Hypothesis Testing Based on the Maximum of Two Statistics from Weighted and Unweighted Estimating Equations

Transcription

1 Hypothesis Testing Based on the Maximum of Two Statistics from Weighted and Unweighted Estimating Equations Takeshi Emura and Hisayuki Tsukuma Abstract For testing the regression parameter in multivariate failure time data analysis, the solution to partial likelihood score equation provides useful test statistic. Proper weight insertion in the score equation allows us to use more powerful test statistic, especially when failure times within a cluster are strongly correlated. In some cases two solutions to weighted and unweighted score equations, which are calculated from the same data, are very different. This means two statistics based on the solutions to weighted and unweighted score equations yield different results of the testing. If one has no prior information in advance to prefer one statistic over the other, it is natural to consider the maximum of the two test statistics for testing the regression parameter. In this paper, we discuss the way to approximate the asymptotic distribution of the maximum statistic, and construct hypothesis testing based on the maximum statistic. Simulation studies are carried out to see the validity of the approximation of the maximum statistic and its power properties. 1 Introduction Let (T i1,..., T in ), i = 1,..., K, be K clusters of n-dimensional multivariate failure time data and let (z i1,..., z in ) be K clusters of the corresponding p-dimensional covariate vectors. Then, marginal Cox (1972)-type hazard models for the jth member in cluster i, with common regression parameter and baseline hazard function, formulates the hazard function of T ij given z ij as λ(t z ij ) = e zt ijβ λ (t), i = 1,..., K, j = 1,..., n. (1.1) Here, β is a p-vector of unknown common regression parameters and λ (t) is an unknown, unspecified common baseline hazard function. Cai and Prentice (1997) introduced estimating equation for the regression parameter β, by incorporating weight matrices into partial likelihood score equation. They proved that the solution to the estimating equation, ˆβ, is consistent and asymptotically normal. Let (C i1,..., C in ) be censoring time. Assume that, given z ij, T ij and C ij are independent for all i, j. Considering the right censoring situation, observed data are X ij = min{t ij, C ij }, δ ij = I(T ij C ij ), and z ij for every i, j. Assume that (T i1,..., T in, C i1,..., C in, z i1,..., z in ) are independent and identically distributed(i.i.d.)from the same probability model. We use only time independent covariates for brevity, but our methods can easily be extended to the case of time dependent covariates with finite variations. Define N ij (t) = I(X ij t, δ ij = 1), Y ij (t) = I(X ij t), then, M ij (t) = N ij (t) t Y ij (s)e zt ijβ λ (s)ds Graduate School of Science and Technology, Chiba University, 1-33 Yayoi-cho, Chiba , Japan; s.s.takeshi@ezweb.ne.jp The institute of Statistical Mathematics, Minami-Azabu, Minato-ku, Tokyo , Japan; tsukuma@ism.ac.jp 1

2 is a martingale with respect to a marginal filtration (see Spiekerman and Lin (1998)). An estimated martingale replaces this martingale with its empirical counterpart, and thus, ˆM ij (β, t) = N ij (t) t t Y ij (s)e zt ijβ dˆλ (β, s), where ˆΛ (β, t) = K n i=1 dn ij(s)/( K n k=1 l=1 Y kl(s)e zt β kl ) and β is an arbitrary p-vector. Then, estimating equation of Cai and Prentice (1997) are defined as i=1 { k=1 } z ik w ikj (β, s) d ˆM ij (β, s) =. Here K weight matrices w i11 (β, t) w i1n (β, t) W i (β, t) =....., i = 1,... K, (1.2) w in1 (β, t) w inn (β, t) are arbitrary functions of β and t and T > is a fixed constant. In case every W i (β, t) is an identity matrix, (1) is partial likelihood score equation, and we call these unweighted estimating equation throughout this paper. Cai and Prentice (1997) showed in their simulation studies that the inverses of some correlation matrices between martingales allow the solution to the estimating equation to have smaller variance than the solution to the unweighted estimating equation. One might not know whether weight insertion achieves great efficiency prior to the analysis of the experiment. It is expected that weight insertion always reduces the variance of the regression estimator when the weight matrices (1.2) are properly selected and are completely known. However, the proper weight matrices have generally to be estimated from data, and the deviation of the weight matrix estimators from the true values often causes the efficiency of the regression parameter estimator to reduce slightly. One might calculate two test statistics from both weighted estimating and unweighed estimating equation. Usually the two analyses yield very similar results, but in some cases, the results can be very different. On the nonparametric test of the equality of two survival curves, the quite similar problem of whether one uses the Mantel s logrank test or the generalized Wilcoxon test is discussed earlier by Tarone (1981). He suggests the maximum of the logrank test statistic and the Wilcoxon test statistic when none of prior information to prefer one statistic over the other is given. In this paper, we consider the problem of the level α hypothesis testing of H : β = β in favor of the alternative H 1 : β β under model (1.1). Let ˆβ 1 be the solution to partial likelihood score equation, or unweighted estimating equation, U 1 (β) = i=1 z ij d ˆM ij (β, s) =, (1.3) and let ˆβ 2 be the solution to weighted estimating equation U 2 (β) = i=1 { k=1 } z ik w ikj (β, s) d ˆM ij (β, s) =. (1.4) For each l(l = 1, 2), asymptotic normality of K(ˆβ l β ) (l = 1, 2) and consistency of their variance estimators ˆΩ l (ˆβ l ), which are described in Cai and Prentice (1997), yields two different Wald type hypothesis testings: One is the test which rejects H whenever Q 1 = K(ˆβ 1 β ) T ˆΩ1 (ˆβ 1 ) 1 (ˆβ 1 β ) > χ 2 p(α) (1.5) 2

3 and the other is the test which rejects H whenever Q 2 = K(ˆβ 2 β ) T ˆΩ2 (ˆβ 2 ) 1 (ˆβ 2 β ) > χ 2 p(α). (1.6) Here χ 2 p(α) is a percentage point of chi square distribution with degrees of freedom p. Since one might not be able to select a statistic prior to the analysis of experiment, we will examine the next test which rejects H whenever Q max = max{q 1, Q 2 } > ˆχ 2 max(α). (1.7) Here ˆχ 2 max(α), which is depend on data, is an estimator of a percentage point we describe later. The use of the third statistic (1.7) is regarded as the direct application of Tarone (1981) s idea to the marginal Cox-type regression analysis. In Section 2, we examine the asymptotic distribution of Q max, and construct an asymptotically level α test based on Q max. In Section 3, simulation studies are provided under several parameter configurations to see the validity of the approximation of the statistic Q max, and to compare the power properties of the three test statistics. In Section 4, it is concluded from our simulation studies that the empirical power of the test (1.7) is higher than the power of either the test (1.5) or the test (1.6) for most of the alternatives. 2 Theoretical results and test based on Q max 2.1 Notation and theoretical results We consider β an arbitrary element in a compact set B( R p ), which contains the true value β. Define a 1,ij (β, t) = z ij and a 2,ij (β, t) = n k=1 z ikw ikj (β, t), and the following notation: S () (β, t) = 1 K S (1) (β, t) = 1 K S (2) l (β, t) = 1 K S (3) l (β, t) = 1 K i=1 i=1 i=1 i=1 E l (β, t) = S(2) l (β, t) S () (β, t) Y ij (t)e zt ijβ, Y ij (t)z ij e zt ijβ, Y ij (t)a l,ij (β, t)e zt ijβ (l = 1, 2), Y ij (t)a l,ij (β, t)z T ije zt ijβ (l = 1, 2), (l = 1, 2), V l (β, t) = S(3) l (β, t) S () (β, t) S(2) l (β, t)s (1) (β, t) T S () (β, t) 2 (l = 1, 2). Under the model assumptions of the listed earlier the regularity conditions outlined in Section 3 of Cai and Prentice (1997), there exist non-random functions, on B [, T ], s (r) l (β, t), (r =, 1, 2, 3, 4), e l (β, t), and v l (β, t) such that sup β,t S (r) l (β, t) s (r) l (β, t) P, sup β,t E l (β, t) e l (β, t) P and sup β,t V l (β, t) v l (β, t) P (K ) for (l = 1, 2), and the matrix A l (β) = v l (β, s)s () (β, s)λ (s)ds 3

4 is assumed to be positive definite and symmetric. Define D l,i = { } a l,ij (β, s) e l (β, s) dm ij (s) (l = 1, 2), then K random 2p-vertors (D T 1,i, D T 2,i) are i.i.d. and the covariance matrices between D l,i and D m,i are { { } } Σ lm (β ) = E a l,1j (β, s) e l (β, s) dm 1j (s) { { } T a m,1j (β, s) e m (β, s) dm 1j (s)} (l, m = 1, 2). Asymptotic distribution of (ˆβ 1, ˆβ 2 ) such that U 1 (ˆβ 1 ) = and U 2 (ˆβ 2 ) = is summarized in the next proposition. Proposition 1 Suppose the model assumptions listed earlier and the regularity conditions A-D of the section 3 in Cai and Prentice (1997, pp 2) hold. Then, ( K(ˆβ 1 β ) T, K(ˆβ 2 β ) T ) d (W T 1, W T 2 ) (K ), where both W 1 and W 2, are p-vectors and (W T 1, W T 2 ) is 2p-dimensional normal distribution with E(W l ) = and E(W l W T m) = A l (β ) 1 Σ lm (β )A m (β ) 1 Ω lm (β ) for l, m = 1, 2. Proof. By the same argument in the Appendix of Cai and Prentice (1997, pp 212), K(ˆβl β ) = 1 K A l (β ) 1 D l,i + o p 1 (1), K i=1 for l = 1, 2, where o p 1 (1) is a p 1 vector with o p (1) elements. To establish multivariate convergence, it is sufficient by the Cramer-Wold device to prove 2 l=1 c T l K(ˆβl β ) d 2 c T l W l for all c 1 R p and c 2 R p. The central limit theorem for i.i.d. random vectors establishes 2 l=1 c T l l=1 1 K A l (β ) 1 d D l,i N(, σ), K i=1 where σ = 2 2 l=1 m=1 ct l A l(β ) 1 Σ lm (β )A m (β ) 1 c l. It is easy to see that the distribution of 2 l=1 ct l W l is equal to N(, σ). If the weight functions are unknown, one needs to replace w ikj (β, s) by its estimator ŵ ikj (β, s). We suspect that, by adding some regularity conditions such as the condition E of Cai and Prentice (1997), Proposition 1 can be extended to the case where weight functions are estimated. Note that the covariance matrix of (W T 1, W T 2 ) can be estimated by the following plug-in estimators and we define them for the estimating equation with both deterministic weights and estimated weights. ˆΩ ll (ˆβ l ) = Âl(ˆβ l ) 1 ˆΣl (ˆβ l )(Âl(ˆβ l ) 1 ) T (l = 1, 2), (2.1) 4

5 ˆΩ lm (ˆβ l, ˆβ m ) = Âl(ˆβ l ) 1 ˆΣlm (ˆβ l, ˆβ m )(Âm(ˆβ m ) 1 ) T (l m), (2.2) Â l (β) = 1 K ˆΣ 11 (β) = 1 K ˆΣ 22 (β) = 1 K ˆΣ 12 (β 1, β 2 ) = 1 K i=1 { i=1 { { i=1 { { i=1 { ˆΣ 21 (β 2, β 1 ) = ˆΣ 12 (β 1, β 2 ) T. ˆV l (β, s)dn ij (s) (l = 1, 2) { zij E 1 (β, s) } d ˆM ij (β, s)} { zij E 1 (β, s) } d ˆM } T ij (β, s), {â2,ij (β, s) Ê2(β, s) } d ˆM ij (β, s)} {â2,ij (β, s) Ê2(β, s) } d ˆM } T ij (β, s), { zij E 1 (β 1, s) } d ˆM ij (β 1, s)} {â2,ij (β 2, s) Ê2(β 2, s) } d ˆM } T ij (β 2, s), We define â 2,ij, Ê 2, and ˆV 2 as a 2,ij, E 2, and V 2 respectively if weights are not estimated, and as a 2,ij, E 2, and V 2 with w ikj (β, s) replaced by its estimator ŵ ikj (β, s) if weights are estimated. The estimators (2.1) are suggested by Cai and Prentice (1997), and they point out that these are consistent estimators for Ω 11 (β ) and Ω 22 (β ). We suggest the estimator (2.2), which is a natural extension of (2.1) to Ω lm (β ) (l, m = 1, 2), and it may be a consistent estimator. 2.2 The test based on the maximum statistics Q max Proposition 1 and the estimator of the covariance matrix, (2.1) and (2.2), allow us to approximate the percentage point of Q max. According to Proposition 1, the distribution of ( K(ˆβ 1 β ) T, K(ˆβ 2 β ) T ) is approximated by (W T 1, W T 2 ). The unknown distribution of (W T 1, W T 2 ) is further approximated by the known 2p-dimensional normal distribution [ Ŵ 1 Ŵ 2 ] ([ N 2p ], [ ˆΩ 11 (ˆβ 1 ) ˆΩ12 (ˆβ 1, ˆβ 2 ) ˆΩ 21 (ˆβ 2, ˆβ 1 ) ˆΩ22 (ˆβ 2 ) A pair of the known random variables (Ŵ T 1 ˆΩ 11 (ˆβ 1 ) 1 Ŵ 1, Ŵ T 2 ˆΩ 22 (ˆβ 2 ) 1 Ŵ 2 ) can approximate the distribution of (Q 1, Q 2 ) as long as ˆΩ ll (ˆβ l ) (l = 1, 2) are positive definite. Exactly the same approximation can be obtained by using standardization of the normal distribution of Ŵ l (l = 1, 2). Define ˆV l = ˆΩ ll (ˆβ l ) 1/2 Ŵ l and ˆΩ ll (ˆβ l ) 1/2 ˆΩll (ˆβ l ) 1/2 = ˆΩ ll (ˆβ l ) 1 (l = 1, 2). Then the distribution of ( ˆV T 1, ˆV T 2 ) is 2p-dimensional normal with mean and covariance matrix given by ) ( Ip ˆΓ12 ˆΓ T 12 where I p is a p p identity matrix and ˆΓ 12 = ˆΩ 11 (ˆβ 1 ) 1/2 ˆΩ12 (ˆβ 1, ˆβ 2 ) ˆΩ 22 (ˆβ 2 ) 1/2. Again, the known distribution ( ˆV T 1 ˆV 1, ˆV T 2 ˆV 2 ) can approximate the distribution of (Q 1, Q 2 ). The percentage 5 I p, ]).

6 point of Q max = max{q 1, Q 2 }, say ˆχ 2 max(α) in (1.7), is approximated by the percentage point of the distribution of max{ŵ T 1 ˆΩ 11 (ˆβ 1 ) 1 Ŵ 1, Ŵ T 2 ˆΩ 22 (ˆβ 2 ) 1 Ŵ 2 } or max{ ˆV T 1 ˆV 1, ˆV T 2 ˆV 2 }, for example, via Monte Carlo simulations. Similar to the fact mentioned in Fleming and Harrington (1991, pp.28) for weighted logrank tests, the test (1.7) is closely related to the p-values of the tests (1.5) and (1.6). Suppose one computes the Wald statistics Q 1 and Q 2, and corresponding p-values p 1 and p 2. The test which rejects H if min{p 1, p 2 } p(α) such that P H (min{p 1, p 2 } p(α)) = α is conceivable as long as p(α) can be computed. From the definition of p-value, we can see min{p 1, p 2 } p(α) Q max χ 2 p(p(α)). and therefore, the value p(α) is approximated by ˆp(α) = 1 F χ 2 p (ˆχ 2 max(α)), where F χ 2 p is the distribution function of χ 2 p. Hence, the test (1.7) is exactly the same as the test which rejects H whenever min{p 1, p 2 } ˆp(α). 3 Simulation Studies We focused on the case where p = 1 in (1.1) throughout our simulation studies and hence β is a scalor in which H : β = β versus H 1 : β β are tested. The models we used were 2- and 3-dimensional models of Clayton and Cuzick (1985): The conditional survival function for (T 1,..., T n ) given (z 1,..., z n ) is { S(t 1,..., t n z 1,..., z n ) = exp ( t j θe z jβ ) 1/θ (n 1)} (n = 2, 3). The parameter θ determines the degree of dependence between T j and T k, (j, k = 1,..., n), and in simulation studies we used θ =.25 (strongly correlated), θ = 1.5 (weakly correlated) and θ = 1 (nearly independent). The covariates z 1,..., z n were considered to be mutually independent, { 1} valued random variable with P (z 1 = 1) = = P (z n = 1) =.5. The censoring variables C i1,..., C in were independently generated from exponential distributions with hazard function being µ 1,..., µ n respectively. The number of clusters, K, was fixed at 1. The unweighted score equation (1.3) was solved by using Newton-Raphson procedure and ˆβ 1 denotes this solution. For weight matrices, we specified w i11 w i12 w i1n w i21 w i22 w i2n = w in1 w in2 w inn 1 ρ i12 (z i1, z i2 ) ρ i1n (z i1, z in ) ρ i12 (z i1, z i2 ) 1 ρ i2n (z i2, z in ) ρ i1n (z i1, z in ) ρ i2n (z i2, z in ) 1 where ρ ijk (z ij, z ik ) = corr{m ij (X ij ), M ik (X ik ) z ij, z ik } is a correlation between two unknown random variables given (z ij, z ik ) for i = 1,..., K, and j, k = 1,..., n. This weight matrices are slightly different from those used by Cai and Prentice (1997) in their simulation studies, in that we do not need to know potential censoring times for the failing individuals. Since these weight matrices are unknown, we replaced these by their empirical estimators. The restriction of z ij to be { 1} valued is crucial since the number of parameters in the all weight matrices reduces significantly. In the case of n = 2, only four parameters, ρ 12 (, ), ρ 12 (, 1), ρ 12 (1, ) and ρ 12 (1, 1) determine the weight matrices, and in the case of n = 3, so do only 12 parameters. Each value of ρ ijk (z ij, z ik ) was estimated by the product moment correlation coefficient of the group of pseudodata ( ˆM i j( ˆβ 1, X i j), ˆM i k( ˆβ 1, X i k)), i R(z ij, z ik ), where R(z ij, z ik ) = {i z i j = z ij, z i k = z ik }. When #R(z ij, z ik ) = or 1, we set ˆρ ijk (z ij, z ik ) =. 6 1,

7 ˆβ 2 denotes the solution to the estimating equation (1.4) with these estimated weights, and this was solved by using Newton-Raphson procedure. Rarely do extremely large values of the estimated weights cause the Newton-Raphson algorithm to stop. We regulated these phenomena by letting the values of Q l (l = 1, 2) be whenever U l (β)/ β <.1 (l = 1, 2). All simulations were based on 1, runs, in which ERR1 and ERR2 indicate the total number of this type of error regulations in both U 1 (β) and U 2 (β) respectively. For each run, the statistics Q 1, Q 2 and Q max and the estimator ˆΓ 12 were computed. The quantiles of the Q max, that is ˆχ 2 max(.5), were approximated by the quantiles based on the 1, random samples from the distribution of ( ˆV 1, ˆV 2 ) for every run. First, our aim in this simulation studies is to see how well the distribution of ( ˆV 1, ˆV 2 ) approximates that of ( K( ˆβ 1 β )/{ˆΩ 1 ( ˆβ 1 ) 1/2 }, K( ˆβ 2 β )/{ˆΩ 2 ( ˆβ 2 ) 1/2 }). Since the distribution of ( ˆV 1, ˆV 2 ) is completely determined by the estimator ˆΓ 12, it is worth seeing the performance of ˆΓ 12 compared with the actual correlation coefficient of ( K( ˆβ 1 β )/{ˆΩ 1 ( ˆβ 1 ) 1/2 }, K( ˆβ 2 β )/{ˆΩ 2 ( ˆβ 2 ) 1/2 }). Mean and standard deviation of ˆΓ 12 and product moment correlation coefficient of ( K( ˆβ 1 β )/{ˆΩ 1 ( ˆβ 1 ) 1/2 }, K( ˆβ 2 β )/{ˆΩ 2 ( ˆβ 2 ) 1/2 }) are listed as Γ 12, SD(ˆΓ 12 ), and r respectively in Table 1. P denotes the probability of censoring calculated from the actual sample. For the test of β =, that is H : β =, the empirical size and power of Q 1, Q 2 and Q max based on these 5% points are given in Tables 2 and 3. Based on the Tables 2 and 3, power functions are drawn in Figures Concluding remarks We have discussed the way to approximate the asymptotic distribution of Q max to construct the hypothesis testing (1.7). As shown in Table 1, the approximation of Q max by ( ˆV 1, ˆV 2 ) performed well in terms of the correlation coefficient of ( K( ˆβ 1 β )/{ˆΩ 1 ( ˆβ 1 ) 1/2 }, K( ˆβ 2 β )/{ˆΩ 2 ( ˆβ 2 ) 1/2 }) and ˆΓ 12. Empirical size of Q max listed in Tables 2 and 3 appears to slightly larger than the nominal 5% level, especially in the case of 3-dimensional Clayton model. At almost every alternatives, the empirical power of the test based on Q max did not become the worst ones compared to Q 1 and Q 2. Therefore, we can conclude that the test (1.7) tends to prevent the power reduction caused by selecting the wrong estimating equation. Acknowledgement The authors would like to thank Dr. Yoshihiko Konno for helpful comments. References [1] Cai, J. and Prentice, R.L. (1997). Regression estimation using multivariate failure time data and a common baseline hazard function model, Lifetime Data Analysis 3, [2] Clayton, D.G. and Cuzick, J. (1985). Multivariate generalizations of the proportional hazards model (with discussion), Journal of the Royal Statistical Society Series A 148, [3] Fleming, T.R. and Harrington, D.P. (1991). Counting Process and Survival Analysis, Wiley: New York. [4] Spiekerman, C.F. and Lin, D.Y. (1998). Marginal regression models for multivariate failure time data, Journal of the American Statistical Association 93,

8 [5] Tarone, R.E. (1981). On the distribution of the maximum of the logrank statistics and the modified Wilcoxon statistics, Biometrics 37,

9 Table 1: Mean and standard deviation of ˆΓ 12 and product moment correlation coefficient of ( K( ˆβ 1 β )/{ˆΩ 1 ( ˆβ 1 ) 1/2 }, K( ˆβ 2 β )/{ˆΩ 2 ( ˆβ 2 ) 1/2 }) 2 dimensional Clayton model θ β P Γ12 SD(ˆΓ 12 ) r ERR1 ERR2 µ 1 = µ 2 = µ 1 = µ 2 = dimensional Clayton model θ β P Γ12 SD(ˆΓ 12 ) r ERR1 ERR2 µ 1 = µ 2 = µ 3 = µ 1 = µ 2 = µ 3 =

10 Table 2: Empirical power of tests based on Q1, Q2, and Qmax for nominal 5 % significance level under 2 dimensional Clayton model (* indicates the worst empirical power among Q1, Q2, and Qmax) µ1 = µ2 =.5 β (true) θ =.25 Q Q Qmax ERR1 ERR2 θ = 1.5 Q Q Qmax ERR1 ERR2 θ = 1. Q Q Qmax ERR1 ERR2 µ1 = µ2 = 1.5 β (true) θ =.25 Q Q Qmax ERR1 ERR2 θ = 1.5 Q Q Qmax ERR1 ERR2 θ = 1. Q Q Qmax ERR1 ERR2 1

11 Figure 1: Comparison of powers of Q 1 (dots), Q 2 (dashes), and Q max (plain) for 5% significance level under 2 dimensional Clayton model: Clayton dependence parameters θ =.25 (upper figure), θ = 1.5 and θ = 1 (lower figure) with µ 1 = µ 2 =.5 11

12 Figure 2: Comparison of powers of Q 1 (dots), Q 2 (dashes), and Q max (plain) for 5% significance level under 2 dimensional Clayton model: Clayton dependence parameters θ =.25(upper figure), θ = 1.5 and θ = 1 (lower figure) with µ 1 = µ 2 =

13 Table 3: Empirical power of tests based on Q1, Q2, and Qmax for nominal 5 % significance level under 3 dimensional Clayton model (* indicates the worst empirical power among Q1, Q2, and Qmax) µ1 = µ2 = µ3 =.5 β (true) θ =.25 Q Q Qmax ERR1 ERR θ = 1.5 Q Q Qmax ERR1 ERR θ = 1. Q Q Qmax ERR1 ERR2 1 µ1 = µ2 = µ3 = 1.5 β (true) θ =.25 Q Q Qmax ERR1 ERR θ = 1.5 Q Q Qmax ERR1 ERR2 1 θ = 1. Q Q Qmax ERR1 ERR2 1 13

14 Figure 3: Comparison of powers of Q 1 (dots), Q 2 (dashes), and Q max,(plain) for 5% significance level under 3 dimensional Clayton model: Clayton dependence parameters θ =.25(upper figure), θ = 1.5 and θ = 1(lower figure) with µ 1 = µ 2 = µ 3 =.5 14

15 Figure 4: Comparison of powers of Q 1 (dots), Q 2 (dashes), and Q max,(plain) for 5% significance level under 3 dimensional Clayton model: Clayton dependence parameters θ =.25(upper figure), θ = 1.5 and θ = 1(lower figure) with µ 1 = µ 2 = µ 3 =