Seeing Confessions Clearly: Diagnosticity, Risk, and Signal Detection: A Signal-Detection Reappraisal of Interrogation Methods and Confession Evidence

Abstract

Diagnosticity, the ratio of the true confession rate to the false confession rate, is commonly used to compare interrogation methods. We reconsider its suitability for that purpose by examining whether it collapses discriminability and response criterion into a single ratio. We applied a signal detection framework to published confession outcomes, re-expressed true and false confession rates as d′ and criterion c, used posterior simulation to assess uncertainty in condition rankings, and conducted Monte Carlo experiments to examine finite-sample behavior. Diagnosticity and d′ produced substantively different rankings. In the reanalyzed data, the condition ranked highest by diagnosticity ranked only third in model-implied discriminability, whereas the Minimization condition showed the greatest separation of guilty from innocent participants. The rankings differed in 87% of posterior draws, indicating that the disagreement was not limited to the observed point estimates. Simulations further showed that, with the small samples and low false confession rates common in this literature, diagnosticity was strongly right-skewed, upwardly biased when defined, and undefined in a meaningful share of samples. Diagnosticity is therefore best understood as a likelihood ratio tied to a single operating point, not as a criterion-free measure of interrogation performance or a posterior probability. It can favor methods that make guilty and innocent suspects alike less likely to confess rather than methods that better separate them. Confession research should report true and false confession rates alongside d′ and c, and any translation into applied risk should state an explicit base rate.