Uncertainty is known to affect choice, but less is known about its effects on the sampling process leading up to choice. While secondary-data studies have found uncertainty to affect sampling behaviour in decision-from-experience tasks, uncertainty was only inferred from decisions but not directly assessed. We designed an experiment with greater control over uncertainty manipulations while also measuring self-reported uncertainty.

One hundred and thirty-seven participants completed 6 decision-from-experience binary lottery trials, freely sampling from two discrete distributions and finally selecting one. We varied across trials the information given on the distributions and their properties (variance, number of distinct outcomes). Self-reported uncertainty was assessed multiple times throughout the sampling process.

We found that people sampled more when presented with a higher-variance distribution, while sampling did not increase when the distribution had a larger number of distinct outcomes. When measured after five samples were drawn, higher self-reported estimation uncertainty—uncertainty about the distribution's average—was associated with more samples being drawn in that trial. However, the same was not found for self-reported structural uncertainty—uncertainty about the distribution’s characteristics. Additionally, more sampling increased the accuracy of the representation of the outcome distribution in a distribution-builder task.

Our findings are in line with theories that see sampling as being driven by uncertainty reduction. They also make an initial case for wider implementation of self-report measures of uncertainty: Currently largely underutilised, they allow for real-time trial-level measurements. To further corroborate our findings, we plan to test computational measurement frameworks from the literature on this data.