PROJECT:
Predicting When People Think Information Is “Worth It”
Information Seeking ❖ Decision Modelling ❖ Gaussian Process Models ❖ Model Comparison
People sometimes gather additional information before deciding, but skip it in other situations. When do people think information is “worth it”, and why?
I designed a controlled behavioural task and built a new quantitative model to conduct causal inference on information seeking. I found that people selectively seek information to resolve conflict between possible actions.
MY ROLE: From a partially collected dataset, I delivered the full modelling pipeline:
analytical strategy
data cleaning and feature engineering
modelling (Gaussian Process regression, model comparison)
visualization and interpretation
insight communication
BEHAVIOURAL INSIGHTS
Research question
When do people seek extra information before deciding, and why? This question has wide implications; understanding the causal trigger for information seeking allows for precisely targeted and efficient decision support.
Study design
In a controlled task, participants chose between two options that yielded probabilistic rewards. Across choices, we manipulated:
which option was more desirable (higher reward)
which option participants believed was likely to yield rewards, and to what extent
Participants could either choose immediately or gather extra information first.
In the experimental task, participants chose between two options. Each option had a different likelihood of yielding rewards, visually indicated by the amount of evidence (ovals). Before deciding, they could gather extra information about how likely each option was to yield a reward.
Finding
Participants sought additional information not when they were most uncertain about which option was likely to reward them, but when the desirable option was currently less likely to reward them. In other words, participants’ information seeking was triggered by a conflict between the two options (the one desirable and the other supported by the current evidence).
Therefore, people strategically seek information to resolve conflict between possible actions. This means that decision support targeted not at general uncertainty, but specifically at decision conflicts, will be most effective.
STATISTICAL MODELLING
Challenge
Prior to this project, there was no established modelling approach to information seeking. I developed a new approach to uncover the relationship between information seeking and the key experimental variables (which option was desirable, which option was expected to be rewarding). It was critical to estimate the relationship without imposing a specific structure (e.g., linear, quadratic), so that I could conduct a valid statistical inference on its shape.
Approach
I used Gaussian Process (GP) logistic regression. GP regression is a principled way to estimate smooth functions without imposing a form. It is a powerful tool in any domain where we need to uncover interpretable behavioural patterns shaped by multiple factors in some unknown manner.
Results
GP regression revealed that participants gathered information the most when the current evidence favoured the undesirable option, supporting our behavioural insight that information seeking is triggered by conflict between possible actions.
Participants’ tendency to gather extra information peaked not when both options were equally likely to yield rewards, but when the undesirable option was slightly more likely to yield rewards.
To further evaluate the statistical evidence for this claim, I conducted model comparison by cross validation (leave-one-participant-out). My original model (driven by a conflict between actions) outperformed alternatives.
This demonstrates that GP regression, combined with proper techniques such as cross validation, supports rigorous statistical inference while remaining flexible and interpretable.
Note: Modelling code and behavioural data are available at https://gitlab.com/kenji.k/beadsVOI/.