Confidence Clouds Evidence-Based Decision Making

Post by Elisa Guma

What's the science?

Confirmation bias is the tendency to search for, interpret, or favour information that confirms or supports one’s prior beliefs. This type of cognitive bias is most evident when opposing parties are very confident in their beliefs and it can have a significant impact on societal functioning by further polarizing beliefs. It can be most troubling when the evidence against one’s position is selectively disregarded. This week in Nature Communications, Rollwage and colleagues sought to investigate the underlying cognitive, computational, and neuronal mechanisms underlying confirmation bias, as they are poorly understood.

How did they do it?

In order to investigate the mechanisms underlying confirmation bias, the authors applied theoretical models to data collected while participants performed a behavioural task and magnetoencephalography (MEG) recordings of brain activity were taken. During the task, participants were shown a series of randomly moving dots (‘pre-decision evidence’) and indicated using arrow keys whether the majority of the dots were moving towards the left or the right of the screen (‘initial decision’) and how confident they were. Reaction time to key press was measured. Next, participants were presented with ‘post-decision evidence’: a second set of moving dots in which the majority of dots were moving in the same direction as the pre-decision evidence. In order to manipulate participants’ confidence without altering reaction times, participants were presented with two conditions for the post-decision evidence; a high positive evidence condition in which the proportion of dots moving in the incorrect direction was 15%, and the dots moving in the correct direction was higher than 15%, or a low positive evidence condition in which only 5% of dots were moving in the incorrect direction, but dots moving in the correct direction were also less than in the high positive evidence condition.  At the end of each trial, participants provided their final decision and another confidence rating. 

The authors hypothesized that confidence would reduce the frequency of participants changing their minds by promoting a bias towards the processing of confirmatory post-decision evidence. To further probe this behaviour, the authors modeled behavioural data using an analysis technique common in decision making studies (drift-diffusion modeling) to test whether there was a drift in response to post-decisional evidence based on selective accumulation of evidence in line with the participant’s initial decision, and how confidence affected this drift rate from the initial decision. This type of modeling allowed the authors to make indirect inferences about how confidence affects evidence accumulation, so they turned to their time-resolved MEG data. They trained a support vector machine classifier (machine learning model) to predict which choice was made on each trial using brain activity in the pre-decision window. The trained classifier was then applied to brain activity in the post-decision window to predict the probability of neural evidence favouring one decision over another.

What did they find?

The authors observed that participants who received a stronger confidence boost from the positive evidence (high positive evidence condition) presented to them in the task showed a greater reduction in changes of mind, independent of the accuracy or reaction time in their responses. This evidence confirmed the authors’ first hypothesis; that confidence is a critical driver of changes of mind. Next, the authors fit the accuracy and reaction time data using theoretical (drift diffusion) modeling. The model with the best fit indicated that when they were highly confident in their choice, participants started the accumulation of evidence process closer to the bound of their initial decision. Further, participants selectively accumulated evidence supporting their initial choice faster (a faster ‘draft rate’) when they were more confident in their choice.

The influence of confidence on drift rate is clear evidence for confirmation bias. Based on their modeling of the MEG data, the authors observed that there was a confidence-induced confirmation bias driven by selective accumulation of choice-consistent information wherein high confidence leads to an amplified neural response to confirmatory evidence and a blunted response to contradictory evidence. Further, they found that centro-parietal MEG sensors contributed most strongly to the individual’s decision-making activity.

What's the impact?

By combining behavioural and neural modeling, this study provides experimental evidence to suggest that high confidence in a decision leads to behavioural confirmation bias and has striking effects on the way our brain processes post-decision evidence. Since the authors observed such biases in a low-level perceptual task, it may suggest that this bias is a core principle of neural information processing. However, most real-world decisions have added motivational, emotional, and social influences that may amplify, or change the way in which confidence affects the processing of post-decisional evidence. Insights gained here may be applied to better understand the drivers of polarization across a range of societal issues.

Rollwage et al. Confidence drives neural confirmation bias. Nature Communications (2020). Access the original scientific publication here.