In early phase clinical trials, we often wish to identify a subgroup with favorable efficacy for further study. This question arises particularly when the overall results are not favorable. However, it is well known that the unadjusted estimation of the subgroup effects could be overly optimistic, due to multiplicity and selection bias. For example, even if there is no treatment effect, the search of a best subgroup among many subgroups could lead to a false positive signal. In this comparative study, we will use simulations to compare the performances of several popular approaches to address this problem. Candidate approaches include a Bayesian hierarchical model with the horseshoe prior, and also a model averaging approach which views the subgroup selection as a model selection problem and then use model averaging to obtain subgroup effects to account for selection uncertainty. Resampling approach is also evaluated, which is to split the dataset into a selection sample and an estimation sample and hence enables us to estimation the selection bias. The applications of such methods are also demonstrated in a case study.