Interpreting treatment-effect estimates with heterogeneity and choice: simulation model results.