Bond, Charles

Three Major Meta-Analyses of Deception

Bond has authored three highly influential meta-analyses of deception detection research. Meta-analysis is a study of prior studies. Statistical conversions are used to make prior findings comparable from study to study, allowing researchers to look for trends and to draw conclusions across large numbers of studies.

Bond's first and most influential meta-analysis of deception detection research was conducted with fellow psychologist Bella DePaulo, published in 2006 in Personality and Social Psychology Review. This meta-analysis summarized the results of more than 200 separate research reports, spanning more than four decades of research and just below 300 unique findings from more than 24,000 research subjects. It was a massive undertaking, and it remains the most authoritative and comprehensive summary of the deception detection literature available.

One of the major findings of the 2006 meta-analysis was that across studies, deception detection accuracy averaged just below 54 percent, slightly higher than the 50 percent expected by pure chance. Findings were normally distributed around the cross-study average, with 90 percent of prior findings falling between 45 percent and 62 percent, and 99 percent of all findings either below 68 percent or above 38 percent. The 54 percent accuracy value was statistically better than chance. Thus, the meta-analysis revealed strong, consistent, and compelling support for the idea that deception detection accuracy is slightly better than chance.

The meta-analysis also provided evidence for truth bias. Depending on the statistical approach used, somewhere between 55 percent and 57 percent of judgments were truthful. Presumably, as a consequence, the veracity effect received strong support. Accuracy for truths was 61 percent, whereas accuracy for lies was 47 percent.

The results showed that the slightly better-than-chance accuracy finding was very robust and that several factors made very little difference. It mattered little whether the research used forced-choice dichotomous truth–lie measurement, or whether honesty judgments were scaled. If anything, the dichotomous measures yielded slightly higher accuracy. It also mattered little if accuracy was calculated as a raw average, or if complex signal detection statistics were used. Different measures correlated nearly perfectly. Differences in communication media made little difference (video, no sound, 51 percent; audio, no visual, 53 percent; both audio and video, 54 percent). Motivated liars (54 percent) yielded about the same degree of accuracy as liars lacking motivation (53 percent). Spontaneous lies (53 percent) and prepared lies (54 percent) produced similar findings.

Watching baseline honest communication first provided little help (53 percent versus 54 percent). The same was true for judge expertise (nonexperts, 53 percent; experts, 55 percent) and interaction (no interaction, 53 percent; interaction between sender and judge, 52 percent; interaction with a third party, 54 percent). Regardless, accuracy was slightly better than chance. The most influential factor affecting accuracy was the size of the study. The larger the study in terms of the number of judgments made, the closer the results to the 54 percent average. Thus, the 2006 Bond and DePaulo study documents slightly better-than-chance accuracy in deception detection experiments, regardless of methodology or an extenuating circumstance.

...