Longitudinal Data Analysis

Peter Diggle

(Lancaster University and University of Liverpool)

Belfast, February 2015

Lecture topics
1. Time series and longitudinal data: similarities/differences

2. Linear models: capturing correlation structure

3. Missing values: Rubin’s hierarchy, informative dropout

4. Generalised linear models: binary and count data

5. Joint modelling: repeated measurement and time-to-event

outcomes
1. Time series and longitudinal data
Bailrigg temperature records
Daily maximum temperatures, 1.09.1995 to 31.08.1996
25
20
temperature
15
10
5
0

0 100 200 300

day
Schizophrenia clinical trial (PANSS)

• randomised clinical trial of drug therapies

• three treatments:

– haloperidol (standard)
– placebo
– risperidone (novel)

• dropout due to “inadequate response to treatment”

Treatment Number of non-dropouts at week

0 1 2 4 6 8
haloperidol 85 83 74 64 46 41
placebo 88 86 70 56 40 29
risperidone 345 340 307 276 229 199
total 518 509 451 396 315 269
Schizophrenia trial data

150
PANSS
100
50
0

0 2 4 6 8

time (weeks since randomisation)

Diggle, Farewell and Henderson (2007)

Time series decomposition

• trend and residual

• autocorrelation

• prediction
Analysis of Bailrigg temperature data

data<-read.table("../data_and_figures/maxtemp.data",header=F)
temperature<-data[,4]
n<-length(temperature)
day<-1:n
plot(day,temperature,type="l",cex.lab=1.5,cex.axis=1.5)
#
# plot shows strong seasonal variation,
# try simple harmonic regression
#
c1<-cos(2*pi*day/n)
s1<-sin(2*pi*day/n)
fit1<-lm(temperature~c1+s1)
lines(day,fit1$fitted.values,col="red")
#
# add first harmonic of annual frequency to check for
# non-sinusoidal pattern
#
c2<-cos(4*pi*day/n)
s2<-sin(4*pi*day/n)
fit2<-lm(temperature~c1+s1+c2+s2)
lines(day,fit2$fitted.values,col="blue")
#
# two fits look similar, but conventional F test says otherwise
#
summary(fit2)
RSS1<-sum(fit1$resid^2); RSS2<-sum(fit2$resid^2)
F<-((RSS1-RSS2)/2)/(RSS2/361)
1-pf(F,2,361)
#
# conventional residual plots
#
# residuals vs fitted values
#
plot(fit2$fitted.values,fit2$resid)
#
# residuals in time-order as scatterplot
#
plot(1:365,fit2$resid)
#
# and as line-graph
#
plot(1:365,fit2$resid,type="l")
#
# examine autocorrelation properties of residuals
#
residuals<-fit2$resid
par(mfrow=c(2,2),pty="s")
for (k in 1:4) {
plot(residuals[1:(n-k)],residuals[(k+1):n],
pch=19,cex=0.5,xlab=" ",ylab=" ",main=k)
}
par(mfrow=c(1,1))
acf(residuals)
#
# exponentially decaying correlation looks reasonable
#
cor(residuals[1:(n-1)],residuals[2:n])
Xmat<-cbind(rep(1,n),c1,s1,c2,s2)
rho<-0.01*(60:80)
profile<-AR1.profile(temperature,Xmat,rho)
#
# examine results
#
plot(rho,profile$logl,type="l",ylab="L(rho)")
Lmax<-max(profile$logl)
crit.val<-0.5*qchisq(0.95,1)
lines(c(rho[1],rho[length(rho)]),rep(Lmax-crit.val,2),lty=2)
profile
#
# Exercise: how would you now re-assess the significance of
# the second harmonic term?
#
# profile log-likelihood function follows
#
AR1.profile<-function(y,X,rho) {
m<-length(rho)
logl<-rep(0,m)
n<- length(y)
hold<-outer(1:n,1:n,"-")
for (i in 1:m) {
Rmat<-rho[i]^abs(hold)
ev<-eigen(Rmat)
logdet<-sum(log(ev$values))
Rinv<-ev$vectors%*%diag(1/ev$values)%*%t(ev$vectors)
betahat<-solve(t(X)%*%Rinv%*%X)%*%t(X)%*%Rinv%*%y
residual<- y-X%*%betahat
logl[i]<- - logdet - n*log(c(residual)%*%Rinv%*%c(residual))
}
max.index<-order(logl)[m]
Rmat<-rho[max.index]^abs(hold)
ev<-eigen(Rmat)
logdet<-sum(log(ev$values))
Rinv<-ev$vectors%*%diag(1/ev$values)%*%t(ev$vectors)
betahat<-solve(t(X)%*%Rinv%*%X)%*%t(X)%*%Rinv%*%y
residual<- y-X%*%betahat
sigmahat<-sqrt(c(residual)%*%Rinv%*%c(residual)/n)
list(logl=logl,rhohat=rho[max.index],sigmahat=sigmahat,betahat=betahat)
}
Longitudinal data

• replicated time series;

• focus of interest often on mean values;

• modelling and inference can and should exploit

replication
2. Linear models

• correlation and why it matters

• exploratory analysis

• linear Gaussian models

Correlation and why it matters

• different measurements on the same subject are

typically correlated

• and this must be recognised in the inferential process.

Estimating the mean of a time series

Y1 , Y2 , ..., Yt , ..., Yn Yt ∼ N(µ, σ 2 )

p
Classical result: Ȳ ± 2 σ 2 /n

But if Yt is a time series:

• E[Ȳ ] = µ
2 −1
P
• Var{Ȳ } = (σ /n) × {1 + n u6=t Corr(Yt , Yu )}

Exercise: is the sample variance unbiased for σ 2 = Var(Yt )?

Correlation may or may not hurt you

Yit = α + β(t − t̄) + Zit i = 1, ..., m t = 1, ..., n

10
8
response
6
4
2

−4 −2 0 2 4
time
Correlation may or may not hurt you

Yit = α + β(t − t̄) + Zit i = 1, ..., m t = 1, ..., n

10
8
response
6
4
2

−4 −2 0 2 4
time
Correlation may or may not hurt you

Yit = α + β(t − t̄) + Zit i = 1, ..., m t = 1, ..., n

10
8
response
6
4
2

−4 −2 0 2 4
time
Correlation may or may not hurt you

Yit = α + β(t − t̄) + Zit i = 1, ..., m t = 1, ..., n

Parameter estimates and standard errors:

ignoring correlation recognising correlation

estimate standard error estimate standard error
α 5.234 0.074 5.234 0.202
β 0.493 0.026 0.493 0.011
A spaghetti plot of the PANSS data
150

150

150
100

100

100
PANSS

PANSS

PANSS
50

50

50
0

0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
time (weeks) time (weeks) time (weeks since randomisation)
Exploring covariance structure: balanced data
(Yij , tj ) : j = 1, ..., n; i = 1, ..., m

• fit saturated treatments-by-times model to mean response

• compute sample covariance matrix of residuals

PANSS data:
SD Y.t0 Y.t1 Y.t2 Y.t4 Y.t6 Y.t8
Y.t0 20.019 1.000 0.533 0.366 0.448 0.285 0.229
Y.t1 20.184 0.533 1.000 0.693 0.589 0.658 0.535
Y.t2 22.120 0.366 0.693 1.000 0.670 0.567 0.678
Y.t4 20.996 0.448 0.589 0.670 1.000 0.718 0.648
Y.t6 24.746 0.285 0.658 0.567 0.718 1.000 0.792
Y.t8 23.666 0.229 0.535 0.678 0.648 0.792 1.000

• modest increase in variability over time

• correlation decays with increasing time-separation
Exploring covariance structure: unbalanced
data
(Yij , tij ) : j = 1, ..., ni ; i = 1, ..., m

The variogram of a stochastic process Y (t) is

1
V (u) = Var{Y (t) − Y (t − u)}
2

• well-defined for stationary and some non-stationary

processes

• for stationary processes,

V (u) = σ 2 {1 − ρ(u)}

• V (u) easier to estimate than ρ(u) when data are

unbalanced
Estimating the variogram
Data: (Yij , tij ) : i = 1, ..., m; j = 1, ..., ni

rij = residual from preliminary model for mean response

• Define
1
vijkℓ = (rij − rkℓ )2
2
• Estimate

V̂ (u) = average of all vijiℓ such that |tij − tiℓ | ≃ u

σ̂ 2 = average of all vijkℓ such that i 6= k.
Example: sample variogram of the PANSS data
Solid lines are estimates from data, horizontal lines are eye-ball
estimates (explanation later)

400
300
Variogram

200
100
0

0 2 4 6 8

Lag
Where does the correlation come from?

• differences between subjects

• variation over time within subjects

• measurement error
General linear model, correlated residuals

i = subjects j = measurements within subjects

E(Yij ) = xij1 β1 + ... + xijp βp

Yi = Xi β + ǫi
Y = Xβ + ǫ

• measurements from different subjects independent

• measurements from same subject typically correlated.

Parametric models for covariance structure
Three sources of random variation in a typical set of
longitudinal data:

• Random effects (variation between subjects)

– characteristics of individual subjects

– for example, intrinsically high or low responders
– influence extends to all measurements on the
subject in question.
Parametric models for covariance structure
Three sources of random variation in a typical set of
longitudinal data:

• Random effects

• Serial correlation (variation over time within subjects)

– measurements taken close together in time typically

more strongly correlated than those taken further
apart in time
– on a sufficiently small time-scale, this kind of
structure is almost inevitable
Parametric models for covariance structure
Three sources of random variation in a typical set of
longitudinal data:

• Random effects

• Serial correlation

• Measurement error

– when measurements involve delicate determinations,

duplicate measurements at same time on same
subject may show substantial variation

Diggle, Heagerty, Liang and Zeger (2002, Chapter 5)

Some simple models

• Compound symmetry

Yij − µij = Ui + Zij

Ui ∼ N(0, ν 2 )
Zij ∼ N(0, τ 2 )

Implies that Corr(Yij , Yik ) = ν 2 /(ν 2 + τ 2 ), for all j 6= k

• Random intercept and slope

Yij − µij = Ui + Wi tij + Zij

(Ui , Wi ) ∼ BVN(0, Σ)
Zij ∼ N(0, τ 2 )

Often fits short sequences well, but extrapolation

dubious, for example Var(Yij ) quadratic in tij
• Autoregressive

Yij − µij = α(Yi,j−1 − µi,j−1 ) + Zij

Yi1 − µi1 ∼ N{0, τ 2 /(1 − α2 )}

Zij ∼ N(0, τ 2 ), j = 2, 3, ...

Not a natural choice for underlying continuous-time

processes
• Stationary Gaussian process

Yij − µij = Wi (tij )

Wi (t) a continuous-time Gaussian process

E[W (t)] = 0 Var{W (t)} = σ 2
Corr{W (t), W (t − u)} = ρ(u)

ρ(u) = exp(−u/φ) gives continuous-time version

of the autoregressive model
Time-varying random effects
intercept and slope
3
2
1
R(t)

0
−1
−2
−3

0 200 400 600 800 1000

t
Time-varying random effects: continued
stationary process
3
2
1
R(t)

0
−1
−2
−3

0 200 400 600 800 1000

t
• A general model

Yij − µij = d′ij Ui + Wi (tij ) + Zij

Ui ∼ MVN(0, Σ)
(random effects)
dij = vector of explanatory variables for random effects
Wi (t) = continuous-time Gaussian process
(serial correlation)
Zij ∼ N(0, τ 2 )
(measurement errors)

Even when all three components of variation are needed

in principle, one or two may dominate in practice
The variogram of the general model
(stationary case)

Yij − µij = Ui + Wi (tij ) + Zij

V (u) = τ 2 + σ 2 {1 − ρ(u)} Var(Yij ) = ν 2 + σ 2 + τ 2

1.0
0.8
0.6
V(u)
0.4
0.2
0.0

0 1 2 3 4 5 6
u
Fitting the model: non-technical summary

• Ad hoc methods won’t do

• Likelihood-based inference is the statistical gold standard

• But be sure you know what you are estimating when

there are missing values
Maximum likelihood estimation (V0 known)
Log-likelihood for observed data y is

L(β, σ 2 , V0 ) = −0.5{nm log σ 2 + m log |V0 |

+σ −2 (y − Xβ)′ (I ⊗ V0 )−1 (y − Xβ)}, (1)

I ⊗ V0 denotes block-diagonal matrix with non-zero blocks V0

Given V0 , estimator for β is

β̂(V0 ) = (X ′ (I ⊗ V0 )−1 X)−1 X ′ (I ⊗ V0 )−1 y, (2)

Explicit estimator for σ 2 also available as

σ̂ 2 (V0 ) = RSS(V0 )/(nm) (3)

RSS(V0 ) = {y − X β̂(V0 )}′ (I ⊗ V0 )−1 {y − X β̂(V0 )}.

Maximum likelihood estimation, V0 unknown

Substitute (2) and (3) into (1) to give reduced log-likelihood

L(V0 ) = −0.5m[n log{RSS(V0 )} + log |V0 |]. (4)

Numerical maximization of (4) then gives V̂0 , hence β̂ = β̂(V̂0 )

and σ̂ 2 = σ̂ 2 (V̂0 ).

1
• Dimensionality of optimisation is 2
n(n + 1) − 1

• Each evaluation of L(V0 ) requires inverse and

determinant of an n by n matrix.
A random effects model for CD4 cell counts
data<-read.table("../data_and_figures/CD4.data",header=T)
data[1:3,]
time<-data$time
CD4<-data$CD4
plot(time,CD4,pch=19,cex=0.25)
id<-data$id
uid<-unique(id)
for (i in 1:10) {
take<-(id==uid[i])
lines(time[take],CD4[take],col=i,lwd=2)
}
# Simple linear model assuming uncorrelated residuals
#
fit1<-lm(CD4~time)
summary(fit1)
#
# random intercept and slope model
#
library(nlme)
?lme
fit2<-lme(CD4~time,random=~1|id)
summary(fit2)
# make fitted value constant before sero-conversion
#
timeplus<-time*(time>0)
fit3<-lme(CD4~timeplus,random=~1|id)
summary(fit3)
tfit<-0.1*(0:50)
Xfit<-cbind(rep(1,51),tfit)
fit<-c(Xfit%*%fit3$coef$fixed)
Vmat<-fit3$varFix
Vfit<-diag(Xfit%*%Vmat%*%t(Xfit))
upper<-fit+2*sqrt(Vfit)
lower<-fit-2*sqrt(Vfit)
#
# plot fit with 95% point-wise confidence intervals
#
plot(time,CD4,pch=19,cex=0.25)
lines(c(-3,tfit),c(upper[1],upper),col="red")
lines(c(-3,tfit),c(lower[1],lower),col="red")
3. Missing values and dropouts

Issues concerning missing values in longitudinal data can be

addressed at two different levels:

• technical: can the statistical method I am using cope with

missing values?

• conceptual: why are the data missing? Does the fact

that an observation is missing convey partial information
about the value that would have been observed?

These same questions also arise with cross-sectional data, but

the correlation structure of longitudinal data can sometimes be
exploited to good effect, by modelling how the probability of
dropout for each person depends on their previously observed
measurements
Rubin’s classification

• MCAR (completely at random): P(missing) depends

neither on observed nor unobserved measurements

• MAR (at random): P(missing) depends on observed

measurements, but not on unobserved measurements

• MNAR (not at random): conditional on observed

measurements, P(missing) depends on unobserved
measurements.

Rubin (1976)
Dropout
Once a subject goes missing, they never return
Example : Longitudinal clinical trial

• completely at random: patient leaves the the study

because they move house

• at random : patient leaves the study on their doctor’s

advice, based on observed measurement history

• not at random : patient misses their appointment

because they are feeling unwell.

Little (1995)
Modelling the missing value process

• Y = (Y1 , ..., Yn ), intended measurements on a single

subject

• t = (t1 , ..., tn ), intended measurement times

• M = (M1 , ..., Mn ), missingness indicators

• for dropout, M reduces to a single dropout time D,

in which case:

– (Y1 , ..., YD−1 ) observed

– (YD , ..., Yn ) missing

A model for data subject to missingness is just a specification

of the joint distribution
[Y, M ]
Modelling the missing value process:
three approaches

• Selection factorisation

[Y, M ] = [Y ][M |Y ]

• Pattern mixture factorisation

[Y, M ] = [M ][Y |M ]

• Random effects
Z
[Y, M ] = [Y |U ][M |U ][U ]dU
Comparing the three approaches

• Pattern mixture factorisation has a natural data-analytic

interpretation
(sub-divide data into different dropout-cohorts)

• Selection factorisation may have a more natural

mechanistic interpretation in the dropout setting
(avoids conditioning on the future)

• Random effects conceptually appealing, especially for noisy

measurements, but make stronger assumptions and
usually need computationally intensive methods
for likelihood inference
Fitting a model to data with dropouts
• MCAR

1. almost any method will give sensible point estimates

of mean response profiles
2. almost any method which takes account of
correlation amongst repeated measurements will
give sensible point estimates and standard errors
• MAR

1. likelihood-based inference implicitly assumes MAR

2. for inferences about a hypothetical dropout-free
population, there is no need to model the dropout
process explicitly
3. but be sure that a hypothetical dropout-free
population is the required target for inference
• MNAR

1. joint modelling of repeated measurements and dropout

times is (more or less) essential
2. but inferences are likely to be sensitive to
modelling assumptiuons that are difficult
(or impossible) to verify empirically
Proof: Partition Y for each subject into observed and missing
components, Y = (Yo , Ym ) and let M denote binary vector of
missingness indicators. Likelihood for observed data is
R
L = g(yo , m) = R f (yo , ym , m)dym
= f (yo )f (ym |yo )p(m|yo , ym )dym

If p(m|yo , ym ) = p(m|yo ), take outside integral to give

L = p(m|yo )f (yo )

and log-likelihood contribution

log L = log p(m|yo ; θ) + log f (yo |θ)

• OK to ignore first term for likelihood inference about θ

• and no loss of efficiency if θ = (θ1 , θ2 ) such that θ1 and

θ2 parameterise p(·) and f (·), respectively.

But is inference about f (·) what you want?

Example

• Model is Yij = µ + Ui + Zij (random intercept)

• Dropout is MAR: logit(pij ) = −1 − 2 × Yi,j−1

5
4
3
2
y
1
0
−1

2 4 6 8 10
time

• Observed means increase over time, but population mean

µ is constant
PJD’s take on ignorability
For correlated data, dropout mechanism can be ignored only if
dropouts are completely random

In all other cases, need to:

• think carefully what are the relevant practical questions,

• fit an appropriate model for both measurement process

and dropout process

• use the model to answer the relevant questions.

Diggle, Farewell and Henderson (2007)

 Schizophrenia trial data

• Data from placebo-controlled RCT of drug treatments

for schizophrenia:

– Placebo; Haloperidol (standard); Risperidone (novel)

• Y = sequence of weekly PANSS measurements

• F = dropout time

• Total m = 516 subjects, but high dropout rates:

week −1 0 1 2 4 6 8
missing 0 3 9 70 122 205 251
proportion 0.00 0.01 0.02 0.14 0.24 0.40 0.49

• Dropout rate also treatment-dependent (P > H > R)

 Schizophrenia data
PANSS responses from haloperidol arm

160
140
120
PANSS

100
80
60
40

0 2 4 6 8

time (weeks)
 empirical cumulative distribution empirical cumulative distribution
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

40
60

60
80

80
100

Y
Y

t=4
t=0

100
120

120
140

140
160

empirical cumulative distribution empirical cumulative distribution

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
40

40
60

60
80

80
Y
Y
100

t=6
t=1

100
120

120
140

140
160

empirical cumulative distribution

0.0 0.2 0.4 0.6 0.8 1.0
40
60
80
Y
t=2

100

subjects about to drop−out

subjects not about to drop−out
120
140
Dropout is not completely at random

160
 Schizophrenia trial data
Mean response (random effects model)

95
90
85
Mean response

80
75

placebo
halperidol
risperidone
70

0 2 4 6 8

Time (weeks)
 Schizophrenia trial data
Empirical and fitted variograms

400
300
Variogram

200
100
0

0 2 4 6 8

Lag
 Schizophrenia trial data: summary

• dropout is not MCAR

• MAR model apparently fits well, but:

– hard to distinguish empirically between different MAR

models;
– and we haven’t formally investigated evidence for
informative dropout

• Fitted means relate to hypothetical, dropout-free

population
Embedding MAR within an MNAR model
1. Diggle and Kenward

Measurement model
General linear model for Yi = {Yit : t = 1, ..., n}
(balanced data)

Dropout model
Logistic regression:

logit P(Di = t|Yi ) = α + βYi,t−1 + γYit

Diggle and Kenward, 1994

2. Barrett, Diggle, Henderson and Taylor-Robinson

Hybrid time-scales

Continuous-time measurement model

Yij = µij + Si (tij ) + Zij

Discrete-time survival model

Uik = {Si (uk ) : k = 1, ..., N }

Linkage
d
!
X
P(Di = d|Di > d − 1, U ) = 1 − Φ µ̃id + γk Uik
k=1

Barrett, Diggle, Henderson and Taylor-Robinson, 2015

4. Generalized linear models
• random effects models

• transition models

• marginal models

Diggle, Heagerty, Liang and Zeger (2002, Chapter 7)

Random effects GLM
Responses Y1 , . . . , Yn on an individual subject conditionally
independent, given unobserved vector of random effects U
U ∼ g(u) represents properties of individual subjects that
vary randomly between subjects

• E(Yj |U ) = µj : h(µi ) = x′j β + U ′ α

• Var(Yj |U ) = φv(µj )

• (Y1 , . . . , Yn ) are mutually independent conditional on U .

Likelihood inference requires evaluation of

n
Z Y
f (y) = f (yj |U )g(U )dU
j=1
Transition GLM
Each Yj modelled conditionally on preceding Y1 , Y2 , . . . , Yj−1 .

• E(Yj |history) = µj
Pj−1
• h(µj ) = x′j β + k=1
′
Yj−k αk

• Var(Yj |history) = φv(µj )

Construct likelihood as product of conditional distributions,

usually assuming restricted form of dependence.
Example: fk (yj |y1 , ..., yj−1 ) = fk (yj |yj−1 )
Need to condition on y1 as model does not directly specify
marginal distribution f1 (y1 ).
Marginal GLM
Let h(·) be a link function which operates component-wise,

• E(y) = µ : h(µ) = Xβ

• Var(yi ) = φv(µi )

• Corr(y) = R(α).

Not a fully specified probability model

May require constraints on variance function v(·) and
correlation matrix R(·) for valid specification
Inference for β uses generalized estimating equations
Liang and Zeger (1986)
What are we estimating?
• in marginal modelling, β measures population-averaged
effects of explanatory variables on mean response

• in transition or random effects modelling, β measures

effects of explanatory variables on mean response of an
individual subject, conditional on

– subject’s measurement history (transition model)

– subject’s own random characteristics Ui
(random effects model)
Example: Simulation of a logistic regression model,
probability of positive response from subject i at time t is pi (t),

logit{pi (t)} : α + βx(t) + γUi ,

x(t) is a continuous covariate and Ui is a random effect

P(y=1)
1.0
0.8
0.6
0.4
0.2
0.0

−10 −5 0 5 10
x
Example: Effect of mother’s smoking on probability of
intra-uterine growth retardation (IUGR).

Consider a binary response Y = 1/0 to indicate whether a baby

experiences IUGR, and a covariate x to measure the mother’s
amount of smoking.
Two relevant questions:

1. public health: by how much might population incidence

of IUGR be reduced by a reduction in smoking?

2. clinical/biomedical: by how much is a baby’s risk of IUGR

reduced by a reduction in their mother’s smoking?

Question 1 is addressed by a marginal model, question 2 by a

random effects model
R software
The following is almost certainly an incomplete list.

• Marginal models
Function gee within package of same name

• Random effects models

Function glmmPQL within MASS package or lmer within lme4
(but note evaluation of likelihood uses approximate meth-
ods that may perform badly if random effects are high-
dimensional). Package glmmBUGS is a Bayesian alternative.

• Transition models
Standard glm function, after computing values of required
functions of lagged responses to be used as explanatory
variables.
Illustration of marginal modelling

set.seed(2346)
x=rep(1:10,50)
logit=0.1*(x-mean(x))
subject=rep(1:50,each=10)
re=2*rnorm(50)
re=rep(re,each=10)
prob=exp(re+logit)/(1+exp(re+logit))
y=(runif(500)<prob)
fit1=glm(y~x,family=binomial)
summary(fit1)
library(gee)
fit2<-gee(y~x,id=subject,family=binomial)
summary(fit2)
5. Joint modelling: repeated measurements
and time-to-event outcomes

• what is it?

• why do it?

• random effects models

Joint modelling: what is it?

• Subjects i = 1, ..., m.

• Longitudinal measurements Yij at times tij , j = 1, ..., ni .

• Times-to-event Fi (possibly censored).

• Baseline covariates xi .

• Parameters θ.

[Y, F |x, θ]
Joint modelling: what is it?
60

F=6.4
55
outcome
50

5<F<6
45

F>7
40

0 2 4 6 8 10
time
Joint modelling: why do it?

• To analyse failure time F , whilst exploiting correlation

with
an imperfectly measured, time-varying risk-factor Y

• To analyse a longitudinal outcome measure Y with

potentially informative dropout at time F

• Because relationship between Y and F is of direct interest

Random effects models

• linear Gaussian sub-model for repeated measurements

• proportional hazards sub-model with time-dependent

fraility for time-to-event

• sub-models linked through shared random effects

R1 F α

R2 Y β
Example: Wulfsohn and Tsiatis, 1997

latent random effect; measurement model; hazard model

Latent random effect

Random intercept and slope: Ui = (U0i , U1i )
Laird and Ware, 1982

Measurement model

Yij = µi (tij ) + U0i + U1i tij + Zij

• Zij ∼ N(0, τ 2 )

• µi (tij ) = X1i (tij )β1

• (U0i , U1i ) ∼ BVN(0, Σ)

Hazard model

hi (t) = h0 (t) exp{θ(U0i + U1i tij )}

• h0 (t) = non-parametric baseline hazard

• θ(U0i + U1i tij ) = linear predictor for hazard,

proportional to random effect
Example: Henderson, Diggle and Dobson, 2000

latent stochastic process; measurement model; hazard model

Latent stochastic process

Bivariate Gaussian process R(t) = {R1 (t), R2 (t)}

• Rk (t) = Dk (t)Uk + Wk (t)

• {W1 (t), W2 (t)}: bivariate stationary Gaussian process

• (U1 , U2 ): multivariate Gaussian random effects

Bivariate process R(t) realised independently between subjects

Measurement model

Yij = µi (tij ) + R1i (tij ) + Zij

• Zij ∼ N(0, τ 2 )

• µi (tij ) = X1i (tij )β1

Hazard model

hi (t) = h0 (t) exp{X2 (t)β2 + R2i (t)}

• h0 (t) = non-parametric baseline hazard

• η2 (t) = X2i (t) + R2i (t) = linear predictor for hazard

Two (relatively) open questions

• Repeated measurements and recurrent events

• Informative follow-up

Note: what constitutes a missing value if follow-up schedule is

not pre-specified?
The joineR package
Exploring the mental data-set
library(joineR)
data(mental)
mental[1:5,]
y<-as.matrix(mental[,2:7]) # convert data to matrix format
means<-matrix(0,3,6)
for (trt in 1:3) {
ysub<-y[mental$treat==trt,]
means[trt,]<-apply(ysub,2,mean,na.rm=TRUE)
}
residuals<-matrix(0,150,6)
for (i in 1:150) {
residuals[i,]<-y[i,]-means[mental$treat[i],]
}
V<-cov(residuals,use="pairwise"); R<-cor(residuals,use="pairwise")
round(cbind(diag(V),R),3)
The joineR package
Setting up a jointdata object
is.data.frame(mental)
mental.unbalanced<-to.unbalanced(mental, id.col = 1,
times = c(0,1,2,4,6,8),Y.col = 2:7, other.col = 8:11)
names(mental.unbalanced)
names(mental.unbalanced)[3]<-"Y"
mental.long<-mental.unbalanced[,1:3]
mental.surv <- UniqueVariables(mental.unbalanced,
var.col=6:7,id.col = 1)
mental.baseline <- UniqueVariables(mental.unbalanced,
var.col=4,id.col = 1)
mental.baseline$treat<-as.factor(mental.baseline$treat) # !!!
mental.joint<-jointdata(longitudinal=mental.long,
survival=mental.surv,baseline=mental.baseline,
id.col="id",time.col="time")
summary(mental.joint)
The joineR package
Fitting a joint model

fit<-joint(mental.joint,long.formula=Y~-1+treat+time,
surv.formula=Surv(surv.time,cens.ind)~treat,
model="intslope")
summary(fit)
set.seed(389712)
fit.se <- jointSE(fit, n.boot = 5)
# use much larger number of bootstrap samples in practice
set.seed(54912)
fit.se100 <- jointSE(fit, n.boot = 100,max.it=2000,tol=0.01,
print.detail=TRUE)
fit.se100
 Take-home messages
• Correlation matters

• Longitudinal designs address a richer set of questions

than cross-sectional designs

• But also raise challenges in formulating a valid, efficient

analysis:

– what, precisely, is the question?

– what explicit and implicit assumptions does the
proposed method of analysis make?
 References
Barrett, J., Diggle, P.J., Henderson, R. and Taylor-Robinson, D. (2013). Joint modelling of re-
peated measurements and time-to-event outcomes: flexible model specification and exact likelihood
inference. Journal of the Royal Statistical Society, B (submitted)
Chatfield, C. (2003). The Analysis of Time Series: an Introduction (6th edition). London: Chapman
and Hall.
Diggle, P.J. (1990). Time Series: a Biostatistical Introduction.
Diggle, P.J. and Al Wasel, I. (1997). Spectral analysis of replicated biomedical time series (with
Discussion). Applied Statistics, 46, 31–71.
Diggle, P.J., Farewell, D. and Henderson, R. (2007). Longitudinal data with dropout: objectives,
assumptions and a proposal (with Discussion). Applied Statistics, 56, 499–550.
Diggle, P.J., Heagerty, P., Liang, K-Y and Zeger, S.L. (2002). Analysis of Longitudinal Data (second
edition). Oxford: Oxford University Press.
Diggle, P.J. and Kenward, M.G. (1994). Informative dropout in longitudinal data analysis (with
Discussion). Appl. Statist. 43, 49–93.
Diggle, P.J. and Sousa, I. (2013). Real-time detection of incipient renal failure in primary care
patients using a dynamic time series model. Biostatistics (submitted)
Durbin, J. and Koopman, S.J. (2001). Time Series Analysis by State Space Methods. Oxford:
Oxford University Press.
Fitzmaurice, G.M., Laird, N.M. and Ware, J.H. (2004). Applied Longitudinal Analysis. New Jersey:
Wiley.
Laird, N.M. and Ware, J.H. (1982). Random-effects models for longitudinal data. Biometrics, 38,
963–74.
Little, R.J.A. (1995). Modelling the drop-out mechanism in repeated-measures studies. Journal
of the American Statistical Association, 90, 1112–21.
Little, R.J. and Rubin, D.B. (2002). Statistical Analysis with Missing Data (seond edition). New
York: Wiley.
Rubin, D.B. (1976). Inference and missing data. Biometrika, 63, 581–92.
Wulfsohn, M.S. and Tsiatis, A.A (1997). A joint model for survival and longitudinal data measured
with error. Biometrics, 53, 330–339.