The Island Problem Revisited

Publisert i

American Statistician 63 (3), 2009, pages 269-273

Sammendrag

Here I revisit the so-called “island problem” in forensic statistics. The problem is how to properly update the probability of guilt when a suspect is found that has the same characteristics as a culprit. In particular, how should the search protocol be accounted for? I present the established results of the literature and extend them by considering the selection effect resulting from a protocol in which only cases with a suspect reach the court. I find that the updated probability of guilt is shifted when properly accounting for the selection effect. Which way this shift goes depends on the exact distribution of all potential characteristics in the population. The shift is only marginal in numerical examples that have any resemblance to real-world forensic cases. The island problem illustrates the general point that the exact protocol through which data are generated is an essential part of the information set that should be used when analyzing nonexperimental data.

Fulltekst

By Halvor Mehlum
Published June 23, 2011 1:25 PM - Last modified June 23, 2011 1:27 PM