Weibull Lindley Distribution
D.A. Magfira, D. Lestaria) and S. Nurrohmah
Department of Mathematics, Faculty of Mathematics and Natural Sciences (FMIPA), Universitas Indonesia,
Depok 16424, Indonesia
a)
Corresponding author: [email protected]
Abstract. In reliability systems, there are known two types of systems namely series systems and parallel systems. In the
series system, failure will occur if any of the possible events happen. Applications of the series system analysis also varies
from inspecting the durability of manufactured products to examining diseases in human. Therefore, several distributions
have been introduced to model failure data in series system. However, these distributions cannot model data with bathtub
shaped hazard function even though it is the one mostly found in real life situation. As a result, distribution which can
model lifetime data in series system with bathtub-shaped hazard function has to be developed. In real life application, there
is condition where failure could occur caused by several independent events and has a bathtub shaped hazard function, for
example engineering cases and competing risk. Weibull Lindley distribution, which was introduced by Asgharzadeh et al.
(2018), is developed to solve the problem. As Weibull Lindley distribution describes lifetime data of an object that can
experience failure caused by 2 possible events. It can model data with increasing, decreasing and bathtub shaped hazard
function. Asgharzadeh et al. (2018) only show the modeling of Weibull Lindley distribution in medical field which is
competing risk data. This paper discusses the process of forming the Weibull Lindley distribution, its properties and
parameter estimation using the maximum likelihood method. In addition, the application of Weibull Lindley distribution
in engineering field which is the lifetime data of machine consists of two independent components paired in series also be
discussed.
Keywords: Bathtub hazard, compounding distribution, hazard function, maximum likelihood method, series system
INTRODUCTION
Lifetime data tells about time needed for an event or failure to happen. For example, lifetime data describes time
until death for human or time until failure for machine. The analysis of lifetime data is very important in various fields,
such as biomedical science and engineering. In the lifetime data analysis, the object will experience only one event
that can lead to failure. But in real life the population is heterogeneous so it can cause more complicated problems
e.g. failure can occur due to several possible unknown events [1].
In reliability systems, there are known two types of systems namely series systems and parallel systems. In the
series system, failure will occur if any of the possible events happen. In real life application of series system, inter-
event seems to compete each other to cause failure because if one event has occurred, the observation time is
considered finish so that another event cannot occur.
Applications of the series system analysis also varies from inspecting the durability of manufactured products to
examining diseases in human. Thus, several distributions have been introduced to model lifetime data in series system.
Some distributions that can be used to model lifetime data in series system are Exponential Geometric distribution [1],
Exponential Poisson distribution [2], Weibull Geometric distribution [3], Pareto Poisson-Lindley distribution [4], and
so on [5]. However, these distributions cannot model data with bathtub-shaped hazard function even though it is the
one mostly found in real life situation. Bathtub shaped hazard function can be found in the data of machines population
observed since the beginning of usage until failure or human population observed since birth until death. Also, there
is a condition where failure can occur caused by several independent events and has a bathtub shaped hazard function
like in machine with component paired in series case and competing risk case. Therefore, distribution which can model
lifetime data in series system with bathtub-shaped hazard function has to be developed to solve those cases. Weibull
Lindley distribution, which was introduced by Asgharzadeh et al. (2018), is developed to solve the problem [5].
The Weibull Lindley distribution is formed by compounding Weibull distribution and Lindley distribution. It can
model increasing, decreasing, and bathtub shaped hazard function from lifetime data in series system. Weibull Lindley
distribution describes lifetime data of an object that can experience failure caused by 2 possible events. In medical
field, Weibull Lindley is suitable for modelling competing risk data with two risk factors. Asgharzadeh et al. (2018)
had modelled Weibull Lindley distribution in medical field which is competing risk data. Whereas, in engineering
field, Weibull Lindley distribution can model a lifetime data of a machine which has two independent components
paired in series [5]. This paper discusses the formation of Weibull Lindley distribution in a detailed manner, its
properties, and the estimation of its parameters. In addition, this paper model Weibull Lindley distribution in
engineering cases.
METHODOLOGY
Formation of Weibull Lindley Distribution
The Weibull Lindley distribution is formed by compounding Weibull distribution and Lindley distribution. Let ܺଵ
denote a Lindley random variable with parameter ߣ Ͳ and ܺଶ denote a Weibull random variable with parameter
ߙǡ ߚ ͲǤ Assume ܺଵ and ܺଶ are independent random variables. Then, Weibull Lindley random variable is defined as
ܺ ൌ ሺܺଵ ǡ ܺଶ ሻ.
The survival function of Weibull distribution with parameter ߙǡ ߚ [6] known as follows:
ഀ
ܵభ ሺݔሻ ൌ ݁ ିሺఉ௫ሻ ݂ ݔݎ Ͳߙǡ ߚ ͲǤ (1)
The survival function of Lindley distribution with parameter ߣ [7] known as follows:
ͳ ߣ ߣି ݔఏ௫
ܵమ ሺݔሻ ൌ ൬ ݁ ൰ ݂ ݔݎ Ͳǡ ߣ ͲǤ (2)
ͳߣ
Therefore, the distribution function of Weibull Lindley distribution is obtained by compounding two independent, i.e.
Weibull dan Lindley random variables based on compounding methods by Adamidis and Loukas [1], that is defined
by,
ܨ ሺݔሻ ൌ ͳ െ ൣܵభ ሺሻܵమ ሺሻ൧Ǥ
So, the distribution function of Weibull Lindley distribution based on Equation 1 and Equation 2 can be written as
follows:
ͳ ߣ ߣି ݔఒ௫ିሺఉ௫ሻഀ
ܨ ሺݔሻ ൌ ͳ െ ൬ ݁ ൰ ǡ ݔ Ͳǡ ߙ Ͳǡ ߚ Ͳǡ ߣ ͲǤ (3)
ͳߣ
Survival function of Weibull Lindley distribution based on Equation 3 can be written as follows:
ͳ ߣ ߣି ݔఒ௫ିሺఉ௫ሻഀ
ܵ ሺݔሻ ൌ ൬ ݁ ൰ Ǣ ݔ Ͳǡ ߙ Ͳǡ ߚ Ͳǡ ߣ ͲǤ (4)
ͳߣ
In general, the probability density function of Weibull Lindley distribution is:
ͳ ഀ
ሾߙߣሺߚݔሻఈ ߙߚሺͳ ߣሻሺߚݔሻఈିଵ ߣଶ ሺͳ ݔሻሿ݁ ିఒ௫ିሺఉ௫ሻ ݔ Ͳǡ ߙ Ͳǡ ߚ Ͳǡ ߣ Ͳǡ
݂௫ ሺݔሻ ൌ ቐͳ ߣ (5)
Ͳݏݎ݄݁ݐǤ
Known from Fig. 1 that the pdf of Weibull Lindley distribution is unimodal when ߙ ߚ and ߙǡ ߚǡ ߣ ͳ.
When ߙ ͳǡ ߣ ൏ ͳ the pdf of Weibull Lindley distribution will be unimodal. Then, the pdf of Weibull Lindley
distribution will be decreasing when ߙ ൏ ͳ.
Therefore, the hazard function of Weibull Lindley distribution based on Equation 4 and Equation 5 can be written
as follows:
ߣଶ ሺͳ ݔሻ
݄ ሺݔሻ ൌ ߙߚሺߚݔሻఈିଵ Ǣ ݔ Ͳǡ ߙ Ͳǡ ߚ Ͳǡ ߣ ͲǤ (6)
ͳ ߣ ߣݔ
Shown in Fig. 2 that the hazard function shape of Weibull Lindley distribution is increasing when ߙ ͳ.
When Ͳ ൏ ߙ ൏ ͳ the hazard function shape of Weibull Lindley distribution will be decreasing. Then, the hazard
function of Weibull Lindley distribution will bathtub shaped when ߙ close to 0 or ߙ close to 1. In conclusion, the
shape of Weibull Lindley hazard function only depends on the value of ߙ.
The ݎ-th moment function of Weibull Lindley can be written as follows:
ஶ ஶ
ሺെͳሻ ߚ ఈ Ȟሺߙ݅ ߙ ݎ ͳሻ ߣଶ ሺെͳሻ ߚ ఈ Ȟሺߙ݅ ݎ ʹሻ
ܧሾܺ ሿ ൌ ߙߚ ఈ ቈ Ȟሺߙ݅ ߙ ݎሻ ఈାାଵ ቈȞሺߙ݅ ݎ ͳሻ Ǥ (7)
݅Ǩ ߣఈାఈା ሺͳ ߣሻ ͳߣ ݅Ǩ ߣ ߣ
ୀ ୀ
FIGURE 1. Pdf shapes of Weibull Lindley distribution.
FIGURE 2. Hazard function shapes of Weibull Lindley distribution.
Then, the first moment or mean and the second moment of Weibull Lindley distribution can be written as follows:
ஶ ஶ
ሺെͳሻ ߚఈ Ȟሺߙ݅ ߙ ʹሻ ߣଶ ሺെͳሻ ߚఈ Ȟሺߙ݅ ͵ሻ
ܧሾܺሿ ൌ ߙߚ ఈ ቈ Ȟሺߙ݅ ߙ ͳሻ ቈȞሺߙ݅ ʹሻ Ǥ (8)
݅Ǩ ߣఈାఈାଵ ሺͳ ߣሻ ͳߣ ݅Ǩ ߣఈାଶ ߣ
ୀ ୀ
ஶ ஶ
ሺെͳሻ ߚఈ Ȟሺߙ݅ ߙ ͵ሻ ߣଶ ሺെͳሻ ߚ ఈ Ȟሺߙ݅ Ͷሻ
ܧሾܺ ଶ ሿ ൌ ߙߚ ఈ ቈ Ȟሺߙ݅ ߙ ʹሻ ቈȞሺߙ݅ ͵ሻ Ǥ (9)
݅Ǩ ߣఈାఈାଶ ሺͳ ߣሻ ͳߣ ݅Ǩ ߣఈାଷ ߣ
ୀ ୀ
Variance of Weibull Lindley distribution based on Equation 8 and Equation 9 can be written as follows:
ஶ ஶ
ሺെͳሻ ߚ ఈ Ȟሺߙ݅ ߙ ͵ሻ ߣଶ ሺെͳሻ ߚ ఈ Ȟሺߙ݅ Ͷሻ
ܸܽݎሺܺሻ ൌ ߙߚ ఈ ቈ Ȟሺߙ݅ ߙ ʹሻ ቈȞሺߙ݅ ͵ሻ
݅Ǩ ߣఈାఈାଶ ሺͳ ߣሻ ͳߣ ݅Ǩ ߣఈାଷ ߣ
ୀ ୀ
ஶ ஶ (10)
ሺെͳሻ ߚఈ Ȟሺߙ݅ ߙ ʹሻ
ఈ
ߣଶ ሺെͳሻ ߚఈ Ȟሺߙ݅ ͵ሻ
െ ߙߚ ఈାఈାଵ ቈ Ȟሺߙ݅ ߙ ͳሻ ቈȞሺߙ݅ ʹሻ ൩Ǥ
݅Ǩ ߣ ሺͳ ߣሻ ͳߣ ݅Ǩ ߣఈାଶ ߣ
ୀ ୀ
Parameter Estimation of Weibull Lindley Distribution
Likelihood function for Weibull Lindley distribution is needed for maximum likelihood estimation. To obtain
likelihood function, first, let ܺଵ ǡ ܺଶ ǡ ǥ ǡ ܺ a random samples of size n from distribution with pdf (5), then the likelihood
function with observed sample ݔis a multiplication of its pdf:
ஶ
ͳ ಮ ಮ ഀ
ܮሺߙǡ ߚǡ ߣǢ ݔሻ ൌൌ ൬ ൰ ෑሾߙߣሺߚݔ ሻఈ ߙߚሺͳ ߣሻሺߚݔ ሻఈିଵ ߣଶ ሺͳ ݔ ሻሿ ݁ ିఒ σసభ ௫ିσసభሺఉ௫ሻ Ǥ (11)
ͳߣ
ୀଵ
Log-likelihood function is needed to help in obtaining solution. The loglikelihood function can be obtained as:
ஶ ஶ ஶ
݈ሺߙǡ ߚǡ ߣǢ ݔሻ ൌ െ݊ ሺͳ ߣሻ ݈݊ሾߙߣሺߚݔ ሻఈ ߙߚሺͳ ߣሻሺߚݔ ሻఈିଵ ߣଶ ሺͳ ݔ ሻሿ െ ߣ ݔ െ ሺߚݔ ሻఈ Ǥ (12)
ୀଵ ୀଵ ୀଵ
Therefore, the first derivative of Equation 12 with respect to each parameter is needed for maximum likelihood
estimation. Estimated value of ߙǡ ߚ and ߣ are obtained by solving the equation as follows:
ஶ
߲݈ሺߙǡ ߚǡ ߣǢ ݔሻ ͳ
ൌ ൛ൣ൫ߣ ߙߣ݈݊ሺߚݔ ሻ൯ሺߚݔ ሻఈ ൧ ൣ൫ߚሺͳ ߣሻ ߙߚሺͳ ߣሻ݈݊ሺߚݔ ሻ൯ሺߚݔ ሻఈିଵ ൧ൟ
߲ߙ ܹሺݔ ሻ
ୀଵ
ஶ (13)
െ ሺߚݔ ሻఈ ݈݊ሺߚݔ ሻ ൌ Ͳǡ
ୀଵ
where ܹሺݔ ሻ ൌ ሾߙߣሺߚݔ ሻఈ ߙߚሺͳ ߣሻሺߚݔ ሻఈିଵ ߣଶ ሺͳ ݔ ሻሿǤ
Equation 13 can be solved by numerical method that will produce an estimated value of ߙ, which is ߙොǤ
ஶ ஶ
߲݈ሺߙǡ ߚǡ ߣǢ ݔሻ ͳ ߙ
ൌ ሺߙ ଶ ߣሺߚݔ ሻఈ ߙߚሺͳ ݔ ሻሺߚݔ ሻఈିଵ ሻ െ ሺߚݔ ሻఈ ൌ ͲǤ (14)
߲ߚ ߚܹሺݔ ሻ ߚ
ୀଵ ୀଵ
Equation 14 can be solved by numerical method that will produce an estimated value of ߚ, which is ߚመ Ǥ
ஶ ஶ
߲݈ሺߙǡ ߚǡ ߣǢ ݔሻ ͳ ݊
ൌ ሺߙሺߚݔ ሻఈ ሻ ሺߙߚሺߚݔ ሻఈିଵ ሻ ൫ʹߣሺͳ ݔ ሻ൯ െ ݔ െ ൌͲ (15)
߲ߣ ܹሺݔ ሻ ͳߣ
ୀଵ ୀଵ
Equation 15 can be solved by numerical method that will produce an estimated value of ߣ, which is ߣመǤ
Data Illustration
In this section, the application of Weibull Lindley distribution in modelling lifetime data of a series system will be
illustrated. The data consists of 61 machines with two independent components paired in series and the machine will
fail if one of the components fails. The data sourced from Tsai-Hung Fan and Tsung-Ming Hsu [8]. The statistical
descriptions of the data are given in Table 1. In Fig. 3, we see that the data has 2 modus or called bimodal data.
Therefore, Weibull Lindley distribution which has characteristics of bimodal shaped pdf is suitable for modelling
machine lifetime data. Moreover, several other distributions are also suitable for modelling machine lifetime data such
as Weibull distribution and Lindley distribution. Table 2 contains the estimated values of the parameters using the
maximum likelihood method by nlm function in RStudio version 1.2.5019.
Figure 4 gives the empirical and fitted cumulative functions. In Fig. 5 the graph of Weibull Lindley pdf is bimodal
which resembles the shape of machine data histogram. Otherwise, the pdf of Weibull is unimodal, and the pdf of
Lindley is increasing. In Fig. 4. we can see that the cumulative function of Weibull Lindley distribution close fits the
empirical cumulative function.
TABLE 1. Descriptive statistics for machine lifetime data.
Standard
Mean Variance Minimum Maximum
deviation
2.51603 2.164 1.4750 0.148 5.44
FIGURE 3. Histogram of machine lifetime data.
TABLE 2. Maximum Likelihood Estimation of machine lifetime data.
Distribution ࣅ ෝ
ࢻ
ࢼ
Weibull Lindley 0.50117 7.6775 0.2267
Lindley 0.6398 - -
Weibull - 17.147 0.3556
FIGURE 4. The empirical and fitted cumulative functions. FIGURE 5. The pdf and histogram of machine lifetime data
TABLE 3. Summary of fitted model.
Kolmogorov Smirnov
Distribution AIC BIC
Test Statistics
Weibull Lindley 0.067171 204.94 211.277
Lindley 0.119717 429.378 429.378
Weibull 0.128655 459.13 463.352
FIGURE 6. The hazard function of Weibull Lindley distribution.
To compare the fitted model Weibull Lindley distribution with the Weibull and Lindley distribution, we used the
Kolmogorov-Smirnov Test Statistics, the Akaike information criterion (AIC) defined by െʹ݈ ܮ݃ ʹݓ, and the
Bayesian information criterion (BIC) defined by െʹ݈ ܮ݃ ݈݃ݓሺ݉ሻǡ where ݓis the number of estimated
parameters and m is the sample size. The best model would give by the lowest values from Kolmogorov-Smirnov Test
Statistics, AIC and BIC. From Table 3 we have Weibull Lindley is more suitable for modelling the machine lifetime
data than the other distribution. Figure 6 shows that the hazard function of Weibull Lindley distribution in machine
lifetime data is increasing.
RESULTS AND DISCUSSION
1.From data illustration, we can conclude that Weibull Lindley distribution is more suitable for modelling
machine lifetime data designed in series than Weibull distribution or Lindley distribution.
2. Theoretically, it is obtained that the pdf of Weibull Lindley distribution can be bimodal. It is also shown in
the machine lifetime data that the pdf shape is bimodal.
3. Theoretically, it is obtained that the hazard function of Weibull Lindley can be increasing, decreasing and
bathtub shaped. The shape of Weibull Lindley hazard function only depends on the value of ߙ as follows:
• the hazard function is increasing when ߙ ͳ,
• the hazard function is decreasing when Ͳ ൏ ߙ ൏ ͳǡ
• the hazard function is bathtub when ߙ being close enough to 0 or ߙ being close enough to 1.
This can be seen from the application of machine lifetime data that the estimated value of ߙ is 7.6775.
Because the estimated value of ߙ ͳ so the shape of hazard function is increasing.
CONCLUSION
In this paper, we compound Weibull distribution and Lindley distribution to create Weibull Lindley distribution
by defining the Weibull Lindley random variables as ܺ ൌ ሺܺଵ ǡ ܺଶ ሻ when ܺଵ is Lindley random variables and ܺଶ
is Weibull random variables. Some properties of our proposed model Weibull Lindley distribution were derived, i.e.
the pdf can be unimodal, monotonically decreasing, or bimodal. Moreover, the hazard rate of Weibull Lindley
distribution can be decreasing or increasing, or bathtub shaped. The estimation of parameters was obtained by
maximum likelihood method. The results of simulation on machine lifetime data show that Weibull Lindley is more
suitable for modelling the machine lifetime data than the other distribution.
ACKNOWLEDGMENTS
Authors wishing to acknowledge assistance from various parties, including lecturers and friends at Faculty of
Mathematics and Natural Sciences of Universitas Indonesia. This work is funded by DRPM Universitas Indonesia.
REFERENCES
1. K. Adamidis and S. Loukas, Stat Probab Lett 39, 35-42 (1998).
2. Kuú, C. A new lifetime distribution. CSDA 51, 4497-4509 (2007).
3. W. Barreto-Souza, A. L. de Morais,and G. M. Cordeiro, G. M. J. Stat. Comput. 81, 645-657 (2011).
4. A. Asgharzadeh, H. S. Bakouch and L. Esmaeili, J. Appl. Stat. 40, 1717-1734 (2013).
5. A. Asgharzadeh, S. Nadarajah and F. Sharafi, Revstat Stat. J. 16, 87-113 (2018).
6. H. Rinne, The Weibull Distribution: A Handbook (Chapman and Hall/CRC, Florida, 2008).
7. M. E. Ghitany, B. Atieh and S. Nadarajah, Math. Comput. Simul. 78, 493-506 (2008).
8. T. Fan and T. M. Hsu. IEEE Trans. Reliab. 64, 376-385 (2014).
