A Comparative Study of Traditional Estimation Methods and Maximum Product Spacings Method in Generalized Inverted Exponential Distribution

Transcription

1 J. Stat. Appl. Pro. 3, No. 2, (2014) 153 Joural of Statistics Applicatios & Probability A Iteratioal Joural A Comparative Study of Traditioal Estimatio Methods ad Maximum Product Spacigs Method i Geeralized Iverted Expoetial Distributio Umesh Sigh, Sajay Kumar Sigh ad Rajwat Kumar Sigh Departmet of Statistics ad DST-CIMS. Baaras Hidu Uiversity, Varaasi , Idia Received: 22 Feb. 2014, Revised: 28 Apr. 2014, Accepted: 30 Apr Published olie: 1 Jul Abstract: I this paper, we propose the method of Maximum product of spacigs for poit estimatio of parameter of geeralized iverted expoetial distributio (GIED). The aim of this paper is to aalyse the small sample behaviour of proposed estimators. Further, we have also proposed asymptotic cofidece itervals of the parameters ad the estimates of reliability ad hazard fuctio usig Maximum Product Spacigs (MPS) method ad compared with correspodig asymptotic cofidece itervals ad the estimates of reliability ad hazard fuctio of Maximum Likelihood estimatio (MLEs). A comparative study amog the method of MLE, method of least square (LSE) ad the method of maximum product of spacigs (MPS) is performed o the basis of simulated sample of GIED. The MPS method outperforms the method of MLE ad the method of LSE. Furthermore, compariso of differet estimatio method have bee proposed o the basis of K-S distace ad AIC. For umerical illustratio oe real data set has bee cosidered. Keywords: GIED, Reliability characteristic, method of Maximum Product Spacigs, method of Maximum Likelihood Estimatio, method of Least Squares Estimates ad Iterval Estimatio. 1 Itroductio I statistical iferece problem, we are give a set of observatios x 1,x 2,,x. These are the values take by some radom pheomea about whose distributio we have some kowledge. For parameter estimatio, various estimatio methods are widely discussed i literature. Oe ofte uses traditioal estimatio methods such as the method of momets, method of least square, method of weighted least square ad maximum likelihood estimatio (MLE). Each of them havig their ow advatages ad limitatios but amog these methods the most popular method of estimatio is maximum likelihood estimatio method. Which ca be justified o the groud of its various useful properties like cosistecy, sufficiecy, ivariace ad asymptotic efficiecy ad its easy computatios. The MLE method works efficietly if each cotributio to the likelihood fuctio is bouded above. It is the situatio with all discrete distributios. However, havig such ice properties ad better applicability it also has some weakess as metioed by various authors i differet cotext. Its greatest weakess is that it ca ot work for heavy tailed cotiuous distributio with ukow locatio ad scale parameters (Pitma, 1979, p. 70). It also creates problem i situatios where there is oly mixture of cotiuous distributio ad the MLE method ca break dow. It was established by some authors that MLE does ot always provide precise estimates for certai distributios such as gamma, Weibull, ad log ormal distributios. I all these cases the critical difficulty is that there are paths i parameter space with locatio parameter teds to smallest observatio alog which the likelihood becomes ifiite. Ufortuately i such situatios estimates of other parameters becomes icosistet. Harter ad Moore [5] suggests a alterative way to use local maxima as a alterative of global maxima, this ca be effective but ot full proof there beig some weakess as poited out by Cheg for this see [1]. I the cotext of Harter ad Moore, Huzurbazar [19] has show that o statioary poit (ad hece o local maximum) ca provide a cosistet estimator, whe the cocer distributio is J-shaped, for example i the case of Correspodig author rajwat37@gmail.com

2 154 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... Weibull ad gamma distributios whe the shape parameter is less tha uity. Thus whether we cosider a global or a local maximum, Maximum Likelihood estimatio is boud to fail. The practical problem is that eve if the distributio is ot J-shaped, so that parameters ca i priciple be cosistetly estimated by local Maximum Likelihood estimatio as the sample size teds to ifiity, it ca happe that, with fixed sample size, a particular radom sample gives rise to a likelihood fuctio with o local maximum at all (Griffiths, [20]), this occur maily whe shape parameter equal to uity. Several authors has suggested alterative methods to MLE, either ivolvig modificatio to MLE method or method of momets or percetiles. Despite of the above problems whe MLE is applied it outperforms the alterative methods. I order to overcome these shortcomigs ad havig better applicability i such types of situatios which possess properties similar to MLE, Cheg ad Ami [1] itroduced the Maximum Product of Spacigs (MPS) method as a alterative to MLE for the estimatio of parameters of cotiuous uivariate distributios. Cheg ad Ami proposed to replace the likelihood fuctio by a product of spacigs ad cojectured that it retais most of the properties of the method of maximum likelihood. Raeby [3] idepedetly developed the same method as a approximatio to the Kullback-Leibler measure of iformatio. The approach of Cheg ad Ami is more ituitively attractive ad ca, to some extet, be regarded as a practical solutio to the problems liked with likelihood (Titterigto, [15]), but that of Raeby is more powerful theoretically ad allows the derivatio of the properties of MPS estimators. It may be oted that MPS method is especially suited to the cases where oe of the parameter has a ukow shifted origi, as it is the case i three parameter logormal, gamma ad Weibull distributios or to the distributios havig J-shape.. I order to make a geeral idea of advatages of MPS estimatio over MLE, we first list some good properties of MPS estimatio, which were showed by Cheg ad Ami [1], icludig sufficiecy, cosistecy ad asymptotic efficiecy. I certai cases, it is possible to obtai the distributioal behaviour of a MPS estimator for all sample sizes. Thus, for the uiform distributio with ukow edpoits, the MPS estimators are precisely the MVU estimators ad so their distributio is kow exactly solved by Cheg ad Ami [1]. The cosistecy of MPS estimators have bee discussed i detail by Cheg ad Ami [16]. I brief, asymptotically MPS are at least as efficiet as MLE estimators whe they exit. For distributio where the ed poits are ukow ad the desity is J-shaped the MLE is boud to fail, but MPS gives asymptotically efficiet estimators. MPS estimators will ot ecessarily be fuctio of sufficiet statistics i geeral. However, for the case whe the support of desity fuctios are kow, MPS estimator will show the same asymptotic properties as ML estimators icludig the oe of asymptotic sufficiecy. 1.1 The Model The radom variable X has a geeralized iverted expoetial distributio with two parameter α ad λ if it has a probability desity fuctio of the form: f(x)= ( )( αλ x 2 exp λ )[ ( 1 exp λ (α 1), x 0, λ,α > 0 (1) x x)] Where α is shape parameter ad λ is scale parameter, ad its CDF is give by [ ( F(x)=1 1 exp x)] λ α, α,λ > 0 (2) The model ca be cosidered as aother useful two-parameter geeralizatio of the Iverted expoetial distributio (IED). This lifetime distributio ca model various shapes of failure rates ad hece various shapes of ageig criteria. It is oted that the GIED is reduced to the IED for α = 1. I literature, estimatio of parameters i the two parameter GIED is discussed extesively, but o oe has performed compariso of MLE ad MPS. Readers are referred to the followig refereces: Abouammah ad Alshigiti [12], Gupta ad Kudu [14], Gupta ad Kudu [13]. Various properties of the GIED like reliability ad hazard fuctio, mea ad mode is discussed extesively by Abouammah ad Alshigiti [12]. I this paper, the method of product of spacigs is applied for estimatig the parameters i a two parameter GIED. The purpose here is to examie MPS estimates of the parameters of the GIED ad we also costruct 95% cofidece iterval usig MLE ad MPS. The method of product of spacigs is compared with the method of Least squares estimates (LSE) ad the method of MLE usig simulatio. MSE ad K-S distace are calculated ad o the basis of K-S distace through maximum product of spacigs method is better fitted tha MLE to the cosidered real data. AIC is

3 J. Stat. Appl. Pro. 3, No. 2, (2014) / calculated for MPS ad MLE ad both are compared. The mai objective of this paper is to aalyse the small sample behaviour of MPS. As we all kow that it is impossible to aalyse the whole data set due various reasos like cost factor, time factor etc. The orgaisatio of the paper is as follows: I sectio 2, Differet estimatio procedures are metioed ad estimates of reliability ad hazard fuctios usig MPS method is proposed ad compared with MLE. I sectio 3, asymptotic cofidece itervals of the parameters usig MPS method is proposed ad compared with MLE. I sectio 4, real data illustratio ad its applicatio is discussed, ad compariso of estimatio procedure based o K-S statistics is proposed. I sectio 5, a compariso is coducted usig simulatio study. Fially cocludig remarks are preseted i sectio 6. 2 Parameter estimatio For the cosidered distributio, we use two very kow ad popular method amely least squares method ad the maximum likelihood estimatio method ad oe which is ot very commo i.e MPS method for estimatig the parameters α ad λ. 2.1 Least square estimatio Let x 1 < x 2 < <x be ordered radom sample of ay distributio with CDF F(x), we get The least squares estimates are obtaied by miimizig Puttig the cdf of GIED i equatio (4) we get P(α,λ)= P(α,λ)= E(F(x i ))=i/() (3) (F(x i ) i/()) 2 (4) ( ( ( 1 1 exp λ )) α 2 i/()) (5) x i I order to miimize Equatio (5), we have to differetiate it with respect to λ ad α, which gives the followig equatio: )( )) α 1 ( ( )) α ) (α) exp ( λ 1 exp ( λ 1 1 exp ( λ xi xi xi i/() = 0 (6) x i i 1 i 1 [( ( ( 1 1 exp λ )) α )( ( i/() 1 exp λ )) α ( ))] l 1 exp ( λxi = 0 (7) x i x i The above Likelihood equatio caot be solved aalytically therefore we ca use ay iterative procedure such as Newto- Rapso method, to get the solutio. 2.2 Maximum likelihood estimators The likelihood fuctio for a sample of size from GIED (1) is give by: ( L(θ)=(α λ ( )exp λ (1/x i )) (1/x 2 i [(1 exp ) λ ))] α 1,t 0, α,λ > 0 (8) x i ad the log likelihood fuctio is give as M = ll(θ)=lλ + lα+ l ( 1/x 2 ) i λ l(1/x i )+(α 1) [ ) ] l (1 exp ( λxi ), (9)

4 156 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... After differetiatig the above equatio with respect to parameter α ad λ ad the equatig them to zero we got the ormal equatio as follows: [ )] (/α) + l 1 exp ( λxi = 0 (10) (/λ) (1/x)+(α 1) (1/x i )exp( λ/x i ) [1 exp( λ/x i )] = 0 (11) The above ormal equatios are ot i ice closed form, therefore we ca use ay iterative procedure such as Newto- Rapso method, to get the solutio. 2.3 Maximum Product of spacigs estimators Here, the method of maximum product of spacigs is described briefly as follows: Cosiderig a uivariate distributio F(x θ) with desity f(x θ) where it is required to estimate θ. The desity is assumed to be strictly positive i a iterval (α, β ) ad zero elsewhere, α ad β may also be elemets of θ, α= - ad β = are icluded. That is F(x θ)=0 ad f(x θ)=0 for x<α:, F(x θ)=1. ad f(x θ)=0 for x>β. Let x 1 < x 2...<x be a complete ordered sample, further defie x 0 = α, x = β. The spacigs are defied as follows: D 1 = F(x 1:,θ), D = 1 F(x :,θ), D i = F(x i:,θ) F(x i 1:,θ),i = 2,3,, as the spacigs of the sample. Clearly the spacigs sum to uity i.e D i = 1. The MPS method is to choose θ which maximizes the geometric mea of the spacigs i.e G = ( D i) 1/, or equivaletly, its logarithm S = log G. The mai aim for maximizig G (or S) is that the maximum, which is bouded above because of the coditio D i = 1, is foud oly whe all D is are equal. Cheg ad Ami [1] showed that maximizig S as a method of parameter estimatio is as efficiet as ML estimatio. Additioally, they showed that ties preset i data would ot be a matter of cocer i parameter estimatio. The CDF of the GIED is give by the equatio (2) ad the spacigs are defied as follows: D 1 = F(x 1 )=1 [ (1 exp( λ/x 1 )) α] (12) Ad the geeral term of spacigs is give by, D i = F(x i ) F(x (i 1) )= D () = 1 F(x )= [ (1 exp( λ/x )) α] (13) [( ) α ] [( ) α ] 1 exp λ/x (i 1) 1 exp λ/x i ) Such that D i = 1, MPS method choose θ which maximizes the product of spacigs or i other words to maximize the geometric mea of the spacigs i.e Takig the logarithm of G we get, Or we may write S as G= ( (14) ) 1/ D i (15) S=1/() ld i (16) { } S= 1 ld 1 + () ld i + ld i=2 { = 1 [ l 1 (1 e λ/x 1 ) α] [ + () l (1 e λ/x i 1 ) α (1 e λ/x i ) α]} i=2 + 1 { l [(1 e x) λ α]} () (17)

5 J. Stat. Appl. Pro. 3, No. 2, (2014) / After differetiatig the above equatio with respect to parameters ad the equatig them to zero we get the ormal equatio as follows: [ S α = 1 1 i=2 S λ = (1 e λ/x 1) α l(1 e λ/x 1) 1 (1 e λ/x 1 ) α l(1 e λ/x )) (1 e λ x i 1 ) α l(1 e λ x i 1 ) (1 e λ x i ) α l((1 e λ (α/x 1)((1 e λ x 1 ) α 1 )(e λ i=2 1 (1 e λ x 1 ) α (1 e λ x i 1 ) α (1 e λ x i ) α x 1 ) ] + x i )) =0 (α/x i 1)((1 e λ x i 1 ) α 1 )(e λ x i 1 ) (α/x i )((1 e λ x i ) α 1 )(e λ [ (α/x )((1 e λ x ) α 1 )(e λ (1 e λ x ) α x (1 e λ i 1 ) α (1 e λ x i ) α ] x ) The above ormal equatios caot be solved aalytically therefore we ca use Newto-Rapso method, i order to get the solutio. = 0 x i 1 ) (18) (19) 2.4 Reliability ad hazard fuctio I this sectio, we propose the estimatio of reliability ad hazard fuctio usig MPS for specified value of time say (t=4). Cheg ad Ami [1] ad Coole ad Newby [17] had metioed i their paper that MPS also shows the ivariace property just like MLE. So o this basis usig the ivariace property we estimate the reliability ad hazard fuctio. The MPS estimates of the reliability ad hazard fuctio is give as: ( ) ˆα ˆR MPS (t)= 1 e ˆλ t, α,λ,t > 0 (20) ( ) ˆα ˆλ e ˆλ t t Ĥ MPS (t)= (1 e 2 ), α,λ,t > 0 (21) ˆλ t respectively, where ˆα = ˆα mp ad ˆλ = ˆλ mp are the MPS estimates of the parameter α ad λ respectively. I equatio (20) ad (21) puttig the estimates of MLE, we ca get the expressio for the reliability ad hazard fuctio usig MLE. 3 Asymptotic cofidece itervals I this sectio, we propose the asymptotic cofidece itervals usig MPS, as it was metioed by Cheg ad Ami [1] ad Staislav Aatolyevi ad Grigory Koseok [18] i their papers that the MPS method also shows asymptotic properties like the Maximum likelihood estimator. Keepig this i mid, we may propose the asymptotic cofidece itervals usig MPS. The exact distributio of the MPS caot be obtaied explicitly. Therefore, the asymptotic properties of MPS ca be used to costruct the cofidece itervals for the parameters. I( ˆα, ˆλ) is the observed Fishers iformatio matrix ad is defie as: [ I( ˆα, ˆλ)= ] S αα S αλ S λ α S λ λ (α= ˆα mp,λ=ˆλ mp ) (22)

6 158 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... Fig. 1 Variatio of sample size with respect to MSE for differet choices of α ad for fixed value of λ = 1 Fig. 2: Mea Square Error of the estimates for α = 0.5,1 ad λ = 1 with variatio of sample size ()

7 J. Stat. Appl. Pro. 3, No. 2, (2014) / The first derivatives of the product of spacigs i.e the fuctio S with respect to parameter α ad λ are give by Equatios (18) ad (19) ad hece the secod derivatives are calculated as follows: The secod derivative of the fuctio S with respect to α is give as: S αα = [ F(x 1,α,λ)F αα (x 1,α,λ) F α (x 1,α,λ) 2 F(x 1,α,λ) 2 { } {F(x i,α,λ) F(x i 1,α,λ)} F αα(x i,α,λ) F αα(x i 1,α,λ) {F(x i,α,λ) F(x i 1,α,λ)} 2 { } F α(x i,α,λ) F α(x 2 i 1,α,λ) i=2 {F(x i,α,λ) F(x i 1,α,λ)} 2 { 1 {1 F(x,α,λ)}F αα(x,α,λ)+ F α(x,α,λ) {1 F(x,α,λ)} 2 ] } 2 (23) The secod derivative of the fuctio S with respect to λ is give as, S λ λ = 1 F(x 1,α,λ)F λ λ (x 1,α,λ) F λ (x 1,α,λ) 2 F(x 1,α,λ) {F(x i,α,λ) F(x i 1,α,λ)} i=2 {F(x i,α,λ) F(x i 1,α,λ)} 2 { } 2 F λ (x i,α,λ) F λ (x i 1,α,λ) 1 {F(x i,α,λ) F(x i 1,α,λ)} 2 { 1 {1 F(x,α,λ)} F λ λ (x,α,λ)+ F λ (x,α,λ) {1 F(x,α,λ)} 2 { F λ λ (x i,α,λ) F λ λ (x i 1,α,λ) ad the secod derivative of the fuctio S with respect to α,λ is give as: [ S αλ = S λ α = 1 F(x1,α,λ)F αλ (x 1,α,λ) F α (x 1,α,λ)F λ (x ] 1,α,λ) F(x 1,α,λ) 2 { } + 1 {F(x i,α,λ) F(x i 1,α,λ)} F αλ (x i,α,λ) F αλ (x i 1,α,λ) i=2 {F(x i,α,λ) F(x i 1,α,λ)} 2 { }{ } 1 F α (x i,α,λ) F α (x i 1,α,λ) F λ (x i,α,λ) F λ (x i 1,α,λ) {F(x i,α,λ) F(x i 1,α,λ)} 2 { }{ } 1 {1 F(x,α,λ)}F αλ (x,α,λ)+ F α (x,α,λ) F λ (x,α,λ) {1 F(x,α,λ)} 2 } 2 } (24) (25) Where F α(x,α,λ)= (1 e λ/x ) α l(1 e λ/x )

8 160 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... F αα (x,α,λ)=(1 e λ/x ) α (l(1 e λ/x )) 2 F λ (x,α,λ)= α x (1 e λ/x ) α 1 e λ/x F λ λ (x,α,λ)= α { α 1 (1 e λ/x ) α 2 (e λ/x ) 2 1 } x x x (1 e λ/x ) α 1 e λ/x F e λ/x λ α (x,α,λ)= (1 e λ/x ) α 1{ } α(α 1)l(1 e λ/x )+1 x The asymptotic cofidece itervals of the parameters of GIED usig MLE is already calculated by A. M.Abouammoh ad Arwa M. Alshigiti [12]. the first derivatives of the log likelihood fuctio of GIED usig MLE with respect to parameters are give by equatio (10) ad (11), ad the secod derivatives are as follows: ad the secod derivative with respect to α,λ is give as: (ll) αα = α 2 (26) e λ x i (ll) λ λ =(α 1) x 2 λ i (1 e x i ) λ 2 (27) (ll) αλ = e λ x i x i e λ x i itervals for the parameters α ad λ is give as, ˆα± γ β 2 V( ˆα) ad ˆλ ± γ β 2 (28) So o the basis of these derivatives, we obtai the iformatio matrix I(α,λ ). The approximate(1 β)100% cofidece V(ˆλ) respectively, where γ β is the upper 2 ( β 2 ) percetile of stadard ormal distributio, ˆα = ˆα mp ad ˆλ = ˆλ mp are the MPS estimates of the parameter α ad λ ad V( ˆα) ad V(ˆλ) are elemets of I 1 ( ˆα, ˆλ). 4 Real data illustratio The data set cosidered i this sectio for illustratio, cotais stregths of glass polished aeroplae widow. The use of this data set is described by Fuller et al. (1994) to predict the lifetime for a glass aeroplae widow. Sice the data were used i a study to predict failure times, such type of study is a form of reliability aalysis. The data are as follows: 18.83, 20.8, , 23.03, 23.23, 24.05, , 25.5, 25.52, 25.8, 26.69, 26.77, 26.78, 27.05, 27.67, 29.9, 31.11, 33.2, 33.73, 33.76, 33.89, 34.76, 35.75, 35.91, 36.98, 37.08, 37.09, 39.58, , 45.29, The above data set is already fitted to GIED ad the K-S statistics, betwee the fitted ad the empirical distributio is also calculated ad estimates of the parameter usig MLE method is calculated by Abouammoh ad Alshigiti [12]. MLE of the parameters of the GIED ˆα ad ˆλ are ad respectively ad for the same data set we have calculated the estimates of the parameter through MPS method ad the estimates are ˆα mp = ad ˆλ mp = respectively, ad correspodig K-S distace calculated usig estimates of MLE ad MPS respectively, ad it comes out ad respectively. So o the basis of estimates ad K-S statistics, for the cosidered data set MPS fits better as compared to MLE. The result of these distaces shows that MPS serve better tha MLE i this data set. As K-S statistics, is extesively used for differet model compariso ad is cosidered oe of the best way of model compariso, but o oe had paid attetio o compariso of differet estimatio procedure based o K-S statistics. I literature several authors have discussed K-S methodology for differet model compariso i terms of distaces for a give data set. Cosiderig the similar approach o the basis of the K-S statistics, here, we propose compariso of estimatio procedure or method as least K-S distace provides the better method of estimatio for give data set. For the above data set we otice that K-S distace through MPS is smaller tha K-S distace through MLE. I the support of the above propositio empirical cumulative distributio fuctio (ECDF) plot has bee give below for MPS ad MLE. AIC

9 J. Stat. Appl. Pro. 3, No. 2, (2014) / Fig. 3: Figure 2: Mea Square Error of the estimates for α = 1.5,2 ad λ = 1 with variatio of sample size () is also calculated usig both MLE ad MPS for this real data set ad it comes out smaller whe estimates were provided from MPS as compare to MLE. Sice MPS provides lesser K-S distace ad AIC, we may say that MPS serve better tha MLE for the cosidered data set. Thus we propose to estimatio method compariso o the basis of K-S statistic. AIC MLE = AIC MPS = Simulatio studies I order to compare the MPS, the MLE ad the LSE methods i parameter estimatio, a set of simulatio was doe based o GIED. We have geerated five thousad samples from GIED for differet parameters settigs. It may be metioed here that the exact expressio of MSE ca ot be obtaied because estimates are ot foud i ice close forms. It may be also oted here that MSE will deped o sample size, scale parameter λ ad shape parameter α respectively. I this study differet variatio of sample size() say (=20,40,60,80,100,120), shape parameter α say α( = 1,2,3,4,5,6,7) ad scale parameter λ say λ (=1,2,3,4,5,6,7) have bee cosidered. To study the effect of variatio of the sample size we have cosidered the simulated MSE for α ad λ. Firstly, sample size is varied i.e differet values of are take for fixed value of scale parameter λ take as 1 ad 2, for differet choices of shape parameter α as α(= 0.5,1,1.5,2,3,4). Secodly, we varied shape parameter α i.e differet values of α for two differet choices of sample sizes (=30 ad =50) for fixed values of scale parameter λ as λ ( = 1,2), thirdly, we varied the scale parameter λ i.e differet values of λ is take for two differet choices of sample sizes (=30 ad =50) for fixed values of shape parameter α as (α = 2,3), correspodig graphs are attached. For all the above cosidered choices graph of MSE is plotted ad attached.

10 162 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... Fig. 4: Mea Square Error of the estimates for α = 3,4 ad λ = 1 with variatio of sample size ()

11 J. Stat. Appl. Pro. 3, No. 2, (2014) / Fig. 5: Variatio of sample size with respect to MSE for differet choices of α ad λ = 2 We have chose sample size = 30 i secod ad third case because at = 30 chages are more visible i terms of MSE. We have also estimates the reliability ad hazard fuctios for α ad λ = 1,2 ad 3, for sample size = 20,30,50,80 cosiderig MLE ad MPS both, for this see Table 1, 2. Further we have costructed the cofidece iterval ad coverage probability for differet choices of α ad λ = 1,2 ad 3 ad for sample size =20,30,50,80 for this see table 2,3,4 ad 5. We have also compared the average legth of cofidece itervals of MPS with correspodig MLEs. R software is used i all computatios. O the basis of the results summarized i graphs ad table, some coclusios ca be draw which are stated as follows: 1: It is observed that as sample size icreases for fixed values of α ad λ the mea square error of the estimates decreases i all the three cosidered methods ad for large size of i.e after =80 all of them are early equivalet but MPS performs better tha other two cosidered method, see Figures 1 ad 2. It is also observed that as icreases MSE of reliability ad hazard fuctio follows the similar tred as stated above ad here also MPS perform better tha MLE (see Table 1). Furthermore, it is oticed that the average legth of cofidece iterval decreases as sample size icreases i both the cosidered case of MLE as well as MPS but the average legth is smaller i case of MPS as compared to MLE. The coverage probability obtaied here fairly attais the prescribed cofidece iterval. This is also true for small sample size, see Tables : From graph ad table, it is observed that as shape parameter α icreases for fixed sample size ad λ the MSE of the estimates of shape parameter α icreases for all the three cosidered cases but MSE of MPS is smaller tha LSE as well as MLE (see figure 3). It is oticed that for smaller values of shape parameter all of them are equally good but for larger value of shape parameter MPS is better tha other two. Geerally, shape parameters are difficult to estimate i most of the cases but here i case of GIED, oe should use MPS for all choices of λ ad. It is also observed that for fixed value of λ as the shape parameter α icreases MSE of reliability decreases but reverse tred is obtaied i the case

12 164 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... Fig. 6: Variatio of α for differet choices of sample size (=30,50) ad λ = 1,2 of hazard, it is true for MPS ad MLE both but here also MPS serve better tha MLE i terms of MSE. The average legth of cofidece iterval icreases as we icrease the shape parameter α i both the case of MPS ad MLE. Further more, it is also observed that the coverage probability obtaied here fairly attais the prescribed cofidece iterval. 3: It is observed that as scale parameter λ icreases for fixed sample size ad shape parameter α the MSE of the estimates of scale parameter λ icreases ad MPS performs better i terms of MSE i compariso to other two methods. It is also observed that for smaller value of λ all of them are equivalet but for larger value of λ MPS is far better tha LSE as well as MLE see figure 4. It is also observed that for fixed value of α as the scale parameter λ icreases MSE of reliability decreases but reverse tred is obtaied i the case of hazard i case of MLE ad MPS both see Table 1. The average legth of cofidece iterval icreases as we icrease the scale parameter λ for fixed value of shape parameter α similar tred has bee observed i both the case of MPS ad MLE (see Table 2-5). Further more, it is also observed that the coverage probability obtaied here fairly attais the prescribed cofidece iterval ad o ay specific tred has bee observed. 4: From the plot of ECDF ad the value of AIC, it is observed that estimates of MPS fits better tha the estimates of MLE. 6 Coclusio This paper ivolved the compariso of estimates obtaied by MPS, MLE, ad LSE method usig GIED. For smaller sample size it is advised to use MPS as it perform better tha LSE as well as MLE. I this paper we have cosidered the problem of poit estimatio, cofidece iterval ad it also itroduces compariso of differet estimatio procedure o

13 J. Stat. Appl. Pro. 3, No. 2, (2014) / Fig. 7: Variatio of λ for differet choices of sample size (=30,50) ad α = 2,3 the basis of K-S statistic, for this a real data has bee icorporated ad ECDF plot is also give. From the graphs, it is observed that all the estimates appears to be cosistet. The average legth of cofidece iterval usig MPS is smaller tha that of MLE. We have foud that MPS method outperforms the other two method with smaller mea square error. The fidigs of this paper will be very useful to researchers, statisticia ad egieers where such types of thigs were required ad also i cases where we have small sample size to aalyse ad exclusively where GIED is used.

14 Table 1: Average estimates ad correspodig MSEs of the reliability ad hazard fuctio usig MLE ad MPS respectively at specified time say(t=4) for differet parameter settigs ad sample size. Reliability ad hazard usig MLE,t=4 =20 =30 =50 =80 Para rt ht rt ht rt ht rt ht 1, , , , , , , , , MAXIMUM PRODUCT SPACINGS R(t) H(t) (MPS) 1, , , , , , , , , U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum...

15 J. Stat. Appl. Pro. 3, No. 2, (2014) / Table 2: Average estimates, coverage probability (i the brackets) ad correspodig cofidece itervals of the parameters α ad λ usig MPS for differet choices of parameter ad for sample sie =20 ad 30. CI AND COVERAGE PROBABILITY MPS =20 =30 Para α(cp) λ(cp) α(cp) λ(cp) 1, (0.9654) (0.964) (0.9636) (0.9548) (0.2395,1.4505) (0.1284,1.2788) (0.8975,1.8539) (0.5718,1.4983) 1, (0.966) (0.9628) (0.966) (0.953) (0.1571,1.4442) (0.1775,2.5652) (0.6109,1.5836) (1.8005,3.6929) 1, (0.9656) (0.9618) (0.965) (0.956) (0.4953,1.7380) (2.0397,5.5381) (0.0923,1.0677) (0.9450, ) 2, (0.9642) (0.9482) (0.9634) (0.9528) (0,2.8277) (0.1480,1.1370) (0.2349,2.4415) (0.5779, ) 2, (0.9714) (0.9566) (0.9648) (0.95) (0,3.0538) (0.6993,2.6476) (0.7629,3.0160) (1.0253,2.6279) 2, (0.966) (0.9596) (0.9666) (0.9484) (1.1893,4.1223) (1.6403,4.6007) (0.5983,2.8156) (0.5318,2.8675) 3, (0.9668) (0.9596) (0.9682) (0.95) (0.3334,5.4452) (0.7535,1.6584) (2.2181,5.9873) (0.6909,1.4533) 3, (0.963) (0.953) (0.967) (0.9474) (0,3.9963) (0.5667,2.3851) (0,3.6252) (0.9020,2.3533) 3, (0.963) (0.953) (0.9698) (0.9456) (0,4.9593) (1.7279,4.5179) (0.7889,4.5274) (1.5604,3.7566) Refereces [1] Cheg, R. C. H. ad Ami, N. A. K. (1983). Estimatig parameters i cotiuous uivariate distributios with a shifted origi. Joural of the Royal Statistical Society B 45, [2] D. L. Fitzgerald (1996). Maximum product of spacigs estimators for the geeralized Pareto ad log-logistic distributios. Stochastic Hydrology ad Hydraulics 10, Table 3: Average estimates, coverage probability (i the brackets) ad correspodig cofidece itervals of the parameters α ad λ usig MPS for differet choices of parameter ad for sample sie =50 ad 80. CI AND COVERAGE PROBABILITY MPS =50 =80 Para α(cp) λ(cp) α(cp) λ(cp) 1, (0.9572) ( ) (0.9518) (0.9450) (0.5423,1.2488) (0.3020,1.0117) (0.5543,1.1285) (0.7195,1.2810) 1, (0.9608) (0.9530) (0.9544) (0.9468) (0.7531,1.4718) (1.2761,2.7254) (0.4942,1.0670) (1.0908,2.2410) 1, (0.9624) (0.9470) (0.9534) (0.9446) (0.6232,1.3199) (2.1661,4.2704) (0.5486,1.1047) (1.8874,3.5803) 2, (0.9632) (0.9480) (0.9550) (0.9438) (0.9364,2.6279) (0.6742,1.2819) (0.7514,2.0445) (0.5774,1.0650) 2, (0.9616) (0.9434) (0.9498) (0.9440) (1.8147,3.4827) (1.9248,3.1392) (1.5378,2.8039) (1.6764,2.6076) 2, (0.9604) (0.9468) (1.9003,0.9538) (2.8732,0.9392) (1.0153,2.7006) (1.2994,3.1588) (1.4479,2.7118) (2.4640,3.8683) 3, (0.9636) (0.9434) (0.9540) (0.9426) (1.3852,4.0453) (0.5759,1.1409) (1.1708,3.2430) (0.5245,0.9668) 3, (0.9606) (0.9448) (0.9600) (0.9460) (0.8577,3.4972) (1.1978,2.3377) (1.8999,4.0433) (1.5154,2.4169) 3, (0.9628) (0.9382) (0.9556) (0.9428) (1.9603,4.6485) (2.0974,3.7866) (1.4215,3.5387) (2.1025,3.4448)

16 168 U. Sigh et. al. : A Comparative Study of Traditioal Estimatio Methods ad Maximum... Table 4: Average estimates, coverage probability (i the brackets) ad correspodig cofidece itervals of the parameters α ad λ usig MLE for differet choices of parameter ad for sample size =20 ad 30. CI AND COVERAGE PROBABILITY MLE =20 =30 Para α(cp) λ(cp) α(cp) λ(cp) 1, (0.9706) (0.9538) (0.9568) (0.9502) (0.4667,1.8570) (0.5008,1.7484) (0.5712,1.6290) (0.5865,1.5788) 1, (0.9724) (0.9476) (0.9622) (0.947) (0.4680,1.8740) (1.0174,3.5427) (0.5745,1.6408) (1.1859,3.1846) 1, (0.9642) (0.9556) (0.9598) (0.9396) (0.4661,1.8591) (1.5103,5.2748) (0.5741,1.6402) (1.7709,4.7559) 2, (0.965) (0.9418) (0.9666) (0.9484) (0.7223,4.1396) (0.5803,1.6307) (1.0060,3.5580) (0.6529,1.4897) 2, (0.9636) (0.9516) (0.9666) (0.943) (0.7196,4.1907) (1.1558,3.2442) (1.0009,3.5325) (1.3039,2.9780) 2, (0.9622) (0.9486) (0.9686) (0.9506) (0.7244,4.1698) (1.7454,4.8951) (1.0042,3.5456) (1.9530,4.4575) 3, (0.9678) (0.9526) (0.9616) (0.94) (0.8246,6.7882) (0.6092,1.5765) (1.3339,5.6402) (0.6762,1.4504) 3, (0.9658) (0.9486) (0.9654) (0.9526) (0.8276,6.7213) (1.2178,3.1552) (1.3316,5.6016) (1.3458,2.8888) 3, (0.964) (0.942) (0.962) (0.9498) (0.8252,6.7323) (1.8293,4.7372) (1.3351,5.6322) (2.0271,4.3473) Table 5: Average estimates, coverage probability (i the brackets) ad correspodig cofidece itervals of the parameters α ad λ usig MLE for differet choices of parameter ad for sample size =50 ad 80. CI AND COVERAGE PROBABILITY MLE =50 =80 Para α(cp) λ(cp) α(cp) λ(cp) 1, (0.9606) (0.9522) (0.9532) (0.9522) (0.6670,1.4436) (0.6708,1.4196) (0.7368,1.3371) (0.7397,1.3274) 1, (0.9602) (0.9482) (0.9524) (0.9458) (0.6675,1.4450) (1.3465,2.8486) (0.7356,1.3355) (1.4749,2.6470) 1, (0.9668) (0.9556) (0.9568) (0.9486) (0.6668,1.4429) (2.0217,4.2784) (0.7355,1.3346) (2.2057,3.9595) 2, (0.9636) (0.9494) (0.9534) (0.9486) (1.2431,3.0811) (0.7256,1.3617) (1.3882,2.7690) (0.7743,1.2698) 2, (0.9532) (0.9478) (0.9556) (0.957) (1.2354,3.0575) (1.439,2.7029) (1.3928,2.7792) (1.5506,2.5420) 2, (0.9582) (0.9446) (0.958) (0.9538) (1.2390,3.0698) (2.1725,4.0796) (1.3953,2.7856) (2.3286,3.8173) 3, (0.9626) (0.9486) (0.964) (0.953) (1.7397,4.7783) (0.7407,1.3280) (2.0087,4.3095) (0.7942,1.2555) 3, (0.9666) (0.9482) (0.9566) (0.9492) (1.7410,4.7812) (1.4922,2.6750) (2.0102,4.3160) (1.5849,2.5055) 3, (0.9602) (0.9462) (0.956) (0.9514) (1.7368,4.7663) (2.2213,3.9826) (2.0133,4.3237) (2.3809,3.7631) [3] Raeby, B., (1984). The Maximum Spacigs Method. A Estimatio Method Related to the Maximum Likelihood Method. Scad. J. Stat., 11, [4] Shah, A. ad Gokhale, D. V., (1993). O Maximum Product of Spacigs Estimatio for Burr XII Distributios. Commu. Stat. Simulat. Computat., 22(3), [5] Harter, H. L. ad Moore, A. H. (1965). Maximum likelihood estimatio of the parameters of Gamma ad Weibull populatios from complete ad from cesored samples. [6] Ghosh, S.R. Jammalamadaka (2001) A geeral estimatio method usig spacigs. Joural of Statistical Plaig ad Iferece 93.

17 J. Stat. Appl. Pro. 3, No. 2, (2014) / [7] Shao, Y. (2001). Cosistecy of the maximum product of spacigs method ad estimatio of a uimodal distributio. Statistica Siica 11, [8] Wog, T. S. T. ad Li, W. K. (2006). A ote o the estimatio of extreme value distributios usig maximum product of spacigs. IMS Lecture Notes Moograph Series 52, [9] Mezbahur Rahma ad Larry M. Pearso (2002): Estimatio i two-parameter Expoetial distributios. Joural of Stat. Computat. Simulat.,70(4), [10] Mezbahur Rahma ad Larry M. Pearso (2003): A ote o estimatig parameters i two-parameter Pareto distributios, Iteratioal Joural of Mathematical Educatio i Sciece ad Techology, 34:2, [11] Mezbahur Rahma, Larry M. Pearso ad Uros R Martiovic (2007): Method of product of spacigs i the two parameter gamma distributio, Joural of Statistical Research Bagladesh Vol. 41, [12] A.M. Abouammoh ad A.M. Alshigiti, Reliability estimatio of geeralized iverted expoetial distributio, J. Statist. Comput. Simul. 79(11) (2009), pp. 1301?1315. [13] R.D. Gupta ad D.Kudu, Geeralized expoetial distributio: Differet methods of estimatios, J. Statist. Comput. Simul. 69(4) (2001), pp. 315?338. [14] R.D. Gupta ad D. Kudu, Geeralized expoetial distributio, Aust. N. Z. J. Statist. 41(2) (1999). [15] Titterigto, D.M. (1985) Commet o Estimatig parameters i cotiuous uivariate distributios. Joural of the Royal Statistical Society Series B 47. [16] Cheg R.C.H.; Ami, N.A.K. 1979: Maximum product-of-spacigs estimatio with applicatios to the logormal distributio, Uiversity of Wales IST, Math Report79-1. [17] F.P.A Coole ad M.J Newby, A ote o the use of the product of spacigs i bayesia iferece. [18] Aatolyev Staislav ad Koseok Grigory, A alterative to maximum likelihood based o spacigs. [19] Huzurbazar, V. S. (1948) The likelihood equatio, cosistecy ad the maxima of the likelihood fuctio. A. Eugeics, 14, [20] Griffiths, D. A. (1980) Iterval estimatio for the three-parameter logormal distributio via the likelihood fuctio. Appl. Statist., 29 (1),