Statistical analysis for explosives detection system test and evaluation

Abstract

The verification of trace explosives detection systems is often constrained to small sample sets, so it is important to support the significance of the results with statistical analysis. As binary measurements, the trials are assessed using binomial statistics. A method is described based on the probability confidence interval and expressed in terms of the upper confidence interval bound that reports the probability of successful detection and its level of statistical confidence. These parameters provide useful measures of the system’s performance. The propriety of combining statistics for similar tests—for example in trace detection trials of an explosive on multiple surfaces—is examined by statistical tests. The use of normal statistics is commonly applied to binary testing, but the confidence intervals are known to behave poorly in many circumstances, including small sample numbers. The improvement of the normal approximation with increasing sample number is shown not to be substantial for the typical numbers used in this type of explosives detection system testing, and binary statistics are preferred. The methods and techniques described here for testing trace detection can be applied as well to performance testing of explosives detection systems in general.