Measuring judicial review of asylum decisions
- Type
- Research report, summary
- Corpus
- Published Dutch asylum rulings
- Primary source
- rechtspraak.nl open data
- Full text
- Dutch
Can published case law carry a statistic?
Public debate about asylum decisions runs largely on anecdote: a striking ruling, a single case, an impression of how often the immigration service gets it wrong. The report asks whether the published corpus of court rulings can support something firmer, namely a counted, source-linked statement about what courts actually upheld and in which kinds of cases.
The answer is a qualified yes. The corpus supports counting, provided every figure stays attached to the ruling it came from and the selection effects are stated with the number rather than buried in a footnote.
What the report concludes
Judicial review is measurable at scale
Published rulings carry enough structure to be parsed automatically into court, date, procedure, country of origin and outcome, with the source passage retained for every value. What used to require manual reading of individual cases can be counted across the full published corpus.
The account itself can be classified
In rulings that assess the asylum account, the underlying story type can be identified reliably enough to aggregate: conversion, LGBTI orientation, political activity, family ties, concealed origin, document fraud, medical grounds, draft evasion, military service, blood feud. Each label is anchored to a quoted passage so it can be checked.
Outcomes differ sharply per story type
The share of decisions the court leaves intact varies by tens of percentage points between story types. That spread is the central empirical finding: not every type of account survives judicial scrutiny equally well.
The limits are structural, not technical
Publication is selective, only contested decisions are litigated, and a legal outcome is not a factual verdict about a person's account. Better models do not remove those limits; they only make it easier to mistake them for absent.
BronFull report, including the proof of concept and its dataset, on the Dutch report page.