710 Abstracts

averaging the last two. Data collected to date on 1,407 patients' baseline visits do not show systematic differences among the three consecutive measures and the intravisit vadability is similar between any pair of measures. Intervisit variability for the mean of measures 1 and 2 is no greater than that for the mean of measures 2 and 3. This suggests that, although two measures will naturally have slightly more variability than three, two measures may be adequate for future trials. If three measures are available, our results suggest that all three be averaged, or, alternatively, that the median can be used to provide resistance to gross errors. We believe our results differ from some previous trial results because we use certified personnel and enforce a rigid protocol including a rest period followed immediately by measurement.

Pl10 SEARCHING FOR HIDDEN PERIODICITIES IN BIOLOGICAL TIME SERIES

David Carr Iphar CRF

Munich, Germany

Biological time series are usually characterized by cyclical variations, including circadian and ultradian rhythmicity. In the analysis of these rhythms, various techniques have been developed to answer two of the primary points of contention: whether the pododicity of the series in question is statistically significant, and if so, what is the best estimate of the pedod of the oscillation. This time series analysis is the statistical basis for the research area of chronobiology, which includes the analysis of time-of-day effects on drug efficacy in chronopharmacology.

The classical methods in chronobiology, include the Enright periodogram, the Cosinor method, and Au- tocorrelation analysis. The various significance levels that have been suggested for these methods will be discussed. These methods have been under strong criticism, and often lead to different results. More recently viable alternatives such as MESA, Cosinor with AR errors, and nonlinear time series models, have appeared which promise to cope with some of the problems. The advantages and disadvantages of these methods will be summarized, and in particular, the adequacy for the analysis of binary time series will be mentioned.

P l l l A COMPUTER PROGRAM FOR THE DESIGN AND THE ANALYSIS OF PHASE II CANCER CLINICAL

TRIALS WITH THE TRIANGULAR TEST

Eric Belllasant, Jacques Benlchou, and Claude Chaatang HOpital Saint-Louis

Paris, France

In cancer, phase II clinical trials are most of the time noncomparative trials and aim at determining whether the efficacy of a new treatment is sufficient to warrant further studies in phase III. The usual end-point is the response rate p and the study should be able to determine if p is greater than a prespecified value Po, defined as the largest response rate for which the investigators consider that phase III studies are not worthwhile. If one specifies the threshold response rate Pa (P= > Po) corresponding to the minimum clinically interesting benefit when compared with Po, and the values of e and 13, one can compute the required sample size N. In practice, this single-stage design is difficult to implement due to both a recruitment-related problem (N is usually too large) and an ethical problem (necessity of stopping an ongoing study when the drug appears clearly ineffective or effective). In a previous work, we have extended to the comparison of an observed percentage to a theoretical percentage the Triangular Test. This group sequential method (analyses can be performed after each group of n subjects) has appropriate statistical properties and leads to important re- ductions in sample sizes.

We have developed a computer program which allows the performance of both the design and the analysis of phase II studies with the Triangular Test. Based on the values chosen for the different parameters Po, Pa c~, [3 and n, the "Design," option allows to obtain, by simulation, the statistical properties of the test (i.e., the values of the type I and II error rates, more generally the probability of accepting the hypothesis of inefficacy for all the values of the response rate p between 0 and 1 by increment of 0.05) as well as the average sample number required to reach the conclusion (under both the null and the alternative hypotheses, and for all the preceding values of p). The "Analysis" option, based on the result of each patient, allows to calculate the values of V and Z for each sequential analysis, and plots the sequential plan and the path.

In conclusion, this computer program will make easier the use of the Triangular Test in oncology.