00 voturi pozitive00 voturi negative

30 vizualizări94 paginiMon 1030 Crowther

Mar 15, 2015

© © All Rights Reserved

PDF, TXT sau citiți online pe Scribd

Mon 1030 Crowther

© All Rights Reserved

30 vizualizări

00 voturi pozitive00 voturi negative

Mon 1030 Crowther

© All Rights Reserved

Sunteți pe pagina 1din 94

stjm

Estimation

Discussion

References

survival data: Implementation in Stata

Statistical Analysis of Multi-Outcome Data

University of Cambridge

30th June - 1st July 2014

Michael J. Crowther

Department of Health Sciences

University of Leicester, UK

michael.crowther@le.ac.uk

Michael J. Crowther

Joint Modelling

1 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

2 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

3 / 56

Background

stjm

Estimation

Discussion

References

Background

I

repeatedly over time, in parallel to the time to an event of

interest, such as death from any cause

Michael J. Crowther

Joint Modelling

4 / 56

Background

stjm

Estimation

Discussion

References

Background

I

repeatedly over time, in parallel to the time to an event of

interest, such as death from any cause

These biomarkers are often measured with error

Michael J. Crowther

Joint Modelling

4 / 56

Background

stjm

Estimation

Discussion

References

Background

I

I

I

repeatedly over time, in parallel to the time to an event of

interest, such as death from any cause

These biomarkers are often measured with error

Issues:

I

Michael J. Crowther

drop-out, e.g. due to death

Can we account for measurement error when looking at

how a time-varying biomarker is associated with an

event of interest?

Joint Modelling

4 / 56

Background

stjm

Estimation

Discussion

References

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

Michael J. Crowther

measurement error

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

measurement error

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

measurement error

I

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

I

measurement error

Were still ignoring the measurement error

time-varying covariate

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

I

measurement error

Were still ignoring the measurement error

time-varying covariate

I

Michael J. Crowther

through

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

I

I

time-varying covariate

I

measurement error

through

I

Michael J. Crowther

Joint Modelling

5 / 56

Background

stjm

Estimation

Discussion

References

I

trajectories to progression to AIDS in HIV positive

patients (Faucett and Thomas, 1996)

Further developed in cancer, particularly modelling PSA

levels and their association with prostate cancer

recurrence (Proust-Lima and Taylor, 2009)

Michael J. Crowther

Joint Modelling

6 / 56

Background

stjm

Estimation

Discussion

References

I

trajectories to progression to AIDS in HIV positive

patients (Faucett and Thomas, 1996)

Further developed in cancer, particularly modelling PSA

levels and their association with prostate cancer

recurrence (Proust-Lima and Taylor, 2009)

I Latent class approach (Proust-Lima et al., 2012)

I Shared parameter models - dependence through shared

random effects (Wulfsohn and Tsiatis, 1997; Henderson

et al., 2000; Rizopoulos, 2012)

Michael J. Crowther

Joint Modelling

6 / 56

Background

stjm

Estimation

Discussion

References

I

trajectories to progression to AIDS in HIV positive

patients (Faucett and Thomas, 1996)

Further developed in cancer, particularly modelling PSA

levels and their association with prostate cancer

recurrence (Proust-Lima and Taylor, 2009)

I Latent class approach (Proust-Lima et al., 2012)

I Shared parameter models - dependence through shared

random effects (Wulfsohn and Tsiatis, 1997; Henderson

et al., 2000; Rizopoulos, 2012)

Michael J. Crowther

Joint Modelling

6 / 56

Background

stjm

Estimation

Discussion

References

Longitudinal submodel

Assume we observe continuous longitudinal marker:

ei (t) N(0, 2 )

where

and

bi N(0, )

Flexibility can be incorporated through fractional polynomials

or splines in Xi and Zi .

Michael J. Crowther

Joint Modelling

7 / 56

Background

stjm

Estimation

Discussion

References

Survival submodel

We assume a proportional hazards survival submodel

hi (t) = h0 (t) exp T v i + mi (t)

where h0 (t) is the baseline hazard function, and v i U i a set

of baseline time-independent covariates with associated vector

of log hazard ratios, .

Michael J. Crowther

Joint Modelling

8 / 56

Background

stjm

Estimation

Discussion

References

Survival submodel

We assume a proportional hazards survival submodel

hi (t) = h0 (t) exp T v i + mi (t)

where h0 (t) is the baseline hazard function, and v i U i a set

of baseline time-independent covariates with associated vector

of log hazard ratios, .

Michael J. Crowther

Joint Modelling

8 / 56

Background

stjm

Estimation

Discussion

References

Our key question here is how are changes in the biomarker

trajectory associated with survival?

hi (t) = h0 (t) exp T v i + mi (t)

where for example

mi (t) = (0 + b0i ) + (1 + b1i )t + u Ti

mi (t) is termed the current value parameterisation

Michael J. Crowther

Joint Modelling

9 / 56

Background

stjm

Estimation

Discussion

References

Our key question here is how are changes in the biomarker

trajectory associated with survival?

hi (t) = h0 (t) exp T v i + mi (t)

where for example

mi (t) = (0 + b0i ) + (1 + b1i )t + u Ti

mi (t) is termed the current value parameterisation

Michael J. Crowther

Joint Modelling

9 / 56

Background

stjm

Estimation

Discussion

References

Interaction effects

hi (t) = h0 (t) exp T vi1 + T {vi2 mi (t)}

where vi1 , vi2 Ui . We now have vector of association

parameters , providing different associations for different

covariate patterns.

Michael J. Crowther

Joint Modelling

10 / 56

Background

stjm

Estimation

Discussion

References

Time-dependent slope

We may be interested in how the slope or rate of change of

the biomarker is associated with survival:

(1)

hi (t) = h0 (t) exp T vi + 1 mi (t) + 2 mi0 (t)

with

dmi (t)

d{X Ti (t) + Z Ti (t)bi }

=

(2)

dt

dt

The added benefit of including the rate of change of CD4

trajectories within a joint model framework to model the risk

of progression to AIDS or death in HIV-positive patients was

conducted by Wolbers et al. (2010).

mi0 (t) =

Michael J. Crowther

Joint Modelling

11 / 56

Background

stjm

Estimation

Discussion

References

Random effects parameterisation

Finally, I define a time-independent association structure

hi (t) = h0 (t) exp T v i + T ( + b i )

(3)

random effect, plus the subject specific deviation, for example

hi (t) = h0 (t) exp T v i + 1 (0 + b0i )

(4)

where exp(1 ) is the hazard ratio for a one unit increase in the

baseline value of the longitudinal outcome i.e. the intercept.

Michael J. Crowther

Joint Modelling

12 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

13 / 56

Background

stjm

Estimation

Discussion

References

I

I

I

I

251 randomised to receive prednisone and 237 to placebo

Outcome is all-cause death

2968 measurements of prothrombin index (Anderson

et al., 1993; Henderson et al., 2002)

Interested in the association between prothrombin index

over time and survival

Michael J. Crowther

Joint Modelling

14 / 56

Background

Estimation

stjm

Discussion

References

.

id

pro

trt

_t0

_t

_d

1

1

1

38

31

27

placebo

placebo

placebo

0.000

0.244

0.381

0.244

0.381

0.413

0

0

1

518

518

518

518

518

95

110

82

88

74

prednisone

prednisone

prednisone

prednisone

prednisone

0.000

0.268

0.517

0.999

2.053

0.268

0.517

0.999

2.053

2.055

0

0

0

0

0

Michael J. Crowther

Joint Modelling

15 / 56

Background

stjm

Estimation

Discussion

References

200

Longitudinal response

100

150

50

Longitudinal response

100

50

0

0

Event

150

200

Censored

5

Measurement time

10

5

Measurement time

10

Michael J. Crowther

Joint Modelling

16 / 56

Background

stjm

Estimation

Discussion

References

200

Longitudinal response

100

150

50

Longitudinal response

100

50

0

0

Event

150

200

Censored

5

Measurement time

10

5

Measurement time

10

Michael J. Crowther

Joint Modelling

17 / 56

Background

stjm

Estimation

Discussion

References

Longitudinal response

100

0

50

Longitudinal response

100

50

0

15

Event

150

200

Censored

150

200

10

5

Time before censoring

15

10

5

Time before event

Michael J. Crowther

Joint Modelling

18 / 56

Background

stjm

Estimation

Discussion

References

Syntax

.

>

>

>

Michael J. Crowther

panel(varname )

survmodel(model )

[options ]

Joint Modelling

19 / 56

Background

stjm

Estimation

Discussion

References

Syntax

.

>

>

>

panel(varname )

survmodel(model )

[options ]

I

Longitudinal submodel:

I

I

I

I

I

I

Michael J. Crowther

ffp(numlist ) - Fixed FPs of time

rfp(numlist ) - Random FPs of time

frcs(#) - Fixed RCS of time

rrcs(#) - Random RCS of time

timeinteraction(varlist ) - covariates to interact

with fixed time variables

covariance(vartype ) - variance-covariance structure

of random effects

Joint Modelling

20 / 56

Background

stjm

Estimation

Discussion

References

Syntax

.

>

>

>

panel(varname )

survmodel(model )

[options ]

I

Survival submodel:

I

I

I

I

I

Michael J. Crowther

exponential, Weibull, Gompertz, splines on the log

hazard scale, Royston-Parmar, 2-component mixtures

survcov(varlist ) - Baseline covariates

df(#) - degrees of freedom for baseline hazard function

knots(numlist ) - internal knot locations

noorthog - suppress default orthogonalisation

Joint Modelling

21 / 56

Background

stjm

Estimation

Discussion

References

Syntax

.

>

>

>

panel(varname )

survmodel(model )

[options ]

I

Association:

I

I

I

Michael J. Crowther

derivassoc - 1st derivative (slope)

intassoc/assoc(numlist) - Random coefficient, e.g.

random intercept

nocoefficient - do not include fixed coefficient in

time-independent associations

assoccov(varlist ) adjust the association

parameter(s) by covariates

Joint Modelling

22 / 56

Background

stjm

Estimation

Discussion

References

Syntax

.

>

>

>

panel(varname )

survmodel(model )

[options ]

I

Maximisation:

I

I

I

I

I

Michael J. Crowther

gk(#) - number of Gauss-Kronrod quadrature nodes

adaptit(#) - number of adaptive quadrature iterations;

default is 5

nonadapt - use non-adaptive Gauss-Hermite quadrature

showadapt - display iteration log for adaptive

sub-routine

Joint Modelling

23 / 56

Background

stjm

Estimation

Discussion

References

Predictions

predict newvarname, option

I Longitudinal:

I

I

I

I

residuals - Subject level residuals

rstandard - Standardised residuals

reffects/reses - Empirical Bayes predictions of

random effects

or user specified times (timevar(varname )), and at specific

covariate patterns using at(varname # ...).

Michael J. Crowther

Joint Modelling

24 / 56

Background

stjm

Estimation

Discussion

References

Predictions

predict newvarname, option

I Survival:

I

I

I

I

I

survival - Survival function

cumhazard - Cumulative hazard function

martingale - Martingale residuals

stjmcondsurv - Conditional survival

or user specified times (timevar(varname )), and at specific

covariate patterns using at(varname # ...).

Michael J. Crowther

Joint Modelling

25 / 56

Background

. stjm

-> gen

-> gen

(where

stjm

Estimation

Discussion

References

double _time_1 = X^(1)

double _time_1_trt = trt * _time_1

X = _t0)

Fitting full model:

-> Conducting adaptive Gauss-Hermite quadrature

-- Iteration 0:

Adapted log

-- Iteration 1:

Adapted log

-- Iteration 2:

Adapted log

-- Iteration 3:

Adapted log

Iteration 0:

log likelihood

(output omitted )

Iteration 3:

log likelihood

-- Iteration 0:

Adapted log

-- Iteration 1:

Adapted log

-- Iteration 2:

Adapted log

Iteration 4:

log likelihood

Michael J. Crowther

likelihood = -14026.287

likelihood = -14026.87

likelihood = -14026.87

likelihood = -14026.87

= -14026.87

= -14023.698

likelihood = -14023.698

likelihood = -14023.698

likelihood = -14023.698

= -14023.698

Joint Modelling

26 / 56

Background

Estimation

stjm

Panel variable: id

Discussion

Number of obs.

=

Number of panels

=

Number of failures =

References

2968

488

292

Log-likelihood = -14023.698

Coef.

Longitudinal

_time_1

_time_1_trt

_cons

Survival

assoc:value

_cons

ln_lambda

trt

_cons

ln_gamma

_cons

Std. Err.

.4213469

.1185318

73.34061

.4670811

.6411165

.9924867

0.90

0.18

73.90

-.038828

.0036241

.1695929

1.12911

.1236181

.2421363

-.052448

.0489213

P>|z|

0.367

0.853

0.000

-.4941152

-1.138034

71.39537

1.336809

1.375097

75.28585

-10.71

0.000

-.0459312

-.0317248

1.37

4.66

0.170

0.000

-.0726941

.6545319

.4118799

1.603689

-1.07

0.284

-.148332

.043436

Estimate

Std. Err.

sd(_time_1)

sd(_cons)

corr(_time_1,_cons)

3.934646

19.30179

.0076874

.3532822

.7980948

.1040056

3.29973

17.79926

-.1936933

4.691729

20.93116

.2084466

sd(Residual)

17.22068

.2603811

16.71783

17.73865

id: Unstructured

Michael J. Crowther

Joint Modelling

27 / 56

Background

Estimation

stjm

Discussion

References

Patient 98

200

1.0

Patient 253

200

1.0

0.8

0.8

0.4

50

0.6

100

0.4

50

0.2

0.0

0

Michael J. Crowther

Survival probability

0.6

100

Biomarker

150

Survival probability

Biomarker

150

10

12

14

0.2

0.0

0

10

12

14

Follow-up time

Follow-up time

Longitudinal response

Predicted conditional survival

95% Confidence interval

Joint Modelling

28 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

29 / 56

Background

Estimation

stjm

Discussion

References

Joint likelihood

The full joint likelihood is

"Z

!

#

ni

N

Y

Y

p(yi (tij )|bi , ) p(bi |)p(Ti , di |bi , ) dbi

i=1

Michael J. Crowther

j=1

Joint Modelling

30 / 56

Background

Estimation

stjm

Discussion

References

Joint likelihood

The full joint likelihood is

"Z

!

#

ni

N

Y

Y

p(yi (tij )|bi , ) p(bi |)p(Ti , di |bi , ) dbi

i=1

j=1

where

(

p(yi (tij )|bi , ) = (2e2 )1/2 exp

2e2

Michael J. Crowther

Joint Modelling

30 / 56

Background

Estimation

stjm

Discussion

References

Joint likelihood

The full joint likelihood is

"Z

!

#

ni

N

Y

Y

p(yi (tij )|bi , ) p(bi |)p(Ti , di |bi , ) dbi

i=1

j=1

where

q/2

p(bi |) = (2|V |)

Michael J. Crowther

Joint Modelling

b 0 V 1 bi

exp i

2

31 / 56

Background

Estimation

stjm

Discussion

References

Joint likelihood

The full joint likelihood is

"Z

!

#

ni

N

Y

Y

p(yi (tij )|bi , ) p(bi |)p(Ti , di |bi , ) dbi

i=1

j=1

where

p(Ti , di |bi , ) = [h0 (Ti ) exp(mi (t) + vi )]di

Z Ti

exp

h0 (u) exp(mi (u) + vi )du

0

Michael J. Crowther

Joint Modelling

32 / 56

Background

Estimation

stjm

Discussion

References

Joint likelihood

The full joint likelihood is

"Z

!

#

ni

N

Y

Y

p(yi (tij )|bi , ) p(bi |)p(Ti , di |bi , ) dbi

i=1

j=1

where

p(Ti , di |bi , ) = [h0 (Ti ) exp(mi (t) + vi )]di

Z Ti

exp

h0 (u) exp(mi (u) + vi )du

0

Michael J. Crowther

Joint Modelling

32 / 56

Background

stjm

Estimation

Discussion

References

Gauss-Hermite quadrature

integrals (Pinheiro and Bates, 1995)

Z

x 2

f (x)dx

m

X

wq f (xq )

q=1

random effects

Michael J. Crowther

Joint Modelling

33 / 56

Background

Estimation

stjm

Discussion

References

u ~ N(0,1)

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

Discussion

References

u ~ N(0,1)

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

Discussion

References

u ~ N(0,s )

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

Discussion

References

u ~ N(0,s )

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

Discussion

References

u2

u1

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

Discussion

References

u1

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

34 / 56

Background

Estimation

stjm

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

Discussion

References

34 / 56

Background

Estimation

stjm

-6

Michael J. Crowther

-4

-2

Joint Modelling

0

u

Discussion

References

34 / 56

Background

stjm

Estimation

Discussion

References

Computation time

intercept and slope model with current value

parameterisation computation time is 21 seconds

Registry based data example of 5,000 patients with

100,000 measurements takes 10 minutes

Michael J. Crowther

Joint Modelling

35 / 56

Background

Estimation

stjm

Discussion

References

(briefly)

In Crowther et al. (2012) we compared non-adaptive and

adaptive Gauss-Hermite quadrature in evaluating the joint

likelihood

NAQ 5 nodes

Bias 95% CP

NAQ 15 nodes

Bias 95% CP

AQ 5 nodes

Bias 95% CP

Parameter

True value

Model

FPM (df=5)

-0.004

64.8

0.003

80.9

0.002

94.7

0.25

FPM (df=5)

-0.002

93.0

-0.002

91.7

-0.005

95.2

Michael J. Crowther

Joint Modelling

36 / 56

Background

Estimation

stjm

Discussion

References

(briefly)

In Crowther et al. (2012) we compared non-adaptive and

adaptive Gauss-Hermite quadrature in evaluating the joint

likelihood

NAQ 5 nodes

Bias 95% CP

NAQ 15 nodes

Bias 95% CP

AQ 5 nodes

Bias 95% CP

Parameter

True value

Model

FPM (df=5)

-0.004

64.8

0.003

80.9

0.002

94.7

0.25

FPM (df=5)

-0.002

93.0

-0.002

91.7

-0.005

95.2

Michael J. Crowther

Joint Modelling

36 / 56

Background

Estimation

stjm

Discussion

References

(briefly)

In Crowther et al. (2012) we compared non-adaptive and

adaptive Gauss-Hermite quadrature in evaluating the joint

likelihood

NAQ 5 nodes

Bias 95% CP

NAQ 15 nodes

Bias 95% CP

AQ 5 nodes

Bias 95% CP

Parameter

True value

Model

FPM (df=5)

-0.004

64.8

0.003

80.9

0.002

94.7

0.25

FPM (df=5)

-0.002

93.0

-0.002

91.7

-0.005

95.2

Michael J. Crowther

Joint Modelling

36 / 56

Background

Estimation

stjm

Discussion

References

1

0

10

10

1

.5

0

.5

Hazard function

1.5

Followup time

1.5

Followup time

Hazard function

.5

Hazard function

1

.5

0

Hazard function

1.5

1.5

Followup time

Michael J. Crowther

Joint Modelling

10

10

Followup time

37 / 56

Background

stjm

Estimation

Discussion

References

Longitudinal process

mij = 0i + 1i tij + ui

where

0i N(0, 1) 1i N(0, 0.252 ) and corr(0i , 1i ) = 0.25

with ui Bin(1, 0.5) and = 0.25. Measurements are taken

at tij = {0, 1, 2, 3, 4}. Measurements are then generated from:

yij N(mij , 0.52 )

Michael J. Crowther

Joint Modelling

38 / 56

Background

stjm

Estimation

Discussion

References

Survival process

Survival times are then generated from a 2-component mixture

Weibull baseline hazard (Crowther and Lambert, 2013)

S0 (t) = p exp(1 t 1 ) + (1 p)exp(2 t 2 )

and

log(h(t|bi )) = log(h0 (t)) + mi (t) + ui

with = 0.25 and = {0.25, 0.25}.

Joint models applied

I Weibull survival submodel

I 2-component mixture Weibull joint model

Michael J. Crowther

Joint Modelling

39 / 56

Background

Estimation

stjm

Discussion

References

Simulation results

Table : Bias and coverage in the association and treatment effects

Scenario Parameter

True

value

0.25

0.25

0.25

Weibull

Bias Coverage

Mixture Weibull

Bias Coverage

-0.067

-0.008

0.060

43.7 0.003

96.4 -0.007

85.6 0.009

96.6

96.4

95.6

-0.25 -0.103

0.25 -0.006

0.25 0.059

29.1 -0.007

96.0 -0.007

86.4 0.008

96.4

96.0

94.2

0.033

-0.008

0.059

72.4 0.003

95.8 -0.006

83.0 0.007

95.2

95.8

93.8

-0.25 -0.063

0.25 -0.006

0.25 0.050

44.4 0.019

96.2 -0.007

85.6 0.002

89.2

96.0

94.0

0.25

0.25

0.25

Michael J. Crowther

Joint Modelling

40 / 56

Background

Estimation

stjm

Discussion

References

Simulation results

Table : Bias and coverage in the association and treatment effects

Scenario Parameter

True

value

0.25

0.25

0.25

Weibull

Bias Coverage

Mixture Weibull

Bias Coverage

-0.067

-0.008

0.060

43.7 0.003

96.4 -0.007

85.6 0.009

96.6

96.4

95.6

-0.25 -0.103

0.25 -0.006

0.25 0.059

29.1 -0.007

96.0 -0.007

86.4 0.008

96.4

96.0

94.2

0.033

-0.008

0.059

72.4 0.003

95.8 -0.006

83.0 0.007

95.2

95.8

93.8

-0.25 -0.063

0.25 -0.006

0.25 0.050

44.4 0.019

96.2 -0.007

85.6 0.002

89.2

96.0

94.0

0.25

0.25

0.25

Michael J. Crowther

Joint Modelling

40 / 56

Background

Estimation

stjm

Discussion

References

Simulation results

Table : Bias and coverage in the association and treatment effects

Scenario Parameter

True

value

0.25

0.25

0.25

Weibull

Bias Coverage

Mixture Weibull

Bias Coverage

-0.067

-0.008

0.060

43.7 0.003

96.4 -0.007

85.6 0.009

96.6

96.4

95.6

-0.25 -0.103

0.25 -0.006

0.25 0.059

29.1 -0.007

96.0 -0.007

86.4 0.008

96.4

96.0

94.2

0.033

-0.008

0.059

72.4 0.003

95.8 -0.006

83.0 0.007

95.2

95.8

93.8

-0.25 -0.063

0.25 -0.006

0.25 0.050

44.4 0.019

96.2 -0.007

85.6 0.002

89.2

96.0

94.0

0.25

0.25

0.25

Michael J. Crowther

Joint Modelling

40 / 56

Background

Estimation

stjm

Discussion

References

Table : Bias and coverage in the association and treatment effects

Scenario Parameter

True

value

0.25

0.25

0.25

Weibull

Bias Coverage

Mixture Weibull

Bias Coverage

-0.067

-0.008

0.060

43.7 0.003

96.4 -0.007

85.6 0.009

96.6

96.4

95.6

-0.25 -0.103

0.25 -0.006

0.25 0.059

29.1 -0.007

96.0 -0.007

86.4 0.008

96.4

96.0

94.2

0.033

-0.008

0.059

72.4 0.003

95.8 -0.006

83.0 0.007

95.2

95.8

93.8

-0.25 -0.063

0.25 -0.006

0.25 0.050

44.4 0.019

96.2 -0.007

85.6 0.002

89.2

96.0

94.0

0.25

0.25

0.25

Michael J. Crowther

Joint Modelling

40 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

41 / 56

Background

stjm

Estimation

Discussion

References

I

mellitus from the General Practice Research Database

(GPRD) (Ara et al., 2012)

Michael J. Crowther

Joint Modelling

42 / 56

Background

stjm

Estimation

Discussion

References

I

mellitus from the General Practice Research Database

(GPRD) (Ara et al., 2012)

Covariates of interest include gender and age (in years) at

baseline, and repeated measures of Systolic Blood

Pressure (SBP), of which we have 107,347 measurements

Michael J. Crowther

Joint Modelling

42 / 56

Background

stjm

Estimation

Discussion

References

I

mellitus from the General Practice Research Database

(GPRD) (Ara et al., 2012)

Covariates of interest include gender and age (in years) at

baseline, and repeated measures of Systolic Blood

Pressure (SBP), of which we have 107,347 measurements

The event of interest in this case is non-fatal stroke, of

which 278 were observed, with maximum follow-up of 22

years

Michael J. Crowther

Joint Modelling

42 / 56

Background

stjm

Estimation

Discussion

References

I

mellitus from the General Practice Research Database

(GPRD) (Ara et al., 2012)

Covariates of interest include gender and age (in years) at

baseline, and repeated measures of Systolic Blood

Pressure (SBP), of which we have 107,347 measurements

The event of interest in this case is non-fatal stroke, of

which 278 were observed, with maximum follow-up of 22

years

Patients enter the cohort at the time of first SBP

measurement

Michael J. Crowther

Joint Modelling

42 / 56

Background

stjm

Estimation

Discussion

References

I

mellitus from the General Practice Research Database

(GPRD) (Ara et al., 2012)

Covariates of interest include gender and age (in years) at

baseline, and repeated measures of Systolic Blood

Pressure (SBP), of which we have 107,347 measurements

The event of interest in this case is non-fatal stroke, of

which 278 were observed, with maximum follow-up of 22

years

Patients enter the cohort at the time of first SBP

measurement

Research aim is to investigate the association between

baseline SBP and the risk of stroke, utilising the repeated

measures (Crowther et al., 2013b)

Michael J. Crowther

Joint Modelling

42 / 56

Background

stjm

Estimation

Discussion

References

Stroke

50

100

150

200

250

Censored

Graphs by Stroke

mellitus. Box plots of baseline SBP for patients who were censored

or who suffered a stroke.

Michael J. Crowther

Joint Modelling

43 / 56

Background

stjm

Estimation

Discussion

References

Preliminary analyses

We can build our longitudinal and survival submodels

separately initially, to investigate the trajectory over time, and

the shape of the baseline hazard function

I We use restricted cubic splines to model the profile of

SBP over time (Durrleman and Simon, 1989)

I We use the Royston-Parmar survival model to model the

risk of stroke (Royston and Parmar, 2002)

I Model selection criteria can be used to guide the selection

of number of spline parameters (Rutherford et al., 2014)

Michael J. Crowther

Joint Modelling

44 / 56

Background

Estimation

stjm

Patient 2

5

10

15

Follow-up time (years)

5

10

15

Follow-up time (years)

Patient 6

SBP

Patient 7

200

175

150

125

100

5

10

15

Follow-up time (years)

5

10

15

Follow-up time (years)

Patient 8

Patient 9

200

175

150

125

100

5

10

15

Follow-up time (years)

SBP

SBP

SBP

5

10

15

Follow-up time (years)

200

175

150

125

100

5

10

15

Follow-up time (years)

200

175

150

125

100

0

200

175

150

125

100

Patient 5

SBP

SBP

Patient 4

200

175

150

125

100

Observed SBP

References

Patient 3

200

175

150

125

100

SBP

200

175

150

125

100

SBP

SBP

Patient 1

Discussion

200

175

150

125

100

5

10

15

Follow-up time (years)

5

10

15

Follow-up time (years)

Predicted SBP

had at least 10 measurements. The dashed line represents the

fitted longitudinal trajectories based on the joint model.

Michael J. Crowther

Joint Modelling

45 / 56

Background

stjm

Estimation

Discussion

References

mi (tij ) = (0 + b0i ) + 1 agei + 2 sexi + 3 BMIi

+ F sF (tij ; kF ) + bR sR (tij ; kR )

(5)

measurement time with corresponding fixed effects, F , with

knot locations kF , and sR (tij ; kR ) is the restricted cubic spline

basis of measurement time with corresponding random effects,

bR , and knot locations kR . AIC/BIC selected 5 degrees of

freedom for sF (tij ; kF ), and 1 degree of freedom for sR (tij ; kR ).

Michael J. Crowther

Joint Modelling

46 / 56

Background

stjm

Estimation

Discussion

References

log Hi (t) = log H0 (t) + 1 agei + 2 sexi

+ 3 BMIi + 2 (0 + b0i )

(6)

(7)

where

AIC/BIC selected 2 degrees of freedom for s(log(t); , k S ).

Michael J. Crowther

Joint Modelling

47 / 56

Background

stjm

Estimation

Discussion

References

Standard RP model

Coefficient

95% CI

Survival model:

Baseline SBP/10 (2 )

Age (years)

Sex (male)

BMI (kg/m2 )

0.048 0.036 0.060

0.011 -0.233 0.254

0.011 -0.015 0.037

Longitudinal model:

Intercept

Age (years)

Sex (male)

BMI (kg/m2 )

RCS1

RCS2

RCS3

RCS4

RCS5

e

RCS - Restricted Cubic Spline

Michael J. Crowther

Joint model

Coefficient

95% CI

Joint Modelling

0.181

0.050

-0.010

0.013

0.102

0.038

-0.254

-0.012

0.261

0.062

0.234

0.039

0.025 0.022 0.029

-0.252 -0.332 -0.171

0.003 -0.005 0.011

-0.080 -0.121 -0.039

-0.006 -0.019 0.006

-0.001 -0.010 0.007

0.003 0.000 0.006

0.000 -0.001 0.001

1.522 1.515 1.528

48 / 56

Background

stjm

Estimation

Discussion

References

Standard RP model

Coefficient

95% CI

Survival model:

Baseline SBP/10 (2 )

Age (years)

Sex (male)

BMI (kg/m2 )

0.048 0.036 0.060

0.011 -0.233 0.254

0.011 -0.015 0.037

Longitudinal model:

Intercept

Age (years)

Sex (male)

BMI (kg/m2 )

RCS1

RCS2

RCS3

RCS4

RCS5

e

RCS - Restricted Cubic Spline

Michael J. Crowther

Joint model

Coefficient

95% CI

Joint Modelling

0.181

0.050

-0.010

0.013

0.102

0.038

-0.254

-0.012

0.261

0.062

0.234

0.039

0.025 0.022 0.029

-0.252 -0.332 -0.171

0.003 -0.005 0.011

-0.080 -0.121 -0.039

-0.006 -0.019 0.006

-0.001 -0.010 0.007

0.003 0.000 0.006

0.000 -0.001 0.001

1.522 1.515 1.528

48 / 56

Background

Estimation

stjm

Discussion

References

Survival probability

1.0

0.9

0.8

0.7

0.6

0

10

Follow-up time (years)

FPM model: SBP = 130

FPM model: SBP = 200

15

20

Joint model: SBP = 130

Joint model: SBP = 200

joint models, for a female, aged 60 years, BMI of 30, with SBP of

90, 130 or 200.

Michael J. Crowther

Joint Modelling

49 / 56

Background

stjm

Estimation

Discussion

References

Table : Results of sensitivity analysis.

Splines of time

Coefficient

95% CI

Survival model:

Baseline SBP/10 (2 )

Age (years)

Sex

BMI (kg/m2 )

0.181

0.050

-0.010

0.013

Longitudinal model:

Age (years)

Sex

BMI (kg/m2 )

Intercept

-0.252 -0.332 -0.171

0.003 -0.005 0.011

13.006 12.629 13.382

Michael J. Crowther

Joint Modelling

0.102

0.038

-0.254

-0.012

Linear time

Coefficient

95% CI

0.261

0.062

0.234

0.039

0.145

0.051

-0.010

0.014

0.066

0.038

-0.253

-0.012

0.224

0.063

0.233

0.039

-0.259 -0.340 -0.179

0.000 -0.007 0.008

13.218 12.843 13.593

50 / 56

Background

stjm

Estimation

Discussion

References

Table : Results of sensitivity analysis.

Splines of time

Coefficient

95% CI

Survival model:

Baseline SBP/10 (2 )

Age (years)

Sex

BMI (kg/m2 )

0.181

0.050

-0.010

0.013

Longitudinal model:

Age (years)

Sex

BMI (kg/m2 )

Intercept

-0.252 -0.332 -0.171

0.003 -0.005 0.011

13.006 12.629 13.382

Michael J. Crowther

Joint Modelling

0.102

0.038

-0.254

-0.012

Linear time

Coefficient

95% CI

0.261

0.062

0.234

0.039

0.145

0.051

-0.010

0.014

0.066

0.038

-0.253

-0.012

0.224

0.063

0.233

0.039

-0.259 -0.340 -0.179

0.000 -0.007 0.008

13.218 12.843 13.593

50 / 56

Background

stjm

Estimation

Discussion

References

Outline

Background

stjm

Estimation

SBP and risk of stroke

Discussion

Michael J. Crowther

Joint Modelling

51 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

sources, as electronic healthcare record linkage moves to

the forefront of life science strategy (Jutte et al., 2011)

Michael J. Crowther

Joint Modelling

52 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

sources, as electronic healthcare record linkage moves to

the forefront of life science strategy (Jutte et al., 2011)

Current cardiovascular risk scores use observed baseline

biomarkers

Michael J. Crowther

Joint Modelling

52 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

sources, as electronic healthcare record linkage moves to

the forefront of life science strategy (Jutte et al., 2011)

Current cardiovascular risk scores use observed baseline

biomarkers

Opportunities to utilise the joint model framework in

prognostic modelling are great

I

I

I

Michael J. Crowther

patients

Computation time is now viable

Gould et al. (2014)

Joint Modelling

52 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

conducted

Michael J. Crowther

Joint Modelling

53 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

conducted

Extensions

I

Michael J. Crowther

longitudinal-survival model

Delayed entry

Joint Modelling

53 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

conducted

Extensions

I

longitudinal-survival model

Delayed entry

Michael J. Crowther

Joint Modelling

53 / 56

Background

stjm

Estimation

Discussion

References

Discussion

I

conducted

Extensions

I

I

I

longitudinal-survival model

Delayed entry

We will be running a course on joint modelling in Stata

next year at the University of Leicester

Michael J. Crowther

Joint Modelling

53 / 56

Background

stjm

Estimation

Discussion

References

References I

Anderson, P. K., Borgan, ., Gill, R. D., and Keiding, N. Statistical Models Based on

Counting Processes. New York, Springer, 1993.

Ara, R., Blake, L., Gray, L., Hernandez, M., Crowther, M., Dunkley, A., Warren, F.,

Jackson, R., Rees, A., Stevenson, M., Abrams, K., Cooper, N., Davies, M., Khunti,

K., and Sutton, A. What is the clinical effectiveness and cost-effectiveness of using

drugs in treating obese patients in primary care? A systematic review. Health

Technol Assess, 16(5):1202, Feb 2012.

Crowther, M. J. and Lambert, P. C. Simulating biologically plausible complex survival

data. Stat Med, 32(23):41184134, 2013.

Crowther, M. J., Abrams, K. R., and Lambert, P. C. Flexible parametric joint

modelling of longitudinal and survival data. Stat Med, 31(30):44564471, 2012.

Crowther, M. J., Abrams, K. R., and Lambert, P. C. Joint modeling of longitudinal

and survival data. Stata J, 13(1):165184, 2013a.

Crowther, M. J., Lambert, P. C., and Abrams, K. R. Adjusting for measurement error

in baseline prognostic biomarkers included in a time-to-event analysis: A joint

modelling approach. BMC Med Res Methodol, 13(146), 2013b.

Durrleman, S. and Simon, R. Flexible Regression Models with Cubic Splines. Stat

Med, 8(5):551561, 1989.

Michael J. Crowther

Joint Modelling

54 / 56

Background

stjm

Estimation

Discussion

References

References II

Faucett, C. L. and Thomas, D. C. Simultaneously modelling censored survival data

and repeatedly measured covariates: a Gibbs sampling approach. Stat Med, 15

(15):16631685, 1996.

Gould, A. L., Boye, M. E., Crowther, M. J., Ibrahim, J. G., Quartey, G., Micallef, S.,

and Bois, F. Y. Joint modeling of survival and longitudinal non-survival data:

current methods and issues. Report of the DIA Bayesian joint modeling working

group. Stat Med, 2014.

Henderson, R., Diggle, P., and Dobson, A. Joint modelling of longitudinal

measurements and event time data. Biostatistics, 1(4):465480, 2000.

Henderson, R., Diggle, P., and Dobson, A. Identification and efficacy of longitudinal

markers for survival. Biostatistics, 3(1):3350, 2002.

Jutte, D. P., Roos, L. L., and Brownell, M. D. Administrative record linkage as a tool

for public health research. Annu Rev Public Health, 32:91108, 2011.

Pinheiro, J. C. and Bates, D. M. Approximations to the log-likelihood function in the

nonlinear mixed-effects model. J Comput Graph Statist, 4(1):pp. 1235, 1995.

Proust-Lima, C. and Taylor, J. M. G. Development and validation of a dynamic

prognostic tool for prostate cancer recurrence using repeated measures of

posttreatment PSA: a joint modeling approach. Biostatistics, 10(3):535549, 2009.

Michael J. Crowther

Joint Modelling

55 / 56

Background

stjm

Estimation

Discussion

References

References III

Proust-Lima, C., Sene, M., Taylor, J. M., and Jacqmin-Gadda, H. Joint latent class

models for longitudinal and time-to-event data: A review. Stat Methods Med Res,

Apr 2012.

Rizopoulos, D. Joint Models for Longitudinal and Time-to-Event Data With

Applications in R. Chapman & Hall, 2012.

Royston, P. and Parmar, M. K. B. Flexible Parametric Proportional Hazards and

Proportional Odds Models for Censored Survival Data, with Application to

Prognostic Modelling and Estimation of Treatment Effects. Stat Med, 21(15):

21752197, 2002.

Rutherford, M. J., Crowther, M. J., and Lambert, P. C. The use of restricted cubic

splines to approximate complex hazard functions in the analysis of time-to-event

data: a simulation study. J Statist Comput Simulation, 2014.

Wolbers, M., Babiker, A., Sabin, C., Young, J., Dorrucci, M., Ch

ene, G., Mussini, C.,

Porter, K., Bucher, H. C., and CASCADE. Pretreatment CD4 cell slope and

progression to AIDS or death in HIV-infected patients initiating antiretroviral

therapythe CASCADE collaboration: a collaboration of 23 cohort studies. PLoS

Med, 7(2):e1000239, 2010.

Wulfsohn, M. S. and Tsiatis, A. A. A joint model for survival and longitudinal data

measured with error. Biometrics, 53(1):330339, 1997.

Michael J. Crowther

Joint Modelling

56 / 56

## Mult mai mult decât documente.

Descoperiți tot ce are Scribd de oferit, inclusiv cărți și cărți audio de la editori majori.

Anulați oricând.