Jourmal of Multivariate Analysis 143 (2016) 194-207 ron Contents lists available at ScienceDirect "ee A Journal of Multivariate Analysis ELSEVIER journal homepage: www.elsevier.com tocate;jmva Multivariate families of gamma-generated distributions with Qn finite or infinite support above or below the diagonal Narayanaswamy Balakrishnan®, Miroslav M. Ristié’* * MeMaster University, Halton, Ontario, L8S 4K, Canada "University of Ns, Faculty of Sciences and Mathematics, Serbia ARTICLE INFO ABSTRACT Artic stor In this paper, we introduce two new families of multivariate distributions with finite Received 29 july 2014 or infinite support above or below the diagonal generated by McKay's bivariate gamma ‘Availabe online 28 September 2015 istribution and show tat their conditional distributions are univariate gamma- and —___peta-generated distributions. We derive the Shannon entropies of the introduced families ‘Ans 2000 subject assfcations: of bivariate distributions We then focus on the special cases of bivariate gamma- 02899 exponentiated exponential distributions, and discuss their properties. Finally, we illustrate Keynords the usefulness ofthe proposed bivariate gamma-exponentiated exponential distributions ‘Mokiy’s bivariate gamma distribution with areal dataset Gamma-generated disibution (© 2015 Elsevier Inc. ll rights reserved, Beta-generateddistrbution Shannon entropy 1. Introduction We introduce here two families of multivariate distributions with finite or infinite support on a < y = b(support above the diagonal) and a < y < x < D (support below the diagonal), respectively, where @ and b can be finite or infinite ‘numbers. Let F be the curnulative distribution function (ed), f be the corresponding probability density function (pdf), and F be the survival function given by F(x) = 1—F(x). The first family of multivariate distributions is given by the joint density 2) = (1 ri) (-logFex))" 7 (v8 a wheter > 1.x = (1.4%. -...8))@ < Xp < ++ Ofori — 1, 2,...,7. The second family of multivariate distributions is given by the joint density Fc) (1 ro) where r > 1, = (X), X95 -++5%7)o0 < Xp < +++ Oforall i € {1, 2, ...,r). These families of distributions have two or more additional parameters that control shapes and skewness properties of the introduced distributions. ‘These families of distributions can be motivated as follows. First motivation is similar to the motivation of Jones and Larsen [11] who considered a possible application of the multivariate families of distributions defined above the diagonal. wart Foy Fo log Fox,))** T(r) Tr. @) * Corresponding author E-mail addresses: balaememasterc (N- Balakrishnan), mivistc72@gmailcom (MM. Risto bups/dxdokorg/10.1016)4jmva.2015.09012 (0047-259X) 2915 Elsevier Inc.All rights reserve.1. Bolakrishnon, MLM. Rist ourna of Multivariate Analysis 143 (2016) 194-207 195, In the case of bivariate distributions defined above the diagonal, Jones and Larsen [11] have noted that one natural scenario to obtain X; < ¥j, forall = 1,2, ....1is to consider 2: as a minimum and Y as a maximum of some quantity. They refer to 10} in which an example with minimum and maximum temperatures is given. Second motivation follows from the joint distribution of the upper and lower record values, respectively. Let Xyy1, Xuinys «++ ANG Xiciys «+» Xyme «+» be the upper and lower record values, respectively, from a sequence of i.i.d. continuous random variables with parent distribution F. Then, the joint pdf of the upper record values Xu)» » Xugm) iS given by (1) with 6; = my, 9; = mj — my, fori . 3,...,7, and m; < mis; fori = 1,2,...,1r — 1. Similarly, the joint pdf of the lower record Values Xin. +» in fs given by (2) with 8, — my =m — my, for! = 2.3,...,r,andm, = my, for 1.2....6- 1 ‘Some other multivariate and bivariate distributions with support above the diagonal have also been studied in the past. MeKay [15] introduced a bivariate gamma distribution, while Mihram and Hultquist [16] generalized this distribution and introduced the Beta-Stacy distribution, Two more forms of multivariate gamma distributions have been introduced by Mathai and Moschopoutos | 14] and Furman (7]. Jones and Larsen [ | introduced a family of multivariate distributions and used it for modelling ordered multivariate data. ‘The rest of this paper is organized as follows. In Section 2, we provide some connections between the introduced families of bivariate distributions and McKay's bivariate gamma distribution. The conditional pds of X, given ¥ = y, and of Y, given X = x, are derived. The Shannon entropies are derived in Section 3. In Section 4, the special case of bivariate gamma- exponentiated exponential distribution is considered and its properties are discussed. In Section 5, we discuss the estimation of the model parameters of bivariate gamma-exponentiated exponential distribution through maximum likelihood method. An illustration with a real dataset is finally made in Section 6. 2. Bivariate families and their properties Let us now consider the fariate case of the fami given in (1) and (2), First bivariate family of distributions is given by the joint pdf of the form 1 Foy" (tog?) Lever Ak, Y) = ————— (— log F ee Fa 7 3) Be AKY) Fore ¢ log F(x) ( nn) Foo 3) where a < x < y < b,d > Oand v > O. Direct integration yields the marginal pdf of the random variable X as 1 sey = (-logFe , d Oand v > 0. We can show that the corresponding marginal pdfs of random variables X and Y are, respectively, given by SVK) = 6) (= log Fo) (-1 logFo)* fw), a > 36.) \iogFO) and the conditional pdf of Y, given. x, is given by 1 Fy)" (ref) | d ( 1) a (— tog tH) | (tog £0) , [losFy(X)] = —32.1. Bolakrishnon, MLM. Rist ourna of Multivariate Analysis 143 (2016) 194-207 199) Which implies that 6; = 52. Here, E,(-) denotes the expectation with respect to the marginal distribution Gy s. Similarly, from Part (c) of Lemma 1, we have v1 = vz. Thus, the bivariate family of distributions in (3) is identifiable, ‘The marginal distribution of the random variable Y, however, depends on the parameter 5 + v which reveals its non- identifiability with respect to these parameters. Following the procedure based on maximum entropy introduced in [21], we can introduce a procedure for the selection of parent distribution among two parent distributions. Let Fy and fF; be two parent distributions with corresponding pdfs fi and fy, and parameters (1... 6) and @,...,6(?), respectively, Let (X, ¥) be the random vector from the bivariate distribution in (3) with parameters 8 and and parent distribution F. We consider the statistic D(F,, F») Hol F) — Hey( Fy). According to Corollary 2, we obtain aX) fa) eles] [lee ‘The statistic 10(F,, F2) can be estimated in the following manner, Suppose we have a bivariate sample ((x1.1)+ (nse). Then, an estimate ofthe statistic (Fy, Fa) is given by L205) 4 Vy OD log tog 20. leah teeny We may then approximate the distribution ofthis statistic and obtain its critical values by using the parametric bootstrap ‘method inthe following wat 1. The null hypothesis is thatthe random vector (X, Y) is from the bivariate distribution in (3) with parameters 6 and v and parent distribution Fy Since in most cases we do not know the true values f the parameters 8 and v, we need to estimate them too. Based on the sample ((x),yi),---» (,Yq)), We estimate by the maximum likelihood method the unknown parameters 0, ..,6{”, 3 and v, and denote them by”, 0), 3, and fy, 2. The alternative hypothesis is that the random vector (X, ¥) is from the bivariate distribution in (3) with parameters 5 and v and parent distribution Fy. According to this, we derive the maximum likelihood estimates of the parameters 6,” 9? 8 and v, and denote them as... 6). B, and. 3. We then obtain the value of the statistic 5(F;, F,) for the obtained estimates. 4. Now, we have two sets of estimates of the parameters 3 and v, but according tothe assumption, they should be the same. So, we can use their variances 5?, and S2,,i € {1,2}, giving more weight for the estimate with smaller variance and less ‘weight forthe estimate with larger variance, according tothe following equations: Shade bi 5 DF. Fa Sh) 4a UT St Sia ‘This way, we will have only one estimate for 8 adv 5. Now, we simulate N (for: example, 10,000) samples from the bivariate distribution in (3) with parent distribution F, and trae vaes ofthe parameters 4)". 4,3 and ¥. For each sample, we obtain the estimates as in Steps 1 and 2 and derive the value of statistic D\(F,, Fa), where i takes values 1, 2, .....N. 6, Finally, we accept the null hypothesis if Diwa/2)(Fi. F2) < DiFi.F:) < Diwr-o/ay(Frs Fa), where Dy (Fs, F2) is the jth order statistic from D,(F), Fa)...» Dy(Fi, Fa), and a is the level of significance. A similar procedure can be provided for the second bivariate family of distributions as well. Now we provide two examples to show how the proposed discrimination procedure works. Example 1. The null hypothesis is that the parent distribution F; is the exponential distribution with the scale parameter while the alternative hypothesis is that the parent distribution F; is the exponentiated exponential distribution with scale parameter # and shape parameter j1. We simulated a sample of size n = 200 from the bivariate distribution in (3) with the exponential parent distribution with true values of the parameters 2 = 0.5,5 = 2 and v = 3. We obtained the following results: 1. According tothe nul hypothesis, we obtained the maximum. likelihood estimates of the parameters as T = 0.5652, 2.0625 and 9; .2955. The variances of the estimates 5; and 7; are 0.0221 and 0.0574, respectively. 2. According to the alternative hypothesis, we obtained the maximum likelihood estimates of the parameters as a 0.5181, fi = 1.6412, 62 = 1.4596 and 2 = 2.9637. The variances of the estimates 42 and 2 are 0.1656 and 0.1123, respectively. 3. The estimated value of the statistic D(F,, Fy) is 1.1805. 4, According to Steps 1 and 2, we obtained the pooled estimates as 5 = 1.9915 and T = 3.1833.200 IN. Baakrshan, MLM. Rs/ Journal of Multivariate Analysis 143 (2016) 194-207 5. Then, we simulated N = 10,000 samples of size n = 200 from the bivariate distribution in (3) with the exponential parent distribution and true values of the parameters i = 0.5652, 5 = 1.9915 and = 3.1833. For each sample, we obtained the estimates as in Steps 1 and 2 and then obtained the value of statistic (Fy, F2), where i takes values 1, 2, N. 6. Forlevel of significance a = 0.05, we obtained Diasa)(Fr. F) = —2.1295 and Denso) (F1, Fz) = 1.5817. Since —2.1295 ~ 1.1805 < 1.5817, we accept the null hypothesis that the parent distribution is the exponential distribution. Example 2. In this case, the null and alternative hypotheses are the same as in the previous example, while we simulated a sample of size n = 200 from the bivariate distribution in (3) with the exponentiated exponential parent distribution with true values of the parameters @ = 0.5, 1 = 10, 8 = 2and v = 3. We obtained the following results: 1, According to the null hypothesis, we obtained the maximum likelihood estimates of the parameters as 7 = 0.6975, 5; = 6.1544 and ¥ = 4.0238, The variances of the estimates 5, and 7; are 0.1954 and 0.0831, respectively. 2, According to the alternative hypothesis, we obtained the maximum likelihood estimates of the parameters as 0.5218, ft = 9.7387, 5, = 2.3296 and f, = 3.0267. The variances of the estimates 5, and 9 are 0.4297 and 0.1181, respectively. 3, The estimated value of the statistic B(F), F,) is 7.7971 4, According to Steps 1 and 2, we obtained the pooled estimates as 5 = 4.9588 and 7 = 3.6120. 5. Then, we simulated N = 10,000 samples of size = 200 from the bivariate distribution in (3) with the exponential parent distribution and true values of the parameters = 0.6975, } = 4.9588 and ¥ = 3.6120, 6, For level of significance @ = 0.05, we obtained Daso)(Fi,F:) = —4.1033 and Dyso(Fi. 2) = 5.9320. Since 7.7971 > 5.9320, we reject the null hypothesis that the parent distribution is the exponential distribution and accept the alternative hypothesis that the parent distribution is exponentiated exponential distribution. 4, Bivariate gamma-exponentiated exponential distributions In this section, we focus on the bivariate distributions for given exponentiated exponential parent distribution F(x) (1—e*)!",x,3, 4 > 08), and derive some oftheir properties. From (3), we obtain the pdf as PE)PW)A—A-e*)*) PvE x (to =e) » O k/ The second expectation easily follows from (12) upon replacing 5 by 5 + v. Thus, the Shannon entropy of the bivariate gamma-exponentiated exponential distribution in (11) is given by Hey(F) ee ees = 1)Y(v) = 5 = 2loga — 210g 4 Lawl tar] Led (an tan wake If we now consider the families of distributions in (5), then we obtain the pdf as D2yShe MINI — erry TOPO —e*) —log(t Ber (y) O9 Substituting F(x) = (1 — e *) and F(y) = (1—e ution u = a-+ log (1 = (1 —e-*)*), we obtain a=)" ea where (b); = b(b + 1)...(b-+j — 1) isthe ascending factorial with (b)p = 1, we obtain bw a) = SO Pita wa ahd BG.v) & Now, by using the series expansion (1 +2) [ we obtain EW =y) =F Seo(“)a eA (sd 0s Ina similar manner, we can show the conditional mgf of Y. given X = x, to be SOK S y! (“) ayy! = 1-a-e*)" eka Cr)! ) Using recurrence relations for the derivatives of the ascending factorial EUS AR (R=1) 95 my km) s x( ; ew 6+)-v™), where W/() isthe digamma function, the conditional moments of X, given ¥ 1 Pe) Spi (kl : Ew ==> | dew ih Je =e! Bit dsr Hog(1— =e"), p=, ahi (0:8 +0: and the conditional moments of Y, given X are given by slay Sv! (“) i 1-a-e%" a a bee aes Cr) In particular, the conditional means are given by EM =y) = oe Ee u(r “J-a- "oF: (vs 8 + v: Hog (1-1 =e"). evix=9= Sort Sane (“')e-a-e9%y. The conditional mgfs can also be used to derive the joint mgf of random variables X and ¥ as follows: £(e**) = £(e%E(@"1X)) wa Liar (M3 (,)coreteta-e Ayu] m= ym]. we obtain re, BG })21,b > 0, and Eq, (2.3.6.1) of Prudnikov et al. [17] stating that a 1B, BFW; 3+ 9 a), Hog(1—(—e)*)). E(e"X =x) are given by E (VK = Letting u = —log(1~ (1 -e)*) intheexpresion of [exa = 1 ( lero ( Using the series expansion for z nen) EE) eye da 1 =e YA], we then obtain which gives the joint mgf of the random variables X and ¥. Finally, by differentiating the mgf with respect tos (p times) and £(q times) and setting s = 0 and ¢ = 0, we obtain expressions for the product moments of random variables X and Y.1. Bolakrishnon, MLM. Rist ourna of Multivariate Analysis 143 (2016) 194-207 203 tablet ean stand evan the AL forthe vane -amaexponetated Ssoonenta dst Tews 0 fi Ba fa 5 omasoa —os4aa2i6) — ososou) wo iez02se)—0s270430) —ostanoes) 200 tbigda0)bsigaios) —osnxo0e) So Ubmsiouas) —datsoues) —O50i(0um) von aioivo) soso) tsima07) 50 13380933) osaga307) —4a100.81 vio Tigress) babrozs) 48500362) 200 1115(0.386) 040.162) 49530250) 500 1#2(0229) os01(0.105)4934(0.165) 1000 1025/0151) o4Peo075) —4.9s9(0.11°) 50 10.430(4035) 05470219) 0510(0.093) too 1026312855) o.524(0.149) —01505(0066) 200 102129422002) 0512(0.103) 05030047) 500 10029(1.262} 0500(0085) 050210030) 1000 10.008(0'00) 050%0.048) —0501(0022) 50 650%4177) 1.928(0551) —_1.922(0.195) 100 546862305) 186110391) 1.991(0.134) 20 © 53191433) 1921(0275) —_L9s(0.094) 500 5.124089) 1880.175) 1 998(0.058) 1000 50571006) 188710125) 19490082) 5, Estimation of model parameters In this section, we discuss the maximum likelihood estimation of the parameters of the bivariate gamma-exponentiated exponential distribution in (11). The log-likelihood function log for a sample (x,y), .» (e.Ya) from the bivariate gamma-exponentiated exponential distribution in (11), is log = 2n log A + 2m log 1 —mlog 1°(s) —nlog Mw) — Yow ty +a —ve(1—™) i » Deere = (1-2 §)")) +0 » Yes (-%« =A =) Doe ( = (= ‘The maximum likelihood estimates of the parameters 2. y4,5, and can be obtained by numerical maximization. ‘To evaluate the performances of ML estimates, we simulated 10,000 samples of size n = 1000 for different values of +0 += DY tog (1%) YY), O, = 0.6022(0.0021), 5, = 3.6669(0.4518), and 7 = 7.8966(0.4654). The pooled estimates are 5 = 0.7520 and ¥' = 0.6350. The value of the test statistic is D(F,,F:) = —7.5338, while the critical values are Daso)(Fi. 2) = —0.1813 and Desoy(F;. F2) = 0.0838. Thus, we accept the alternative hypothesis that the parent distribution is the exponentiated exponential distribution. Based on the above results of discrimination procedures, we use the exponentiated exponential distribution asthe parent distribution to compare the bivariate gamma-exponentiated exponential distribution in (11) and the bivariate gamma- exponentiated exponential distribution in (13) with the bivariate beta-exponentiated exponential distribution of Jones and Larsen 11] with paf wrt b+y) Terra) x(a eye yy, defined for 0 < x < y < coanda > 0,f > 0,y > 0,8 > Oand p> 0. In the case of bivariate gamma-exponentiated exponential distribution in (13), we switch the sample ((X, ¥)} tothe sample ((¥;, X)} to providea sample satisfying Xi < Ys for alli = 1,2,....1. We now follow the approach used by Balakrishna and Shiji [2] First, we derive the maximum likelihood estimates of the parameters ofthe considered distributions. For the bivariate gamma-exponentiated exponential distribution in (11), we obtain the maximum likelihood estimates as 6 = 0.1674, 7t = 12948.15, 8 = 4.2571, and = 7.0697. Now, we use the obtained estimates to derive the 95% and 99% bootstrap confidence intervals for the parameters 4, 1,3 and v, respectively. We simulated 10,000 samples of size 196 from the bivariate gamma-exponentiated exponential distribution in (11) with true values of the parameters taken to be @ = 0.1674, i_= 12948.15, 5 = 4.2571, and 7 = 7.0697. For each obtained sample, we obtained the maximum likelihood estimates 9°, 7,3, and 1, where i € {1,2,..., 10,000}, and we used the true values of estimates as starting values for the maximum likelihood estimation. For the 95% bootstrap confidence intervals, we took the 250th and 9750th ordered estimates, while for the 99% bootstrap confidence intervals we took the ‘50th and 9950th ordered estimates. In this manner, we obtained the 95% bootstrap confidence intervals for parameters 8,1, 8 and v as [0. 1509, 0.1890], [12946.40, 14243,92], (2.8388, 6.0644] and [6.3015, 8.0205", respectively. The corresponding 99% bootstrap confidence intervals are, respectively, [0.1459, 0.1979], [12944.62, 14245.15], (2.4128, 6.8272] and (6.0348, 8.4009], For the bivariate gamma-exponentiated exponential distribution in (13), we obtained the maximum likelihood estimates as @ = 0.0281, j@ = 103.0442, 5 = 36669, and 7 = 7.8966, The 95% bootstrap confidence intervals for the parameters are [0.0227, 0.0338]. [78.6934, 144.4509], [2.5100, 5.6197] and [6.6835, 9.5939], while the 99% bootstrap confidence intervals are [0.0210, 0.0355], [72.5924, 161.3087], [2.2430, 6.6114] and [6.3581, 10.2016]. Finally, the maximum likelihood estimates of the parameters of the bivariate beta-exponentiated exponential distribution are obtained as @ = 0.0285, 7? = 0.5928, @ = 169.0745, 8 = 7.6471, and 7 = 3.5831, The 95% bootstrap confidence intervals for the parameters #, j2, a, and y are obtained as [0.0229, 0.0343}, [0.4460, 0.8315], [161.4783, 180.4587], {6.4874, 9.3055] and (2.4541, 5.6365]. respectively. Finally, the 99% bootstrap confidence interval for the parameters of this bivariate distribution are [0.02 12, 0.0363}, [0.407, 0.9482], 144.8345, 197.5025], 6.1835, 9.8740] and [2.1726, 6.5145], We now use the Kolmogorov-Smirnov test to check if the marginals of these bivariate distributions follow gamma- exponentiated exponential and beta-exponentiated exponential distributions, respectively. Since we have estimated parameters, we need to adjust the Kolmogorov-Smirnov test in order to account for the estimated parameters. We adjust the test by using the bootstrap approach as considered in Stute et al. [19]. and described in detail by Castro [5]. First, we obtain the value of the Kolmogorov-Smimnov statistic Dy = max, lFy(X,) — F{(X)] for the random variable X based on the sample (x), %2,...,X1pg) and the value of the Kolmogorov-Smirnov statistic D, = maxy, |Fy(y?) ~ F70%)| for the random variable Y based on the sample (yi. Jz...» Jaye). The function Fy is the empirical cumulative distribution function and Bg and Fy are the cumulative distribution functions whose true parameters are the estimated parameters , 7,6: and @. 7, 3545 respectively obtained as maximum ikelinood estimates based on sample data (x1. 42,» 96) aN (Yas Yine) ease: A191 — POH — My1. Bolakrishnon, MLM. Rist ourna of Multivariate Analysis 143 (2016) 194-207 205 Fitted CDF for (11) Fitted CDF for (11), so 60 70 a0 00H a ©6100 12040 tt Fitted COF for (13) Fitted COF for (19) 80 100-120, 140180180 5 60 78D 8100110 Fitted CDF for beta family Fitted CDF for beta family so 60 70) 8] 100 tH0 8 © 10012480180 Fig. 2. ‘The empirical marginal cumulative éisibution functions along withthe fitted marginal éistibuton functions Let figure ae forthe random Variable X, whi ce right figures are forte anéom variable. respectively. Thus, for the bivariate gamma-exponentiated exponential distribution in (11) and the estimates @ = 0.1674, Fi = 12948.15, 5 = 4.2571, and 1 = 7.0697, we obtained d, = 0.0654 and dy = 0.0518. Now, we derive the p- values for these statistics. For this purpose, we simulated 10,000 samples (x, x3... Xjgg) aNd (YY. «++ Higa) OF size 196 {from the marginals of the bivariate gamma-exponentiated exponential distribution in (1) with true values of parameters = 0.1674, = 12948.15, 5 = 4.2571, and 7 = 7.0697. For each simulated sample, we obtained the maximum likelihood estimates #, 7,5 and 7 and then obtained the values of the Kolmogorov-Smirnov statistics D,j = max [F,(%)) ~ F(s)| and Dy, = maxy, |F/(ys) — Fy(DhoJ € {1. 2, .-, 10,000), where F; is the empirical cumulative distribution function based con the simulated sample and Fy and Fy are the cumulative distribution functions whose true parameters are the estimated soa) parameters @, ji, 3; and 8, ji,5 +1, respectively. Finally, the p-values were calculated as p forx and y samples, respectively. In this case, we obtained the p-values as 0.0215 and 0.246 1, respectively. For the bivanate gamma-exponentiated exponential distribution in (13), we obtained the corresponding values as 0.0646 (0.0122? and 0.0300 (048693), respectively. Finally, for the bivariate beta-exponentiated exponential distribution, we obtained 0.0268 (0.9446) and 0.0645 (0.0112), respectively. The empirical marginal cumulative distribution functions along with the fitted marginal distribution functions are presented in Fig. 2, the Q-Q plots for the marginal distributions are given in rig. 3, and the corresponding P-P plots are presented in Fig. 4. Now, we can compare the models with the use of AIC and BIC values. For the bivariate gamma-exponentiated exponential distribution in (1), we obtain AIC = 3123.208 and BIC = 3136.320. The corresponding values for the bivariate gamma-exponentiated exponential distribution given in (13) are AIC = 3098.46 and BIC = 3111.574, Finally, for the bivariate beta-exponentiated exponential distribution, we obtain AIC = 3101.292 and BIC = 3117.683, From these results, we can conclude thatthe best fit for the diastolic and systolic blood pressure data of women is obtained by the bivariate gamma-exponentiated exponential distribution in ( 13). Note that swapping the order of observations produces different fits for the bivariate gamma-exponentiated exponential distributions and their marginals. ‘Since we are not swapping the order of the parameters 5 and v whose estimates take significantly different values, (13) is not a reflection of (11). Acknowledgments The authors are grateful to the referees for their many constructive comments and suggestions which resulted in a significant improvement in the manuscript. The first author thanks the Natural Sciences and Engineering Research Council of Canada for funding this research through an Individual Discovery Grant 5-36028 and the second author acknowledges the grant of MNIR 174013 for carrying out this research,206 1. Bokrshnan, RLM. Rs Journal of Multivariate Analy 143 (2016) 194-207 appt or) 0-apottor(t) 2s] ele ae] ge] zed eu a 5 ate - - 1 0 0 9 100 cr er a “Theoret Quant Theoetcal uerle -2plt or (1) 0-0piot or (13) gq ie és] é. get ae wom] o 7 6 0 100 Theat Quanta Thenecal Quanle (0-0 plot for bet family (0-0 plot or bet faiy bol ae | Fel eee — 5 ale, : : + 0 0 0 10 wo tate] Teretcal Ouarle Theoetcal Ouerle Fig. TheQ-Q plots for the marginal distributions lef Figures are for che random variable X, while the right Figures are forthe random variable ¥ PP plat or (11) P-P plot for (11) Be BSy eat mt esa fece Ppl) Plt 6° 6e4 Soon Seca Observed 09 04 08 ao 02 of 08 08 10 Cr er Expected Expected Fig. TheP-P ples fr the marginal distributions Left figure are forthe random variable X, while the right figures are fr the random variable .1. Bolakrishnon, MLM. Rist ourna of Multivariate Analysis 143 (2016) 194-207 207 References [1] 15. Akhundov, V.8,Nevzorov, Conditional distribution of recor vats end chsracterizavons of distributions Mah, Si 145 2) (2007) 4837-4892 [2] N- Balakrishna, K Shi On a clas of bivariate exponential distibutions. Statist. Probab, Let. 85 (2014) 153-160. {| :1 Carraco. jover, Estimating the generalized concordance correlation cocticient through vasiance components, iomstrcs 59 (2003) 849-858, [| J Carrasco, JP. Martinez, ecerm: Concordance Correlation Coefficient for Repeated (and Non-Repeated) Measures, R Package Version 12.1, 2015, drip] |CRAN-projectorgipackage-ceerm, [5] R. Castro, Goodness-of-fit (GoF) Tests, 2013, htp:iwavwin.tueal-rmcasto[AppStat20:3iflesjlectures23.p¢ 16] M. Franco, ;M. Ruiz, Characerizatios of disrete distributions based on conditional expectations of record Values, Stats. Papers 2 (1) (2001) 101-110, {| E Furman, Ona multivariate gamma distribution, Statist. Probab. Lett. 78 (2008) 2353-2360. [8] RD. Gupta, D-Kund, Theory and methods: Cenevalized exponential distributions, Aust, NZJ Stat. 41 (2)(1999) 173-188 [9] ETT. aynes, information theory and statistical mechanics, Phys. Re 105 (1957) 620-630. [10] 1 jolie, PB Hope, Bounded bivariate distebutions with nearly normal marginal, An. Stats, 5D (1996) 17-20, [11] MC Jones, pv Latsen, Multivaniate distributions with support above the diagonal, Biometrika 91 (2008) 375-986. [12] C Keseling, Conditional cstibations of generalized orer statistics and some characterization, Metriks 491) (1999) 27-40. [1] S Kotz, N: balakrishnan, NI_Johncon, Continuods Multivariate Distrbutions Volume I: Models snd Applications, second ed, Job Wiley and Sons, New York, 2000. [14] AML Matai PC. Moschopoulos. form of multivariate gamma distribution, An. inst. Statist Math 44 (1) (1992) 97-106. [15] AT. McKay, Sampling from batches. R Stat. Soc. (1934) 207-216, (Supplement). [16] GA Mihrar, RA, Hultquist A bivariate warning-imeyfaliure-time dstbution, | Amer. Statist Assoc. 62 (318) (1967) 989-599. [17] AP. Prudnikov. A. Bryetkov, 01. Marchey, Integrals and Series, vol. 1, Gordon and Breach Amsterdam, The Netherlands, 1986, [18] MAL. Ristc,N Balakrishnan, The Gamma-exponentated exponential distribution, Stat, Comput, Simul. 82 (2012) 1191-1206. [19] W. Site, W. Gonziles-Manteiga, MP. Quindiml, Bootstrap based goodness of i-tests,Metrika 40 1853) 243-256, [20 Kc Zagratos, Generalized beta generated I distbutions in: Modem Matinematical Tools and Techniques in Capturing Compleat, Springer, Beri, 2011, pp. 141-153 [21] K.2ogfatos, Balakrishnan, On families of beta- and generalized gamma-generated distributions and associated inference, Stat. Methodol 6 (2009) 3a,