Eye Movement Models of Decision Making

Eye-tracking has become a quantitative instrument for decision research because where someone looks—and for how long—is structured by the same cognitive process that produces the choice itself. The challenge for psychometrics has been to write down formal models that link the millisecond-resolution stream of fixations and saccades to latent variables: attention weights, accumulating evidence, value signals, and strategy states. Wedel, Pieters, and van der Lans (2023) survey this enterprise in Psychometrika, organizing four decades of disparate models around a unifying account of search and choice tasks in which task switching and strategy switching are treated as the central explanatory variables.

Why eye movements are a measurement instrument, not just a behavior

Modern eye-trackers record fixation locations at sampling rates of 60-2000 Hz, producing tens of thousands of data points per minute per participant. Two features of those data are exploitable. First, attention is approximately overt during decision-making: people look at the option they are currently evaluating, with covert attention shifts the exception rather than the rule. Second, the dwell-time distribution at each option is informative about how much weight that option receives in the eventual choice. Wedel et al. (2023) frame this informativeness as the engine of psychometric inference: by writing a probabilistic generative model that links latent task strategy to observable fixation patterns, the analyst can recover decision parameters that behavior alone (response time and choice) cannot identify.

The Wedel-Pieters-van der Lans framework

The 2023 review’s organizing framework partitions decision-making into search (locating information in the visual field) and choice (integrating that information into a selection). Each task admits multiple strategies—exhaustive scan, satisficing, lexicographic comparison, gaze-driven evaluation—and the cognitive system can switch strategies within a single trial. The framework’s central claim is that strategy switching produces measurable structure in the eye-movement record that simpler models (which assume a fixed strategy) cannot capture.

This motivates the review’s emphasis on hidden Markov models (HMMs) as a natural psychometric tool for eye-movement data. In an HMM formulation, each fixation is treated as an emission from a latent strategy state, and transitions between states encode within-trial strategy switches. The model parameters—the emission probabilities for each state and the transition matrix—are estimable from fixation sequences and recover, post hoc, what strategies the participant used and when they switched. The same generative apparatus extends to point-process models of fixation timing and to econometric formulations in which dwell times enter a discrete-choice utility function.

Four model families the review consolidates

The Wedel et al. review situates its psychometric contribution against four older traditions whose models are the building blocks for any modern formulation.

Saliency-based bottom-up attention

The Itti and Koch (2000) saliency model formalized visual attention as the output of feature contrast computations across the visual field. A saliency map—built from local contrasts in luminance, color, orientation, and motion—predicts the next fixation as a winner-take-all selection biased by inhibition of return. The model is purely stimulus-driven and ignores task; it succeeds on free-viewing data and fails predictably whenever participants have an explicit goal. Its lasting value for decision-making research is as a null model: any deviation between observed fixations and saliency predictions is diagnostic of top-down task control.

Active vision and task-driven gaze

Hayhoe and Ballard (2005) crystallized the empirical case against pure saliency models by documenting that, in natural tasks (sandwich-making, ball-catching, driving), almost every fixation is task-relevant and almost no fixation is predicted by image saliency alone. Their account—active vision—treats gaze as a sequence of information-gathering actions selected to reduce uncertainty about task-relevant variables. The implication for decision modeling is that fixation locations are evidence about the participant’s current sub-goal, not just about the stimulus. Wedel et al.’s strategy-switching framework is the formal psychometric counterpart of the active-vision intuition.

Gaze cascade and preference feedback

Shimojo, Simion, Shimojo, and Scheier (2003) demonstrated that gaze and preference influence one another in real time: as a binary choice approaches resolution, fixations shift increasingly toward the option that will eventually be selected, and experimentally manipulating gaze duration biases the choice itself. The gaze cascade is not a separate model class so much as a robust empirical regularity that any decision model must accommodate. It also resolves an identifiability concern: dwell time and choice are not independent observables, and any psychometric formulation must specify whether gaze is causing, reflecting, or co-evolving with the value signal.

Attention drift-diffusion and gaze-weighted accumulation

The most influential synthesis came from Krajbich, Armel, and Rangel (2010), whose attention drift-diffusion model (aDDM) embeds gaze cascade into a sequential sampling architecture. In the aDDM, evidence accumulates faster for whichever option is currently fixated; the choice is determined by which decision boundary is crossed first; and the model parameters are identifiable from the joint distribution of choice, response time, and gaze pattern. The aDDM made quantitative predictions—for example, that longer total gaze on an option increases its choice probability in a way that decomposes into value and gaze components—that have been replicated across food choice, financial choice, and multi-attribute decision tasks.

Thomas, Molter, Krajbich, Heekeren, and Mohr (2019) extended this with the gaze-weighted linear accumulator model (GLAM), which generalizes to many-alternative choice and provides individual-level parameter estimates. Their key result is that the gaze bias parameter—how strongly fixation modulates evidence accumulation—varies systematically across people and predicts behavioral signatures (e.g., choice consistency, response-time distributions) that simpler value-only models miss. GLAM and related architectures are the contemporary state of the art for psychometric modeling of multi-alternative choice with gaze data.

What the 2023 review adds

Wedel et al.’s contribution is not a new model class but a unifying conceptual framework that absorbs each of the above into a hierarchy of tasks, strategies, and observable processes. The framework predicts when each model family will perform well: bottom-up saliency for free viewing of unfamiliar stimuli; aDDM/GLAM for two- to many-alternative value-based choice with stable strategies; HMMs and switching models when strategy itself is the explanatory variable of interest, as in complex consumer decision tasks where participants alternate between feature-based comparison and brand-based heuristics.

The review is explicit about open problems. Individual differences in gaze parameters are large, and most existing models treat them as nuisance variance to be marginalized out rather than as substantive measurement targets. Stimulus complexity is poorly handled: aDDM-class models work cleanly when options have a small number of structured attributes but degrade when stimuli are naturalistic images or text. Strategy identifiability remains hard because multiple strategies can produce statistically similar fixation distributions, requiring auxiliary identifying constraints (e.g., process-tracing self-reports, or multi-task designs that exploit cross-task parameter sharing).

Reliability and replicability of gaze-derived parameters

A measurement-model concern that the 2023 review surfaces, and that has accumulated independent evidence elsewhere, is the test-retest reliability of eye-movement summaries used as individual-difference measures. Aggregate gaze metrics—total dwell time per option, first-fixation location, transition counts—are typically more reliable than the model-derived parameters that depend on them, because the parameter estimates inherit uncertainty from both the gaze data and the choice data. For aDDM and GLAM applications, this means that group-level inferences (e.g., mean gaze bias across a sample) are on firmer footing than individual-level claims (e.g., a specific participant’s gaze bias as a personality-like trait). The Wedel et al. (2023) framework explicitly recommends multi-task designs and hierarchical Bayesian estimation to stabilize person-level parameters when individual differences are the substantive target.

Implications for decision-research methodology

The practical takeaway for researchers planning eye-tracking studies of decision-making is twofold. First, the choice of model class should be driven by the substantive question rather than by tradition. If the goal is to characterize how value signals are integrated, an aDDM/GLAM formulation is appropriate. If the goal is to detect strategy switches—e.g., the moment a shopper abandons feature-by-feature comparison for a brand heuristic—a hidden Markov or switching state-space model is closer to the structure of the question. Saliency models are diagnostic baselines, not explanatory accounts of goal-directed gaze.

Second, gaze-and-choice data are a richer measurement instrument than choice data alone, but only if the experimenter records them at sufficient temporal resolution and pairs them with stimuli whose attribute structure is known. Naturalistic stimuli buy ecological validity at the cost of identifiability; structured multi-attribute stimuli buy identifiability at the cost of generalization. The Wedel et al. (2023) framework makes that tradeoff explicit and gives the analyst a principled way to choose a position on it.

Frequently asked questions

Why use eye movements to study decision-making?

Where someone looks—and for how long—is structured by the same cognitive process that produces the choice itself. Modern eye-trackers record fixation locations at 60–2000 Hz, and the dwell-time distribution at each option is informative about the weight that option receives in the eventual choice. Linking fixation patterns to latent decision parameters lets researchers recover information that behavior alone (response time and choice) cannot identify.

What is the attention drift-diffusion model (aDDM)?

Krajbich, Armel, and Rangel (2010) introduced the aDDM, a sequential-sampling architecture in which evidence accumulates faster for whichever option is currently fixated. The choice is determined by which decision boundary is crossed first, and the model parameters are identifiable from the joint distribution of choice, response time, and gaze pattern. The aDDM has been replicated across food, financial, and multi-attribute decision tasks.

What is the gaze cascade?

Shimojo et al. (2003) demonstrated that gaze and preference influence each other in real time: as a binary choice approaches resolution, fixations shift increasingly toward the option that will eventually be selected, and experimentally manipulating gaze duration biases the choice itself. Any decision model must accommodate this regularity and specify whether gaze is causing, reflecting, or co-evolving with the value signal.

How do hidden Markov models help in eye-tracking decision research?

Hidden Markov models treat each fixation as an emission from a latent strategy state, with transitions encoding within-trial strategy switches. The Wedel et al. (2023) framework recommends HMMs when strategy itself is the explanatory variable of interest—for example, when a participant alternates between feature-based comparison and brand-based heuristics during a consumer decision.

Are eye-tracking-derived parameters reliable as individual-difference measures?

Aggregate gaze metrics (total dwell time, first-fixation location, transition counts) are typically more reliable than model-derived parameters that depend on them, because parameter estimates inherit uncertainty from both gaze and choice data. Group-level inferences are on firmer footing than individual-level claims. Multi-task designs and hierarchical Bayesian estimation help stabilize person-level parameters when individual differences are the substantive target.

When should I use saliency models instead of value-based models?

Bottom-up saliency models (Itti & Koch, 2000) succeed on free-viewing data without explicit goals and fail predictably whenever participants have a task. They are diagnostic baselines, not explanatory accounts of goal-directed gaze: any deviation between observed fixations and saliency predictions is informative about top-down task control. For decision research, aDDM-, GLAM-, or HMM-style models are appropriate.