Bayesian Hurdle Beta Workflow Separates Gaze Visitation and Processing Density

A new statistical methodology developed for eye-tracking research separates the discrete decision to look at an area of interest from the continuous duration of visual processing. In a methodological study published in Behavior Research Methods, Kuangzhe Xu introduces the Bayesian hierarchical hurdle beta workflow to handle the bounded and sparse topologies inherent to visual attention metrics. The framework models structural zeros, which occur when a participant completely skips a designated visual region during an experimental trial, alongside continuous fixation densities within that same region.

Visual attention measures frequently exhibit a combination of complete non-visitation and bounded continuous gaze durations that standard statistical tools fail to characterize appropriately. Conventional linear mixed models mischaracterize this underlying generative data process by assuming continuous distributions that do not accommodate discrete zeros. Although general-purpose bounded regressions address some distributional constraints of proportion data, they frequently fail to distinguish the discrete decision to orient toward a target from the subsequent continuous duration of cognitive processing.

The newly introduced modeling framework grounds its statistical architecture in dual-process theories of gaze control. Within this formulation, the model explicitly distinguishes the probability of visiting an area of interest as a decision component from the conditional processing density as an intensity component. To facilitate reproducible implementation and customization across experimental settings, the author provides an open-source toolkit programmed in R and Stan for researchers analyzing sparse visual data.

Methodological comparisons and non-parametric bootstrapping evaluations demonstrated key differences between estimation techniques when handling sparse gaze metrics. While maximum likelihood estimation of hurdle models suffers from numerical instability under conditions of low data sparsity, the Bayesian hierarchical approach achieves improved computational stability and superior predictive performance. These empirical improvements are realized through prior regularization, which constrains parameter estimation and avoids optimization failures in challenging distributional topologies.

The utility of the statistical workflow was further demonstrated through an empirical application analyzing multimodal gaze and mouse dynamics. This analysis quantified distinct lock-and-process strategies across participants, capturing behavioral patterns that standard aggregate models systematically obscure. By evaluating gaze alongside concurrent mouse movements, the model isolated distinct spatial orienting behaviors that traditional statistical averages conceal within combined metric summaries.

External validation conducted via participant confidence ratings further supported the model’s structural separation of visual components. The observed spatial decoupling between orientation decisions and continuous processing durations was modulated in a pattern consistent with an underlying cognitive strategy. This alignment suggests that non-fixation reflects a distinct cognitive state rather than mere measurement noise, demonstrating the value of modeling structural zeros directly in visual cognition research.

Subject of Research: Statistical modeling of sparse eye-tracking metrics using Bayesian hurdle beta regression

Article Title: A Bayesian hierarchical hurdle beta workflow for sparse AOI-based eye-tracking data

Article References: Xu, K. (2026). A Bayesian hierarchical hurdle beta workflow for sparse AOI-based eye-tracking data. Behavior Research Methods, 58(12), Article 351. https://doi.org/10.3758/s13428-026-03197-5

Image Credits: AI Generated

DOI: 10.3758/s13428-026-03197-5

Keywords: eye-tracking, Bayesian hierarchical modeling, hurdle beta model, area of interest, visual attention, structural zeros, Stan, R programming, gaze control, multimodal tracking, mouse dynamics, computational stability, prior regularization, Behavior Research Methods, cognitive strategy, bounded regression