Between 2014 and 2025, 25,712 deaths were recorded in 1,562 incidents along Mediterranean migration routes. The question is how many deaths were not recorded.
The Bayesian time-stratified latent-class framework accommodates incidents that were never observed at all. The model treats the recorded incidents as a partial sample of a larger population, with the detection probability varying by incident characteristics. The central finding: reporting likelihood increases with incident severity. A shipwreck that kills 200 is almost certain to be documented. A boat that sinks at night with 8 people aboard may leave no trace.
The estimated true fatality count is 30,426 to 39,172, meaning the recorded data captures 66% to 85% of actual deaths. The unreported fatalities are concentrated in 2014-2016, before surveillance networks expanded. Data collection has improved, but the improvement is itself informative: the gap between observed and estimated narrows over time, suggesting that the early period's undercount was severe.
The statistical method is multiple-systems estimation adapted for sparse humanitarian data. The model can't identify specific unobserved incidents — it estimates how many there were. The uncertainty is large because the unobserved population is, by definition, unobserved. But the lower bound of the credible interval exceeds the observed count by nearly 5,000, meaning the undercount is real regardless of modeling assumptions.
The hardest deaths to count are the ones where nobody survived to report them.