Вы находитесь на странице: 1из 16

Biostatistics (2000), 1, 2, pp.

141156
Printed in Great Britain

Generalized linear mixture models for handling


nonignorable dropouts in longitudinal studies
GARRETT M. FITZMAURICE AND NAN M. LAIRD
Department of Biostatistics, Harvard School of Public Health, 655 Huntington Avenue,
Boston, MA 02115, USA

fitzmaur@hsph.harvard.edu
S UMMARY
This paper presents a method for analysing longitudinal data when there are dropouts. In particular, we
develop a simple method based on generalized linear mixture models for handling nonignorable dropouts
for a variety of discrete and continuous outcomes. Statistical inference for the model parameters is based
on a generalized estimating equations (GEE) approach (Liang and Zeger, 1986). The proposed method
yields estimates of the model parameters that are valid when nonresponse is nonignorable under a variety
of assumptions concerning the dropout process. Furthermore, the proposed method can be implemented
using widely available statistical software. Finally, an example using data from a clinical trial of contracepting women is used to illustrate the methodology.
Keywords: Discrete data; Generalized estimating equations; Missing data; Nonresponse; Repeated measures.

1. I NTRODUCTION
The problem of missing or incomplete responses is ubiquitous in longitudinal research. Although most
longitudinal studies are designed to collect data on every individual in the sample at each time of followup, commonly not all responses are observed at all occasions. In general, an individuals response can
be missing at one follow-up time and then be measured at a later follow-up time, resulting in a large
class of distinct missingness patterns. Alternatively, longitudinal studies often suffer from the problem
of attrition or dropouts, i.e. some individuals drop out of the study before the intended completion
time and thus have incomplete responses. There are a variety of reasons for dropout, including lack of
efficacy and removal from the study due to adverse events. Note, however, that the decision to discontinue
is commonly made by either the individual, her physician, or the investigator. In this paper, we restrict
attention to the monotone missing data patterns that result from dropouts.
Consider the following example from a longitudinal clinical trial of contracepting women reported by
Machin et al. (1988). In this trial women received an injection of either 100 mg or 150 mg of depotmedroxyprogesterone acetate (DMPA) on the day of randomization and three additional injections at 90day intervals. Throughout the study each woman completed a menstrual diary that recorded any vaginal
bleeding pattern disturbances. The diary data were used to generate a binary sequence for each woman
according to whether or not she had experienced amenorrhea in the four successive intervals. However,
a feature of this clinical trial that complicated the analyses is that there were a substantial number of
dropouts for reasons that were thought likely to be related to the outcome. More than one third of the
women dropped out before the completion of the trial; 17% dropped out after receiving only one injection
of DMPA, 13% dropped out after receiving only two injections, and 7% dropped out after receiving three
c Oxford University Press (2000)


142

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

injections. The primary focus of this paper is on methods for analysing longitudinal data when there are
dropouts for reasons that are thought to be related to the unobserved outcome.
Many of the recently developed methods for analysing longitudinal data that have been implemented
in commercially available statistical software packages focus on marginal models for describing the mean
response and its time trend, and assume that dropouts are ignorable. That is, it is assumed that the probability of dropout does not depend upon the unobserved responses (Rubin, 1976; Little and Rubin, 1987).
In this paper, we develop methods for handling dropouts in longitudinal studies where the dropouts are
assumed to be nonignorable. That is, the probability of dropout is assumed to depend on the unobserved
responses. We consider a class of mixture models that are closely related to models first proposed by
Little (1993, 1994) and develop generalized linear mixture models for handling nonignorable dropout for
a variety of discrete and continuous longitudinal outcomes. This work is a natural extension of the models
for repeated continuous responses developed in Hogan and Laird (1997) to the case where the response
variable can be discrete, ordinal, continuous or count data. An attractive feature of the proposed method is
that it can be implemented, with only minor modification, using existing statistical software for analysing
repeated responses (e.g. using PROC GENMOD in SAS, 1999).

2. N ONIGNORABLE DROPOUT
With nonignorable dropouts the probability of nonresponse is assumed to depend on the unobserved
responses. In general, models for nonignorable dropout are fundamentally nonidentifiable unless some
arbitrary model constraints are imposed. That is, inference is possible only once some modeling assumptions have been made. However, it is worth stressing that these assumptions are not verifiable from the
data at hand. Therefore, it is an inescapable fact that all methods for handling nonignorable dropout have
to make some unverifiable assumptions. In this section we review two broad classes of models that have
been proposed for handling nonignorable dropout.
In order to correct for the bias that can arise when there is nonignorable dropout in longitudinal studies,
joint models for the multivariate outcomes and dropout indicators have been proposed. Little (1993, 1995)
identified two broad classes of joint models: selection models and mixture models. In selection models,
one uses a complete data model for the longitudinal outcomes, and then the probability of dropout is
modeled conditional on the possibly unobserved outcomes. To ensure identifiability, the set of outcomes
is usually restricted in some way and arbitrary constraints on the dropout model (e.g. setting certain
interactions to zero) usually have to be imposed. One of the primary attractions of selection models is
that they directly model the marginal distribution of the longitudinal outcomes. A drawback of selection
models is that they are computationally intractable in many cases and cannot be implemented using existing statistical software. In contrast, with mixture models, one uses a model for the distribution of the
longitudinal outcomes given the dropout patterns and then a model for the marginal distribution of the
dropout patterns. With mixture models it is immediately clear that the distribution of the outcomes given
patterns of nonresponse is not completely identifiable, since for some patterns, certain outcomes are not
observed. Hence restrictions must be built into this model (Little, 1993, 1994). The main focus of this
paper is on mixture models for handling nonignorable dropout in longitudinal studies; the term mixture
model is used in this context to emphasize that the distribution of the longitudinal outcomes depends upon
dropout time. However, it should be kept in mind that the primary focus of statistical inference is on the
marginal distribution of the outcomes. That is, the main scientific goal is to make inferences concerning
the marginal expectation of the longitudinal outcomes. Within the mixture model approach, the latter is
attained by averaging over the distribution of the dropout times.
In much of the recent methodological work on mixture models for nonignorable nonresponse in longitudinal studies two main avenues of research can be distinguished. The first is the mixture modeling

Generalized linear mixture models

143

approaches that have focussed on linear mixed effects models for the repeated responses using dropout
time as a covariate. For example, Wu and Bailey (1989) have described a conditional linear model,
given dropout time, which can be fit to the individuals least squares slopes, and then averaged over the
distribution of the dropout times. Recently, Follmann and Wu (1995) have extended the conditional
linear model to generalized linear models. Extensions to permit censored dropout times have also been
considered by Hogan and Laird (1997), who use a full ML approach, implemented via the EM algorithm,
to estimate the model parameters. Finally, there is related work by Pawitan and Self (1993) who consider
a fully parametric likelihood-based approach for modeling a longitudinal response conditional on a timeto-event outcome.
The second main avenue of research has focussed on pattern-mixture models. Little (1993, 1994) and
Ekholm and Skinner (1998) have considered pattern-mixture models which stratify the incomplete data by
the pattern of missing values and formulate distinct models within each stratum. In pattern-mixture models
additional assumptions about the missing data mechanism have to be made in order to yield supplemental
restrictions that ensure the models are identifiable. For example, consider the simple case of bivariate
monotone missing data discussed in Little (1994). That is, let Yi = (Yi1 , Yi2 ) be a bivariate continuous
response, where Yi2 may be only partially observed (e.g. due to dropout prior to the second measurement
occasion), and assume that there are no covariates. Little (1994) assumes that, given the two distinct
missing data patterns (i.e. completers and noncompleters), Yi has a bivariate Gaussian distribution. Note
that this model implies that marginally Yi is a mixture of two bivariate Gaussian distributions. Clearly, for
the completers, the five model parameters (two means and three variances and covariances) are identifiable
from the data at hand. However, it is equally clear that only two model parameters are identifiable from
the data for the other missing data pattern. The remaining three model parameters are identifiable only by
making supplemental restrictions (or by having prior information). In order to ensure that all of the model
parameters are identifiable, Little (1994) assumes that missingness in Yi2 depends on an arbitrary function
of (Yi1 + Yi2 ), for some known . Note that, in effect, Little (1994) relies on a selection model for
dropout to motivate this set of identification restrictions for the pattern-mixture model. This assumption is
sufficient to ensure that all of the model parameters are identifiable. However, will not usually be known
and the data provides no information about . Thus, Little (1994) recommends conducting a sensitivity
analysis for a range of plausible values of . The mixture model methods proposed in Section 3 of this
paper are more closely related to the models considered by Wu and Bailey (1989) and others than they are
to the pattern-mixture models proposed by Little (1994).

3. A MIXTURE MODEL FOR NONIGNORABLE DROPOUT


In this section we develop a mixture model for nonignorable nonresponse in longitudinal studies which
can be considered a generalization of the Wu and Bailey (1988, 1989) approach. The proposed model is
a standard marginal regression model except that the parameters which characterize change over time
depend also upon time of dropout. That is, a suitable linear or nonlinear function of the mean response
is assumed to be linear in time and dropout time, and their interactions. However, before describing the
model in greater detail, we need to establish some notation and describe the data typically collected in
longitudinal settings.
We assume that N individuals are to be observed at the same set of m occasions, {t1 , t2 , . . . , tm }.
Then, let Yi j denote the outcome variable for the ith individual at occasion t j . The outcome Yi j can be
either discrete, ordinal, continuous or count data and we can form the m 1 complete response vector,
Yic = (Yi1 , . . . , Yim ) . In addition, each individual has a set of covariates, and in the most general case,
each repeated outcome may have a p 1 vector of covariates, say X i j , measured at t j , j = 1, . . . , m. We
let X i = (X i1 , . . . , X im ) denote the matrix of covariates for Yic .

144

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

Subjects are followed for occurrence of nonignorable study dropout, i.e. dropout that is assumed to be
related to the unobserved responses measured by the Yi j s. Also, we implicitly assume that dropout implies
that no subsequent repeated measures are made. Thus, each individual has an event time Di , denoting
nonignorable dropout time, which is thought to be related to Yic . If Di tm , the ith individual is a dropout;
otherwise, the ith individual is a completer. Thus, all observed dropouts fall in the interval t1 < Di tm ,
but we also assume a subset of subjects are completers, i.e. time of nonignorable nonresponse is not
relevant for Di > tm , but the proportion with Di > tm is. We assume that t1 is the start of the study,
e.g. randomization, and thus subjects cannot drop out prior to t1 and Y1 is observed for all subjects. Note
also that Di is often discrete, since it occurs in the interval after t j but before t j+1 for some j. In such cases
we simply consider Di to be discrete and occurring at t j+1 (i.e. the response at t j+1 is not observed). In
this setting, {t1 , t2 , . . . , tm } are the set of ordered dropout times and we let j = pr(Di = t j ). It is assumed
that i1 = 0 and we include an additional category i,m+1 for the study completers. The distribution of
the dropout time, Di , must be correctly specified and thus we allow Di to depend on covariates, typically a
subset of X i . Finally, we note that any intermediate missingness in outcome can be easily accommodated
provided that it is missing completely at random (MCAR) given dropout time.
Letting Yi denote the m i 1 vector of the responses observed on the ith individual (i.e. the observed
portion of Yic ), the observed data for each subject consists of (Yi , Di , X i ). Note that the covariates in X i
will generally include time (ti j ). Thus, in a randomized clinical trial, X i may include time (as ti j or as
dummy indicators) and treatment, while the distribution of Di may depend only on treatment. In principle
we could allow time-varying covariates for Di . In the example from the clinical trial of contracepting
women, Yi is a vector of observed indicators of whether a woman experienced amenorrhea during each
of the four successive intervals, Di is the dropout time, X i is a matrix of covariates that includes time (or
indicators of time trends) and a treatment indicator for whether the woman received 100 mg or 150 mg of
DPMA, and Di simply depends on the treatment indicator.
As discussed in Section 2, joint models for the longitudinal outcomes and the response indicators
generally fall into two categories: selection and mixture models. It is easiest to represent these two
models in terms of the complete data, Yic = (Yi1 , . . . , Yim ) and Di . For the selection model the joint
distribution is factored as follows,
f (Yic , Di |X i ) = f Y (Yic |X i ) f DY (Di |Yic , X i ).
Recall that in longitudinal studies the main focus is on inferences about f Y (|), and in selection models
f DY (|) plays the role of nuisance parameters, which can be ignored only if f (Di |Yic , X i ) does not
depend upon any missing Yi j s. In contrast, for the mixture model the joint distribution is factored as
follows,
f (Yic , Di |X i ) = f D (Di |X i ) f Y D (Yic |Di , X i ).
Note that inferences about f Y D (|) are not of main interest. Rather, inferences about the marginal distribution of the outcomes, f Y (|), obtained by averaging over the distribution of the dropout times, are of
primary interest. In particular, the main substantive interest is in the marginal means of the repeated outcomes (and their dependence on covariates) averaged over the distribution of the dropout times. Although
we have described the mixture model in terms of the factorization above, note that the methods described
later avoid making full distributional assumptions about f Y D (Yic |Di , X i ).
In order to clarify the main ideas, consider the following simple example where we have two treatment
groups (say, X i = 0 and X i = 1) and it is of interest to compare the marginal means in the two treatment
groups. That is, we are interested in comparing 0 j = E(Yi j |X i = 0) with 1 j = E(Yi j |X i = 1) (for
j = 1, . . . , m). Note that in any model for the conditional mean of Yi j given the dropout time and X i , the
regression parameters are not directly of interest. This is because the parameters in a model for the conditional means, say 0 j = E(Yi j |Di , X i = 0) and 1 j = E(Yi j |Di , X i = 1), have a somewhat unappealing

Generalized linear mixture models

145

interpretation due to the stratification or dependence on patterns of dropout which may themselves depend
upon the outcome. Rather, the quantity of primary interest concerns the marginal mean of the repeated
outcomes, averaged over the distribution of the dropout times. That is, we are primarily interested in
making inferences concerning the comparison of
E(Yi j |X i = 0) = 0 j =

m+1


l 0 j ,

l=2

with
E(Yi j |X i = 1) = 1 j =

m+1


l 1 j ,

l=2

where l can also depend on X i .


Thus, we propose to consider models for Yi , conditional on the time of dropout, that are of the following
general form:
g(E[Yi j |Di , X i j ]) = Z i j ,

(1)

where g() is a known link function and the design vector Z i j depends on the dropout time, the covariates
X i j , and also incorporates their interactions. Thus, the conditional mean of Yi j might depend on the time
of dropout, Di , and any other covariates of interest (e.g. treatment group, time), and their interactions.
With a continuous response, the identity link function is a natural choice; with a binary response, the logit
link function is a natural choice. However, in principle, any suitable link function g() can be chosen.
Recall that the parameter of interest is not . Rather, the target of primary interest concerns the marginal
expectation of the repeated outcomes, averaged over the distribution of the dropout times,
E(Yi j |X i j ) = i j =

m+1


l g 1 (Z i j ),

(2)

l=2

where Z i j depends on the dropout patterns and X i j , and l depends on X i (or some subset of X i ). Furthermore, note that because the final estimates will have been averaged over the distribution of the dropout
times, f (Di ), the marginal means will not, in general, follow the link function model that has been assumed in (1). For this reason, we recommend saturating Z i j in treatment effects (and in any other covariate
effects of primary interest). Treatment comparisons will be made directly in terms of the marginal means.
Finally, it is important to note that in (1) may be nonidentifiable unless some suitable structure is
specified for the conditional mean of each response. For example, consider the case where m = 3 and
there is dropout at the second and third planned measurement occasions. This gives rise to three distinct
dropout patterns: the study completers (say Di = 4), those individuals who drop out at the third occasion
(say Di = 3), and those individuals who drop out at the second occasion (say Di = 2). For simplicity, we
assume that there are no covariates. A model with separate intercepts and slopes for each of the distinct
dropout patterns is clearly not identifiable. That is, the model
(l)

(l)

g(E[Yi j |Di ]) = 0 + 1 t j

(l = 2, 3, 4),

is not identifiable since the distinct slopes are not estimable for one of the incomplete data patterns.
Estimation of the slope requires at least two repeated outcomes, but those individuals who drop out at the
second occasion (Di = 2) have only a single outcome. On the other hand, the following model
(l)

g(E[Yi j |Di ]) = 0 + 1 t j

(l = 2, 3, 4),

146

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

which has separate intercepts for each of the distinct dropout patterns, but has a common slope, is identifiable. Also, if time is treated as a qualitative covariate or factor, a model with separate intercepts and a
common time factor (indexed by k),
(l)

g(E[Yi j |Di ]) = 0 + k

(l = 2, 3, 4),

is identifiable. Alternatively, a model with intercepts and slopes that depend linearly on the dropout times,
g(E[Yi j |Di ]) = 0 + 1 t j + 2 Di + 3 (t j Di ),
is also identifiable.
4. E STIMATION
In the preceding, the only assumption that has been made concerns the mean of the distribution of
(Yi j |Di ), that is, the conditional mean of each response separately. However, in order to account for the
time-dependence or covariance among the repeated outcomes some additional assumptions about the joint
distribution of the m responses must be made. When the response variable is continuous and assumed to
be approximately Gaussian, there is a general class of linear models that are suitable for analyses. Ware
(1985) provides a comprehensive description of these models for the ignorable case, while Hogan and
Laird (1997) have developed an extension of this general class of linear models to handle nonignorable
dropouts. However, when the response variable is categorical, ordinal or count data, fewer techniques are
available. This is due in part to the lack of a discrete multivariate analogue of the multivariate Gaussian for
the joint distribution of the responses. Also, unlike the multivariate Gaussian distribution for continuous
responses, the joint distribution of a categorical response variable cannot, in general, be represented by
the first two moments of Yi alone. For example, with a discrete response having C response categories,
the fully parameterized distribution has C m 1 nonredundant parameters. However, if the main interest
is in , and the time dependence can be considered a nuisance characteristic of the data, valid estimates
of can be obtained without having to completely specify the joint distribution of Yi .
In companion papers, Liang and Zeger (1986) and Zeger and Liang (1986) introduced a general method
for incorporating within-subject correlation in generalized linear models. In their proposed generalized estimating equations (GEE), the within-subject correlation among the repeated outcomes can be accounted
for by introducing a working covariance matrix. That is, generalized estimating equations only require
specification of the form of the first two moments, the mean and covariance of the vector of responses.
Furthermore, valid estimates of are obtained regardless of the choice of working covariance matrix.
Indeed, a working independence assumption may actually be preferred when there are missing data and
the sample size is relatively small (say N < 50) (H. Lee, N. M. Laird and G. Johnston, unpublished
manuscript). Note that while the GEE approach leaves the joint distribution of Yi completely unspecified, it does yield valid estimates of with often only modest losses of efficiency for many standard
longitudinal designs (Fitzmaurice et al., 1993; Diggle et al., 1994). Note also that the GEE approach can
easily accommodate any intermediate missingness in outcome provided that it is missing completely at
random (MCAR) given dropout time. In fact, it can accommodate each subject having a distinct set of
measurement times.
The generalized estimating equations for are given by
u () =

N


G i Vi1 [Yi E(Yi |Di ; )] = 0,

(3)

i=1

where G i = E(Yi |Di ; )/, and Vi = Vi (, ) is the m i m i working covariance matrix of Yi


(Liang and Zeger, 1986). The elements of Vi are specified in terms of known functions of the marginal

Generalized linear mixture models

147

means, E(Yi j |Di ; ), in addition to a set of association parameters, say (e.g. the pairwise correlations
or odds ratios). Note that Vi depends upon Di only through the marginal means, E(Yi j |Di ; ). However, in principle, this restriction could easily be relaxed for simple models for the correlations, e.g. by
allowing separate exchangeable correlations for the completers and dropouts. Typically is unknown,
but can be estimated using method of moments or with a set of estimating equations similar to (3). Then,
 ) has an asympusing Taylor series expansions similar to Prentice (1988), it can be shown that N 1/2 (
totic distribution which is multivariate Gaussian with mean vector 0. An expression for the asymptotic
 ) is given by
covariance matrix of N 1/2 (
1 

1

N
N
N



(4)
V = lim N
G i Vi1 G i
G i Vi1 cov(Yi )Vi1 G i
G i Vi1 G i
.
N

i=1

i=1

i=1

Consistent estimates of the asymptotic covariance of the estimated regression parameters can be obtained
 first suggested by Cox (1961), and later proposed by Huber (1967),
using the empirical estimator of cov()
White (1982) and Royall (1986).
Next, we consider estimation of the dropout probabilities. With a small number of discrete covariates,
the multinomial probabilities for dropout, l , can simply be estimated as the sample proportion with
each dropout time, stratified by the covariate patterns (e.g. stratified by treatment group and, perhaps, by
other relevant covariates). Letting denote the vector of probabilities, and the corresponding vector of
sample proportions, it is easily shown that the asymptotic covariance matrix of N 1/2 ( ) is given by
V = diag( )  .

(5)

Note that when the sample proportions are obtained from strata defined by treatment group, or by other
relevant covariates, the number of individuals within each of the strata, and not N , is the appropriate
denominator for the sample proportions (and the corresponding covariance matrix). However, when either
the number of dropout times or the number of covariates is relatively large, fully nonparametric estimation
of may not be feasible in moderate size samples. In that case, parametric models (e.g. multinomial loglinear regression or discrete survival regression) for the can be adopted.
Thus, the model parameters in (1) can be estimated using widely available statistical software for
GEE (e.g. PROC GENMOD in SAS) by simply including as covariates the dropout times (or indicator
variables for the distinct dropout times) and the interactions of dropout times with other covariates. This
provides a valid estimate of under a set of reasonable models for the dropout process, without requiring
the complete specification of the entire joint distribution of the repeated outcomes. Furthermore, the
multinomial probabilities for dropout, , can simply be estimated as the sample proportion with each
dropout time (stratified by treatment group and, perhaps, by other relevant covariates).
Then, various quantities of interest, such as the marginal mean at occasion t j , can be estimated by

i j =

m+1




l g 1 (Z i j ).

l=2

Standard errors for 


i j can be obtained using the -method (Rao, 1965). In general, for any function

h(, ) the asymptotic covariance matrix of N 1/2 [h(,
) h(, )] can be approximated by V  ,
where  is the Jacobian evaluated at (, ), and


V 0
V =
.
0 V

Alternatively, resampling methods, such as the bootstrap or the jackknife, can be used (Efron, 1982).

148

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD


5. A PPLICATION : CLINICAL TRIAL OF CONTRACEPTING WOMEN

In this section we consider an application of the proposed methodology for handling dropout in the
study of contracepting women introduced in Section 1. Recall that this study was a randomized clinical
trial of the efficacy of two doses of a contraceptive, 100 mg or 150 mg of depot-medroxyprogesterone
acetate (DMPA), given at 90-day intervals. Women participating in the study received an injection of
either 100 mg or 150 mg of DMPA on the day of randomization and three additional injections at 90day intervals. There was a final follow-up visit 90 days after the fourth injection, i.e. 1 year after the
first injection. Throughout the study each woman completed a menstrual diary that recorded any vaginal
bleeding pattern disturbances. The outcome of interest is a binary response indicating whether or not a
woman experienced amenorrhea, the absence of menstrual bleeding for a specified number of days. Note
that in clinical trials of modern hormonal contraceptives, pregnancy is exceedingly rare (and would be
regarded as a failure of the contraceptive method), and is not the main outcome of interest in this study
(Machin et al., 1988).
A total of 1151 women completed the menstrual diaries and the diary data were used to generate a
binary sequence for each woman according to whether or not she had experienced amenorrhea in the
four successive intervals. However, a feature of this clinical trial is that there was substantial dropout
for reasons that were thought likely to be related to the outcome. More than one third of the women
dropped out before the completion of the trial; 17% dropped out after receiving only one injection of
DMPA, 13% dropped out after receiving only two injections, and 7% dropped out after receiving three
injections. When the dropout rates are broken down by dose group, the rates were marginally higher in
the higher dose group. For those women randomized to 100 mg of DMPA; 37% dropped out before the
completion of the trial; 17% dropped out after receiving only one injection of DMPA, 12% dropped out
after receiving only two injections, and 8% dropped out after receiving three injections. For those women
randomized to 150 mg of DMPA; 39% dropped out before the completion of the trial; 17% dropped out
after receiving only one injection of DMPA, 15% dropped out after receiving only two injections, and
6% dropped out after receiving three injections. For women who dropped out before the end of the 90day injection interval, a determination of whether or not they experienced amenorrhea was made, on a
proportionate basis, using their existing menstrual diary data for that interval.
Letting Yi j = 1 if the ith women experienced amenorrhea in the jth injection interval, we first explored the usual model, without conditioning on dropout patterns, under the assumption that dropout is
completely at random. That is, we considered the following logistic regression model for the marginal
probabilities (hereafter referred to as model 1)


pr[Yi j = 1]
log
= 0 + 1 time + 2 time2 + 3 dose + 4 time dose + 5 time2 dose;
pr[Yi j = 0]
where time = 0, 1, 2, 3 for the four consecutive 90-day injection intervals, and dose = 1 if randomized
to 150mg of DMPA, and dose = 0 otherwise. If dropout could be assumed to be completely at random,
then estimates of the marginal regression parameters can be obtained using a standard GEE approach.
The GEE parameter estimates, assuming an unstructured working correlation matrix, and their empirical
standard errors are displayed at the top of Table 1 (we note that the empirical and model-based standard
errors were very similar in all of the reported analyses, as might be expected given the unstructured
working correlation assumption). These results suggest that the rates of amenorrhea in the second and
third injection intervals are significantly higher for those women who received the higher dose of DMPA,
although these differences tend to decline over time (see Table 2). For example, during the third injection
interval (69 months post-randomization) the predicted rates of amenorrhea are 0.497 in the 150 mg dose
group and 0.388 in the 100 mg dose group. However, by the final follow-up visit there is no longer a
discernible treatment difference, with predicted rates of amenorrhea of 0.569 in the 150mg dose group

Generalized linear mixture models

149

Fig. 1. Plot of estimated probability of amenorrhea from model 2.

and 0.517 in the 100mg dose group. Note, however, that if dropout is nonignorable the standard GEE
analysis of the available data can yield biased estimates of the effects of treatment.
Next, we considered a mixture modelling approach where the conditional probabilities of amenorrhea,
given dropout patterns, are related to the covariates by a logit link function,


pr[Yi j = 1|Di ]
log
= 0 + 1 Di1 + 2 Di2 + 3 Di3 + 4 time + 5 time2
pr[Yi j = 0|Di ]
+6 dose + 7 time dose + 8 time2 dose;
where Di j = 1 if the ith woman dropped out after the jth injection, and Di j = 0 otherwise (i.e.
Di1 = Di2 = Di3 = 0 for the completers). For this simple model (hereafter referred to as model
2), the effect of dropout is to raise or lower the overall level of response, but not the shape of the response curve. Estimates of the regression parameters were obtained using the standard GEE approach,
with an unstructured working correlation matrix, and are displayed in the middle of Table 1. Finally,
E(Di j |dose) = pr(Di j = 1|dose) was allowed to depend on dose and was estimated nonparametrically
as the sample proportion with each dropout time, stratified by dose group. The estimates of 1 , 2 , and
3 indicate that the overall rates of amenorrhea are higher for women who drop out than for women who
complete the study. Furthermore, the rates of amenorrhea are higher for women who drop out earlier in
the study (see Figure 1). Given estimates of (, ), the marginal means (averaged over the distribution
of the dropout times) in the two dose groups at each occasion were estimated using (2) and are displayed
in the middle of Table 2. When compared to the predicted rates from model 1, the overall results for
the effects of treatment are not discernibly different except that the underlying rates of amenorrhea are
higher. This suggests that those women who drop out have slightly higher rates of amenorrhea. However,
the treatment comparisons are relatively unaffected by this dropout process. This result might have been
anticipated since the fitted model has not allowed the shape of the response curve to vary across dropout
patterns (see Figure 1). A comparison of the estimated means from model 2 indicate that during the second and third injection intervals the rates of amenorrhea are higher in the 150 mg dose group. However,
by the end of the study there is no discernible treatment difference in the rates of amenorrhea.
As noted earlier, models for nonignorable dropout are, in general, nonidentifiable unless model constraints are imposed. However, since these assumptions cannot be verified from the available data, we
recommend examining the sensitivity of inference to model specifications. To assess the sensitivity of
results to model specification we also considered a substantially more elaborate model for the conditional

150

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD


Table 1. Parameter estimates and standard errors under three different modelling assumptions about dropout
Model

Parameter

Estimate

S.E.

Model 1

Intercept

1.4904

0.1069

13.95

time

0.5120

0.1319

3.88

time2

0.0023

0.0400

0.06

dose

0.1091

0.1485

0.74

time dose

0.4341

0.1883

2.30

time2 dose

0.1337

0.0570

2.35

Intercept

1.7182

0.1252

13.72

Di1

0.6865

0.1857

3.70

Di2

0.4520

0.1583

2.86

Di3

0.3845

0.1847

2.08

time

0.6080

0.1393

4.37

0.0132

0.0414

0.32

dose

0.1006

0.1500

0.67

time dose

0.4496

0.1918

2.34

time2 dose

0.1374

0.0578

2.38

Intercept

1.5416

0.1367

11.28

0.3472

0.2450

1.42

Model 2

time2

Model 3

Di1
Di2

0.0744

0.2599

0.29

Di3

0.1887

0.3120

0.60

time

0.4709

0.1537

3.06

time2

0.0153

0.0447

0.34

0.1543

0.1984

0.78

0.6094

0.2277

2.68

time2 dose

0.1694

0.0656

2.58

Di time
Di time2
Di dose
Di time dose
Di time2 dose

0.0882

0.4501

0.20

0.2506

0.2198

1.14

0.5624

0.3032

1.86

0.1147

0.5985

0.19

0.0522

0.3049

0.17

dose
time dose

Generalized linear mixture models

151

Table 2. Marginal rates of amenorrhea under three different modelling assumptions


about dropout
Model

Time

Model 1

Model 2

Model 3

100mg

150mg

Difference

90 days

0.184

0.201

0.017

180 days

0.274

0.363

270 days

0.388

360 days

S.E.

0.023

0.73

0.4629

0.089

0.025

3.54

0.0004

0.497

0.109

0.030

3.65

0.0003

0.517

0.569

0.052

0.036

1.46

0.1437

90 days

0.184

0.200

0.016

0.023

0.72

0.4734

180 days

0.288

0.379

0.091

0.026

3.57

0.0004

270 days

0.414

0.526

0.112

0.030

3.71

0.0002

360 days

0.546

0.599

0.053

0.035

1.52

0.1275

90 days

0.185

0.201

0.016

0.023

0.70

0.4864

180 days

0.282

0.382

0.100

0.029

3.39

0.0007

270 days

0.457

0.562

0.105

0.041

2.55

0.0108

360 days

0.648

0.673

0.025

0.056

0.44

0.6614

probabilities of amenorrhea, given dropout patterns,




pr[Yi j = 1|Di ]
log
= 0 + 1 Di1 + 2 Di2 + 3 Di3 + 4 time + 5 time2 + 6 dose
pr[Yi j = 0|Di ]
+7 time dose + 8 time2 dose + 9 Di time + 10 Di time2
+11 Di dose + 12 Di time dose + 13 Di time2 dose;

152

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

Fig. 2. Plot of estimated probability of amenorrhea from model 3.

where Di = 1 if the ith woman dropped out at any occasion, and Di = 0 otherwise (i.e. Di = 0 for the
completers). In this model (hereafter referred to as model 3), the effect of dropout changes both the overall
level of response and the shape of the response curves for dropouts and completers, separately for each
treatment group (see Figure 2). Estimates of the regression parameters were obtained using the standard
GEE approach, with an unstructured working correlation matrix, and are displayed at the bottom of Table
1. We note that in an analysis similar to the above, Machin et al. (1988) also allowed the dose-response
trend to vary as a function of dropout.
Finally, the marginal means (averaged over the distribution of the dropout times) in the two dose
groups at each occasion were estimated using (2) and are displayed at the bottom of Table 2. Although
the estimated rates of amenorrhea are somewhat higher in the third and fourth injection intervals, overall
the predicted rates are very similar to those obtained from model 2 and the main substantive conclusions
have not changed. In Figure 3 the predicted rates of amenorrhea in the two treatment groups are compared
to those obtained from model 1. The mean response profiles in Figure 3 indicate that the underlying rates
of amenorrhea are somewhat higher than those suggested by model 1, but that the treatment comparisons
are relatively unaffected by this nonignorable dropout process.
Finally, we note that a test of H0 : 1 = 2 = 3 = 0 in model 2, or of H0 : 1 = 2 = 3 = 9 =
10 = 11 = 12 = 13 = 0 in model 3, provides a simple test of the hypothesis that dropouts are
completely at random (Little, 1988; Park and Davis, 1993; Park and Lee, 1997). For both models 2 and 3,
the null hypothesis that dropouts are completely at random is overwhelmingly rejected, with X 2 = 20.02,
3 df ( p < 0.0002) and X 2 = 29.68, 8 df ( p < 0.0003) respectively. With models of the form given by
(1), a missing at random (MAR) dropout mechanism cannot be expressed in terms of restrictions on the
conditional distribution of each response, Yi j , given the time of dropout. However, Molenberghs et al.
(1998) have shown that MAR dropout corresponds to a set of restrictions on the conditional distributions
of the responses, given previous responses (and time of dropout).
6. C ONCLUSIONS
In this paper we have considered methods for analysing longitudinal data when there are dropouts. In
particular, we have described a simple method based on generalized linear mixture models for handling
nonignorable dropouts. The proposed model can be considered a generalization of the Wu and Bailey
(1988, 1989) approach and is also a natural extension of the model for repeated continuous responses

153

0.5
0.4
0.3

Model 1: 100mg
Model 1: 150mg
Model 3: 100mg
Model 3: 150mg

0.1

0.2

Probability of Amenorrhea

0.6

0.7

Generalized linear mixture models

10

12

Time in Months

Fig. 3. Plot of estimated probability of amenorrhea.

developed in Hogan and Laird (1997) to the case where the response variable can be discrete, ordinal,
continuous or count data. An attractive feature of the proposed method is that it is suitable for analysing a
variety of discrete and continuous outcomes and does not require specification of the joint distribution of
the longitudinal response vector. The main assumption that is made concerns the mean of the distribution
of (Yi j |Di ), that is, the conditional mean of each response separately. Statistical inference is based on a
generalized estimating equations approach (Liang and Zeger, 1986) and can be implemented using widely
available statistical software for GEE (e.g. Stata (StataCorp., 1997), SUDAAN, or PROC GENMOD in
SAS).
As discussed in the earlier sections, there are two broad classes of models for dropouts that have been
proposed in the statistical literature: selection models and mixture models. Selection models specify
complete data models for the longitudinal outcomes and model the probability of dropout conditional
on the longitudinal outcomes. Two recently developed approaches for handling dropout within a GEE
framework are the imputation method of Paik (1997) and the weighted estimating equations approach
of Robins et al. (1995). Both of these approaches specify complete data models for the longitudinal
outcomes, but differ in how they model the dropout times. In the imputation method, the missing data
on the longitudinal outcomes are imputed or filled-in based on an assumed model for dropout given
the observed data. In the weighted estimating equations approach, contributions to the standard GEE are
weighted inversely by the probability of dropout. In the latter, the weights are estimated based on an
assumed model for dropout. Note, however, that both of these approaches provide valid estimates only
when dropout is ignorable (i.e. MCAR or MAR), but not otherwise. When dropout is nonignorable, the
focus of this paper, the estimators proposed by Paik (1997) and Robins et al. (1995) will, in general, be

154

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

biased.
Selection and mixture models for nonignorable dropouts each have their own advantages and disadvantages. With selection models, the dropout process is modeled conditional on the repeated outcomes,
making it very easy to formulate hypotheses about the dropout process (see, for example, Diggle and
Kenward, 1994; Molenberghs et al., 1997). Thus, selection models may seem more intuitive to most
investigators. However, establishing whether such models are identifiable in considerably more difficult.
To ensure that selection models are identifiable, arbitrary constraints on the dropout model (e.g. setting
certain interactions to zero) need to be imposed and the set of outcomes is usually restricted in some way.
A drawback of selection models is that it is unclear how these restrictions on the dropout process translate
into assumptions about the distribution of the unobserved outcomes. In summary, selection models are a
very natural way to model the dropout process, but suffer from lack of identifiability, model sensitivity,
and are computationally intractable in many cases. In contrast, with mixture models one uses a model for
the conditional distribution of the repeated outcomes given the dropout patterns. In this case it is patently
clear that this conditional distribution is nonidentifiable since various subsets of the outcomes are not
observed for the different dropout patterns. Therefore, restrictions have to be built into the mixture models.
However, unlike selection models, mixture models make explicit assumptions about the distribution of the
unobserved outcomes (see, for example, Figures 1 and 2). Thus, with mixture models, it is relatively easy
to explore the sensitivity of results to model specification. Finally, mixture models are easy to implement
and estimation is relatively straightforward.
The main drawback of mixture models is that the natural parameters of interest are not immediately
available; they require marginalization of the distribution of outcome over dropout time. Furthermore, in
general, it is not possible to parsimoniously describe the effects of covariates on the marginal distribution
of the outcome in terms of regression coefficients. This may not be so problematic in a simple randomized
clinical trial where the main parameter of interest is often a simple difference in treatment means (or
change from baseline) at the completion of the trial. However, in many other applications where there are
a number of quantitive covariates of interest, or where there is a set of potential confounding variables
that need to be adjusted for in the analysis, a simple summary of the mixture model results is not readily
available.
ACKNOWLEDGEMENTS
The authors thank the Associate Editor and the referees for their helpful comments and suggestions.
This research was supported by grants GM 29745 and MH 17119 from the National Institutes of Health.
R EFERENCES
C OX , D. R. (1961). Tests of separate families of hypotheses. In Proceedings of the Fourth Berkeley Symposium on
Mathematics, Probability and Statistics, Vol. 1, pp. 105123. Berkeley: University of California Press.
D IGGLE , P. J. AND K ENWARD , M. G. (1994). Informative dropout in longitudinal data analysis (with discussion).
Applied Statistics 43, 4993.
D IGGLE , P. J., L IANG , K. Y. AND Z EGER , S. L. (1994). Analysis of Longitudinal Data. New York: Oxford University Press.
E FRON , B. (1982). The Jackknife, the Bootstrap, and Other Resampling Plans. Philadelphia: Society for Industrial
and Applied Mathematics.
E KHOLM , A. AND S KINNER , C. (1998). The Muscatine childrens obesity data reanalysed using pattern-mixture
models. Applied Statistics 47, 251263.

Generalized linear mixture models

155

F ITZMAURICE , G. M., L AIRD , N. M., AND ROTNITZKY, A. G. (1993). Regression models for discrete longitudinal
responses (with discussion). Statistical Science 8, 248309.
F OLLMANN , D. AND W U , M. (1995). An approximate generalized linear model with random effects for informative
missing data. Biometrics 51, 151168.
H OGAN , J. W. AND L AIRD , N. M. (1997). Mixture models for the joint distribution of repeated measures and event
times. Statistics in Medicine 16, 239257.
H UBER , P. J. (1967). The behavior of maximum likelihood estimates under nonstandard conditions. In Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, Vol 1, pp. 221-233. Berkeley:
University of California Press.
L IANG , K. Y. AND Z EGER , S. L. (1986). Longitudinal data analysis using generalized linear models. Biometrika 73,
1322.
L ITTLE , R. J. A. (1988). A test of missing completely at random for multivariate data with missing values. Journal
of the American Statistical Association 83, 11981202.
L ITTLE , R. J. A. (1993). Pattern-mixture models for multivariate incomplete data. Journal of the American Statistical
Association 88, 125134.
L ITTLE , R. J. A. (1994). A class of pattern-mixture models for normal incomplete data. Biometrika 81, 471483.
L ITTLE , R. J. A. (1995). Modelling the drop-out mechanism in repeated-measures studies. Journal of the American
Statistical Association 90, 11121121.
L ITTLE , R. J. A. AND RUBIN , D .B. (1987). Statistical Analysis with Missing Data. New York: Wiley.
M ACHIN , D., FARLEY, T., B USCA , B., C AMPBELL , M. AND D A RCANGUES , C. (1988). Assessing changes in
vaginal bleeding patterns in contracepting women. Contraception 38, 165179.
M OLENBERGHS , G., K ENWARD , M. G. AND L ESAFFRE , E. (1997). The analysis of longitudinal ordinal data with
nonrandom drop-out. Biometrika 84, 3344.
M OLENBERGHS , G., M ICHELS , B., K ENWARD , M. G. AND D IGGLE , P. J. (1998). Missing data mechanisms and
pattern-mixture models. Statistica Nederlandica 52, 153161.
PAIK , M. (1997). The generalized estimating equation approach when data are not missing completely at random.
Journal of the American Statistical Association 92, 13201329.
PARK , T. AND DAVIS , C. S. (1993). A test of missing data mechanism for repeated categorical data. Biometrics 49,
631638.
PARK , T. AND L EE , S.-Y. (1997). A test of missing completely at random for longitudinal data with missing observations. Statistics in Medicine 16, 18591871.
PAWITAN , Y. AND S ELF, S. (1993). Modeling disease marker processes in AIDS. Journal of the American Statistical
Association 88, 719726.
P RENTICE , R. L. (1988). Correlated binary regression with covariates specific to each binary observation. Biometrics
44, 10331048.
R AO , C. R. (1965). Linear Statistical Inference and its Application. New York: Wiley.
ROBINS , J. M., ROTNITZKY, A. AND Z HAO , L. P. (1995). Analysis of semiparametric regression models for repeated outcomes in the presence of missing data. Journal of the American Statistical Association 90, 106121.
ROYALL , R. M. (1986). Model robust confidence intervals using maximum likelihood estimators. International Statistical Review 54, 221226.
RUBIN , D .B. (1976). Inference and missing data. Biometrika 63, 581592.
SAS I NSTITUTE I NC . (1999). SAS/STAT Users Guide: Version 7-1. Cary, NC: SAS Institute Inc.
S TATA C ORPORATION (1997). Stata Statistical Software: Release 5.0. College Station, Texas: Stata Corporation.

156

G ARRETT M. F ITZMAURICE AND NAN M. L AIRD

WARE , J. H. (1985). Linear models for the analysis of longitudinal studies. American Statistician 39, 95101.
W HITE H. (1982). Maximum likelihood estimation of misspecified models. Econometrica 50 126.
W U , M. C. AND BAILEY, K. R. (1988). Analysing changes in the presence of informative right censoring caused by
death and withdrawal. Statistics in Medicine 7, 337346.
W U , M. C. AND BAILEY, K. R. (1989). Estimation and comparison of changes in the presence of informative right
censoring: conditional linear model. Biometrics 45, 939955.
Z EGER , S. L. AND L IANG , K. Y. (1986). Longitudinal data analysis for discrete and continuous outcomes. Biometrics 42, 121130.

[Received July 15, 1999; revised November 4, 1999; accepted for publication November 22, 1999]

Вам также может понравиться