Emmanuelle Comets (1), France Mentré (1), Ryosei Kawai (2), Peter Marbach (2) and Jacky Vonderscher (2).
(1) Inserm U 436, CHU Pitié-Salpetrière, 91 Bd de l'Hopital, 75 013 Paris, France (2) Novartis Pharma AG, Basel CH-4002, Switzerland
Our study was motivated by the analysis of dissolution tests performed For several batches of a long-acting repeatable formulation of octreotide, a somatostatine analogue used in acromegaly. Seven measurements of the cumulative fraction dissolved were collected over 24 hr. For each batch, six dissolution curves were recorded, and analysed with a three-parameter Weibull model. A first method to compare two batches is to estimate the individual parameters for each curve by non linear regression, and to compare them between batches using the non parametric Wilcoxon test. In a second method, population analysis (NONMEM) is used to estimate the model, assuming either identical or different parameters for the two batches, and a likelihood ratio test is performed. STS (standard two-stage) and GTS (global two-stage) are alternative methods toestimate population parameters for each batch from the individual estimates. We derived the corresponding Wald tests, based on the estimated asymptotic information matrices. The four approaches were evaluated by simulations: first, assuming two identical batches, we performed 200 replications of 10 curves for each batch and evaluated the nominal level of the tests. Then, we assessed thepower of the tests to detect differences under several alternative hypotheses. The type I error of the Wilcoxon test was close to 5%. For the Population approach, we corrected the nominal level of the test to 0.5% and obtained a type I error close to 5%. The two-stage approaches, on the other hand, proved disappointing: type I error was higher than 20% for both, and surprisingly, GTS did not improve STS results. Compared to Wilcoxon, the population method provided the highest power under the various alternatives. As an illustration, the different tests were applied to real data for four batches to detect process variations.
Reference: PAGE 9 (2000) Abstr 107 [www.page-meeting.org/?abstract=107]
Poster: poster