Reliability Engineering and System Safety 73 (2001) 73±81

www.elsevier.com/locate/ress

Models involving two inverse Weibull distributions

R. Jiang a,*, D.N.P. Murthy b, P. Ji a
a
Department of Manufacturing Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, People's Republic of China
b
Department of Mechanical Engineering, The University of Queensland, Q 4072, Australia
Received 22 December 2000; accepted 17 March 2001

Abstract
In this paper, we look at three models (mixture, competing risk and multiplicative) involving two inverse Weibull distributions. We study
the shapes of the density and failure-rate functions and discuss graphical methods to determine if a given data set can be modelled by one of
these models. q 2001 Elsevier Science Ltd. All rights reserved.
Keywords: Inverse Weibull; Mixture model; Competing risk model; Multiplicative model; WPP and IWPP plots

1. Introduction meter) and with the inverse Weibull transformations. Refs.

[4,8,9] also deal with the inverse Weibull distribution.
The basic Weibull model is the standard 2-parameter In this paper, we study three different models involving
Weibull distribution function given by two inverse Weibull distributions and these are as follows:
b (i) mixture model, (ii) competing risk model and (iii) multi-
F t† ˆ 1 2 e2 t=a† ; a; b . 0; 1† plicative model. Our focus is on issues relating to the suit-
ability of these models to model a given data set. If the data
where a is the scale parameter and b is the shape parameter.
set is large, one can plot the empirical plots of the density and
A variety of models have been derived from this basic
failure rate functions. By comparing these with the shapes of
model. One such is the following. Let X denote the random
the density and failure rate functions for the model under
variable from the Weibull model. De®ne Y as follows:
consideration one can decide on whether the model is appro-
a2 priate or not. An alternate approach is to transform the data
Yˆ : 2† before plotting. One such is the Weibull Probability Paper
X
(WPP) plot where if the plot is roughly a straight-line then
Then the distribution function for Y is given by the basic Weibull model can be used to model the data.
b In this paper, we study the shapes of the density and
F t† ˆ e2 a=t† : 3† failure rate functions for these three models and look at
alternate plotting techniques to determine if a given data
This is the basic `inverse Weibull' model for obvious
set can be adequately modelled by one of these models.
reason and the associated distribution is the inverse Weibull
The outline of the paper is as follows. In Section 2, we
distribution.
look at the basic inverse model and review the relevant
The inverse Weibull distribution has received some atten-
results available in the literature. Following this we look
tion. Keller and Kamath [10] study the shapes of the density
at the three models in the next three sections (Mixture
and failure rate functions for the basic inverse model. Calab-
model Ð Section 3, Competing risk model Ð Section 4
ria and Pulcini [1,2] deal with parameter estimation of the
and Multiplicative model Ð Section 5). Section 6 presents a
model. Drapella [3] calls the basic inverse Weibull model as
numerical example. We conclude with some comments and
the `complementary Weibull' model and discusses a graphi-
topics for future research in Section 7.
cal plotting technique to determine the suitability of the
model to model a given data set. Ref. [7] deals with the
inverse Weibull mixture model (with a common shape para-
2. The basic inverse Weibull model
* Corresponding author. Fax: 1852-2362-5267.
E-mail address: [email protected] (R. Jiang). In this section, we ®rst look at some properties of the
0951-8320/01/$ - see front matter q 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0951-832 0(01)00030-8
74 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81

Table 1 On differentiating Eq. (7), we have

Fractiles t0.001 and t0.999
 
b t0.001/a t0.999/a 0 2b21 hb b
r t† ˆ r t†t 2 b 1 1†t : 9†
1 2 e2z
Weibull Inverse Weibull Weibull Inverse Weibull
From this it is easily seen that the failure rate is unimodal.
1 0.0010 0.1448 6.908 999.5 The failure rate has a peak at t ˆ tM given by the solution of
1.5 0.0100 0.2757 3.627 99.97
2 0.0316 0.3805 2.628 31.61
the following equation:
2.5 0.0631 0.4616 2.166 15.85
3 0.1000 0.5251 1.905 9.998 z tM †
3.5 0.1390 0.5757 1.737 7.196 ˆ 1 1 1=b: 10†
4 0.1779 0.6168 1.621 5.623
1 2 e2z tM †
4.5 0.2155 0.6508 1.537 4.641 This is in contrast to the basic Weibull model where the
5 0.2512 0.6794 1.472 3.981 failure rate is either decreasing (for b , 1), constant (for
5.5 0.2848 0.7037 1.421 3.511
6 0.3165 0.7246 1.380 3.162 b ˆ 1) or increasing (for b . 1).
6.5 0.3455 0.7428 1.346 2.894
7 0.3728 0.7587 1.318 2.683 2.3. Moments

The kth moment of Y is

basic inverse Weibull model following which we discuss the
graphical plotting technique to determine the suitability of Z1 Z1
the model to model a given data set. Mk ˆ tk f t† dt ˆ hbt k2b21† e2z t† dt
0 0
De®ne h ˆ ab ; z ˆ z t† ˆ a=t†b and u ˆ t=a†b :
Z1  z 2k=b
ˆ e2z dz ˆ ak G 1 2 k=b†: 11†
2.1. Density function 0 h

Differentiating F(t) given by Eq. (3) yields the density The integral does not have a ®nite value for k > b . The
function reason for this is that the inverse Weibull distribution has a
heavy right tail.
f t† ˆ F 0 t† ˆ hbt2b21 e2z t† : 4†
2.4. Transformed plots
To study the shapes for f(t) we differentiate it and this
yields We consider two transforms.

f 0 t† ˆ f t†t2b21 ‰hb 2 b 1 1†tb Š: 5† 2.4.1. WPP transform

It is positive for t , tm and negative for t . tm, where The Weibull transform is given by

tm ˆ a 1 1 1=b†21=b : 6† x ˆ ln t†; y ˆ ln 2ln 1 2 F t†††: 12†

This implies that the density function is unimodal. The Weibull model (given by Eq. (1)) gets transformed to
This is in contrast to the basic Weibull model where the
shape is either monotonically decreasing (for b # 1) or y ˆ bx 2 b ln a†: 13†
uni-modal (for b . 1).
The plot of y vs. x is called the WPP plot. Note that the
An interesting comparison is the right and left tails for the
plot is a straight-line.
Weibull and inverse Weibull distributions. Table 1 shows
The inverse Weibull model (given by Eq. (3)) gets
the fractiles t0.001 and t0.999 (scaled by the shape parameter a )
transformed to
for the two distributions for a range of the shape parameter
b . It indicates that the left tail for the inverse Weibull distri-
bution is lighter and the right tail is heavier. y ˆ ln 2ln 1 2 e 2z ††; 14†

2.2. Failure rate function where

The failure rate function is given by

z ˆ a=t†b ˆ e‰b ln a†2bxŠ : 15†
2b21 2z 2z
r t† ˆ hbt e = 1 2 e †: 7†
Note that y is a highly non-linear function of x. Fig. 1
It is easily shown that shows the WPP plots for the two models (for a ˆ 5.0 and
b ˆ 3.5).
lim r t† ˆ lim r t† ˆ 0: 8† The results for small and large t, for the inverse Weibull
t!0 t!1
 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81 75

and this is a highly nonlinear function. Fig. 2 shows the

IWPP plots for both models for a ˆ 5.0 and b ˆ 3.5.
The results for small and large t, for the Weibull model,
are as follows:

² For small t (or small u) we have

y ˆ 2ln 2ln 1 2 e2u †† < 2ln 2ln u††

ˆ 2ln 2b x 2 ln a†††: 21†

² For large t (or large u) we have

Fig. 1. WPP plots for Weibull and inverse Weibull models.
y ˆ 2ln 2ln 1 2 e2u †† < 2ln e2u † ˆ u

model, are as follows: ˆ exp‰b x 2 ln a††Š: 22†

² For small t (or large z) we have

y ˆ ln 2ln 1 2 e2z †† < ln e2z † ˆ 2z ˆ 2e2b x2ln a†† : 2.4.3. Comparison between WPP and IWPP plots
The plot of the inverse Weibull on WPP and the plot of
16†
the Weibull on IWPP are similar in their shape although one
is convex and the other is concave. To note that one can be
² For large t (or small z) we have viewed as a mirror image of the other. This is an interesting
feature.
y ˆ ln 2ln 1 2 e2z †† < ln 2ln z†† ˆ ln‰b x 2 ln a††Š:
Given a data set one plots it on the WPP (see Ref.
17† [12] for more on the plotting details as it depends on
the type of data Ð complete, censored or grouped). If
2.4.2. IWWP transform the plot has a shape similar to that in Fig. 1 then
The inverse Weibull transform is given by inverse Weibull model can be viewed as a potential
x ˆ ln t†; y ˆ 2ln 2ln F t†††: 18† model to model the given data set. In this case, the data
can be plotted on the IWPP to see whether it is roughly
The above transformation was ®rst proposed by Drapella scattered along a straight-line in which case the inverse
[3]. The plot y vs. x is called the inverse Weibull Probability Weibull can be used to model the data.
Paper (IWPP) plot.
Under this transform, the inverse Weibull model (given
by Eq. (3)) becomes a straight line: 3. Mixture model involving two inverse Weibull
distributions
y ˆ b x 2 ln a††: 19†
The model is given by
In contrast, the basic Weibull model gets transformed into
F t† ˆ pF1 t† 1 qF2 t†; 23†
y ˆ 2ln 2ln 1 2 e21=z †† 20†
where F1(t) and F2(t) are the two basic inverse Weibull
distributions of the form given by Eq. (3) and p and q are
positive with p 1 q ˆ 1: Let a 1 and b 1 denote the scale and
shape parameters for the sub-population F1(t) and a 2 and b 2
denote the same for the sub-population F2(t), respectively.

3.1. Density function

Since the density function for each sub-population is

unimodal, the density function for the mixture model can
be one of the following two shapes: (i) unimodal and (ii)
bimodal. Fig. 3 illustrates both these shapes. The parameters
for the bi-modal shape are p ˆ 0.4, b 1 ˆ 1.5, a 1 ˆ 3,
b 2 ˆ 3.5, a 2 ˆ 5 and for the unimodal shape are p ˆ 0.25,
Fig. 2. IWPP plots for Weibull and inverse Weibull models. b 1 ˆ 1.5, a 1 ˆ 6, b 2 ˆ 3.5, a 2 ˆ 5.
76 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81

with coordinates xI and yI. They are given by

b1 ln a1 † 2 b2 ln a2 †
xI ˆ ;
b1 2 b2
24†
b b ‰ln a1 † 2 ln a2 †Š
yI ˆ 1 2 :
b1 2 b2
We prove some properties of the IWPP plot:

1. Since min{F1 t†; F2 t†} , F t† , max{F1 t†; F2 t†}; and

log transformation being monotonic, we have min
{y1 x†; y2 x†} , y x† , max{y1 x†; y2 x†}: This implies
Fig. 3. Plots of the density function for mixture model. that the IWPP plot is always within the cone de®ned by
L1 and L2. The IWPP plot passes through the intersection
point given by Eq. (24).
3.2. Failure rate 2. The location of the intersection point is in¯uenced by the
parameters of the two sub-populations. yI , 0 if a 1/a 2
Since the failure rates of the two sub-populations are (the ratio of the two scale parameters) .1, ˆ 0 if
unimodal, the possible shapes for the failure rate for the a1 =a2 ˆ 1 and .0 if a1 =a2 , 1: The sign of xI depends
mixture model are: (i) unimodal and (ii) bimodal. Fig. 4 on both the scale and shape parameters. The slope at the
illustrates these two shapes and the parameter values are intersection is given by
the same as for Fig. 3. y 0 xI † ˆ pb1 1 qb2 : 25†

3.3. IWPP plot 3. The asymptotes are as follows. For small and large x (or
t), we have the following the limits:
We need to consider the following two cases separately: lim z2 =z1 ˆ 1; lim z2 =z1 ˆ 0: 26†
t!0 t!1
Case (i). The shape parameters are different and without
loss of generality we assume that b 1 , b 2,
Case (ii). The two shape parameters are the same so that
b1 ˆ b2 ˆ b† and without loss of generality we assume The ®rst limit shows that F1 t† q F2 t† when t ! 0.
that a 1 , a 2. Therefore, the left asymptote is given by
Let L1 and L2 denote the IWPP plots for the two sub- L1 :
populations. These are given by Eq. (19) with the
parameters corresponding to the two sub-populations. The y ˆ 2ln 2ln pF1 t† 1 qF2 t††† < 2ln 2ln pF1 t†††
two intersect when the shape parameters are different and
are parallel when they are the same. De®ne zi ˆ zi t† ˆ ! b1 x 2 ln a1 ††: (27)
t=ai †2bi ; i ˆ 1; 2:
Case (i). b 1 , b 2 For large x (or t), we have
In this case L1 and L2 intersect and let I denote this point 2ln F t†† ˆ 2ln p e2z1 1 q e2z2 †

< 2ln p 1 2 z1 † 1 q 1 2 z2 †† < pz1 1 qz2 : 28†

As a result, the right asymptote is given by
LR :

y ˆ 2ln pz1 1 qz2 † ˆ b1 x 2 ln a1 †† 2 ln p 1 qz2 =z1 †

! b1 x 2 ln a1 †† 2 ln p† (29)

that is parallel to L1 but is displaced vertically by ln(1/p).

Fig. 5(a) shows the IWPP plot with the following values
for the model parameters: p ˆ 0.4, b 1 ˆ 1.5, a 1 ˆ 3,
b 2 ˆ 3.5, a 2 ˆ 5. Note that for this parameter set a1 =a2 ,
1: In this case, the left in¯ection point can be used to esti-
Fig. 4. Plots of failure rate for mixture model. mate the parameter p as indicated later.
 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81 77

right asymptote yields the estimates of b 1 and a 1 using

the estimate of p. Locate point I (the intersection of L1 ˆ
^ with the ®tted IWPP plot) and obtain the
LR 1 ln p††
coordinates of the point. Using this in Eq. (24) yields
estimates of b 2 and a 2.

Comment. The graphical method involves a degree of

subjectivity and yields crude estimates. One needs to use
statistical methods (such as maximum likelihood method) to
get better estimates. When the statistical method involves an
iterative procedure, the estimates based on the graphical
method serve as good initial values for iteration process.
Case (ii). b1 ˆ b2 ˆ b†
In this case L1 and L2 are parallel to each other.
For small x (or t) we have
" b #
a1 2 ab2
F2 t†=F1 t† ˆ exp z1 t† 2 z2 t†† ˆ exp ! 0:
tb
30†

As a result, the left asymptote is given by

L1 :

y ˆ 2ln 2ln pF1 t† 1 qF2 t††† < 2ln 2ln pF1 t††† 31†
Fig. 5. (a) IWPP plot for mixture model Ð Case (i) and a1 =a2 , 1: (b)
! b x 2 ln a1 ††:
IWPP plot for mixture model Ð Case (i) and a1 =a2 . 1:

For large x (or t), we have

Fig. 5(b) shows the IWPP plot with the model parameters:
p ˆ 0.4, b 1 ˆ 1.5, a 1 ˆ 3, b 2 ˆ 3.5, a 2 ˆ 1. For this para- 2ln F t†† ˆ 2ln p e2z1 1 q e2z2 †
meter set a1 =a2 . 1: In this case, the right in¯ection point
can be used to estimate the parameter q as indicated later. < 2ln p 1 2 z1 † 1 q 1 2 z2 †† < pz1 1 qz2 : 32†
We now consider the problem of parameter estimation
based on the IWPP plot. Since the IWPP plots for the This implies that the right asymptote is given by
inverse Weibull mixture and the WPP plot for Weibull LR :
mixture are mirror images of each other, the parameters
can be estimated using a procedure similar to that in y ˆ 2ln pz1 1 qz2 † ˆ b x 2 ln a1 †† 2 ln p 1 qz2 =z1 †
Ref. [5]. We omit the details and present the ®nal results. 33†
 b !
We need to consider the following three cases separately: a
! b x 2 ln a1 †† 2 ln p 1 q 2
a1
1. a1 =a2 < 1 (Well-mixed case). From the ®t to the left and
right asymptotes obtain estimates of b 1, a 1, and p. Locate
point I (the intersection of L1 with the ®tted IWPP plot)
and obtain the coordinates of the point. Using this in
Eq. (24) yields estimates of b 2 and a 2.
2. a1 =a2 p 1 (Well-separated case). From the ®t to the left
asymptote obtain estimates of b 1 and a 1. Locate point I
(the intersection of L1 with the ®tted IWPP plot) and
obtain the coordinates of the point. Using this in
Eq. (24) yields estimates of b 2 and a 2. Locate the left
in¯ection point of the plot and obtain the y-coordinate of
this point. Let Y denote the value of this co-ordinate, then
the estimate of p is given by exp(2exp(2Y)).
3. a1 =a2 q 1 (Well-separated case). The estimate of q (or
1 2 p) is given by exp(2exp(2Y)) where Y is the y
co-ordinate of the right in¯ection point. A ®t to the Fig. 6. IWPP plot for mixture model Ð Case (ii).
78 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81

that is parallel to L1 but is displaced vertically by ln p 1

q a2 =a1 †b †:
A typical IWPP plot is as shown in Fig. 6 (with the para-
meters p ˆ 0.4, b 1 ˆ b 2 ˆ b ˆ 3.5, a 1 ˆ 3, a 2 ˆ 5).
If a smooth ®t to the plot of a given data set has a shape
indicated in Fig. 6, then the data set can be modelled by a
mixture model with same shape parameter. The model para-
meters can be estimated from the plot using the following
procedure. From the left and right asymptotes the estimates
of b 1, a 1, and p 1 qk (where k ˆ a2 =a1 †b ) can be obtained.
Denote p 1 qk as c1. Locate the intersection of the IWPP
plot and x-axis and this yields an estimate of x0, the x-coor-
dinate of this point. De®ne c2 ˆ exp‰2 a1 =ex0 †b Š: Since Fig. 8. Plots of failure rate function for the competing risk model.
y x0 † ˆ 2ln 2ln pc2 1 qck2 †† ˆ 0; then p and k can be
obtained by solving the following equations: are unimodal, the possible shapes of the density func-
tion for the competing risk model are: (i) unimodal and
p 1 qk ˆ c1 ; pc2 1 qck2 21
ˆe : 34† (ii) bimodal.
Estimate of a 2 can be obtained from the estimate of k. Fig. 7 illustrates both these shapes. The parameter values
Comment. If a1 =a2 p 1; it may be hard to ®t the right for the bi-modal shape are b 1 ˆ 1.5, a 1 ˆ 3, b 2 ˆ 3.5,
asymptote LR and hence the parameter p cannot be estimated a 2 ˆ 5 and for the unimodal shape are b 1 ˆ 1.5, a 1 ˆ 6,
by the above method. In this case, one can locate the in¯ec- b 2 ˆ 3.5, a 2 ˆ 5.
tion point and obtain its y-coordinate. Let Y denote the esti-
mated value of this co-ordinate, then an estimate of p is 4.2. Failure rate
given by exp(2exp(2Y)). In this case the failure rate r t† ˆ r1 t† 1 r2 t† where
r1(t) and r2(t) are the failure rates for the two sub-popu-
4. Competing risk model involving two inverse Weibull lations. Since both r1(t) and r2(t) are unimodal, the two
distributions possible shapes of the failure rate are: (i) unimodal and
(ii) bimodal.
The model is given by Fig. 8 illustrates these two shapes. The parameter values
for the bi-modal shape are b 1 ˆ 1.5, a 1 ˆ 2, b 2 ˆ 3.5,
F t† ˆ F1 t† 1 F2 t† 2 F1 t†F2 t†; 35† a 2 ˆ 5 and for the unimodal shape are b 1 ˆ 1.5, a 1 ˆ 6,
where the two sub-populations are inverse Weibull distribu- b 2 ˆ 3.5, a 2 ˆ 5.
tions of the form given by Eq. (3). Note that in this case we
can have F1 t† ; F2 t†: We can assume without loss of 4.3. IWPP plot
generality that b1 # b2 and a1 # a2 when b1 ˆ b2 : A typical IWPP plot is as shown in Fig. 9 (with the para-
meter values b 1 ˆ 1.5, a 1 ˆ 3, b 2 ˆ 3.5, a 2 ˆ 5).
4.1. Density function
This is similar to the WPP plot for the competing risk
The density function for this model is given by model involving two Weibull distributions discussed in
Jiang and Murthy [6], i.e. both are convex.
f t† ˆ f1 t†‰1 2 F2 t†Š 1 f2 t†‰1 2 F1 t†Š: 36† For small x (or t), F1 t† q F2 t†: As a result, Eq. (36) can
Since the density functions for the two sub-populations

Fig. 7. Plots of density function for competing risk model. Fig. 9. IWPP plot for competing risk model.
 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81 79

be approximated by F t† < F1 t†: This results in the follow-

ing left asymptote:
L1 :

y ˆ 2ln 2ln F1 t† 1 F2 t† 2 F1 t†F2 t†††

! 2ln 2ln F1 t††† ˆ b1 x 2 ln a1 ††: (37)

For large x (or t), we have

zi t† ! 0; F t† < 1 2 z1 z2 ; ln 1 2 z1 z2 † < 2z1 z2 :
As a result, the right asymptote is given by
LR : Fig. 10. Plots of density function for multiplicative model.

y < ln z1 z2 † ˆ b1 x 2 ln a1 †† 1 b2 x 2 ln a2 ††
the density function plots for a range of parameter values
  indicated that the shape is always unimodal. Fig. 10 shows
b1 ln a1 † 1 b2 ln a2 †
ˆ b1 1 b2 † x 2 : (38) the plot for the following parameter values: b 1 ˆ 1.5,
b1 1 b2
a 1 ˆ 3, b 2 ˆ 3.5, a 2 ˆ 5.
Note that the coordinates of the intersection of L1 and LR An explanation for why the plots are always unimodal is
are (ln(a 2), b 1 ln(a 2/a 1)). Also, there is a link between the as follows. For the density function to be bi-modal, the two
IWPP plot for the inverse Weibull competing risk model sub-populations must be well separated. Without loss of
and the WPP plot for the Weibull multiplicative model. generality, assume that tm1 p tm2 ; where tm1 and tm2 are
The left asymptote for the former is the same as the right the locations of the modes for the two sub-populations.
asymptote for the latter, and the right asymptote for the Over the interval t # tm1 and in region close to t ˆ tm1 ;
former is the same as the left asymptote for the latter. F2 t† < 0 and f2 t† < 0; and as a result f t† < 0: Thus, the
If the IWPP plot of a given data set has a shape similar to density function cannot have a peak near the region close to
that shown in Fig. 9, then the data can be modelled by the t ˆ tm1 : This implies that f(t) is unimodal.
competing risk model. In this case, the model parameters The above results are in contrast to the Weibull multi-
can be estimated as follows. The left asymptote yields esti- plicative model which exhibits bimodal shape for the
mates of b 1, a 1 and the right asymptote and Eq. (38) yield density function. This is because the Weibull model does
the estimates of b 2 and a 2. We need to ensure that the two not have a light left tail (as for the inverse Weibull) so that
asymptotes intersect at the point ln(a 2), b 1 ln(a 2/a 1)). the condition F2 t† < 0 and f2 t† < 0 for t # tm1 do hold.

5.2. Failure rate

5. Multiplicative model
The failure rate function is given by
The model is given by
r t† ˆ r1 t†p t† 1 r2 t†q t†; 41†
F t† ˆ F1 t†F2 t†; 39†
where the two sub-populations are inverse Weibull distribu- where p t† ˆ F2 t†‰1 2 F1 t†Š= 1 2 F1 t†F2 t†† and q t† ˆ
tions of the form given by Eq. (3). If b1 ˆ b2 then the two F1 t†‰1 2 F2 t†Š= 1 2 F1 t†F2 t††:
sub-populations can be merged into one sub-population
with shape parameter b 1 and scale parameter ab1 1 1
ab2 1 †1=b1 : In this case, the model gets reduced to a simple
inverse Weibull distribution. Hence, we assume, without
loss of generality, that b1 , b2 :

5.1. Density function

The density function is given by

f t† ˆ f1 t†F2 t† 1 f2 t†F1 t†: 40†
Since the density functions for the two sub-populations
are unimodal, one can expect the possible shapes of the
density function for the multiplicative model to be (i)
unimodal and (ii) bimodal. However, an examination of Fig. 11. Plots of failure rate function for multiplicative model.
80 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81

Since the failure rates of two sub-populations are uni- ² For large t we have
modal, the possible shapes of the failure rate are unimodal
and bimodal. However, an examination of the failure rate F t† < F1 t†: 46†
function plots for a range of parameter values indicated that
the shape is always unimodal. Fig. 11 shows the plot for the These imply that the two asymptotes for the IWPP are
following parameter values: b 1 ˆ 1.5, a 1 ˆ 3, b 2 ˆ 3.5, given by L2 as x ! 2 1 (or t ! 0) and by L1 as x ! 1 (or
a 2 ˆ 5. t ! 1).
A typical IWPP plot is as shown in Fig. 12 (with the
5.3. IWPP plot parameter values b 1 ˆ 1.5, a 1 ˆ 3, b 2 ˆ 3.5, a 2 ˆ 5).
This is similar to the WPP plot for a multiplicative model
Under the inverse Weibull transform, the IWPP plot is involving two Weibull distributions discussed in Ref. [6],
given by i.e. both are concave. Also, there exists a link between the
y ˆ 2ln‰z1 x† 1 z2 x†Š 42† IWPP plot for the inverse Weibull multiplicative model and
the WPP plot for the Weibull competing risk model
and this is a nonlinear relationship between y and x. It is discussed in Ref. [6]. That is, the left asymptote for the
easily shown that former is the same as to the right asymptote for the latter,
and the right asymptote for the former is the same as the left
b1 z 1 1 b 2 z 2 2 b1 2 b2 †2 z1 z2 asymptote for the latter.
y 0 x† ˆ ; y 00 ˆ ,0
z1 1 z2 z1 1 z2 †2 Let (xI, yI) denote the co-ordinates of the intersection of L1
43† and L2, then we have
so that the IWPP plot is concave. b1 1 b2
y xI † ˆ yI 2 ln 2†; y 0 xI † ˆ : 47†
From Eq. (39) we have 2
F t† ˆ exp 2z1 2 z2 † ˆ exp 2z1 1 1 z2 =z1 †† If the IWPP plot of a given data set has a shape similar to
that shown in Fig. 12, then the data can be modelled by the
ˆ exp 2z2 1 1 z1 =z2 ††: 44† multiplicative model. In this case, the model parameters can
be estimated as follows. The left asymptote yields estimates
It is easily seen that of b 2, a 2 and the right asymptote yields the estimates of b 1
and a 1. One should ensure that Eq. (47) can be roughly
ab1 1 b2 2b1
lim z1 =z2 ˆ lim t ˆ 0; satis®ed when ®tting the asymptotes.
t!0 t!0 ab2 2 Comment. If a1 =a2 p 1 ‰q 1Š; it may be hard to ®t the
right [left] asymptote L1 [L2]. In this case, one can locate the
ab2 2 b1 2b2
lim z2 =z1 ˆ lim t ˆ 0: intersection of the plot and the x-axis and obtain x-coordi-
t!1 t!1 ab1 1 nate, x0, and the slope at this point, y 0 (x0). From Eqs. (42)
and (43), estimates of b 1 and a 1 [b 2 and a 2] are obtained by
Using these limits in Eq. (44) yields the following results:
solving the following set of equations:
² For small t we have z1 x0 † 1 z2 x0 † ˆ 1; b1 z1 x0 † 1 b2 z2 x0 † ˆ y 0 x0 †: 48†
F t† < F2 t†: 45†

6. A numerical example

The data (reproduced below) is from Example 1 of Ling

and Pan [11]

9.6598 13.276 17.452 20.862 25.836 27.789

30.518 34.779 35.437 36.087 37.904 38.324
38.455 38.458 39.403 39.408 39.612 40.178

They use a competing risk model involving two 3-para-

meter Weibull distributions to ®t the data. The parameters
are estimated by minimizing the maximum absolute differ-
ence between the empirical distribution and the model
Fig. 12. IWPP plot for multiplicative model. distribution function. Their parameter estimates are as
 R. Jiang et al. / Reliability Engineering and System Safety 73 (2001) 73±81 81

of the density and failure rate functions and the IWPP

plots for these three models. Also, we have discussed when
one of these models can be used to model a given data set
and the estimation of the model parameters based on the
IWPP plots.
Some topics for future study are as follows:

² Extending the model to include k (.2) sub-populations

each being an inverse Weibull distribution.
² Models involving a Weibull distribution and an inverse
Weibull distribution.
² More re®ned statistical methods for parameter estimation
and model validation for the models studied in this paper.
Fig. 13. IWPP plot for the Example.

follows:
Acknowledgements
a^ 1 ˆ 36:3701; b^ 1 ˆ 1:3308; t^ 1 ˆ 9:4121;
The authors would like to thank the referee for the
a^ 2 ˆ 2:3729; b^ 2 ˆ 1:7183; t^ 2 ˆ 36:9330: extremely helpful comments. The ®rst and third authors
wish to acknowledge the Hong Kong Polytechnic Univer-
The corresponding maximum absolute difference
sity for the ®nancial support of the project (No. G-YW41).
between the two distributions is 0.04, which is less than
the allowable critical value 0.31 corresponding to signi®-
cance level of 0.05. The corresponding sum of absolute References
differences is 0.54.
An IWPP plot of the data is show in Fig. 13. As can be [1] Calabria R, Pulcini G. Con®dence limits for reliability and tolerance
seen, a smooth ®t to the data suggests that the inverse limits in the inverse Weibull distribution. Reliability Engng Syst
Safety 1989;24:77±85.
Weibull competing risk model might be appropriate to
[2] Calabria R, Pulcini G. On the maximum likelihood and least-squares
model the data set. Using the approach discussed in Section estimation in the inverse Weibull distribution. Statistica Applicata
4 and yielded the following estimates for the model 1990;2(1):53±66.
parameters: [3] Drapella A. The complementary Weibull distribution: unknown or
just forgotten? Quality Reliability Engng Int 1993;9:383±5.
[4] Erto P. Genesis, properties and identi®cation of the inverse Weibull
a^ 1 ˆ 32:4; b^ 1 ˆ 0:98; a^ 2 ˆ 38:2; b^ 2 ˆ 47:2: lifetime model (in Italian). Statistica Applicata 1989;1(2):117±28.
[5] Jiang R, Murthy DNP. Modeling failure-data by mixture of two
The corresponding maximum absolute difference is
Weibull distributions: a graphical approach. IEEE Trans Reliability
0.12, which is also less than the allowable critical value 1995;44(3):477±88.
0.31. The corresponding sum of the absolute difference is [6] Jiang R, Murthy DNP. Reliability modeling involving two Weibull
0.54. distributions. Reliability Engng Syst Safety 1995;47:187±98.
As can be seen both models are appropriate to model the [7] Jiang R, Zuo MJ, Li H-X. Weibull and inverse Weibull mixture
models allowing negative weights. Reliability Engng Syst Safety
data set. However, the advantage of our model (inverse
1999;66(3):227±34.
Weibull multiplicative model) is that it has only four para- [8] Johnson NL, Kotz S, Balakrishnan N. Continuous univariate distribu-
meters as opposed to the model of Ling and Pan which has tions. 2nd ed., vol. 1. New York: Wiley, 1994.
six parameters. [9] Keller AZ, Giblin MT, Farnworth NR. Reliability analysis of
commercial vehicle engines. Reliability Engng 1985;10(1):15±25.
[10] Keller AZ, Kamath ARR. Alternative reliability models for mechan-
7. Conclusions ical systems. Proceedings of the Third International Conference on
Reliability and Maintainability. 1982. p. 411±5.
In this paper, we have carried out a study of three [11] Ling J, Pan J. An optimization method of parameter estimation for
mixed Weibull distributions. Reliability and quality in design.
models (mixture model, competing risk model and Proceedings of the fourth ISSAT International Conference, 1998
multiplicative model) involving two inverse Weibull August 12±14; Seattle, Washington, USA. p. 237±43.
distributions. We have given a complete characterization [12] Nelson W. Applied life data analysis. New York: Wiley, 1982.