An alpha level of .05 was used. Effect sizes were estimated using Cohen's d (for differences between means) and Cohen's w (for differences between proportions). Small, medium, and large effects – respectively – are suggested by cut-offs of 0.2, 0.5, and 0.8 for d and .1, .3, and .5 for w.
Both manipulations had the desired effect. A lower proportion of participants reported that the confession was given voluntarily in the coercion condition (.35) than the no coercion condition (.65), χ2(1, N = 129) = 15.50, p < .001, w = .34. The number of inconsistencies recalled was greater in the inconsistent condition (M = 1.79, SD = 1.00) than the consistent condition (M = 0.05, SD = 0.21), t(67.18) = −13.56, p < .001, d = 2.43. Likewise, ratings of the degree of consistency between the confession and the facts of the crime were lower for participants in the inconsistent condition (M = 2.72, SD = 1.77) than the consistent condition (M = 5.59, SD = 1.30), t(109.58) = 10.38, p < .001, d = 1.86. (Two participants did not provide a rating of perceived consistency.)
The consistency manipulation affected participants' verdicts, but the pattern of results differed from expectations (see Table 1). A 2 (Consistency) × 2 (Coercion) × 2 (Verdict) hierarchical log-linear analysis yielded a significant Coercion × Consistency interaction on verdicts, χ2(1, N = 129) = 4.38, p = .036. Follow-up chi-square tests indicated that, when coercion was present, the proportion of guilty verdicts was higher in the consistent condition that the inconsistent condition, χ2(1, N = 64) = 5.16, p = .023. Likewise, with no coercion present, there was a higher proportion of guilty verdicts in the consistent condition that the inconsistent condition, χ2(1, N = 65) = 22.30, p < .001. But, contrary to our hypothesis, the effect of consistency on verdicts was – if anything – greater when there was no coercion (w = .59) than coercion (w = .28; Table 1 shows 95% CIs around these differences in proportions).
Table 1. The effect of inconsistencies in confession evidence on the proportion of ‘guilty’ verdicts [with 95% CIs] rendered by mock jurors in 'EXPERIMENT 1' and 2
| ||Consistent||Inconsistent||Difference in proportions|
| 'EXPERIMENT 1' |
|No coercion||.91 [.76, .97]||.34 [.20, .52]||.57 [.34, .72]|
|30 of 33||11 of 32|| |
|Coercion||.61 [.44, .75]||.32 [.19, .50]||.28 [.04, .48]|
|20 of 33||10 of 31|| |
| 'EXPERIMENT 2' |
|No motive||.96 [.82, .99]||.40 [.25, .58]||.56 [.34, .72]|
|27 of 28||12 of 30|| |
|Motive||.80 [.63, .91]||.18 [.08, .36]||.62 [.37, .77]|
|24 of 30||5 of 28|| |
These results do not support the notion that jurors will discount inconsistencies in a confession unless there is a salient explanation for them. When an explanation (coercion) was made salient, the effect of inconsistencies on verdicts was actually smaller than when no explanation was made salient. This unexpected pattern of results may indicate a floor effect, whereby the effect of inconsistencies is constrained if a proportion of participants convict regardless of how problematic a confession is. From a broader perspective, the fact that inconsistencies significantly reduced the proportion of guilty verdicts in both coercion conditions runs contrary to the idea that jurors will render a guilty verdict despite the presence of inconsistencies in a confession. These data provide evidence that, in at least some circumstances, jurors' verdicts are influenced by inconsistencies between a confession and the facts of the crime.
We further explored the effect of confession inconsistencies on participants' verdicts by testing whether the effect was mediated by the number of inconsistencies detected by participants (using the INDIRECT macro for SPSS; Preacher & Hayes, 2008). This would indicate that the relationship operates in a graded manner, as opposed to an all-or-none manner whereby the magnitude of the effect does not vary with the number of inconsistencies detected.
The effect of the consistency manipulation on verdicts was fully mediated by the number of inconsistencies detected by participants. The consistency manipulation significantly affected the number of inconsistencies detected, with more inconsistencies detected in the inconsistent condition than the consistent condition, B = 1.75, SE = .13, t = 13.85, p < .001. In turn, the number of inconsistencies detected was related to participants' verdicts, such that more detected inconsistencies were associated with fewer guilty verdicts, B = −1.06, SE = .31, z = −3.38, p < .001. Bootstrap confidence intervals (5,000 samples) indicated a significant indirect effect of the consistency manipulation on verdicts via the number of inconsistencies detected, ab = −1.86, 95% CI [−3.26, −0.83], SE = .63. When the mediator was controlled for, the effect of the consistency manipulation on verdicts became trivial, B = −0.13, SE = .63, z = .20, p = .839. These results suggest that the effect of inconsistencies on verdicts operates in a graded manner.