TL;DR
Sensation-seeking is traditionally viewed as a drive for intense, often risky experiences, but this paper suggests a different underlying motive: a hunger for information. By reanalyzing decision-making data with Bayesian models, the authors show that high sensation-seekers aren't necessarily "liking" aversive stimuli like electric shocks; rather, they are significantly less avoidant of the uncertainty surrounding those stimuli.
What Problem Does This Paper Address?
Sensation-seeking (SS) is a personality trait linked to everything from extreme sports to substance abuse and antisocial behavior. Historically, researchers thought sensation-seekers simply assigned a higher "economic value" to intense sensations—meaning they might actually enjoy things others find aversive.
However, the authors of this paper argue that this "value-based" explanation is incomplete. They propose that SS might instead be driven by the exploration-exploitation trade-off. In this view, sensation-seekers are "information-seekers" who are willing to endure aversive outcomes (like a mild electric shock) if it helps them resolve uncertainty about their environment. The gap the authors tackle is whether SS is rooted in a distorted valuation of the stimulus itself or a strategic preference for informative, albeit risky, environments.
Methods
The researchers reanalyzed a dataset from a previous study (Norbury et al., 2015) using a suite of computational models to see which best explained participant behavior.
The task was a "two-armed bandit" game. Participants chose between different visual stimuli (bandits) that yielded points. In the test phase, some bandits were associated with a 75% chance of a mild electric stimulation (MES).
The authors compared four types of models:
- The Original Model: Assigned a fixed, binary value to the shock.
- Rescorla-Wagner (RW): A standard reinforcement learning model that updates the expected probability of a shock based on trial and error.
- Bayesian Learning (Mean Policy): Modeled shock probabilities as Beta distributions, focusing on the expected mean probability.
- Bayesian Learning (Uncertainty Policy - BLvar): This novel addition modeled not just the probability of a shock, but the uncertainty (variance) of that probability.
They used Hierarchical Bayesian Modeling (HBM) implemented in STAN to estimate parameters and used Leave-One-Out Information Criteria (LOO-IC) for model comparison.
Datasets
The study pooled data from two previous experiments:
- Study 1: 39 healthy subjects (after exclusions) performing the sensation-seeking task.
- Study 2: 28 healthy male subjects who performed the task under three conditions: baseline, placebo, and a 2.5mg dose of haloperidol (a dopamine antagonist).
- Total N: 73 subjects for the primary analysis.
Key Insights and Ideas
- Information Over Value: The winning model (BLvar) showed that choices were guided by the uncertainty of the shock rather than just its presence or its mean probability. Most people are "uncertainty-avoidant" (they dislike not knowing if a shock is coming), but high sensation-seekers show significantly reduced avoidance.
- The "Less Avoidant" Sensation-Seeker: High SS scores didn't correlate with "liking" the shock. Instead, they correlated with the $\omega$ (omega) parameter, which represents the weight given to uncertainty. High sensation-seekers were simply less bothered by the "unknown" of the shock.
- Dopamine's Role is Noisy: Interestingly, blocking dopamine with haloperidol didn't change the information-seeking strategy ($\omega$). Instead, it increased the $\tau$ (tau) parameter, making participants' choices more "stochastic" or random. This suggests dopamine might regulate the "noise" in our decisions rather than the core strategy of seeking information.
Key Results
- Model Fit: The BLvar model (Bayesian learning with an uncertainty-based choice rule) outperformed all other models, including the original economic model.
- Correlation with SS: There was a significant positive correlation ($\rho = 0.29, p = .02$) between self-reported sensation-seeking and the $\omega$ parameter. High SS individuals had "flatter" response curves, meaning their probability of choosing a bandit didn't drop as sharply when uncertainty increased.
- Subscale Specifics: The link to uncertainty was strongest for the "Experience Seeking" and "Disinhibition" subscales of the Sensation-Seeking Scale (SSS-V).
- Drug Effects: Haloperidol significantly increased choice stochasticity ($p < .001$) but had no significant effect on uncertainty avoidance ($p = .32$).
Limitations and Future Work
The authors acknowledge a major hurdle: collinearity. In the original task design, bandits without shocks (CS-) had zero variability, while bandits with shocks (CS+) had high variability. This makes it difficult to perfectly tease apart whether a participant is seeking the shock itself or seeking the information provided by the probabilistic nature of the shock.
Future research should aim to "orthogonalize" these factors—creating scenarios where high-value rewards and high-uncertainty shocks are decoupled. The authors also suggest that studying "extreme" groups (like professional skydivers or clinical populations) might provide a clearer signal than the healthy student populations used here.
Summary updated