Finally, a by-product of the method proposed here is that the individual genes can be estimated to have exactly zero effect on the response. The list of genes with estimated nonzero effects then comprise a gene list that investigators can do further validation work on. However, in our simulations (data not shown), we found that the method had difficulty in selecting the correct v variables. This attests to the fact that variable selection in the situation of large p and small n is quite difficult. An alternative to the method proposed here is Bayesian variable selection methods (Lee et al [24]). We are currently exploring an adaptation of the algorithm described here to a Bayesian approach.
