A personalized medicine may only benefit a patient population defined by a predictive biomarker. However, there is great uncertainty about the biomarker effect in designing a confirmatory trial and it is logical to take a two-stage approach. The first stage de-selects (or prunes) non-performing subpopulations at an interim analysis, and the second stage pools the remaining ones at the final analysis. The endpoints used at the two stages are different in general. A key issue of interest is the statistical property of the test-statistics and point estimate at the final analysis. Pruning and pooling generally inflates type I error. Previous research has focused on type I error control and power calculation. This presentation will investigate estimation bias of the treatment effect, which is implied with type I error inflation. While the previous work has treated treatment effect on the intermediate endpoint as a nuisance parameter in order to provide the most conservative type I error control, we take the same approach to explore the maximum bias. The methodology is applied to two designs. In the first design, patients with different biomarker levels are enrolled in a study and the treatment effect is assumed to increase with the biomarker level. The goal of the interim analysis is to identify a biomarker subpopulation in ascending order. In the second design, patients with different tumor types but same biomarker signature are included in the same trial applying a basket design. The goal of the interim analysis is to identify a subset of tumor types in absence of order. Closed form equations are provided for estimation of the bias under the two designs, and simulations are conducted under various scenarios to validate the analytic results. Worked examples are presented. Extensions and operational considerations are discussed.