Metamodel-based bootstrap methods for characterizing input model uncertainty have disadvantages for settings where there are a large number of input distributions, or when using empirical distributions to drive the simulation. Early direct bootstrapping of empirical distributions did not take into account the distinction between intrinsic and extrinsic variations in the resampled quantities. When the intrinsic uncertainty is large, the result is overcoverage of the bootstrap percentile intervals. We explore ways of accounting for both sources in direct bootstrap characterization of input model uncertainty, and study the impact on confidence interval (CI) coverage. Four new bootstrap-based CIs for the expected simulation output under the unknown true distribution are proposed, basic shrinkage CI, percentile shrinkage CI, basic hierarchical bootstrap CI, and percentile hierarchical bootstrap CI, and their empirical performances are demonstrated using an example.