Bootstrap Stability and Learner Sensitivity in Bayesian Network Modeling of near Midair Collision Incidents Using Aviation Safety Narratives and Operational Data

Near midair collisions (NMACs) remain a complex aviation safety challenge shaped by operational conditions and human factors. Aviation Safety Reporting System (ASRS) narratives provide useful information about communication, visual acquisition, workload, and other event circumstances, but relationships identified from voluntary reports may vary across analytical approaches and over time. This study examined 13,111 NMAC-related ASRS reports from 1988 to 2025 to identify stable report-level associations among operational conditions, encounter proximity, narrative topics, and reported evasive action. Across two Bayesian-network approaches and 500 paired bootstrap samples, traffic-pattern operations/training and tower communication/sequencing were the strongest variables directly associated with reported evasive action. Their recurrence was moderate rather than universal, however, and their inclusion changed under different relationship-selection thresholds. This indicates that the local structure around reported evasive action was sensitive to modeling decisions, even though the two variables consistently ranked as the leading response-adjacent candidates. Additional robustness checks preserved these two response parents in the full-data models while showing changes in the broader network structure. A temporal evaluation produced a different pattern. Models trained on reports from 1988 to 2019 and evaluated on reports from 2020 to 2025 instead identified airspace structure/radar-contact issues and final-approach/base-to-final conflicts as the response-adjacent variables. Because the narrative-topic features were derived from the full corpus, this analysis represents a temporal evaluation conditional on a fixed topic representation, rather than a fully independent validation of the complete modeling pipeline. The changed response parents therefore indicate temporal sensitivity but do not establish nonstationarity. The findings describe associations within voluntary safety reports and should not be interpreted as evidence of causality, collision risk, or operational warning performance. The study contributes a reproducible, systematically stress-tested framework for evaluating Bayesian-network findings in aviation safety narratives. It identifies communication, sequencing, visual acquisition, and traffic-pattern operations as priorities for expert review and prospective measurement while cautioning against treating relationships found in historical narratives as stable operational predictors.