Abstract
Understanding which variables are most strongly associated with operator error in processcontrol environments, and whether those associations are stable across different decision supportconfigurations, is a central challenge for human reliability research. This paper presents agroup-stratified analysis of error-variable associations in a simulated formaldehyde productionplant, using data from a human-in-the-loop experiment involving 92 participants distributedacross four support configurations: no support (G1), alarm rationalisation (G2), digital on-screenprocedures (G3), and an AI-based decision support system (G4). Each participant completedthree scenarios of increasing task complexity. A set of behavioural, physiological, and subjectivevariables, including operational metrics from simulator logs, physiological signals, NASA-TLXworkload subscales, SPAM situational awareness probes, and SART ratings , was analysed usingthree complementary measures: normalised conditional mutual information (NMI) estimatedfrom continuous data via the Kraskov 𝑘-nearest-neighbour estimator, sensitivity range (SR)derived from Bayesian Network inference conditioned on group and scenario, and normalisedlikelihood with Bayes factors (BF) for state-level evidential assessment. The results demonstratethat error-variable associations are substantially heterogeneous across support configurations.In the unsupported group, error is most strongly linked to objective operational variables,response time, reaction time, and alarm count, with almost no subjective variable reachingsignificance. As decision support increases, this profile shifts: in G2, situational awarenessdimensions (SPAM Monitoring, SART Complexity) become the dominant predictors; in G3,cognitive load and attentional distribution variables co-emerge with operational measures; in G4,the decisive error signals are mainly affective and arousal-based, frustration, arousal, instability,and skin temperature, with fewer operational variable reaching decisive evidential weight inscenario 3. These findings indicate that decision support technologies do not simply reduceerror rates but redirect the observable correlates of failure toward higher-order cognitive andaffective dimensions, with direct implications for the design of error monitoring systems and thecalibration of human reliability assessments across different operational configurations.

This publication has 15 references indexed in Scilit: