Statistical Modelling 2006; 6: 265–277

Generalization of the Weibull distribution: the odd

Weibull family
Kahadawala Cooray
Department of Mathematical Sciences, University of Nevada, Las Vegas, USA

Abstract: A three-parameter generalization of the Weibull distribution is presented to deal with general
situations in modeling survival process with various shapes in the hazard function. This generalized Weibull
distribution will be referred to as the odd Weibull family, as it is derived by considering the distributions of
the odds of the Weibull and inverse Weibull families. As a result, the odd Weibull family is not only useful
for testing goodness-of-fit of the Weibull and inverse Weibull as submodels, but it is also convenient for
modeling and fitting different data sets, especially in the presence of censoring. The model parameters for
uncensored data are estimated in two different ways because of the fact that the inverse transformation
of the odd Weibull family does not change its density function. Adequacy of the model for the given
uncensored data is illustrated by using the plot of scaled fitted total time on test (TTT) transforms.
Furthermore, simulation studies are conducted to measure the discrepancy between empirical and fitted
TTT transforms by using a previously proposed test statistic. Three different examples are, respectively,
provided based on data from survival, reliability and environmental sciences to illustrate increasing, bathtub
and unimodal failure rates.

Key words: confidence band; goodness-of-fit; maximum likelihood; total time on test transform

Data and code link available from: http://stat.uibk.ac.at/SMIJ

Received May 2005; revised February, April 2006; accepted May 2006

1 Introduction
The Weibull distribution, having exponential and Rayleigh as submodels, is often used
for modeling lifetime data. When modeling monotone hazard rates, the Weibull distri-
bution may be an initial choice because of its negatively and positively skewed density
shapes. However, the Weibull distribution does not provide a reasonable parametric fit
for some practical applications where the underlying hazard rates may be bathtub or
unimodal shapes. In order to achieve these behaviors from a single distribution and
to model the data that are in the entire positive real-line, well-known three-parameter
Weibull extensions, such as generalized Gamma distribution (Stacy, 1962) and expo-
nentiated Weibull family (Mudholker et al., 1995), have been proposed. In another
sense, these interesting Weibull extensions do not support the modeling of comfort-
able bathtub-shaped failure rate data often encountered in real-life data analyses. Such
bathtub hazard curves have nearly flat middle portions, and the corresponding densities

Address for correspondence: Kahadawala Cooray, Department of Mathematical Sciences, University of

Nevada, Las Vegas, NV 89154, USA. E-mail: [email protected]

© 2006 SAGE Publications 10.1191/1471082X06st116oa

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016
266 K Cooray

have a positive anti-mode. A variety of comfortable bathtub-shaped failure rate life

distributions and their applications can be found in Rajarshi and Rajarshi (1988) and
Murthy et al. (2004).
In this paper, a generalization of the Weibull family is derived by considering the
distributions of the odds of the Weibull and inverse Weibull families. This generalized
Weibull distribution given in Section 2 is henceforth referred to as the odd Weibull family.
This family is not only convenient for modeling comfortable bathtub-shaped failure rates
data but is also suitable for testing goodness-of-fit of the Weibull and inverse Weibull as
submodels. Furthermore, the odd Weibull family accommodates all five major hazard
shapes: constant, increasing, decreasing, bathtub and unimodal failure rates. The inverse
transformation of the odd Weibull family does not change its density function. Hence,
this property is used in Section 3, to estimate the odd Weibull parameters under the
maximum likelihood procedure for a given data set in two different ways. The first
method parameters are estimated by considering them to be initially positive. In the
other method, the same parameters are estimated by taking the inverse sample of the
data set and considering the parameters to be initially negative. Moreover, the total time
on test (TTT) transform procedure is used as a tool to identify the hazard behavior
of the proposed distribution. To measure the discrepancy between empirical and fitted
TTT transforms, a previously proposed test statistic (Aarset, 1987) is used. Simulation
studies are carried out to obtain the upper percentage points of this statistic for the odd
Weibull family. Section 4 provides an analysis of three different examples to illustrate:
increasing, bathtub and unimodal failure rates. In each of these examples, the scaled
fitted TTT graph is plotted along with the scaled empirical TTT graph by providing an
approximate pointwise confidence band to the scaled fitted TTT graph. Finally, Section 5
concludes the article.

2 Odd Weibull family

The odd Weibull family is considered as a suitable answer to the following two questions
found in survival analysis:
1) What are the odds that an individual will die prior to time X, if X follows a certain
life distribution W?
2) If these odds follow some other life distribution L, then what is the corrected
distribution of X?
Obviously, the answer to the first question is very straightforward and depends on the
distribution of W. However, the answer to the second question will vary due to the choice
of both L and W. Let’s answer the first question by representing odds that an individual
will die prior to time X in terms of its survival function SX (x) as (1 − SX (x))/SX (x),
where SX (x) = Pr(X ≥ x), x ∈ (0, ∞). Here one can denote this ratio, the odds of death,
by y(y ∈ (0, ∞)), and it can be considered as a random variable. Suppose that we are
interested in modeling the randomness of the ‘odds of death’ using an appropriate
parametric distribution, say, FY (y). Then, we can write


1 − SX (x)
Pr(Y ≤ y) = FY (y) = FY (2.1)
SX (x)

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 267

Let us consider the log-logistic distribution to model this randomness with its cdf
given by FY (y) = 1 − (1 + yγ )−1 ; 0 < γ < ∞. Perhaps, it is a desirable candidate to
model this randomness, as the analog power transformation exists as follows


1/γ
1 − SX (x) 1 − SY (y)
= (2.2)
SX (x) SY (y)

where SY (y) = 1 − FY (y), and γ can be considered as a correction parameter of the W

distribution.
Now suppose the lifetime random variable X follows the Weibull distribution with
its survival function SX (x) = e−(x/θ ) ; 0 < x < ∞, 0 < α, 0 < θ. Then the cdf of the
α

corrected distribution of X is
  α
γ −1
FX (x) = 1 − 1 + e(x/θ ) − 1 ; 0 < x < ∞, 0 < α, 0 < γ , 0 < θ (2.3)

If the lifetime random variable X follows the inverse Weibull distribution with its
survival function, SX (x) = 1 − e−(x/θ ) , 0 < x < ∞, α < 0, 0 < θ. Then we can write
α

the corrected distribution of X as

 −γ −1
(x/θ )α
FX (x) = 1 − 1 + e −1 ; 0 < x < ∞, α < 0, 0 < γ , 0 < θ (2.4)

One can easily combine Equations (2.3) and (2.4) by writing the correction parameter
β = ±γ , γ > 0, to obtain the cdf of the odd Weibull family as

 β −1
(x/θ )α
F(x; α, β, θ ) = 1 − 1 + e −1 ; 0 < x < ∞, 0 < θ, 0 < αβ (2.5)

The corresponding pdf, hazard function, and the quantile function are, respectively,

  β−1
β −2
αβ x α (x/θ )α  (x/θ )α 
(x/θ)α
f (x; α, β, θ ) = e e −1 1+ e −1 (2.6)
x θ

  β−1
β −1
αβ x α (x/θ )α  (x/θ )α 
(x/θ)α
h(x; α, β, θ ) = e e −1 1+ e −1 (2.7)
x θ

and 
1/β 
−1 1/α u
Q(u) = F (u) = θ ln 1+ ; 0<u<1 (2.8)
1−u

It should be noted that there may be some other ways to derive this distribution
by variable transformations and parameter addition (removal) of existing distributions

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

268 K Cooray

Figure 1 Odd Weibull hazard curves. The dark dotted line (α = 9, β = 0.7, θ = 85), the dashed line
(α = 0.5, β = 0.3, θ = 100), the dot dashed line (α = 1, β = 1, θ = 50), the dark solid line (α = 8, β = 0.01, θ = 45)
and the solid line (α = −1.5, β = −0.1, θ = 75), respectively, represent increasing, decreasing, constant,
bathtub and unimodal failure rates

because of its simple form as a life distribution. It is clear that 1/x transformation
of the odd Weibull family does not change its density form. Furthermore, the odd
Weibull gives the Weibull density when β = 1, and gives the inverse Weibull density when
β = −1. Clearly, the odd Weibull family is asymptotically equivalent to the log-logistic
distribution for larger values of θ .
Figures 1 and 2, respectively, show the odd Weibull hazard and corresponding density
curves for different parameter values.
When both shape parameters, α and β, of the odd Weibull family are negative, the
hazard function given in Equation (2.7) is unimodal. When parameters α and β are posi-
tive, the major shapes of the hazard function are separated by the boundary line of the
space of shape parameters, α = 1 and αβ = 1. Specifically, when (α > 1, αβ > 1), (α < 1,
αβ < 1), (α > 1, αβ ≤ 1) and (α < 1, αβ ≥ 1), the shapes of the hazard function are,

Figure 2 Odd Weibull density curves correspond to the hazard curves given in Figure 1

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 269

respectively, increasing, decreasing, bathtub and unimodal. In addition, when α and β

are positive, some other shapes of the hazard function may appear in the two regions
(α > 1, αβ > 1) and (α < 1, αβ < 1). However, because of complexities of the deriva-
tives of the hazard function, the boundary lines of the parameter space are obtained
numerically.

3 Parametric inference
3.1 Likelihood procedure

The standard maximum likelihood procedure can be used to estimate the odd Weibull
parameters, to the given data, by maximizing the log-likelihood function (l(θ) =
ln L(x1 , x2 , . . . , xn ; α, β, θ )), where x1 , x2 , . . . , xn are an ordered random sample (ie,
x1 ≤ x2 ≤ · · · ≤ xn ) from the odd Weibull family.
The log-likelihood function of the odd Weibull family is given by


n   x α  
δj ln(αβθ −α ) + (α − 1) ln xj +
j α
l(θ ) = + (β − 1) ln e(xj /θ) − 1
θ
j=1


n
 β 
(xj /θ )α
− (1 + δj ) ln 1 + e −1 (3.1)
j=1

where
0 if jth observation is right censored, j = 1, 2, . . . , n
δj =
1 else

The likelihood procedures are available in Cox and Oakes (1984), Kalbfleisch and
Prentice (2002), Rao (1973) and Lawless (2003). In our numerical examples given in
Section 4, the asymptotic standard errors (SE’s) of the estimates are obtained by inverting
the Hessian matrix of the log-likelihood function.

3.1.1 ML parameter estimation for complete data using the inverse property

Let X1 , X2 , . . . , Xn be a random sample from the positive region (ie, α > 0, β > 0) of
the odd Weibull family. Without loss of generality, we can assume that this is an ordered
sample, that is x1 ≤ x2 ≤ · · · ≤ xn . Let us denote the corresponding log-likelihood
function l+ (θ ) = ln L(x1 , x2 , . . . , xn ; α+ , β+ , θ+ ), which can easily be obtained from
Equation (2.9), where α+ = α, β+ = β, θ+ = θ.
Now suppose the inverse random sample of X1 , X2 , . . . , Xn , that is 1/Xn , 1/Xn−1 , . . . ,
1/X1 are coming from the negative region (ie, α < 0, β < 0) of the odd Weibull
family. As before, we can assume that this is an ordered sample, that is 1/xn ≤
1/xn−1 ≤ · · · ≤ 1/x1 . Let us denote the corresponding log-likelihood function

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

270 K Cooray

l− (θ ) = ln L(1/xn , 1/xn−1 , . . . , 1/x1 ; α− , β− , θ− ), which can easily be obtained from

Equation (2.9), where α− = α, β− = β, θ− = θ. Then the following four relations hold.


n
α+ = −α− , β+ = −β− , θ+ = 1/θ− , l+ (θ) = l− (θ) − 2 ln xj (3.2)
j=1

According to these relations, the odd Weibull family allows us to estimate its
parameters in two different ways. Moreover, one can easily prove that this density is
strictly unimodal and positively skewed when α < 0, β < 0. Therefore, this two-way esti-
mation method is useful to avoid some computational issues involved in the likelihood
procedure, especially when the odd Weibull densities are non-unimodal.

3.1.2 Goodness-of-fit

Comparing the goodness-of-fit of a Weibull model is complicated by the large magnitude

of the class of alternatives. By restricting the alternatives to the odd Weibull family, we
can use the usual likelihood ratio statistics for testing the adequacy of the Weibull
and inverse Weibull submodels. The null hypotheses, H011 : β = 1, H012 : (α = 1,
β = 1), H021 : β = 1 and H022 : (α = −1, β = −1), respectively, correspond to the
Weibull, exponential, inverse Weibull and inverse exponential submodels of the odd
Weibull family. The likelihood ratio statistics for H0ij (i = 1, 2; j = 1, 2) are:

supR0ij L(α, β, θ )
= ; i = 1, 2; j = 1, 2 (3.3)
supUR L(α, β, θ )

where, R0ij is the restricted parametric space corresponding to H0ij , i = 1, 2; j = 1, 2.

UR is the unrestricted parameter space.
In terms of the ML estimates, the likelihood ratio statistics reduce to:
L (αw , β = 1, θw ) L (α = 1, β = 1, θe )
11 = ; 12 =
L (α, β, θ ) L (α, β, θ)
L (αiw , β = −1, θiw ) L (α = −1, β = −1, θie )
21 = ; 22 = (3.4)
L (α, β, θ ) L (α, β, θ)

Under the null hypothesis, −2 ln(11 ), −2 ln(12 ), −2 ln(21 ) and −2 ln(22 ),

respectively, follow χ 2 distribution with degrees of freedom 2, 1, 2 and 1. The use of
the odd Weibull family for modeling and testing goodness-of-fit hypotheses is given in
Section 4.

3.2 TTT transforms

In the analysis of lifetime data by the odd Weibull family, predetermination of the sign
and range of α and β can be obtained by using the empirical TTT procedure. For this

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 271

purpose, the empirical TTT transform (HF−1 (r/n)) can be used as a tool to identify the
hazard shape for the given data set.
The scaled empirical TTT is given by

r HF−1 (r/n) r

+ (n − r) Xr:n
i=1 Xi:n
φn = = (3.5)
HF−1 (1)
n
n i=1 Xi

where r = 1, . . . , n and Xi:n , i = 1, . . . , n represent the order statistics of the sample.

If the empirical TTT transform is convex, concave, convex then concave and concave
then convex, the shape of the corresponding hazard function for such failure data is
decreasing, increasing, bathtub and unimodal, respectively, (Barlow and Campo, 1975;
Aarset, 1987; Mudholkar et al., 1996).
The scaled TTT transform for the odd Weibull family can be defined as

HF−1 (u)
φF (u) = ; 0<u<1 (3.6)
HF−1 (1)

 F−1 (u)
where HF−1 (u) = 0 (1 − F(x))dx. The F−1 (u) is given in Equation (2.8).
Adequacy of the model for the given uncensored data can be illustrated by using the
plot of φF (u). The φF (u) of the odd Weibull family is indeed a function of parameters
α and β for a given value u. The typical shapes of φF (u) of the odd Weibull family are
illustrated in Figure 3.
Note that, the ML estimates were used to plot the fitted φF (u) curves for the examples
given in Section 4. In these cases, we graphically investigate the effect of the substitution
of ML estimates for the unknown parameters by constructing an approximate pointwise

Figure 3 Typical shapes of φF (u) of the odd Weibull family. The above light dashed line
(α = 25.745, β = 0.149), the below light dashed line (α = 0.44, β = 0.98), the dark solid lines (above:
α = 7.967, β = 0.098, below: α = 3.5, β = 0.1), and the light solid lines (above: α = −0.683, β = −2.91, below:
α = −2.55, β = −0.7), respectively, represent increasing, decreasing, bathtub and unimodal failure rates

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

272 K Cooray

confidence band for φF (u) (Figures 4–7). This is done by calculating the asymptotic
variance of φF (u) for given u (0 < u < 1), using the delta method.
For the present purpose, to test the goodness-of-fit based on the TTT plot, one can
check the values of the following test statistic (Aarset, 1987).
 1
Rn = Tn2 (u) du (3.7)
0
√
where Tn (u) = n{φn (r/n) − φF (u)}, and the asymptotic distribution of Rn under
exponentiality is obtained. The null hypothesis is rejected when Rn is large, and we
can use it to measure the discrepancy between an empirical and a fitted TTT plot.

3.3 Simulation studies

To gain some insight into the adequacy of the above test statistic (3.7) for the odd
Weibull family, simulation studies are conducted as follows: 100 000 random samples
are generated from the odd Weibull family and Rn values are calculated. 90, 95 and 99%
upper percentage points of Rn are recorded. This procedure is repeated for different
parameter values and sample sizes. The results are given in Table 1. Parameters of the
odd Weibull family are chosen in a way to produce different types of failure rates. From
this table, one can see that the Rn values are decreased with sample size, except when
the parameters of the odd Weibull family represent unimodal failure rates.

Table 1 Upper percentage points of Rn for the odd Weibull family based on 100 000 simulations

Sample size

90% 95% 99%

α β 10 20 50 10 20 50 10 20 50
−3.0 −0.5 1.5324 1.8996 2.5966 1.9166 2.4047 3.3786 2.6724 3.4445 5.1617
−2.0 −2.0 0.1408 0.1517 0.1609 0.1981 0.2169 0.2303 0.3628 0.4188 0.4469
−2.0 −1.0 0.7115 0.8647 1.0894 0.9285 1.1687 1.5323 1.4408 2.1047 3.2768
−1.0 −2.0 0.6123 0.7300 0.9146 0.8147 0.9991 1.3089 1.3054 1.9234 3.0607
−1.0 −1.0 3.5976 5.0896 8.0137 4.0986 5.8790 9.4595 4.9987 7.3592 12.040
−0.5 −3.0 1.0544 1.3146 1.8132 1.3208 1.6927 2.4129 1.8628 2.5363 4.4061
0.5 0.5 1.0202 0.8437 0.7333 1.3784 1.1978 1.0469 2.1045 2.0433 1.8313
0.5 1.0 1.0111 0.8948 0.8138 1.3820 1.2712 1.1605 2.1901 2.1687 2.0180
0.5 3.0 0.4248 0.4307 0.4482 0.5883 0.6024 0.6330 0.9633 1.0105 1.0851
0.5 5.0 0.2002 0.2008 0.2049 0.2807 0.2847 0.2903 0.4843 0.4858 0.5064
1.0 0.5 0.4233 0.3273 0.2823 0.6598 0.4968 0.4143 1.2840 0.9991 0.7731
1.0 1.0 0.3652 0.3210 0.3002 0.5455 0.4715 0.4301 1.0246 0.8799 0.7728
1.0 5.0 0.0544 0.0526 0.0522 0.0783 0.0753 0.0735 0.1404 0.1321 0.1272
5.0 0.1 0.0576 0.0412 0.0336 0.1006 0.0674 0.0505 0.2651 0.1564 0.1032
5.0 0.5 0.0582 0.0497 0.0459 0.0968 0.0779 0.0668 0.2110 0.1633 0.1292
5.0 1.0 0.0304 0.0280 0.0271 0.0475 0.0413 0.0389 0.0994 0.0833 0.0713
8.0 0.1 0.0503 0.0361 0.0288 0.0912 0.0616 0.0452 0.2513 0.1496 0.0983
8.0 0.5 0.0309 0.0264 0.0252 0.0517 0.0418 0.0366 0.1130 0.0898 0.0717
8.0 1.0 0.0139 0.0129 0.0124 0.0219 0.0193 0.0178 0.0465 0.0386 0.0329
8.0 5.0 0.0010 0.0010 0.0009 0.0015 0.0014 0.0013 0.0028 0.0025 0.0024

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 273

4 Examples
Example 1: Increasing failure rate illustration by using the mice mortality data. This
example contains 208 data points, which represent the ages at death in weeks for male
mice exposed to 240r of gamma radiation (Kimball, 1960). The empirical TTT plot for
this data set indicates an increasing hazard rate (dark dotted line in Figure 4). The odd
Weibull family becomes versatile because it expands the Weibull family into a larger
family because of an additional shape parameter. Therefore, it is important to identify
the parameter space with increasing failure rate. From Section 2, when α > 1, αβ > 1,
the hazard function of the odd Weibull family is increasing. Therefore, one can easily
make an initial guess to estimate the parameters.
Table 2 provides the estimated parameter values and the log-likelihood values of the
fitted odd Weibull family to the original and inverse sample of mice data. Specifically,
the inverse sample of mice data is analyzed by using the negative region (Section 3.1.1)
of the odd Weibull family.
Furthermore, the fitness of the odd Weibull family to the original mice data is illus-
trated by the scaled fitted TTT graph (dark solid line in Figure 4) with its 5% confidence
band (solid lines). The calculated Rn value of the original sample of mice data is 0.0107,
which is smaller than 90% upper percentage point of Rn , 0.0249. Therefore, we do not
reject the odd Weibull fit for the mice data.
Table 3 provides the likelihood ratio statistics of the odd Weibull model for the two
null hypotheses H011 : β = 1 to the Weibull submodel (11 ) and H021 : β = −1 to the

Figure 4 TTT transforms of the mice data [f = φF (u) or φn (r /n)]. Scaled empirical (dark dotted line), scaled
odd Weibull fit (dark solid line), 5% confidence band (light solid lines)

Table 2 Estimated values of fitted odd Weibull family for the mice death data


α ± SE
α  ± SE 
β 
θ ± SE l(
θ)
β θ

Original sample 6.2278 ± 0.8326 0.7495 ± 0.1221 131.45 ± 1.9535 −988.89

Inverse sample −6.2278 ± 0.8328 −0.7495 ± 0.1220 0.0076 ± 0.0001 993.25

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

274 K Cooray
Table 3 Likelihood ratio tests of subhypotheses for mice data

Null hypothesis χ12 p-value

H011 −2 ln(11 ) = 2.62 0.11

H021 −2 ln(21 ) = 150.34 0.00

Figure 5 TTT transforms of the device failure data [f = φF (u) or φn (r /n)]. Scaled empirical (dark dotted line),
scaled odd Weibull fit (dark solid line), scaled generalized Weibull fit (light dashed line), 5% confidence band
for the odd Weibull fit (light solid lines)

inverse Weibull submodel (21 ). These values indicate an adequate fit to the Weibull
model, whereas the inverse Weibull fit is inappropriate.

Example 2: Bathtub-shaped failure rate illustration by using the simulated device failure
data. This example represents time to failure of 50 devices put on a life test at time zero
(Aarset, 1987). The empirical TTT procedure indicates a bathtub hazard shape for this
data set (dark dotted line in Figure 5). As before, when modeling bathtub-shaped failure
rate data, it is important to identify the parameter space of the odd Weibull family, which
actually produces a bathtub hazard shape. It could be observed that the major shape
of the hazard function of the odd Weibull family is bathtub-shaped when 1 < α < ∞,
0 < αβ ≤ 1.
As in the previous example, the estimated values are given in Table 4.
In addition, the adequacy of the fit is strengthened by illustrating the scaled fitted
TTT graph (dark solid lines in Figure 5) with its 5% confidence band (solid lines).
This indicates that the data is better-fit with the odd Weibull family. The calculated Rn

Table 4 Estimated values of fitted odd Weibull family for device failure data


α ± SE
α  ± SE 
β 
θ ± SE l(
θ)
β θ

Original sample 6.9657 ± 0.5903 0.0921 ± 0.0171 53.509 ± 2.4733 −215.88

Inverse sample −0.0921 ± 0.0140 −0.0921 ± 0.0140 0.0187 ± 0.0006 92.02

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 275
Table 5 Likelihood ratio tests of subhypotheses for device data

Null hypothesis χ12 p-value

H011 −2 ln(11 ) = 50.65 0.00

H021 −2 ln(21 ) = 98.28 0.00

value of the device failure data is 0.0482, which is slightly smaller than the 95% upper
percentage point of Rn , 0.0485. Therefore, it is hard to reject the odd Weibull fit for the
device failure data.
Mudholkar et al. (1996) used this example to illustrate the flexibility of their
generalized Weibull family under the non-regular case. For comparison purposes, their
generalized Weibull fit is illustrated in Figure 5 (light dashed line).
Meanwhile, the likelihood ratio statistics given in Table 5 of the odd Weibull family
for the null hypotheses H011 : β = 1 and H021 : β = −1 corresponding to the Weibull and
inverse Weibull submodels, indicate their inappropriateness.

Example 3: Unimodal failure rate illustration by using the alluvial diamond data. This
twin data set consists of alluvial diamonds from the Bougban and Damaya deposits
in Guinea of West Africa (Beirlant et al., 1996). The sampling program on Bougban
recovered 683 stones, whereas the Damaya sampling recorded 444 stones. In fact, the
empirical TTT procedure gives unimodal hazard shapes for the twin data set (dark
dotted lines in Figures 6 and 7). From Section 2, we observed that the hazard function
of the odd Weibull family is unimodal when its shape parameters, α and β, are both
negative or α < 1 and αβ ≥ 1. Therefore, to estimate its parameters for unimodal failure
rate data, one can pick the values from one of these ranges as an initial guess. The best-
fitting parameters for the odd Weibull family can be obtained by choosing the maximum
value out of the following two likelihood functions, one from the range of both shape
parameters α < 0, β < 0 and the other function is from α < 1, αβ ≥ 1.

Figure 6 TTT transforms of the Bougban data [f = φF (u) or φn (r /n)]. Scaled empirical (dark dotted line),
scaled odd Weibull fit (dark solid line), 5% confidence band (light solid lines)

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

276 K Cooray

Figure 7 TTT transforms of the Damaya data [f = φF (u) or φn (r /n)]. Scaled empirical (dark dotted line), scaled
odd Weibull fit (dark solid line), 5% confidence band (light solid lines)

Table 6 Estimated values of fitted odd Weibull family for diamond data

Original sample Inverse sample

Bougban Damaya Bougban Damaya

α ± SE
α −1.3431 ± 0.1372 −0.6107 ± 0.1811 1.3430 ± 0.1372 0.6107 ± 0.1811
 ± SE 
β −1.2884 ± 0.1501 −2.9853 ± 0.9173 1.2885 ± 0.1502 2.9853 ± 0.9173
β

θ ± SE 0.2008 ± 0.0064 0.2642 ± 0.0460 4.9807 ± 0.1594 3.7844 ± 0.6585
θ
l(
θ) 113.49 −167.06 −1564.67 −773.02

The estimated values are given in Table 6. Specifically, the inverse samples of diamond
data are analyzed by using the positive region (Section 3.1.1) of the odd Weibull
family.
The odd Weibull fit to the original diamond data using scaled fitted TTT graphs
(dark solid lines) with their 5% confidence bands (light solid lines) are illustrated in
Figure 6 (Bougban) and Figure 7 (Damaya). The calculated Rn values of the Bougban
and Damaya data are, respectively, 0.1111 and 0.0464, which are quite smaller than their
90% upper percentage points of 4.9550 and 2.6762.
Table 7 gives the likelihood ratio statistics of the odd Weibull family for the twin
data set. These values indicate that the Weibull and inverse Weibull submodels are
inappropriate for the diamond data.

Table 7 Likelihood ratio tests of subhypotheses for diamond data

Bougban Damaya
Null hypothesis χ12 p-value χ12 p-value

H011 −2 ln(11 ) = 483.19 0.00 −2 ln(11 ) = 230.90 0.00

H021 −2 ln(21 ) = 5.90 0.02 −2 ln(21 ) = 46.79 0.00

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016

 Generalization of the Weibull distribution 277

5 Conclusions
A generalization of the Weibull distribution, the odd Weibull family, has been presented
for modeling different types of failure rate data. Its applicability and better fitness to
increasing, comfortable bathtub and unimodal failure rates data have been demon-
strated and emphasized using well-known examples by illustrating the scaled empirical
and scaled fitted TTT plots. The upper percentage points of a test statistic, which mea-
sure the goodness-of-fit based on TTT plot, have been tabulated for different parameter
values of the odd Weibull family. Furthermore, permissibility of testing the goodness-
of-fit of the Weibull and inverse Weibull as submodels of the odd Weibull family was
also demonstrated in the examples given in Section 4. The inverse transformation of the
odd Weibull family is the same as the original distribution and is uncommon among the
distributions having bathtub-shaped failure rates. Using this property and the maximum
likelihood procedure, parameters of the odd Weibull family were able to estimate in two
different ways for uncensored samples. This is actually useful to avoid some computa-
tional difficulties involved in the likelihood function, especially when the densities of the
odd Weibull family are non-unimodal.

Acknowledgement
The comments of two referees, an associate editor, and the editor were very helpful in
revising an earlier version.

References
Aarset MV (1987) How to identify a bathtub Lawless JF (2003) Statistical models and methods
hazard rate. IEEE Transactions on Reliability for lifetime data. John Wiley.
R-36, 106–8. Mudholkar GS, Srivastava DK and Freimer M
Barlow RE and Campo R (1975) Total time (1995) The exponentiated Weibull family: a
on test processes and applications to reanalysis of the bus-motor-failure data.
failure data analysis. In Barlow RE, Technometrics 37, 436–45.
Fussell JB and Singpurwalla ND, editors, Mudholkar GS, Srivastava DK and Kollia GD
Reliability and fault tree analysis. Society (1996) A generalization of the Weibull
for Industrial and Applied Mathematics, distribution with application to the analysis
451–81. of survival data. Journal of the American
Beirlant J, Teugels JL and Vynckier P (1996) Statistical Association 91, 1575–83.
Practical analysis of extreme values. Leuven Murthy DNP, Xie M and Jiang R (2004) Weibull
University Press. models. John Wiley.
Cox DR and Oakes D (1984) Analysis of survival Rajarshi S and Rajarshi MB (1988) Bathtub
data. Chapman & Hall. distributions: a review. Communications in
Kalbfleisch JD and Prentice RL (2002) The Statistics-Theory and Methods 17, 2597–621.
statistical analysis of failure time data. John Rao CR (1973) Linear statistical inference and its
Wiley. applications. John Wiley.
Kimball AW (1960) Estimation of mortality Stacy EW (1962) A generalization of the gamma
intensities in animal experiments. Biometrics distribution. Annals of Mathematical
16, 505–21. Statistics 33, 1187–92.

Downloaded from smj.sagepub.com at PENNSYLVANIA STATE UNIV on May 11, 2016