By Bharath Chandra Talluri & Tobias Donner.
Confirmation bias is the tendency to interpret new information such that it agrees with one’s pre-existing beliefs. This cognitive bias is ubiquitous in human reasoning and has been known for a long time. Take, for example this quote from Dante’s Divine Comedy (14th century):”Opinion—hasty—often can incline to the wrong side, and then affection for one’s own opinion binds, confines the mind”. Yet, the underlying mechanisms have remained elusive. One reason is that it has, so far, been challenging to study the phenomenon under laboratory conditions that afford precise experimental control over the phenomenon over many repeated trials – the basis for thorough and quantitative analysis. Another reason is that previous studies into confirmation bias have used tasks requiring higher-level, abstract judgments, the neural bases of which are currently elusive.
In this study, we show that confirmation bias arises through selective overweighting of evidence consistent with a previous choice, akin to feature-based attention. To this end, we have developed a novel task protocol that allowed us to induce confirmation bias in a systematic and reproducible fashion and to quantify it in individual participants by means of psychophysical and computational modelling techniques. The task protocol is generic, and can be applied to many different forms of decisions. Here, we used two versions of the protocol: a low-level perceptual task entailing judgment of visual motion direction as well as a higher-level cognitive task (numerical integration). We founding robust signatures of confirmation bias in both tasks. Establishing confirmation bias in low-level judgments about visual motion (obviously rather neutral for participants) is particularly striking, and points to a mechanism deeply built into the brain’s decision-making machinery. Further, because visual motion decisions have been extensively studied in computational neuroscience and neurophysiology over the past two decades, the discovery of confirmation bias in this form of judgment provides insights into the mechanistic basis.
Subjects saw two intervals of noisy random dot motion stimuli (panel A in Figure 1 below), and were required to give a continuous estimate of the average direction of motion in the two intervals. Critically, subjects were also asked to make a fine discrimination binary choice about the direction of dots w.r.t a reference line after the first interval. The position of the reference line was fixed throughout the trial, but changed randomly between trials. The signal dots were sampled from five possible directions (-20°, -10°, 0°, 10°, 20° w.r.t the reference line) independently in the two intervals. The task allowed us to quantify the influence of the binary choice on stimulus information from the second interval.
Figure 1
Using extensive computational model-based and model-free analyses, we found that subjects assigned greater weight to stimulus in interval 2 that was consistent with their binary choice, compared to an inconsistent stimulus (panels B & C in Figure 1). We ruled out several alternative explanations through formal model comparisons as well as – more importantly – through the targeted analysis of diagnostic features in the behavioural data.
We believe our work opens doors to pursue several exciting questions in the future, with the rigorous tools of cognitive computational neuroscience. Our task protocol allows us to quantify confirmation bias in low-level sensory decisions whose neural basis is well established. The multiplicative modulation coupled by the top-down nature of the mechanism strongly suggests that choices act as cues to direct attention selectively to certain stimulus features.
Confirmation biases are pervasive in daily life, and can have unintended consequences especially in cases of critical significance like evaluating job candidates, or making policies that impact a large section of the society. Even scientific hypothesis testing is susceptible to this bias, where researchers tend to favour results that support their prior hypotheses thus questioning the generality of such findings. We hope our study inspires fellow scientists to acknowledge this societally important cognitive phenomenon, and to pursue its mechanistic and neural basis. Finally, our results highlight the recurrent interplay between decision-making and selective attention, two cognitive phenomena often studied in isolation but intricately linked in real life.