Statistical Significance of Veterans Administration Vasodilator Heart Failure Trial Results Jay N. Cohn, MD
hestatistical significanceof the mortality reduction observedin the Veterans Administration Vasodilator Heart Failure Trial (V-HeFT) from hydralazine-isosorbidedinitrate treatment of heart failure’ has been the subject of some controversy. Since this could impact on the mandate for vasodilator therapy of heart failure and on the design of subsequenttrials, the issuedeservesfurther consideration. The urgency of this reconsiderationhas been hastenedby the August article regarding the Studies of Left Ventricular Dysfunction (SOLVD) protocol,2 in which an unfortunate typographical error identified the p value for V-HeFT as 0.93 rather than 0.093. V-HeFT had 3 treatment arms: placebo, hydralazine-nitrate or prazosin added to digoxin and diuretics. The study was designedto compare placebo with combined vasodilator regimens if the results with the 2 treatments were similar, or with the individual vasodilator regimens if they were not. A 2-tailed test of significancewas chosen (although in retrospect a l-tailed test would have been appropriate becausewe were not concernedwith an adverseeffect of vasodilators) and a logrank mortality analysis was designated. It became apparent early to the Data Monitoring Committee-which remained blinded to the treatment arms-that 1 group was faring better than the others, and therefore all analysesperformed compared individual rather than combined treatment arms. Comparison of the overall placebo and hydralazine-isosorbide dinitrate survival curves at the end of the trial revealed a difference with a p value of 0.093 using the specified log-rank test. Using a generalized Wilcoxon test, which gives somewhat more weight to the early events, the p value was 0.046. Some slight adjustment upward of these p values would be appropriate because of 4 “looks” at the data before termination of the study, but the magnitude of the adjustment would not be substantial becausethe stopping rules for the study before termination were extremely stringent. Consideration also could be given to the 2 possiblecomparisonsproposedcombinedvasodilator groups versusplacebo and hydralazine-nitrate versus placebo-which might also impact on the significance of the p value. On the other hand, of course, a l-sided test would have halved the p value.
T
From the Department of Medicine, University of Minnesota Medical School, Minneapolis, Minnesota. Manuscript received and accepted August 20,199O. Address for reprints: Jay N. Cohn, MD, Cardiovascular Division, Box 488 UMHC, University of Minnesota Medical School, Minneapolis, Minnesota 55455.
The 2-year mortality also was designatedin advance as an end point for the study, becauseit was anticipated that a positive treatment effect might wane over time as high-risk patients, perhaps kept alive by the treatment, could contribute to a progressivelyhigher late mortality in the treatment group. Furthermore, comparison of survival curves in modest-sizedstudies usually includes adjustment for baseline differences between the treatment groups of variables that are shown in the trial to be related to survival. This adjustment requires modeling of the data using a Cox life-table regressionmodel, which closely fit the data in V-HeIT. This modeling of the data with adjustment for baseline differences revealed an overall mortality reduction from hydralazineisosorbide dinitrate of 28% (95% confidence interval 3 to 46%) and a 2-year mortality reduction of 34% (95% confidence interval 4 to 54%, p = 0.028).3 It is appropriate to compare this somewhat “borderline” statistical significance in V-HeFT with the prominent (p
hestatistical significanceof the mortality reduction observedin the Veterans Administration Vasodilator Heart Failure Trial (V-HeFT) from hydralazine-isosorbidedinitrate treatment of heart failure’ has been the subject of some controversy. Since this could impact on the mandate for vasodilator therapy of heart failure and on the design of subsequenttrials, the issuedeservesfurther consideration. The urgency of this reconsiderationhas been hastenedby the August article regarding the Studies of Left Ventricular Dysfunction (SOLVD) protocol,2 in which an unfortunate typographical error identified the p value for V-HeFT as 0.93 rather than 0.093. V-HeFT had 3 treatment arms: placebo, hydralazine-nitrate or prazosin added to digoxin and diuretics. The study was designedto compare placebo with combined vasodilator regimens if the results with the 2 treatments were similar, or with the individual vasodilator regimens if they were not. A 2-tailed test of significancewas chosen (although in retrospect a l-tailed test would have been appropriate becausewe were not concernedwith an adverseeffect of vasodilators) and a logrank mortality analysis was designated. It became apparent early to the Data Monitoring Committee-which remained blinded to the treatment arms-that 1 group was faring better than the others, and therefore all analysesperformed compared individual rather than combined treatment arms. Comparison of the overall placebo and hydralazine-isosorbide dinitrate survival curves at the end of the trial revealed a difference with a p value of 0.093 using the specified log-rank test. Using a generalized Wilcoxon test, which gives somewhat more weight to the early events, the p value was 0.046. Some slight adjustment upward of these p values would be appropriate because of 4 “looks” at the data before termination of the study, but the magnitude of the adjustment would not be substantial becausethe stopping rules for the study before termination were extremely stringent. Consideration also could be given to the 2 possiblecomparisonsproposedcombinedvasodilator groups versusplacebo and hydralazine-nitrate versus placebo-which might also impact on the significance of the p value. On the other hand, of course, a l-sided test would have halved the p value.
T
From the Department of Medicine, University of Minnesota Medical School, Minneapolis, Minnesota. Manuscript received and accepted August 20,199O. Address for reprints: Jay N. Cohn, MD, Cardiovascular Division, Box 488 UMHC, University of Minnesota Medical School, Minneapolis, Minnesota 55455.
The 2-year mortality also was designatedin advance as an end point for the study, becauseit was anticipated that a positive treatment effect might wane over time as high-risk patients, perhaps kept alive by the treatment, could contribute to a progressivelyhigher late mortality in the treatment group. Furthermore, comparison of survival curves in modest-sizedstudies usually includes adjustment for baseline differences between the treatment groups of variables that are shown in the trial to be related to survival. This adjustment requires modeling of the data using a Cox life-table regressionmodel, which closely fit the data in V-HeIT. This modeling of the data with adjustment for baseline differences revealed an overall mortality reduction from hydralazineisosorbide dinitrate of 28% (95% confidence interval 3 to 46%) and a 2-year mortality reduction of 34% (95% confidence interval 4 to 54%, p = 0.028).3 It is appropriate to compare this somewhat “borderline” statistical significance in V-HeFT with the prominent (p
