### Paper Menu >>

### Journal Menu >>

Applied Mathematics, 2010, 1, 274-278 doi:10.4236/am.2010.14034 Published Online October 2010 (http://www.SciRP.org/journal/am) Copyright © 2010 SciRes. AM On Robustness of a Sequential Test for Scale Parameter of Gamma and Exponential Distribu tions Parameshwar V. Pandit1, Nagaraj V. Gudaganava r2 1Department of Statistics Bangalore University, Bangalore, India 2Department of Statistics Anjuman Arts, Science and Commerce, Dharwad, India E-mail: panditpv12@gmail.com Received May 8, 2010; revised July 30, 2010; accepted August 4, 2010 Abstract The main aim of the present paper is to study the robustness of the developed sequential probability ratio test (SPRT) for testing the hypothesis about scale parameter of gamma distribution with known shape parameter and exponential distribution with location parameter. The robustness of the SPRT for scale parameter of gamma distribution is studied when the shape parameter has undergone a change. The similar study is con- ducted for the scale parameter of exponential distribution when the location parameter has undergone a change. The expressions for operating characteristic and average sample number functions are derived. It is found in both the cases that the SPRT is robust only when there is a slight variation in the shape and location parameter in the respective distributions. Keywords: Gamma Distribution, Sequential Probability Ratio Test, Operating Characteristic Function, Average Sample Number Function, Robustness 1. Introduction The robustness of sequential probability ratio test (SPRT) has been studied by several authors for various probabil- ity distributions when the distribution under considera- tion has undergone a change. Barlow and Proschan [1], Harter and Moore [2], Montagne and Singpurwalla [3], and others have studied this problem for various prob- ability models. In this paper, the problem of testing simple hypothesis against simple alternatives for scale parameter of the gamma distribution assuming shape parameter to be known is considered. The gamma distribution plays im- portant role in many areas of the Statistics including ar- eas of life testing and reliability. It is used to make real- istic adjustment to exponential distribution in life-testing situations. The fact that a sum of independent exponen- tially distributed random variables has a gamma distribu- tion, leads to the appearance of gamma distribution in the theory of random counters and other topics associated with precipitation processes. In Section 2, we state the problem and develop SPRT for testing of hypothesis giving expressions for Operat- ing Characteristic (OC) and Average Sample Number (ASN) functions. In Section 3, robustness of the devel- oped SPRT with respect to OC functions when the dis- tribution considered here has undergone a change in the shape parameter, has been studied. In Section 4 we have studied the robustness of SPRT for the scale parameter of exponential distribution with respect to OC function when the location parameter has undergone a change. 2. Materials and Methods The set-up of the problem and SPRT Let X1, X2, ….. be a sequence of random variables from a gamma distribution with scale parameter θ (> 0) and shape parameter λ (> 0), whose density function is given by 1/ :,, 0 x xe fx x (1) where it is assumed that λ is known. Suppose we want to test the null hypothesis H0: θ = θ0 against the alternative H1: θ = θ1 (> θ0). For this problem following SPRT is developed. 0 1 0101 (; )11 Let InIn (; ) fx Z x fx (2) Two numbers A and B (0 < B < 1 < A) are chosen. At P. V. PANDIT ET AL. Copyright © 2010 SciRes. AM 275 nth stage accept H0 if 1 In n i i zB , reject H0 if 1 In n i i zA , otherwise, continue sampling by taking (n + 1) th observation. Here Zi is obtained by replacing X by Xi in (2). Let (α, β) be the desired strength of SPRT, then according Wald [4], A and B are approximately given by 1,1 AB , where α (0,1) and β (0,1). The Operating Characteristic (OC) function L(θ) is given by () () () 1 h hh A LAB (3) Where h(θ) is the non-zero solution of () 1 hz Ee (4) From (1) and (2), since () () 0 01 1 11 1() h hz Ee h we get from (4) that () 0 1 01 1 11 () h h (5) The Average Sample Number (ASN) function is given by ()ln[1 ()]ln |() LBL A EN EZ (6) Provided E(Z) 0, where 0 101 11 lnEZ (7) The maximum value of ASN occurs in the neighbour- hood of , say where is the solution of () 0EZ and this value is given by 2 ln .ln () () A B EN EZ (8) It is easy to see that 10 01 ln( /) 11 (9) and 2 20 1 () lnEZ (10) Table 1 contains the values of 0 EN , 1 EN and EN for different values of α, β and ratios of and ratios of θ0 and θ1. 3. Results and Discussions 3.1. Robustness of the SPRT for Scale Parameter of Gamma Distribution Let us suppose that there is misspecification for the shape parameter λ in the probability distribution. Then the pdf (1) becomes ;, *fx . To study the robust- ness of the SPRT developed in Section 2 with respect to OC function, consider h(θ) as the solution of the equa- tion. () *1 hz Ee i.e. () 1 0 0 (;, );, *1 (;, ) h fx fx dx fx giving () * 0 1 01 1 11 () h h (11) For different values of θ, h(θ) is evaluated and the OC function is obtained. The robustness of SPRT with re- spect to ASN function can be studied by replacing the denominator of (7) by 0 * 101 0 11 ;,* In*EZ zfxdx (12) We consider the cases * and * to study the robustness of the SPRT. In Table 2, we present the ASN function for the cases * for the SPRT of testing the hypothesis H0: θ = 45 against H1: θ = 50. The values of OC function for the cases * and * respectively are plotted Figure 1 and Figure 2. 3.2. Robustness of SPRT for Exponential Distribution The random variable X is said to follow exponential dis- tribution with location parameter and scale parameter P. V. PANDIT ET AL. Copyright © 2010 SciRes. AM 276 Table 1. ASN values of SPRT for scale parameter of gamma distribution. λ = 5, α = 0.01, β = 0.01 λ = 5, α = 0.05, β = 0.05 θ0/θ1 E0(N) E1(N) Eθ(N) θ0/θ1 E0(N) E1(N) Eθ(N) 0.1 0.64 0.13 0.80 0.1 0.38 0.08 0.33 0.2 1.11 0.38 1.63 0.2 0.65 0.22 0.67 0.3 1.79 0.80 2.91 0.3 1.05 0.47 1.20 0.4 2.85 1.54 5.03 0.4 1.68 0.91 2.07 0.5 4.66 2.94 8.79 0.5 2.74 1.73 3.61 0.6 8.13 5.78 16.18 0.6 4.78 3.40 16.64 0.7 15.89 12.53 33.20 0.7 19.35 17.37 13.63 0.8 38.92 33.54 84.81 0.8 22.90 19.73 34.82 0.9 168.02 156.62 380.42 0.9 198.87 192.16 156.20 λ = 5, α = 0.01, β = 0.01 λ = 5, α = 0.05, β = 0.05 θ0/θ1 E0(N) E1(N) Eθ(N) θ0/θ1 E0(N) E1(N) Eθ(N) 0.1 0.41 0.12 0.51 0.1 0.60 0.09 0.51 0.2 0.72 0.35 1.05 0.2 1.03 0.24 1.05 0.3 1.15 0.74 1.88 0.3 1.66 0.52 1.88 0.4 1.84 1.43 3.24 0.4 2.64 1.00 3.24 0.5 3.01 2.72 5.66 0.5 4.33 1.90 5.66 0.6 5.25 5.36 10.42 0.6 7.54 3.73 10.42 0.7 10.27 11.62 21.38 0.7 14.74 18.10 21.38 0.8 25.15 31.11 54.61 0.8 36.10 21.67 54.61 0.9 108.58 145.27 244.96 0.9 155.84 101.22 244.96 Table 2. ASN function for λ < λ* (H0: θ = 45, H1: θ = 50, α = β = 0.05, λ = 5. θ/λ* 5.0 5.05 5.10 5.15 5.20 5.25 41.00 41.30 44.11 47.3251.03 55.33 60.39 42.00 48.87 52.92 57.6663.26 69.93 77.94 43.00 59.62 65.72 73.0281.81 92.38 104.97 44.00 75.50 84.94 96.24109.58 124.72 140.68 45.00 99.02 112.80 128.09143.50 156.49 163.82 46.00 129.74 144.44 155.87160.87 157.93 147.84 47.00 154.07 157.58 153.15142.24 127.85 112.80 48.00 149.13 137.97 123.70109.01 95.54 83.89 49.00 120.96 106.59 93.4182.03 72.48 64.57 50.00 92.35 81.06 71.5963.74 57.25 51.83 51.00 71.40 63.49 56.9551.51 46.96 43.11 52.00 57.12 51.57 46.9443.03 39.70 36.85 53.00 47.25 43.23 39.8236.89 34.36 32.16 54.00 40.18 37.15 34.5432.27 30.28 28.52 55.00 34.91 32.55 30.4928.68 27.06 25.62 if its probability density function is given by ;,, 0, 0 x fxe x (13) where it is assumed that is known. Given a sequence of observations X1, X2, …, from (13), we wish to test the null hypothesis H0: = 0 against the alternative H1: = 1 (> 0). We propose the following SPRT. 11 10 00 (; ) Let InIn (; ) fx Zx fx (14) We choose two numbers A and B such that 0 < B < 1 < A. At nth stage, accept H0 if 1 In n i i zB , reject H0 if 1 ln n i i Z A , otherwise continue sampling by taking (n + 1)th observation. Here Zi is obtained by replacing x by xi in (3.2). If and are the probabilities of type I and type II errors respectively, then 1,1 AB The OC function L( ) is given by P. V. PANDIT ET AL. Copyright © 2010 SciRes. AM 277 0 0.2 0.4 0.6 0.8 1 1.2 41 42 43 4445 4647 484950 51525354 55 54.95 4.9 4.85 4.8 4.75 Figure 1. Graph of OC function for gamma distribution < *. 0 0.2 0.4 0.6 0.8 1 1.2 41 4243 44 4546 4748 49 5051 5253 54 55 55.05 5.1 5.15 5.2 5.25 Figure 2. Graph of OC function for gamma distribution > *. () () () 1, h hh A LAB (15) Where h( ) is the solution of 1 hZ Ee . Here, 1 0 h hZ Ee and hence we obtain 0 1 h . Let is suppose that the location parameter has under- gone a change and the pdf (3.1) becomes ;*;fx . To study the robustness of the SPRT developed here with respect to OC function, we consider h( ) as the solution of the equation () 1 0 0 (; ,);*; 1 (; ,) h fx fx dx fx giving *1 10 0 ln ()()ln h (16) 0 0.2 0.4 0.6 0.8 1 1.2 0.10.5 0.9 1.31.722.3 2.7 0-0.05 -0.1 -0.15 -0 .2 Figure 3. Graph of OC function for exponential distr i bution θ < θ*. 0 0. 2 0. 4 0. 6 0. 8 1 1. 2 0.1 0.50.9 1.3 1.722.3 2.7 00.05 0.1 0.15 0.2 Figure 4. Graph of OC function for exponential distr i bution θ > θ*. For different values of , h( ) is evaluated and OC function is obtained with the help of Equation (15). The values of OC function for the cases * and * are plotted in Figures 3 & 4 respectively. 4. Conclusions Some Remarks 1) The values of OC function for the cases * and * respectively are plotted Figure 1 and Fig- ure 2. The figures indicate that for * (* ), the OC curve shifts to the right side (left side) of the curve for when * . From the figure it is clear that SPRT is robust for *0.05 as the deviation in OC function is insignificant. However, for other values of * the test becomes highly non-robust. 2) We have considered the SPRT for testing the hypothesis H0: = 1 versus H1: = 2 for = = 0.05. The values of OC function for the cases * and * are plotted in Figures 3 & 4 respectively. It is exhibited from the table that the SPRT is non-robust even for * = 0.05. The ASN function for the case * is tabulated in Table 2. P. V. PANDIT ET AL. Copyright © 2010 SciRes. AM 278 5. References [1] R. E. Barlow and E. Proschan, “Exponential Life Test Procedures When the Distribution under Test has Mono- tone Failure Rate,” Journal of American Statistic Asso- ciation, Vol. 62, No. 318, 1967, pp. 548-560. [2] L. Harter and A. H. Moore, “An Evaluation of Exponen- tial and Weibull Test Plans,” IEEE Transactions on Re- liability, Vol. 25, No. 2, 1976, pp. 100-104. [3] E. R. Montagne and N. D. Singpurwalla, “Robustness of Sequential Exponential Life-Testing Procedures,” Jour- nal of American Statistic Association, Vol. 80, No. 391, 1985, pp. 715-719. [4] A. Wald, “Sequential Analysis,” John Wiley, New York, 1947. |