We show that survey and administrative data-based estimates of a panel data model of earnings, employment, and outmigration yield very different qualitative and quantitative predictions. Survey-based estimates substantially overpredict outmigration, in particular for lower performing immigrants. Consequently, employment and earnings of immigrants who remain in the country are overpredicted relative to model predictions from administrative data. Importantly, estimates from both data sources find opposite self-selection mechanisms into outmigration. Differences hold despite using the same cohort, survey period, and observable characteristics. Differences in predictions are driven by difficulties of properly separating non-random sample attrition from selective outmigration in survey data.