Comparing measures of sensitivity for vigilance tasks: Bayesian-estimated and log-linear corrected measures outperform A!

Abstract

Signal detection analyses often attribute the vigilance decrement to a combination of bias shifts and sensitivity losses. In many vigilance experiments, however, false alarm rates are at or near zero, complicating the analysis of sensitivity. Here, we report Monte Carlo simulations comparing three measures of sensitivity that can be calculated even with extreme hit and false alarm rates: A’, an estimate of the area under the curve that is commonly but mistakenly described as nonparametric; A_z calculated using the log-linear correction, a statistic that adjusts individual observers’ data to protect against low false alarm rates; and, 4_z estimated using a Bayesian hierarchical procedure, a measure that protects against extreme false alarm rates by sharing information between observers. Results confirm that bias shifts produce spurious changes in A’, and demonstrate that, 4_z estimated with either a log-linear correction or through hierarchical Bayesian modeling is more robust against low false alarm rates.