Weibull Extension of Bivariate Exponential Regression

Stat Papers (2009) 50:2949
DOI 10.1007/s00362-007-0057-4
REGULAR ARTI CLE
Weibull extension of bivariate exponential regression
model with different frailty distributions
David D. Hanagal
Received: 2 August 2006 / Revised: 15 February 2007 / Published online: 7 March 2007
Springer-Verlag 2007
Abstract We propose bivariate Weibull regression model with frailty which
is generated by a gamma or positive stable or power variance function distri-
bution. We assume that the bivariate survival data follows bivariate Weibull of
Hanagal (Econ Qual Control 19:8390, 2004; Econ Qual Control 20:143150,
2005a; Stat Pap 47:137148, 2006a; Stat Methods, 2006b). There are some inter-
esting situations like survival times in genetic epidemiology, dental implants
of patients and twin births (both monozygotic and dizygotic) where genetic
behavior (which is unknown and random) of patients follows known frailty
distribution. These are the situations which motivate to study this particular
model.
Keywords Bivariate Weibull Frailty Gamma Positive stable
Power variance function Parametric regression Survival times
1 Introduction
The shared gamma frailty models was suggested by Clayton (1978) for the
analysis of the correlation between clustered survival times in genetic epidemi-
ology. An advantage is that without covariates its mathematical properties are
convenient for estimation [See (Oakes 1982, 1986)]. However, when adjusting
David D. Hanagal is on leave from Department of Statistics, University of Pune, Pune 411007,
India.
D. D. Hanagal (B)
Department of Statistics and Probability, Michigan State University,
East Lansing, MI 48824, USA
e-mail: hanagal@stt.msu.edu; david
hanagal@yahoo.co.in
30 D. D. Hanagal
for environment risk factors the analysis of the clustering is more difcult [See
(Parner 1998)]. Until recently, a lack of theory and reliable software had pre-
vented widespread use of the model.
In a frailty model, it is absolutely necessary to be able to include explana-
tory variables. The reason is that the frailty describes the inuence of common
unknown factors. If some common covariates are included in the model, the
variation owing to unknown covariates should be reduced. Common covariates
are common for all members of the group.
For monozygotic twins, examples are sex and any other genetically based
covariate. Both monozygotic and dizygotic twins share date of birth and com-
mon pre-birth environment. By measuring some potentially important cova-
riates, we can examine the inuence of the covariates, and we can examine,
whether they explain the dependence, that is, whether the frailty has no effect
(or more correctly, no variation), when the covariate is included in the model.
It is not possible in practice to include all relevant covariates. For example,
we might know that some given factor is important, but if we do not know
the value of the factor for each individual, we can not include the variable in
the analysis. For example, it is known that excretion of small amounts of albu-
min in the urine is a diagnostic marker for increased mortality, not only for
diabetic patients, but also for general population. However, we are unable to
include this variable, unless we actually obtain urine and analyze samples for
each individual under study. It is furthermore possible that we are not aware
that there exist variables that we ought to include. For example, this could
be a genetic factor, as we do not know all possible genes having inuence on
survival. This consideration is true for all regression models, not only survival
models. If it is known that some factor is important, it makes sense to try to
obtain the individual values, but if it is not possible, the standard is to ignore
the presence of such variables. In general terms, we let the heterogeneity go
into the error term. This will, of course, lead to an increase in the variabil-
ity of the response compared to the case, when the variables are included. In
the survival data case, however, the increased variability implies a change in
the form of the hazard function, as will be illustrated by some more detailed
calculations.
There are some situations where T
1
and T
2
are dependent, for example,
paired organs like kidneys, lungs, eyes, ears, dental implants, etc. are depen-
dent on each other. So, we are interested mainly on paired data with common
shared frailty. Hence we have two types dependencies. The rst one is due
to the dependence of life times of the components in such a way that failure
of one component increases the load on the other component and hence in-
crease in the failure rate of survived component or the dependence may be
due to a common shock which results simultaneous failure of components.
The second type of dependence is due to frailty, the unobserved covariate, for
example, the inuence of genes on the survival. There may be several type of
dependencies in the components of a system. One has to take into account
all these type of dependencies in order to get correct model without losing
information.
Weibull extension of bivariate exponential regression model 31
The regression model is derived conditionally on the shared frailty (Y).
Conditionally on Y, the hazard function of lifetimes (T
1
, T
2
) is assumed to be
of the form
Y M(t
1
, t
2
)
where M(t
1
, t
2
) is the bivariate cumulative or integrated hazard function of the
(T
1
, T
2
) and that the value of Y is common to two components in a group. When
there is no variability in the distribution of Y, that is, when Y has a degener-
ate distribution. When the distribution is not degenerate, the dependence is
positive. The value of Y can be considered as generated from unknown values
of some explanatory variables. Conditional on Y = y, the bivariate survival
function is
S(t
1
, t
2
| y) = e
yM(t
1
,t
2
)
(1)
When T
1
and T
2
are independent, M(t
1
, t
2
) = M
1
(t
1
) + M
2
(t
2
), where M
i
(t
i
),
i = 1, 2 are the integrated hazards of T
1
and T
2
, respectively. From this, we
immediately derive the bivariate survival function by integrating Y out
S(t
1
, t
2
) = Ee
YM(t
1
,t
2
)
= L
_
M(t
1
, t
2
)
_
(2)
where L() is the Laplace transformof the distribution of Y. Thus, the bivariate
survivor function is easily expressed by means of the Laplace transform of the
frailty distribution, evaluated at the total integrated conditional hazard.
The natural parametric distribution to consider is the Weibull, because it
allows for both the proportional hazard model and the accelerated failure time
model. There is no unique natural extension of Weibull distribution in the
bivariate or multivariate situation. So, we have different versions of bivariate
or multivariate Weibull distributions, each has its own merits and demerits.
These distributions have been derived from the exponential distribution by
taking power transformation.
Lu(1989) extendedWeibull extensions of FreundandMarshallOlkinbivari-
ate exponential models. Later, Hanagal (1996) extendedtomultivariate Weibull
(MVW) distribution which is the extension of bivariate Weibull of Lu (1989)
and also the extension of the multivariate exponential of Marshall and Olkin
(1967). Hanagal (2004, 2005a, 2006a,b) proposed bivariate Weibull (BVW)
regression models for the survival data without using frailty distribution. These
BVW models are based on the extension of bivariate exponential of Marshall
and Olkin (1967) and Freund (1961). Recently Hanagal (2005b, 2006c) pro-
posed BVW regression model (Weibull extension of bivariate exponential of
MarshallOlkin) with positive stable and gamma frailties respectively and also
obtained maximum likelihood estimators (MLEs) of the parameters in the
model.
32 D. D. Hanagal
Lee (1979) extended BVE of Gumbel (1960) model to bivariate Weibull
(BVW) distribution by introducing shape parameter. Hanagal (2004, 2005a,
2006a) proposed bivariate regression analysis based on Weibull model with
identical covariates for both components. There are some situations where
we nd some non-identical covariates in addition to identical covariates for a
paired components in a system. For example, failure times of a pair of den-
tal implants in a jaw of a patient. Here the identical covariates are age and
sex of a patient and non-identical covariates may be (1) different materials
(ceramic, metal) of dental implant, (2) different shapes (screw, anchor, pillar,
hollow cylinder) of dental implant, (3) dental implants in different locations
(front, premolar, molar) of a jaw, (4) dental implants in different jaws (lower,
upper).
Dental implant study in medical science is one branch called Implantology.
Implantology studies the scientic technique of installation of dental implant
through surgical procedure after loss of natural tooth. After installation of
dental implant, one usually waits three to six months for the placement of an
articial tooth on this implant. With some procedures, placement of an arti-
cial tooth after the installation of the implant is immediate. Implants within a
patient are dependent and their life times are correlated. See (Haas et al. 1996;
Ivanoff et al. 1999).
We propose BVW regression model based on identical and non-identical
covariates on both components. In this paper, we consider bivariate Weibull
BVW of Hanagal (2004, 2005a, 2006a,b) with covariates with three different
frailty distributions namely gamma, positive stable and power variance func-
tion distributions. We introduce the BVWregression model with gamma frailty,
positive stable frailty and power variance function (PVF) frailty distributions
in Sects. 2, 3 and 4, respectively.
2 Gamma frailty
Gamma distributions have been used for many years to generate mixtures in
exponential and Poisson models. From a computational point of view, gamma
models t very well to survival models, because it is easy to derive the for-
mulas for any number of events. This is due to simplicity of the derivatives of
the Laplace transform. This is also the reason why this distribution has been
applied in most of the applications published until now. The probability density
function (PDF) of gamma distribution as
f (y) =
y
1
e
y
/ () (3)
For many calculations, it makes sense to restrict the scale parameter, and
the standard restriction is = , as this implies a mean of 1 for Y, when
, the distribution becomes degenerate. The PDF and Laplace trans-
form of gamma distribution (or unconditional bivariate survivor function) are,
respectively, as follows:
f (y) =
y
1
e
y
/ () (4)
S
(t
1
, t
2
) = L
_
M(t
1
, t
2
)
_
=

_
+M(t
1
, t
2
)
_
=
_
1 +
M(t
1
, t
2
)
. (5)
The various families of distributions have some theoretical advantages. If the
frailty distribution is a natural exponential family, selection, that is truncation
(updating when no events have happened), implies that the conditional distri-
bution of the frailty is still within the same family. The cross-ratio function is
dened as [See (Clayton 1978; Oakes 1989)]
(t
1
, t
2
) =

_
t
1
| T
2
= t
2
_
_
t
1
| T
2
> t
2
_
=
_
2
S
(t
1
,t
2
)
t
1
t
2
_
_
S
(t
1
, t
2
)
_
_
S
(t
1
,t
2
)
t
1
_ _
S
(t
1
,t
2
)
t
2
_
=
_
1 +
1
_
+

12
(t
1
, t
2
)
_
+M(t
1
, t
2
)
_
1
(t
1
, t
2
)
2
(t
1
, t
2
)
(6)
where
i
(t
1
, t
2
) =
M(t
1
, t
2
)
t
i
, i = 1, 2
12
(t
1
, t
2
) =

2
M(t
1
, t
2
)
t
1
t
2
The above expression is true for any underlying bivariate distribution of (T
1
, T
2
)
and Y has gamma with PDF given in Eq. (4). When T
1
and T
2
are independent,
then M(t
1
, t
2
) = M
1
(t
1
) +M
2
(t
2
), (t
1
, t
2
) = 0 and
(t
1
, t
2
) = 1 +
1
.
If (t
1
, t
2
) = 0, then T
1
and T
2
may not necessarily be independent. An exam-
ple for this is Weibull extension of bivariate exponential of Marshall and Olkin
(1967) discussed in Sect. 2.2 in which (t
1
, t
2
) = 0.
The expression (6) can be interpreted as the relative risk for an individ-
ual if the other one has experienced the event rather than being event free at
a given time [See (Liang 1991)]. Therefore, it is an association function such
that (t
1
, t
2
) > 1 represents positive association, (t
1
, t
2
) < 1 indicates negative
association and (t
1
, t
2
) = 1 implies no association.
34 D. D. Hanagal
2.1 Weibull extension of BVE of Gumbel
The Weibull extension of BVE-type I of Gumbel (1960) with survival function
is given by
S(t
1
, t
2
) = P
_
T
1
> t
1
, T
2
> t
2
_
= e
1
t
c
1
1

2
t
c
2
2

1
2
t
c
1
1
t
c
2
2
, t
1
, t
2
> 0 (7)
where
1
,
2
, c
1
, c
2
> 0, 0 1
Here the marginal distribution of T
1
and T
2
are distributed as Weibull with
scale parameters
1
and
2
and shape parameters c
1
and c
2
, respectively. The
parameter corresponds to the dependence parameter in BVW model.
Now the conditional survival function of BVW given the frailty (Y = y) is
given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
1
+
2
t
c
2
2
+
1
2
t
c
1
1
t
c
2
2
_
(8)
where Y follows gamma distribution given in (4).
Integrating over Y, we get unconditional survival function and is given by
S
(t
1
, t
2
) =
_
1 +

1
t
c
1
1
+
2
t
c
2
2
+
1
2
t
c
1
1
t
c
2
2
(9)
This is the bivariate Burr distribution, generalizing the Pareto power distri-
bution. The marginal distributions are Burr distribution with survival functions
given by
S
(t
i
) =
_
1 +

i
t
c
i
i
, i = 1, 2 (10)
The unconditional bivariate survival function in (9) has two types of depen-
dencies, one parameter () is due to dependence parameter and the other
parameter () is due to frailty. These two parameters are identiable.
As we have seen in the univariate Weibull regression, the scale parameter
of the univariate Weibull distribution can be expressed in terms of regression
coefcients. If is the scale parameter of the Weibull distribution, then = e
z
or = e
z
where is the vector of regression parameters and z is the vector
of regressors or covariates. In the similar manner, the scale parameters
1
and
2
can be expressed in terms of regression parameters in the following way.
1
= e
(
0
z
0
+
1
z
1
)
2
= e
(
0
z
0
+
2
z
2
)
(11)
where
0
=
_
01
, . . . ,
0p
_
, <
0
<
1
=
_
11
, . . . ,
1q
_
, <
1
<
2
=
_
21
, . . . ,
2q
_
, <
2
<
z
0
=
_
z
01
, . . . , z
0p
_
z
1
=
_
z
11
, . . . , z
1q
_
z
2
=
_
z
21
, . . . , z
2q
_
The exponent terms in the above expressions, we can take either positive or
negative but in either case
1
,
2
> 0.
0
z
0
corresponds to the term containing
identical covariates for both components.
1
z
1
corresponds to the termcontain-
ing covariates for rst component and
2
z
2
corresponds to the term containing
covariates for second component.
Now the survival function of BVW in terms of covariates is given by
S
(t
1
, t
2
) =
_
1 +
t
c
1
1
e
0
z
0
+
1
z
1
_
+t
c
2
2
e
0
z
0
+
2
z
2
_
+
t
c
1
1
t
c
2
2
e
_
2
0
z
0
+
1
z
1
+
2
z
2
_
. (12)
The another Weibull extension of BVE-type II of Gumbel (1960) with sur-
vival function is given by
S(t
1
, t
2
) = S
1
(t
1
)S
2
(t
2
)
_
1 +
_
1 S
1
(t
1
)
__
1 S
2
(t
2
)
_ _
, t
1
, t
2
> 0 (13)
where 1 +1,
The above survival function is derived from the given two independent mar-
ginal distributions. This relation is true for any bivariate distribution. Suppose
T
i
, i = 1, 2 follows Weibull (
i
, c
i
) distribution then BVW of (T
1
, T
2
) with
survival function is given by
S(t
1
, t
2
) = (1 +)e
1
t
c
1
1

2
t
c
2
2
e
2
1
t
c
1
1

2
t
c
2
2
e
1
t
c
1
1
2
2
t
c
2
2
+e
2
1
t
c
1
1
2
2
t
c
2
2
, (14)
where t
1
, t
2
> 0,
1
,
2
, c
1
, c
2
> 0, 1 +1.
The above survival function is the weighted combination of four bivariate
Weibull distributions with weights (1 + ), , , . Here the marginal distri-
bution of T
1
and T
2
are distributed as Weibull with scale parameters
1
and
2
1
and c
2
, respectively. The parameter corresponds
36 D. D. Hanagal
to the dependence parameter in BVW model. When = 0, T
1
and T
2
are
independent.
Nowthe conditional survival distribution of the above BVWgiven the frailty
(Y = y) is given by the weighted combination of the four conditional survival
functions given the frailty, that is,
S(t
1
, t
2
| y) = (1 +)e
y
_
1
t
c
1
1
+
2
t
c
2
2
_
e
y
_
2
1
t
c
1
1
+
2
t
c
2
2
_
e
y
_
1
t
c
1
1
+2
2
t
c
2
2
_
+e
y
_
2
1
t
c
1
1
+2
2
t
c
2
2
_
, (15)
S
(t
1
, t
2
) = (1 +)
_
1 +

1
t
c
1
1
+
2
t
c
2
2
_
1 +
2
1
t
c
1
1
+
2
t
c
2
2
_
1 +

1
t
c
1
1
+2
2
t
c
2
2
+
_
1 +
2
1
t
c
1
1
+2
2
t
c
2
2
(16)
The above bivariate survival function is the weighted combination of the
four bivariate Burr distributions with weights (1 + ), , , . The marginal
distributions are Burr distribution with survival functions given by
S
(t
i
) =
_
1 +

i
t
c
i
i
, i = 1, 2 (17)
The parameters
1
and
2
here also expressed in terms of the regression param-
eters and covariates as we did in the Eq. (11).
The another Weibull extension of BVE-type III of Gumbel (1960) with sur-
vival function is given by
S(t
1
, t
2
) = e
1
t
c
1
/
1
+
2
t
c
2
/
2
_
, t
1
, t
2
> 0 (18)
where
1
,
2
, c
1
, c
2
> 0, 0 1
1
and T
2
are distributed as Weibull with
scale parameters
1
and
2
1
and c
2
, respectively. The
parameter corresponds to the dependence parameter in BVW model.
Nowthe conditional survival distribution of the above BVWgiven the frailty
(Y = y) is given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
/
1
+
2
t
c
2
/
2
_
, t
1
, t
2
> 0 (19)
S
(t
1
, t
2
) =
_
1 +
_
1
t
c
1
/
1
+
2
t
c
2
/
2
_
(20)
The marginal distributions are Burr distribution with survival functions given
by
S
(t
i
) =
_
1 +

i
t
c
i
i
, i = 1, 2 (21)
The parameters
1
and
2
eters as we did in the Eq. (11). The survival function in terms of covariates is
given by
S
(t
1
, t
2
) =
1 +
_
t
c
1
/
1
e
0
z
0
+
1
z
1
_
+t
c
2
/
2
e
0
z
0
+
2
z
2
_
_
(22)
2.2 Weibull extension of BVE of MarshallOlkin
Hanagal (1996) proposed k +2 parameter family of k-variate Weibull distribu-
tion. Specifying k = 2, we get BVW with survival function given by
S(t
1
, t
2
) = e
1
t
c
1
2
t
c
2
3
t
c
(2)
(23)
where t
(2)
= max(t
1
, t
2
),
1
,
2
,
3
, c > 0.
The BVW model has important properties. Both marginal distribution of
BVW are univariate Weibull. More specically speaking, we can state that
(T
1
, T
2
) BVW(
1
,
2
,
3
, c)T
1
Weibull
_
(
1
+
3
), c
_
T
2
Weibull
_
(
2
+
3
), c
_
T
(1)
min(T
1
, T
2
)Weibull
_
(
1
+
2
+
3
), c
_
where (
1
, +
3
), (
2
+
3
) and (
1
+
2
+
3
) are the scale parameters of T
1
,
T
2
and T
(1)
, respectively and c is the common shape parameter. This BVW
model is not absolutely continuous with respect to Lebesque measure in R
2
. It
has singularity on the diagonal T
1
= T
2
. Here the parameter
3
corresponds
to the dependence between the two variables (T
1
, T
2
) and
3
= 0 implies
T
1
and T
2
and independent. The probability of simultaneous failures, that is,
P[T
1
= T
2
] is
3
/(
1
+
2
+
3
) which is the correlation between T
1
and T
2
. The
38 D. D. Hanagal
probability in the remaining two regions are P[T
1
< T
2
] =
1
/(
1
+
2
+
3
)
and P[T
1
> T
2
] =
2
/(
1
+
2
+
3
).
given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
(24)
S
(t
1
, t
2
) =
_
1 +
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
(25)
This is the bivariate Burr distribution, generalizing the Pareto power distri-
bution. The marginal distributions and T
(1)
= min(T
1
, T
2
) are Burr distribution
with survival functions given by
S
(t
i
) =
_
1 +
(
i
+
3
)t
c
i
, i = 1, 2
S
(t
(1)
) =
_
1 +
(
1
+
2
+
3
)t
c
(1)
.
(26)
dencies, one is due to simultaneous failures and the other is due to frailty.
Now we develop a regression model for the two component system. The scale
parameters
1
and
2
can be expressed in terms of regression parameters and
covariates as in Eq. (11) and the parameter
3
expressed as follows.
3
= e
0
z
0
+
1
z
1
+
2
z
2
_
(27)
Note: From here onwards,
1
,
2
and
3
are expressed in terms of covariates
from Eqs. (11) and (27) when one wants to express survival function in the
form of regression model.
2.3 Weibull extension of BVE of BlockBasu
Block and Basu (1974) proposed an absolutely continuous BVE as an alterna-
tive model to Marshall and Olkin (1967) which has a singular component on
the diagonal arising out of possible simultaneous failures. MarshallOlkin is not
absolutely continuous with respect to Lebesque measure in R
2
, there are some
situations when this model is not appropriate. The survival function of BVE of
Block and Basu (1974) is given by
S(x
1
, x
2
) =

1
+
2
e
1
x
1
2
x
2
3
x
(2)
1
+
2
e
x
(2)
(28)
where x
(2)
= max(x
1
, x
2
),
1
,
2
,
3
> 0, =
1
+
2
+
3
.
Taking the transformation T
1
= X
c
1
and T
2
= X
c
2
, c > 0, we get BVW and its
survival function is given by
S(t
1
, t
2
) = P
_
T
1
> t
1
, T
2
> t
2
_
=

1
+
2
e
1
t
c
1
2
t
c
2
3
t
c
(2)
1
+
2
e
t
c
(2)
(29)
where t
(2)
= max(t
1
, t
2
),
1
and T
2
are now not Weibull but a
weighted combination of two Weibull with scales (
i
+
3
) and , i=1,2 with
weights 1 +
3
/(
1
+
2
) and
3
/(
1
+
2
).
The above model turns out to be the absolutely continuous part of BVW of
Hanagal (1996, 2004) and also a special case of BVW of Hanagal (2006a). The
independence of two components corresponds to
3
= 0.
given by the weighted combination of the conditional survival functions given
the frailty, that is,
S(t
1
, t
2
| y) =

1
+
2
e
y
_
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
1
+
2
e
y
_
t
c
(2)
_
(30)
S
(t
1
, t
2
) =

1
+
2
_
1 +

1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
1
+
2
_
1 +
t
c
(2)
(31)
The above bivariate survival function is weighted combination of bivari-
ate Burr distribution and univariate Burr distribution with weights [1 +
3
/
(
1
+
2
)] and
3
/(
1
+
2
), respectively. The marginal distributions are
weighted combinations of two Burr distributions with weights [1+
3
/(
1
+
2
)]
and
3
/(
1
+
2
). The distribution of T
(1)
= min(T
1
, T
2
) is Burr distribution
with survival function given by
S
(t
(1)
) =
_
1 +
(
1
+
2
+
3
)t
c
(1)
(32)
dencies, one is due to the dependence parameter
3
and the other is due to
frailty.
40 D. D. Hanagal
2.4 Weibull extension of BVE of Freund and ProschanSullo
Freund (1961) proposed BVE as a model for failure time distribution of a sys-
temwith life-times (X
1
, X
2
) operating in the following manner. Initially X
1
and
X
2
are independent exponential with failure rates
1
and
2
, respectively,
1
,
2
, > 0. The interdependence of the components is such that failure of a com-
ponent changes the failure rate of other component from
1
to
11
(
2
to
22
).
The BVE of Freund (1961) with its joint PDF is given by
f (x
1
, x
2
) =
_
22
e
22
x
2
(
1
+
2
22
)x
1
, 0 < x
1
< x
2
<
11
e
11
x
1
(
1
+
2
11
)x
2
, 0 < x
2
< x
1
<
(33)
where
1
,
2
,
11
,
22
> 0.
Proschan and Sullo (1974) proposed BVE which is the combination of both
MarshallOlkin and Freund models and the two component system operate in
the following manner. Initially X
1
and X
2
follow BVE of Marshall and Olkin
(1967). When a component fails the failure of a component changes the failure
rate of other component from
1
+
3
to
11
+
3
(
2
+
3
to
22
+
3
). The
BVE of Proschan and Sullo (1974) with its PDF is given by
f (x
1
, x
2
) =
1
(
22
+
3
)e
(
22
+
3
)x
2
(
1
+
2
22
)x
1
, 0 < x
1
< x
2
<
2
(
11
+
3
)e
(
11
+
3
)x
1
(
1
+
2
11
)x
2
, 0 < x
2
< x
1
<
3
e
(
1
+
2
+
3
)x
, 0 < x
1
= x
2
= x <
(34)
where
1
,
2
,
3
,
11
,
22
> 0.
Taking transformation T
1
= X
c
1
and T
2
= X
c
2
, c > 0 we get bivariate Weibull
model (BVW) which was introduced by Hanagal (2005a,b) with PDF given by
f (t
1
, t
2
) =
1
(
22
+
3
)c
2
(t
1
t
2
)
c1
e
(
22
+
3
)t
c
2
(
1
+
2
22
)t
c
1
, 0 < t
1
< t
2
<
2
(
11
+
3
)c
2
(t
1
t
2
)
c1
e
(
11
+
3
)t
c
1
(
1
+
2
11
)t
c
2
, 0 < t
2
< t
1
<
3
e
(
1
+
2
+
3
)t
c
, 0 < t
1
= t
2
= t <
(35)
Re-parameterize
11
=
1
1
,
22
=
2
2
and rewrite the above PDF as
f (t
1
, t
2
) =
1
(
2
2
+
3
)c
2
(t
1
t
2
)
c1
e
(
2
2
+
3
)t
c
2
(
1
+
2
2
)t
c
1
,
0 < t
1
< t
2
<
2
(
1
1
+
3
)c
2
(t
1
t
2
)
c1
e
(
1
1
+
3
)t
c
1
(
1
+
2
1
)t
c
2
,
0 < t
2
< t
1
<
3
e
(
1
+
2
+
3
)t
c
, 0 < t
1
= t
2
= t <
(36)
As we know in BVE of Proschan and Sullo (1974), the marginals are weighted
combinations of two exponential distributions. Here in the BVWalso, the marg-
inals are weighted combinations of two Weibull distributions with same weights.
The min(T
1
, T
2
) is Weibull with scale parameter (
1
+
2
+
3
) and shape param-
eter c. When
3
= 0, the BVW in Eq. 36 reduces to BVW of Hanagal (2004)
and when
1
=
2
= 1, it reduces to BVW of Hanagal (2006a). When
i
= 1,
i = 1, 2 and
3
= 0 then T
1
and T
2
are independent. The probabilities in the
three regions are given by
P
_
T
1
< T
2
_
=
1
/(
1
+
2
+
3
),
P
_
T
1
> T
2
_
=
2
/(
1
+
2
+
3
) and P
_
T
1
= T
2
_
=
3
/(
1
+
2
+
3
).
The survival function of this BVW is given by
S(t
1
, t
2
) =
2
(1
2
)e
(
1
+
2
+
3
)t
c
2
1
+
2
(1
2
)
+
1
e
(
1
+
2
(1
2
))t
c
1
(
2
2
+
3
)t
c
2
1
+
2
(1
2
)
, 0 < t
1
t
2
1
(1
1
)e
(
1
+
2
+
3
)
t
c
1
2
+
1
(1
1
)
+
2
e
(
2
+
1
(1
1
))t
c
2
(
1
1
+
3
)t
c
1
2
+
1
(1
1
)
, 0 < t
2
t
1
(37)
given by the weighted combination of the conditional survival functions given
the frailty, that is,
S(t
1
, t
2
| y) =
2
(1
2
)e
y(
1
+
2
+
3
)t
c
2
1
+
2
(1
2
)
+
1
e
y
_
(
1
+
2
(1
2
))t
c
1
+(
2
2
+
3
)t
c
2
_
1
+
2
(1
2
)
, 0 < t
1
t
2
1
(1
1
)e
y(
1
+
2
+
3
)t
c
1
2
+
1
(1
1
)
+
2
e
y
_
(
2
+
1
(1
1
))t
c
2
+(
1
1
+
3
)t
c
1
_
2
+
1
(1
1
)
, 0 < t
2
t
1
(38)
S
(t
1
, t
2
) =
2
(1
2
)
_
1+
(
1
+
2
+
3
)t
c
2
1
+
2
(1
2
)
+
1
_
1+
(
1
+
2
(1
2
))t
c
1
+(
2
2
+
3
)t
c
2
1
+
2
(1
2
)
, 0 < t
1
t
2
1
(1
1
)
_
1+
(
1
+
2
+
3
)t
c
1
2
+
1
(1
1
)
+
2
_
1+
(
2
+
1
(1
1
))t
c
2
+(
1
1
+
3
)t
c
1
2
+
1
(1
1
)
, 0 < t
2
t
1
(39)
42 D. D. Hanagal
The above bivariate survival function is the weighted combination of the
bivariate Burr distribution and univariate Burr distribution with same weights
as in BVE of Proschan and Sullo (1974). The unconditional bivariate survival
function in (39) has three types of dependencies, one is due to simultaneous
failures and second is due to load on one component due to the failure of
another component and the third is due to frailty.
Substituting
3
= 0 in all the above expressions we get the corresponding
expressions for the Weibull extension of BVE of Freund (1961).
3 Positive stable frailty
Inpractice, the gamma frailty specicationmay not t well (Shih(1998); Glidden
(1999); Fan et al. (2000). The positive stable model (Hougaard 2000) is a useful
alternative, in part because it has the attractive feature that predictive haz-
ard ratio decrease to one over time (Oakes 1989). The property is observed in
familial associations of the ages of onset of diseases with etiologic heterogeneity,
where genetic cases occur early and long-term survivors are weakly correlated.
The gamma model has predictive hazard ratios which are time invariant and
may not be suitable for these patterns of failures (Fine et al. 2003). The PDF of
positive stable distribution with two parameters and is given by
f (y) =
1
y
k=1
(k +1)
k!
(y
/)
k
sin(k), y > 0, 0 < < 1 (40)
with Laplace transform
E{e
sY
} = e
s
/
. (41)
[See Hougaard (2000, p. 503)]
The unconditional bivariate survival function with positive stable frailty is
given by
S
,
(t
1
, t
2
) = e
[M(t
1
,t
2
)]
/
(42)
The PDF of positive stable distribution with ( = ) is given by
f (y) =
1
k=1
(k +1)
k!
_
1
y
_
k+1
sin(k), y > 0, 0 < < 1 (43)
with Laplace transform
E{e
sY
} = e
s
. (44)
The unconditional bivariate survival function with positive stable frailty is
given by
S
(t
1
, t
2
) = e
[M(t
1
,t
2
)]
. (45)
When = 1, the frailty distribution is degenerate at Y = 1.
The stable Weibull model The Weibull distributionis particularly well suited
to the positive stable frailty model. The bivariate Weibull model is obtained by
assuming M
i
(t) =
i
t
c
i
, i = 1, 2. This means that conditionally on Y = y, the
distribution of T
i
is Weibull (
i
, y, c
i
), i = 1, 2. When T
1
and T
2
are independent,
the bivariate survival function is
S
(t
1
, t
2
) = e
1
t
c
1
1
+
2
t
c
2
2
_
. (46)
The advantage of this model is that the marginal distributions are also of
Weibull form. Also, the time to the rst event, T
(1)
= min(T
1
, T
2
) is of Weibull
form when c
1
= c
2
.
The positive stable model has the advantage that it ts proportional hazards
which means that if the conditional model has proportional hazards, so does
the marginal distribution. This is an advantage, when considering the model as
a random effects model.
3.1 Weibull extension of BVE of Gumbel
Nowthe conditional survival distribution of Weibull extension of the BVEtype
I of Gumbel (1960) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
1
+
2
t
c
2
2
+
1
2
t
c
1
1
t
c
2
2
_
(47)
where
1
,
2
, c
1
, c
2
> 0, 0 1
where Y follows positive stable distribution given in (43).
S
(t
1
, t
2
) = e
1
t
c
1
1
+
2
t
c
2
2
+
1
2
t
c
1
1
t
c
2
2
_
(48)
The above bivariate survival function has two types of dependencies, one
is due to dependence parameter and other is due to frailty. The marginal
distributions are also Weibull and are given by
S
(t
i
) = e
i
t
c
i
i
_
, i = 1, 2 (49)
The parameters
1
and
2
eters as we did in the Eq. (11).
44 D. D. Hanagal
Now the conditional survival distribution of Weibull extension of the BVE
type II of Gumbel (1960) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) = (1 +)e
y
_
1
t
c
1
1
+
2
t
c
2
2
_
e
y
_
2
1
t
c
1
1
+
2
t
c
2
2
_
e
y
_
1
t
c
1
1
+2
2
t
c
2
2
_
+e
y
_
2
1
t
c
1
1
+2
2
t
c
2
2
_
, (50)
S
(t
1
, t
2
) = (1 +)e
1
t
c
1
1
+
2
t
c
2
2
_
_
2
1
t
c
1
1
+
2
t
c
2
2
_
1
t
c
1
1
+2
2
t
c
2
2
_
+e
_
2
1
t
c
1
1
+2
2
t
c
2
2
_
, (51)
The marginal distributions are also of Weibull form with survival functions
given by
S
(t
i
) = e
i
t
c
i
i
_
, i = 1, 2 (52)
Now the conditional survival distribution of Weibull extension of BVE type III
of Gumbel (1960) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
/
1
+
2
t
c
2
/
2
_
, t
1
, t
2
> 0 (53)
where
1
,
2
, c
1
, c
2
> 0, 0 1
S
(t
1
, t
2
) = e
(
1
t
c
1
/
1
+
2
t
c
2
/
2
)
, (54)
The marginal distributions are also of Weibull form with survival functions
given by
S
(t
i
) = e
(
i
t
c
i
i
)
, i = 1, 2 (55)
3.2 Weibull extension of BVE of MarshallOlkin
Suppose Y has a positive stable distribution which we assume as a frailty distri-
bution. Now the conditional survival function of Weibull extension of BVE of
Marshall and Olkin (1967) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) = e
y
_
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
(56)
S
(t
1
, t
2
) = e
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
(57)
The above bivariate survival function has two types of dependencies, one is
due to simultaneous failures and other is due to frailty. When = 1, the frailty
distribution is degenerate at Y = 1.
The main advantage of this model is that the marginal distributions are also
of Weibull and T
(1)
= min(T
1
, T
2
) is also of Weibull form and corresponding
survival functions are given by
S
(t
i
) = e
(
i
+
3
)
t
c
i
, i = 1, 2
S
(t
(1)
) = e
(
1
+
2
+
3
)
t
c
(1)
(58)
3.3 Weibull extension of BVE of BlockBasu
Nowthe conditional survival function of Weibull extension of BVEof Marshall
and Olkin (1967) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) =

1
+
2
e
y(
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
)
1
+
2
e
yt
c
(2)
(59)
S
(t
1
, t
2
) =

1
+
2
e
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
1
+
2
e
(t
c
(2)
)
(60)
The above bivariate survival function has two types of dependencies, one is
due to dependence parameter
3
and other is due to frailty. When = 1, the
frailty distribution is degenerate at Y = 1.
The marginal distributions are of weighted combinations of two Weibull and
T
(1)
= min(T
1
, T
2
) is of Weibull form and corresponding survival function of
T
(1)
is given by
S
(t
(1)
) = e
(
1
+
2
+
3
)
t
c
(1)
(61)
46 D. D. Hanagal
3.4 Weibull extension of BVE of Freund and ProschanSullo
Nowthe conditional survival function of Weibull extension of BVEof Proschan
and Sullo (1974) given the frailty (Y = y) is given by
S(t
1
, t
2
| y) =
2
(1
2
)e
y
_
1
+
2
+
3
)t
c
2
1
+
2
(1
2
_
+
1
e
y[(
1
+
2
(1
2
))t
c
1
+(
2
2
+
3
)t
c
2
]
1
+
2
(1
2
)
, 0 < t
1
t
2
1
(1
1
)e
y(
1
+
2
+
3
)t
c
1
2
+
1
(1
1
)
+
2
e
y[(
2
+
1
(1
1
))t
c
2
+(
1
1
+
3
)t
c
1
]
2
+
1
(1
1
)
, 0 < t
2
t
1
(62)
S
(t
1
, t
2
) =
2
(1
2
)e
((
1
+
2
+
3
)t
c
2
)
1
+
2
(1
2
)
+
1
e
[(
1
+
2
(1
2
))t
c
1
+(
2
2
+
3
)t
c
2
]
1
+
2
(1
2
)
, 0 < t
1
t
2
1
(1
1
)e
((
1
+
2
+
3
)t
c
1
)
2
+
1
(1
1
)
+
2
e
[(
2
+
1
(1
1
))t
c
2
+(
1
1
+
3
)t
c
1
]
2
+
1
(1
1
)
, 0 < t
2
t
1
(63)
Substituting
3
4 Power variance function frailty
This distribution is a three-parameter family uniting gamma and positive stable
distributions. The distribution is denoted PVF(,
, ). For = 0, the gamma

distributions are obtained, with same parametrization. Some formulas are valid,
but many are others are different in this case. For = 0, the positive stable dis-
tributions are obtained. For = 1/2, the inverse Gaussian distributions are
obtained. For = 1, the non-central gamma distribution of shape parameter
zero is obtained. For = 1, a degenerate distribution is obtained.
The parameter set is ( 1,
> 0), with ( 0 for > 0), and ( > 0 for

0). The distribution is concentrated on the positive numbers for 0, and
is positive or zero for < 0. In the case > 0, the PDF of PVF is given by [See
Hougaard (2000, p. 504)]
f (y) = e
y+
/
1
k=1
(k +1)
k!
(
1
y
)
k+1
sin(k), y > 0 (64)
In the case < 0, the term in the density is not necessarily dened,
and therefore we can use the alternative expression for p.d.f. of PVF as [See
Hougaard (2000, p. 504)]
f (y) = e
y+
/
1
y
k=1
(y
/)
k
k!(k)
, y > 0 (65)
This expression is valid for all values, except 0 and 1, with the convention
that when the -function in the denominator is undened (which happens when
k is a positive integer), the whole term in the sum is zero. For < 0, there
is probability exp(
/) of the random variable being zero. For 0, the

distribution is unimodal.
If Y
1
and Y
2
are independent, and Y
i
follows PVF(,
i
, ), i = 1, 2 the distri-
bution of Y
1
+Y
2
is PVF(,
1
+
2
, ). So, PVF distribution is innitely divisible.
When > 0, all (positive) moments exist, and the mean is
1
. The variance
is
(1 )
2
.
The Laplace transform of PVF distribution is
L(s) = e
{(+s)
}/
. (66)
The unconditional bivariate survival function with PVF frailty is given by
S
,
,
(t
1
, t
2
) = e
{+M(t
1
,t
2
)}
/+
/
. (67)
4.1 Weibull extension of BVE Models
The unconditional survival function of BVW with PVF frailty distribution cor-
responding to BVE type I of Gumbel (1960) is given by
S
,
,
(t
1
, t
2
) = e
_
+
1
t
c
1
1
+
2
t
c
2
2
+
1
2
t
c
1
1
t
c
2
2
_
/+
/
(68)
The unconditional survival function of BVW with PVF frailty distribution
corresponding to BVE type II of Gumbel (1960) is given by
S
,
,
(t
1
, t
2
) = (1 +)e
_
+
1
t
c
1
1
+
2
t
c
2
2
_
/+
/
e
_
+2
1
t
c
1
1
+
2
t
c
2
2
_
/+
/
e
_
+
1
t
c
1
1
+2
2
t
c
2
2
_
/+
/
+e
_
+2
1
t
c
1
1
+2
2
t
c
2
2
_
/+
/
, (69)
48 D. D. Hanagal
corresponding to BVE type III of Gumbel (1960) is given by
S
,
,
(t
1
, t
2
) = e
{+(
1
t
c
1
1
+
2
t
c
2
2
)
/+
/
, (70)
corresponding to BVE of Marshall and Olkin (1967) is given by
S
,
,
(t
1
, t
2
) = e
_
+
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
/+
/
(71)
corresponding to BVE of Block and Basu (1974) is given by
S
,
,
(t
1
, t
2
) =

1
+
2
e
_
+
1
t
c
1
+
2
t
c
2
+
3
t
c
(2)
_
/+
1
+
2
e
_
+t
c
(2)
_
/+
/
(72)
corresponding to BVE of Proschan and Sullo (1974) is given by
S
(t
1
, t
2
) =
2
(1
2
)e
(+t
c
2
)
/+
1
+
2
(1
2
)
+
1
e
_
+
_
1
+
2
(1
2
)
_
t
c
1
+
_
2
+
3
_
t
c
2
_
_
+
1
+
2
(1
2
)
, t
1
t
2
1
(1
1
)e
(
+t
c
1
)
/+
2
+
1
(1
1
)
+
2
e
_
+
_
2
+
1
(1
1
))t
c
2
+(
1
1
+
3
_
t
c
1
_
/+
2
+
1
(1
1
)
, t
2
t
1
(73)
where =
1
+
2
+
3
.
Substituting
3
Acknowledgements I thank both the referees for the suggestions and comments.
References
Block HW, Basu AP (1974) A continuous bivariate exponential extension. J Am Stat Assoc 69:
10311037
Clayton DG (1978) A model for association in bivariate life tables and its applications to epidemi-
ological studies of familial tendency in chronic disease incidence. Biometrika 65:141151
Fan JJ, Hsu L, Prentice RL (2000) Dependence estimation over a nite bivariate failure region.
Lifetime Data Anal 6:343355
Fine JP, Glidden DV, Lee KE (2003) A simple estimator for a shared frailty regression model.
J R Stat Soc B 65:317329
Freund JE (1961) A bivariate extension of exponential distribution. J Am Stat Assoc 56:97177
Glidden DV (1999) Checking the adequacy of the gamma frailty model for multivariate failure
times. Biometrika 86:381394
Gumbel EJ (1960) Bivariate exponential distribution. J Am Stat Assoc 55:698707
Haas R, Mensdorff-Pouilly N, Mailath G, Watzek G(1996) Survival of 1,920 IMZimplants followed
for up to 100 months. Int J Oral Maxillofac Implants 11:581588
Hanagal DD (1996) A multivariate Weibull distribution. Econ Qual Control 11:193200
Hanagal DD(2004) Parametric bivariate regression analysis based on censored samples: AWeibull
model. Econ Qual Control 19:8390
Hanagal DD (2005a) A bivariate Weibull regression model. Econ Qual Control 20:143150
Hanagal DD (2005b) A positive stable frailty regression model in bivariate survival data. J Ind Soc
Probab Stat 9:3544
Hanagal DD (2006a) Bivariate Weibull regression model based on censored samples. Stat Pap
47:137148
Hanagal DD (2006b) Weibull extensions of bivariate exponential regression model for a survival
data. Stat Methods 8:110121
Hanagal DD (2006c) Weibull extension of bivariate exponential regression model with gamma
frailty for survival data. Econ Qual Control 21:165174
Hougaard P (2000) Analysis of multivariate survival data. Springer, New York
Ivanoff CT, Grondahl K, Sennerby L, Bergstorm C, Lekholm W (1999) Inuence of variations in
implant diameters: a 3 to 5 year retrospective clinical report. Int J Oral Maxillofac Implants
14:173180
Lee L (1979) Multivariate distributions having Weibull properties. J Multivariate Anal 9:267277
Liang K-Y (1991) Estimating effects of probands characteristic in familial risk: 1. Adjustments for
censoring and correlated ages at onset. Genet Epidemiol 8:329338
Lu JC (1989) Weibull extensions of the Freund and MarshallOlkin bivariate exponential models.
IEEE Trans Reliabil 38:615619
Marshall AW, Olkin I (1967) A multivariate exponential distribution. J Am Stat Assoc 62:3044
Oakes D (1982) A model for association in bivariate survival data. J R Stat Soc B 44:414422
Oakes D (1986) Semiparametric inference in bivariate survival data. Biometrika 73:353361
Oakes D (1989) Bivariate survival models induced by frailties. J Am Stat Assoc 84:487493
Parner E (1998) Asymptotic theory for correlated gamma-frailty model. Ann Stat 26:183214
Proschan F, Sullo P (1974) Estimating the parameters of bivariate exponential distributions in
several sampling situations. In: Proschan F, Sering RJ (eds) Reliability and biometry.SIAM,
Philadelphia, pp 423440
Shih JH (1998) A goodness-of-t test for association in a bivariate survival model. Biometrika
85:189200

Weibull Extension of Bivariate Exponential Regression

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

Weibull Extension of Bivariate Exponential Regression

Загружено:

Авторское право:

Доступные форматы

Stat Papers (2009) 50:2949

, ). For = 0, the gamma

> 0), with ( 0 for > 0), and ( > 0 for

/) of the random variable being zero. For 0, the

Вам также может понравиться