Results
A Bayesian Joint Dispersion Model
With Flexible Links
Background HIV/AIDS Application Results
Summary
1
Background
Longitudinal studies
Outline
2
HIV/AIDS Application
Data
Exploratory analysis
Joint Model
3
Results
Results
Questions of interest
Biomarker, Y1
•
e.g. CD4 counts, collected repeatedly over time (longitudinal data)
time to an event of interest, Y2
•
e.g. death from any cause (survival data)
Separate Analysis
•
does treatment affect survival?
•
are the average longitudinal evolutions different between males
and females?
Joint Analysis
Background HIV/AIDS Application Results Longitudinal studies Outline
Questions of interest
Biomarker, Y1
•
e.g. CD4 counts, collected repeatedly over time (longitudinal data)
time to an event of interest, Y2
•
e.g. death from any cause (survival data)
Separate Analysis
•
does treatment affect survival?
•
are the average longitudinal evolutions different between males
and females?
Joint Analysis
•
what is the effect of the missing information due to drop-out in
assessing the trends of the repeated measures?
•
what is the effect of the longitudinal evolution of CD4 cell count in
Results
Questions of interest
Biomarker, Y1
•
e.g. CD4 counts, collected repeatedly over time (longitudinal data)
time to an event of interest, Y2
•
e.g. death from any cause (survival data)
Separate Analysis
•
does treatment affect survival?
•
are the average longitudinal evolutions different between males
and females?
Joint Analysis
Background
HIV/AIDS Application Results
Longitudinal studies Outline
Correlated data in longitudinal studies
0 1 2 3 4 5 0.0 2.5 5.0 7.5 10.0 Time Biomar k er v alues Longitudinal Trajectory 0 0.1 0.2 0.3 0.4 0.5 0.0 2.5 5.0 7.5 10.0 Time hazard Hazard Function Latent Structure
Longitudinal Marker (Y1) Time-To-Event (Y2)
If the two processes are associated ⇒ define a model for their joint probability
distribution:
f (y
1, y
2)
Results Joint Model
Analysing HIV/AIDS data through
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Data
network of 88 laboratories located in every state in Brazil during
Jan 2002 – Dec 2006;
Sample: n = 500 individuals; 2757 repeated measurements;
Outcomes:
CD4
+
T lymphocyte counts, Y ,
and
time-to-death, T
;
Covariates:
age
(<50=0, ≥50=1);
sex
(Female=0, Male=1);
PrevOI
(previous opportunistic infection at study entry=1, no
previous infection=0);
measurement times
;
date of diagnosis
;
date of death
;
failure indicator, δ
;
Patients: 34 deaths. 88% between 15 and 49 years old; 60%
males. 61% no previous infection. Initial CD4 median: 269
Results Joint Model
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Intra-individual variance (dispersion)
0 100 200 300 400 500 0 10 20 30 40
Ordered subjects by std. deviation
C D 4 ●●● ● ● ● ●●● ● ●●●● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●●●● ● ● ● ● ●● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ●● ●● ● ● ● ●●● ●● ● ● ● ● ● ● ●● ●● ● ● ●● ●● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ●●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●●●● ● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ●● ●● ● ● ● ●●● ● ●● ●● ● ● ● ●● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●● ●●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ●●● ● ●● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ●● ● ● ●● ● ● ●●● ● ● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ●● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ●● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●
lowess smooth of the CD4counts connecting the standard deviations
Individuals values of the
√
CD4 vs Std. Deviation (ordered) suggests considerable
within-subject variance heterogeneity. Individuals with higher
√
CD4 values are
associated with a higher variability.
Results Joint Model
Longitudinal specification
Mixed-effects dispersion model
(McLain et al. 2012)
y
ij|b
i, σ
i2∼ N m
i(t
ij), σ
i2,
j = 1, . . . , n
i(1)
m
i(t
ij) = β
> 1x
1i(t
ij) + b
> 1iw
1i(t
ij),
(2)
σ
i2= σ
02exp{ β
> 2x
2i(t
ij) + b
> 2iw
2i(t
ij) },
(3)
•
y
i= (y
i1, . . . , y
ini) → n
iobserved repeated measures,
√
CD4
•
t
i= (t
i1, . . . , t
ini) →
visiting times
•
x
1i, x
2i, w
1iand w
2i→ individual covariates (time-dependent?)
•
β
1and β
2→ population parameters
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Longitudinal specification
Mixed-effects dispersion model
(McLain et al. 2012)
y
ij|b
i, σ
i2∼ N m
i(t
ij), σ
i2,
j = 1, . . . , n
i(1)
m
i(t
ij) = β
> 1x
1i(t
ij) + b
> 1iw
1i(t
ij),
(2)
σ
i2= σ
02exp{ β
> 2x
2i(t
ij) + b
> 2iw
2i(t
ij) },
(3)
•
y
i= (y
i1, . . . , y
ini) → n
iobserved repeated measures,
√
CD4
•
t
i= (t
i1, . . . , t
ini) →
visiting times
•
x
1i, x2i, w1iand w2i
→ individual covariates (time-dependent?)
•
β
1and β2
→ population parameters
Results Joint Model
Longitudinal outcome
√
CD4
ij
|b
i
, σ
2
i
∼ N m
i
(t
ij
), σ
i
2
Longitudinal mean
m
i(t
ij) = β
10+β
11sex
i+β
12age
i+β
13PrevOI
i+β
14t
ij+b
1i,1+b
1i,2t
ij(4)
Dispersion model (3) may assume:
σ
2i= σ
2
0
exp{β
21sex + β
22age + β
23PrevOI + b
2i},
(5)
σ
2i= σ
2 0exp{b
2i},
(6)
σ
2i,
(7)
σ
2i= σ
2 0.
(8)
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Longitudinal outcome
√
CD4
ij
|b
i
, σ
2
i
∼ N m
i
(t
ij
), σ
i
2
Longitudinal mean
m
i(t
ij) = β
10+β
11sex
i+β
12age
i+β
13PrevOI
i+β
14t
ij+b
1i,1+b
1i,2t
ij(4)
Dispersion model (3) may assume:
σ
2i= σ
2
0
exp{β
21sex + β
22age + β
23PrevOI + b
2i},
(5)
σ
2i= σ
2 0exp{b
2i},
(6)
σ
2i,
(7)
σ
2i= σ
2 0.
(8)
Priors
•
β
1p∼ N (0, 100); p = 0, . . . , 4 and β2q
∼ N (0, 100); q = 1, . . . , 3
•
b
i|Σ ∼ N
p(0, Σ);
Σ
−1∼ Wish(R, ξ)
•
log(σ0
) ∼ U (−100, 100); or log(σ
i) ∼ U (−100, 100)
Results Joint Model
Survival specification
Time-dependent coefficients (Penalized cubic B-Splines)
h
i(t|b
i, σ
i) = h
0(t) exp{β
>
3
x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(9)
•
C
i{.} → specifies which components of the longitudinal process
are related to h
i(.)
•
Link → Shared parameters
I
b
i, σ
i•
x
3i→ baseline covariates
•
β
3→ population parameters
•
h
0(t) →
parametric (e.g. Weibull); P-Splines; Piecewise constant
function.
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Survival specification
Time-dependent coefficients (Penalized cubic B-Splines)
h
i(t|b
i, σ
i) = h
0(t) exp{β
>
3
x
3i+ C
i{
b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(9)
•
C
i{.} → specifies which components of the longitudinal process
are related to h
i(.)
•
Link → Shared parameters
I
bi, σi
•
x
3i→ baseline covariates
•
β
3→ population parameters
•
h
0(t) →
parametric (e.g. Weibull); P-Splines; Piecewise constant
function.
•
g(t) = (g
1(t), . . . , g
L(t)) →
suitable vector of smooth functions
Results Joint Model
Survival specification
Time-dependent coefficients (Penalized cubic B-Splines)
h
i(t|b
i, σ
i) = h
0(t) exp{β
>
3
x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(9)
•
C
i{.} → specifies which components of the longitudinal process
are related to h
i(.)
•
Link → Shared parameters
I
b
i, σ
i•
x
3i→ baseline covariates
•
β
3→ population parameters
•
h
0(t) →parametric (e.g. Weibull); P-Splines; Piecewise constant
function.
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Time-to-death
h
i(t|b
i, σ
i) = h
0(t) exp{β
> 3x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(10)
all models
β
3
>
x
3i
= β
31sex
i
+ β
32age
i
+ β
33PrevOI
i
(11)
C
i
(.)
may assume:
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)b
2i
(12)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)σ
i
(13)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
(14)
g
1
(t), g
2
(t), g
3
(t) →
Penalized Splines with 19 internal knots.
•
g
l(t) =
P
19 q=1γ
lqB
lq(t),
l = 1, 2, 3
•
γ
l1∼ N (0, 1000), γ
lq|τ
l2∼ N (γ
l,q−1, τ
l2),
q = 2, . . . , 19
•
1/τ
2 l∼ G(0.001, 0.001), l = 0, 1, 2, 3.
3 scenarios for h
0
(t) = log(g
0
(t)) →
Weibull,
Penalized Splines
Results Joint Model
Time-to-death
h
i(t|b
i, σ
i) = h
0(t) exp{β
> 3x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(10)
all models
β
3
>
x
3i
= β
31
sex
i
+ β
32
age
i
+ β
33
PrevOI
i
(11)
C
i
(.)
may assume:
C
i
(.) = g1(t)b1i,1
+ g2(t)b1i,2
+ g3(t)b2i
(12)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)σ
i
(13)
C
i
(.) = g1(t)b1i,1
+ g2(t)b1i,2
(14)
g
1
(t), g
2
(t), g
3
(t) →
Penalized Splines with 19 internal knots.
•
g
l(t) =
P
19
q=1
γ
lqB
lq(t),
l = 1, 2, 3
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Time-to-death
h
i(t|b
i, σ
i) = h
0(t) exp{β
> 3x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(10)
all models
β
3
>
x
3i
= β
31
sex
i
+ β
32
age
i
+ β
33
PrevOI
i
(11)
C
i
(.)
may assume:
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)b
2i
(12)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)σ
i
(13)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
(14)
g
1(t), g2(t), g3(t) →
Penalized Splines with 19 internal knots.
•
g
l(t) =
P
19 q=1γ
lqB
lq(t),
l = 1, 2, 3
•
γ
l1∼ N (0, 1000), γ
lq|τ
l2∼ N (γ
l,q−1, τ
l2),
q = 2, . . . , 19
•
1/τ
2 l∼ G(0.001, 0.001), l = 0, 1, 2, 3.
3 scenarios for h
0
(t) = log(g
0
(t)) →
Weibull,
Penalized Splines
Results Joint Model
Time-to-death
h
i(t|b
i, σ
i) = h
0(t) exp{β
> 3x
3i+ C
i{b
i, σ
i; g(t)}} = h
0(t) exp{%
i(t)}
(10)
all models
β
3
>
x
3i
= β
31
sex
i
+ β
32
age
i
+ β
33
PrevOI
i
(11)
C
i
(.)
may assume:
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)b
2i
(12)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
+ g
3
(t)σ
i
(13)
C
i
(.) = g
1
(t)b
1i,1
+ g
2
(t)b
1i,2
(14)
g
1
(t), g
2
(t), g
3
(t) →
Penalized Splines with 19 internal knots.
•
g
l(t) =
P
19
q=1
γ
lqB
lq(t),
l = 1, 2, 3
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Joint likelihood
We consider:
y
i
|b
i
⊥ T
i
|b
i
;
y
ij
|b
i
⊥ y
il
|b
i
, j 6= l
non-informative right censoring
L(θ, b, σ | D) =
N
Y
i=1
n
iY
j=1
p(y
i
(t
ij
)|θ, b
i
, σ
2
i
)
p(T
i
, δ
i
|θ, b
i
, σ
i
)
where
D = {D
i
}
N
i=1
= {(y
i
, t
i
, T
i
, δ
i
)}
N
i=1
→ observed data for the N
independent individuals
θ →
other parameters;
Results Joint Model
Joint likelihood
We consider:
y
i
|b
i
⊥ T
i
|b
i
;
y
ij
|b
i
⊥ y
il
|b
i
, j 6= l
non-informative right censoring
L(θ, b, σ | D) =
N
Y
i=1
n
iY
j=1
p(y
i
(t
ij
)|θ, b
i
, σ
2
i
)
p(T
i
, δ
i
|θ, b
i
, σ
i
)
where
p(y
i
(t
ij
) | θ, b
i
, σ
i
2
) =
1
p2πσ
2
i
exp
(
−
[y
i
(t
ij
) − m
i
(t
ij
)]
2
2σ
2
i
)
Background HIV/AIDS Application Results Data Exploratory analysis Joint Model
Joint likelihood
We consider:
y
i
|b
i
⊥ T
i
|b
i
;
y
ij
|b
i
⊥ y
il
|b
i
, j 6= l
non-informative right censoring
L(θ, b, σ | D) =
N
Y
i=1
n
iY
j=1
p(y
i
(t
ij
)|θ, b
i
, σ
2
i
)
p(T
i
, δ
i
|θ, b
i
, σ
i
)
where
p(T
i
, δ
i
| θ, b
i
, σ
i
) =h(T
i
|θ, b
i
, σ
i
)
δi
× S(T
i
|θ, b
i
, σ
i
)
=[h
0
(T
i
) exp{β
>
3
x3i
+ C
i
{b
i
, σ
i
; g(t)}}]
δi
×
exp
(
−
Z
Ti
0
h
0
(u) exp{β
3
>
x
3i
+ C
i
{b
i
, σ
i
; g(t)}}du
Results
Models comparison
MCMC simulation within WinBUGS.
Longitudinal model
Survival model
m
iσ
i2%
i(t)
h
0Weibull P-Spline Piecewise
(4)
(5)
(11) + (12)
14671
12573
14317
(4)
(6)
14700
12848
14483
(4)
(5)
(11) + (13)
14307
12605
13365
(4)
(6)
14452
12917
13571
(4)
(7)
13134
12104 ♣ 12921
(4)
(8)
13956
12887
13533
(4)
(5)
(11) + (14)
14811
13334
14463
(4)
(6)
14923
13688
14599
(4)
(7)
14314
13144
13968
(4)
(8)
14627
13553
14355
(4)
(8)
(11)+g
1b
1i,1+g
2b
1i,216984
15779
16383
Background HIV/AIDS Application
Results
Posterior estimates for the time-dependent coefficients
0 1 2 3 4 5 −3 −2 −1 0 t g0 (t) 0 1 2 3 4 5 −2 −1 0 1 2 3 t g1 (t) 0 1 2 3 4 5 −2 −1 0 1 2 t g2 (t) 0 1 2 3 4 5 −1 0 1 2 3 4 5 t g3 (t)
Figure 1:
Posterior mean estimates, together with the corresponding 95% Credible Bands (CB),
for the selected model ♣. The top left panel shows g
0= log (h
0)
and the subsequent panels have
BREZGER, A.,ANDLANG, S.
Simultaneous probability statements for bayesian p-splines. Statistical Modelling 8, 2 (2008), 141–168.
FAUCETT, C. L.,ANDTHOMAS, D. C.
Simultaneously modelling censored survival data and repeatedly measured covariates: a gibbs sampling approach. Statistics in Medicine 15, 15 (Aug 1996), 1663–1685.
IBRAHIM, J. G., CHEN, M. H.,ANDSINHA, D.
Bayesian Survival Analysis. Springer-Verlag, 2001.
LANG, S.,ANDBREZGER, A.
Bayesian p-splines.
Journal of Computational and Graphical Statistics 13, 1 (2004), 183–212.
LUNN, D., THOMAS, A., BEST, N.,ANDSPIEGELHALTER, D.
Winbugs – a bayesian modelling framework: Concepts, structure, and extensibility. Statistics and Computing 10 (2000), 325–337.
MCLAIN, A. C., LUM, K. J.,ANDSUNDARAM, R.
A joint mixed effects dispersion model for menstrual cycle length and time-to-pregnancy. Biometrics 68, 2 (Feb 2012), 648–656.
RIZOPOULOS, D.
Joint Models for Longitudinal and Time-to-Event Data With Applications in R. Chapman and Hall/CRC, 2012.