This post describes my latest paper: Urai AE, de Gee JW, Tsetsos K, Donner TH. (2019) Choice history biases subsequent evidence accumulation. eLife, 8:e46331, and was reposted from the DonnerLab website.
To study the mechanisms of decision-making, researchers often treat individual decisions as isolated events. However, as we go around the world, our decisions can be strongly influenced by previous experiences. Even in cases where it may be best to focus only on the current situation, our choices can be hard to separate from contextual factors. For example, studies of perceptual decision-making use well-controlled, but artificial, stimuli about which observers have no prior knowledge. Still, it has long been known that the choice an observer made seconds ago influences what she will report next.
This phenomenon, called choice history bias, is increasingly being understood as resulting from the interplay of two factors: the accumulation of momentary sensory evidence for each decision, and slow statistical learning across multiple decisions. This learning process leads to the ongoing updating of prior expectations from one decision to the next, which then bias the current choice. We, and others, have previously found that these choice history biases exhibit several hallmark signatures of a rational and adaptive learning process – for example, they are flexibly adjusted to the stability of the environment, and modulated by the confidence in previous decisions (Urai et al. 2017; Braun et al. 2018).
Here, we set out to answer an open question. How is this bias (or prior expectation) generated in the slow learning process combined with current sensory information to make people repeat or alternate their decisions? In other words, how do previous choices affect the dynamics of a decision, from seeing a stimulus to reporting what you saw?
The process of forming a choice is described by a process of evidence accumulation: information that arrives over time is integrated, a process which averages out noise in the sensory stimulus. This integrated ‘decision variable’ grows towards one of two ‘decision bounds’, that when reached indicate that enough information has been gathered to commit to a certain choice.
We fit several variants of such sequential sampling models of decision-making to a large number of datasets (n = 194 people, doing 6 different perceptual choice tasks). In such models, a biased decision process can arise in two principal ways – illustrated schematically below for a widely used instance of bounded-accumulation model called the drift diffusion model (DDM; Figure 1a). Previous choices could shift the starting point of the next evidence accumulation process (Figure a, left), in other words: and additive offset to the decision variable. Alternatively, previous choices could change the rate (‘drift’) at which evidence for one vs. the other option is accumulated, by adding an evidence-independent constant to the drift. Because that constant is integrated over time, along with the sensory evidence (see equation), the impact of the bias on choice grows with elapsed time. Combining information from choices and response times enabled us to disentangle those two scenarios in our data.
People’s individual tendency to repeat or alternate their own choices was captured by a change in their history-dependent drift bias (Figure b). Not only was this main finding present regardless of whether the previous choice was correct or not, we also found that the effect of choice history lasts multiple trials into the past (Figure c, left). We then fit a set of more complex models to the data, showing that choice history bias most likely arises from unbalanced input (think of the stimulus-selective neuronal responses in sensory cortex) to the accumulators that govern behavioral choice (Figure c, right). This is intriguingly similar to the mechanisms by which our visual system can selectively pay attention to only some parts of an image, depending on
We are excited about this work for a number of reasons. By using both choice data and response times, we were able to get a close look at the fine-scale dynamics that govern our choices, and how these dynamics change with previous decisions. This calls for the revision of assumptions in previous models, that mostly assumed choice history would affect the starting point of the next decision. This project also shows the strength of using multiple datasets, which allowed us to establish the generality of the findings. It highlights the importance of convergent evidence from different approaches: fitting several different separate implementations of bounded-accumulation models convinced us of the robustness of our interpretations. Lastly, these findings also open up exciting new avenues for finding neurophysiological correlates of this biasing process, and pinpointing how the brain uses such choice history in the evidence integration process.