What do detection dogs know and how do we know they know it?

Detection dogs do their jobs in a double-blind world, where neither the evaluator (the dog) nor the handler knows whether the substance of interest is present or its identity.  For dogs to be considered reliable, double-blind testing must occur as part of the end-stage and ongoing evaluation process.  This testing is the only way accurate to determine false positive and false negative rates.  These rates determine the predictive value (external validity) of the positive test, the predictive value (external validity) of the negative test, and likelihood ratios that will tell us whether the dog is reliable and how sensitive or specific the evaluation process is.  Sensitive tests detect very low levels of the target odor, so sensitivity is a measure of what you could miss. When dogs are tested there are usually very few target odors compared with blanks so a dog could do extremely well with respect to sensitivity if he never finds anything.  For this reason, sophisticated assessment is needed. Calculation of sensitivity, specificity and likelihood ratios can be used to improve a program and justify investment in it.  Such an approach is scientifically robust, but not commonly used for detection dogs. Using a model that set a priori constraints for false positive and false negative error rates, we show that dogs trained and tested on a large array system can achieve unambiguous, high and repeatable levels of reliability that are statistically robust.

Authors: Karen L. Overall, Rolf von Krogh, Ann Brinck, Arthur E. Dunham