- Email: [email protected]
The use of factorial designs when the outcome variable is binary: What happens when low level interactions are present
Abstracts
323
to age five and then to randomize 410 children to either penicillin or placebo for a minimum of two years.
Design of a T w o - T i e r e d C l i n i c a l Trial: Studies of Left Ventricular Dysfunction James D. H o s k i n g , Clarence E. Davis, Salim Y u s u f for the S O L V D I n v e s t i g a t o r s University of North Carolina, Chapel Hill, North Carolina (21) Clinical trials are usually planned to address multiple questions, with conflicting demands on study design. To have good power to detect reductions of 15-20% in relatively rare events (e.g., mortality) thousands of subjects are required, but the necessary data are simple. To study the mechanisms of treatment effect, extensive (and expensive) measurements of physiologic variables (e.g., ventricular function, arrhythmias) are necessary, but if the measurements are continuous, far fewer subjects are required. To balance these needs, the SOLVD project combines two large trials with limited data collection on a total of 7100 subjects, with seven detailed substudies on subsets of 40 to 500 subjects. The design of the trials and substudies will be presented, and the implications for study operations and data analysis will be presented.
Design of the POSCH Kidney Stone Trial J o h n P. Marts, Laurie L. Fitch, J o h n M. L o n g a n d the P O S C H G r o u p
University of Minnesota, Minneapolis, Minnesota (22) During the course of the Program on the Surgical Control of the Hyperlipidemias (POSCH), it was found that the surgery group patients had an increased incidence of kidney stones. Several methods of therapy were tried in the surgery patients but to no avail. Next, a case-control study was performed to gather information on possible etiology. Based on this information, it was decided to piggyback onto POSCH a double blind randomized trial of potassium citrate versus placebo for the prevention of kidney stones using POSCH surgery patients. Due to a limitation of funds, baseline x-rays and laboratory chemistries were obtained from each patient's local physician. Other data were gathered by phone. After randomization into the kidney stone trial, each patient's followup schedule had to be synchronized with their original POSCH foliowup schedule. The design and implementation of this trial will be presented as well as a discussion of the advantages and disadvantages, problems and solutions of conducting a "trial within a trial."
A Multi-Center Phase I Trial of Chemotherapy in Operable Breast Cancer M. Levine, K. Pritchard for the O n t a r i o Clinical O n c o l o g y G r o u p
McMaster University, Hamilton, Ontario, Canada (23) A phase I study is in progress to determine the most intensive regimen of cyclophosphamide, epirubicin and five fluorouracil which is tolerable in patients with Stage II breast cancer. Outcome measures are admission to hospital for febrile neutropenia and change in cardiac function as assessed by radionucleide angiography. To improve the efficiency, this study is multi-center rather than using the traditional single-center approach. Data are forwarded to a coordination center in a timely fashion, monitored carefully and dosages are increased according to predetermined guidelines. To date, 110 consecutive eligible patients have been entered into this study and the third level of an escalating dose schedule has been reached. Our experience shows that it is feasible to conduct a multi-center Phase I dose finding study in cancer. This should lead to a quicker establishment of the maximum, tolerable regimen to be evaluated in a subsequent randomized trial in patients with Stage II breast cancer.
The Use of Factorial Designs When The Outcome Variable Is Binary: What Happens When Low Level Interactions Are Present Bruce T h o m p s o n , Genell K n a t t e r u d
Maryland Medical Research Institute, Baltimore, Maryland (24) Often, in the course of planning a clinical trial with a binary dependent variable, there are several interventions that could be tested. Rather than conduct a trial for each of these treatments, it has become increasingly common to combine the treatments into one clinical trial using a factorial
324
Abstracts design. This design is advantageous over two independent designs because it provides a cost savings, a reduced sample size to address both primary aims, and the opportunity ascertain whether there are treatment interactions. It was found that the test for interaction may not have sufficient power to detect important low level interactions. This could affect main effect comparisons. Using power calculations in the presence of low level interactions, the magnitude of this problem is investigated.
When To Randomize? Sylvain Durrleman, Richard Simon
National Cancer Institute, Bethesda, Maryland (25) In many complex clinical trials, an important design issue is to decide w h e n to randomize patients. For example, in cancer therapeutics this question will arise in a trial comparing chemo-therapy versus the same chemutherapy followed by radiotherapy. More general examples include trials comparing maintenance therapy to no maintenance therapy. In these trials, one may randomize patients when treatment begins, or one may delay randomization until the two arms differ. In the first case, patients who fail the first part of treatment, or who refuse the assigned maintenance therapy, are problematic for the analysis. In order to ensure unbiased comparisons, the trial should be analyzed according to the "intent to treat" principle. Because of the dilution of treatment effect, this unbiased analysis has reduced power. With the late randomization design, early failures are not a problem, but there is often a higher proportion of patients who refuse randomization. We present a model to compare the efficiency of these two designs and discuss the characteristics of the pragmatic approach (early randomization) versus the explanatory attitude. Other approaches to analysis are also discussed.
Defining the Universe of Patients B a r b a r a S. H a w k i n s for t h e C o l l a b o r a t i v e O c u l a r M e l a n o m a S t u d y G r o u p The Johns Hopkins University, Baltimore, Maryland (26) In a set of clinical trials of treatment for a rare eye cancer, the Collaborative Ocular Melanoma Study Group has made a special effort to document all patients with choroidal melanoma examined in study centers during the anticipated 10-year patient accrual period. Each patient with this diagnosis is evaluated clinically and reported to be eligible or ineligible for one of the COMS randomized trials. As is common in such trials, extensive baseline and follow-up data are collected for eligible patients who are enrolled. However, unlike most trials in more common conditions, selected "baseline" information also is collected for all ineligible patients and for eligible patients who do not enroll. This information permits comparisons of (1) eligible patients who enroll with those who do not, (2) eligible and ineligible patients, and (3) clinical centers regarding possible selection biases. Furthermore, the assumptions made in the planning stage regarding study feasibility may be checked. The benefits of such information must be weighed against the cost of achieving comparable data quality for all three groups of patients, as illustrated by experience with more than 1100 cases reported during the first two years of patient accrual.
Quantifying Ascertainment Bias in Vaccine Efficacy Trials C o n o r P. F a r r i n g t o n
PHLS Communicable Disease Surveillance Center, London, England (27) The case definition used in a vaccine efficacy trial may be insensitive or nonspecific, thus biasing the efficacy estimates. Bias may also occur as a result of u n k n o w n prior disease and, in the case of open trials, through selective ascertainment of cases. The interactions of these various sources of ascertainment bias may be studied and quantified, thus enabling experimenters to evaluate the sensitivity of efficacy estimates obtained in a trial to the various bias parameters. Consideration of ascertainment bias is also important at the design stage of a study and may influence the choice of case definition. The consequences of ascertainment bias on age-specific vaccine efficacy estimates may also be quantified. Results are illustrated with data from pertussis vaccine studies.
323
to age five and then to randomize 410 children to either penicillin or placebo for a minimum of two years.
Design of a T w o - T i e r e d C l i n i c a l Trial: Studies of Left Ventricular Dysfunction James D. H o s k i n g , Clarence E. Davis, Salim Y u s u f for the S O L V D I n v e s t i g a t o r s University of North Carolina, Chapel Hill, North Carolina (21) Clinical trials are usually planned to address multiple questions, with conflicting demands on study design. To have good power to detect reductions of 15-20% in relatively rare events (e.g., mortality) thousands of subjects are required, but the necessary data are simple. To study the mechanisms of treatment effect, extensive (and expensive) measurements of physiologic variables (e.g., ventricular function, arrhythmias) are necessary, but if the measurements are continuous, far fewer subjects are required. To balance these needs, the SOLVD project combines two large trials with limited data collection on a total of 7100 subjects, with seven detailed substudies on subsets of 40 to 500 subjects. The design of the trials and substudies will be presented, and the implications for study operations and data analysis will be presented.
Design of the POSCH Kidney Stone Trial J o h n P. Marts, Laurie L. Fitch, J o h n M. L o n g a n d the P O S C H G r o u p
University of Minnesota, Minneapolis, Minnesota (22) During the course of the Program on the Surgical Control of the Hyperlipidemias (POSCH), it was found that the surgery group patients had an increased incidence of kidney stones. Several methods of therapy were tried in the surgery patients but to no avail. Next, a case-control study was performed to gather information on possible etiology. Based on this information, it was decided to piggyback onto POSCH a double blind randomized trial of potassium citrate versus placebo for the prevention of kidney stones using POSCH surgery patients. Due to a limitation of funds, baseline x-rays and laboratory chemistries were obtained from each patient's local physician. Other data were gathered by phone. After randomization into the kidney stone trial, each patient's followup schedule had to be synchronized with their original POSCH foliowup schedule. The design and implementation of this trial will be presented as well as a discussion of the advantages and disadvantages, problems and solutions of conducting a "trial within a trial."
A Multi-Center Phase I Trial of Chemotherapy in Operable Breast Cancer M. Levine, K. Pritchard for the O n t a r i o Clinical O n c o l o g y G r o u p
McMaster University, Hamilton, Ontario, Canada (23) A phase I study is in progress to determine the most intensive regimen of cyclophosphamide, epirubicin and five fluorouracil which is tolerable in patients with Stage II breast cancer. Outcome measures are admission to hospital for febrile neutropenia and change in cardiac function as assessed by radionucleide angiography. To improve the efficiency, this study is multi-center rather than using the traditional single-center approach. Data are forwarded to a coordination center in a timely fashion, monitored carefully and dosages are increased according to predetermined guidelines. To date, 110 consecutive eligible patients have been entered into this study and the third level of an escalating dose schedule has been reached. Our experience shows that it is feasible to conduct a multi-center Phase I dose finding study in cancer. This should lead to a quicker establishment of the maximum, tolerable regimen to be evaluated in a subsequent randomized trial in patients with Stage II breast cancer.
The Use of Factorial Designs When The Outcome Variable Is Binary: What Happens When Low Level Interactions Are Present Bruce T h o m p s o n , Genell K n a t t e r u d
Maryland Medical Research Institute, Baltimore, Maryland (24) Often, in the course of planning a clinical trial with a binary dependent variable, there are several interventions that could be tested. Rather than conduct a trial for each of these treatments, it has become increasingly common to combine the treatments into one clinical trial using a factorial
324
Abstracts design. This design is advantageous over two independent designs because it provides a cost savings, a reduced sample size to address both primary aims, and the opportunity ascertain whether there are treatment interactions. It was found that the test for interaction may not have sufficient power to detect important low level interactions. This could affect main effect comparisons. Using power calculations in the presence of low level interactions, the magnitude of this problem is investigated.
When To Randomize? Sylvain Durrleman, Richard Simon
National Cancer Institute, Bethesda, Maryland (25) In many complex clinical trials, an important design issue is to decide w h e n to randomize patients. For example, in cancer therapeutics this question will arise in a trial comparing chemo-therapy versus the same chemutherapy followed by radiotherapy. More general examples include trials comparing maintenance therapy to no maintenance therapy. In these trials, one may randomize patients when treatment begins, or one may delay randomization until the two arms differ. In the first case, patients who fail the first part of treatment, or who refuse the assigned maintenance therapy, are problematic for the analysis. In order to ensure unbiased comparisons, the trial should be analyzed according to the "intent to treat" principle. Because of the dilution of treatment effect, this unbiased analysis has reduced power. With the late randomization design, early failures are not a problem, but there is often a higher proportion of patients who refuse randomization. We present a model to compare the efficiency of these two designs and discuss the characteristics of the pragmatic approach (early randomization) versus the explanatory attitude. Other approaches to analysis are also discussed.
Defining the Universe of Patients B a r b a r a S. H a w k i n s for t h e C o l l a b o r a t i v e O c u l a r M e l a n o m a S t u d y G r o u p The Johns Hopkins University, Baltimore, Maryland (26) In a set of clinical trials of treatment for a rare eye cancer, the Collaborative Ocular Melanoma Study Group has made a special effort to document all patients with choroidal melanoma examined in study centers during the anticipated 10-year patient accrual period. Each patient with this diagnosis is evaluated clinically and reported to be eligible or ineligible for one of the COMS randomized trials. As is common in such trials, extensive baseline and follow-up data are collected for eligible patients who are enrolled. However, unlike most trials in more common conditions, selected "baseline" information also is collected for all ineligible patients and for eligible patients who do not enroll. This information permits comparisons of (1) eligible patients who enroll with those who do not, (2) eligible and ineligible patients, and (3) clinical centers regarding possible selection biases. Furthermore, the assumptions made in the planning stage regarding study feasibility may be checked. The benefits of such information must be weighed against the cost of achieving comparable data quality for all three groups of patients, as illustrated by experience with more than 1100 cases reported during the first two years of patient accrual.
Quantifying Ascertainment Bias in Vaccine Efficacy Trials C o n o r P. F a r r i n g t o n
PHLS Communicable Disease Surveillance Center, London, England (27) The case definition used in a vaccine efficacy trial may be insensitive or nonspecific, thus biasing the efficacy estimates. Bias may also occur as a result of u n k n o w n prior disease and, in the case of open trials, through selective ascertainment of cases. The interactions of these various sources of ascertainment bias may be studied and quantified, thus enabling experimenters to evaluate the sensitivity of efficacy estimates obtained in a trial to the various bias parameters. Consideration of ascertainment bias is also important at the design stage of a study and may influence the choice of case definition. The consequences of ascertainment bias on age-specific vaccine efficacy estimates may also be quantified. Results are illustrated with data from pertussis vaccine studies.