Approximate Confidence Intervals for Estimation of Matching Error Rates of Biometric Identification Devices

Abstract

Assessing the matching error rates of a biometric identification devices is integral to understanding its performance. Here we propose and evaluate several methods for creating approximate confidence intervals for matching error rates. Testing of biometric identification devices is recognized as inducing intra-individual correlation. In order to estimate error rates associated with these devices, it is necessary to deal with these correlations. In this paper, we consider extensions of recent work on adjustments to confidence intervals for binomial proportions to correlated binary proportions. In particular we propose a Agresti-Coull type adjustment for estimation of a proportion. Here that proportion represents an error rate. We use an overdispersion model to account for intra-individual correlation. To evaluate this approach we simulate data from a Beta-binomial distribution and assess the coverage for nominally 95% confidence intervals.