Joint Modelling of Longitudinal Measurements

and Survival Outcomes

Peter Diggle
Medical Statistics Unit, Lancaster University
and
Department of Biostatistics, Johns Hopkins University
Copenhagen, April 2006

Outline

Joint modelling:
what is it?
why do it?
Modelling longitudinal measurements
Modelling survival outcomes
Joint modelling

transformation models
random effects models
Examples

Conclusions

Joint modelling: what is it?

Subjects i = 1, ..., m.
Longitudinal measurements Yij at times tij .

Times-to-event Si (possibly censored).

Baseline covariates xi.
Parameters .

[Y, S|x, ]

AIDS data

Data from RCT of three different drug regimes

for HIV+ patients.
1 = zidovudine, 600mg
2 =didanosine, 500mg
3 = didanosine, 750mg
Total n = 913 subjects.
S = time of progression to AIDS or death
Y = time-sequence of CD4 measurements
(weeks 0, 2, 6, 12, 18, 24)
S > 24 for most subjects (long-term follow-up)

334 observed event-times, 579 censored

Schizophrenia data
Data from placebo-controlled RCT of drug treatments
for schizophrenia:
Placebo
Haloperidol (standard)
Risperidone (novel)
Total n = 517 subjects.
Y = PANSS measurement (weeks -1, 0, 1, 2, 4, 6, 8)
S = dropout time

High dropout rates:

week
missing
proportion

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

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

Heart surgery data

RCT to compare two types of artificial heart-valve

homograft
stentless
Y = time-sequence of left-ventricular-mass-index (LVMI)
S = time of death
Is a patients longitudinal LVMI profile predictive of their survival
prognosis?

Joint modelling: why do it?

To analyse survival time S, whilst exploiting correlation with
an imperfectly measured, time-varying risk-factor Y
Example: AIDS data
interest is in time to progression/death
but long latency period implies heavy censoring

hence, joint modelling improves inferences about

marginal distribution [S]

To analyse a longitudinal outcome measure Y with

potentially informative dropout at time S
Example: Schizophrenia data
interest is reducing mean PANSS score

but informative dropout process would imply that

modelling only [Y ] may be misleading

Because the relationship between Y and S is of intrinsic interest

Example: heart surgery data
long-term build-up of left-ventricular muscle mass
may increase hazard for fatal heart-attack
hence, interested in modelling relationship between
survival and subject-level LVMI

Scientific goal affects choice of statistical model/method?

Modelling longitudinal measurements

The Gaussian linear model
(Yij , tij ) : j = 1, ..., ni; i = 1, ..., m

Yi = (Yi1, ..., Yini )

ti = (ti1, ..., tini )

ij = E[Yij ]

Y = (Y1, ..., Ym)

t = (t1, ..., tm)

Y MVN{X, V ()})

V1
0
V () =
..
0

0 . . . 0

.
.
.
.
.
V2

...
0

. . . 0 Vm

Modelling longitudinal measurements

Specifying the covariance structure

Yij = ij + Ui + Wi(tij ) + Zij

Corresponds to within-subject variance matrices

Vi = 2J + 2R(ti) + 2I
where R(ti) has elements Rjk = (|tij tik |).

Modelling longitudinal measurements

The variogram

V (u) = Var{Y (t) Y (t u)}

2
useful data-analytic tool for irregularly spaced,
incomplete data
theoretical form for model on previous slide
V (u) = 2 + 2{1 (u)}

Modelling survival outcomes

Fundamental tool is the hazard function,

h(s) = f (s)/{1 F (s)}

Modelling strategies

Proportional hazards
hi(s) = h0(s)i i = exp(x0i)
Accelerated life
Fi(s) = F0(is) i = exp(x0i)
Frailty
hi(s) = h0(s)i Ui

i = exp(x0i)

Ui G(u)

Modelling strategies (continued)

Parametric
S f (s; )

i = exp(x0i)

S Gamma(, ) : proportional hazards ( known)

log S N(, 2) : accelerated life
S Weibull(, ) : proportional hazards and accelerated life

Joint modelling
Considerations to inform choice of approach
focus on questions of primary scientific interest
interpretability of model parameters
statistical efficiency

diagnostic checks for assumptions about effects

of primary interest

robustness to departures from assumptions about effects

not of primary interest
ease of implementation
reduction to standard methods when there is no association

Joint modelling
Transformation models

(log S, Y ) MVN(, )
= (S , Y )
=

V ()

subjects provide independent replicates of (log S, Y )

Transformation models: the likelihood function

Standard result:
2
log S|Y N(S|Y , S|Y
)

S|Y = S + 0V ()1(Y Y )
2
= 2 0V ()1
S|Y

Likelihood contribution from ith subject:

uncensored Si:
[Yi] [log Si|Yi]

(multivariate Gaussian pdf )

censored Si:
[Yi] {1 ((log Si S|Yi )/S|Y )}

(pdf times tail probability)

Joint modelling
Random effects models
R1

R2

Formulation of random effects models

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) is realised independently between subjects

Measurement sub-model
Yij = i(tij ) + R1i(tij ) + Zij
Zij N(0, 2)

i(tij ) = X1i(tij )1
Hazard sub-model
hi(t) = h0(t)F {X2(t)2 + R2i(t)}
h0(t) = non-parametric baseline hazard
2(t) = X2i(t) + R2i(t) = linear predictor for hazard

typical choice F () = exp()

Random effects models: the likelihood function

conditional independence: S Y |R
standard Gaussian marginal: [Y ]],L1(; Y )

Gaussian conditional: [R|Y ]

standard conditional: [S|R], L2(; S|R)

selection factorisation

[Y, S] = [Y, S, R]dR

Z
= [Y ][R|Y ][S|R, Y ]dR
Z
= [Y ] [R|Y ][S|R]dR
L(; Y, S) = L1(; Y ) ER|Y [L2(; S|R)]

Evaluating the likelihood function

L(; Y, S) = L1(; Y ) ER|Y [L2(; S|R)]

R is infinite-dimensional, but
non-parametric specification of h0() implies we
only need R at event-times
Monte Carlo evaluation of expectation term
explicit EM evaluation possible in useful special cases
(eg Wulfsohn and Tsiatis, 1997)

Examples

AIDS data (transformation model)

Schizophrenia data (random effects model)

Heart surgery data (random effects model?)

rag replacements

AIDS data
Mean square-root CD4
Y
S
R1
R2

10.0
9.5

3
2

2
2

2
3
1

3
1

2
1

9.0

CD4

10.5

mean response by treatment and time

8.5

10

15

20

Time (weeks)

Fitted lines are quadratics with common intercept

Covariance structure

Variance matrix of OLS residuals from saturated

treatments-by-times model
17.94
18.26
18.53
18.80
18.47
18.01

18.26
22.86
21.02
21.32
20.94
20.57

18.53
21.02
23.95
22.47
22.02
21.76

18.80
21.32
22.47
25.01
22.60
22.54

18.47
20.94
22.02
22.60
24.67
22.20

18.01
20.57
21.76
22.54
22.20
24.64

g replacements

Y
S
R1
R2
Time (weeks)

CD4

Variogram tells similar story: dominant source of variation

is between subjects

15
10
0

V (u)

20

25

residual variogram

10

15

R(t) = residual at time t

V (u) = 12 Var{R(t) R(t u)}

20

frag replacements

Y
S
R1
R2
Time (weeks)

CD4

AIDS data
Baseline CD4 and time to progression/death

0.6
0.4
0.2
0.0

S(t)

0.8

1.0

Fitted S(t) at quartiles of baseline CD4

100

200

300

400

t (days)

500

600

700

Model formulation

exchangeable covariance structure for Y

non-parametric specification of Cov(log S, Y )

mean CD4 quadratic in time within each treatment group

linear covariate adjustments

Maximum likelihood estimation

Treatment contrasts:
21, 31 : contrasts in mean log S

21
, 31
: contrasts in mean Y at 24 weeks
Scale parameter estimate std. error correlation
log S
21
0.351
0.129
31
0.222
0.126
0.486
Y

21

31

0.981
1.003

0.439
0.362

0.423

 Covariance structure
Var(Y ) = 2I + 2J
Var(log S) = 2
Corr(log S, Yj ) = j
Parameter estimate std. error
2
2.432
0.057
2
16.723
0.808
1.653
0.145
2
0.428
0.038
1
0.468
0.035
2
0.467
0.035
3
4
0.457
0.036
0.520
0.035
5
6
0.509
0.038

PSfrag replacements

Goodness-of-fit
Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)

Focus on conditional distribution of log S given Y

0.0

0.4

0.8

Gaussian P-P and Q-Q plots with multiple imputation

of censored log S

0.0

0.4

0.8

Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)

40

60

80

100

120

PANSS

140

160
0

t (weeks)

Schizophrenia data
PANSS responses from haloperidol arm

replacements

90
85
80

Mean PANSS

PANSS

75

Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)
t (weeks)

placebo
halperidol
risperidone

70

PSfrag replacements

95

Schizophrenia data
Mean response

Time (weeks)

300
200
100

Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)
t (weeks)

V (u)

PSfrag replacements

400

Empirical and fitted variograms

Apparently fits well, but hard to distinguish empirically

PANSS
from random
intercept and slope model.
Time (weeks)
Mean PANSS

PSfrag replacements

Meanresponse by dropout cohort

Mean PANSS
u
V (u)

100
90
80
70

PANSS
Time (weeks)

mean response by dropout cohort

mean response

Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)
t (weeks)

Mean PANSS

Time (weeks)

Model formulation
Measurement sub-model
For subject in treatment group k,
i(t) = 0k + 1k t + 2k t2
Yij = i(tij ) + R1i(tij ) + Zij
Hazard sub-model
For subject in treatment group k,
hi(t) = h0(t) exp{k + R2i(t)}

Latent process
Two choices for measurement process component
R1(t) = U1 + W1(t)
R1(t) = U1 + U2t
And for hazard process component
R2(t) = 1R1(t)
R2(t) = 1(U1 + U2t) + 2U2

Y
S
R1
R2
Time (weeks)

CD4
t (days)
S(t)
t (weeks)

Mean PANSS

1
2
3
4
5
Time from operation (/years)

4.8

log(LVMI)
5.0
5.2

Stentless valve

4.6

4.8
4.6

4.8
4.6

Mean PANSS
u
V (u)
Time (weeks)

Homograft valve

log(LVMI)
5.0
5.2

overall

5.4

5.4

5.4

Mean profile (band = 3 months)

log(LVMI)
5.0
5.2

PANSS
Time (weeks)

Heart surgery data

Mean log-LVMI response profiles

1
2
3
4
5
Time from operation (/years)

1
2
3
4
5
Time from operation (/years)

PSfrag replacements

Heart surgery data

SurvivalY curves adjusted for baseline covariates

Mean PANSS

1.0
0.8
0.6
0.4
0.2

Mean PANSS
u
V (u)
Time (weeks)

observed
fitted

0.0

PANSS
Time (weeks)

Cox Proportional model, for baseline covariates

Survival

S
R1
R2
Time (weeks)

CD4
t (days)
S(t)
t (weeks)

6
Time

10

Heart surgery data

Hypothesis
subjects who regress after surgery (increasing LVMI) are
at greater risk of heart attack
Exploratory analysis
overall mean log(LVMI) decreases initially after surgery,
then remains approximately constant
but some patients regress (increasing LVMI)
time-averaged log(LVMI) is associated with increased hazard

Random effects joint model for heart surgery data

(tij ) + Ai + Wi(tij ) + Zj
: t
Yij =
(tij ) + (Ai + {Bi(t )} + Wi(tij ) + Zj : t >
hi(t) = h0(t) exp(x0i + Bi)
surgery has immediate beneficial effect on all patients
patient outcomes diverge after time post-surgery

hazard for survival depends on slope of LVMI after time

Conclusions

Choice of model/method should relate to scientific purpose.

Simple models/methods are useful when exploring a range
of modelling options, for example to select from amongst
potential covariates.
Complex models/methods are useful when seeking to understand
subject-level stochastic variation.
Flat likelihoods are common: different models may fit
the data almost equally well.
We need an R library.

References
Andersen, P.K., Borgan, ., Gill, R.D. and Keiding, N. (1993). Statistical Models Based on Counting Processes. New York: SpringerVerlag.
Cox, D.R. (1972). Regression models and life tables (with Discussion). Journal of the Royal Statistical Society B, 34, 187220.
Cox, D.R. (1999). Some remarks on failure-times, surrogate markers, degradation, wear and the quality of life. Lifetime Data Analysis,
5, 30714.
Diggle, P.J., Heagerty, P., Liang, K-Y and Zeger, S.L. (2002). Analysis of Longitudinal Data (second edition). Oxford : Oxford
University Press.
Faucett, C.L. and Thomas, D.C. (1996). Simultaneously modelling censored survival data and repeatedly measured covariates: a
Gibbs sampling approach. Statistics in Medicine, 15, 16631686.
Finkelstein, D.M. and Schoenfeld, D.A. (1999). Combining mortality and longitudinal measures in clinical trials. Statistics in Medicine,
18, 13411354.
Henderson, R., Diggle, P. and Dobson, A. (2000). Joint modelling of longitudinal measurements and event time data. Biostatistics,
1, 46580.
Henderson, R., Diggle, P. and Dobson, A. (2002). Identification and efficacy of longitudinal markers for survival. Biostatistics, 3,
3350.
Hogan, J.W. and Laird, N.M. (1997a). Mixture models for the joint distribution of repeated measures and event times. Statistics in
Medicine 16, 239257.
Hogan, J.W. and Laird, N.M. (1997b). Model-based approaches to analysing incomplete longitudinal and failure time data. Statistics
in Medicine 16, 259272.
Little, R.J.A. (1995). Modelling the drop-out mechanism in repeated-measures studies. Journal of the American Statistical Association, 90, 111221.
Pawitan, Y. and Self, S. (1993). Modelling disease marker processes in AIDS. Journal of the American Statisical Association 88,
719726.

Taylor, J.M.G., Cumberland, W.G. and Sy, J.P. (1994). A stochastic model for analysis of longitudinal AIDS data. Journal of the
American Statistical Association 89, 727736
Tsiatis, A.A., Degruttola, V. and Wulsohn, M.S. (1995). Modelling the relationship of survival to longitudinal data measured with
error. Applications to survival and CD4 counts in patients with AIDS. Journal of the American Statistical Association 90, 2737.
Wulfsohn, M.S. and Tsiatis, A.A. (1997). A joint model for survival and longitudinal data measured with error. Biometrics, 53,
3309.
Xu, J. and Zeger, S.L. (2001). Joint analysis of longitudinal data comprising repeated measures and times to events. Applied Statistics,
50, 375387.