410 Abstracts

A27 BETA ERRORS REVISITED

D. Moher, G.A. Wells and C.S. Dulberg Loeb Medical Research Institute

Ottawa, Canada

Randomized controlled trials (RCTs) have been accepted as the gold standard by which the medical community assesses the efficacy of interventions. Fifteen years ago, in a landmark study, Freiman and colleagues reviewed 71 "negative" RCTs: the vast majority had inflated Type II errors. We replicated and extended this study to assess whether the proportion of negative RCTs and inflated Type II errors had changed over time. A structured data collection form as well as guidelines to classify trials as positive or negative and to determine primary outcomes were developed. We reviewed all RCTs in three leading general medical journals (JAMA, Lancet, and NEJM), published in 1975, 1980, 1985 and 1990. Of the 383 RCTs, 102 (27%) were classified as negative: 22, 22, 21, and 37 over time. The number of RCTs nearly tripled from 1975 to 1990, with the proportion of negative trials remaining fairly stable. Only 15% and 34% of the negative two group parallel design trials (n=73) had sufficient statistical power (80%) to detect, respectively, a 25% or 50% relative difference. These results have not improved over time: 12%, 12%, 6%, 24% for a 25% relative difference; and 25%, 44%, 25%, 40% for a 50% relative difference. While overall, only 32 % of the negative trials reported sample size calculations, this percentage has improved significantly over time: 0%, 32%, 48%, 43%. The majority of these negative trials (70%) did not provide sufficient information to replicate sample size calculations. There is still much that needs improvement. One suggestion is to report RCTs using a structured format.

A28 SHOULD THERE BE RESTRICTIONS ON INDICATOR

LESIONS FOR PHASE II TRIALS IN ONCOLOGY?

Hilary Franklin on behalf of Early Clinical TriaLs Group of the EORTC

New Drug Development Office Amsterdam, The Netherlands

As tumor response is the main end point of phase II cancer clinical trials, it is essential that the chosen indicator lesions are reliable parameters for the activity of a potential anti-cancer drug. Recent proposals are to restrict the target groups of patients, for example to only those who have bi- dimensionally measurable disease of more than a certain diameter. Since these restrictions will inevitably decrease the rate of patient accrual, it is important to know whether they will in fact give a more reliable assessment of drug activity.

Over the last three years the Early Clinical Trials Group has gathered data on more than 750 patients in 27 multi-centric phase I1 trials concerning seven drugs and ten different tumor types. The following factors, besides the drug itself, which could influence the response of tumor lesions, have been recorded: site of lesion (primary, lung-, liver-, lymphnode-, soft tissue-, bone-metastasis), size of lesion, method of evaluation, tumor-type, prior irradiation, measurability (bi-, uni-dimensional), institution. Simulation of different groups of selection criteria was applied to these data to assess the "cost/benefit" in terms of loss of accrual/overall response rate in different situations, using multivariate techniques we could study the characteristics of an "optimal" balance between the speed/reliability of anti-cancer drug testing.

A29 OPTIMIZING PATIENT SELECTION IN

THERAPEUTIC EXPERIMENTS

K.G.M. Moons, G.A. van Es and D.E. Grobbee Erasmus UniversiLy

Rotterdam, Netherlands

To participate in a therapeutic experiment (trial) patients have to satisfy pre-defined inclusion and exclusion criteria. These criteria are often evaluated over a period of time, called selection period. The goal of this paper is to propose a methodology for optimizing the selection period in terms of efficiency, i.e., to obtain relevant information for recruitment with a minimum of measurements. The approach is to determine whether variables initially measured in the selection period have a predictive value for exclusions later on in this period, is evaluated by an univariate sensitivity and specificity analysis and subsequently by multivariate logistic regression models in