Asylum is a legal protection granted by a state to individuals who demonstrate a well-founded fear
of persecution or who face real risk of being subjected to torture in their country. However, asylum
adjXdicaWion ofWen dependV on Whe deciVion makeU¶V VXbjecWiYe aVVeVVmenW of Whe applicanW¶V
credibility. To investigate potential sources of bias in asylum adjudication practices researchers
have used statistics and machine learning models, finding significant sources of variation with
respect to a number of extra-legal variables. In this paper, we analyse an original dataset of Danish
asylum decisions from the Refugee Appeals Board to understand the variables that explain Danish
Adjudication. We train a number of classifiers and, while all classifiers agree that candidate
credibility is the single most important variable, we find that performance and variable importance
change significantly depending on whether data imbalance and temporality are taken into account.
We discuss the implications of our findings with respect to the theory and practice of predicting and
explaining asylum adjudication.