|
Abstract:
|
In FDA drug safety evaluation, there are thousands of reported adverse events (AE) associated with thousands of drugs under the adverse event reporting system (AERS). The data is in the form of a large I × J table, with nij being the reported number of AE’s for the ith AE and jth drug. The data are collected for a large number of users of the drugs over multiple years. As no adverse events have been observed for a lot of drugs for many years, the challenges are how to handle the large number of excessive zero counts, and incorporate potential covariates. To handle these problems, we propose a panel count model which assumes a non-homogenous Poisson process yij=yij(t) for counts in time interval (0,t] having conditional meanE(yij(t)|zj,xj )=G(t)exp(?Txij+?izij ),where xij=(xij1,...,xijd)?Rd are covariates ,yij(t)is the event count in (0,t], zj is the vector of length I indicating the AE positive status: zij = I(nij > 0); ? is the regression coefficients vector, and G(·) ? 0 is an unspecified monotone increasing function. The EM algorithm and isotonic technique are used to estimate parameters. Simulation studies are conducted to evaluate the performance of the method.
|
