Joint Modeling Approaches Using Random Effects

Joint Modeling Approaches in Longitudinal Studies
Using Random Eects

Dimitris Rizopoulos1
Geert Verbeke2
Geert Molenberghs2
d.rizopoulos@erasmusmc.nl
geert.verbeke@med.kuleuven.be
geert.molenberghs@uhasselt.be
Department of Biostatistics, Erasmus Medical Center, Rotterdam, the Netherlands

2
Interuniversity Institute for Biostatistics and statistical Bioinformatics (I-BioStat)
Katholieke Universiteit Leuven & Universiteit Hasselt, Belgium
www.erasmusmc.nl/biostatistiek & www.ibiostat.be
JSM, Vancouver, BC, August 3, 2010
Learning Objectives
After this course participants will be able to identify settings in which

a joint modeling approach is required. From the course it will become
clear which joint models can be used depending on the actual
research questions to be answered, and which model-building
strategies are currently available. Further, participants should be able
to construct and t an appropriate joint model, correctly interpret the
obtained results, and extract additional useful information (e.g.,
plots) that can help communicate the results better.
The course will be explanatory rather than mathematically rigorous.
Therefore emphasis is given on sucient detail in order for
participants to obtain a clear view on the dierent joint modeling
approaches, and how they should be used in practice. To this end, we
rst motivate joint modeling using real data sets, and then illustrate
in detail the virtues and drawbacks of each of the presented joint
modeling approaches. For completeness and throughout the course,
references are provided to material with more technical information.
Contents
Related References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Motivation for Joint Modeling
Outcomes in Longitudinal Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Approaches to Simultaneously Analyze Multiple Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
Motivating Datasets and Key Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Generalized Estimating Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
The Generalized Linear Mixed Model (GLMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Joint Modeling Approaches in Longitudinal Studies Using Random Eects: JSM 2010
Inverse Probability Weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
II
Joint Models for the Longitudinal and Dropout Processes
Sensitivity Happens . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
A Latent-variable Mixture Model as a Basis for Sensitivity Analysis in Incomplete Longitudinal Data . . . . . . . . . . . . . . . . 78
III
Joint Models for Longitudinal and Time-to-event Data
Longitudinal Studies with Time-to-event Primary Endpoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10
Time-dependent Covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
11
Joint Models for Longitudinal and Time-to-Event Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
12
Connection with Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
13
Further Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
54
94
ii
IV
Joint Models for Multivariate Longitudinal Data
14
Random-eects Models for High-dimensional Multivariate Longitudinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144
Case Study: A Joint Model for Multivariate Longitudinal Data and a Time to Event
15
Renal Transplant Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165
143
164
iii
Chapter 0
Related References
Aerts, M., Geys, H., Molenberghs, G., and Ryan, L.M. (2002). Topics in
Modelling of Clustered Data. London: Chapman and Hall.
Brown, H. and Prescott, R. (1999). Applied Mixed Models in Medicine.
New-York: John Wiley & Sons.
Davidian, M. and Giltinan, D.M. (1995). Nonlinear Models For Repeated
Measurement Data. London: Chapman and Hall.
Diggle, P.J., Heagerty, P.J., Liang, K.Y. and Zeger, S.L. (2002). Analysis of
Longitudinal Data. (2nd edition). Oxford: Oxford University Press.
Fitzmaurice, G., Davidian, M., Verbeke, G. and Molenberghs, G. (2009).
Longitudinal Data Analysis, Chapman & Hall/CRC Handbooks of Modern

Statistical Methods. Boca Raton: Chapman & Hall/CRC
Goldstein, H. (1995). Multilevel Statistical Models. London: Edward Arnold.
Kalbeisch, J. and Prentice, R. (2002). The Statistical Analysis of Failure Time
Data. (2nd edition). New York: Wiley.
Longford, N.T. (1993). Random Coecient Models. Oxford: Oxford University
Press.
Molenberghs, G. and Kenward, M.G. (2007). Missing Data in Clinical Studies.
Chichester: John Wiley.
Molenberghs, G. and Verbeke, G. (2005). Models for Discrete Longitudinal Data.
New York: Springer.
Pinheiro, J.C. and Bates D.M. (2000). Mixed eects models in S and S-Plus,
Springer Series in Statistics and Computing. New-York: Springer.

Searle, S.R., Casella, G., and McCulloch, C.E. (1992). Variance Components.
New-York: Wiley.
Tsiatis, A. and Davidian, M. (2004). Joint modeling of longitudinal and
time-to-event data: an overview. Statistica Sinica 14, 809834.
Verbeke, G. and Molenberghs, G. (1997). Linear Mixed Models In Practice: A
SAS Oriented Approach, Lecture Notes in Statistics 126. New-York: Springer.
Verbeke, G. and Molenberghs, G. (2000). Linear Mixed Models for Longitudinal
Data. Springer Series in Statistics. New-York: Springer.
Vonesh, E.F. and Chinchilli, V.M. (1997). Linear and Non-linear Models for the
Analysis of Repeated Measurements. Marcel Dekker: Basel.
Part I
Motivation for Joint Modeling
Chapter 1
Outcomes in Longitudinal Studies
Often in longitudinal studies dierent types of outcomes are collected

Explicit outcomes
multiple longitudinal responses (e.g., markers, blood values)
time-to-event(s) of particular interest (e.g., death, relapse)
Implicit outcomes
missing data (e.g., dropout, intermittent missingness)
visit times
Example: patients in the waiting list for primary renal transplantation
Explicit outcomes:
markers for renal graft functioning: glomerular ltration rate, haematocrit,
blood pressure, proteinuria
events of interest: transplantation & death
Example: a longitudinal study on HIV infected patients
Explicit outcomes
markers for the progression of the disease: CD4 cell count, anti-p24 antibody,
beta2-microglobulin
time-to-death
Implicit outcomes
patients die or leave the study dropout: planned measurements on the
markers are not taken
1.1
Focus of Inference
Depending on the questions of interest, dierent types of statistical analysis are

required
We should distinguish between two general types of analysis
separate analysis per outcome
joint analysis of outcomes
Focus on each outcome separately
does treatment aect survival?
are the average longitudinal evolutions dierent between males and females?
...
Focus on multiple outcomes

Complex hypothesis testing: does treatment improve the average longitudinal
proles in all markers?
Complex eect estimation: what is the eect of the longitudinal evolution of
CD4 cell count in the hazard rate for death?
Association structure among outcomes:
* how the association between markers evolves over time (evolution of the
association)
* how marker-specic evolutions are related to each other (association of the
evolutions)
Prediction: can we improve prediction for the time to death by considering all
markers simultaneously?
Handling implicit outcomes: focus on a single longitudinal but with dropout
10
1.2
Recent Developments
Until recently emphasis has been

restricted or coerced to separate analysis per outcome
or given to naive types of joint analysis (e.g., last observation carried forward)
Main reasons
lack of appropriate statistical methodology
lack of ecient computational approaches & software
11
However, recently there has been an explosion of statistics and biostatistics

literature of joint modeling approaches
Many dierent approaches have proposed that

can handle dierent types of outcomes
can be utilized in pragmatic computing time
can be rather exible
most importantly: can answer the questions of interest
12
1.3
What Is This Course About?
Purpose of this course is to present the state of the art in

Joint Modeling Techniques
In particular, we will present

the challenges in joint modeling
how joint modeling approaches can be applied in practice
what are the recent trends & a look into the future
13
Caveat: we will not present all available joint modeling approaches

we will focus on three
interdependencies captures via random eects
14
Agenda:
Part I: Approaches to analyze multiple outcomes simultaneously
Part II: Joint Models for the Longitudinal and Dropout Processes
Part III: Joint Models for Longitudinal and Time-to-event Data
Part IV: Joint Models for Multivariate Longitudinal Data
Part V: Connections & Extensions
15
Chapter 2
Approaches to Simultaneously Analyze Multiple Outcomes
Introduction and notation

Multivariate models
Conditional models
Shared-parameter models
Random-eects models
Methods based on dimension reduction
16
2.1
Introduction and Notation
Let Y1 and Y2 be two outcomes measured on a number of subjects for which joint
modeling is of scientic interest.
Both can be measured longitudinally
E.g., CD4 cell count and viral load in HIV patients
One longitudinal outcome with one survival outcome
E.g., Haematocrit and time to graft failure in renal transplant patients
We will discuss various approaches possible to construct a joint density f (y1, y2)
of (Y1, Y2)
Extensions to more than 2 outcomes are (relatively) straightforward
17
2.2
Multivariate Models
General idea: Specify f (y1, y2) directly

Advantages:
Allows for direct inferences for marginal characteristics of Y1, Y2, and their
associations
Symmetric treatment of Y1 and Y2
Disadvantages:
Dicult with Y1 and Y2 of a dierent type
Dicult for unbalanced data since association between Y1 and Y2 needs to be
modeled directly
Dicult to extend to higher dimensions
18
2.3
Conditional Models
General idea: Factorize f (y1, y2) as

f (y1, y2) = f (y1|y2)f (y2) = f (y2|y1)f (y1)
Advantage:
Modeling tasks reduced to specifying models for each of the outcomes
separately
Disadvantages:
With Y1 and Y2, specifying f (y1|y2) requires careful reection about plausible
associations between response Y1 and time-varying covariate Y2
No direct marginal inferences.
19
For example, based on f (y1, y2) = f (y1|y2)f (y2), E(Y1) requires

]
[
E(Y1) = E[E(Y1|Y2)] =
y1f (y1|y2) dy1 f (y2) dy2,
E(Y1) not necessarily of the same parametric form as E(Y1|Y2)

(e.g., logistic)
Eects on Y1 may be attenuated by conditioning on Y2.
Compatible specication of f (y1|y2)f (y2) and f (y2|y1)f (y1) often requires
direct specication of f (y1, y2)
Higher dimensions: Many possible factorizations
20
Typical example:
Selection models for longitudinal data subject to informative dropout
Marginal model f (y1) for the longitudinal outcome Y1
Conditional model for dropout time Y2:
P [Y2 = t | Y1(t), Y1(t 1), . . .]
21
2.4
Shared-parameter Models
General idea: Latent variables, shared by Y1 and Y2 imply associations

Let b denote a vector of random eects, with density f (b) (often normal)
Assume conditional independence: Y1Y2|b
Joint density of (Y1, Y2) obtained from
f (y1, y2) =
f (y1, y2|b) db =
f (y1|b)f (y2|b)f (b) db
22
Advantages:
Y1 and Y2 can be of dierent type
Parameters in joint model f (y1, y2) have the same interpretation as in the
univariate models f (y1) and f (y2)
Extension to higher dimensions very straightforward
Disadvantage:
Very strong assumptions about the association between Y1 and Y2
Assume random-intercepts models for Y1 and Y2:
Y1(t) = 1 + b + 2t + e1(t)
Y2(t) = 3 + b + 4t + e2(t)
with e1(t) N (0, 12), e2(t) N (0, 22), and b N (0, b2)
23
Implied association structure (s = t):

b2
Corr{Y1(s), Y1(t)} = 2
b + 12
2b2
Corr{Y2(s), Y2(t)} = 2 2
b + 22
b2
Corr{Y1(s), Y2(t)} = 2
2
b + 1 2b2 + 22
= Corr{Y1(s), Y1(t)} Corr{Y2(s)Y2(t)}

Association between Y1 and Y2 directly follows from association structures for
Y1 and Y2
In some cases this is problematic, especially in higher dimensions
24
Typical example:
Longitudinal outcome Y1 and time-to-event Y2
Assume a mixed model for Y1:
Y1(t) = (1 + b1) + (2 + b2)t + e1(t)
with b = (b1, b2) N (0, D)
Assume proportional hazard model for Y2 with hazard depending on b:
lim Pr{t T < t + h|T t, b} = = 0(t) exp{g(b)}
h0
for some pre-specied function g(b).

Hazard depends on some feature g(b) of the longitudinal trajectories.
Relation between Y1 and Y2 addressed via inference for
25
2.5
Random-eects Models
General idea: Separate but correlated latent variables for Y1 and Y2

Let b1 and b2 denote vectors of random eects, with joint density f (b1, b2)
(often normal)
Assume conditional independence: Y1Y2|(b1, b2)
Joint density of (Y1, Y2) obtained from
f (y1, y2) =

f (y1, y2|b) db =
f (y1|b1)f (y2|b2)f (b1, b2) db1db2
26
Advantages:
Same as with shared-parameter models
Less strict assumtptions about the association between Y1 and Y2
Assume random-intercepts models for Y1 and Y2:
Y1(t) = 1 + b1 + 2t + e1(t)
Y2(t) = 3 + b2 + 4t + e2(t)
with e1(t) N (0, 12), e2(t) N (0, 22), and (b1, b2) N (0, D)
Similar expressions for Corr{Y1(s), Y1(t)} and Corr{Y2(s), Y2(t)} as before, but
Corr{Y1(s), Y2(t)} = Corr(b1, b2) Corr{Y1(s), Y1(t)} Corr{Y2(s), Y2(t)}
Corr{Y1(s), Y1(t)} Corr{Y2(s), Y2(t)}

does not directly follow from the association structures for Y1 and Y2
27
Disadvantage:
Dimensionality of b increases with the number of outcomes modeled, hence
also the integration in
f (y1, . . . , yk ) =
=
...
f (y1|b1) . . . f (yk |bk )f (b1, . . . , bk ) db1 . . . dbk
Computational problems (addressed later)
28
2.6
Methods Based on Dimension Reduction
General idea: Use a factor-analytic, or principal-component type, analysis to

rst reduce the dimensionality of the response vector.
In a second stage, the principal factors are analyzed using any of the classical
(longitudinal) models.
Advantage:
Standard techniques can be used to analyze the principal factors
Disadvantages:
Inferences about principal factor(s), not about original outcome variables.
Very strong restrictions needed in cases of highly unbalanced longitudinal data:
unequal numbers of measurements for dierent subjects
observations taken at arbitrary time points.
29
Chapter 3
Motivating Datasets and Key Models
The toenail study

Missing data concepts
The linear mixed model
Generalized estimating equations
The generalized linear mixed model
30
3.1
Toenail Data: Continuous and Discrete Outcome
De Backer, De Keyser, De Vroey, Lesare (British Journal of Dermatology 1996)
Toenail Dermatophyte Onychomycosis: Common toenail infection, dicult to

treat, aecting more than 2% of population
Classical treatments with antifungal compounds need to be administered until the
whole nail has grown out healthy
New compounds have been developed that reduce treatment to 3 months
Randomized, double-blind, parallel group, multicenter study for the comparison of
two such new compounds (A and B) for oral treatment
31
Design:
2 189 patients randomized, 36 centers
48 weeks of total follow up (12 months)
12 weeks of treatment (3 months)
Complication: Dropout (24%)
Measurements at months 0, 1, 2, 3, 6, 9, 12
General research question:

Are both treatments equally eective for the treatment of TDO?
32
Unaected nail length (mm)?
Severity relative to treatment of

TDO?
As response is related to toe size, we

restrict to patients with big toenail as
target nail = 150 and 148 subjects
30 randomly selected proles, in each
group:
33
3.2
Longitudinal and Incomplete Data Notation
Subject i at occasion (time) j = 1, . . . , ni

Measurement Yij
Missingness indicator
Rij =
if Yij is observed,
otherwise.
Group Yij into a vector

Y i = (Yi1, . . . , Yini ) = (Y oi, Y m
i )
Yo
contains Yij for which Rij = 1,
i
Y m contains Y for which R = 0.
ij
ij
i
Group Rij into a vector Ri = (Ri1, . . . , Rini )
i
Di: time of dropout: Di = 1 + nj=1
Rij
34
3.3
Framework
f (Y i, Di|Xi, , )
Selection Models: f (Y i|Xi, ) f (Di|Xi, Y oi, Y m
i , )
MCAR
f (Di|Xi, )
MAR
f (Di|Xi, Y oi, )
MNAR
f (Di|Xi, Y oi, Y m
i , )
Pattern-Mixture Models: f (Y i|Xi, Di, )f (Di|Xi, )

Shared-Parameter Models: f (Y i|Xi, bi, )f (Di|Xi, bi, )
35
f (Y i, Di|Xi, , )
Selection Models: f (Y i|Xi, ) f (Di|Xi, Y oi, Y m
i , )
MCAR
MAR
MNAR
CC ?
direct likelihood !
joint model !?
LOCF ?
expectation-maximization (EM).
sensitivity analysis ?!
imputation ?
...
multiple imputation (MI).

(weighted) GEE !
36
3.4
Direct Likelihood Maximization

MAR
: f (Y oi|) f (Di|Y oi, )
Mechanism is MAR
and distinct
Interest in
Use observed information matrix
ignorability
Likelihood inference is valid
Outcome type
Modeling strategy
Software
Gaussian
Linear mixed model
SAS: MIXED
Non-Gaussian Generalized/non-linear mixed model SAS: GLIMMIX, NLMIXED

37
3.5
A Simple Linear Mixed Model
yi1
yi = ...
yini
+ 2
2
...
...
...
N Xi,
2
2 + 2
Marginal: Vi = 2Jni + 2Ini PD
yij = xij + bi + ij
bi N (0, 2)
ind.
ij N (0, 2)
Hierarchical:
{ 2
>0
2 0
38
3.6
The General Linear Mixed Model
[
]
yi1
yi = ... N Xi , Vi = ZiDZi + i
yini
Marginal: Vi = ZiDZi + i PD
yi = Xi + Zibi + i
bi N (0, D)
ind.
i N (0, i)
Hierarchical:
{
i PD
D PD
39
3.7
Analysis of the Toenail Data
Mixed model formulation without measurement error:

{
Yij =
(A0 + bi) + A1tij + A2t2ij + (2)ij
group A
(B0 + bi) + B1tij + B2t2ij + (2)ij
group B
Mixed model formulation with measurement error:

{
(A0 + bi) + A1tij + A2t2ij + (1)ij + (2)ij
Yij =
(B0 + bi) + B1tij + B2t2ij + (1)ij + (2)ij
group A
group B
40
Missing data mechanism under ignorability:

{
f (yio, di) =
f (yio)f (di)
under MCAR
f (yio)f (di|yio) under MAR,
Missing data mechanism under non-ignorability:
di1
P
(D
=
d
|D
d
,
y
)
i
i
i
i
i
k=2 [1 P (Di = k|Di k, yi )] di ni
ni
f (di|yi) =
[1 P (Di = k|Di k, yi)]

di > ni.
k=2
logit [P (Di = j | Di d, yi)] = 0 + 1yij + 2yi,j1

= 4.26 + 0.47yij 0.46yi,j1
LR: MAR versus MNAR: p < 0.0001
0.035
Sensitivity!
41
42
Chapter 4
Generalized Estimating Equations
Liang and Zeger (1986)
S() =
[Di]T [Vi()]1 (y i i) = 0
i=1
V i(.) is not the true variance of Y i but only a plausible guess

the score equations are solved in a standard way
Asymptotic distribution:
I0 =
i=1
) N (0, I 1I1I 1)
N (
0
0
DiT [Vi()]1Di
I1 =
DiT [Vi()]1Var(Y i)[Vi()]1Di
i=1
43
4.1
Application to the Toenail Data
Yij : severe infection (yes/no) at occasion j for patient i

tij : measurement time for occasion j
Ti: treatment group
ij
log
1 ij
Yij Bernoulli(ij ),
Eect
)
= 0 + 1Ti + 2tij + 3Titij
Par.
IND
EXCH
UN
Int.
-0.557(0.109;0.171)
-0.584(0.134;0.173)
-0.720(0.166;0.173)
Ti
0.024(0.157;0.251)
0.012(0.187;0.261)
0.072(0.235;0.246)
tij
-0.177(0.025;0.030)
-0.177(0.021;0.031)
-0.141(0.028;0.029)
Ti tij
-0.078(0.039;0.055)
-0.089(0.036;0.057)
-0.114(0.047;0.052)
0.1515
0.1208
0.0275
p-value for 3
estimate (model-based s.e.; empirical s.e.)

44
Chapter 5
The Generalized Linear Mixed Model (GLMM)
{
}
f (yij |ij , ) = exp [yij ij (ij )] + c(yij , )
1
ij = xij + zij bi
bi N (0, D)
The likelihood function for , D, and :
L(, D, ) =
fi(yi|, D, ) =
i=1
ni
N
i=1
fij (yij |bi, , ) f (bi|D) dbi
j=1
Integral:
Linear mixed model: closed form
In general: approximation of:
integrand
data
integral
45
5.1
The Toenail Data
Yij is binary severity indicator for subject i at visit j.

Model:
(
Yij |bi Bernoulli(ij ),
log
ij
1 ij
)
= 0 + bi + 1Ti + 2tij + 3Titij
Notation:
Ti: treatment indicator for subject i
tij : time point at which jth measurement is taken for ith subject
Adaptive as well as non-adaptive Gaussian quadrature, for various Q.
46
Results:
Quadrature (Num. Int.)
Parameter
AD Q = 50
NOAD Q = 3
PQL
MQL
GEE
Intercept A
1.63 (0.44)
1.52 (0.31)
0.72 (0.24)
0.56 (0.17)
0.72 (0.17)
Intercept B
1.75 (0.45)
1.91 (0.32)
0.72 (0.24)
0.53 (0.17)
0.65 (0.17)
Slope A
0.40 (0.05)
0.32 (0.03)
0.29 (0.03)
0.17 (0.02)
0.14 (0.03)
Slope B
0.57 (0.06)
0.41 (0.04)
0.40 (0.04)
0.26 (0.03)
0.25 (0.04)
Var. R.I.
15.99 (3.02)
5.11 (0.54)
4.71 (0.60)
2.49 (0.29)
|GEE| < |MQL| < |PQL| < |QUAD|
47
Chapter 6
Inverse Probability Weighting
Generalized estimating equations

Weighted generalized estimating equations
Pseudo-likelihood
Weighted pseudo-likelihood
Analysis of the toenail data
48
6.1
Weighted GEE
ni
i =
(1 pi)
=2
di 1
i = (1 pi) pidi
=2
pi = P (Di = |Di , Yi , Xi )
Ri = 1 if subject i is complete
Ri = 0 if subject i is incomplete
N
Ri i 1
S() =
Vi (y i i ) = 0
i
i=1
N
1 oi o 1 o
o
S() =
(Vi ) (y i i ) = 0.

i=1 i
49
6.2
Marginal Pseudo-likelihood
Full likelihood
First pseudo-likelihood
Second pseudo-likelihood
ln f (yi1, . . . , yin)
pi = j<k ln f (yik , yij )
pi = j<k ln f (yik , yij )/(ni 1)
Pairwise Likelihood: Complete Cases

N
Ri
U (yij , yik )
U naive, CC =
i=1
U IPWCC
U IPWCC,dr
j<k
Ri
=
U (yij , yik )
i
i=1
j<k

(
)
N
Ri
R
i
=
U (yij , yik )
U (yij , yik ) + 1
EY m|y o
i
i
i
i
i=1
j<k
j<k
50
Pairwise Likelihood: Available Cases
U naive, AC =
i=1
U IPWAC =
U (yij , yik ) +
(ni di + 1)U (yij )
j=1
N [
Rij
Rik
U (yij ) +
U (yik |yij )
ij
ik
i=1
U IPW, exch
di 1
j<k<di
i=1 j<k
U IPWAC,dr =
di 1
U i(yij , yik ) +
(ni di + 1) U i(yij )
j=1
j<k<di
E[U i(yik |yij )] +

E[U i(yij , yik )]
j<di k
di =j<k
di 1
N
U (yij , yik ) +
(ni di + 1)U (yij )
= U naive, AC =
i=1
j<k<di
j=1
51
6.3
Analysis of Toenail Data
Consider model:
Yij |bi N [bi + 0 I(Ti = 0) + 1 I(Ti = 1)2tj I(Ti = 0) + 3tj I(Ti = 1), 2]
bi N (0, 2)
Four forms of pseudo-likelihood:
Three ways of dealing with

missingness:
full likelihood
complete cases
complete pairs
available cases
naive
singly robust
doubly robust
52
Eect
Int.A
Int.B
Sl.A
Sl.B
R.I.v.
Res.v.
Par.
0
1
2
3
2
2
U full.lik.
2.52(0.247;0.228)
2.77(0.243;0.249)
0.56(0.023;0.045)
0.61(0.022;0.043)
6.49(0.628;0.633)
6.94(0.248;0.466)
U naive, CC
2.77(0.086;0.272)
2.82(0.083;0.271)
0.55(0.011;0.046)
0.60(0.011;0.044)
6.71(0.226;0.731)
7.31(0.150;0.520)
U naive, CP
2.70(0.081;0.248)
2.81(0.078;0.254)
0.56(0.011;0.045)
0.61(0.011;0.043)
6.67(0.213;0.680)
7.13(0.140;0.483)
U naive, AC
2.56(0.075;0.231)
2.77(0.073;0.250)
0.57(0.011;0.045)
0.61(0.010;0.043)
6.41(0.200;0.645)
7.05(0.137;0.472)
Eect
Int.A
Int.B
Sl.A
Sl.B
R.I.v.
Res.v.
Par.
0
1
2
3
2
2
U wt.lik.
1.85(0.092;0.303)
2.65(0.089;0.517)
0.68(0.014;0.068)
0.73(0.013;0.101)
6.21(0.235;1.032)
5.05(0.088;0.603)
U IPWCC
2.71(0.074;0.266)
2.78(0.073;0.265)
0.54(0.010;0.046)
0.60(0.010;0.044)
6.66(0.195;0.717)
7.29(0.130;0.513)
U IPWCP
2.77(0.079;0.270)
2.82(0.077;0.269)
0.53(0.010;0.044)
0.59(0.010;0.044)
6.72(0.209;0.753)
7.59(0.142;0.562)
U IPWAC
2.59(0.069;0.237)
2.77(0.069;0.249)
0.55(0.010;0.045)
0.60(0.010;0.043)
6.44(0.187;0.669)
7.35(0.130;0.514)
Eect
Int.A
Int.B
Sl.A
Sl.B
R.I.v.
Res.v.
Par.
0
1
2
3
2
2
U IPWCC,dr = U IPWCP,dr = U IPWAC,dr

2.52(0.074;0.226)
2.77(0.072;0.247)
0.56(0.011;0.046)
0.61(0.011;0.044)
6.23(0.197;0.636)
7.09(0.139;0.483)
53
Part II
Joint Models for the Longitudinal and Dropout Processes
54
Chapter 7
Sensitivity Happens
The Slovenian Public Opinion Survey

Bodyguards for MNAR models
55
7.1
The Slovenian Plebiscite
Rubin, Stern, and Vehovar (1995)

Slovenian Public Opinion (SPO) Survey
Four weeks prior to decisive plebiscite
Three questions:
1. Are you in favor of Slovenian independence ?
2. Are you in favor of Slovenias secession from Yugoslavia ?
3. Will you attend the plebiscite ?
Political decision: ABSENCENO
Primary Estimand: : Proportion in favor of independence
56
Slovenian Public Opinion Survey Data:

Independence
Attendance
Yes
Yes
Yes
1191
21
No
107
Yes
158
68
29
No
14
18
43
31
Yes
90
109
No
25
19
96
No
No
Secession
57
7.2
MAR in 3 Frameworks
Selection models
Pattern-mixture models
f (r i|y i, ) = f (r i|y oi, )
o
m o
f (y m
i |y i , r i , ) = f (y i |y i , )
Shared-parameter models
58
7.3
MAR in Selection Models
Diggle and Kenward (ApStat 1994)
f (r i|y i, ) = f (r i|y oi, )

Longitudinal data:
logit [P (Di = j|Di j, yij , yi,j1)] = 0 + 1yi,j1 + 2yij
2 = 0
2 = 0
1 = 2 = 0
MNAR
MAR
MCAR
No dependence on the future (NFD): built in

59
7.4
MAR in Pattern-mixture Models
Molenberghs, Michiels, Kenward, and Diggle (Statistica Neerlandica 1998)

Thijs, Molenberghs, Michiels, Verbeke, and Curran (Biostatistics 2002)
o
m o
f (y m
i |y i , r i , ) = f (y i |y i , )
For longitudinal data: ACMV: available case missing value restrictions:

t 2, s < t : f (yit|yi1, , yi,t1, di = s) = f (yit|yi1, , yi,t1, di t)
Practical implementation: doable!
)
n (
f (y , . . . , yi,s1)
ni d d i1
f (yit|yi1, , yi,t1, di = s) =
fd(ys|yi1, . . . , yi,s1)
d=s d fd (yi1 , . . . , yi,s1 )
d=s
60
7.5
Non-future Dependence in Pattern-Mixture Models
Kenward, Molenberghs, and Thijs (Biometrika 2003)
Within every pattern:

Past: Build a model for the observed data
Present, given past: For the rst unobserved time, given the past: Free
choice!
Future, given past and present: Use ACMV-type restrictions
Named NFMV: non-future missing values
Equivalence:
SeM: NFD
PMM: NFMV
61
7.6
MAR in Shared-parameter Models
Creemers, Hens, Aerts, Molenberghs, Verbeke, and Kenward (2008)
Conventional
Extended
MAR
f (y i|Xi, bi, ) f (r i|Xi, bi, )
o
f (y oi|g i, hi, j i, i)f (y m
i |y i , g i , hi , ki , mi ) f (r i |g i , j i , ki ni )
f (y oi|g i, hi, j i)f (y m

|y oi, g i, hi, ki)f (r i|g i, j i, ki)f (bi) dbi
i
f (y oi|g i, j i)f (r i|g i, j i)f (bi) dbi

=
o
f (y oi|g i, hi)f (y m
i |y i , g i , hi )f (bi ) dbi
f (y oi)
62
7.7
MAR in Shared-parameter Models
Extended
o
f (y oi|g i, hi, j i, i)f (y m
|y
i
i , g i , hi , ki , mi ) f (r i |g i , j i , ki ni )
f (y oi|g i, hi, j i)f (y m

|y oi, g i, hi, ki)f (r i|g i, j i, ki)f (bi) dbi
i
f (y oi|g i, j i)f (r i|g i, j i)f (bi) dbi

MAR
=
o
|y
f (y oi|g i, hi)f (y m
i , g i , hi )f (bi ) dbi
i
f (y oi)
Sub-class MAR
o
f (y oi|j i, i)f (y m
i |y i , mi )f (r i |j i , ni )
63
7.8
Every MNAR Model Has Got an MAR Counterpart
Molenberghs, Beunckens, Sotto, and Kenward (JRSSB 2008)

Fit an MNAR model to a set of incomplete data

Change the conditional distribution of the unobserved outcomes, given the
observed ones, to comply with MAR
Resulting new model has exactly the same t as the original MNAR model
The missing data mechanism has changed
This implies that denitively testing for MAR versus MNAR is not possible
64
7.9
MAR Counterpart to Pattern-mixture Models
b )
b = f (y io|r i, )
b f (r i|)
b f (y im|y io, r i, )
b
f (y io, y im, r i|,
b )
b
b )
b = f (y io|r i, )
b f (r i|)
b f (y im|y io, ,
h(y io, y im, r i|,
Starting from PMM is natural: clear separation into:
fully observable components
entirely unobserved component
65
7.10
MAR Counterpart to Selection Models
b )
b = f (y io, y im|)
b f (r i|y io, y im, )
b
b )
b
b )
b = f (y io|r i, ,
b )
b f (r i|,
b )
b f (y im|y io, r i, ,
b )
b = f (y io|r i, ,
b )
b f (r i|,
b )
b f (y im|y io, ,
b )
b
h(y io, y im, r i|,
66
7.11
MAR Counterpart to Shared-parameter Models
o
f (y io, y im, r i|bi) = f (y oi|g i, hi, j i, i) f (y m
|y
i
i , g i , hi , ki , mi ) f (r i |g i , j i , ki ni )
o
h(y io, y im, r i|bi) = f (y oi|g i, hi, j i, i) h(y m
i |y i , mi ) f (r i |g i , j i , ki ni )
with
o
h(y m
i |y i , mi ) =
g i hi ki
o
f (y m
i |y i , g i , hi , ki , mi )dg i dhi dki
67
7.12
Slovenian Public Opinion Survey: Counterpart Added
Model
Structure
d.f.
loglik
C.I.
bMAR
BRD1
(, )
-2495.29
0.892
[0.878;0.906]
0.8920
BRD2
(, j )
-2467.43
0.884
[0.869;0.900]
0.8915
BRD3
(k , )
-2463.10
0.881
[0.866;0.897]
0.8915
BRD4
(, k )
-2467.43
0.765
[0.674;0.856]
0.8915
BRD5
(j , )
-2463.10
0.844
[0.806;0.882]
0.8915
BRD6
(j , j )
-2431.06
0.819
[0.788;0.849]
0.8919
BRD7
(k , k )
-2431.06
0.764
[0.697;0.832]
0.8919
BRD8
(j , k )
-2431.06
0.741
[0.657;0.826]
0.8919
BRD9
(k , j )
-2431.06
0.867
[0.851;0.884]
0.8919
Model 10
(k , jk )
-2431.06
[0.762;0.893]
[0.744;0.907]
0.8919
Model 11
(jk , j )
-2431.06
[0.766;0.883]
[0.715;0.920]
0.8919
Model 12
(jk , jk )
10
-2431.06
[0.694;0.904]
0.8919
68
7.13
Slovenian Public Opinion Survey: Incomplete Data
Observed
BRD7 BRD7(MAR)
BRD9 BRD9(MAR):
BRD1 BRD1(MAR):
BRD2 BRD2(MAR):
1439
78
159
16
16
32
144
1381.6 101.7
8.1
1402.2 108.9
159.0
15.6
22.3
136
182.9
41.4
24.2
54
179.7 18.3
136.0
181.2 16.8
136.0
32.0
69
7.14
Slovenian Public Opinion Survey: Complete-data

Prediction
BRD1 BRD1(MAR):
1381.6 101.7
24.2 41.4
170.4 12.5
3.0 5.1
176.6 13.0
3.1 5.3
121.3 9.0
2.1 3.6
BRD2:
1402.2 108.9
15.6 22.3
147.5 11.5
13.2 18.8
179.2 13.9
2.0 2.9
105.0 8.2
9.4 13.4
BRD2(MAR):
1402.2 108.9
15.6 22.3
147.7 11.3
13.3 18.7
177.9 12.5
3.3 4.3
121.2 9.3
2.3 3.2
BRD7:
1439
16
78
16
3.2 155.8
0.0 32.0
142.4 44.8
1.6 9.2
0.4 112.5
0.0 23.1
BRD9:
1439
16
78
16
150.8 8.2
16.0 16.0
142.4 44.8
1.6 9.2
66.8 21.0
7.1 41.1
BRD7(MAR) BRD9(MAR):
1439
16
78
18
148.1 10.9
11.8 20.2
141.5 38.4
2.5 15.6
121.3 9.0
2.1 3.6
70
7.15
Slovenian Public Opinion Survey: Collapsed

(Marginalized) Predictions
BRD1 BRD1(MAR):
1849.9 136.2
32.4 55.4
b = 89.2%
BRD2:
1833.9 142.5
40.2 57.5
b = 88.4%
BRD2(MAR):
1849.0 142.0
34.5 48.5
b = 89.2%
BRD7:
1585.0 391.1
17.6 80.3
b = 76.4%
BRD9:
1799.7 152.0
40.7 82.3
b = 86.7%
BRD7(MAR) BRD9(MAR):
1849.9 136.3
30.4 57.4
b = 89.2%
71
7.16
Toenail Data: Unaected Nail Length
We opt for the following SPM:

E(Yij |gi, Ti, tj , ) = 0 + gi + 1Ti + 2tj + 3Titj
logit [P (Rij = 1|Ri,j1 = 0, gi, Ti, tj , )] = 0 + 01gi + 1Ti + 2tj + 3Titj
with
Yij : unaected nail length for subject i at occasion j
tj : time at which the jth measurement is made
Ti: treatment indicator for subject i
gi: normal random eect
72
Parameter estimates (standard errors):
Eect
Unaected nail length
Dropout
Parameter Estimate (s.e.)
Parameter Estimate (s.e.)
Mean structure parameters

Intercept
2.510 (0.247)
-3.127 (0.282)
Treatment
0.255 (0.347)
-0.538 (0.436)
Time
0.558 (0.023)
0.035 (0.041)
Treatment-by-time
0.048 (0.031)
0.040 (0.061)
Variance-covariance structure parameters

Residual variance
6.937(0.248)
Scale factor
Rand. int. variance
6.507 (0.630)
01
-0.076 (0.057)
2 2
01
0.038 (0.056)
73
Graphical representation of predictions for incomplete portions:

o
m
2
MNAR model: Y m
i |y i , gi N (Xi + Zi gi , Ii )
o
2
MAR counterpart: Y m
i |y i N (Xi , dJi + Ii )
(dashed lines)
(solid lines)
74
7.17
Toenail Data: Severity of Infection
gi1i2rt = g i1|g i2|i1gt r|g

Variable
Index
Complete rst measurement
i1
non-severe
severe
Incomplete last measurement
i2
non-severe
severe
Dropout indicator
dropout
completer
Treatment arm
standard
experimental
Latent class
class 0
class 1
75
7.18
Toenail Data: Severity of Infection
eg
g =
1 + e
i1|g
Model
i2|i1gt
e(0+1g)i1
=
1 + e0+1g
r|g
e(0+1i1+2g+3i1g+4t)i2
=
1 + e0+1i1+2g+3i1g+4t
e(0+1g)r
=
1 + e0+1g
Restriction
Mechanism
Bin1
1 = 0
MNAR
Bin1
Bin1(MAR)
Bin2
2 = 3 = 0
MAR
Bin2
Bin2(MAR)
Implication
76
Standard treatment
Completers
Experimental treatment
Dropouts
Completers
Dropouts
Observed data
77
10
79
11
42
42
Fit of Model Bin1

76.85
5.66
9.04
0.34
9.38
81.21
2.43
9.36
0.15
9.51
40.60
7.99
4.62
0.90
5.52
45.62
3.63
5.19
0.41
5.60
Fit of Model Bin1(MAR)

77.12
5.39
8.77
0.61
9.38
81.32
2.32
9.24
0.26
9.51
40.61
7.98
4.62
0.91
5.52
45.63
3.63
5.18
0.41
5.59
Fit of Model Bin2Bin2(MAR)

75.86
5.58
9.72
0.72
10.44
80.16
2.40
10.27
0.31
10.58
41.50
8.15
3.74
0.73
4.47
46.61
3.72
4.20
0.34
4.53
77
Chapter 8
A Latent-variable Mixture Model as a Basis for Sensitivity
Analysis in Incomplete Longitudinal Data
Depression trial
Model formulation
Parameter estimation
Analysis of the depression trial
Sensitivity analysis for the depression trial
78
8.1
Depression Trial
Clinical trial: experimental drug versus standard drug

170 patients
Response: change versus baseline in HAM D17 score
20
5 post-baseline visits: 48
-4
-6
Change versus Baseline
-8
-10
-20
-10
Change
10
-2
Standard Drug
Experimental Drug
6
=Visit
8
4
Visit
79
8.2
Latent Class Mixture Model
Latent variable: Qi = (Qi1, ..., Qig )

{
Qik =
1 subject i belongs to group k

0 otherwise
P (Qik = 1) = k prior probabilities with
k=1 k
=1
Measurement process: heterogeneity linear mixed model

(k)
Y i|qik = 1, bi N (Xi k + Zibi, i ) with

g
Shared parameter bi k=1 k N (k , Dk )
bi|qik = 1 N (k , Dk )
80
8.3
Dropout process: logistic regression for

gij (wij , bi, qik ) = P (Di = j|Di j, wij , bi, qik = 1)
= logit[gij (wij , bi, qik )] = wij k + bi
f (di
|qik = 1, bi)
di 1
[1 gij (wij , bi, qik )] if incomplete
gidi (widi , bi, qik )

j=2
=
ni
[1 gij (wij , bi, qik )]

if complete
j=2
81
8.4
Joint likelihood of measurement and dropout process:

f (y i, di) =
f (y i|qik = 1, bi)f (di|qik = 1, bi)fk (bi)dbi
k=1
Log-Likelihood function:
(|y, d)
=
i=1
ln
{ g
f (y i|, bi, qik = 1)f (di|, bi, qik = 1)fk (bi|)dbi
k=1
82
8.5
Maximization of Log-Likelihood
Latent variable Qi is considered missing

Incomplete data: (Y i, Di)
Complete data: (Y i, Di, Qi)
Joint density
fi(y i, di, Qi1 = qi1, . . . , Qig = qig ) =
[k fik (y i, di|, , )]qik
k=1
Log-likelihood
(|y, d, q) =
g
N
qik {ln k + ln fik (y i, di|, , )}
i=1 k=1
Maximization: EM algorithm
83
8.6
Analysis of Depression Trial
Measurement model:
(k)
Yij = 0 + 1 trti + 2 tj + 3 trti tj + 4 baselinei + 5 t2j + bi + ij

Dierent number of latent groups and dropout model
Model
Dropout Model
# Par
AIC
BIC
0,k + 1,k tj
10
4676.07
4696.08
4727.44
0,k + 1,k tj
14
4662.37
4690.37
4734.27
0,k + 1,k tj
18
4662.03
4698.03
4754.48
0,k + 1,k tj + bi
11
4669.12
4691.12
4725.61
0,k + 1,k tj + bi
15
4662.02
4692.02
4739.06
84
8.7
Model 2
Eect
Estimate
s.e.
p-value
Intercept : 0
23.17
3.75
< 0.0001
Treatment : 1
2.69
1.49
0.072
Time : 2
-6.18
1.18
< 0.0001
Time Treatment : 3
-0.52
0.24
0.028
Baseline : 4
-0.42
0.07
< 0.0001
Time Time : 5
0.41
0.10
< 0.0001
Measurement Error :
4.24
0.13
< 0.0001
Measurement Model
85
8.8
Eect
Estimate
s.e.
p-value
-8.58
3.57
0.009
0.83
0.44
0.056
Intercept Group 2 : 0,2
-1.35
1.28
0.292
Time Group 1 : 1,2
-0.05
0.20
0.793
-3.64
0.43
< 0.0001
Variance Shared Intercept : d
2.67
0.50
< 0.0001
Prior probability Group 1 : 1 =
0.48
0.10
< 0.0001
Dropout Model
Intercept Group 1 : 0,1
Time Group 1 : 1,1
Shared Eects
Mean Shared Intercept Group 1 : 1
Log-likelihood
-2331.18
86
8.9
Analysis of the Depression Trial
79 patients in group 1: lower and more completers
20
10
0
-20
-10
0
-10
-20
10
20
91 patients in group 2: higher and more dropout
Visit
Visit
87
8.10
Acute
Chronic
proles over time

Association between latent groups and dropout pattern
age
baseline value
treatment
Independence between latent groups and
gender
origin
88
8.11
The Depression Trial: How Sure is Classication?

i1
Classication
# patients
0.80 1.00
Clearly Group 1
61
0.60 0.80
Group 1
0.55 0.60
Doubtful, more likely Group 1
0.45 0.55
Uncertain
0.40 0.45
Doubtful, more likely Group 2
0.20 0.40
Group 2
19
0.00 0.20
Clearly Group 2
64
There are only 8/170 in the [0.45; 0.55] range
89
8.12
The Depression Trial: Sensitivity Analysis
Latent-class mixture model (Model 2)

Shared-parameter model (Model 4)
Pattern-mixture model (pattern-specic intercepts and slopes)
An MAR selection model (Diggle & Kenward, 1994)
logit [P (Di = j | Di j, Yi,j1, Yij )] = 0 + 1Yi,j1
An MNAR selection model (Diggle & Kenward, 1994)
logit [P (Di = j | Di j, Yi,j1, Yij )] = 0 + 1Yi,j1 + 2Yij
90
8.13
The Depression Trial: Sensitivity Analysis
Treatment at Endpoint
Treatment Time
Model
Estimate
s.e.
p-value
Estimate
s.e.
p-value
LCMM
-1.44
0.91
0.114
-0.52
0.23
0.028
SPM
-1.69
0.93
0.069
-0.50
0.24
0.035
PMM
-2.01
1.20
0.096
-0.55
0.31
0.077
MAR SeM
-2.17
1.25
0.082
-0.58
0.32
0.068
MNAR SeM
-2.16
1.24
0.081
-0.57
0.31
0.068
91
8.14
Conclusion: Correspondence Between Model Families
Molenberghs, Michiels, Kenward, and Diggle (Statistica Neerlandica 1998)

Kenward, Molenberghs, and Thijs (Biometrika 2003)
NFD
general MNAR
NFMV
general MNAR
=
interior
: MCAR Theorem 1 Theorem 2 general MNAR
Subfamily 1 Subfamily 2
MAR
ACMV
SPM
: MCAR
PMM : MCAR
SeM
92
8.15
Conclusion: Counterparts to Models
Molenberghs, Beunckens, Sotto, and Kenward (JRSSB 2008)

Verbeke and Molenberghs (2008)
MNAR model
MAR model:
Observed data: same t

Unobserved data given observed data: MAR prediction
Holds more generally with all unobservables / latent structures.
93
Part III
Joint Models for Longitudinal and Time-to-event Data
94
Chapter 9
Longitudinal Studies with Time-to-event Primary Endpoints
In many longitudinal studies we collect dierent types of outcomes

Time-to-event outcomes
time until an event of particular interest occurs (e.g., death, relapse, dropout)
Covariates
baseline: treatment, gender, family history, etc.
time-dependent: treatment dose, exposure (e.g., smoking), longitudinal
markers (e.g., blood values)
95
AIDS Data Set: longitudinal study on 467 HIV infected patients who had failed
or were intolerant of zidovudine therapy
Outcomes
randomized treatment: didanosine (ddI) and zalcitabine (ddC)
time to death
longitudinal measurements of CD4 cell counts (baseline, 2, 6, 12 & 18 months)
Interest could be either on the longitudinal or event time outcome

focus on the survival outcome: what is the eect of CD4 cell count on the
time to death
96
We have already seen how we can analyze longitudinal outcomes

mixed eects models, GEE, . . .
When focus on the event time outcome

predominant relative risk models (Cox Model)
special care is required when dealing with time-dependent covariates
97
9.1
A (Brief) Review of Relative Risk Models
The most important complication in the analysis of time-to-event outcomes is

Censoring partial information for the event times
Notation (i denotes the patient)

Ti true time-to-event
Ci the censoring time (e.g., the end of the study or a random censoring time)
Available data for each patient
observed event time: Ti = min(Ti, Ci)
event indicator: i = 1 if event; i = 0 if censored
98
Relative Risk Models assume a multiplicative eect of covariates in the hazard

scale, i.e.,
hi(t) = h0(t) exp(1wi1 + 2wi2 + . . . + pwip)
log hi(t) = log h0(t) + 1wi1 + 2wi2 + . . . + pwip
where
hi(t) denotes the hazard for an event for patient i at time t
h0(t) denotes the baseline hazard
wi1, . . . , wip a set of covariates
99
The baseline hazard h0(t) represents the hazard for an event when all the
covariates or all the s are 0
That is, h0(t) represents the instantaneous risk of experiencing the event at time
t, without the inuence of any covariate
Therefore,
if a covariate has a benecial eect, decreases h0(t)
if it has a harmful eect, increases h0(t)
<0
>0
100
Cox Model:
we make no assumptions for the baseline hazard function
Parameter estimates and standard errors are based on the log partial likelihood
function
n
[
{
]
}
exp( wi)
p() =
i ( wi) log
i=1
Tj Ti
where only patients who had an event contribute
101
Chapter 10
Time-dependent Covariates
One of the basic assumptions of the Cox model is the Proportional Hazards
assumption
In practice, PH means that the eect of a covariate in the risk for an event is
constant over time
However, PH is not always reasonable

in the AIDS data set we have repeated measurements of the CD4 cell count,
which is a marker for the condition of the immune system
102
10
8
6
CD4
4
2
Time
103
10
8
6
CD4
Time to
Death
Time
103
10
6
Time to = 2.9 months

Death
CD4
Patient Died
Time
103
In the AIDS data set we have repeated measurements of the CD4 cell count,
which is a marker for the condition of the immune system
As the CD4 cell count starts deteriorating there is a higher risk for event
Possible research questions of interest:

how strong is the association between the value of CD4 cell count at t and the
risk for an event at the same time point?
what is the treatment eect on survival after adjusting for the CD4 cell count?
is CD4 cell count a good surrogate for the time-to-death?
...
104
To answer these questions we need a model that relates

the longitudinal evolution of CD4 cell count with the
time-to-event
105
10.1
External versus Internal Time-Dependent Covariates
There are two types of time-dependent covariates

External (a.k.a. exogenous): the value of the covariate at time point t is not
aected by the occurrence of an event at time point u, with t > u
Internal (a.k.a. endogenous): not External
In practice, we can distinguish between the two types using the concept of
predictability
a predictable process is one whose value at any time t is known innitesimally
before t
106
External time-dependent covariates are predictable

Internal are not
Examples
in the time-dependent treatment dose example
* the dose for time t is known (based on past values)
* the dose for time t is not related to the failure status
* knowing the dose at time t does not mean that the patient is still alive at t
107
in the CD4 cell count example,

* the CD4 count at time t is not available (based on past values)
* the CD4 count at time t is directly related to the failure status
* being able to measure CD4 cell count at time t means that the patient is
still alive at time t
It is very important to distinguish between these two types of time-dependent

covariates, because it determines the type of analysis that it should be followed
If we treat internal covariates as external, we may produce spurious results
108
10.2
Handling External Covariates
The Cox model can be easily extended to handle external time-dependent

covariates
hi(t) = h0(t) exp{ wi + mi(t)}
where
wi denotes the baseline covariates as we had so far
mi(t) denotes the value of the time-dependent covariate at time t
quanties the eect of this covariate at time t to the hazard for an event at
the same time point
109
Note: the time-dependent Cox model relaxes the PH assumption

The hazard ratio is
hi(t)
= exp{ wi + mi(t)}
h0(t)
which is not constant in time
When we want to t the Cox model taking into account the eect of external
time-dependent covariates we need to use the counting process formulation
this is a rather technical subject which we will not describe in detail here
110
10.3
Handling Internal Covariates
Let us turn our attention to internal covariates

what is exactly the problem with such covariates and why we cannot use the
time-dependent Cox model
Internal covariates have a stochastic nature

we do not have the complete history of past values available at any time point
they contain measurement error
111
10
Values measured
with error
CD4
Event Time
Underlying
true evolution
Time
112
Chapter 11
Joint Models for Longitudinal and Time-to-Event Data
To deal with these features of internal covariates a special class of models has
been developed
Joint Models for Longitudinal and Time-to-Event Data
113
11.1
Formulation
Step 1: lets assume that we know mi(t), i.e., the true & unobserved value of
CD4 cell count at any time t
Then, we can dene a standard Proportional Hazards model

hi(t | Mi(t)) = h0(t) exp{ wi + mi(t)},
where
Mi(t) = {mi(s), 0 s < t} true CD4 cell count history
quanties the eect of CD4 cell count on the hazard for death
114
Step 2: using the observed longitudinal responses yi(t) reconstruct the

covariate history Mi(t) for each subject
Mixed-eects model (time-dependent nature)

yi(t) | bi = mi(t) + i(t)
= x
i (t) + z i (t)bi + i (t),
i(t) N (0, 2)
where
xi(t) and : xed-eects part
z i(t) and bi: random-eects part
115
Step 3: the two processes are associated dene a model for their
joint distribution
Joint models for such joint distributions are of the following form
f (y i, Ti, i) =
f (y i | bi)
hi(Ti | bi) S(Ti | bi)

i
f (bi) dbi
where
y i: longitudinal measurements;
Ti: time-to-event;
i: event indicator
bi a vector of random eects that explains the interdependencies

f () density function;
S() survival function
116
Key assumption: full conditional independence given the random eects bi

the longitudinal outcome is independent of the time-to-event outcome
the repeated measurements in the longitudinal outcome are independent of
each other
f (y i, Ti, i | bi) = f (y i | bi) f (Ti, i | bi)
f (y i | bi) =
f {yi(tij ) | bi}
117
10
Patient i
Patient j
CD4
End of the Study
Time
118
Time to
AIDS
CD4
10
End of the Study
Time to
AIDS
0
Time
118
Time to
=?
AIDS
CD4
10
End of the Study
Time to
= 2.9 months
AIDS
0
Time
118
6
4
CD4
10
End of the Study
Patients who show steep decrease

are more likely to dropout
Time
118
15
5
10
End of the Study
Patients who show steep increase

are more likely to dropout
Time
118
15
5
10
End of the Study
Random Effects
Time
118
11.2
Details in Joint Models Specication
The survival function, which is a part of the likelihood of the model, depends on
the whole longitudinal history
( t
)
Si(t | bi) = exp
h0(s) exp{ wi + mi(s)} ds
0
Therefore, care in the denition of the design matrices of the mixed model:
capture possible nonlinearities
Random-eects distribution bi N (0, D)

many authors have reported joint models are rather robust to misspecication
this is especially the case as ni increases
119
Assumptions for the baseline hazard function h0(t)

parametric possibly restrictive
unspecied within JM framework underestimates standard errors
It is advisable to use parametric but exible models for h0(t)
splines
step-functions: piecewise-constant baseline hazard often works satisfactorily
h0(t) =
q I(vq1 < t vq )
q=1
where 0 = v0 < v1 < < vQ denotes a split of the time scale
120
Dierent parameterizations
time-dependent:
hi(t | bi) = h0(t) exp{ wi + mi(t)}
| {z }
x
i (t) + z i (t)bi
easily interpretable in high-dimensional random-eects structures
can be extended to time-dependent derivative: mi(t) = mi(t)/t
random eects:
hi(t | bi) = h0(t) exp( wi + bi)
time-independent
not easily interpretable in high-dimensional random-eects structures
121
11.3
Estimation
Estimation of joint models is a computationally challenging task

computation of the log-likelihood entails integration wrt the random eects
() =
log
f (y i | bi; )
hi(Ti | bi; ) Si(Ti | bi; )

i
f (bi; ) dbi
computation of the survival function entails integration wrt time

( t
)
Si(t | bi) = exp
h0(s) exp{ wi + mi(s)} ds
0
122
Combination of optimization and numerical integration
Numerical integration
Gaussian quadrature
Monte Carlo
Laplace (especially useful in high-dimensional settings)
Optimization
EM algorithm
Newton-Raphson or quasi-Newton algorithm
hybrid algorithms (combination of EM & quasi-Newton)
123
11.4
Joint Modeling Analysis of the AIDS Data Set
Patient longitudinal trajectories

0
ddC
12
18
ddI
Kaplan-Meir estimate for the time to

death
2
1
0
2
12
18
ddC
ddI
0.2
0.4
0.6
0.8
Months
1.0
0.0
CD4
10
Months
15
20
124
Longitudinal sub-model
xed eects: time eect + interaction of treatment with time
random eects: intercept + time eect
Survival sub-model
treatment eect + underlying CD4 cell count eect
piecewise-constant baseline hazard in 7 intervals
125
Survival
Longitudinal
value (s.e.)
Random Eects
value (s.e)
value
Treat
0.35 (0.15)
Inter
2.56 (0.04)
Inter
0.87
CD4
1.10 (0.12)
Time
0.04 (0.005)
Time
0.04
Treat:Time
0.01 (0.01)
0.07
0.38
a very strong association between the underlying CD4 cell count at time t, and
the risk for death at t
analysis was done with the R package JM freely available from
http://cran.r-project.org more info available at:
http://rwiki.sciviews.org/doku.php?id=packages:cran:jm
126
11.5
A Comparison of JM versus naive TD Cox
To illustrate the virtues of joint modeling, we compare with the standard

time-dependent Cox model
i.e., we ignore the measurement error in the CD4 cell count
Joint Model
Naive TD Cox
value (s.e.)
value (s.e.)
Treat
0.35 (0.15)
0.33 (0.15)
CD4
1.10 (0.12)
0.72 (0.08)
Clearly, there is a considerable eect of ignoring the measurement error, especially

for the eect of CD4!
127
10
6
CD4
Event Time
Joint Model
timedependent Cox
Time
128
Chapter 12
Connection with Missing Data
So far we have attacked the problem from the survival point of view
However, some times, we may also have interest on the longitudinal outcome
Issue: when patients experience the event, they dropout from the study
a direct connection with the missing data area
Dropout must be taken into account when deriving

inferences for the longitudinal outcome
129
12.1
Missing Data Mechanism for Joint Models
To show this connection more clearly

Ti: true time-to-event
y oi: longitudinal measurements before Ti
ym
:
longitudinal
measurements
after
T
i
i
Important to realize that the model we postulate for the longitudinal responses
is for the complete vector {y oi, y m
i }
implicit assumptions about missingness
130
Missing data mechanism:

f (Ti | y oi, y m
i )=
f (Ti | bi) f (bi | y oi, y m

i ) dbi
still depends on y m
i , which corresponds to nonrandom dropout
Intuitive interpretation: patients who dropout show dierent longitudinal
evolutions than patients who do not
observed data do not constitute a random sample from the target population
this feature considerably complicates the validation of the joint models
assumptions using standard residual plots
131
12.2
A Comparison of MAR versus MNAR for the AIDS

Data
The strong association suggests nonrandom dropout a comparison between

linear mixed eect model MAR
joint model MNAR
is warranted
MAR assumes that missingness depends only on the observed data
o
f (Ti | y oi, y m
)
=
f
(T
|
y
i
i
i)
132
The strong association suggests nonrandom dropout

as sensitivity analysis we compare MAR versus MNAR
LMM (MAR)
JM (MNAR)
value (s.e.)
value (s.e)
Inter
2.55 (0.037)
2.51 (0.043)
Time
0.04 (0.005)
0.04 (0.004)
0.01 (0.007)
0.01 (0.006)
Treat:Time
Minimal sensitivity in parameter estimates & standard errors
Warning:
this does not mean that this is always the case!
133
12.3
Problems with Residuals
Standardized marginal residuals

1/2
ri = V i (y i X i),
+ 2I
V i = Z iDZ
i
they predict the marginal errors y i X i = Z ibi + i
Systematic trend explained

deterioration of CD4 cell count implies higher risk for death
for small tted values we mainly observe patient who did not die
however, the joint models tted average accounts for dropout
134
Residuals
Marginal Residuals vs Fitted Values
1.8
2.0
2.2
Fitted Values
2.4
135
Chapter 13
Further Topics
Joint modeling of longitudinal and time-to-event data is a very active area of

current Biostatistics research
several extensions of the standard joint model have been already proposed
Some of these are
recurrent events & competing risks
more than one longitudinal outcomes
semiparametric modeling of longitudinal proles
semiparametric modeling of random-eects distribution
136
13.1
Dynamic Predictions
Increasing interest in individualized predictions
Within the Joint Modeling framework
Dynamic Prediction:
estimate the probability for an event for a specic patient

update probabilities for an event dynamically as longitudinal information is
recorded
Example: we will consider Patient 7 that has provided us with 4 measurements

of CD4 cell count
137
More formally, we have available measurements up to time point t

Yi(t) = {yi(s), 0 s t}
and we are interested in
{
}
i(u | t) = Pr Ti u | Ti > t, Yi(t), Dn

where
where u > t, and
Dn denotes the sample on which the joint model was tted
138
It is convenient to proceed using a Bayesian formulation of the problem. More

specically, i(u | t) can be written as
{
}
Pr Ti u | Ti > t, Yi(t), Dn =
Pr
Ti
u|
Ti
> t, Yi(t); f ( | Dn) d
The rst part of the integrand reduces to

Pr
Ti
Ti
u|
{
}
Si u | Mi(u, bi, );
{
} f (bi | Ti > t, Yi(t); ) dbi
Si t | Mi(t, bi, );
> t, Yi(t); =
139
0.8
0.6
0.4
0.0
0.2
0.4
0.2
0.0
Time
12
18
Time
12
18
0.8
0.6
0.2
0.0
0.0
0.2
0.4
0.6
0.8
1.0
Subject 7
1.0
Subject 7
0.4
Pr(Ti u | Ti > t, ~
y i(t), Dn)
0.6
0.8
1.0
Subject 7
1.0
Subject 7
Time
12
18
Time
12
18
140
Patient 20
12
Patient 7
4.0
CD4
3.5
3.0
2.5
2.0
12
Months
141
0.6
0.7
0.8
0.9
1.0
Patient 7
extra 2 months survival
Patient 7
Patient 7
Patient 20
Patient 20
Patient 20
after 12 m
after 6 m
after 2 m
baseline
after 12 m
after 6 m
after 2 m
baseline
0.6
0.7
0.8
0.9
1.0
0.6
0.7
0.8
0.9
1.0
Survival Probability
142
Part IV
Joint Models for Multivariate Longitudinal Data
143
Chapter 14
Random-eects Models for High-dimensional Multivariate
Longitudinal Data
Examples
A random-eects model
A pairwise model tting approach
Applications
144
14.1
Example 1: Hearing Data
Threshold sound pressure levels (dB), on both ears,

11 frequencies: 125 8000 Hz
Observations from 603 males, with up to 15 obs./subject.
603
145
Research questions:
Is the relation between hearing loss and age the same for all frequencies?
How are subject-specic evolutions for the dierent frequencies related?
146
14.2
Example 2: Fitness Data
Intervention study on 105 elderly participants

Randomization:
classical tness: 3 weekly visits to gym
distance coaching program with emphasis on incorporating physical activities in
daily life
Aim is to study the eect on psycho-cognitive functioning
147
Psycho-cognitive functioning: 106 dichotomised items, 7 dierent questionnaires,

each measuring a latent component of psycho-cognitive functioning:
1. Physical well-being (10)
2. Psychological well-being (14)
3. Self-esteem (10)
4. Physical self-perception (30)
5. Degree of opposition to physical activities (21)
6. Perceived self-ecacy towards physical activity (5)
7. Motivation for intervention program (16)
148
Research questions:
Is there an overall treatment eect?
How are the various components of psycho-cognitive functioning associated?
149
14.3
A Random-eects Model
We now consider the setting of modeling multivariate longitudinal data

Let Y1i(t), . . . , Ymi(t) be the m outcomes measured on subject i, at time point t
Outcomes can be of dierent types:
continuous
binary
counts
...
150
Our requirements:
Inferences for original outcomes
Direct marginal inferences
Separate univariate models are
implied by multivariate model
Dierent types of outcomes possible
= Random-eects approach
}
No restriction on dimensionality
= Computational problem !
151
As an example, re-consider the hearing data:

Linear mixed model for each outcome separately:
Yi(t) = (1 + 2 Fagei + 3 Fage2i + ai)
+ (4 + 5 Fagei + bi) t + 6 visit1(t) + i(t)
Joint model:
Y1i(t) = 1(t) + a1i + b1it + 1i(t)
Y2i(t) = 2(t) + a2i + b2it + 2i(t)
..
Y (t) = (t) + a + b t + (t)

22i
22
22i
22i
22i
152
Distributional assumptions:
(a1i, a2i, . . . , a22i, b1i, b2i, . . . , b22i) N (0, D4444)
(1i(t), 2i(t), . . . , 22i(t)) N (0, 2222) , for all t
Full multivariate joint model:

44 44 covariance matrix for random eects
22 22 covariance matrix for error components
990 + 253 = 1243 covariance parameters
= Computational problems!
153
As an example, re-consider the tness data:

Random-eects logistic regression for each outcome:
logit{P (Yij = 1)} = + DCi + bi
Joint model:
logit{P (Yij1 = 1)} = 1 + 1DCi + bi1
logit{P (Yij2 = 1)} = 2 + 2DCi + bi2
..
logit{P (Y = 1)} = + DC + b
ij7
7
7
i
i7
154
Distributional assumptions:
(bi1, bi2, . . . , bi7) N (0, D77)
Full multivariate joint model:

Only (?) 28 parameters in covariance matrix
Numerical integration over 7-dim. random-eects distribution!
= Computational problems!
155
14.4
A Pairwise Model Fitting Approach
General idea:
Estimation of all parameters does not require tting the full multivariate model
It is sucient to t the implied model for all pairs, i.e., all bivariate models
Fit all bivariate models:
(Y1, Y2), (Y1, Y3), . . . , (Y1, Ym), (Y2, Y3), . . . , (Y2, Ym), . . . , (Ym1, Ym)
Straightforward using standard software (e.g., SAS)
Equivalent to maximizing pseudo (log-)likelihood:
p() = (Y1, Y2|1,2) + (Y1, Y3|1,3) + . . . + (Ym1, Ym|m1,m)
Fieuws & Verbeke, Biometrics 2006

156
p,q is the vector of parameters in the bivariate model for (Yp, Yq )

Let be the vector obtained from stacking all p,q
Asymptotic properties (from pseudo likelihood theory):
b ) M V N (0, J 1KJ 1)
N (
J and K consist of rst and second-order derivatives of p.
Some of the p,q contain the same parameters.
Estimates for these parameters are obtained by averaging pair-specic estimates
Inference is based on:
b A) M V N (0, AJ 1KJ 1A)

N (A
157
Properties of pairwise estimators:

Consistent, high agreement with MLE (ICC > 0.95)
Correct estimation of sampling variability, corrected for misspecied association
structure
Relative eciency versus MLE:
In general, minor loss of eciency (RE > 0.9), unless with shared parameters
RE independent of number of outcomes
158
14.5
Example 1: Hearing Data
Example: Interaction between the linear time eect and age.

Estimates and standard errors:
210 = 90.4, p < 0.0001
210 = 110.9, p < 0.0001
159
Association between underlying random eects: D4444 of interest

PCA on correlation matrix of random slopes, left side:
160
14.6
Example 2: Fitness Data
Treatment eects (s.e.s) from univariate and multivariate models:
Physical well-being
Psychological well-being
Self-esteem
Physical self-perception
Degree of opposition
Self-ecacy
Motivation
p < 0.05
Univariate
Models
0.13 (0.37)
1.22 (0.61)
0.43 (0.42)
0.58 (0.24)
0.06 (0.24)
0.24 (0.33)
0.35 (0.16)
Multivariate
Model
0.12 (0.37)
1.00 (0.68)
0.49 (0.39)
0.52 (0.25)
0.07 (0.24)
0.22 (0.33)
0.34 (0.16)
161
No gain from multivariate analysis if interest is in inferences for each outcome

separately.
Wald test for overall treatment eect: 26 = 16.66, p = 0.011
Correlation matrix (with variances) of random intercepts:
Physical well-being:
2.55
Psychological well-being: 0.75
4.41
Self-esteem:
0.55
0.76
3.43
Physical self-perception: 0.66
0.46
0.53
1.83
Degree of opposition:
0.19
0.12
0.23
0.38
1.16
Self-ecacy:
0.29
0.24
0.25
0.36
0.23
1.33
Motivation:
0.42
0.31
0.28
0.40
0.47
0.30
0.33
162
Special association models:
logit{P (Yij1 = 1)} = 1 + 1DCi + bi1

Original model:
logit{P (Y = 1)} = + DC + b
ij7
7
7
i
i7
logit{P (Yij1 = 1)} = 1 + 1DCi + bi
logit{P (Y = 1)} = + DC + b
ij2
2
2
i
2 i
(dev. = 533.7)
Special case 1:
logit{P (Yij7 = 1)} = 7 + 7DCi + 7bi
logit{P (Yij1 = 1)} = 1 + 1DCi + bi

Special case 2:
logit{P (Y = 1)} = + DC + b
ij7
7
7
i
i
(dev. = 758.4)
163
Part V
Case Study: A Joint Model for Multivariate Longitudinal
Data and a Time to Event
164
Chapter 15
Renal Transplant Data
Introduction of the data

A joint model
Results from the analyses
Conclusions
165
15.1
Introduction of the Data
Patients with kidney transplant between 1983 and 2000 at U.H.Leuven

Clinical interest:
Continuous prediction of long-term success of graft (> 10 years)
Conditional on:
not losing graft during rst year
not dying in the rst 10 years for reasons not related to transplantation.
166
Information: 949 patients, with 1-78 visits per patient

341 patients with functioning graft after 10 years
91 patients with a graft failure before 10 years
517 patients with functioning graft, but FU < 10 yrs
Prediction based on longitudinal measurements of:
Haematocrit Level
Filtration Rate
Proteinuria
Blood Pressure
167
Haematocrit level:
Non-failures
Failures
168
Glomerular ltration rate (GFR):
Non-failures
Failures
169
Presence of proteinuria:
Non-failures
Failures
170
Mean blood pressure:
Non-failures
Failures
171
Let Fi be the time of renal failure and let Y t

i be all the longitudinal information
gathered up to and including time point t.
Aim of the analysis:
Hi(t) = Pi(t Fi 120 | y t

i ),
Specication of conditional distribution for (Fi | Y t

i ) problematic due to:
Unbalanced nature of the longitudinal data
The dierent outcome types in Y t
i
The running time t
172
15.2
A Joint Model for Fi and Y t

i
A pattern-mixture modeling approach will be followed:

f (Fi, Y i) = f (Y i|Fi)f (Fi)
Joint model for Fi and Y t
i obtained from marginalizing the above
Bayes rule to predict failure, conditional on Y t
i :
Hi(t) = Pi(t Fi 120 | y t
i )
=
fi (y t
i
fi (y t
i | t Fi 120)P (Fi 120|Fi t)
| t Fi 120)P (Fi 120|Fi t) + fi (y t
i | Fi > 120)P (Fi > 120|Fi t)
173
Components to be specied:
fi(y t
i
fi(y t
i | t Fi 120)P (Fi 120|Fi t)
| t Fi 120)P (Fi 120|Fi t) + fi(y t
i | Fi > 120)P (Fi > 120|Fi t)
Prior probabilities P (Fi 120|Fi t) and P (Fi > 120|Fi t)

Model fi(y t
i | Fi > 120) for non-failures
Model fi(y t
i | t Fi 120) for failures
Prior probabilities can be obtained from any standard survival analysis.
174
Kaplan-Meier estimate for time to renal failure:
t
For the longitudinal models fi(y t
i | Fi > 120) and fi (y i | t Fi 120), a
random-eects approach is followed, allowing separate model building for each
outcome.
175
Models for the non-failures: fi(y t

i | Fi > 120):
Model for all measurements up to 10 yrs, and assumed independent of Fi
Haematocrit:
Y1i(t) = 01 + b01i + (11 + b11i)t + 1i(t)
GFR:
Y2i(t) = 02 + b02i + (12 + b12i)t + 2i(t)
Proteinuria:
logit{P (Y3i(t))} = 03 + b03i + 13t
Mean Blood Pressure:
Y4i(t) = 04 + b04i + (14 + b14i)t + 4i(t)
176
Models for the failures: fi(y t

i | t Fi 120):
Models will depend on actual failure time Fi
In order to be able to include Fi as a covariate, use approximation:
fi(y t
i | t Fi 120)
=
119
fi(y t
i | k Fi k + 1)P (k Fi k + 1)/P (t Fi 120)
k=t
119
k=t
1
fi(y t
|
F
=
k
+
) P (k Fi k + 1)/P (t Fi 120)
i
i
{z
}
2 |
from prior probabilities
177
Haematocrit:
Y1i(t) = 01 + 01Fi + b01i + (11 + 11Fi + b11i)t + 1i(t)
GFR:
+ b + [t ] + (t) if t
0
02i
12
2
2i
2
Y2i(t) =
+ b + [t ] + (t) if t >
0
02i
32
2
2i
2
with 0 = 02 + 02Fi and 2 = 22 + 22Fi
Proteinuria:
logit{P (Y3i(t))} = 03 + 03Fi + b03i + (13 + 13Fi)t
Y4i(t) = 04 + 04Fi + b04i + (14 + 14Fi + b14i)t + 4i(t)
178
Summary of the mixed models:
Non-Failures
Failures
Haematocrit:
LMM (2)
LMM (2)
GFR:
LMM (2)
NLMM (1)
GLMM (1)
GLMM (1)
LMM (2)
LMM (2)
Proteinuria:
2 mixed models with many random eects (7 & 6)

computational diculties
pairwise model tting approach
179
15.3
Results from the Analyses
L.R. test for no association in each pairwise model:

Markers
1,2:
1,3:
1,4:
2,3:
2,4:
3,4:
Non-failures
dev. #
p
82.8
4
< 0.0001
18
2
0.0001
6.4
4
0.17
22.4
2
< 0.0001
8.1
4
0.09
7.4
2
0.025
Failures
dev. #
p
18.9
2
< 0.0001
7.2
2
0.027
2.6
4
0.63
0.3
1
0.58
0.1
2
0.95
6.1
2
0.047
Does the association matter for our classication purposes ?
180
In order to validate the classication, the dataset is split in a training dataset and
a validation dataset (50% of patients)
Models tted based on training dataset
Posterior probabilities calculated for all patients in the validation dataset
Each time new information on the outcomes of a patient becomes available, the
posterior probability Hi(t) = Pi(t Fi 120 | y t
i ) for that patient to fail can
be updated.
We rst consider prediction based on GFR only
181
Illustration for one patient from the validation dataset:
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
182
This can be repeated for all failures in the validation dataset, and the median
posterior probability, based on GFR can be calculated:
183
Failure is not detected until very close to the actual failure time.
If failure is to be predicted based on all 4 markers, various strategies could be used:
Decision based on highest posterior probability
Joint model assuming uncorrelated markers
Joint model allowing markers to be correlated
Lets evaluate them all, for failures as well as non-failures separately.
We rst consider the failures:
46 patients in validation set who fail
Median posterior probabilities to fail within the remaining period
As a function of time: years before failure
184
Median posterior probability for failures, based on Haematocrit only:
185
Median posterior probability for failures, based on GFR only:
186
Median posterior probability for failures, based on Proteinuria only:
187
Median posterior probability for failures, based on Mean Blood Pressure only:
188
Median posterior probability for failures, based on highest posterior probability:
189
Median posterior probability for failures, based on model with uncorrelated

markers:
190
Median posterior probability for failures, based on model with correlated markers:
191
The full multivariate model allowing for correlation between the markers
considerably improves early detection of failure.
This should not be at the expense of many false positives.
We therefore perform the same validation exercise for the non-failures:
171 patients in validation set who do not fail
Median posterior probabilities to fail within the remaining period
As a function of time: years since transplantation
192
Median posterior probability for non-failures, based on Haematocrit only:
193
Median posterior probability for non-failures, based on GFR only:
194
Median posterior probability for non-failures, based on Proteinuria only:
195
Median posterior probability for non-failures, based on Mean Blood Pressure only:
196
Median posterior probability for non-failures, based on highest posterior

probability:
197
Median posterior probability for non-failures, based on model with uncorrelated

markers:
198
Median posterior probability for non-failures, based on model with correlated

markers:
199
15.4
Conclusions
Discriminant analysis based on many outcomes, measured longitudinally, in an

unbalanced design, is technically possible
Allowing the longitudinal markers to be correlated considerably improves
predictions
This requires joint modeling of all longitudinally measured markers and failure time
A pattern-mixture approach allows for continuous updating of posterior
probabilities
Various mixed models can be combined and tted using pairwise tting approach
200

Joint Modeling Approaches Using Random Effects

Cargado por

Información del documento

Derechos de autor

Formatos disponibles

Compartir este documento

Compartir o incrustar documentos

Opciones para compartir

¿Le pareció útil este documento?

¿Este contenido es inapropiado?

Copyright:

Formatos disponibles

Joint Modeling Approaches Using Random Effects

Cargado por

Copyright:

Formatos disponibles

Joint Modeling Approaches in Longitudinal Studies

Using Random Eects

Department of Biostatistics, Erasmus Medical Center, Rotterdam, the Netherlands

JSM, Vancouver, BC, August 3, 2010

After this course participants will be able to identify settings in which

Motivation for Joint Modeling

Outcomes in Longitudinal Studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Approaches to Simultaneously Analyze Multiple Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

Motivating Datasets and Key Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

Generalized Estimating Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

The Generalized Linear Mixed Model (GLMM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

Inverse Probability Weighting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

Joint Models for the Longitudinal and Dropout Processes

Joint Models for Longitudinal and Time-to-event Data

Longitudinal Studies with Time-to-event Primary Endpoints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

Time-dependent Covariates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102

Joint Models for Longitudinal and Time-to-Event Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

Connection with Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

Further Topics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

Joint Models for Multivariate Longitudinal Data

Random-eects Models for High-dimensional Multivariate Longitudinal Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144

Renal Transplant Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Longitudinal Data Analysis, Chapman & Hall/CRC Handbooks of Modern

Springer Series in Statistics and Computing. New-York: Springer.

Often in longitudinal studies dierent types of outcomes are collected

Example: patients in the waiting list for primary renal transplantation

Example: a longitudinal study on HIV infected patients

Depending on the questions of interest, dierent types of statistical analysis are

Focus on multiple outcomes

Until recently emphasis has been

However, recently there has been an explosion of statistics and biostatistics

Many dierent approaches have proposed that

What Is This Course About?

Purpose of this course is to present the state of the art in

In particular, we will present

Caveat: we will not present all available joint modeling approaches

Introduction and notation

Introduction and Notation

General idea: Specify f (y1, y2) directly

General idea: Factorize f (y1, y2) as

For example, based on f (y1, y2) = f (y1|y2)f (y2), E(Y1) requires

y1f (y1|y2) dy1 f (y2) dy2,

E(Y1) not necessarily of the same parametric form as E(Y1|Y2)

General idea: Latent variables, shared by Y1 and Y2 imply associations

f (y1|b)f (y2|b)f (b) db

Implied association structure (s = t):

= Corr{Y1(s), Y1(t)} Corr{Y2(s)Y2(t)}

for some pre-specied function g(b).

General idea: Separate but correlated latent variables for Y1 and Y2

f (y1|b1)f (y2|b2)f (b1, b2) db1db2

Corr{Y1(s), Y2(t)} = Corr(b1, b2) Corr{Y1(s), Y1(t)} Corr{Y2(s), Y2(t)}

Corr{Y1(s), Y1(t)} Corr{Y2(s), Y2(t)}

f (y1|b1) . . . f (yk |bk )f (b1, . . . , bk ) db1 . . . dbk

Computational problems (addressed later)

Methods Based on Dimension Reduction

General idea: Use a factor-analytic, or principal-component type, analysis to

The toenail study

Toenail Data: Continuous and Discrete Outcome

De Backer, De Keyser, De Vroey, Lesare (British Journal of Dermatology 1996)

Toenail Dermatophyte Onychomycosis: Common toenail infection, dicult to

General research question:

Unaected nail length (mm)?

Severity relative to treatment of

As response is related to toe size, we