Bayesian Methodology

Divide-and-conquer methods provide scalable Bayesian inference by partitioning large datasets, running independent Markov chain Monte Carlo (MCMC) algorithms on each subset, and recombining subposterior draws to approximate the full posterior distribution. A widely used strategy, Consensus Monte Carlo recombination (CMC), aggregates subposterior samples under Gaussian assumptions. This recombination strategy has proven effective for multivariate normal posteriors, but its performance under censored likelihoods remains poorly understood. We investigate the performance of CMC for Bayesian models with left-, right-, and interval-censored data. Censoring arises in many cognitive and behavioral settings, including time-to-event data as well as bounded measurement instruments such as Likert scales, which impose natural lower and upper limits on the observable responses. Some Bayesian software (e.g., JAGS) can express censoring directly at the likelihood level, but it is unclear how well subposterior recombination preserves the resulting posterior structure, particularly when censoring induces asymmetry or boundary effects. Because censoring appears in many real-world cognitive measurements, understanding this interaction is critical for valid large-scale inference. In this work we conduct simulation studies that vary censoring type and proportion, partition size, and number of subsets, and evaluate posterior recovery and divergence from full-data MCMC. We then illustrate performance on an empirical rating-scale dataset. Our results identify scenarios when CMC yields reliable inference under censoring and identify conditions under which recombination bias or variance distortion emerges, informing the use of scalable Bayesian methods in cognitive modeling and psychometrics.

This is an in-person presentation on July 18, 2026 (10:40 ~ 11:00 EDT).

Security Operations Centre (SOC) operators sustain concurrent demands absent from standard lab paradigms: vigilance monitoring, working memory updating, conflict resolution, and decision-making under time pressure and escalating threat. Understanding how these stressors interact to produce performance collapse has implications for both cognitive capacity theory and applied cybersecurity design. We present agent-based simulations of a three-task SOC environment (PVT-like alarm dismissal, SART-like log triage, WCST-like phishing email sorting) implemented on the CogFlow platform’s multi-window SOC Dashboard with overlapping subtask schedules and varying urgency. Synthetic agents draw performance profiles from empirical RT/accuracy distributions, then face manipulated time pressure (session duration, response windows) and hazard escalation (threat severity, miss costs) in a 2×2 factorial design. Key findings from Bayesian hierarchical modelling: (1) Time pressure drives speed-accuracy trade-offs–16% faster RT but 46% more misses; (2) Critical hazard massively increases vigilance failures (102% more misses) with minimal RT effect; (3) Combined stressors show sub-additive interaction (OR=0.83), suggesting strategic adaptation rather than linear degradation. Operators under dual stress shift to conservative strategies, not collapse. Sub-additive interaction effects in realistic multi-tasking suggest adaptive mechanisms that warrant investigation through empirical studies and formal cognitive modelling (e.g., capacity architectures). Surrogate findings inform initial interface design priorities (hazard visibility over speed pressure) pending human validation.

This is an in-person presentation on July 18, 2026 (11:00 ~ 11:20 EDT).

Empirical findings in psychology typically rely on a single population, design, and analytic pipeline, which leaves many reasonable alternatives unexplored. Many-analysts projects, multiverse studies, and robustness analyses make this uncertainty visible by demonstrating how results can vary across defensible analytic choices applied to the same dataset. Beyond revealing heterogeneity, these approaches offer a powerful conceptual tool for acknowledging and quantifying epistemic uncertainty in empirical research. Yet practical guidance on how to analyze and interpret the outcomes of such projects remains limited. In this talk, we present methodological developments to move many-analysts and multiverse projects beyond merely describing analytic variability. Our approach is grounded in a Bayesian (cognitive) modeling framework. Within this framework, we (1) provide structured opportunities for analysts to report their beliefs about an effect’s plausibility, their confidence, their perceived robustness, and their critical reflections on the dataset and design, and (2) implement a single-dataset meta-analytic model that yields point and interval estimates of the average effect across analytic approaches and between-analyst heterogeneity. These estimates can be complemented by Bayesian hypothesis tests that assess evidence for an overall effect and for the presence of heterogeneity. Together, these models complement the qualitative evaluation of many-analysts and multiverse studies and help prevent overconfident inferences. We also highlight emerging initiatives that aim to make routine reanalysis through alternative, plausible pipelines a standard component of psychological science. Together, these methodologies point to an emerging and still underrecognized direction in methodological reform: embracing analytic uncertainty as a core component of rigorous psychological science.

This is an in-person presentation on July 18, 2026 (11:20 ~ 11:40 EDT).

Attentional control is a central construct in cognitive individual-differences research, yet its conceptual unity and psychometric coherence remain debated (Rey-Mermet et al., 2025). Conventional operationalizations—typically difference scores of behavioral data—have been criticized for weak cross-task convergence, casting doubt on whether different conflict tasks reflect a domain-general capacity. To address these limitations, we used variants of a sequential sampling model with time-varying drift rate (Diffusion Model for Conflict Tasks; Ulrich et al., 2015), aiming to isolate process-pure measures of attentional control on an individual level. Participants completed confound-minimized four-choice versions of arrow, letter, and number flanker paradigms, along with a matrix reasoning test. Following a principled modeling workflow, individual parameters were estimated using Amortized Bayesian Inference, and structural equation modeling was applied to evaluate the latent structure across task variants and associations with reasoning ability. Parameters representing attentional control processes demonstrated robust convergence across flanker tasks and loaded onto stable latent factors. Their correlations with reasoning ability suggest that individuals with higher reasoning ability exhibit stronger initial automatic activation from irrelevant information, coupled with more rapid decay of this activation and more efficient accumulation of task-relevant evidence compared to individuals with lower reasoning ability. Taken together, these findings indicate that computational modeling yields psychometrically coherent and theoretically informative indicators of attentional control. Moreover, computationally derived parameter estimates capture meaningful variance in reasoning ability, underscoring the potential of this process-level approach for refining the conceptualization and measurement of attentional control in individual-differences research.

This is an in-person presentation on July 18, 2026 (11:40 ~ 12:00 EDT).