The Poisson distribution is suitable to model a rare event in a variety of fields, such as biology, commerce, quality control,
and so on. Real examples include a breast cancer study in Ng and Tang (2005) and an evaluation of a new vaccine in Chan
and Wang (2009). Recently, the Poisson distribution has been used to model the number of mapping reads for each gene in
an RNA-seq experiment. See Wang et al. (2010).
In comparing two Poisson distributions, the asymptotic test can be too liberal in finite sample cases. When the
sample scales are not large, the exact testing procedure is more appropriate for prevention of an inflated type I error
rate. One major challenge in the development of an exact test is the presence of nuisance parameters. When the null
hypothesis states the equality of the two populations, the classical conditional test uses conditioning to get rid of the
nuisance parameter(s). However, with a null hypothesis of non-superiority, the conditioning fails to eliminate the nuisance
parameter(s) completely.
One easy way to deal with the nuisance parameters is to estimate the p-value by plugging in some consistent estimates
of the nuisance parameters. Krishnamoorthy and Thomson (2004) first introduced the use of the restricted maximum
likelihood estimate (RMLE) of the common Poisson mean under the null hypothesis of equal means. Ng et al. (2007) extended
the idea to the problems with a nonzero difference null hypothesis and proposed the numerical approximation p-value.
Chan and Wang (2009) use the RMLEs at the boundary of the null space for stratified data. In this study, we will propose
an exact method under the estimation principle as well. The RMLEs of the nuisance parameters by taking the null space of
non-superiority into account is employed in estimating an exact p-value. Although the estimated p-value is easy and quick
to implement, it does not guarantee a well-controlled type I error rate.
The Poisson distribution is suitable to model a rare event in a variety of fields, such as biology, commerce, quality control,
and so on. Real examples include a breast cancer study in Ng and Tang (2005) and an evaluation of a new vaccine in Chan
and Wang (2009). Recently, the Poisson distribution has been used to model the number of mapping reads for each gene in
an RNA-seq experiment. See Wang et al. (2010).
In comparing two Poisson distributions, the asymptotic test can be too liberal in finite sample cases. When the
sample scales are not large, the exact testing procedure is more appropriate for prevention of an inflated type I error
rate. One major challenge in the development of an exact test is the presence of nuisance parameters. When the null
hypothesis states the equality of the two populations, the classical conditional test uses conditioning to get rid of the
nuisance parameter(s). However, with a null hypothesis of non-superiority, the conditioning fails to eliminate the nuisance
parameter(s) completely.
One easy way to deal with the nuisance parameters is to estimate the p-value by plugging in some consistent estimates
of the nuisance parameters. Krishnamoorthy and Thomson (2004) first introduced the use of the restricted maximum
likelihood estimate (RMLE) of the common Poisson mean under the null hypothesis of equal means. Ng et al. (2007) extended
the idea to the problems with a nonzero difference null hypothesis and proposed the numerical approximation p-value.
Chan and Wang (2009) use the RMLEs at the boundary of the null space for stratified data. In this study, we will propose
an exact method under the estimation principle as well. The RMLEs of the nuisance parameters by taking the null space of
non-superiority into account is employed in estimating an exact p-value. Although the estimated p-value is easy and quick
to implement, it does not guarantee a well-controlled type I error rate.
การแปล กรุณารอสักครู่..
