Most statistical approaches summarize the probe-level expression data into gene-level measures, which then are used for downstream statistical analyses. However, there are some limitations in using the gene level data for analysis, such as nonhomogeneous probe effects and the interaction effect (e.g., alternative splicing). In this paper, we consider a two-stage binomial test with a weighted probe rank approach to determine differentially expressed genes. Using a series of benchmark gene array datasets, we show the two-stage binomial test approach yielded a higher positive predictivity and a higher sensitivity than the conventional RMA, GCRMA, Dchip, and ANOVA approaches. In data application, the two-stage binomial test identified a subset of genes strongly related to cell proliferation in the prolactin study, and a subset of genes associated with lymph node metastasis in the breast cancer dataset. In addition, by exploring the proportion of probes with expression changes and the probe expression plot, the two-stage binomial test helped detect an alternative splicing form of the prolactin gene in the prolactin study. In the breast cancer dataset, the approach also identified one potential alternative splicing gene.