- Email: [email protected]
Borrowing information in hypothesis testing
Abstracts
251 A Meta-Analysis of Clinical Trials of the Use of Antiplatelet or Anticoagulant
Therapy After Coronary Artery Bypass Surgery W i l l i a m G. H e n d e r s o n , S t e v e n G o l d m a n , Jack C o p e l a n d , T o m Moritz, a n d L a u r e n c e A. H a r k e r VA Cooperative Studies Program, Hines, Illinois (30) We performed a meta-analysis on 12 controlled clinical trials investigating the efficacy of antiplatelet or anticoagulant drugs in the prevention of saphenous vein graft occlusion following coronary bypass surgery. Eleven of the studies were single-center clinical trials, whereas the last study reported was a large-scale Veterans Administration cooperative study. Five studies were clearly positive, three clearly negative, and four equivocal. All had positive treatment effect sizes. A test of homogeneity showed the effect sizes could have arisen from the same population. The overall effect size was significant (95% C.I. = 0.23-0.41). Also, there was a strong relationship between effect size and time of initiation of treatment (r = 0.62). The meta-analysis suggests that the treatment is effective, particularly w h e n started early.
To Bonferroni or not Bonferroni? B.G. W h i t e
Otsuka Pharmaceutical Co., Ltd. Rockville, Maryland (31) The use of the Bonferroni method in reporting cardiovascular research results with simultaneous multiple comparisons has become widespread following the recommendation by Wallenstein et al. (Circulation Research, 1980.) This approach can lead to setting clinically and statistically unreasonable limits to indicate the size of the "residual doubt." If m is the number of comparisons between groups, should the adjustment of the "residual doubt" be a division of the P value by m? Or should it be the n u m b e r of time points of preplanned interest multiplied by m, or this value multiplied further by the n u m b e r of variables, or should m be set equal to 1? This article will examine current use of the Bonferroni method in reporting cardiovascular research. Borrowing Information in Hypothesis Testing K e n t R. Bailey
Mayo Clinic, Rochester, Minnesota (32) Suppose one is interested in testing the null hypothesis/4o: 81 = 0 but one has data (Z1, Z2) on 82 as well as 81, where 81 and 82 are "related," e.g., they are treatment effects in subsets. Rather than performing a test at the a level irrespective of 82, consider letting the test depend on Z2 as well as Z1. Let a(82) be the probability of a type I error, as a function of 82. [It is clearly desirable to b o u n d sup 82 c~(82).] Some simple rejection rules and their operating characteristics are studied.
Stage-Shift Cancer Screening Model R o b e r t J. C o n n o r a n d K e n n e t h C. C h u
Biometry Branch, National Cancer Institute, Bethesda, Maryland (33) A stage-shift cancer screening model is developed in the context of a randomized controlled trial (RCT) of cancer screening. In the model detection by screening causes the time of diagnosis of the cancer to be advanced so that either the stage at diagnosis is shifted from one stage to the next lower one or the stage of diagnosis is unchanged but the cancer is diagnosed earlier in the stage. These are called external and internal stage shifts, respectively. At each stage the extent of the external and internal shifts as well as the associated "cure" and survival benefits are estimated. Further, the model allows the interrelationships of these benefits within and between stages to be delineated. This then allows us to better understand the results of the RCT. Data from a completed breast cancer screening RCT are used to illustrate the application of the model and its value in improving our understanding of the trial's results. Effects of Allocation of Patients Using Minimization on Subsequent Analyses D. G i l b e r t s o n , S. C h a , a n d S. W i e a n d
Mayo Clinic, Rochester, Minnesota (34) We consider the effects of minimization, an essentially n o n r a n d o m process for allocating patients to treatments, on subsequent analyses comparing the treatments. Our results are consistent with those of Birkett (1985) and Kalish and Begg (1987) in that they indicate minimization may result
251 A Meta-Analysis of Clinical Trials of the Use of Antiplatelet or Anticoagulant
Therapy After Coronary Artery Bypass Surgery W i l l i a m G. H e n d e r s o n , S t e v e n G o l d m a n , Jack C o p e l a n d , T o m Moritz, a n d L a u r e n c e A. H a r k e r VA Cooperative Studies Program, Hines, Illinois (30) We performed a meta-analysis on 12 controlled clinical trials investigating the efficacy of antiplatelet or anticoagulant drugs in the prevention of saphenous vein graft occlusion following coronary bypass surgery. Eleven of the studies were single-center clinical trials, whereas the last study reported was a large-scale Veterans Administration cooperative study. Five studies were clearly positive, three clearly negative, and four equivocal. All had positive treatment effect sizes. A test of homogeneity showed the effect sizes could have arisen from the same population. The overall effect size was significant (95% C.I. = 0.23-0.41). Also, there was a strong relationship between effect size and time of initiation of treatment (r = 0.62). The meta-analysis suggests that the treatment is effective, particularly w h e n started early.
To Bonferroni or not Bonferroni? B.G. W h i t e
Otsuka Pharmaceutical Co., Ltd. Rockville, Maryland (31) The use of the Bonferroni method in reporting cardiovascular research results with simultaneous multiple comparisons has become widespread following the recommendation by Wallenstein et al. (Circulation Research, 1980.) This approach can lead to setting clinically and statistically unreasonable limits to indicate the size of the "residual doubt." If m is the number of comparisons between groups, should the adjustment of the "residual doubt" be a division of the P value by m? Or should it be the n u m b e r of time points of preplanned interest multiplied by m, or this value multiplied further by the n u m b e r of variables, or should m be set equal to 1? This article will examine current use of the Bonferroni method in reporting cardiovascular research. Borrowing Information in Hypothesis Testing K e n t R. Bailey
Mayo Clinic, Rochester, Minnesota (32) Suppose one is interested in testing the null hypothesis/4o: 81 = 0 but one has data (Z1, Z2) on 82 as well as 81, where 81 and 82 are "related," e.g., they are treatment effects in subsets. Rather than performing a test at the a level irrespective of 82, consider letting the test depend on Z2 as well as Z1. Let a(82) be the probability of a type I error, as a function of 82. [It is clearly desirable to b o u n d sup 82 c~(82).] Some simple rejection rules and their operating characteristics are studied.
Stage-Shift Cancer Screening Model R o b e r t J. C o n n o r a n d K e n n e t h C. C h u
Biometry Branch, National Cancer Institute, Bethesda, Maryland (33) A stage-shift cancer screening model is developed in the context of a randomized controlled trial (RCT) of cancer screening. In the model detection by screening causes the time of diagnosis of the cancer to be advanced so that either the stage at diagnosis is shifted from one stage to the next lower one or the stage of diagnosis is unchanged but the cancer is diagnosed earlier in the stage. These are called external and internal stage shifts, respectively. At each stage the extent of the external and internal shifts as well as the associated "cure" and survival benefits are estimated. Further, the model allows the interrelationships of these benefits within and between stages to be delineated. This then allows us to better understand the results of the RCT. Data from a completed breast cancer screening RCT are used to illustrate the application of the model and its value in improving our understanding of the trial's results. Effects of Allocation of Patients Using Minimization on Subsequent Analyses D. G i l b e r t s o n , S. C h a , a n d S. W i e a n d
Mayo Clinic, Rochester, Minnesota (34) We consider the effects of minimization, an essentially n o n r a n d o m process for allocating patients to treatments, on subsequent analyses comparing the treatments. Our results are consistent with those of Birkett (1985) and Kalish and Begg (1987) in that they indicate minimization may result