Bài viết này trình bày một frequentist so sánh hiệu suất của khoảng thời gian con-fidence và độ tin cậy cho difference của các tỷ lệ hai từ hai mẫu độc lập. So sánh tiến hành xem xét ba frequentist tiêu chuẩn. Nó được tìm thấy rằng khoảng thời gian với tốt nhất perfor-mance, về phạm vi bảo hiểm xác suất, Bayesians; trong điều kiện độ dài dự kiến và phương sai của chiều dài, khoảng Newcombe cho thấy hiệu suất tốt nhất. Như một nhận xét ngoài, nó đã được tìm thấy khoảng thời gian truyền thống như Wald và điều chỉnh Wald có một hiệu suất kém.Từ khóa: Confidence khoảng thời gian, khoảng tin cậy, Difference của hai tỷ lệ...ResumenEste artículo presenta una comparación del comportamiento de interva los de confianza frecuentistas y de credibilidad bayesianos para la diferencia de dos proporciones provenientes de muestras aleatorias independientes. La comparación se lleva cabo considerando tres criterios frecuentistas con los cuales se concluyó que el mejor comportamiento, en términos de la proba bilidad de cobertura, lo tienen los intervalos bayesianos, y en términos de la longitud esperada y varianza de la longitud el mejor comportamiento está dado por el intervalo frecuentista de Newcombe. Como resultado de esta trong vestigación se encontró que los intervalos frecuentistas más populares como Wald y Wald tienen ajustado un comportamiento deficiente.Palabras clave: intervalos de confianza, intervalos de credibilidad, diferen-cia de dos proporciones.aDocente investigadora. E-mail: hanwenzhang@usantotomas.edu.co bDirector. E-mail: hugogutierrez@usantotomas.edu.co cProfesor asociado. E-mail: ecepedac@unal.edu.co6364 Hanwen Zhang, Hugo Andrés Gutiérrez Rojas & Edilberto Cepeda Cuervo1. nềnA common problem in practical statistics is estimatig the difference of two proportions by means of interval estimation. This topic is especially important in clinical trials where it is necessary to investigate cure rates of two drugs or treatments. The theoretical background of this research is as follows: suppose that X1,...,Xn1 and Y1,...,Yn2 are two independent samples such that Xi ∼ Bernoulli(p1) and Yj ∼ Bernoulli(p2), with i = 1,...,n1 and j = 1,...,n2. It is necessary to construct a confidence interval or a credibility interval for the difference of the proportions p1 − p2. The most popular method for estimating p1 −p2 by means of frequentist confi- dence interval is the Wald interval, which is presented in most introductory statis- tics textbooks in spite of its poor performance. Many modifications have been made to the Wald interval in order to improve it. One of them is the adjusted Wald interval obtained by widening the Wald interval to increase the coverage probability. This improvement is especially meaningful when the sample sizes are small. Another important interval is the score interval (Wilson 1927), obtained by inverting the score test statistics. This interval was first obtained for one proportion, and thereafter was to be extended to deal with the difference of two proportions. However, in that case, the interval lacks a closed form (Pan 2002) and must be computed by numerical approximations. Agresti & Caffo (2000) analyzed the score interval, and derived the Adding-4 method: add 2 successes and 2 failures to sample observation. A considerable number of authors agree that Agresti and Caffo method has a very good performance (Pan 2002, Correa & Sierra 2003, Agresti et al. 2008). Another interval obtained by modifying the score method is the Newcombe interval (Newcombe 1998a, 1998b), and it seems to have a similar performance to the Agresti and Caffo interval (Correa & Sierra 2003). In the Bayesian approach, Pham-Gia & Turkkan (1993) used the hypergeo- metric Appell function and derived the posterior distribution of p1 −p2 when beta priors are used for each proportion. Given the exact posterior distribution, an exact Bayesian credibility interval for p1 − p2 can be found. However the compu- tational procedures are somewhat tedious, therefore new computational methods such as the Markov Chain Monte Carlo (MCMC), can be used to make it easier to evaluate posterior distributions for p1 − p2, as Agresti & Min (2005) argued. In the literature, many comparisons between confidence intervals have been done (Newcombe 1998a, Newcombe 1998b, Agresti & Caffo 2000, Pan 2002, Correa & Sierra 2003). The aim of this research is to take into account Bayesian credibility intervals jointly with frequentist confidence intervals. After a brief introduction, Section 2 presents some frequentist and Bayesian intervals for p1 −p2. Traditional confidence intervals such as the Wald and adjusted Wald are considered, as well as Bayesian credibility intervals with two noninformative priors. Section 3 deals with the comparison criteria for the considered intervals: the coverage probability, the expected length, and the variance of the length are used in order to evaluate the performance of the intervals. Sect
đang được dịch, vui lòng đợi..
![](//viimg.ilovetranslation.com/pic/loading_3.gif?v=b9814dd30c1d7c59_8619)