Self-Consistency in Sequential Decision-Making

Human decisions are rarely made in isolation. We typically have to make a sequence of decisions to reach a goal. Studies in economics and cognitive psychology have shown that making a decision may result in several biases in subsequent judgments. Similar biases have also recently been found in human percepts of low-level stimuli such as motion direction. What lacking is a principled framework that can account for several sequential dependencies between judgments. Towards that goal, in my thesis, I propose and experimentally test a self-consistent Bayesian observer model that assumes humans maintain self-consistency along the inference process. In Chapter 2, I first demonstrate that after having made a categorical decision on stimulus orientation, subjects’ estimate of the stimulus is systematically biased away from the decision boundary. Two additional experiments suggest that the bias occurs because subjects treat their first decision as a fact and use that to constrain the subsequent estimation. Model fit to the data in my experiments and data in previous studies show that the self-consistent Bayesian model can quantitatively account for human behaviors in a wide range of experimental settings. In Chapter 3, using the same decision-estimation tasks, I probed the post-decision sensory representation by providing feedback on the categorical decision. I found that subjects’ sensory representation is kept intact and the self-consistency is implemented by conditioning the prior distribution on the categorical decision. The results also suggest another interesting form of self-consistency when subjects’ decision was incorrect: they reconstructed the sensory measurement to make it consistent with the given feedback. In Chapter 4, I found that the choice-induced bias also occurs in human judgment of number. The bias is similar for both non-symbolic (cloud of dots) and symbolic (sequence of Arabic numerals) forms of number. Finally, I propose in the general discussion how the self-consistent Bayesian framework may account for other biases in sequential decision-making such as the halo effect and sunk-cost fallacy