ates (usually but not always psychology students); military trainees; other workplace volunteers; and research subjects recruited through employment agencies. Although samples drawn from these sources are not demographically representative of any population on which polygraph testing is routinely performed, neither is there a specific reason to believe such collections of examinees would be either especially susceptible or refractory to polygraph testing. Since the examinees thus selected usually lack experience with polygraph testing, we will loosely refer to the subjects from these studies as “naïve examinees, untrained in countermeasures.” The degree of correspondence between polygraph responsiveness of these examinees and the special populations of national security employees for whom polygraph screening is targeted is unknown.
Many of the studies collected data and performed comparative statistical analyses on the chart scores or other quantitative measures taken from the polygraph tracings; however, they almost invariably reported individual test results in only two or three decision classes. Thus, 34 studies reported data in three categories (deception indicated, inconclusive, and no deception indicated, or comparable classifications), yielding two possible combinations of true positive (sensitivity) and false positive rates, depending on the treatment of the intermediate category. One study reported polygraph chart scores in 11 ranges, allowing extraction of 10 such combinations to be used to plot an empirical receiver operating characteristic (ROC) curve. The remaining 17 used a single cutoff point to categorize subjects relative to deception, with no inconclusive findings allowed. The median sample size of the 52 datasets from laboratory studies was 48, with only one study having fewer than 20 and only five studies having as many as 100 subjects.
Figure 5-1 plots the 95 combinations of observed sensitivity (percent of deceptive individuals judged deceptive) and false positive rate (percent of truthful people erroneously judged deceptive), with straight lines connecting points deriving from the same data set. The results are spread out across the approximately 30 percent of the area to the upper left. Figure 5-2 summarizes the distribution of accuracy indexes (A) that we calculated from the datasets represented in Figure 5-1. As Figure 5-2 shows, the interquartile range of values of A reported for these data sets is from 0.81 to 0.91. The median accuracy index in these data sets is 0.86. The two curves shown in the Figure 5-1 are ROC curves with values of the accuracy index (A) of 0.81 and 0.91.1
Three conclusions are clearly illustrated by the figures. First, the data (and their errors of estimate; see Appendix H, Figure H-3) clearly fall above the diagonal line, which represents chance accuracy. Thus, we conclude that features of polygraph charts and the judgments made from them are correlated with deception in a variety of controlled situations