- Email: [email protected]
Partial factorial designs for randomized clinical trials
Abstracts
221
cross-sectional percentages at a given time in the trial. These results may be difficult to interpret as participants may be withdrawn from the study prior to its conclusion and before they ever complain of an adverse effect; if they remained in the study they might have complained at a later time. Withdrawals are participants who died, were~lost to follow-up, or were never asked about an adverse effect. However, survival analysis methods can be used to compare the distributions of "time to first complaint" in the active and placebo treatment groups, taking into account withdrawals. Particpants in the Beta-Blocker Heart Attack Trial were monitored for possible adverse effects. On each follow-up visit, they were asked whether they experienced any of seven conditions (blacking out, depression, fatigue, nightmares, hallucinations, bronchospasm, and decrease in sexual activity) since their previous visit about 3 months earlier. The patients were followed for up to 38 months. For fatigue and bronchospasm, the complaint-free time was significantly longer in the placebo versus active (propranolol) treatment group (p < .005).
Response Conditional Two-Period Crossover Design B. W h i t e
Otsuka Pharmaceutical Co., Ltd., Rockville, MD (04) To balance the conflicting interests of medicine, ethics, and economics inherent in clinical investigation, this paper presents a design intended for use in prospective, controlled studies on patients in which the best standard is compared with an experimental treatment. Proposed is a two-period design with random assignment to treatment and response conditional crossover to a second treatment. Maximum likelihood estimation procedures are used as a means of deriving inferences. Precision comparisons are made with the single-period parallel groups design. Relative precision is measured by asymptotic variances of the estimated difference in recovery rates for the two treatments. Computer simulations of situations that often occur in clinical practice-at least in trials of chronic neurological disorders--are used to assess small sample accuracy and relative efficiency of parameter estimation. Data are presented from an actual clinical trial utilizing this design, and statistical inferences are derived from a likelihood analysis of the results.
"Skewed" Incomplete Block Designs for Phase I Studies Barrett ScoviUe a n d Takashi Y o k o y a m a
Otsuka Pharmaceutical Co., Ltd., Frankfurt, Federal Republic of Germany (05) Two incomplete block designs are presented for early human studies incorporating randomized treatment assignments to several dosages or placebo, with "skewing" of treatment assignments so that dosage assignments proceed from lower to higher doses as the experiment progresses. In Design I, for pilot work when no previous human data exist, an initial presumed safe dose and a target plasma drug concentration are selected. Subjects are randomized in groups of three treatments--placebo, a low, and a higher dose. In successive groups, the high dose of the previous group becomes the low dose until the target plasma concentration is achieved or toxicity emerges. In Design II, when initial pilot data exist, power calculations are used to estimate the number of subjects desired per dose. An initial group of volunteers receive two treatments, placebo or dose 2D, assumed a mid-range dose. Depending on the outcome of the first treatment assignments, further assignments include 4D or 1D, to achieve the desired study N. Successful applications of these designs are presented. Partial Factorial Designs for Randomized C l i n i c a l Trials D a v i d Byar, S t e v e n Piantadosi, a n d A g n e s H e r z b e r g
Clinical and Diagnostic Trials Section, National Cancer Institute, National Institutes of Health, Bethesda, MD (06) Recently there has been increased interest in considering factorial designs for randomized clinical trials where there is a genuine interest in studying two or more treatments. Such designs may offer impressive gains in efficiency, especially in the absence of interactions between the treatments. When interactions are of interest, factorial designs provide one sensible approach for their study, although larger sample sizes may be required because tests for interactions have lower power than those for main effects. Sometimes it may not be reasonable, interesting, feasible, or ethical to study all treatment combinations in a balanced factorial design, and yet a classical fractional factorial design might be inappropriate because of aliasing. In this paper, a new class
222
Abstracts of partial factorial designs is presented and shown to avoid aliasing and provide a reasonable compromise in some instances.
Design Features of the VA Cooperative Study of Aspirin in Unstable Angina D o n a l d Archibald, H. Daniel Lewis, Jr. a n d J a m e s W. Davis
Cooperative Studies Program Coordinating Center, Veterans Administration Medical Center, West Haven, CT (07) Most of the trials to evaluate aspirin in preventing death or acute myocardial infarction (MI) have been chronic (post-Ml), long term, and high dose (900 to 1500 mg daily). The VA Cooperative Study (CS), which randomized 661 patients to aspirin and 677 patients to placebo and showed a 51% reduction in death or MI with aspirin (p = 0.0005), differed from previous studies in several important design features. CS used a low dose (325 mg daily) which had minimal side effects. CS was an acute, short-term (12-week) study of unstable angina (UA). Since UA is a potentially reversible syndrome with increased short term risk, randomization was prompt (within 51 hours of hospitalization and 3 hours of screening). There was frequent and intensive followup (FU) for acute MI which was diagnosed (DX) by both ECG and cardiac specific isoenzymes (ISO) analyzed at a central laboratory. ISO increased the specificity of the study (72 patients were DX with acute MI at entry and analyzed separately). ISO increased the sensitivity of DX of acute MI over ECG alone by 112% and over Investigator DX by 36%. Prompt randomization proved important because in placebo, 38% of events occurred within the first 2 weeks. The short term FU resulted in high compliance, low dropout to surgery, and was cost effective because the annualized placebo mortality rate during FU was 55% higher than that during the next 9 months. In conclusion, unstable angina is an excellent model for studying the efficacy of aspirin and other antiplatelet drugs and ISO greatly increased sensitivity for diagnosis of acute MI.
Surrogate Endpoints in Clinical Trials L a w r e n c e F r i e d m a n a n d Salim Y u s u f
Clinical Trials Branch, NHLBI, Bethesda, MD (08) Surrogate endpoints may be used in clinical trials either as a primary outcome instead of morbid and fatal events or a secondary outcome of interest. Although such an approach has the advantage of making clinical trials potentially less expensive or time consuming, various important methodologic issues should be addressed: First, the surrogate endpoint should correlate strongly with major clinical events and the modification of the former should reasonably lead to a similar change in the latter. Second, missing data due to death as a competing risk, noncompliance to a procedure such as angiography, or incomplete patient follow-up could have a considerable impact on the appropriate analyses and conclusions. Third, the risk, costs, and technical problems of the specialized investigations, standardization procedures, and persuasiveness of the results should be considered. The decision to use a particular surrogate endpoint should be based on an evaluation of the potential advantages and disadvantages. Plans to address the various methodologic pitfalls should be formulated in advance.
Current Treatments of Cardiovascular Diseases are Unlikely to Produce More than Moderate Reductions in Mortality S. Yusuf, R. Collins, L. F r i e d m a n , C. Furberg, a n d R. Peto NHLBI, Bethesda, Maryland and Oxford, England (09) Review of the mortality results from over 250 cardiovascular clinical trials indicates that the risk reductions for even the more promising interventions are likely to be only about 5%, 10% or 20%. This is perhaps due to the heterogeneity of each disease, unpredictability of survival even among apparently similar patients, and the variety of different mechanisms that can lead to death, only one of which may be appreciably influenced by a particular therapy. Therefore, for most treatments of common vascular diseases, whose effects on major end points are still uncertain, the chief need is to be able to distinguish reliably between whether there is no material benefit or whether there is a moderate effect. Consequently, such trials should aim for several times more endpoints than are currently usual, if they are to provide definitive results. This can be achieved by appropriate designs whereby a several-fold increase in trial size is possible, by enrolling high-risk patients and/or basing judgment on the combined data from all trials of an agent.
221
cross-sectional percentages at a given time in the trial. These results may be difficult to interpret as participants may be withdrawn from the study prior to its conclusion and before they ever complain of an adverse effect; if they remained in the study they might have complained at a later time. Withdrawals are participants who died, were~lost to follow-up, or were never asked about an adverse effect. However, survival analysis methods can be used to compare the distributions of "time to first complaint" in the active and placebo treatment groups, taking into account withdrawals. Particpants in the Beta-Blocker Heart Attack Trial were monitored for possible adverse effects. On each follow-up visit, they were asked whether they experienced any of seven conditions (blacking out, depression, fatigue, nightmares, hallucinations, bronchospasm, and decrease in sexual activity) since their previous visit about 3 months earlier. The patients were followed for up to 38 months. For fatigue and bronchospasm, the complaint-free time was significantly longer in the placebo versus active (propranolol) treatment group (p < .005).
Response Conditional Two-Period Crossover Design B. W h i t e
Otsuka Pharmaceutical Co., Ltd., Rockville, MD (04) To balance the conflicting interests of medicine, ethics, and economics inherent in clinical investigation, this paper presents a design intended for use in prospective, controlled studies on patients in which the best standard is compared with an experimental treatment. Proposed is a two-period design with random assignment to treatment and response conditional crossover to a second treatment. Maximum likelihood estimation procedures are used as a means of deriving inferences. Precision comparisons are made with the single-period parallel groups design. Relative precision is measured by asymptotic variances of the estimated difference in recovery rates for the two treatments. Computer simulations of situations that often occur in clinical practice-at least in trials of chronic neurological disorders--are used to assess small sample accuracy and relative efficiency of parameter estimation. Data are presented from an actual clinical trial utilizing this design, and statistical inferences are derived from a likelihood analysis of the results.
"Skewed" Incomplete Block Designs for Phase I Studies Barrett ScoviUe a n d Takashi Y o k o y a m a
Otsuka Pharmaceutical Co., Ltd., Frankfurt, Federal Republic of Germany (05) Two incomplete block designs are presented for early human studies incorporating randomized treatment assignments to several dosages or placebo, with "skewing" of treatment assignments so that dosage assignments proceed from lower to higher doses as the experiment progresses. In Design I, for pilot work when no previous human data exist, an initial presumed safe dose and a target plasma drug concentration are selected. Subjects are randomized in groups of three treatments--placebo, a low, and a higher dose. In successive groups, the high dose of the previous group becomes the low dose until the target plasma concentration is achieved or toxicity emerges. In Design II, when initial pilot data exist, power calculations are used to estimate the number of subjects desired per dose. An initial group of volunteers receive two treatments, placebo or dose 2D, assumed a mid-range dose. Depending on the outcome of the first treatment assignments, further assignments include 4D or 1D, to achieve the desired study N. Successful applications of these designs are presented. Partial Factorial Designs for Randomized C l i n i c a l Trials D a v i d Byar, S t e v e n Piantadosi, a n d A g n e s H e r z b e r g
Clinical and Diagnostic Trials Section, National Cancer Institute, National Institutes of Health, Bethesda, MD (06) Recently there has been increased interest in considering factorial designs for randomized clinical trials where there is a genuine interest in studying two or more treatments. Such designs may offer impressive gains in efficiency, especially in the absence of interactions between the treatments. When interactions are of interest, factorial designs provide one sensible approach for their study, although larger sample sizes may be required because tests for interactions have lower power than those for main effects. Sometimes it may not be reasonable, interesting, feasible, or ethical to study all treatment combinations in a balanced factorial design, and yet a classical fractional factorial design might be inappropriate because of aliasing. In this paper, a new class
222
Abstracts of partial factorial designs is presented and shown to avoid aliasing and provide a reasonable compromise in some instances.
Design Features of the VA Cooperative Study of Aspirin in Unstable Angina D o n a l d Archibald, H. Daniel Lewis, Jr. a n d J a m e s W. Davis
Cooperative Studies Program Coordinating Center, Veterans Administration Medical Center, West Haven, CT (07) Most of the trials to evaluate aspirin in preventing death or acute myocardial infarction (MI) have been chronic (post-Ml), long term, and high dose (900 to 1500 mg daily). The VA Cooperative Study (CS), which randomized 661 patients to aspirin and 677 patients to placebo and showed a 51% reduction in death or MI with aspirin (p = 0.0005), differed from previous studies in several important design features. CS used a low dose (325 mg daily) which had minimal side effects. CS was an acute, short-term (12-week) study of unstable angina (UA). Since UA is a potentially reversible syndrome with increased short term risk, randomization was prompt (within 51 hours of hospitalization and 3 hours of screening). There was frequent and intensive followup (FU) for acute MI which was diagnosed (DX) by both ECG and cardiac specific isoenzymes (ISO) analyzed at a central laboratory. ISO increased the specificity of the study (72 patients were DX with acute MI at entry and analyzed separately). ISO increased the sensitivity of DX of acute MI over ECG alone by 112% and over Investigator DX by 36%. Prompt randomization proved important because in placebo, 38% of events occurred within the first 2 weeks. The short term FU resulted in high compliance, low dropout to surgery, and was cost effective because the annualized placebo mortality rate during FU was 55% higher than that during the next 9 months. In conclusion, unstable angina is an excellent model for studying the efficacy of aspirin and other antiplatelet drugs and ISO greatly increased sensitivity for diagnosis of acute MI.
Surrogate Endpoints in Clinical Trials L a w r e n c e F r i e d m a n a n d Salim Y u s u f
Clinical Trials Branch, NHLBI, Bethesda, MD (08) Surrogate endpoints may be used in clinical trials either as a primary outcome instead of morbid and fatal events or a secondary outcome of interest. Although such an approach has the advantage of making clinical trials potentially less expensive or time consuming, various important methodologic issues should be addressed: First, the surrogate endpoint should correlate strongly with major clinical events and the modification of the former should reasonably lead to a similar change in the latter. Second, missing data due to death as a competing risk, noncompliance to a procedure such as angiography, or incomplete patient follow-up could have a considerable impact on the appropriate analyses and conclusions. Third, the risk, costs, and technical problems of the specialized investigations, standardization procedures, and persuasiveness of the results should be considered. The decision to use a particular surrogate endpoint should be based on an evaluation of the potential advantages and disadvantages. Plans to address the various methodologic pitfalls should be formulated in advance.
Current Treatments of Cardiovascular Diseases are Unlikely to Produce More than Moderate Reductions in Mortality S. Yusuf, R. Collins, L. F r i e d m a n , C. Furberg, a n d R. Peto NHLBI, Bethesda, Maryland and Oxford, England (09) Review of the mortality results from over 250 cardiovascular clinical trials indicates that the risk reductions for even the more promising interventions are likely to be only about 5%, 10% or 20%. This is perhaps due to the heterogeneity of each disease, unpredictability of survival even among apparently similar patients, and the variety of different mechanisms that can lead to death, only one of which may be appreciably influenced by a particular therapy. Therefore, for most treatments of common vascular diseases, whose effects on major end points are still uncertain, the chief need is to be able to distinguish reliably between whether there is no material benefit or whether there is a moderate effect. Consequently, such trials should aim for several times more endpoints than are currently usual, if they are to provide definitive results. This can be achieved by appropriate designs whereby a several-fold increase in trial size is possible, by enrolling high-risk patients and/or basing judgment on the combined data from all trials of an agent.