N (0; 2(2) ); ei N (0; e2 )
(i)
Here the level structure is patients within hospitals within the pseudo
unit level. Z1 and Z2 are columns 1 and 2 of Z. 2(3);1 and 2(3);2 are
both estimates of the between neighbourhood variation, therefore we
constrain them to be equal. Thus we can use the standard IGLS hierarchical algorithm to dene and estimate the correct covariance structure
for a cross-classied model. Now if we had 200 hospitals and 100 neighbourhoods, we would have to form 100 dummy variables for the neighbourhoods, allow them all to have variances at level 3 and constrain the
variances to be equal. Details of this algorithm are given in Rasbash
and Goldstein, 1994 and Bull et al., 1999 and it will be refered to as the
RG algorithm in later sections.
2.2.3
MCMC. The MCMC estimation methods (see Chapter 3
of this book for a fuller description) aim to generate samples from the
joint posterior distribution of all unknown parameters. They then use
these samples to calculate point and interval estimates for each individual parameter. The Gibbs sampler algorithm produces samples from
the joint posterior by generating in turn from the conditional posterior distributions of groups of unknown parameters. In chapter 3 the
Gibbs sampling algorithm for a Normally distributed response hierarchical model is given.
As we have seen in the notation section we can describe our model
as a set of additive terms, one for the xed part of the model and one
11
for each of the random classications. The MCMC algorithm works on
each of these terms seperately and consequently the algorithm for a crossclassied model is no more complicated than for a hierarchical model.
For illustration we present the steps for the following cross-classied
model based on the variance components hospitals by neighbourhoods
model and refer the interested reader to Browne et al., 2000 for more
general algorithms. Note that if the response is dichotomous or a count
then as in chapter 3 we can use the Metropolis-Gibbs hybrid method
discussed there.
The basic two level cross-classied model (with hospitals + neighbourhoods) can be written as :
(2)
(3)
yi = X + hosp
+ nhbd
+ ei
(i)
(i)
(2)
(3)
hosp
N (0; 2(2) ); nhbd
N (0; 2(3) ); ei N (0; e2 )
(i)
(i)
We can split our unknown parameters into 6 distinct sets : the xed ef(2)
fects, , the hospital random eects, hosp
, the neighbourhood random
(i)
(3)
2
eects, nhbd(i) , the hospitals variance, (2) the neighbourhood variance,
2(3) and the residual variance, e2 .
Then we need to generate random draws from the conditional distribution of each of these six groups of unknowns. MCMC algorithms
are generally used in a Bayesian context and consequently we need
to dene prior distributions for our unknown parameters. For generality we will use a multivariate Normal prior for the xed eects,
Npf (p; Sp ), and scaled inverse (SI) 2 priors for the three variances. For the hospital variance 2(2) SI2 (2 ; s22 ), for the neighbourhood variance 2(3) SI2 (3 ; s23 ) and for the residual variance
e2 SI2 (e; s2e ). The steps are then as follows:
In step 1 of the algorithm the conditional posterior distribution in the
Gibbs update for the xed eects parameter vector is multivariate
normal with dimension pf (the number of xed eects) :
p( j y; (2) ; (3) ; 2(2) ; 2(3) ; e2 ) Npf (b; Db ); where
h
i;1
T
Db = hPNi=1 (Xi)e Xi + Sp;1 i and
T
b = Db Pi (Xi)e di + Sp;1p ; where
(2)
(3)
di = yi ; hosp
; nhbd
:
(i)
(i)
2
2
In step 2 we update the hospital residuals, k(2) , using Gibbs sampling
with a univariate Normal full conditional distribution :
12
p(k(2) j y; ; (3) ; 2(2) ; 2(3) ; e2 ) N (bk(2) ; Db k(2) ); where
;1
bDk(2) = nk + 1
and
e
(2)
ubk = Db k
(2)
P
2
2
(2)
d(2)
i;hosp(i)=k ie2
; where
(3)
d(2)
i = yi ; Xi ; nhbd(i) :
(2)
In step 3 we update the neighbourhood residuals, k(3) , using Gibbs sampling with a univariate Normal full conditional distribution :
p(k(3) j y; ; (2) ; 2(2) ; 2(3) ; e2 ) N (bk(3) ; Db k(3) ); where
;1
Db k(3) = nke + 1
and
(3)
P
2
2
(3)
b (3) i;nhbd(i)=k di ; where
ub(3)
k = Dk
e
(2)
d(3)
=
y
;
X
;
i
i
i
hosp(i) :
(3)
3
Note that in the above two steps n(kc) refers to the number of individuals in the kth unit of classication c.
In step 4 we update the hospital variance 2(2) using Gibbs sampling
and a Gamma full conditional distribution for 1=2(2) :
p(1=2(2) j y; ; (2) ; (3) ; 2(3) ; e2 ) Gamma
hn
2+ 2
2
2
In step 5 we update the neighbourhood variance 2(3) using Gibbs sampling and a Gamma full conditional distribution for 1=2(3) :
p(1=2(3) j y; ; (2) ; (3) ; 2(2) ; e2 ) Gamma
hn
3+ 3
2
hN
+ e
2
3
i
; 12 Pi e2i + es2e :
The above 6 steps are repeatedly sampled from in sequence to produce
correlated chains of parameter estimates from which point and interval
estimates can be created as in chapter 3.
2.2.4
i
; 21 Pnj =1 (j(3) )2 + 3 s23 :
In step 6 we update the observation level variance e2 using Gibbs sampling and a Gamma full conditional distribution for 1=e2 :
p(1=e2 j y; ; (2) ; (3) ; 2(2) ; (3) ) Gamma
i
; 21 Pnj =1 (j(2) )2 + 2 s22 :
AIP method. The Alternating Imputation Prediction (AIP)
method is a data augmentation algorithm for estimating cross-classied
13
models with large numbers of random eects. Comprehensive details of
this algorithm are given in Clayton and Rasbash, 1999. We now give an
overview.
Data augmentation has been reviewed by Schafer, 1997. Tanner and
Wong, 1987 introduced the idea of data augmentation as a stochastic version of the EM algorithm for maximum likelihood estimation in
problems involving missing data. Corresponding to the E and M steps
of Tanner and Wong we have
I(mputation) step - impute missing data by sampling the distribution of
the missing data conditional upon the observed data and current values
of the model parameters.
P(osterior) step - sample parameter values from the complete data posterior distribution; these will be used for the next I-step.
In the context of random eect models, the random eects play the
role of missing data. If the observed data are denoted by y, the random
eects by and the model parameters by and if we denote the probability distribution of y conditional on X as p(yjX) then the algorithm is
specied (at step t) by
I step - Draw a sample (t) from p( j y; = (t;1) )
P step - Draw a sample (t) from p( j y; = (t) )
Repeated application of these two steps delivers a stochastic chain
with equilibrium distribution p(; j y) in a similar way to the MCMC
algorithm. Now lets look at how we can adapt this method to t a
crossed random eects model when the only estimating engine we have
at our disposal is one optimized for tting nested random eects.
An n-way cross-classied model can be broken down into n sub-models
each of which is a 2 level hierarchical model. For example, patients
nested within a cross-classication of neighbourhood by hospital can be
broken down into a patient within hospital sub-model and a patient
within neighbourhood sub-model.
Take the simple model
(2)
(3)
yi = Xi + nbd
+ hosp
+ ei
(i)
(i)
where neighbourhood and hospital are cross-classied. This crossclassied model can be portioned into two hierarchical sub models :
patients within neighbourhoods (model N) and patients within hospitals
(model H). An informal description of the AIP algorithm is :
1. Start by tting model N using an estimation procedure for 2 level
models.
14
2. Sample the model parameters from an approximation to their joint
posterior distribution. That is sample the xed eects, the neighbourhood level variance and the patient level variance; denote
these samples by [0;2] , 2[0;2] and e2[0;2] respectively. Here [0,2]
labels a term as belonging to AIP iteration 0, for classication
number 2, that is neighbourhood. This is the P-step for the neighbourhood classication.
3. Next sample a set of neighbourhood level random eects(o[0;2] )
from p([0;2] j y; [0;2] ; 2[0;2] ; e2[0;2] ) . This is the I-step for the
neighbourhood classication.
4. Oset o[0;2] from y, that is form y = y ; o[0;2], re-sort the data
according to hospitals and t model H using the new oset response
y .
5. Next sample [0;3] , 2[0;3] and e2[0;3], from this second model, H.
This is the P-step for the hospital classication.
6. Sample a set of hospital level random eects(o[0;3] ) from p([0;3] j
y; [0;3]; 2[0;3] ; e2[0;3] ). This is the I-step for the hospital classication.
This completes one iteration of the AIP algorithm, this is an ImputationPosterior algorithm that Alternates between the neighbourhood and hospital classications. We proceed by forming y = y ; o[0;3], that is osetting the sampled hospital residuals from y and using that as a response
in step 1. After T iterations the procedure delivers the following two
chains, that can be used for inference
f ; ; ; ; e ; g; f ; ; ; ; e ; g : : : f T; ; T; ; e T; g
f ; ; ; ; e ; g; f ; ; ; ; e ; g : : : f T; ; T; ; e T; g
[0 2]
[0 3]
2
[0 2]
2
[0 3]
2
[0 2]
2
[0 3]
[1 2]
[1 3]
2
[1 2]
2
[1 3]
2
[1 2]
2
[1 3]
[
2]
[
3]
2
[
2
[
2]
3]
2
[
2
[
2]
3]
Note that we get two sets of estimates for both the xed eects and
the level 1 variance with the AIP algorithm and these should be approximately equal.
2.2.5
Other Methods. Raudenbush, 1993 considers an empirical Bayes approach to tting cross-classied models based on the EM
algorithm. He considers the specic case of two classications where one
of the classications has many units whilst the other has far fewer and
shows two educational examples to illustrate the method.
Two other recent approaches that can be used for tting cross-classied
models, in particular with non-Normal responses are Gauss-Hermite
quadrature within PQL estimation Pan and Thompson, 2000 and the
15
HGLM model framework as described in Lee and Nelder, 2000. Neither
of these approaches has been designed with speed of estimation in mind
and so they are currently not feasible for the size of some of the problems
that are considered in practice.
2.2.6
Comparison of estimation methods. The RG method
when it works is generally fairly quick to converge where all or all but one
of the crossed classications have small numbers of units. When there
are multiple crossed classications with large numbers of untis then the
speed of the RG algorithm deteriorates and memory usage is greatly increased, often exhausting the available memory. The AIP method does
not have these memory problems but will be slower for structures that
are almost hierarchical. Although this method works reasonably well, if
the response is a binary variable and quasi-likelihood methods need to
be used, then this method like the RG method is still aected by the bias
that is inherent in quasi-likelihood methods for binary response multilevel models (See Goldstein and Rasbash, 1996). The MCMC methods
have no bias problems although there are still issues on which prior distributions to use for the variance parameters. They also, like the AIP
methods do not have any memory problems. They are however generally
computationally a lot slower as they are estimating the whole distribution and not simply the mode, although as the structure of the data
becomes more complex the ratio of speed dierence is reduced.
2.2.7 An example analysis of a two way cross-classication:
primary schools crossed with secondary schools. We will here
consider tting the RG method using the IGLS algorithm, the MCMC
method based on Gibbs sampling (Browne et al., 2000) and the AIP
method to an educational example from Fife in Scotland. Here we have
as a response the exam results of 3,435 children at age 16. We know for
each child both the primary school and secondary school that they attended and we are interested in partitioning the variance between these
two sources and individual pupil level variation. The classication diagram is shown in gure 1.4. There are 148 primary schools that feed
into 19 secondary schools in the dataset. Of the 148 primary schools, 59
are nested within a single secondary school, whilst another 62 have at
most 3 pupils that do not go to the main secondary school so we have
an almost nested structure. This structure is particularly suited for the
RG algorithm.
We will t the following model to the dataset
16
Primary School
Secondary School
Pupil
Figure 1.4
Classication Diagram for the Fife educational example
Table 1.5
Point estimates for the Fife educational dataset.
Parameter
Mean achievement (0 )
Secondary school variance (2(2) )
Primary school variance (2(3) )
Individual level variance (e2 )
IGLS
5.50 (0.17)
0.35 (0.16)
1.12 (0.20)
8.10 (0.20)
MCMC
5.50 (0.18)
0.41 (0.21)
1.15 (0.213)
8.12 (0.20)
AIP
5.51 (0.19)
0.34 (0.15)
1.11 (0.20)
8.11 (0.20)
(2)
(3)
yi = 0 + SEC
+ PRIM
+ ei
(i)
(i)
(2)
(3)
SEC
N (0; 2(2) ); PRIM
N (0; 2(3) ); ei N (0; e2 ):
(i)
(i)
The results are shown in table 1.5.
From table 1.5 we can see that in this example there is more variation
between primary schools than between secondary schools. The MCMC
17
estimates replicate the IGLS estimates with slightly greater higher level
variances (mean versus mode estimates) due to the skewness of the posterior distribution. The AIP method gives very similar results to the
IGLS method. A further discussion of these results is given in Goldstein, 1995.
2.3
MODELS FOR MORE COMPLEX
POPULATION STRUCTURES
In this section we will consider expanding the simple two cross-classied
structure to accomodate more classications and more complex structures.
2.3.1
Example scenarios. Lets take the situation described in
the classication diagram drawn in gure 1.3(i) where patients lie within
a cross-classication of hospitals by areas. We may have information on
the doctor that treated each patient and doctors may be nested within
hospitals. The classication diagram for this structure is shown in gure
1.5.
Hospital
Doctor
Neighbourhood
Patient
Classication Diagram for two crossed hierarchies (patients within doctors
within hospitals)*(patients within neighbourhoods).
Figure 1.5
A variance components model for this structure is written as
18
(2)
(3)
(4)
yi = 0 + nhbd
+ hosp
+ doct
+ ei
(i)
(i)
(i)
If doctors work across hospitals and are therefore not nested within
hospital we then have a three way cross-classication which is drawn in
gure 1.6.
Hospital
Neighbourhood
Doctor
Patient
Classication Diagram for three crossed hierarchies (patients within hospitals)* (patients within doctors)*(patients within neighbourhoods).
Figure 1.6
Note that the variance components model for the structure in gure 1.6 is also described by the same equation. This is a reection of
the fact that the model notation for describing the random eects simply lists the classications that are sources of variation for the response
we are modelling. In the variance components model we only have an
intercept term which varies across all four classications present. Suppose we had another explanatory variable, x1 and we wished to allow its
coecient to vary across the doctor classications; we would write this
model as
(2)
(3)
(4)
(4)
yi = 0 + nhbd
+ hosp
+ doct
+ 1 x1i + doct
x + ei
(i)
(i)
(i);0
(i);1 1i
or alternatively we can express the model as :
19
yi = 0i + 1i x1i + ei
(2)
(3)
(4)
0i = 0 + nhbd
+ hosp
+ doct
(i)
(i)
(i);0
(4)
1i = 1 + doct
(i);1
It may be that the scenario described in gure 1.6 is further complicated because hospitals, doctors and neighbourhoods are all nested
within regions. In this case the classication diagram becomes as in
gure 1.7.
Region
Hospital
Neighbourhood
Doctor
Patient
Classication Diagram for three crossed hierarchies nested within a higher
level classication.
Figure 1.7
Extending the last model to incorporate a simple random eect for
the region classication we have
yi = 0i + 1i x1i + ei
(2)
(3)
(4)
(5)
0i = 0 + nhbd
+ hosp
+ doct
+ reg
(i)
(i)
(i);0
(i)
(4)
1i = 1 + doct
(i);1
These few example scenarios indicate how the classication diagrams
and simplied notation can extend to describe patterns of crossings of
arbitrary complexity.
2.3.2 An example analysis of a complex cross-classied structure : Articial Insemination data. We consider a data set con-
20
cerning articial insemination by donor. Detailed description of this
data set and the substantive research questions addressed by modelling
it within a cross-classied frame work are given in Ecochard and Clayton, 1998. The data was re-analysed in Clayton and Rasbash, 1999 as an
example case study demonstrating the properties of the AIP algorithm
for estimating cross-classied models.
The data consists of 1901 women who were inseminated by sperm
donations from 279 donors. Each donor made multiple donations, there
were 1328 donations in all. A single donation is used for multiple inseminations. Each woman receives a series of monthly inseminations, 1
insemination per ovulatory cycle. The data contain 12100 cycles within
the 1901 women.
There are two crossed hierarchies, a hierarchy for donors and a hierarchy for women. Level 1 corresponds to measures made at each ovulatory cycle. The response we analyse is the binary variable indicating
if conception occurs in a given cycle. The hierarchy for women is cycles within women. The hierarchy for donors is cycles within donations
within donors. Within a series of cycles a women may receive sperm from
multiple donors/donations. The classication diagram for this structure
is given in gure 1.8. The model tted to the data is
Donor
Donation
Woman
Cycle
Figure 1.8
Classication Diagram for the articial insemination example model.
21
Table 1.6
Results for the Articial insemination example.
Parameter
MCMC
AIP
intercept (0 )
-3.92 (0.21)
-3.90 (0.21)
azoospermia (1 )
0.21 (0.09)
0.22 (0.10)
semen quality (2 )
0.18 (0.03)
0.18 (0.03)
womens age > 35 (3 )
-0.29 (0.12)
-0.27 (0.12)
Sperm count (4 )
0.002 (0.001)
0.002 (0.001)
Sperm motility (5 )
0.0002 (0.0001) 0.0002 (0.0001)
Insemination too early (6 ) -0.69 (0.17)
-0.67 (0.17)
Insemination too late (7 )
-0.27 (0.09)
-0.25 (0.09)
Donor variance (2(4) )
0.11 (0.06)
0.10 (0.06)
2
Donation variance ((3) )
0.36 (0.074)
0.34 (0.065)
Women variance (2(2) )
1.02 (0.15)
1.01 (0.11)
yi Bernouilli(i )
logit(i ) = 0 + azooi 1 + semenqi 2 + age > 35i 3 +
spermcounti 4 + spermmoti 5 + iearlyi 6 +
(2)
(3)
ilatei 7 + woman
+ donation
+ ((4)
(i)
(i)
donor(i)
(3)
(4)
(2)
2
2
2
woman
N
(0
;
)
;
N
(0
;
)
;
(2) donation(i)
(3) donor(i) N (0; (4) )
(i)
(1.1)
Note that azoospermia (azoo) is a dichotomous variable indicating
whether the fecundability of the woman is impaired (0 impaired, 1 not
impaired). The results of tting this model from the MCMC and AIP
estimation procedures are given in table 1.6. This model could not be
tted using the RG algorithm. This is because if the data is sorted according to women then we need to t 279 dummy variables for donors
and 1328 dummy variables for donations. Alternatively, if we sort the
data according to donations within donors we have to t 1901 dummy
variables for women. Either way, the size of these data matrices cause
problems of insucient memory. Even if these memory problems can be
worked around the numerical instability of the constraining procedure,
that attempts to constrain over a thousand seperately estimated variances to be equal, causes the adapted IGLS algorithm to fail to converge.
After inclusion of covariates there is considerably more variation in
the probability of a successful insemination attributable to the women
22
hierarchy than the donor hierarchy. Both the AIP and MCMC methods
give similar estimates for all parameters. The xed eect estimates
show that the probability of conception is increased with azoospermia
and increased sperm quality, count and motility but decreased with the
age of the woman and with inseminations that are too early or late.
3.
MULTIPLE MEMBERSHIP MODELS
3.1
A BASIC STRUCTURE FOR TWO
LEVEL MULTIPLE MEMBERSHIPS
As we have seen from the previous section, allowing classications to
be crossed gives rise to a large family of additional model structures
that can be estimated. The other main restriction of the basic multilevel model is the need for observations to belong to a unique classication unit i.e. every pupil belongs to a particular class, every patient
is treated at a particular hospital. Often however, over time a patient
may be treated at several hospitals and depending on the response of
interest all of these hospitals may have inuence. In this section we will
rstly introduce the idea of multiple membership and give some example
scenarios where it may occur. We will then discuss the possible estimation procedures that can be used to t multiple membership models and
nish the chapter with a simulated example from the eld of education.
Supose we have data on a large number of patients that attend their
local hospital and during the course of their hospital stay they are treated
by several nurses and we regard the nurses as an important factor on
the patients outcome of interest. Now typically each patient will be seen
by more than one nurse during their stay (although some will only see
1) but there are many nurses and so we will treat nurses as a random
classication rather than as xed eects. To illustrate this table 1.7
shows the nurses seen by the rst 4 patients.
We can consider this structure in a unit diagram as shown in gure 1.9.
Here each line in the diagram corresponds to a tick mark in the table.
Again as our dataset gets larger such unit diagrams become impractical
as there will be too many nodes and so we will resort to using the classication diagrams introduced earlier for cross-classied models. If we
wish to include multiple membership classications in such diagrams we
use the convention of a double arrow to represent multiple membership.
This will lead to the classication diagram shown in gure 1.10 for the
above patients and nurses example.
23
Table 1.7
Table of patients that are seen by multiple nurses.
Patient 1
Patient 2
Patient 3
Patient 4
Nurse
Patient
Figure 1.9
Nurse 1 Nurse 2 Nurse 3
p
p
p
p
p
N1
P1
p
N2
P2
p
N3
P3
P4
Unit Diagram for multiple membership patients within nurses example.
3.1.1 Example scenarios. Many studies have multiple membership structure, here are a few examples :
Education : pupils change school/class over the course of their
education and each school/class has an eect on their education.
Health : patients are seen by several doctors and nurses during the
course of their treatment.
Survey data : Over their lifetime individuals move household and
each household has a bearing on their lifestyle, health, salary etc.
3.1.2
Constructing a statistical model. Returning to our example of patients being seen by multiple nurses, we have patient 1's
response being aected by nurse 1 and nurse 3 while patient 2 is only
aected by nurse 1. As we are treating nurse as a random classication
24
Nurse
Patient
Figure 1.10
example.
Classication Diagram for multiple membership patients within nurses
we would like each patient's response to have equal eect on the nurse
classication variance so we generally weight the random eects to sum
to 1. For example let's assume patient 1 has been treated by nurse 1
for 2 days and nurse 3 for 1 day. Then we may give nurse 1 a weight
of 32 and nurse 3 a weight of 13 . Often we do not have information on
the amount of time patients are seen by each nurse and so we commonly
allocate equal weights (in this case 21 ) to each nurse.
We can then write down a general two level multiple-membership
model as
X (2) (2)
yi = X +
wi;j j + ei
j 2nurse(i)
j(2) N (0; 2(2) ); ei N (0; e2 )
(2)
nurse(i) is the set of nurses seen by patient i and wi;j
is the weight
given to nurse j for patient i. Here we assume that
X (2)
wi;j = 18i:
j 2nurse(i)
25
If we wish to write out this model for the rst four patients from the
example we get
y1 = X1 + 12 1(2) + 12 3(2) + e1
y2 = X2 + 1(2) + e2
y3 = X3 + 21 2(2) + 12 3(2) + e3
y4 = X4 + 21 1(2) + 12 2(2) + e4
3.2
ESTIMATION ALGORITHMS
3.2.1
An IGLS algorithm for multiple membership models.
There are two main algorithms for multiple membership models, an
adaption of the Rasbash and Goldstein, 1994 algorithm described earlier
and the MCMC method. The AIP method has not been extended to
cater for multiple membership models.
Earlier we described how to t a cross-classied model by absorbing
one of the cross-classications into a set of dummy variables (The RG
method). A slight modication is required to allow this technique to be
used to t multiple membership models. First lets consider a two level
hierarchical model for patients within nurses:
(2)
yi = 0 + nurse
+ ei
(i)
(2)
nurse
N (0; 2(2) ); ei N (0; e2 ):
(i)
We can reparamaterise this simple two level model as
(2)
(2)
(2)
(2)
yi = 0 +zi;1nurse
+zi;2 nurse
+zi;3 nurse
+: : :+zi;J nurse
+ei
(i);1
(i);2
(i);3
(i);J
2
66 nurse i ;
66 nurse i ;
66 66 nurse. i ;
4
(2)
(2)
(2)
( )1
( )2
( )3
(2)
nurse
(i);J
3
2
0
0
77
66 0 ; 0
;
77
66 0
0
;
77 N (0; ); = 66
...
...
77
66 ...
...
...
5
4 ...
2
(2) 1
(2)
2
(2) 2
(2)
0
ei N (0; e2 )
0
2
(2) 3
0
...
0
...
0
...
0
... ...
... ...
... 2(2);J
3
77
77
77
77
5
26
where zi;j is a dummy variable which is 1 if patient i is seen by nurse
j , 0 otherwise and J is the total number of nurses. Also we add the
constraint 2(2);1 = 2(2);2 = : : : = 2(2);J . Now these two models will
deliver the same estimates, however the second formulation will take
much longer to compute. The advantage of the second model formulation
is that it is straightforward to extend it to the multiple membership
case. Suppose patients are not nested within a single nurse but are
multiple members of nurses with membership probabilities, i;j . We can
simply replace zi;j with i;j in the second formulation and estimation
can proceed in an identical fashion but will now deliver estimates for the
multiple membership model.
3.2.2
MCMC. Once again we will use a Gibbs sampling algo-
rithm that relies on updating groups of parameters in turn from their
conditional posterior distributions. For illustration we present the steps
for the following simple multiple membership model based on the variance components model patients within nurses described earlier. We
once again refer the interested reader to Browne et al., 2000 for more general algorithms and note that if the response is dichotomous or a count
then as in chapter 3 we can use the Metropolis-Gibbs hybrid method
discussed there.
The basic two level multiple membership model (patients within nurses)
can be written as :
X (2) (2)
yi = X +
wi;j j + ei
j 2nurse(i)
j(2) N (0; 2(2) ); ei N (0; e2 )
We can split our unknown parameters into 4 distinct sets : the xed
eects, , the nurse random eects, j(2) , the nurse level variance, 2(2)
and the patient level residual variance, e2 .
We then need to generate random draws from the conditional distribution of each of these four groups of unknowns. We will dene prior distributions for our unknown parameters as follows: For generality we will use
a multivariate Normal prior for the xed eects, Npf (p ; Sp ), and
scaled inverse 2 priors for the two variances. For the nurse level variance
2(2) SI2 (2 ; s22 ), and for the patient level variance e2 SI2 (e ; s2e ).
The steps are then as follows:
In step 1 of the algorithm the conditional posterior distribution in the
Gibbs update for the xed eects parameter vector is multivariate
normal with dimension pf (the number of xed eects) :
27
p( j y; (2) ; 2(2) ; e2 ) Npf (b; Db ); where
h
i;1
T
Db = hPNi=1 (Xi)e Xi + Sp;1 i and
T
b = Db Pi (Xi)e di + Sp;1p ; where
(2) (2)
di = yi ; Pj 2nurse(i) wi;j
j :
2
2
In step 2 we update the nurse residuals, k(2) , using Gibbs sampling with
a univariate Normal full conditional distribution :
p(k(2) j y; ; 2(2) ; e2 ) N (bk(2) ; Db k(2) ); where
#;
P
w
and
Db k = i;k2nurse i + "
#
ubk = Db k Pi;k2nurse i w d ; where
di;k = yi ; Xi ; Pj2nurse i ;j 6 k wi;j j :
(2)
"
( )
(2)
1
(2)
( i;k )2
1
2
2
e
(2)
(2) (2)
(2)
i;k i;k
2
( )
e
(2)
(2) (2)
( ) =
In step 3 we update the nurse level variance 2(2) using Gibbs sampling
and a Gamma full conditional distribution for 1=2(2) :
p(1=2(2) j y; ; (2) ; e2 ) Gamma
hn
2+ 2
2
i
; 21 Pnj =1(j(2) )2 + 2 s22 :
2
In step 4 we update the patient level variance e2 using Gibbs sampling
and a Gamma full conditional distribution for 1=e2 :
p(1=e2 j y; ; (2) ; 2(2) ) Gamma
hN
+ e
2
i
; 12 Pi e2i + e s2e :
The above 4 steps are repeatedly sampled from in sequence to produce
correlated chains of parameter estimates from which point and interval
estimates can be created as in chapter 3.
3.2.3
Comparison of estimation methods. As in the compar-
ison for cross-classied models there are benets for both methods. The
RG method is fairly quick but the number of level 2 units determines
the size of some of the matrices involved and the number of constraints
that the method has to apply. These dependencies lead to numerical
instability or memory exhaustion in situations with more than a few
hundred level 2 units. The MCMC methods although again computationally slower do not suer from these memory problems.
3.2.4 An example analysis of a two level multiple membership model : Children moving school . We consider a simulated
28
Table 1.8
Results for the multiple membership schools example.
Parameter
RG RIGLS estimates MCMC Estimates
intercept (0 )
0.002 (0.040)
0.003 (0.040)
LRT eect (1 )
0.565 (0.012)
0.565 (0.013)
2
School variance ((3) )
0.093 (0.018)
0.096 (0.020)
Pupil variance (2(2) )
0.570 (0.013)
0.571 (0.013)
data example based on the problem in education of adjusting for the
fact that pupils move school during the course of their studies. We will
consider a study with 4059 students from 65 schools taken from Rasbash
et al., 2000. The actual data in the study has each child belonging to
1 school but we will assume that over their education 10% of children
moved school so we will choose at random for 10% of the children a
second school. We will assume that information about when the move
occured is unavailable and so for these children we will allocate equal
weights of 0.5 to each school. Browne et al., 2000 considered this as
the basis for a simulation experiment by generating 1000 datasets with
this structure to show the bias and coverage properties of the MCMC
method. We will instead consider the true response on our modied
structure. We have as a response the pupil's total (normalised) exam
score in all GCSE exams taken at age 16 and as a predictor the pupil's
(standardised) score in a reading test taken at age 11. As we are interested in progress from age 11-16 it makes sense to consider the eect of
all schools attended in this period.
We will consider the following model
normexami = 0 + 1 standlrti +
X
j 2school(i)
(2) (2)
wi;j
j + ei
j(2) N (0; 2(2) ); ei N (0; e2 )
We t this model using both the RG and MCMC methods and the
results can be seen in table 1.8
From the table we can see that both methods give similar results.
If we compare the results here with the results in Rasbash et al., 2000
we see only slight changes to the estimates with the level 2 variance
slightly decreased and the level 1 variance slightly increased. However
in cases where there is greater amounts of multiple membership the
29
variance estimates can be altered if this multiple membership is ignored,
for example if we randomly assigned every pupil to a second school the
variances change to 0.088 and 0.609 at levels 1 and 2 respectively.
4.
COMBINING MULTIPLE MEMBERSHIP
AND CROSS-CLASSIFIED STRUCTURES
IN A SINGLE MODEL
Consider two of our earlier examples in the eld of education, rstly
pupils in a crossing of primary schools and secondary schools and secondly pupils who are moving from school to school. We could assume
that these two structures occur simultaneously and we will then end
up with a model structure that contains both a multiple membership
classication (secondary schools) and a second classication (primary
schools) that is crossed with the rst. This scenario can be represented
by a classication diagram as in gure 1.11. Browne et al., 2000 refer to
models that contain both multiple memberships and cross classications
as multiple membership multiple classication (MMMC) models.
P. School
S. School
Pupil
Figure 1.11
model.
Classication Diagram for the neighbours/schools multiple membership
30
4.1
EXAMPLE SCENARIOS
Many studies have both cross-classied and multiple membership classications in their structure, a few examples are the following :
Education : pupils can be aected by the crossing of the neighbourhood they live in and the school they attend. They could also
change class over their period of education and so this multiple
membership class classication will be crossed with the neighbourhood classication.
4.2
Health : patients are seen by several doctors during their treatment and may visit several hospitals. Doctors who are specialists
may move from hospital to hospital and so are crossed with the
hospitals.
Survey Data : individuals will belong to many households over
the course of their lives and will reside in several properties. An
entire household may move to a new property so households can
be crossed with properties and all the households/properties can
have an eect on the individual. See Goldstein et al., 2000 for
more details.
Spatial Data : individuals will belong to a particular area but will
also be aected by multiple neighbouring areas.
CONSTRUCTING A STATISTICAL
MODEL
If we return to our example of pupils attending multiple secondary
schools but coming from one primary school we need to combine the multiple membership and cross classied model structures into one model.
As we are treating the secondary schools as a random classication we
would like each pupil to have an equal eect on the secondary school
classication so we will use weights that add to 1 when a pupil attends
more than one secondaryschool. We will let second(i) be the list of
secondary schools that child i has attended.
We can then write down a general two classication MMMC model
as
X (2) (2) (3)
yi = X +
wi;j j + prim + ei
j 2second(i)
(3)
j(2) N (0; 2(2) ); prim
N (0; 2(3) ); ei N (0; e2 ):
31
(2)
Here wi;j
is the weight given to secondary school j for pupil i. Here
P
(2)
we assume that j 2second(i) wi;j
= 18i. Both the RG algorithm and
the MCMC method can be used to t these models that combine both
multiple membership and cross classication.
4.3
AN EXAMPLE ANALYSIS : DANISH
POULTRY FARMING
Rasbash and Browne, 2001 consider an example from veterinary epidemiology concerning the outbreaks of salmonella typhimurium in ocks
of chickens in poultry farms in Denmark between 1995 and 1997. The
response of interest is whether salmonella typhimurium is present in a
ock and in the data collected 6.3% of ocks had the disease. At the
observation level, each observation represents a ock of chickens. For
each ock the response variable is whether or not there was an instance
of salmonella in that ock. The basic data have a simple hierarchical
structure as each ock is kept in a house on a farm until slaughter. As
ocks live for a short time before they are slaughtered several ocks will
stay in the same house each year. The hierarchy is as follows 10,127
child ocks within 725 houses on 304 farms.
Each ock is created from a mixture of parent ocks (up to 6) of which
there are 200 in Denmark and so we have a crossing between the child
ock hierarchy and the multiple membership parent ock classication.
The classication diagram can be seen in gure 1.12. We also know the
exact makeup of each child ock (in terms of parent ocks) and so can
use these as weights for each of the parent ocks. We are interested
in assessing how much of the variability in typhoid incidence can be
attributed to houses, farms and parent ocks.
There are also 4 hatcheries in which all the eggs from the parent
ocks are hatched. We will therefore t a variance components model
that allows for dierent average rates of salmonella for each year with
hatchery included in the xed part as follows :
salmonellai Bernouilli(i )
logit(i ) = 0 + Y 96 1 + Y 97 2 + hatch2 3 +
P
(2)
(3)
(4) (4)
hatch3 4 + hatch4 5 + House
+ Farm
+ j 2P:flock(i) wi;j
j
(i)
(i)
(2)
(3)
House
N (0; 2(2) ); Farm
N (0; 2(3) ); (4) N (0; 2(4) )
(i)
(i)
(1.2)
32
Farm
House
Parent Flock
Child Flock
Figure 1.12
Classication diagram for the Danish poultry model.
The results of tting model 1.2 using both the Rasbash and Goldstein
method with 1st order MQL estimation and the MCMC method can be
seen in table 1.9. The quasi-likelihood methods are numerically rather
unstable and we could not get either 2nd order MQL or PQL to t this
model.
We can see here that there are large eects for the year the chickens
were born suggesting that salmonella was more prevalent in 1995 than
the other years. The hatchery eects were also large suggesting chickens
produced in hatcheries 1 and 3 had a larger incidence of salmonella.
There is a large variability for the parent ock eects and for the farm
eects which are of similar magnitude. There is less variability between
houses within farms.
4.3.1 Method comparison. The MCMC results were run for
50,000 iterations after a burn-in of 20,000 (This took just under 2 hours
on a 733MHz PC) as we used arbitrary starting values and so the chain
took a while to converge. From table 1.9 we can see reasonable agreement between the two methods, although the xed eects in MQL are all
smaller as is the farm level variance. This behaviour was shown in simulations on a nested 3 level binary response data structure in Rodriguez
33
Table 1.9
Results for the Danish poultry example.
Parameter
intercept (0 )
1996 eect (1 )
1997 eect (2 )
Hatchery 2 eect (3 )
Hatchery 3 eect (4 )
Hatchery 4 eect (5 )
Parent ock variance (2(4) )
Farm variance (2(3) )
House variance (2(2) )
1st MQL
MCMC Estimates
-1.862 (0.184) -2.322 (0.213)
-1.004 (0.138) -1.239 (0.162)
-0.852 (0.159) -1.165 (0.187)
-1.458 (0.222) -1.733 (0.255)
-0.250 (0.209) -0.211 (0.252)
-1.007 (0.353) -1.062 (0.388)
0.892 (0.184)
0.895 (0.179)
0.639 (0.121)
0.927 (0.197)
0.206 (0.096)
0.208 (0.108)
and Goldman, 1995 with the improvements of the MCMC method shown
in Browne and Draper, 2000 and so this suggests that the MCMC results
should be more accurate.
4.4
COMPLEX RANDOM EFFECTS
Model 1.2 is essentially another variance components model but we
could t a model that has complex variation at one of the higher classications. To illustrate this we will modify the farm level variance to
account for dierent variability between years at the farm level that is
(3)
we replace the simple farm level random eects, Farm
with 3 sets of
(i)
eects one for each year. Our expanded model is then as follows :
salmonellai Bernouilli(i )
logit(i ) = 0 + Y 96 1 + Y 97 2 + hatch2 3 +
(2)
(3)
hatch3 4 + hatch4 5 + House
+ Y 95 Farm
+
(i)
(i);1
P
(3)
(3)
(4) (4)
Y 96 Farm
+ Y 97 Farm
+ j 2P:flock(i) wi;j
j
(i);2
(i);3
(2)
(3)
House
N (0; 2(2) ); Farm
N3(0; (3) ); (4) N (0; 2(4) )
(i)
(i)
(1.3)
34
Table 1.10
Estimates for the parameters in model 1.3.
Parameter
MCMC Estimates
intercept (0 )
-2.544 (0.240)
1996 eect (1 )
-1.149 (0.256)
1997 eect (2 )
-1.003 (0.293)
Hatchery 2 eect (3 )
-1.788 (0.265)
Hatchery 3 eect (4 )
-0.143 (0.252)
Hatchery 4 eect (5 )
-1.065 (0.383)
Parent ock variance (2(4) )
0.878 (0.180)
Farm year95 variance ((3) [1; 1])
1.416 (0.341)
Farm 95/96 covariance ((3) [1; 2])
0.514 (0.262)
Farm 95/97 covariance ((3) [1; 3])
0.415 (0.226)
Farm year96 variance ((3) [2; 2])
1.239 (0.463)
Farm 96/97 covariance ((3) [2; 3])
0.750 (0.321)
Farm year97 variance ((3) [3; 3])
1.017 (0.482)
2
House variance ((2) )
0.271 (0.119)
The parameter estimates for this extended model are given in table 1.10. We see that the xed eects estimates are fairly similar to
model 1.2. It is interesting to see that all the covariances in the farm
level variance matrix are positive. This suggests that after adjusting for
other factors, if a farm has an incidence of salmonella in 1995 then it
is more likely to have an incidence again in 1996 and in 1997. In fact
the corresponding correlation estimates are 0.39, 0.35 and 0.67 respectively showing that in particular there is a strong correlation between
salmonella infection in farms in 1996 and 1997. The numerical instabilities of the quasi-likelihood methods mean that comparitive estimates
could not be calculated for this model.
5.
CONSEQUENCES OF IGNORING
NON-HIERARCHICAL STRUCTURES
Analysing only hierarchical components of populations which have
additional non-nested structures has two potentially negative consequences. Firstly, the model is under-specied because there are sources
of variation that have not been included in the model. This underspecication can lead to an underestimation of the standard errors of
35
Table 1.11
Eects of ignoring a cross-classied structure.
Parameter
intercept
VRQ eect
primary school variance
secondary school variance
pupil variance
Model I
Model II
Model III
5.97 (0.07) 6.02 (0.07) 5.98 (0.07)
0.16 (0.003) 0.16 (0.003) 0.16 (0.003)
0.28 (0.06)
0.27 (0.06)
0.05 (0.02) 0.01 (0.02)
4.25 (0.10) 4.48 (0.11) 4.25 (0.10)
the parameters and therefore to incorrect inferences. Secondly, the variance components obtained from the simple hierarchical model, or sets of
separate hierarchical models, can not be trusted. They may change substantially if the additional non-nested structures are included in a single
model. For example, we may wish to know about the relative importance
of general practices and hospitals on the variation in some patient level
outcome. If patients are cross-classied by hospital and general practice,
we need to t the full cross-classied model including patients, general
practices and hospitals in order to address this question. Looking at two
separate hierarchical analysis one of patients within hospital, the other
of patients within general practices, is not sucient.
A numerical example of this is shown in table 1.11 which shows results
for three models tted using the RG method to the educational attainment data from Fife in Scotland, where pupils are contained within a
cross-classication of primary schools by secondary schools. Model I ts
pupils within primary schools and ignores secondary school, model II ts
pupils within secondary schools and ignores primary school and model
III ts the cross-classication. The response is an attainment score at
age 16, the explanatory variable vrq is a verbal reasoning measure taken
at age 11. When one side of the cross-classication is ignored, the released variance is split between the classication left in the model and
the pupil level variance, inating both estimates. This has the most
drastic eect when the primary school hierarchy is ignored, in this case
(model II) the inated estimate of the between secondary school variance is 2.5 times its standard error as opposed to 0.5 times its standard
error in the full model.
References
Browne, W. J. and Draper, D. (2000). A comparison of Bayesian and
likelihood methods for tting multilevel models. Submitted.
Browne, W. J., Goldstein, H., and Rasbash, J. (2000). Fitting complex model structures to large datasets: a Monte Carlo Markov chain
(MCMC) algorithm to t multiple membership multiple classication
models. Submitted.
Bull, J. M., Riley, G. D., Rasbash, J., and Goldstein, H. (1999). Parallel
Implementation of a Multilevel Modelling Package. Computational
Statistics and Data Analysis, 31:457{474.
Clayton, D. G. and Rasbash, J. (1999). Estimation in large crossed
random-eects models by data augmentation. Journal of the Royal
Statistical Society, Series A, 162:425{436.
Ecochard, R. and Clayton, D. G. (1998). Multilevel modelling of conception in articial insemination by donor. Statistics in Medicine,
17:1137{1156.
Goldstein, H. (1986). Multilevel mixed linear model analysis using iterative generalised least squares. Biometrika, 73:43{56.
Goldstein, H. (1995). Multilevel Statistical Models. Edward Arnold, London, 2 edition.
Goldstein, H. and Rasbash, J. (1996). Improved Approximations for Multilevel Models with Binary Responses. Journal of the Royal Statistical
Society, Series A, 159:505{513.
Goldstein, H., Rasbash, J., Browne, W. J., Woodhouse, G., and Poulain,
M. (2000). Multilevel modelling in the study of dynamic household
structures. European Journal of Population, pages {.
Lee, Y. and Nelder, J. (2000). Hierarchical Generalized linear models:
a synthesis of generalized linear models, random eects models, and
structured dispersion. Technical report, Department of Mathematics,
Imperial College, London.
37
38
Pan, J. X. and Thompson, R. (2000). Generalized linear mixed models:
An improved estimating procedure. In Bethlehem, J. G. and van der
Heijden, P. G. M., editors, COMPSTAT: Proceedings in Computational Statistics, 2000., pages 373{378. Physica-Verlag.
Rasbash, J. and Browne, W. J. (2001). Non-hierarchical multilevel models. In Leyland, A. and Goldstein, H., editors, Multilevel modelling of
health statistics. Wiley.
Rasbash, J., Browne, W. J., Goldstein, H., Yang, M., Plewis, I., Healy,
M., Woodhouse, G., Draper, D., Langford, I., and Lewis, T. (2000). A
User's Guide to MLwiN. Institute of Education, London, 2.1 edition.
Rasbash, J. and Goldstein, H. (1994). Ecient analysis of mixed hierarchical and crossed random structures using a multilevel model.
Journal of Behavioural Statistics, 19:337{350.
Raudenbush, S. W. (1993). A crossed random eects model for unbalanced data with applications in cross-sectional and longitudinal research. Journal of Educational Statistics, 18:321{350.
Rodriguez, G. and Goldman, N. (1995). An Assessment of Estimation
Procedures for Multilevel Models with Binary Responses. Journal of
the Royal Statistical Society, Series A, 158:73{89.
Schafer, J. (1997). Analysis of Incomplete Multivariate Data. Chapman
& Hall, London.
Tanner, M. and Wong, W. (1987). The calculation of posterior distributions by data augmentation (with discussion). Journal of the American
Statistical Association, 82:528{550.