A methodological framework for occupational cancer surveillance: Addressing multiplicity and unmeasured confounding in Italian registry data

• Large-scale registry linkages improve occupational cancer surveillance in Italy. • Framework controls multiplicity and selective inference in cancer risk profiling. • E-values assess robustness to unmeasured confounding in surveillance studies. • Results guide prevention policies and methods in occupational epidemiology. The Big Data and Deep Learning in the surveillance of occupational cancers (BEST) project aimed to enhance occupational cancer surveillance in Italy by using large-scale registry linkages. It addressed two methodological challenges: controlling for multiple testing/selective inference when profiling occupational categories and cancers, and assessing robustness of inference in presence of unmeasured confounding. Data were obtained through record linkage between the Italian cause-of-death registry (2005–2018) and the National Social Insurance Agency (INPS) database (1974–2018). Male blue-collar workers were classified by their longest held occupational sector, censoring the last 5 years before death. The study used a proportional cancer mortality design. Logistic regression models were fitted to estimate cause-specific mortality odds ratios (CMORs) for selected cancers and occupational sectors, adjusting for age, education, last region of residence, and year of death using the service sector as reference. Multiplicity was addressed through Q-Q plots with guide rails, q-values and control of false discovery rate, hierarchical Bayesian models and posterior rankings. E-values were calculated to assess the potential influence of unmeasured confounding. Elevated CMORs emerged for several cancers and industries, including lung cancer in construction and fishing, pleural cancer in shipyards, and sinonasal cancers in leather and woodworking trades. Findings remained consistent after multiple testing adjustments and sensitivity analyses using E-values. We propose an integrated methodological framework that combines multiplicity-aware profiling and E-values to address selective inference from multiple testing and sensitivity to unmeasured confounding. The framework improves the interpretation and prioritization of signals in occupational cancer surveillance, providing robust insights to guide prevention strategies.