A non-standard method for estimating accuracy of lie detection techniques demonstrated on a self-validating set of field polygraph examinations

Abstract

A unique method for estimating field accuracy of the Comparison Question Test (CQT) – a polygraph technique – is presented, based on a combined probabilistic and algebraic model. It is built on paired examinations in criminal cases in which two opposing versions per case have been subjected to polygraph tests. The developed model is ground-truth free, thus there was no need to rely on external criteria of deception (e.g., confessions or physical evidence) in estimating the accuracy of the CQT. Results indicate an accuracy rate of 0.94 in detecting guilty examinees (Sensitivity) with a 0.06 False Negative rate and an accuracy rate of 0.835 (Specificity) with False Positive of 0.165 for the innocents. These figures excluded 20% of the cases that were ruled inconclusive. When no inconclusive calls were allowed, the accuracy rate dropped down to 0.8 with 0.2 error rates for both the guilty and the innocent examinees. The importance of this research stems from its being a field study that due to the unique methodology was not subjected to weaknesses usually found in polygraph field validity studies. This method is applicable to other techniques of deception detection and with some necessary adaptations may be also to eyewitness situations.

Keywords:

• polygraph
• CQT
• field paired examinations
• accuracy rates
• self-validating examinations
• detection of deception

View correction statement:

Corrigendum

Notes

This article was originally published with errors. This version has been corrected. Please see Corrigendum (http://dx.doi.org/10.1080/1068316X.2013.765137).

1. In order to convict a suspect the guilt must be proven beyond a reasonable doubt. Thus, although not perfect, one can count on convictions to a very high degree to be correct. This is not the case with acquittals, since a reasonable doubt is enough to acquit a suspect and obviously this approach results in quite a number of guilty suspects with false negative error acquittal verdicts.

2. The Innocence Project is an US national litigation and public policy organization dedicated to exonerating wrongfully convicted people through DNA testing and reforming the criminal justice system to prevent future injustice. http://www.innocenceproject.org/

3. IDENTA 85 – The international congress on techniques for criminal identification & counter terrorism. Jerusalem, Israel 1985.

4. In 1988, the research model was presented with some preliminary results, in a close forum, however, it has never been published in the scientific literature, and remains unheard and unknown to most of the relevant community.

5. The probability of getting independently 26 records belong to 25 different pairs or more, is about 0.21. So, though it is relatively a rare event it is still within acceptable limits.

6. One can raise the point that also in a large proportion of single examinee cases, the suspect is accused by someone who is very unlikely to be a suspect, providing the same anxiety as in the paired cases. But as a matter of fact all these cases are potentially paired examinees cases, because the accused suspect's denial of being guilty, might means that the accusing person become an alleged suspect of giving false testimony or complain.

7. A NDI outcome can be a True Negative, p=0.835, or a False Negative, p=0.06. Since the prior probability of telling the truth or being deceptive is 0.5, the weight given to both options are the same. Thus, the conditional probability that this specific outcome is actually correct equals 0.835/0.835+0.06=0.835/0.895=0.93. A DI outcome can be a True Positive, p=0.94, or a False Positive, p=0.165. Thus, the conditional probability that this specific outcome is actually correct equals 0.94/0.94+0.135=0.94/1.105=0.85.

8. Personal knowledge based on internal annual reports of Israel Police Polygraph Labs.

9. Two Opposite Outcomes can either be two correct results or two mistakes. Probability of getting two correct results is 0.8² =0.64. Probability of getting two incorrect results is 0.2²=0.04. Probability of two correct results, given two Opposite Outcomes is ^0.64/0.64+0.04 =^0.64/0.68=0.94.

10. Two opposite outcomes can either be two correct results or two mistakes. Probability of getting two correct results is 0.94*0.835=0.785. Probability of getting two incorrect results is 0.06*0.165=0.0099. Probability of two correct results, given two Opposite Outcomes is ^0.785/0.785+0.0099=^0.785/0.7949 =0.987.