Choice history biases subsequent evidence accumulation

Anne E Urai, Jan Willem de Gee, Konstantinos Tsetsos, Tobias H. Donner

TL;DR

This paper challenges the prevailing view that past choices primarily influence future decisions by altering the starting point of evidence accumulation. Instead, the authors demonstrate that choice history biases the rate at which sensory evidence is accumulated, effectively biasing the interpretation of current sensory input. This finding has significant implications for understanding how our brains integrate past experiences with present information to make decisions.

What Problem Does This Paper Address?

When we make decisions, especially in perceptual tasks, our choices aren't made in isolation. They're influenced by what we've done before. This "choice history bias" can lead us to repeat or alternate our previous choices. The prevailing computational models for decision-making, like the Drift Diffusion Model (DDM), explain this bias in two main ways: either by shifting the starting point of the evidence accumulation process or by changing the rate at which evidence is accumulated (the "drift rate"). However, it was unclear which of these mechanisms was primarily responsible for choice history biases observed in actual behavior. This paper aims to disentangle these two mechanisms and determine how past choices truly shape ongoing decision-making.

Methods

The researchers employed a sophisticated modeling approach using variants of the Drift Diffusion Model (DDM) to analyze human behavioral data (choices and response times). They fitted these models to data from six different perceptual decision-making tasks across various sensory modalities. Crucially, they compared models where history biases affected only the starting point, only the drift rate, or both. They then analyzed the model parameters to see how individual differences in history biases correlated with the estimated history shifts in starting point and drift bias. To further investigate the dynamics of this bias, they also explored more complex models, including leaky accumulators and models with collapsing bounds, to understand how the bias evolves over time and across multiple past trials.

Datasets

The study utilized data from six distinct perceptual decision-making experiments. These datasets covered a range of task types (2AFC, 2IFC, yes/no), sensory modalities (visual motion, visual contrast, auditory tone), and included variations in experimental design such as fixed duration vs. response time tasks, and the presence or absence of single-trial performance feedback. The specific datasets are detailed in the "Materials and methods" section, including their origins and participant numbers.

Key Insights and Ideas

  • Drift bias, not starting point bias, is the key: The most significant finding is that choice history biases are predominantly explained by a history-dependent change in the evidence accumulation rate (drift bias), rather than a shift in the starting point of the decision process. This challenges the long-held assumption that history biases primarily act as an initial offset.
  • Biasing sensory interpretation: This drift bias suggests that past choices influence how we interpret current sensory evidence. It's akin to our attention being subtly directed towards or away from a particular interpretation of the sensory input based on our previous decision.
  • Long-lasting effects: The influence of past choices on evidence accumulation isn't limited to the immediately preceding trial. The study found that biases could extend over several past trials, with the effect gradually decaying.
  • Rational, but based on faulty assumptions: The observed dynamics of the effective bias signal suggest that participants might be combining prior information (choice history) with current evidence in a rational way, perhaps by assuming environmental stability that doesn't actually exist in the controlled lab setting.

Key Results

  • Model comparison: Models incorporating history-dependent drift bias provided a significantly better fit to the behavioral data across multiple datasets compared to models that only considered starting point biases.
  • Correlation with behavior: Individual differences in choice repetition probability were strongly correlated with the estimated history shift in drift bias, but not with the history shift in starting point.
  • RT distributions: Conditional bias functions revealed that drift bias impacts choices across the entire range of response times (RTs), while starting point bias primarily affects faster responses. The data strongly supported the drift bias explanation, especially for slower responses.
  • Multi-trial effects: History kernels showed that choice history influenced drift bias for up to 2-4 previous trials, with a slower decay than for starting point effects.

Limitations and Future Work

  • Starting point bias still present: While drift bias was the primary driver of individual differences in repetition behavior, a history-dependent shift in the starting point was still observed in a majority of participants, often leaning towards alternation. The authors suggest this might reflect motor planning dynamics or more complex decision dynamics not fully captured by the standard DDM.
  • Alternative scenarios: The authors acknowledge that other decision-making processes, such as fast guesses or post-accumulation biases, might also contribute to behavior, though their models focused on bounded accumulation.
  • Future directions: The paper suggests future work could explore the neural mechanisms underlying these biases, potentially using multi-area neurophysiological recordings. Investigating how these biases interact with explicit attention cues and how they adapt to environments with varying levels of stability are also promising avenues. The authors also propose testing these ideas by manipulating decision speed and response delays.

Summary updated

Abstract

Perceptual choices depend not only on the current sensory input but also on the behavioral context, such as the history of one's own choices. Yet, it remains unknown how such history signals shape the dynamics of later decision formation. In models of decision formation, it is commonly assumed that choice history shifts the starting point of accumulation toward the bound reflecting the previous choice. We here present results that challenge this idea. We fit bounded-accumulation decision models to human perceptual choice data, and estimated bias parameters that depended on observers' previous choices. Across multiple task protocols and sensory modalities, individual history biases in overt behavior were consistently explained by a history-dependent change in the evidence accumulation, rather than in its starting point. Choice history signals thus seem to bias the interpretation of current sensory input, akin to shifting endogenous attention toward (or away from) the previously selected interpretation.
BibTeX
@inproceedings{urai2019choice,
  author = {Anne E Urai and Jan Willem de Gee and Konstantinos Tsetsos and Tobias H. Donner},
  title = {{Choice history biases subsequent evidence accumulation}},
  booktitle = {eLife},
  year = {2019},
  doi = {10.7554/elife.46331},
}
Figure