Heterogeneity in acute PM₂.₅ mortality: big data strategies in the Mexico City Metropolitan Area

Abstract

Background and aim. Identification of subgroups at greater risk of mortality from PM₂.₅ may inform targeted public health interventions. The case-crossover design is a powerful tool for studying acute PM₂.₅ mortality associations in large registries where statistical interaction indicates differentially susceptible subgroups. We explore shrinkage with resampling to identify these groups in a big data context. Methods. We used a time-stratified case-crossover design to investigate the association between PM₂.₅ with 1.5 million ICD-10 coded non-external 18+ years-old death records from the Mexico City Metropolitan Area, 2004–2019. Daily municipal-level exposures for PM₂.₅ and temperature came from population-weighting 1km satellite-based model predictions, and were scaled to z-scores. We assessed whether the subgroup-specific association of PM₂.₅ exposure (lag₀₁) with daily mortality varied from the overall association for PM₂.₅ by including weighted effect coded interaction terms of lag₀₁ PM₂.₅ by cause-of-death:age-group:sex groupings (56 interaction terms), adjusting for temperature. We fit conditional logistic regression with shrinkage for regularization (ridge regression), and cross-validation to select the value of λ with the minimum mean square error. Finally, bootstrap resampling was used for coefficient confidence intervals (95%CI). Results. Overall, a 10μg/m³ higher lag₀₁ PM₂.₅ was associated with 1.46% higher mortality (95%CI: 1.30-1.77). Subgroups exhibited substantial variation (point estimates from 1.1% to 1.91%) with the strongest associations seen for: circulatory system in females ≥80 years-old (1.91%, 95%CI: 1.51%-2.92%); respiratory system in females ≥80 years-old (1.86%, 95%CI: 1.44%-3.30%); and respiratory system in females 65-79 years-old (1.83%, 95%CI: 1.56%-3.61%). Conclusion. Regularization techniques and inclusion of multi-way interactions in big data analyses aid in systematically identifying specific subpopulations with enhanced response to PM₂.₅ relative to the population-average association. Keywords PM₂.₅, penalized regression, case crossover, mortality, big data