Personal.soton.ac.uk

Open Journal of Statistics, 2012, 2, 48-59
doi:10.4236/ojs.2012.21006 Published Online January 2012 (http://www.SciRP.org/journal/ojs)
Minimum MSE Weights of Adjusted Summary
Estimator of Risk Difference in Multi-Center Studies
Chukiat Viwatwongkasem1*, Jirawan Jitthavech2, Dankmar Böhning3, Vichit Lorchirachoonkul2
1Department of Biostatistics, Faculty of Public Health, Mahidol University, Bangkok, Thailand 2School of Applied Statistics, National Institute of Development Administration, Bangkok, Thailand 3Applied Statistics, School of Biological Sciences, University of Reading, Reading, UK Received October 14, 2011; received November 18, 2011; accepted November 30, 2011 ABSTRACT
The simple adjusted estimator of risk difference in each center is easy constructed by adding a value c on the number of
successes and on the number of failures in each arm of the proportion estimator. Assessing a treatment effect in
multi-center studies, we propose minimum MSE (mean square error) weights of an adjusted summary estimate of risk
difference under the assumption of a constant of common risk difference over all centers. To evaluate the performance
of the proposed weights, we compare not only in terms of estimation based on bias, variance, and MSE with two other
conventional weights, such as the Cochran-Mantel-Haenszel weights and the inverse variance (weighted least square)
weights, but also we compare the potential tests based on the type I error probability and the power of test in a variety
of situations. The results illustrate that the proposed weights in terms of point estimation and hypothesis testing perform
well and should be recommended to use as an alternative choice. Finally, two applications are illustrated for the practi-
cal use.
Keywords: Minimum MSE Weights; Optimal Weights; Cochran-Mantel-Haenszel Weights; Inverse Variance Weights;
1. Introduction
V  p   n p 1 p  n  2c . They concluded that the 1 n  2 minimizes the Bayes risk (the It is widely known that the conventional proportion esti- average MSE of ˆp ) in the class of all estimators of the mator, ˆp  X n , is a maximum likelihood estimator form  X  c n  2c with respect to uniform prior on (MLE) and an uniformly minimum variance unbiased [0,1] and Euclidean loss function; furthermore, the esti- estimator (UMVUE) for the binomial parameter p 1 n  2 has smaller MSE than X n in where the binomial random variable X is the number the approximate interval 0.15, 0.85 of p . For another of successes out of the number of patients n . However, argumentation in the Bayesian approach, Casella and Agresti and Coull [1], Agresti and Caffo [2], Ghosh [3], Berger [7] showed that  X    n      is a Bayes and Newcombe [4,5] highlighted the point that ˆp might estimator of p under the conditional binomial sampling not be a good choice for p when the assumption of n p and the prior beta distribution np  5 and n 1 ˆp  5 was violated; this violation p ~ beta ,   . Note that in case of     1 the beta often occurs when the sample size n is small, or the distribution has a special case as the uniform distribution estimated probability ˆp is close to 0 or 1 (close to the over [0,1]. Consequently, the estimator  X  c n  2c boundaries of parameter space), leading to the problem derived from the Bayesian approach and the Bayes risk of the zero estimate of the variance of ˆp . The estimated approach under the above mentioned criteria provides the variance of ˆp , provided by V ( ˆp)  ˆp 1 ˆp n , is zero in the occurrence of any case: X  0 or X  n . Böh- With the idea of ˆp   X  c n  2c , the extension ning and Viwatwongkasem [6] proposed the simple ad-   ˆp  ˆp , the adjusted risk difference esti- justed proportion estimator by adding a value c on the mator between two independent binomial proportions, number of successes and the number of failures; cones- for estimating a common risk difference  where quently, ˆp   X  c n  2c is their proposed esti- mate of p with the non-zero variance estimate are proportion estimators for treatment and control arms. In a multi-center study of size k , the parameter of in- Copyright 2012 SciRes. OJS
terest is also a common risk difference  that is as- To obtain the minimum Q subject to a constraint sumed to be a constant across centers. We concern about  f 1, we form the auxiliary function  to seek a combination of several adjusted risk difference estima-   ˆp  ˆp from the th j center  j  1, 2,, k  into the adjusted summary estimator of risk difference of   E  f       f 1 f  where f are the weights where  is a Lagrange multiplier. The weights f we would propose the optimal weights f as an alter- and  are derived by solving the following equations native choice based on minimizing the MSE of ˆ Section 2, then state the well-known candidates such as the Cochran-Mantel-Haenszel (CMH) weights and the inverse variance (INV) weights in Section 3. A simula- details are presented in Appendix. The result of the tion plan for comparing the performance among weights in terms of estimation and hypothesis testing is presented in Section 4. The results of the comparison among the potential estimators based on bias, variance, and MSE and also the evaluations among tests related the men- tioned weights through the type I error probability and the power criteria lie on Section 5. Some numerical ex- amples are applied in Section 6. Finally, conclusion and 2. Deriving Minimum MSE Weights of
Adjusted Summary Estimator
Under the assumption of a constant of common risk dif- ference  across k centers, we combine several ad-justed risk difference estimators ˆ j  1, 2,, k  arrive at an adjusted summary estimator f are non-random weights subject to the constraint  f 1. Please observe that for a single center ˆp2 In the particular case of c  c  0 , our estimator  f 1 is a shrinkage estimator of a   ˆp  ˆp . Our minimum popular inverse-variance weighted estimator. Under a MSE weights f of the adjusted summary estimator common risk difference  over all centers, the variance  were derived by following Lagrange’s method [8]  in the case of non-random weights f are ob- under the assumption of a constant of common risk dif- ference over all centers with the pooling point estimator to estimate  . Lui and Chang [9] proposed the optimal weights proportional to the reciprocal of the variance with the Mantel-Haenszel point estimator under the as- sumption of noncompliance. It was observed that both of optimal weights provided the different formulae because of different assumptions even though they were derived from the same method of Lagrange. Now, we wish to Suppose that a normal approximation is reliable, the present the proposed weights minimizing the MSE of  E    E  f    Copyright 2012 SciRes. OJS
for testing H :   we have the normal approximate test der the null hypothesis of H : OR  1   0 . With the null criterion, Mantel-Haenszel’s weight stated by Sanchez-Meca and Marin-Martine [13] was equivalent to n  n 1 . Since the minor difference between the conditional Mantel-Haenszel weight and the We will reject H at  level for two-sided test if unconditional Cochran weight is in the denominators, thus the two are often referred to interchangeably as the percentile of the standard normal distribution. Alterna- Cochran-Mantel-Haenszel weight. In this study, we use tively, H is rejected when the p-value ( p ) is less than or equal to   p    where p  2 1 and   Z  is the standard cumulative normal distribu- 3.2. Inverse Variance (INV) or Weighted Least
Square (WLS) Weights
3. Other Well-Known Weights
Fleiss [14] and Lipsitz et al. [15] showed that the in- verse-variance weighted (INV) estimator or the weighted- 3.1. Cochran-Mantel-Haenszel (CMH) Weights
least-square (WLS) estimator for  was in the summary Cochran [10,11] proposed a weighted estimator of cen- estimator of the weighted mean (linear, unbiased estima- ter-specific sample sizes for a common risk difference based on the unconditional binomial likelihood as   ˆp  ˆp  X n  X defined by the reciprocal of the variance as The non-random and non-negative weights w yield widely used as a standard non-random weight derived by the minimum variance of the summary estimator ˆ the harmonic means of the center-specific sample sizes.  w is also Cochran’s weight However, the weights w cannot be used in practice come common practice to replace them by their sample    p 1 p n  p 1 p suming that a normal approximation is reliable, the Cochran’s Z-statistic for testing H :   is provided This weight was suggested in several textbooks of epidemiology such as Kleinbaum et al. [16] or in text- books of meta-analysis such as Petitt [17]. We assume that a normal approximation is reliable; the inverse-variance  j   j 1 j2 weighted test statistic for testing H :   is  H  ˆp 1 ˆp The rejection rule of H follows the same as the previ- Alternatively, Mantel and Haenszel [12] suggested the test based on the conditional hypergeometric likelihood  V  H . Also, the rule of H rejection for a common odds ratio among the set of k tables un- follows the same as the above standard normal test. Copyright 2012 SciRes. OJS
4. Monte Carlo Simulation
bution over 0.1, 0.8 . Binomial random variables X We perform simulations for estimating a common risk are drawn with parameters n , p n , p , respectively. All proposed test statistics are difference  and testing the null hypothesis H :   then computed. The procedure is replicated 5,000 times. Parameters Setting: Let the common risk difference From these replicates, the empirical power 1  of test  be some constants varying from 0 to 0.6, with incre- mental steps of 0.1. Baseline proportion risks p Number of rejections of H when H is true j  1, 2,, k  are generated from a uniform distribution over 0, 0.95   . The correspondent proportion risks 5. Results
 . For example, if   0.2 , then Since it is difficult to present all enormous results from ~ U 0, 0.75 and p  p the simulation study, we just have illustrated some in- sample sizes n and n are varied as 4, 8, 16, 32, stances. Nevertheless, the main results are concluded 100. The number of centers k takes values 1, 2, 4, 8, 16, Statistics: Binomial random variables X and X 5.1. Results for Estimating Risk Differences
in treatment and control arms are generated with pa- Table 1 presents some results according to point estima-
tion of a common risk difference  . However, we can Estimation: All summary estimates of  are com- puted in a variety of different weights. The procedure is  The number of centers, k , can not change the order replicated 5000 times. From these replicates, bias, vari- of the MSE of all weighted estimators, even though ance, and MSE (mean square error) are computed in the an increase in k can decrease the variance and the MSE of all estimators, leading to the increasing effi- Type I Error: From the above parameter setting, we ciency. Also, increasing n and n can decrease assign    under a null H :   , so all tests are the variance of all estimators while fixing k . The computed. The replication is treated 5000 times. From unbalanced cases of n and n for center j have these replicates, the number of the null hypothesis reject- a rare effect on the order of the MSE of all estimates. tions is counted for the empirical type I error  .  For most popular situations used under   0 , Number of rejections of H when H is true cluding adjusted by c  2 is the best choice with the The evaluation for two-sided tests in terms of the type I probability is based on Cochran limits [18] as follow. c  0.5 and the inverse-variance (INV) weighted es- At   0.01 , the  value is between 0.005, 0.015 . timator c  0 are close together and are the second At   0.05 , the  value is between 0.04, 0.06 . choice with smaller MSE. The Cochran-Mantel- At   0.10 , the  value is between 0.08, 0.12 . Haenszel (CMH) weight performs the worst in this simulation setting. This finding is very useful in gen- Cochran limits, then the statistical test can control type I eral situations of most clinical trials and most causal relations between a disease and a suspected risk factor Power of Tests: Before evaluating tests with their since the risk difference is often less than 0.25 [19]. powers, all comparative tests should be calibrated to have  For   0.4 , the proposed estimator ˆ the same type I error rate under H ; then any test whose by c  1 performs best; for   0.5 , the proposed power hits the maximum under H would be the best  adjusted by c  0.5 performs best; test. To achieve the alternative hypothesis, we assume for   0.6 , the INV weighted estimator ( c  0 )   0.1U  0.1 m2U   5.2. Results for Studying Type I Error
where U as an effect of between centers is assigned to m m for a given m 0, 0.  Table 2 presents some results for controlling the empiri-
equivalently, U is an uniform variable over 0,  cal type I error. We can conclude the performance of That is, E    0.1 and Var   m several tests according to the empirical alpha under H we have p    p where p be uniform distri- Copyright 2012 SciRes. OJS
Table 1. Mean, variance, MSE for estimating θ .
Mean: –0.001700 –0.000850 –0.001130 –0.000850 –0.000570 0.171245 0.042811 0.076109 0.042811 0.019027 MSE: 0.171250 0.042813
0.042813 0.019028
Mean: –0.000800 0.000400 –0.000640 –0.000530 –0.000400 0.088874 0.053058 0.056879 0.039499 0.022219 MSE: 0.088875 0.053058 0.056880 0.039500 0.022219
1 8 8 Mean: 0.002625 0.001965 0.002333 0.002100 0.001750 0.042575 0.035480 0.033641 0.027249 0.018923 MSE: 0.042584 0.035483 0.033647 0.027254 0.018926
1 16 16 Mean: –0.000050 0.000328 –0.000047 –0.000044 –0.000040 0.021759 0.020761 0.019275 0.017193 0.013926 MSE: 0.021759 0.020761 0.019275 0.017193 0.013926
Mean: –0.001900 –0.001950 –0.001840 –0.001790 -0.001690 0.010805 0.010674 0.010160 0.009572 0.008538 MSE: 0.010809 0.010678 0.010164 0.009575 0.008540
100 Mean: 0.000566 0.000572 0.000560 0.000555 0.000544 0.003482 0.003478 0.003413 0.003346 0.003219 MSE: 0.003482 0.003478 0.003413 0.003347 0.003219
16 2 2 Mean: 0.102200 0.051100 0.068133 0.051100 0.034067 0.178755 0.044689 0.079446 0.044689 0.019861 MSE: 0.178759 0.047080
0.047080 0.024210
16 4 4 Mean: 0.101900 0.071067 0.081520 0.067933 0.050950 0.093292 0.056358 0.059708 0.041462 0.023323 MSE: 0.093295 0.057194 0.060047 0.042490 0.025729
16 4 8 Mean: 0.091175 0.073915 0.078964 0.069820 0.056883 0.068527 0.048536 0.047903 0.036184 0.023445 MSE: 0.068605 0.049217 0.048345 0.037095 0.025305
16 4 16 Mean: 0.096425 0.086770 0.087330 0.080322 0.069865 0.057752 0.041273 0.040889 0.032469 0.024164 MSE: 0.057764 0.041448 0.041048 0.032856 0.025072
16 4 32 Mean: 0.103087 0.094537 0.095306 0.089488 0.080958 0.052651 0.037007 0.037127 0.030458 0.025400 MSE: 0.052662 0.037037 0.037149 0.030568 0.025763
16 8 8 Mean: 0.105625 0.091604 0.093890 0.084500 0.070417 0.047621 0.041375 0.037626 0.030478 0.021165 MSE: 0.047653 0.041446 0.037664 0.030718 0.022040
16 8 16 Mean: 0.100700 0.094838 0.093524 0.087382 0.077367 0.035620 0.031899 0.029404 0.024987 0.019128 MSE: 0.035620 0.031926 0.029445 0.025147 0.019641
16 8 32 Mean: 0.097381 0.093334 0.092488 0.088258 0.081217 0.028539 0.025407 0.023764 0.020808 0.017542 MSE: 0.028546 0.025452 0.023820 0.020945 0.017895
16 16 Mean: 0.099100 0.094834 0.093271 0.088089 0.079280 0.023792 0.023050 0.021075 0.018798 0.015227 MSE: 0.023793 0.023077 0.021120 0.018941 0.015656
32 32 Mean: 0.100794 0.099611 0.097741 0.094866 0.089594 0.011022 0.010951 0.010364 0.009764 0.008709 MSE: 0.011023 0.010951 0.010369 0.009790 0.008817
100 Mean: 0.100052 0.099934 0.099061 0.098092 0.096204 0.003728 0.003725 0.003654 0.003583 0.003446 MSE: 0.003728 0.003725 0.003655 0.003587 0.003461
Copyright 2012 SciRes. OJS
Table 2. Empirical type I error for testing H : θ = θ at 5% significance level.
0
0
4.16 4.16
16 4.26 4.26 6.54
4.74 4.30
32 4.34 4.34 5.58 4.22 5.10
100 5.02 5.02 6.58
16 4.74 4.74 4.48 4.48 3.38
32 4.50 4.50 4.94 4.44 3.90
100 5.02 5.02 5.30 4.58 5.10
32 5.04 5.04 4.66 4.34 3.88
100 5.22 5.22 5.16 4.46 4.34
100 4.74 4.74 4.60 4.40 4.14
4.56 4.56
16 4.84 4.84 4.66 4.66 3.54
32 4.52 4.52 5.00 4.52 4.10
100 5.46 5.46 5.66 4.72 5.26
32 4.74 4.74 4.42 4.18 3.92
100 5.34 5.34 5.48 4.74 4.46
100 5.04 5.04 4.98 4.86 4.64
8 4.24 4.24 7.6
4.66 4.66
16 5.18 5.18 5.76 5.04 4.06
32 5.66 5.66 5.82 5.40 5.30
100 58.6 5.86 62.0 4.84 4.88
32 5.72 5.72 5.64 4.96 4.44
100 5.88 5.88 5.44 5.20 4.82
100 5.22 5.22 5.16 5.10 4.82
4.36 4.36 8.00
8 4.66 4.66 8.58 5.38 5.38
5.56 5.60
5.40 5.88
32 5.46 5.46 5.40 5.46 5.08
100 5.56 5.56 5.26 5.22 48.8
100 5.34 5.34 5.16 5.10 5.22
16 5.78 5.78 5.92 5.16 7.04
32 5.96 5.96 5.78 5.94 6.28
100 5.92 5.92 5.80 6.04 6.72
100 5.68 5.68 5.34 5.14 5.48
Bold values denote that the statistical tests can control the type I error. Copyright 2012 SciRes. OJS
 The increasing k cannot change the order of the spectively. Also, the estimated standard errors of those of empirical type I error rates of all tests. Also, the un- overall differences are 0.014, 0.013, 0.014, respectively. balanced cases of n and n for center j have a slight effect on the order of the empirical type I error tests at c  1 reject the null hypothesis at 5% level for  None of tests can control type I error rates when sam- two-sided test and lead to the conclusion of a significant ple size of treatment or control arm is very small difference between the placebo and metoprolol mortality  4 ). There exists few tests that can control type I error when sample size is small ( n  8 Turner et al. [21] presented data from clinical trials to  For   0 , almost all tests can control type I error study the effect of selective decontamination of the di- rates when the sample size is moderate to large gestive tract on the risk of respiratory tract infection of patients in intensive care units. See data and weights in curs in practical use of H :  0 . Table 5. The estimated overall differences and their es-
 For   0.2 ,   0.4 , and   0.6 , almost all tests timated standard errors are 0.152 (0.012), 0.140 (0.011), can control type I error rates when the sample size is 0.162 (0.012) for the CMH, the INV, and the proposed large to very large ( n  32 or n weights at c  1, respectively. All tests reject the null hypothesis with Z 5.3. Results for Studying Power of Tests
13.719 and lead to the conclusion of a significant Table 3 shows some more details of the powers. Fortu-
difference between treatment effect of selective decon- nately, almost all tests under H :  0 can control type tamination of the digestive tract on the risk of respiratory I error rates when the sample size is moderate to large parative tests when sample size is very small ( n  4 or 7. Conclusions and Discussion
 4 ) since all of tests can not control type I error In most general situations used by the risk difference rates. The performance of several weighted tests accord- lying on [0, 0.25], the results have confirmed that the ing to the powers under H :  0.1U can be con- minimum MSE weight of the proposed summary esti- adjusted by c  c  c  1 (including The empirical powers yield a similar pattern of results c  c  c  2 ) is the best choice with the smallest MSE like the MSE. An increase in the number of centers, under a constant of common risk difference  over all k , can increase the power but it can not change the k centers. The number of centers, k , cannot change the order of the MSE of all weighted estimators, even  Overall, the proposed weights adjusted by c 1 in- though an increase in k can decrease the variance and cluding c  2 perform best with the highest power the MSE of all weighted estimators. Also, increasing n in a multi-center study of size k  2 when n  16 can decrease the variance of all estimators while fixing k . The unbalanced cases of n and n The INV weight and the CMH weight are achieved for center j have a slight effect on the order of the with the highest powers in one center study when MSE of all estimates. The minimum MSE weight is de- signed to yield more precise estimate relative to the When the sample size is large to very large ( n  32 CMH and INV weights. Another benefit of the proposed weight is easy to compute because of its closed-form formula. With the basis of smallest MSE and the 6. Numerical Examples
easy-to-compute formula, we have been solidly sug- Two examples are presented to illustrate the implementa- gested to use the proposed weight. In addition, the vari- tion of the related methodology. Pocock [20] presented ous choices for c have been considered again. The use data from a randomized trial studying the effect of pla- of c  0.5 as a conventional correction term [22] should cebo and metoprolol on mortality after heart attack (AMI: be revised. The better value of c in adding on the Acute Myocardial Infarction) classified by three strata of number of successes and the number of failures is sug- age groups, namely, 40 - 64, 65 - 69, 70 - 74 years. Ta-
gested with at least for c  1 (including c  2 ). This ble 4 shows the data and weights corresponding to the
result is supported by the ideas of Böhning and Viwat- CMH, the INV, and the proposed strategies. The esti- wongkasem [6], Agresti and Coull [1], and Agresti and mated summary differences based on the CMH, the INV, Caffol [2] that recommended to use the appropriate val- and the proposed weights are 0.031, 0.024, 0.030, re- ues of c greater than or equal to 1. Copyright 2012 SciRes. OJS
Table 3. Empirical power (percent) at m = 0.04 after controlling the estimated type I error at the nominal 5% level.
X X X X 11.2 11.2 10.6 10.6
16.4 16.4 15.4 14.8 14.6
X X X X 17.6 17.6 16.5 16.4
21.4 21.4 21.2 20.8 20.3
36.8 36.8 36.8 36.5 36.1
26.9 29.7
29.5 32.8
33.1 35.2
40.6 43.6
46.8 48.9
62.8 64.6
87.2 87.8
53.9 59.0
64.3 68.5
74.5 77.1
76.9 80.4
89.1 90.4
99.1 99.1
68.3 77.5
77.1 82.1
89.2 92.0
93.8 94.8
99.2 99.3
100.0 100.0 100.0 100.0 100.0
X X 81.8 83.2 92.7 95.6
95.1 96.7
94.5 95.0 97.5
99.9 99.9
99.9 99.9 99.9
100.0 100.0 100.0 100.0 100.0
100.0 100.0 100.0 100.0 100.0
Copyright 2012 SciRes. OJS
Table 4. Mortality data over three strata of age groups following Pocock.
Table 5. Respiratory tract infections following Turner et al.
In terms of type I error estimates, when sample size is when sample size is moderate to large ( n  16 or type I error rates. In addition, there exists few tests that In terms of power, we ignore to evaluate the power can control type I error rates when sample size is small when sample size is very small ( n  4 or n  8 ). This result is consonant with the because all tests can not control type I error rates. The comments of Lui [23] that none of conventional results illustrate the same pattern like the MSE. The pro- tests/weights under sparse data is appropriate. This inap- posed weights adjusted by c  1 including c  2 per- propriateness under sparse data can cope with the mini- form best with the highest power in a multi-center study mum MSE weights from this finding. The further work of size k  2 when n  16 or n to seek some appropriate tests/weights in sparse data weight and the CMH weight are achieved with the high- challenges for investigators to develop an innovation or est powers in one center study when n  16 or to improve much more reasonable tests/weights. In gen- 16 . When sample size is large to very large eral results, almost all tests can control type I error rates Copyright 2012 SciRes. OJS
strongly recommend to use the minimum MSE weight as Common Chi-Square Test,” Biometrics, Vol. 10, No. 4, an appropriate choice because of its highest power. [12] N. Mantel and W. Haenszel, “Statistical Aspects of the 8. Acknowledgements
Analysis of Data from Retrospective Studies of Disease,” Journal of the National Cancer Institute, Vol. 22, 1959, We would like to thank the editors and the referees for comments which greatly improved this paper. This study [13] J. Sanchez-Meca and F. Marin-Martinez, “Testing the was partially supported for publication by the China Significance of a Common Risk Difference in Meta- Medical Board (CMB), Faculty of Public Health, Mahi- Analysis,” Computational Statistics & Data Analysis, Vol. [14] J. L. Fleiss, “Statistical Methods for Rates and Propor- REFERENCES
tions,” John Wiley & Sons Inc., New York, 1981. [1] A. Agresti and B. A. Coull, “Approximate Is Better than [15] S. R. Lipsitz, K. B. G. Dear, N. M. Laird and G. Molen- Exact for Interval Estimation of Binomial Proportions,” berghs, “Tests for Homogeneity of the Risk Difference American Statistical Association, Vol. 52, 1998, pp. 119- When Data Are Sparse,” Biometrics, Vol. 54, No. 1, 1998, [2] A. Agresti and B. Caffo, “Simple and Effective Confi- [16] D. G. Kleinbaum, L. L. Kupper and H. Morgenstern, dence Intervals for Proportions and Differences of Pro- “Epidemiologic Research: Principles and Quantitative portions Result from Adding Two Successes and Two Methods,” Lifetime Learning Publications, Belmont, 1982. Failures,” The American Statistician, Vol. 54, No. 4, 2000, [17] D. B. Petitti, “Meta-Analysis, Decision Analysis and Cost-Effectiveness Analysis: Methods for Quantitative [3] B. K. Ghosh, “A Comparison of Some Approximate Con- Synthesis in Medicine,” Oxford University Press, Oxford, fidence Interval for the Binomial Parameter,” Journal of the American Statistical Association, Vol. 74, No. 368, [18] W. G. Cochran, “The Chi-Square Test of Goodness of Fit,” Annals of Mathematical Statistics, Vol. 23, No. 3, [4] R. G. Newcombe, “Two-Sided Confidence Intervals for the Single Proportion: Comparison of Seven Methods,” [19] K. J. Lui and C. Kelly, “Tests for Homogeneity of the Statistics in Medicine, Vol. 17, No. 8, 1998, pp. 857-872. Risk Ratio in a Series of 2 × 2 Tables,” Statistics in Medi- cine, Vol. 19, No. 21, 2000, pp. 2919-2932. doi:10.1002/1097-0258(20001115)19:21<2919::AID-SI [5] R. G. Newcombe, “Interval Estimation for the Difference between Independent Proportions: Comparison of Eleven [20] S. J. Pocock, “Clinical Trials: A Practical Approach,” Methods,” Statistics in Medicine, Vol. 17, No. 8, 1998, pp. [21] R. Tuner, R. Omar, M. Yang, H. Goldstein and S. Thompson, “A Multilevel Model Framework for Meta- Analysis of Clinical Trials with Binary Outcome,” Statis- [6] D. Böhning and C. Viwatwongkasem, “Revisiting Pro- tics in Medicine, Vol. 19, No. 24, 2000, pp. 3417-3432. portion Estimators,” Statistical Methods in Medical Re- search, Vol. 14, No. 2, 2005, pp. 147-169. [22] F. Yates, “Contingency Tables Involving Small Numbers [7] G. Casella and R. L. Berger, “Statistical Inference,” Dux- and the Chi-Squared Test,” Journal of the Royal Statisti- cal Society (Supplement), Vol. 1, 1934, pp. 217-235. [8] A. E. Taylor and W. R. Mann, “Advanced Calculus,” [23] K. J. Lui, “A Simple Test of the Homogeneity of Risk Difference in Sparse Data: An Application to a Multicen- [9] K. J. Lui and K. C. Chang, “Testing Homogeneity of Risk ter Study,” Biometrical Journal, Vol. 47, No. 5, 2008, pp. Difference in Stratified Randomized Trials with Non- compliance,” Computational Statistics and Data Analysis, [24] C. A. Rencher, “Linear Models in Statistics,” Wiley Se- ries in Probability and Mathematical Statistics, New York, [10] W. G. Cochran, “The Combination of Estimates from [25] A. Sen and M. Srivastava, “Regression Analysis: Theory, Different Experiments,” Biometrics, Vol. 10, No. 1, 1954, Methods, and Applications,” Springer-Verlag, New York, [11] W. G. Cochran, “Some Methods for Strengthening the Copyright 2012 SciRes. OJS
Appendix
Under a true common risk difference  over all k centers ( j  1, 2,, k ), the mean square error of Solving for  by taking summation on f , it yields To obtain the optimal weights  f f 1  0 , we form the auxiliary function  by following Lagrange’s method to seek   E  f        E   f        a E  f E   1 b f E   V   f        a E  f E   aE  1 aV f    f E  a E  b   a E  b  E  E  cj . The partial de- rivatives with respect to  and f yield  2 f V  2 f E   E   aV f    f E  1  aV f   f E  f E   f E   1  Substitute each of the subscript j and rearrange j  1 ; (aV  E  ) f  f E   f E    f E   1  j  2 ; f E   aV  E   f E   f E   1  j  3 ; f E   f E   aV  E  f  . . .  f E   1  f E   f E   f E   . . .  (aV  E  ) f It can be written in the matrix form as H f  y
Copyright 2012 SciRes. OJS
t     
f    f f f
y  1  1  1   1 
e  E E  E
The matrix H can be illustrated as
The inverse of H is suggested in several textbooks
of linear model such as Rencher [24] and Sen and H  D  t 
= D + t 
e  = D 
Therefore, f = D y 
f   
In practice, we have to estimate the adjusted summary estimator by replacing the sample estimates for the un- known quantities: E , V , p , p ,  . Copyright 2012 SciRes. OJS

Source: http://www.personal.soton.ac.uk/dab1f10/OJS20120100008_25739500.pdf