Received: 21 July 2016 Revised: 25 November 2016 Accepted: 7 April 2017 Heliyon 3 (2017) e00296
The Extended Erlang-Truncated Exponential distribution: Properties and application to rainfall data I.E. Okorie a,β , A.C. Akpanta b , J. Ohakwe c , D.C. Chikezie b a
School of Mathematics, University of Manchester, Manchester M13 9PL, UK
b
Department of Statistics, Abia State University, Uturu, Abia State, Nigeria
c
Department of Mathematics & Statistics, Faculty of Sciences, Federal University Otuoke, P.M.B 126 Yenagoa,
Bayelsa, Nigeria * Corresponding author. E-mail addresses:
[email protected] (I.E. Okorie),
[email protected] (A.C. Akpanta),
[email protected] (J. Ohakwe),
[email protected] (D.C. Chikezie).
Abstract The Erlang-Truncated Exponential πππ distribution is modiο¬ed and the new lifetime distribution is called the Extended Erlang-Truncated Exponential ππππ distribution. Some statistical and reliability properties of the new distribution are given and the method of maximum likelihood estimate was proposed for estimating the model parameters. The usefulness and ο¬exibility of the ππππ distribution was illustrated with an uncensored data set and its ο¬t was compared with that of the πππ and three other three-parameter distributions. Results based on the minimized log-likelihood Μ Akaike information criterion (AIC), Bayesian information criterion (BIC) and (βπ΅), the generalized CramΓ©rβvon Mises π β statistics shows that the ππππ distribution provides a more reasonable ο¬t than the one based on the other competing distributions. Keywords: Mathematics, Applied mathematics
http://dx.doi.org/10.1016/j.heliyon.2017.e00296 2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
1. Introduction Erlang-Truncated Exponential (πππ) distribution was originally introduced by El-Alosey [1] as an extension of the standard one parameter exponential distribution. The πππ distribution results from the mixture of Erlang distribution and the left truncated one-parameter exponential distribution. The cumulative distribution function (cdf) G(x), and probability density function (pdf) g(x) of the πππ distribution are given by; πΊ(π₯) = 1 β πβπ½(1βπ
βπ )π₯
; 0 β€ π₯ < β, π½, π > 0,
(1)
and π(π₯) = π½(1 β πβπ )πβπ½(1βπ
βπ )π₯
; 0 β€ π₯ < β, π½, π > 0,
(2)
respectively, where π½ is the shape parameter and π is the scale parameter. The πππ distribution collapses to the classical one-parameter exponential distribution with parameter π½ when π β β. Unfortunately, the πππ distribution share the same limitation of constant failure rate property with the exponential distribution which makes it unsuitable for modelling many complex lifetime data sets that have nonconstant failure rate characteristics. Generally speaking, research has shown that the standard probability distributions are largely inadequate for modelling complex lifetime data sets and various excellent ways of overcoming this shortcoming have been proposed in the literature; for instance: Beta exponential G distributions, due to Alzaatreh et al. [2]; Beta extended G distributions, due to Cordeiro et al. [3]; Beta G distributions, due to Eugene et al. [4]; Exponentiated exponential Poisson G distributions, due to RistiΔ and Nadarajah [5]; Exponentiated generalized G distributions, due to Cordeiro et al. [6]; MarshallβOlkin G distributions, due to Marshall and Olkin [7]; Transmuted family of distributions, due to Shaw and Buckley [8]; and so on. Mainly, by introducing extra shape parameter(s) to standard distribution a robust and more ο¬exible distribution is derived. For a comprehensive list of methods of generating new distributions readers are encouraged to see Nadarajah and Rocha [9], AL-Hussaini, Ahsanullah [10], Ali et al. [11], Cordeiro et al. [12], Alzaatreh et al. [13] and Pescim et al. [14]. To motivate our new distribution, we consider the time to failure of a component in series arrangement of a certain device, denoted by π where π1 , π2 , β― , ππΌ are independent and identically distributed πππ random variables. The device fails (stops functioning) if one of its component fails. Hence, the probability that the device will stop functioning before or exactly at a speciο¬ed time say x is given by;
2
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
P(max(π1 , π2 , β― , ππΌ ) β€ π₯) = =
πΌ β π=1 πΌ β π=1
P(ππ β€ π₯) πΉππ (π₯)
= (πΉππ (π₯))πΌ ,
(3)
this formulation ο¬ts exactly into the framework of Gupta and Kundu [15] (Exponentiated family of distributions). The new distribution is called the Extended Erlang-Truncated Exponential (ππππ) distribution. The ππππ distribution has a tractable pdf whose shape is either decreasing or unimodal. The failure rate function (frf ) is characterized by decreasing, constant and increasing shapes and the new three-parameter distribution demonstrates a superior ο¬t when compared with some other well-known threeparameter distributions, as we shall see later. Related works are: the Transmuted Erlang-Truncated Exponential distribution, due to Okorie et al. [16], MarshallβOlkin generalized Erlang-truncated exponential distribution, due to Okorie et al. [17] and the generalized Erlang-Truncated Exponential distribution, due to Nasiru et al. [18]. The remaining part of this paper is organized as follows. In Section 2, we present the closed form mathematical expression for the pdf and cdf of the new probability distribution ππππ and its statistical and reliability properties. In Section 3, the parameters of the ππππ distribution are estimated through the method of maximum likelihood estimation. In Section 4, we perform a Monte-Carlo simulation study to assess the stability of the maximum likelihood estimates of the parameters of the ππππ distribution. And we introduce a real data set, methods of model selection, application of the ππππ distribution to the data and the results are also presented. In Section 5, we present the discussion of results, and lastly, in Section 6, we give the concluding remarks.
2. Model The cdf F(x) and pdf f(x) of the ππππ distribution are given by; )πΌ ( βπ πΉ (π₯) = 1 β πβπ½(1βπ )π₯ ; 0 β€ π₯ < β, πΌ, π½, π > 0,
(4)
and π (π₯) = πΌπ½(1 β πβπ )πβπ½(1βπ
βπ )π₯
(
1 β πβπ½(1βπ
βπ )π₯
)πΌβ1
; 0 β€ π₯ < β, πΌ, π½, π > 0, (5)
where πΌ and π½ are the shape parameters and π is the scale parameter.
3
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
Figure 1. Possible shapes of the probability density function π (π₯) (left) and cumulative distribution function πΉ (π₯) (right) of the ππππ distribution for ο¬xed parameter values of π½ and π.
The ππππ distribution reduces to the πππ distribution when πΌ = 1. The plots of the cdf and pdf are shown in Figure 1.
2.1. Reliability characteristics The reliability function π
(π₯) is an important tool for characterizing life phenomenon. π
(π₯) is analytically expressed as π
(π₯) = 1 β πΉ (π₯). Under certain predeο¬ned conditions, the reliability function π
(π₯) gives the probability that a system will operate without failure until a speciο¬ed time π₯. The reliability function of the ππππ distribution is given by; )πΌ ( βπ π
(π₯) = 1 β 1 β πβπ½(1βπ )π₯ ; 0 β€ π₯ < β, πΌ, π½, π > 0.
(6)
Another important reliability characteristics is the failure rate function. The frf gives the probability of failure for a system that has survived up to time π₯. The frf h(x) is mathematically expressed as β(π₯) = π (π₯)βπ
(π₯). The frf of the ππππ distribution is given by; ( )πΌβ1 βπ βπ πΌπ½(1 β πβπ )πβπ½(1βπ )π₯ 1 β πβπ½(1βπ )π₯ β(π₯) = ; 0 β€ π₯ < β, πΌ, π½, π > 0. ( )πΌ 1 β 1 β πβπ½(1βπβπ )π₯ (7) The plots of the reliability function and frf are shown in Figure 2.
4
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
Figure 2. Possible shapes of the reliability function π
(π₯) (left) and failure rate function β(π₯) (right) of the ππππ distribution for ο¬xed parameter values of π½ and π.
2.2. Asymptotics and shapes In this section we present the asymptotics and shape characteristics of the pdf and frf of the ππππ distribution. The following asymptotic behaviors are observed β§ if πΌ < 1, βͺβ, βͺ π (0) = β(0) = β(β) = β¨π½[1 β πβπ ], if πΌ = 1; and π (β) = 0; β πΌ > 0. βͺ if πΌ > 1, βͺ0, β© Theorem 2.1. A random variable X with pdf f(x) is said to be unimodal if the pdf is log-concave, i.e.; log π β²β² (π₯) β€ 0. If X βΌ ππππ (πΌ , π½ , π) with pdf in Equation (5) then the distribution could take either of the two shapes β§ βͺdecreasing, lim π (π₯) = β¨ π₯ββ βͺunimodal, β©
if πΌ β€ 1, if πΌ > 1.
Proof. Taking the log of the pdf in Equation (5) and diο¬erentiating w.r.t x gives log π (π₯) = log(πΌ) + log(π½[1 β πβπ ]) β π½[1 β πβπ ]π₯ + (πΌ β 1) log(1 β πβπ½[1βπ log π β² (π₯) = βπ½[1 β πβπ ] +
5
(πΌ
βπ β 1)π½[1 β πβπ ]πβπ½[1βπ ]π₯
1 β πβπ½[1βπβπ ]π₯
.
βπ ]π₯
)
(8)
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
From Equation (8) it is clear that the pdf of the ππππ distribution is decreasing as π₯ β β when πΌ < 1 and π β² (π₯) has a single root, denoted by π₯0 when πΌ > 1 and the root is located at π₯0 =
log(πΌ) ; πΌ > 1, 2π½[1 β πβπ ]
(9)
this implies that there exist some π₯ < π₯0 such that π (π₯) is increasing and π₯ > π₯0 such that π (π₯) is decreasing, then π₯0 is said to be the critical point at which the pdf is maximized (i.e.; the mode). The second derivative log π β²β² (π₯) =
β(πΌ β 1)(π½[1 β πβπ ])2 πβπ½[1βπ
βπ ]π₯
(1 β πβπ½[1βπβπ ]π₯ )2
| | < 0| , | |πΌ>1
completes the proof. β‘ Theorem 2.2. If X follows the ππππ distribution with frf in Equation (7) then the shape of the frf h(x) are β§ βͺconstant(π½[1 β πβπ ]), if πΌ = 1, βͺ lim β(π₯) = β¨decreasing, if πΌ < 1, π₯ββ βͺ if πΌ > 1. βͺincreasing, β© Proof. The ο¬rst case is satisο¬ed when Equation (7) is evaluated at πΌ = 1. The ο¬rst derivative of the log of β(π₯) is log β(π₯) = log(πΌ) + log(π½) + log(1 β πβπ ) β π½[1 β πβπ ]π₯ + (πΌ β 1) log(1 β πβπ½[1βπ log ββ² (π₯) = βπ½[1 β πβπ ] + +
(πΌ
βπ ]π₯
) + πΌ log(1 β πβπ½[1βπ
βπ ]π₯
)
βπ β 1)π½[1 β πβπ ]πβπ½[1βπ ]π₯
1 β πβπ½[1βπβπ ]π₯
βπ πΌπ½[1 β πβπ ]πβπ½[1βπ ]π₯
1 β πβπ½[1βπβπ ]π₯
.
(10)
It is clear from Equation (10) that the failure rate of the ππππ distribution is decreasing when πΌ < 1 and increasing when πΌ > 1 and Equation (10) has no unique root. β‘
2.3. The πππ quantile function The ππ‘β quantile function of the ππππ distribution is given by; 1
β log(1 β π πΌ ) π₯π = ; π β (0, 1). π½(1 β πβπ )
(11)
Random samples from the ππππ distribution can be simulated through the inversion of cdf method by simply substituting π with a uniform π (0, 1) variates. It is easy to
6
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
obtain the median of the ππππ distribution by substituting π = 1β2 in Equation (11). For any given values of πΌ, π½ and π, the median of the ππππ distribution is given by; 1
π₯0.5
β log(1 β 0.5 πΌ ) = . π½(1 β πβπ )
2.4. The πππ crude moment Theorem 2.3. If the ππ‘β crude moment of any random variable π exists then other essential characteristics of the distribution could be calculated; for example the mean, variance, coeο¬cient of variation, skewness and kurtosis statistics. The ππ‘β crude moment of a continuous random variable π is deο¬ned by πΈ(π π ) = β ππβ² = β«ββ π₯π π (π₯)ππ₯. Hence, it follows that the ππ‘β crude moment of the ππππ distribution is ππβ² =
β β (β1)π+2π πΌπΌ!π! . [π½(1 β πβπ )]π π,π=0 π!(πΌ β π)!(π + π + 1)π+1
Proof. Using Equation (5) and the deο¬nition of ππβ² we have β
ππβ²
=
β«
π₯π πΌπ½(1 β πβπ )πβπ½(1βπ
βπ )π₯
(
1 β πβπ½(1βπ
βπ )π₯
)πΌβ1
ππ₯
0
β
= πΌπ½(1 β πβπ )
π₯π πβπ½(1βπ
β«
βπ )π₯
(
1 β πβπ½(1βπ
βπ )π₯
)πΌβ1
ππ₯,
(12)
0
substituting π¦ = π½(1 β πβπ )π₯ into Equation (12) we have β
ππβ²
πΌ = π¦π πβπ¦ (1 β πβπ¦ )πΌβ1 ππ¦, [π½(1 β πβπ )]π β«
(13)
0
the series expansion of (1 β π¦)πβ1 for |π¦| < 1 and π β πΉ+ (non-integer) could be expressed as (1 β π¦)πβ1 = (1 β π¦)π β
(1 β π¦)β1 ( ) β ( ) β β β π π π π 1+πβ1 = (β1) (β1) π¦ β
(β1)π π¦π π π π=0 π=0 ( ) β β β β π π+π = (β1)π+2π π¦ π π=0 π=0 = π!
β β π,π=0
7
(β1)π+2π
π¦π+π π!(π β π)!
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
β β
= Ξ(π + 1)
(β1)π+2π
π,π=0
π¦π+π . Ξ(π + 1) β
Ξ(π β π + 1)
If (π > 1) β πΉ+ (integer) then the sum terminates at π β 1; thus, the series representation of Equation (13) is given by; ππβ²
β ( ) β ( ) β β β πΌ πΌ π βπ¦ π βπ¦π π 1+πβ1 = π¦ π π (β1) (β1) (β1)π πβπ¦π ππ¦ π π [π½(1 β πβπ )]π β« π=0 π=0 0
( ) β β β πΌ πΌ (β1)π+2π π¦π πβπ¦(π+π+1) ππ¦ = π β« [π½(1 β πβπ )]π π,π=0
(14)
0
and by substituting π§ = π¦(π + π + 1) into Equation (14) we have ππβ²
( ) β β β (β1)π+2π πΌ πΌ = π§π πβπ§ ππ§ [π½(1 β πβπ )]π π,π=0 (π + π + 1)π+1 π β« 0
=
πΌπΌ!π! [π½(1 β πβπ )]π
β β
(β1)π+2π
π,π=0
π!(πΌ β π)!(π + π + 1)π+1
.
β‘
(15)
The mean is a very important characteristics of a distribution not only in statistics but also in reliability engineering as it is referred to as the mean time to failure (ππππ
). Under some predeο¬ned conditions ππππ
could be interpreted as the expected length of time a non-repairable item can operate before it fails. The mean of the ππππ distribution is given by; π1β² =
β β (β1)π+2π πΌπΌ! π½(1 β πβπ ) π,π=0 π!(πΌ β π)!(π + π + 1)2
and the variance π (π) could be obtained by β β (β1)π+2π 2πΌπΌ! π (π) = [π½(1 β πβπ )]2 π,π=0 π!(πΌ β π)!(π + π + 1)3 ( )2 β β (β1)π+2π πΌπΌ! β . π½(1 β πβπ ) π,π=0 π!(πΌ β π)!(π + π + 1)2
The coeο¬cient of variation (πΆπ ), skewness (πΎ1 ) and kurtosis (πΎ2 ) statistics of the ππππ distribution could be obtained by evaluating β β β² βπ πΆπ = β 2 β 1, π1β²2 πΎ1 =
π3β² β 3π2β² π1β² + 2π1β²3 3
(π2β² β π1β²2 ) 2
,
and
8
http://dx.doi.org/10.1016/j.heliyon.2017.e00296
2405-8440/Β© 2017 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
Article No~e00296
Table 1. The ο¬rst four order moments of 200 random samples from the ππππ distribution with selected parameter values. πΆ, π·, π β 3.5,5,0.2 0.01,2,3 0.45,5,10 0.001,0.5,20 3,12,8 0.5,0.5,0.5 28,7,1 0.8,1,50 4,0.45,.01 1,5,5
πβ² π 2.1692 0.0086 0.1132 0.0033 0.1528 3.1195 0.8875 0.8622 465.2816 0.2014
πβ² π 6.4053 0.0066 0.0395 0.0096 0.0328 28.0785 0.8699 1.6514 287494.7 0.0811
πβ² π 24.4149 0.0095 0.0225 0.0519 0.0093 408.9417 0.9458 4.8698 226080693 0.049
πβ² π 115.6766 0.0191 0.0176 0.3982 0.0033 8151.614 1.1448 19.3318 219160499642 0.0395
Table 2. The ο¬rst four order moments of 200 random samples from the πππ distribution with selected parameter values. π·, π β 5,0.2 2,3 5,10 0.5,20 12,8 0.5,0.5 7,1 1,50 0.45,0.01 5,5
πΎ2 =
πβ² π 1.1033 0.5262 0.2 2 0.0834 5.083 0.226 1 223.3352 0.2014
πβ² π 2.4347 0.5538 0.08 8 0.0139 51.6735 0.1021 2 99757.21 0.0811
πβ² π 8.0588 0.8742 0.048 48 0.0035 787.9679 0.0693 6 66837885 0.049
πβ² π 35.566 1.84 0.0384 384 0.0012 16020.93 0.0626 24 59709005539 0.0395
π4β² β 4π3β² π1β² + 6π2β² π1β²2 β 3π1β²4 (π2β² β π1β²2 )2
respectively. Comparing the results in Table 1 with those in Table 2 we observe that β π½ and π, ππβ² EETE = ππβ² ETE when πΌ = 1, ππβ² EETE < ππβ² ETE for all πΌ < 1, and β πΌ > 1 ππβ² EETE > ππβ² ETE . The following π code was used to produce the results in Table 1, and the results in Table 2 can be reproduced on slight modiο¬cation of this code. moments