Online Processing Interventions

Belief bias is a contextual bias; that is, it manifests predominantly not when a problem is considered alone, but in the context of other problems, some of which have believable conclusions, and others with unbelievable conclusions. When manipulated within-participants, the belief bias effect is often diminished (e.g. Evans & Pollard, 1990) or eliminated (Trippas, Verde, & Handley, 2014). This trend is also found when participants are asked to choose the valid syllogism from an unbelievable and a believable syllogism presented side by side (Trippas et al, 2014, Experiments 1 and 2).

Milkman, Chugh, and Bazerman (2009) discuss an approach to debiasing in decision-making research, which involves providing individuals with an alternative heuristic to the one which is the source of their bias, with the replacement being one which results in more accurate responding. However, given the strength of the belief bias effect, this seems unlikely to prove useful. Alternatively, as belief bias may be caused by an over-reliance on Type 1 responses, providing reasoners with a

strategy which explicitly engages a Type 2 response may produce improvements in performance. This approach has been found in other areas of cognition to lead to novice task performance becoming more accurate than that of experts (Dawes, 1971). The use of simultaneous presentation to encourage a shift from Type 1 to Type 2 responding discussed above has also been found in other domains (e.g. Bazerman, Loewenstein, & White, 1992). However, in everyday life, it is rare that logical arguments are encountered in such a specific context, and so alternative ways of altering processing are still important to examine.

The use of disfluent texts has been argued by some to be effective in this regard; Thompson, Prowse-Turner, et al. (2013) found that participants of higher cognitive ability were more accurate on the Cognitive Reflection Task (CRT; Frederick, 2005) when the text of the task was presented in a font which was difficult to read. They concluded that this was because it promoted extra processing. However, Trippas, Handley, and Verde (2014) found that, in the case of syllogistic reasoning, disfluency actually led to lower accuracy for higher ability participants, concluding that on complex tasks, disfluent fonts leads to additional processing which would otherwise be used to deduce a correct answer. Meyer et al. (2015) reported the results of research in which 16 separate experiments attempted to replicate the effect of disfluent fonts on improving performance. They found no effects of disfluency at all, even when cognitive ability was accounted for, and concluded the only consistent change in performance was an increase in response time.

Feedback has been used in other domains such as memory and decision making to improve task performance. When rapid feedback is available on a large number of decisions for which ratings of confidence have been given, individuals show a stronger link between accuracy and confidence in that domain (Wagenaar & Keren, 1986). Feedback and confidence ratings have also been used as a debiasing technique in a classroom setting, with participants only showing an increase in accuracy after being given both immediate feedback and having to give confidence ratings (Renner & Renner, 2001). Indeed, it has been argued that response confidence acts

as a metacognitive cue for reasoners in determining whether they give an answer consistent with Type 1 responding or Type 2 responding (Thompson, Prowse- Turner, et al., 2013). Therefore, asking reasoners to give a rating of confidence in a given answer, followed by evaluative feedback may lead to better calibrated confidence and higher accuracy.

Specific training can be an effective debiasing technique, as rather than attempt- ing to modify behaviours, it simply gives people “those tools needed to arrive at correct answers” (Arkes, 1991). The benefits of undertaking a course in statistics, or training in statistical principles in the lab, have been shown to be transferable to a later task involving statistical inference (Fong, Krantz, & Nisbett, 1986). As dis- cussed earlier, inhibition may play an important role in the suppression of intuitive response on conflict problems. Houd´e et al. (2000) attempted to debias participants completing a matching bias task. Participants were given a pre- and post-test, with an inhibition training task in-between. It was found that participants showed lower levels of matching bias on the post-test, and increased activation in areas of the brain thought to be associated with inhibitory control. One problem with taking this as evidence of the success of inhibition training, however, is that Houd et al did not include a control condition in their experiment. Moutier, Angeard, and Houd´e (2002) compared the effects of inhibition training and general logic training on performance on the matching bias task. It was found that inhibition training improved performance, but logic training had no effect.

Later research by Moutier and Houd´e (2003) examined debiasing of the con- junction fallacy. The conjunction fallacy is often demonstrated by participants’ response to the Linda problem (Tversky & Kahneman, 1983):

“Linda is 31 years old, single, outspoken, and very bright. She majored in philosophy. As a student, she was deeply concerned with issues of discrimination and social justice, and also participated in anti-nuclear demonstrations.

Linda is a bank teller.

Linda is a bank teller and is active in the feminist movement.”

When given such a task, individuals tend to choose the second option, fitting in with the extra information that they have been given. However, logically, the first option is more likely given that membership of a single category cannot be probabilistically more likely than membership of that category in addition to another one. Moutier & Houde (2003) found that inhibition training was the key to enhanced performance on this task. When participants were given training which shared many underlying features with the main task, little improvement in performance was shown compared to a control group who were given no such training, even when the underlying logic of the task was explicitly discussed, and it was checked that the participants understood it fully. However, an additional condition combined this logic training with an inhibition component, in which the experimenter explained why the participant’s intuitive response may be misleading, and had the participant repeat the explanation back to them until it was clear that they had a complete understanding of it. It was found that in this condition participants showed significantly less bias on a post-test than the control or logic- trained group, indicating that the demonstration of principles of inhibition were key to debiasing responses on this task.

Logic training alone is thought to fail, as people struggle to apply abstract the concepts they have been trained in to the specific tasks they are required to complete. Specific task training alone is also ineffective; to get any benefit, individuals must be simultaneously trained in both the specific task they will be tested on, and the underlying logical principles (Bransford, Sherwood, Vye, & Rieser, 1986; Cheng, Holyoak, Nisbett, & Oliver, 1986).

Task-specific training has been shown to improve performance in syllogistic reasoning tasks; Prowse-Turner and Thompson (2009) found that a short training session in which participants were taught how to construct diagrams to represent syllogisms, and then training on just 5 practice syllogisms subsequently improved

performance on a later task in which 16 syllogisms were evaluated. A more detailed approach was taken compared to the typical binary response design, which showed participants became better at categorising syllogisms on the basis of logical necessity. However, this study did not use belief-oriented materials, and it is unclear whether this technique would be sufficient to override an effect of such magnitude. In addition, Prowse-Turner & Thompson added further nuance to the issue of confidence rating data; item-by-item confidence was no better calibrated with training. However, a single overall estimate of performance was found to be significantly more accurate for the group who did receive training.