A Study of the Classical Asymptotic Noninferiority Test for Two Binomial Proportions
- PDF / 4,430,003 Bytes
- 5 Pages / 612 x 792 pts (letter) Page_size
- 29 Downloads / 163 Views
167
A Study of the Classical Asymptotic Noninferiority Test for Two Binomial Proportions
Felix Almrndra-Arao, PhD UPIITA del Instituto Politecnico Nacional, Mhico, DE Mexico
Key Words Noninferiority test; Classical asymptotic; Significance level; Binomial proportions; Hypothesis test Correspordrro Address Felix Almendra-Arao, Hidalgo #20,Tequisistldn. Tezoyuca. estado de Mhico, Mexico (email: [email protected]).
In a recent article by Li and Chuang-Stein (I), an evaluation was made of the peP.fonnanceof twofiequentb used methods-the classical asymptotic normal appraximation and the same method with the Hauck-Anderson continuity correction-to test noninfwiority between two proportions. The evaluation, using simulation, estimated type 1 errors and power. Continuing the research of Li and ChuangStein, this study evaluates the performance of these two meth-
INTRODUCTION When comparing a new drug with a standard active control, the researcher is often interested in demonstrating that the new drug, which has fewer side effects, is less costly, or can be more easily applied, is not much worse than the standard active control. To test this hypothesis, noninferiority tests are used. Many noninferiority tests have been proposed to compare two binomial proportions (2-11). However, as a nuisance parameter appears, calculation of levels of significance is computationally intensive, and comparing these tests is complicated. The present work was conducted to analyze significance level performance of classical asymptotic noninferiority tests for the difference between two binomial independent proportions, with and without continuity correction (cc), for the noninferiority margins, 0.1 and 0.15, most frequently used in antibiotic trials (1). Li and Chuang-Stein (1)do not compute the levels of significance for the asymptotic tests with and without cc; they compute actual levels of significance, namely the power at observed values. Moreover, type I errors and power were calculated in (1)using simulation. However, the contention of this article is that without levels of significance and using simulation, solid con-
ods. However, here type 1 errors are not estimated fim simulation, and lev& of sign$cance are computed by enumemting all possible cases rather than through simulation. And finally, instcad of the confidence intervals approach adopted in Li and Chuang-Stein, our work uses the hypothesis tests approach. The results show a comparison of the two methods that is differentfiom that obtained by Li and Chuang-Stein.
clusions on performance of levels of significance are difficult to obtain. Therefore, none of the computations in our investigation-type I errors and the levels of significance-are estimated using simulation. Because of technical difficulties specified in the section 'Significance Level," these values are calculated for a slight modification of the original statistics. In this way, our study resulted in a comparison that differs from that obtained by Li and Chuang-Stein (1) and, therefore, different conclusions on
Data Loading...
