Analyzing Experiments with Correlated Multiple Responses

CHIH-HUA CHIAO
Soochow University, Taipei, Taiwan

MICHAEL HAMADA
Los Alamos National Laboratory, Los Alamos NM 87545
Statistically designed experiments have been employed extensively to improve product or process quality and to make products and processes robust. In this paper, we consider experiments with correlated multiple responses whose means, variances, and correlations depend on experimental factors. Analysis of these experiments consists of modeling distributional parameters in terms of the experimental factors and nding factor settings which maximize the probability of being in a specication region, i.e., all responses are simultaneously meeting their respective specications. The proposed procedure is illustrated with three experiments from the literature.

Introduction
method for improving quality A proactiveproducts and processes robust is toand making use statistically designed experiments. Such experiments allow one to assess a variables impact on a product or processs quality. For most products, quality is multidimensional, so it is common to observe multiple responses on the experimental units. There are numerous examples of experiments with multiple responses that appear in the literaturesee, for example, Symposia on T aguchi M ethods (American Supplier Institute (1984-1993)). When analyzing multiple response data, analyzing each response separately is not very satisfactory, especially when the multiple responses are correlated. For example, separate analyses may lead to conicting recommendations regarding the levels of important factors, because a factor level may improve one responses quality but degrade anothers. Moreover, when correlations between the responses are ignored, one may miss nding settings which simultaneously improve the quality of all the responses.

Dr. Chiao is an Associate Professor in the Department of Business Mathematics. His email address is chchiao@bmath. scu.edu.tw. Dr. Hamada is a Technical Sta Member in Statistical Sciences.

Derringer and Suich (1980) propose using desirability functions which turn the multiple response problem into a single response problem; that is, the response that is analyzed is the total desirability D = (d1 (Y1 ) dm (Ym ))1/m , where (Y1 , . . . , Ym ) are m responses and d1 , . . . , dm are individual desirabilities. There are two drawbacks to this approach: rst, D will likely be harder to model because it is a complex function of the m responses; second, it is hard to say what the dierence between expected values of D means except that the higher one is better. Other papers account for a common variancecovariance matrix of the multiple responses in a criterion to determine optimal settings (Khuri and Conlon (1993) Khuri (1996) and Vining (1998)). In this paper, we consider the variance-covariance matrix as dependant on experimental factors as suggested by Pignatiello (1993). This might be viewed as an extension of modeling variances (Bartlett and Kendall (1946)), which has been promoted by Taguchi (1986). We propose a simple method to properly handle multiple responses. First, the parameters of the multiple response distribution are modeled in terms of the experimental factors. Then, factor settings that optimize a suitable criterion are found. The criterion considered in this paper is the probability that all responses simultaneously meet their respective specications, a criterion that is easily interpreted. The paper is organized as follows. First, we con-

Vol. 33, No. 4, October 2001

451

www.asq.org

452

CHIH-HUA CHIAO AND MICHAEL HAMADA

FIGURE 1. Specication Regions for a Bivariate Response.

sider a specication region determined by each responses own specications, and then we consider a multivariate normal distribution for the multiple responses. Next, two classes of experimental designs are discussed that are used to improve quality and to make quality robust. The multiple response data from such experiments can then be analyzed by estimating the parameters of the multivariate normal distribution and then modeling them in terms of the experimental factors. This proposed analysis methodology is illustrated with data from three experiments for which the goals were quality improvement and robust quality improvement. The paper concludes with a discussion.

A Specication Region and Model for Multiple Responses

First we consider a specication region for multiple responses or, in other words, a multivariate re-

sponse. For m responses, let Y = (Y1 , Y2 , , Ym )T denote the m 1 vector of multiple responses or an m dimensional multivariate response. Each component response Yi has a specication (li , ui ) consisting of lower and upper limits within which it is desirable for Yi to be. Using the component specications, we dene a specication region S for the multivariate response as the m dimensional hypercube whose sides are the m component specications. Figure 1 displays the nine possible specication regions for a bivariate response (m = 2) as indicated by the shaded areas. Given a specication region, a measure of quality is the probability that the m component responses are simultaneously meeting their respective specications or the proportion of conformance (which Wang and Lam (1996) introduced for a univariate response) P (Y S). (1)

In order to calculate the value in Equation (1) we

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES

453

FIGURE 2. Some 95% Contours and a Specication Region.

need a model for the multivariate response Y. We assume that Y follows an m dimensional multivariate normal distribution with mean = (1 , 2 , , m )T and variance-covariance matrix (the diago2 2 2 nal elements of which are the variances 1 , 2 , , m and the o-diagonal elements of which are the covariances k k , 1 k < l m, where k is the correlation between Yk and Yl ). The joint probability density function for Y, f (Y; , ), is given by f (Y; , ) = (2)(m/2) ||1/2 1 exp (Y ) 1 (Y ) . 2 (2) Assuming that the multivariate normal distribution for Y in Equation (2) holds, the proportion of confor-

mance in Equation (1) can be evaluated for a given and . Figure 2 presents comparisons of two bivariate response distributions with dierent and . The 95% contours for Y (i.e., the probability content of a contour which is centered at its is 0.95.) indicate the shape of the distribution. Figure 2 illustrates three situations in which the specication region is a rectangle. In Figure 2a, only the means of both Y1 and Y2 are dierent, and case (a1) is better; the two cases, (a1) and (a2), have proportions of conformance of 0.336 and 0.119, respectively. In Figure 2b, only the variances of Y2 are dierent. Case (b2) is preferred because the Y2 variance for this case is smaller. The two cases (b1) and (b2) have proportions of conformance of 0.366 and 0.911, respectively. In Figure 2c, the correlations between Y1 and Y2 are dierent; there is no correlation in (c1) and nega-

Vol. 33, No. 4, October 2001

www.asq.org

454

CHIH-HUA CHIAO AND MICHAEL HAMADA

FIGURE 3. The Inverse Hyperbolic Tangent Transformation of the Correlation (Rho).

tive correlation ( = 0.9) in case (c2). Case (c2) is preferred because of its proportion of conformance of 0.501, which is comparatively better than 0.366 for case (c1). Note that cases (a1), (b1), and (c1) are the same for purposes of comparison. Figures 2b and 2c show that dierent variances and correlations can impact the proportion of conformance. As can be seen in Figure 2, the proportion of conformance is determined by the means, variances and correlations of the multivariate response distribution. In order to improve quality or make quality robust, one needs to rst identify what variables impact these distributional parameters and then nd the values of these variables which make these goals achievable. In the next section, we consider models for these distributional parameters in terms of these variables which will be used to analyze experiments with multiple responses.

of the m dimensional multivariate response: k = xk ,

2 log(k ) = xk ,

k = 1, , m, k = 1, , m, 1 k < l m.

(3) (4)

and tanh1 (k ) = xk , (5)

Note that k , k , and k are column vectors. The nonzero components of k , k , and k determine the impact of the associated variables. We refer to as the vector of all the coecients from Equations (3) - (5). Note that the standard log-linear model for variances in Equation (4) ensures positive values for the variances. All real values are allowed on the log scale so that the additive model given above is well dened. A similar transformation is needed for correlations () which range between 1 and +1; here we use the inverse hyperbolic tangent transformation in Equation (5) (Rao (1973)), which is dened as tanh1 () = 1 (1 + ) log 2 (1 )

Distributional Parameter Models

Let x denote a row vector of covariates associated with a set of factors which may include an intercept, main eects, two-factor interactions, etc. Consider the following models for the distribution parameters

and is displayed in Figure 3. Perhaps a better explanation for using tanh1 () is as follows: note that 0 ( + 1)/2 1 is a proportion, that taking the

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES

455

logit transformation of a proportion is common practice (i.e., the basis of logistic regression), and that one can show that tanh1 () = 0.5logit(( + 1)/2).

goal for robust quality experiments is to maximize Pnoise (Y S | xcontrol ) = P (Y S | (xcontrol , xnoise )dF (xnoise ), (7) where Pnoise indicates that the variation in the noise factors is accounted for.

Experiments and Their Goals

Statistically designed experiments provide proactive means for identifying the important factors that impact quality and for determining factor levels that lead to quality gains. Fractional factorial designs are used when a large number of factors need to be studied in a relatively small number of runs. For example, to study factors F1 , . . . , Fr , each at two levels, a 2rp fractional factorial design might be used. In order to achieve robust quality in Taguchis robust parameter design paradigm, one needs to nd the levels of the control factors (which are controllable) so that the quality is insensitive to the variation of so-called noise factors (which are hard or impossible to control). Crossed array designs are often used in robust quality experiments. A crossed array consists of a control array, the array of settings for control factors F1 , , Fr , and a noise array, the array of settings for noise factors N1 , , Ns . In the experiment, each noise factor setting is performed with each control factor setting. The form of the covariate vector x depends on the type of experiment. For fractional factorial designs, the x consists of covariates corresponding to an intercept, main eects, and two-factor interactions involving the factors F1 , . . . , Fr . The crossed array design allows an intercept, all Control main eects (and possibly some Control Control interactions), all Control N oise interactions, and all N oise main eects to be estimated; Control and N oise refer to the control factors F1 , , Fr and noise factors N1 , , Ns , respectively. For quality improvement, the experimental goal is to maximize the proportion of conformance in Equation (1), which is the proportion of products which simultaneously meet all specications. For robust quality improvement, the experimental goal is to maximize the proportion of conformance in Equation (1) while accounting for variation in the noise factors. To do this, one must rst identify the distributional parameter models in Equations (3)(5) for the multivariate normal response and use them to nd the factor settings that accomplish the goal. Thus, the proportion of conformance depends on the factor levels x, which is expressed as P (Y S | x). (6)

Analysis Methodology for Multiple Responses

Consider experiments in which there are n replicates at each experimental run, i.e., each factor setting in the experimental design. We propose a simple method in which one rst estimates the distributional parameters of the multivariate normal response, and , at each run and then identies the experimental factor models (3)-(5). We use sample means, variances, and correlations to estimate and as follows:
n

Yki k =
n 2 k i=1

(Yki k )2 n1 ,

i=1

and
n n

(Yki k )(Ylj l )
n

k =

i=1 j=1 n i=1

, (Yli l )2

(Yki k

)2
j=1

1 k < l m, where Yki is the ith replicate of the k th response for a given run, i =1, , n. The experimental factor 2 models (3)-(5) can then be t using the (k , k , k ) to obtain least square estimates (LSEs) for the factorial eects given in these models. Standard methods for analyzing unreplicated experiments apply here, such as normal or half-normal plots to determine the signicance of the eects (Daniel (1959)). Since estimating variances and correlations is much harder than estimating means, concern arises when the number of replicates in an experiment is small. Hamada and Balakrishnan (1998) study formal tests of eect signicance. It turns out that log 2 and tanh1 are approximately normal with constant variance: 2/(n 1) for log 2 (Bartlett and Kendall (1946)) and 1/(n 3) for tanh1

The control and noise factors in a robust quality experiment are denoted by x = (xcontrol , xnoise ), so the

Vol. 33, No. 4, October 2001

www.asq.org

456

CHIH-HUA CHIAO AND MICHAEL HAMADA TABLE 1. Factors and Levels for Wheel Cover Component Experiment

Level Factor F1 : F2 : F3 : F4 : F5 : F6 : F7 : mold temperature close time booster time plunger time pack pressure hold pressure barrel 1 80 16 1.88 4 1300 1100 490 +1 110 21 1.7 8 1425 700 505

here is that the number of replicates may be too small. Checking for multivariate normality is illustrated, however, in the third example in the next section.

Examples
The proposed methodology is illustrated with three experiments from the literature. The rst two focus on quality improvement. The third experiment considers robust quality improvement. Example 1: Wheel Cover Component Experiment Harper, Kosbe, and Peyton (1987) report on an experiment to nd the optimum combination of injection molding parameters to minimize the imbalance of a plastic wheel cover component. Seven factors thought to be potentially important to the components balance are listed in Table 1 as well as the levels for each factor denoted by 1 and +1. The relevant quality characteristics of the wheel cover component are the total weight (Y1 in grams) and the balance (Y2 in inch-ounces) of the component. The two-sided specication region for Y1 and Y2 is dened by (710, 715) and (0.3, 0.4), respectively. The data for eight runs of a 274 fractional factorial design are shown in Table 2 in which ve components were measured in each run (Box, Hunter, and Hunter (1978)). Only main eects of the seven factors can be studied in this experiment. Estimates for , which includes all main eects {p , p , p } for p = 1, , 7 and their intercepts, respectively, were obtained using the least squares method. The least square estimates (LSEs) of the effects have standard deviations of 2/[m(n 1)] for log 2 and 1/[m(n 3)] for tanh1 , where m = 8 and n = 5. Using these standard deviations for the LSEs of the eects for k (k=1,2) and tanh1 and Lenths approach for k (k=1,2), we identify the following experimental factor models: 1 = 720.763 + 1.873x1 + 5.318x5 3.408x7 , 2 = 0.967 + 0.113x1 + 0.328x5 0.174x7 ,
2 log 1 = 0.944 0.509x2 + 1.189x4

(Fisher (1970)). For a two-level, four-factor, full factorial design (24 ), the number of runs m is 16. The LSEs of the eects then have a standard deviation of 2/[m(n 1)] for log 2 and 1/[m(n 3)] for tanh1 . The results of Hamada and Balakrishnan (1998) for the Lenth method (Lenth (1989)) indicate that the power of detecting an eect one and a half times the standard deviation is about 0.70 and that the power is below 0.40 for an eect equal to the standard deviation. The fact that the transformed sample variances and correlations are approximately normal with constant and known variances allows a simple formal test: since the LSEs are normal, the LSEs divided by their standard deviations are normal and can be compared with 1.96 for a 5% test. The power of such tests is easily computed: 0.99 for an eect one and a half times the standard deviation and 0.81 for an eect equal to the standard deviation. Thus, these simple tests are much more powerful than the Lenth method, and the power increases for larger designs with run size m, say 32 runs. Also, smaller eects can be detected for more replicates n since the standard deviation of log 2 is 2/(n 1) and that of tanh1 is 1/(n 3). Based on the estimates of the signicant eects, factor settings can then be found that maximize the proportions of conformance in Equations (6) and (7) in quality improvement and robust quality improvement experiments, respectively. When there are three or more responses, the predicted variancecovariance matrix needs to be checked for positive deniteness, i.e. all its eigenvalues are positive so Equations (6) and (7) can be evaluated. As with any data analysis, it is worth checking model assumptions. Andrews et al. (1971, 1973) provides a comprehensive review of tests for multivariate normality. The limitation for the experiments discussed
Journal of Quality Technology

+ 1.196x5 0.487x7 , (8) 2 log 2 = 4.797 0.692x2 , 1 tanh (12 ) = 0.116 + 0.490x1 0.505x5 , where x1 , x2 , x4 , x5 and x7 are the covariates corresponding the ve signicant factors.
Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES TABLE 2. Experimental Design and Bivariate Responses for Wheel Cover Component Experiment

457

Run 1

F1 1 1 1 1

F2 1 1

F3 1 1

F4 1

F5 1

F6 1

F7 1

Y Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 711.9 .59 725.0 .70 711.6 .56 733.7 1.50 725.4 1.25 728.7 1.17 726.6 1.52 714.3 .57 713.4 .59 720.1 .91 711.7 .44 724.1 1.55 721.6 1.36 721.1 .97 731.4 1.58 714.4 .51

Replicate 712.3 .47 711.8 1.13 711.3 .46 732.0 1.38 722.6 1.51 722.9 .98 731.4 1.61 713.6 .44 712.4 .71 723.9 .79 712.1 .53 732.7 1.45 723.1 1.22 723.0 .97 729.6 1.40 716.3 .44 711.9 .63 720.9 .78 711.7 .46 733.3 1.45 721.1 1.25 719.7 .73 731.3 1.57 714.6 .56

+1 1

+1 1

+1

+1

+1

+1

+1 1 1

+1 1

+1 1 1

+1

+1 1

+1

+1

+1

+1 1

+1 1 1

+1

+1 1 1

+1 1

+1

+1

+1

+1 1

+1 1

+1

+1

+1

+1

Using the experimental factor models, the proportion of conformance in Equation (6) can be evaluated for all 128 (= 27 ) possible factor settings, with each factor taking on levels 1 and +1. We use the IMSL FORTRAN subroutine bnrdf to evaluate the proportion of conformance (IMSL 1999). Based on values of the proportion of conformance, the optimal setting is xopt = (1, +1, , 1, 1, , +1), where an insignicant factor is denoted by . That is, the optimal setting is level 1 for F1 , F4 , and F5 and level +1 for F2 and F7 , for which (1 , 2 , 1 , 2 , ) = (710.16, 0.35, 0.30, 0.06, 0.10). The proportion of conformance at the optimal setting is 0.515. This proportion of comformance seems low and suggests

that exploration outside the current experimental region be considered for further improvement. To illustrate the potential benets of the proposed procedure, consider what would result from using two existing approaches which evaluate (i) only the means and (ii) only the means and variances. The best three settings are presented in Table 3. The rank vector orders the settings by proportion of conformance for approaches (i) and (ii). (i) Identify signicant factors using the mean model (3), and nd the optimal settings whose means are the nearest to the center of the specication region. We use the experimental factor

TABLE 3. Best Settings for Wheel Cover Component Experiment

Setting 1 2 3

F1 1 +1 +1

F2 +1 1 +1

F4 1 1 1

F5 1 1 1

F7 +1 +1 +1

P (Y S|x) 0.515 0.068 0.003

1 710.16 713.91 713.91

2 0.35 0.58 0.58

1 0.30 0.49 0.30

2 0.06 0.13 0.06

0.10 0.71 0.71

Rank (1,2,4) (7,1,2) (9,1,1)

Vol. 33, No. 4, October 2001

www.asq.org

458

CHIH-HUA CHIAO AND MICHAEL HAMADA

models for the means in Equation (8) to evaluate (1 , 2 ) for all possible factor settings. The setting x = (+1, 1, 1, 1, 1, 1, +1) produces the minimum squared distance between (1 , 2 ) and the center of the specication region ((710+715)/2, (0.3+0.4)/2), which is 2.04; however, its proportion of conformance is only 0.068 based on all of the experimental factor models in Equation (8), which is much smaller than the 0.515 for the optimal setting found by the proposed procedure. (ii) Identify signicant factors using the mean and variance models ((3) and (4)), and nd the optimal settings whose means are the nearest to the center of the specication region with minimum variance. We use the experimental factor models for the means and variances in Equa2 2 tion (8) to evaluate (1 , 2 , 1 , 2 ) for all possible factor settings. The criterion used to evaluate the settings is the sum of the squared distance given in (i) and the two variances 2 (k , k = 1, 2). The optimal setting using this criterion is x = (+1, +1, +1, 1, 1, +1, +1) whose value is 2.13. The corresponding proportion of conformance is only 0.003 based on all of the experimental factor models in Equation (8), which is again much smaller than the 0.515 for the optimal setting found by the proposed procedure. Example 2: Plasma Enhanced Chemical Vapor Deposition Experiment Tong and Su (1997) report on an experiment to improve a plasma enhanced chemical vapor deposition (PECVD) process in the fabrication of integrated circuits. In this experiment, eight factors are selected, the names and corresponding levels of which are given in Table 4.

The two quality characteristics are the deposition thickness (Y1 ) in Angstroms (A) and a refractive index (Y2 ). The specications are (950, 1050) for Y1 and (1.90, 2.10) for Y2 . An L18 orthogonal array is used to accommodate the one two-level factor and seven three-level factors studied in the experiment, and each of the 18 runs was replicated ve times (Taguchi (1986)). The experimental design and bivariate response data are displayed in Table 5. The experimental factor models in Equations (3)(5) consisting of the main eects are t, and their estimates appear in Table 6. Note that there are two estimates for each of the three-level factors corresponding to linear and quadratic eects which appear in this order. The covariates in models (3)-(5) for threelevel factors corresponding to levels 1-3 are (1, 0, 1) for the linear eect and (1, 2, 1) for the quadratic eect. Here, the standard deviations of the LSEs are 2/[m(n 1)] for log 2 and 1/[m(n 3)] for tanh1 , where m = 18 for the two-level factor effect, m = 12 for the three-level factor linear eect, and m = 36 for the three-level factor quadratic effect. The signicant eects are identied by asterisks in Table 6. Using the experimental factor models in Equations (3)(5) with the estimates of signicant eects given in Table 6, the proportion of conformance in Equation (6) can be evaluated at all 4,374(= 2 37 ) factor settings. The ve best settings appear in Table 7, and xopt = (2, 1, 3, 1, 2, 2, 2, 3) is the optimum setting; i.e., level 1 for F2 and F4 , level 2 for F1 , F5 , F6 , and F7 , and level 3 for F3 and F8 . At this optimal setting the proportion of conformance is 0.719. The original setting, xoriginal = (2, 1, 2, 2, 2, 2, 2, 2), has a proportion of conformance of 0.530, so the proposed method can produce an improvement of more than 36%.

TABLE 4. Factors and Levels for Chemical Vapor Deposition Experiment

Factor F1 : F2 : F3 : F4 : F5 : F6 : F7 : F8 : cleaning method chamber temperature batch after cleaned chamber ow rate of SiH4 ow rate of N2 chamber pressure R. F. power deposition time

1 No 100 C 1st 6% 30% 160 mtorr 30 watt 11.5 min

Level 2 Yes 200 C 2nd 7% 35% 190 mtorr 35 watt 12.5 min

3 300 C 3rd 8% 40% 220 mtorr 40 watt 13.5 min

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES TABLE 5. Experimental Design and Bivariate Responses for Chemical Vapor Deposition Experiment

459

Run 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

F1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2

F2 1 1 1 2 2 2 3 3 3 1 1 1 2 2 2 3 3 3

F3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3 1 2 3

F4 1 2 3 1 2 3 2 3 1 3 1 2 2 3 1 3 1 2

F5 1 2 3 2 3 1 1 2 3 3 1 2 3 1 2 2 3 1

F6 1 2 3 2 3 1 3 1 2 2 3 1 1 2 3 3 1 2

F7 1 2 3 3 1 2 2 3 1 2 3 1 3 1 2 1 2 3

F8 1 2 3 3 1 2 3 1 2 1 2 3 2 3 1 2 3 1

Y Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 Y1 Y2 694 2.118 918 2.205 936 2.677 765 2.096 709 2.032 795 1.860 711 2.012 580 1.834 590 1.719 917 2.097 1389 1.927 865 1.963 827 1.903 787 2.103 739 2.182 724 2.274 771 1.942 712 2.077 839 1.919 867 2.240 954 2.643 828 1.997 743 2.007 785 1.838 816 1.909 644 1.760 812 1.707 1142 1.911 1405 1.860 914 1.881 884 1.829 805 2.020 779 2.080 721 2.166 806 1.905 781 1.961

Replicate 728 1.985 861 2.234 930 2.714 842 1.949 753 1.943 846 1.842 1085 1.797 602 1.760 627 1.676 1126 1.889 1219 1.945 993 1.812 884 1.788 780 2.011 745 2.071 690 2.215 785 1.909 749 1.985 688 2.085 874 2.165 1058 2.456 768 2.046 752 2.003 722 1.999 787 1.930 607 1.782 595 1.704 916 2.014 2063 1.539 838 1.923 851 1.863 776 2.107 724 2.179 1023 2.103 869 1.916 692 2.101 704 2.056 851 2.275 958 2.565 801 2.000 989 1.845 833 1.858 1150 1.819 811 1.744 609 1.675 966 1.960 1392 1.867 893 1.899 1066 1.767 976 1.968 976 1.968 915 2.203 859 1.900 760 1.980

Vol. 33, No. 4, October 2001

www.asq.org

460

CHIH-HUA CHIAO AND MICHAEL HAMADA TABLE 6. Parameter Estimates for Chemical Vapor Deposition Experiment

Parameter 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8

Y1 859.011 63.078 (165.345 , 50.636) (38.844, 60.121) (25.597, 7.974) (42.073, 65.823) (114.363, 52.147) (99.608, 10.233) (59.845, 76.247) 8.716 0.305 (0.183, 0.141) (0.593 , 0.161) (0.021, 0.276 ) (0.199, 0.447 ) (1.040 , 0.839 ) (0.224, 0.239 ) (0.302, 0.350 )

Y2 1.994 0.0169 (0.126 , 0.027) (0.031, 0.044) (0.105 , 0.027) (0.035, 0.049) (0.148 , 0.002) (0.030, 0.004) (0.060, 0.064) 5.642 0.108 (0.987 , 0.260 ) (0.229, 0.368 ) (0.307, 0.069) (0.769 , 0.232 ) (0.982 , 0.454 ) (0.356, 0.164) (0.300, 0.140) 1.288 0.096 (0.358, 0.036) (0.031, 0.134) (0.073, 0.271 ) (0.467 , 0.208) (0.370, 0.235 ) (0.434 , 0.179) (0.106, 0.339 )

Intercept F1 F2 F3 F4 F5 F6 F7 F8 Intercept F1 F2 F3 F4 F5 F6 F7 F8 Intercept F1 F2 F3 F4 F5 F6 F7 F8

indicates signicant eects

To illustrate why it is necessary to model distributional parameters in terms of the experimental factors, observe what happens when one assumes a constant variance-covariance matrix for this example. Using the models for the means in (3) and optimizing the proportion of conformance for a constant

variance-covariance matrix, the optimal setting is (1, 1, 1, 1, 3, 2, 2, 3). The actual proportion of conformance for this setting (evaluated under all of the experimental factor models (3)(5)) is only 0.502, even though the location (1 = 975.93 and 2 = 2.01) is close to the middle of the specication region. Com-

TABLE 7. Best Settings for Chemical Vapor Deposition Experiment

Rank 1 2 3 4 5

F1 2 1 1 2 2

F2 1 1 1 1 1

F3 3 3 3 3 2

F4 1 1 1 1 1

F5 2 2 2 2 2

F6 2 2 2 2 2

F7 2 2 3 3 2

F8 3 1 1 3 3

P (Y S|x) 0.719 0.719 0.691 0.691 0.690

1 975.93 975.93 975.93 975.93 975.93

2 2.01 2.01 2.01 2.01 2.01

1 33.47 33.47 38.74 38.74 41.27

2 0.07 0.07 0.07 0.07 0.05

0.81 0.81 0.89 0.89 0.81

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES TABLE 8. Cross Array Design for the LED Experiment

461

the degradation paths stabilize by then: R(t) = L(t3 ) L(t) , L(t3 ) (9)

Noise Array Control Array F1 F2 F3 1 +1 1 +1 1 +1 1 +1 1 1 +1 +1 1 1 +1 +1 1 1 1 1 +1 +1 +1 +1 N 1 +1

Run 1 2 3 4 5 6 7 8

where L(t) denotes the luminosity in lumens at time t. To illustrate the proposed methodology for making quality robust, consider a bivariate response consisting of Y1 = R(500) and Y2 = R(1000) for R(), which is dened in Equation (9). Scatter plots of (Y1 , Y2 ) for the thirty LEDs at each of the sixteen factor level combinations in the experiment are displayed in Figure 4. For purposes of illustration, the specications for (Y1 , Y2 ) are taken to be (0.0, 0.05) and (0.0, 0.15), although the ideal R(t) would be zero for all t. The specications for (Y1 , Y2 ) are based on the expectation that the degradation rates should be below 5% and 15% at the rst 500 and 1,000 hours of operation, respectively. Before tting the experimental factor models in Equations (3)(5), consider a simple procedure to check the bivariate normality assumption. For each LED in a run, dene the Mahalanobis distance as d2 = (Y )T (Y ), (10)

See Figure 4 for Bivariate Response Plots

pare this with the value 0.719 for the optimal setting found with the proposed method. Example 3: LED Experiment An experiment is performed to improve the reliability of light emitting diodes (LEDs) and make reliability robust. The luminosity of an LED degrades over time, and the slower its degradation, the higher the reliability. An LED is composed of a die, a solidstate component, which is attached by silver epoxy to the lead frame which carries the electric current. The control factors chosen for the experiment are die vendor (F1 ), type of epoxy material (F2 ), and lead frame design (F3 ). In the manufacturing process, the location on the lead frame where the die is attached varies. One goal considered in Chiao and Hamada (1996) for this experiment is how to make the LED reliability robust to the die location variation; hence, the die location (F actor N ) is a noise factor. The experimental design used is a crossed array consisting of an eight-run control array and a two-run noise array as displayed in Table 8 (Taguchi (1986)). Note that all factors are studied at two levels (denoted by 1 and +1), giving eight dierent LED designs (as specied by the eight-run control factor array), each of which uses two dierent die locations (as specied by the two levels in the noise factor array); for the experiment, the die location needs to be specially controlled so that information about die location can be obtained. This yields 16 dierent control and noise factor level combinations performed in the experiment in which a sample of thirty LEDs is tested. In the experiment, the luminosities of the LEDs are inspected ten times. The observed relative degradation rate R(t) is dened in terms of the luminosity at the third inspection (at t3 = 192 hours) because
Vol. 33, No. 4, October 2001

where and are estimated from the replicates from the run. Then d2 follows approximately a 2 distribution with m degrees of freedom for an m dimensional multivariate normal response. Thus, a chi-squared probability plot of the Mahalanobis distances should be a straight line if the multivariate normal assumption holds. Figure 5 displays the chisquared probability plot (with 2 degrees of freedom) of the Mahalanobis distances for the LED experiment and suggests that the bivariate normal assumption is reasonable. Note that a number of replicates are required for each run in order to estimate (, ) needed in Equation (10). For a bivariate response there are ve parameters, so at least ve replicates are required. Next, we t the experimental factor models in Equations (3)(5) for the LED experiment. The covariates used correspond to three Control main eects, three Control Control interactions, one N oise main eect, and three Control N oise interactions. The LSEs of the eects for log 2 have a standard deviation of 2/[m(n 1)] and those for tanh1 have a standard deviation of 1/[m(n 3)], where m = 16 and n = 30. The halfnormal plots (Daniel, 1959) given in Figure 6 are used to identify the following experimental factor models (i.e., those eects whose absolute estimates fall above the line through the smallest absolute estimates
www.asq.org

462

CHIH-HUA CHIAO AND MICHAEL HAMADA

FIGURE 4. Scatter Plots of Bivariate Responses For LED Experiment.

FIGURE 5. Mahalanobis Distance Plot For LED Experiment.

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES

463

FIGURE 6. Half-Normal Plots For LED Experiment.

Vol. 33, No. 4, October 2001

www.asq.org

464

CHIH-HUA CHIAO AND MICHAEL HAMADA TABLE 9. Best Settings for LED Experiment

Run F1 1 2 3 4 5 6 7 8 1 +1 1 +1 1 +1 1 +1

F2 1 1 +1 +1 1 1 +1 +1

F3 Pnoise (Y S|xcontrol ) Rank 1 1 1 1 +1 +1 +1 +1 0.928 0.286 0.028 0.002 0.952 0.386 0.582 0.008 2 5 6 8 1 4 3 7

of the multivariate response distribution are modeled in terms of experimental factors. Once such models are identied, optimal settings can then be determined using a particular criterion; in this paper, we have used the proportion of conformance, or the probability that all responses simultaneously meet their respective specications. Other criteria might be considered. It is important that practitioners be aware of the methods limitations. This method depends on the normality assumption which can be checked by the chi-squared probability plot of Mahalanobis distances. Given these concerns, it is important to run conrmation experiments at the recommended settings. With the tting of multiple models and potentially poor parameter estimation, it is prudent to check whether or not the predictions at those settings are valid.

corresponding to zero eects): 1 = 0.077 + 0.027x1 + 0.032x2 0.023x3 + 0.018x1 x2 0.021x2 x3 0.019xN , 2 = 0.162 + 0.041x1 + 0.062x2 0.043x3 0.043x2 x3 0.042xN ,
2 log(1 )

Acknowledgments
C. H. Chiao was a Visiting Fellow at the University of Michigan when research for this paper was performed and is grateful for nancial support from the Ministry of Education of Taiwan. M. Hamadas research was supported by NSF grant DMS-9704649. The authors thank the referees whose insightful comments helped to improve a previous version of this paper.

= 8.385 + 0.483x1 + 1.363x2 0.481x3 0.571x2 x3 + 0.528xN , = 7.670 + 1.414x2 0.569x2 x3 + 1.017xN ,

2 log(2 )

tanh

(12 ) = 1.197 + 0.471x2 ,

where x1 , x2 , x3 are the covariates corresponding to the three control factors and xN is the covariate corresponding to the noise factor. Based on these identied models, the proportion of conformance in Equation (7) for each setting specied in the control array is computed. Here, for purposes of illustration, we assume that the distribution of the noise factor is N (0, (1/3)2 ) so that the bulk of the distribution is within the experimental region [1, 1]. Table 9 presents the proportions of conformance for each control array setting, and it shows that the optimum settings are (1, 1, +1) for factors F1 , F2 , and F3 , respectively; that is, run 5 has the largest performance of conformance with a value of 0.952.

References
Andrews, D. F.; Gnanadesikan, R.; and Warner, J. L. (1971). Transformations of Multivariate Data. Biometrics 27, pp. 825840. Andrews, D. F.; Gnanadesikan, R.; and Warner, J. L. (1973). Methods for Assessing Multivariate Normality in Multivariate Analysis 3 edited by P. R. Krishnaiah. John Wiley and Sons, New York, NY, pp. 825840. Bartlett, M. S. and Kendall, D. G. (1946). The Statistical Analysis of Variance-Heterogeneity and the Logarithmic Transformation. Journal of the Royal Statistical Society B 8, pp. 128138. Box, G. E. P.; Hunter, W. G.; and Hunter, J. S. (1978). Statistics for Experimenters. John Wiley and Sons, New York, NY. Chiao, C. H. and Hamada, M. (1996). Using Degradation Data from an Experiment to Achieve Robust Reliability for Light Emitting Diodes. Quality and Reliability Engineering International 12, pp. 8994. Daniel, C. (1959). Use of Half-Normal Plots in Interpreting Factorial Two-Level Experiments. Technometrics 1, pp. 311340. Derringer, G. and Suich, R. (1980). Simultaneous Optimization of Several Response Variables. Journal of Quality Technology 12, pp. 214219.

Conclusions
In this paper, we have proposed a simple method for improving quality and for achieving robust quality using statistically designed experiments with multiple responses. While the experimental plans are the same ones used for improving any quality characteristic, it is the simultaneous analysis of multiple responses that requires new consideration. Both the mean vector and the variance-covariance matrix

Journal of Quality Technology

Vol. 33, No. 4, October 2001

ANALYZING EXPERIMENTS WITH CORRELATED MULTIPLE RESPONSES

Fisher, R. A. (1970). Statistical Methods for Research Workers. Hafner Press, New York, NY. Hamada, M. and Balakrishnan, N. (1998). Analyzing Unreplicated Factorial Experiments: A Review with Some New Proposals. Statistica Sinica 8, pp. 141. Harper, D.; Kosbe, M.; and Peyton, L. (1987). Optimization of Ford Taurus Wheel Cover Balance (by Design of ExperimentsTaguchis Method). Fifth Symposium on Taguchi Methods, Romulus, MI: American Supplier Institute, pp. 527539. IMSL (1999). IMSL Fortran Subroutines for Statistical Applications Manual, Stat/Library, Visual Numerics. Khuri, A. I. (1996). Multiresponse Surface Methodology. in Handbook of Statistics: Design and Analysis of Experiments 13. edited by S. Ghosh and C. R. Rao, Elsevier Science B. V., New York, NY, pp. 377406. Khuri, A. I. and Conlon, M. (1981). Simultaneous Optimization of Multiple Responses Represented by Polynomial Regression Functions. Technometrics 23, pp. 363375. Lenth, R. V. (1989). Quick and Easy Analysis of Unreplicated Factorials. Technometrics 31, pp. 469473. Pignatiello, J. J., Jr. (1993). Strategies for Robust Mul-

465

tiresponse Quality Engineering. IIE Transactions 25, pp. 5 15. Rao, C. R. (1973). Linear Statistical Inference and Its Application. John Wiley and Sons, New York, NY. Symposia on Taguchi Methods (1984-1993). Romulus, MI: American Supplier Institute, Taguchi, G. (1986). Introduction to Quality Engineering. Asian Productivity Organization, Tokyo. Tong, L. and Su, C. (1997). Optimizing Multiresponse Problems in the Taguchi Method by Fuzzy Multiple Attribute Decision Making. Quality and Reliability Engineering International 13, pp. 2534. Vining, G. G. (1998). A Compromise Approach to Multiresponse Optimization. Journal of Quality Technology 30, pp. 309313. Wang, C. M. and Lam, C. T. (1996). Condence Limits for Proportion of Conformance. Journal of Quality Technology 28, pp. 439445.

Key Words: Factorial Designs, Fractional Factorial Designs, Multivariate Normal Distribution, Robust Parameter Design.

Vol. 33, No. 4, October 2001

www.asq.org