Learning

How People Search Before They Choose
Why No Single Model Captures Choice Across all Environments

How People Search Before They Choose

Every decision rests on information that must first be gathered—either recalled from memory or found in the external environment. From browsing restaurant reviews to comparing investment options online, modern life increasingly offers opportunities to explore the information ecology freely before committing to a consequential choice.

What drives exploratory behavior? Theoretical accounts have long assumed that people search primarily to reduce "estimation uncertainty"—the error in judging an option's average value. Yet this narrow view has rarely been tested against large-scale empirical data. ARC addressed this gap by analyzing a database of over 1,000,000 information-search decisions made by more than 2,500 individuals to uncover how people allocate samples to options and when they decide to stop searching (Spektor & Wulff, 2024).

The central finding was that predecisional information search is predominantly discovery-driven. That is, participants deciding which option to sample next overwhelmingly favored options with undiscovered outcomes over those with higher estimated values or greater estimation uncertainty. Notably, this frequently involved sampling from options with zero experienced variance, at odds with uncertainty-driven search. At the aggregate level, 63% of sampling decisions aligned with discovery-driven search, and a classification analysis revealed that nearly half of all participants were best described by this strategy. Crucially, once all possible outcomes had been observed, implying no further outcomes can be discovered, participants shifted toward reducing estimation uncertainty—suggesting that they made adaptive, sequential transitions between search modes.

The analysis also uncovered signatures of fixed-search strategies: tendencies to stop at round sample sizes (e.g., 10, 20, or 40) and to allocate samples evenly across options, simplifying the computational demands of tracking frequencies. These patterns point to a third type of uncertainty reduction: computational uncertainty. Together, the results demonstrate that how people search before they choose involves trading off multiple objectives in sequence, guided by explicit mental representations of what has and has not been observed.

Key Reference

Why No Single Model Captures Choice Across all Environments

Choice behavior under described risk (known probabilities) and experienced uncertainty (unknown probabilities) are qualitatively different, as are simple lotteries (with one or two outcomes each) and complex lotteries (three or more outcomes each). Spiliopoulos and Hertwig (2023) investigated how the interactions of these two dimensions affect choice behavior. Using a large within-subjects experiment that improved on earlier, less powerful between-participant and cross-experiment comparisons, the study stress-tested five decision models—cumulative prospect theory (non-cumulative prospect theory, CPT; expected utility; mean-variance-skewness, MVS; and disappointment-elation-skewness, DES)—across four environmental niches (described–simple, described–complex, experienced–simple, and experienced–complex). Of the decision models, CPT, non-cumulative prospect theory, and expected utility theory rely on subjective probabilities, whereas MVS and DES are driven by moment-based preferences.

The study replicated the well-known decision–experience gap and brought to light a new finding: The gap was significantly moderated by the skewness of lotteries in complex environments. Bayesian hierarchical mixture modeling showed that the widely used CPT was the clear dominant decision model in only the described–simple environment. It was significantly outperformed by the MVS and DES models in the two complex environments. This implies that indiscriminately applying CPT to more complex lotteries and uncertainty is problematic. Crucially, choice behavior was found to be context-dependent: No single model performed best across all environments. 

The authors also exploited participants' elicited subjective beliefs to show that they were more attuned to the variance and skewness of lottery distributions in complex and experienced environments than to explicit subjective probabilities (as assumed by CPT). Finally, the abundance of within-participant data allowed a powerful test of whether decision-makers are heterogeneous, not just in the models they use, but also in specific parameter values within those models. The data showed that participants were indeed heterogeneous in how they arrive at decisions—an important finding for future empirical work.

The most important conclusion from this research is that seeking a single universal theory of decision-making across different environments is not viable. Instead, researchers should explore context-specific models that capture decision-makers' adaptation to environmental characteristics.

Key Reference