Unexpected questions in deception detection interviews: Does question order matter?

Abstract

Purpose Unexpected questions have been shown to increase cues to deception, without reducing the information given by truth tellers. Two studies investigated whether the detail given by an interviewee is affected by whether the expected or unexpected questions are asked first. Methods In Study 1, participants (N = 85) were interviewed about their own intentions, and in Study 2, participants (N = 84) were given an intention by the experimenter. They were then interviewed. Results Results showed that in both studies, differences between the expected‐first and the unexpected‐first order were minimal and lie detection accuracy was not improved by asking the unexpected questions first. Conclusions These results offer important information for forensic interviewers, showing that there is no need to ask unexpected questions at a certain point in the interview. Link to associated OSF page: https://osf.io/93g7h/?view_only=586daff060d846efb760c8155478ce9e.

What to do with all these Bayes factors: How to make Bayesian reports in deception research more informative

“…Thus, P (H 1 ) = 0.08 and P(H 0 ) = 0.92. This interpretation of Luke's finding can now be used as a context to evaluate the moderate BF 10 = 3.58 as reported in Warmelink et al (2019) using the Equation (1). A direct calculation shows that the nominal support for H 1 of 8% is then increased to 24%, P(H 1 |data) = 3.58 9 0.08/(0.92 + 3.58 9 0.08), which leaves a posterior probability of 76% in support for H 0 .…”

“…These cut-offs, however, should be used with caution. Let us take the results of Warmelink et al (2019) as a concrete example. In total, they report 22 Bayes factors.…”

“…This distinction is typically ignored within frequentist statistics, where 'non-significant' p values (p > .05) are incorrectly interpreted as an absence of an effect (Altman & Bland, 1995). For this reason alone, Warmelink, Subramanian, Tkacheva, and McLatchie (2019) should be commended for reporting Bayes factors, which can quantify the evidence for the alternative (presence of an effect) over the null hypothesis (absence of an effect), and vice versa. For instance, a Bayes factor BF 10 = 8 implies that the observations are eight times more likely under the hypothesis that the effect is present than under the hypothesis that the effect is absent.…”

Reply to Mac Giolla and Ly (2019): On the reporting of Bayes Factors in Deception Research

“…However, vague prior models are not very useful for testing theories, as it is unlikely that the vague prior model represents a scientific theory of interest. In contrast, Warmelink et al (2019) used prior models that represented their scientific theory (i.e., the prior model is a 'model of H1', see Dienes, 2019, box 4) was specified using a half-normal distribution (indicated by the subscript H in the notation and the solid black line in Figure 1A and 1B). The half-normal distribution represents several aspects of the theory: (I) only effects in one direction are considered plausible, these are represented as positive in the analysis, (II) smaller effect sizes are considered more plausible than larger effect sizes, and (III) effect sizes up to twice the approximate scale-of-effect are considered plausible (the approximate scale-of-effect was 7.05, based on the results of Warmelink, 2012).…”

“…Consider again the Bayes factor discussed above: BH(0, 7.05) = 3.64. Warmelink et al (2019) concluded that they had obtained moderate evidence for an effect, when specifying their model of H1 with a scale-of-effect of 7.05 and a minimum cut-off of B = 3.…”