Daniels M.J., Hogan J.W. Missing Data in Longitudinal Studies: Strategies for Bayesian Modeling and Sensitivity Analysis

Подождите немного. Документ загружается.

110 MISSING DATA MECHANISMS

of dropout does and does not depend on future observations is given in our

analysis of the Growth Hormone trial in Section 10.2.

From the point of view of modeling, mixture models treat dropout or miss-

ingness as a source of variation in the full data distribution. Specifying a

diﬀerent distribution for each dropout time or missing data pattern may seem

cumbersome, but it frequently has the advantage of making explicit the pa-

rameters that cannot be identiﬁed by observed data. This can be seen as

follows (again leaving aside covariates). Recall the mixture model speciﬁca-

tion is

p(y

obs

, y

mis

, r | ω)=p(y

mis

, y

obs

| r, α) p(r | φ).

Further decompose the ﬁrst term on the right-hand side as

p(y

obs

, y

mis

| r, α)=p(y

mis

| y

obs

, r, α

) p(y

obs

| r, α

), (5.20)

where α

= λ

(α), α

= λ

(α)arefunctions of the mixture component

parameter α.Theparameters α

and α

may overlap. The subscript E in-

dicates that α

indexes an extrapolation distribution; i.e., the distribution of

missing responses given the observables under the assumed full-data model.

The subscript O indicates that α

indexes a model for observables. In general,

cannot be fully identiﬁed from data alone.

Now suppose there exists a partition (α

E:I

, α

E:NI

)ofα

such that the

observed-data likelihood is a function of α

E:I

but not of α

E:NI

(i.e., α

E:I

is identiﬁed from data but α

E:NI

is not). A parameter α

E:NI

satisfying this

condition makes a suitable basis for formulating sensitivity analysis or for

encoding the conditional distribution of missing responses using informative

prior distributions; in general, mixture models lend themselves well to pa-

rameterizations having these properties. To illustrate, we revisit Example 5.9.

Example 5.11. Mixture of bivariate normals with missingness in Y

We generalize the speciﬁcation of Example 5.9 to allow separate variance

matrices by pattern of missingness; speciﬁcally let Y =(Y

)

,with

Y | R =1 ∼ N (µ

(1)

, Σ

(1)

)

Y | R =0 ∼ N (µ

(0)

, Σ

(0)

)

R ∼ Ber(φ),

where µ

(r)

=(µ

(r)

,µ

(r)

)

and Σ

(r)

has elements σ

(r)

=(σ

(r)

,σ

(r)

,σ

(r)

). For

pattern r,deﬁne β

(r)

, β

(r)

,andσ

(r)

2|1

to be, respectively, the intercept, slope,

and residual variance for the regression of Y

on Y

.Under this reparameter-

ization, the full-data model parameters are

α =



(r)

,σ

(r)

,β

(r)

,β

(r)

,σ

(r)

2|1

: r =0, 1



The nonidentiﬁed parameters are seen to be those indexing the conditional

FULL-DATA MODELS UNDER MNAR 111

distribution of Y

given Y

among those with R =0.Hence the distribution

p(y

obs

, y

mis

| r, α)canbeparameterized as

p(y

obs

, y

mis

| r, α)=p(y

mis

| y

obs

,r,α

) p(y

obs

| r, α

). (5.21)

To simplify notation, deﬁne p

(y)=p(y | r). Then the two factors of (5.21)

can be written as

mis

| y

obs

, α



| y

, β

(0)

,σ

(0)

2|1

)



1−r

obs

| α



| µ

(0)

,σ

(0)

)



1−r



| y

, β

(1)

,σ

(1)

2|1

) p

| µ

(1)

,σ

(1)

)



Hence α

can be partitioned as (α

E:I

, α

E:NI

), with

E:I

= ∅

E:NI

=(β

(0)

,σ

(0)

2|1

Furthermore,

=(µ

(1)

, β

(1)

,σ

(1)

,µ

(0)

,σ

(0)

It is a simple matter to show that the observed data likelihood contribution

for individual i does not depend on α

E:NI

L(α

,φ| y

i,obs

) ∝



(1 − φ) p

| µ

(0)

,σ

(0)

)



1−r



φp

| y

, β

(1)

,σ

(1)

2|1

) p

| µ

(1)

,σ

(1)

)



In this model, it is possible to show that setting β

(0)

= β

(1)

and σ

(0)

2|1

= σ

(1)

2|1

yields MAR; hence, in general, a function that maps identiﬁed parameters and

sensitivity parameters to the space of nonidentiﬁed parameters can be used

to quantify departures from MAR.

For example, ifweconstrainα

E:NI

such that β

(0)

= β

(1)

and σ

(0)

2|1

= σ

(1)

2|1

but let

(0)

= β

(1)

+∆,

then assigning a point mass prior at ∆ = 0 implies MAR. Fixing ∆ at a

nonzero value implies MNAR. Formalpriorson∆also can be used. More

detail on this approach, including a general framework for model parameter-

ization and prior speciﬁcation, is given in Chapter 9. Illustrations using the

Growth Hormone trial and the OASIS study are provided in Chapter 10. 2

By contrast with the selection model forbivariate response in Example 5.10,

this observed-data likelihood is completely free of parameters indexing the

conditional distribution of Y

mis

given Y

obs

.Forthepurposes of posterior

112 MISSING DATA MECHANISMS

inference, information about this distribution will derive at least in part from

parameters that are informed solely by prior distributions.

To summarize, mixture models have the advantage that in many cases, one

can ﬁnd full-data parameters indexing the distribution of missing responses

that are not identiﬁed by observed data. In one sense, this makes inferences

about the full-data more transparent in that the source of information about

the posterior is clear. A potential downside of the models relates to practical

implementation: in the previous example, we conﬁne attention to a bivari-

ate distribution with no covariates, and even in this simple case there are

three unidentiﬁed parameters. As the dimension of Y increases, so will the

dimension of the unidentiﬁed parameter. In our case studies in Chapter 10,

we address this concern by illustrating sensible ways to reduce the dimension

of nonidentiﬁed parameters indexing the extrapolation distribution.

5.9.3 Shared parameter models

Athirdapproach to specifying the full data distribution is to use an explicitly

multilevel formulation, frequently called a shared parameter model,whereran-

dom eﬀects b are modeled jointly with Y and R.Insomecases the random

eﬀects can be used to explain the association or to capture multiple sources

of variation. Key papers tracing the development of these models include Wu

and Carroll (1988); Follmann and Wu (1995); Pulkstenis et al. (1998), and

Henderson et al. (2000). De Gruttola and Tu (1994), Wulfsohn and Tsiatis

(1997), and Faucett and Thomas (1996) use similar formulations but with the

objective of modeling a survival process as a function of stochastic longitudinal

covariates.

The general form of the full-data model using a shared parameter ap-

proach is

p(y, r | x, ω)=



p(y, r, b | x, ω) db. (5.22)

Speciﬁc shared parameter models are formulated by making assumptions

aboutthe joint distribution under the integral sign. Notice in (5.22) that

the full-data parameter ω also includes parameters indexing the distribution

of the random eﬀects.

Advantages to the SPM model include simpliﬁed speciﬁcation for the re-

sponse and missingness components. When Y is measured with error, SPMs

provide a useful way for missingness to depend on the error-free version of

Y ,represented in terms of the random eﬀects (this is a type of ‘random-

coeﬃcient-dependent missingness’; see Wu and Carroll, 1988, and Little, 1995).

Another potential advantage is that through the use of random eﬀects, SPMs

can be used to handle high-dimensional or multilevel response data. A disad-

vantage is that except in simple settings, the underlying missing data mecha-

FULL-DATA MODELS UNDER MNAR 113

nism can be diﬃcult to understand and may not even have a closed form: the

representation of p(r | y, ω)requiresintegration over the random eﬀects.

Example 5.12. Random coeﬃcients selection model.

Wu and Carroll (1988) are concerned with estimating the population slope of

longitudinal lung function measurements using data from a study where the

dropout proportion was appreciable. They assume the lung function measures

follow a linear random eﬀects model

| x

, b

∼ N(x

β + w

, Σ

(φ)),

where w

⊆ x

are the random eﬀects covariates, with rows w

=(1,t

);

hence each individual has a random slope and intercept. The random eﬀects

=(b

)

are assumed to follow a bivariate normal distribution

∼ N (0, Ω).

The hazard of dropout is Bernoulli, with

| R

i,j−1

=1, b

∼ Ber(π

where hazard of dropout depends on the random eﬀects governing Y

via

g(π

)=ψ

+ ψ

The model is seen as a special case of (5.22) where the joint distribution

under the integral is factored as

p(y, r, b | x, ω)=p(r | y, b, x, ψ) p(y | b, x, β, φ) p(b | x, Ω)

= p(r | b, ψ) p(y | b, x, β, φ) p(b | Ω),

with ω =(β, φ, Ω, ψ). The ﬁrst equality is just a factorization of the joint

distribution, and the second equality follows from the key assumption that

dropout R is independent of both Y

obs

and Y

mis

,conditionally on random

eﬀects. However, integrating over the random eﬀects induces dependence be-

tween R and Y

mis

conditionally on Y

obs

;hence the model characterizes an

MNAR mechanism. 2

The‘conditional linear model’ (Wu and Bailey, 1989; Hogan and Laird,

1997a) also can be viewed as a version of an SPM, though its more standard

representation is in terms of a mixture model. The conditional linear model

specializes (5.22) to

p(y, r, b | x)=p(y | r, b, x) p(b | r, x) p(r | x).

As an example, Hogan and Laird (1997a) assume that mixture components

p(y | r, b, x), r ∈ R in the ﬁrst factor follow random eﬀects models where

the regression coeﬃcients β depend on r,andtherandom eﬀects have a

distribution that does not depend on r; i.e., p(b | r, x)=p(b | x). Hogan

et al. (2004a) generalize the model to allow continuous dropout time, using

114 MISSING DATA MECHANISMS

amixtureofvarying coeﬃcient models. For the case where p(y | r, b, x)

and p(b | r, x)follow normal distributions and E(Y | r, b, x)islinear in b,

integrating the random eﬀects yields a standard mixture-of-normals model for

the full-data distribution (as in Example 5.11).

5.10 Summary

In this chapter, we have provided a formal framework for conceptualizing infer-

ence from incomplete data, based on specifying a full-data model and drawing

inference from an observed-data likelihood (and associated priors). In order to

make clear the limits of drawing inference from incomplete data, we provided

formal deﬁnitions for missing data mechanisms and applied them to settings

with dropout. We then showed how to apply the notions of model speciﬁ-

cations and missing data assumptions intermsofselection models, mixture

models, and shared parameter models.

In the next chapter, we turn our focus to drawing inference about full-

data models for longitudinal responses under the ignorability assumption. For

longitudinal data, the key issue lies in specifying the dependence structure,

because the extrapolation model under ignorability, p(y

mis

| y

obs

), depends

critically on the dependence model.

5.11 Further reading

Models that accommodate dropout and intermittent missingness

Several papers have discussed approaches to handle dropout and intermittent

missingness simultaneously. Albert (2000) and Albert et al. (2002) constructed

models with a multinomial random variable speciﬁed at each time indicating

still in study, intermittent missing, or dropout. Lin, McCulloch, and Rosen-

heck (2004) deal with intermittent missingness by grouping the missing data

patterns using a latent pattern mixture approach. Troxel et al. (1998) use a

pseudo-likelihood approach that models the missingness indicator at time t

conditional on only the potentially missing response at time t.

Multiple cause dropout and dropout due to death

Forageneral discussion of multiple cause dropout, see Rotnitzky et al. (2001).

There have been several important papers recently that address the issues

involved with dropouts due to death. Frangakis and Rubin (2002) illustrate

the use of principal stratiﬁcation to estimate causal eﬀects of treatment in

thepresence of death. Kurland and Heagerty (2005) address dropout due to

death in observational studies by directly modeling the mean conditional on

surviving past a speciﬁed time.

CHAPTER 6

Inference about Full-Data Parameters

under Ignorability

6.1 Overview

In this chapter, we discuss inference about the full-data distribution under

the ignorability assumption. Under the full-data model factorization

p(y, r, | x, ω)=p(y | x, θ(ω)) p(r | y, x, ψ(ω)),

(6.1)

missing at random implies

p(r | y, x, ψ(ω)) = p(r | y

obs

, x, ψ(ω)). (6.2)

Ignorability arises under the additional restriction of aprioriindependence

between the parameters of the full-data response model θ and the parameters

of the missing data mechanism ψ.Under ignorability, we only need to specify

the full-data response model, not the missing data mechanism. In the following

we suppress the dependence on x to maintain clarity.

Recall that the full-data posterior of θ is p(θ | y)=p(θ | y

obs

, y

mis

). Under

ignorability, inference is based on the observed-data posterior,

p(θ | y

obs



p(θ | y)p(y

mis

| y

obs

) dy

mis

∝



p(y

obs

, y

mis

| θ)p(θ)dy

mis

. (6.3)

This integral, and hence p(θ | y

obs

), depends on the assumed full data response

model p(y | θ), where y =(y

obs

, y

mis

). Thus, valid inference depends on its

correct speciﬁcation. We must keep in mind that neither the full-data response

model nor the ignorability assumption itself is veriﬁable from the observed

data. In longitudinal data, one of the most important aspects in specifying a

full-data response model is the association (dependence) structure.

The main focus of this chapter is methods for posterior inference. Sampling

from the observed data posterior can often be simpliﬁed by supplementing the

Gibbs sampler with a data augmentation step. In particular, we can sample

from

1. p(y

mis

| y

obs

, θ)usingadataaugmentation step, and

2. p(θ | y)usingGibbs sampling.

115

116 INFERENCE UNDER MAR

The second step of this algorithm, which samples the full-data response model

parameters, is the same as if there were no missing data (Section 3.4). This

often simpliﬁes computations considerably. We provide more details on this

in Section 6.2.

The remainder of the chapterisorganized as follows. First, issues of model

speciﬁcation are discussed, emphasizing the importance of correctly specifying

the dependence structure in p(y | θ). Second, the use of data augmentation to

simplify posterior sampling will be described in the setting of missing data and

its implementation in speciﬁc models shown. Third, we will illustrate modeling

of dependence by giving examples of convenient parsimonious structures and

computationally tractable ways to allow the dependence to be a function of

model covariates. Fourth, we will illustrate the use of joint modeling of several

processes as a way to incorporate information on auxiliary stochastic time-

varying covariates (recall Section 5.4.4 from Chapter 5 and Deﬁnition 5.11).

The chapter closes with some suggested approaches for assessing model ﬁt and

model checking. Throughout the chapter, we often refer the full-data response

model and posterior as the full-data model and posterior without any loss of

clarity.

6.2 General issues in model speciﬁcation

Inference under ignorability relies on correct speciﬁcation of the full-data re-

sponse model p(y | θ). It is convenient to view the the full-data response model

in terms of assumptions about the mean, variability and (serial) dependence,

and distribution. The following development focuses on the consequences of

mis-speciﬁcation of the dependence.

6.2.1 Mis-speciﬁcation of dependence

In longitudinal settings, when primary interest is in the mean, the dependence

is often viewed as a nuisance, with correct speciﬁcation only leading to gains

in eﬃciency. However, when dealing with incomplete data, correctly specifying

the dependence takes on added importance.

To illustrate the consequences of mis-speciﬁcation of dependence under

ignorability, we return to the bivariate normal model (Example 5.7) where

is observed and Y

is possibly missing. Recall, Y

∼ N(µ, Σ), with the

j, k element of Σ denoted as σ

.Toﬁllin missing values of Y

within the

sampling algorithm, we sample from the (conditional) normal distribution

p(y

| y

, µ, Σ), where θ =(µ, Σ). It has mean µ

+ φ

− µ

)andis

afunction of the covariance parameter through φ

= σ

/σ

.Bycontrast,

with complete data,the data augmentation step is not necessary and the

posterior of µ

is not afunction of φ

The simplest mis-speciﬁcation would be to assume φ

=0when,intruth,

ISSUES IN MODEL SPECIFICATION 117

=0.Alessobvious, but important mis-speciﬁcation would be to assume

is constant when, in truth, φ

is a function of modelcovariates via φ

h(x)forsomefunction h.

Of course,whenJ>2, there are more dependence parameters and covari-

ance model simpliﬁcation is often necessary. For example, in a multivariate

normal model, when Σ is large, a common strategy is to assume a simpler,

parsimonious form for Σ;wemight(incorrectly) assume an AR-1 form for Σ.

Or we might introduce structure via random eﬀects, but assume the wrong

modelfor the random eﬀects covariance matrix (Daniels and Zhao, 2003).

Several of these cases are explored in detail below. We start by examining

the consistency of the posterior distribution of the mean when the dependence

is mis-speciﬁed.

Example 6.1. Posterior distribution of the mean for a bivariate normal un-

der monotone missingness that is ignorable (continuation ofExample5.7).

Consider a bivariate response from a randomized controlled trial with binary

treatment and response vector Y =(Y

)

. Y

is always observed, but Y

may be missing. The missingness indicator R takes value 0 if Y

is missing.

Conditional on a single scalar covariate x,weassumethefull-data response

model is a bivariate normal distribution

Y | x, µ, Σ ∼ N(µ(x), Σ(x)),

with Σ(x)={σ

(x)} and φ

(x)=σ

(x)/σ

(x). For j =1, 2, deﬁne

(x)=E(Y

| x)

(x)=E(Y

| x, R =1)

(x)=E(Y

| x, R =0).

It can be shown that under ignorability,

(x) − φ

(x) µ

(x)=µ

(x) − φ

(x) µ

(x)

= µ

(x) − φ

(x) µ

(x). (6.4)

We now examine the posterior distribution of µ

(x) under mis-speciﬁcation

of the dependence φ

(x); in particular, we incorrectly assume φ

(x)=φ

Deﬁne



β =(



)

to be the ordinary least squaresestimatesforregressing

on Y

for those with R =1andX = x where φ

(x)istheslope. Assume

there are n

individuals with X = x.

Consider the conditional posteriormeanofµ

(x)given{y

obs

,µ

(x)};we

do not need to condition on µ

(x)here, but it will make the arguments more

clear. We examine the behavior of the posterior mean of µ

(x)asn

→∞(and

the dropout proportion remains constant); in particular, we will determine

whether the posterior distribution of µ

(x)isconsistent (recall Deﬁnition 3.2).

Under ignorability and a correctly speciﬁed covariance, condition (6.4) can

118 INFERENCE UNDER MAR

be used to show that

E{µ

(x) | y

obs

,µ

(x)} =



(x)

→ µ

(x) − φ

(x)µ

(x)+φ

(x)µ

(x)

= µ

(x). (6.5)

However, under the mis-speciﬁcation φ

(x)=0,

E{µ

(x) | y

obs

,µ

(x)} =¯y

(x)

→ µ

(x)

= µ

(x) − φ

(x){µ

(x) − µ

(x)},

where

(x)=



{i:x

=x,r

=1}

and m

is the number of subjects with {x

= x, r

=1}.So, under mis-

speciﬁcation of the dependence structure, the posterior distribution of µ

(x)

will be inconsistent with bias

−φ

(x){µ

(x) − µ

(x)}.

Clearly, the bias increases with themagnitude of the dependence parame-

ter. The bias is also a function of the diﬀerence in the means between the

completers and dropouts at time 1.

Using similar arguments, bias in treatment eﬀect at time 2, µ

(1) − µ

(0),

is equal to

−φ

(1){µ

(1) − µ

(1)} + φ

(0){µ

(0) − µ

(0)}.

This example illustrates that even in the simplest case, a bivariate normal with

ignorable monotone missingness, mis-speciﬁcation of the dependence leads to

an inconsistent posterior for themeanparameters. Similar biases occur when

the mis-speciﬁcation is φ

(x)=φ

=0.Theseresults readily generalize

to higher-dimensional multivariate normals and other models discussed in

Chapter 2.

6.2.2 Orthogonal parameters

Amoregeneral way to understand the importance of correctly modeling the

dependence is by exploring the diﬀerence in the form of the information matrix

for the mean and dependence parameters for complete data vs. incomplete

data (under ignorability). Before going into the details, we ﬁrst introduce the

concept of orthogonal parameters. Let β represent the mean parameters, α

the dependence parameters, and let  =logL.

ISSUES IN MODEL SPECIFICATION 119

Deﬁnition 6.1. Orthogonal parameters.

For parameters β and α,letβ

and α

be their true values, respectively.

Then β and α are orthogonal (Cox and Reid, 1987) if, for any component β

of β and any component α

of α,

,α

(β

, α

)=−E



∂

(β, α)

∂β

∂α



, α

=0. (6.6)

Theblock diagonality of the information matrix implies that the maximum

likelihood estimates of the orthogonal parameters are asymptotically indepen-

dent (Cox and Reid, 1987).

In the Bayesian paradigm, in addition to the full-data likelihood, we also

have a prior. Typically, the prior will be O

(1), whereas the likelihood is

(n). We now state a similar deﬁnition to Deﬁnition 6.1 for the Bayesian

setting.

Deﬁnition 6.2. Orthogonal parameters in Bayesian inference.

Consider parameters β and α with prior p(β, α), and let β

and α

be their

true values. Then β and α are orthogonal if, for any component of β

of β

and any component α

of α,



,α

(β

, α

)=−E

∂

{(β, α)+log(p(β, α))}

∂β

∂α

, α

=0. (6.7)

For identiﬁed parameters, if condition (6.6) holds, then condition (6.7) holds

exactly under independent priors on β and α.Ifp(β, α) = p(β)p(α), but

satisﬁes p(β, α)=O

(1), then condition (6.7) holds asymptotically.

When condition (6.7) holds, the joint posterior distribution of (β, α)is

equal to the product of the marginal posteriors of β and α (asymptotically).

Fororthogonality involving non- or weakly identiﬁed parameters, the apriori

independence of β and α is necessary for orthogonality.

Orthogonality of the mean and dependence parameters is only a necessary

condition for the posterior distribution of the mean parameters to be consis-

tent if the dependence is mis-speciﬁed. A suﬃcient condition for the poste-

rior of the mean parameters to be consistent is given next. This condition is

stronger than orthogonality conditions given in Deﬁnitions 6.1 and 6.2.

Theorem 6.1. Consistency of the posterior for mean parameters.

Let β

be the true value for the mean parameters and α



be any value for

the dependence parameters. If

,α

(β

, α



)=−E



∂

(β, α)

∂β∂α



, α



= 0, (6.8)

then the mle’s of the mean parameters are consistent even if the dependence

is mis-speciﬁed (Firth, 1987). The analogous condition for the consistency of