Background: One of the largest water contamination episodes by PFAS (poly/perfluoroalkyl substances) occurred in a vast area of Veneto (Italy), first detected in 2013, though likely began decades earlier. A voluntary biomonitoring program (2017) measured 12 PFAS, including PFOA, among residents in 30 municipalities (~135,000 inhabitants) supplied by the contaminated aquifer. Human exposure has been associated with multiple health problems, including testicular and kidney cancers. Available data are aggregated geographically. Disease rates show underlying spatial structure, commonly attributed to unobserved confounders. In causal inference terms, spatial autocorrelation violates Rubin’s SUTVA (Stable Unit Treatment Value Assumption) principle, leading to spatial interference, motivating spatial models with random effects of responses and exposures.
?Objectives: This study evaluates, at the aggregate level, the relationship between PFOA exposure and testicular cancer incidence in 21 municipalities within the PFAS-contaminated area of Veneto, applying causal inference methods for ecological regression.
Methods: PFOA exposure was measured via two variables: geometric mean blood concentration from 2017 biomonitoring data; water concentrations (from 2013 to 2017) interpolated via spatial kriging. Lacking a randomized experiment, both exposures could depend on measured/unmeasured variables and exhibit hidden spatial structure. Testicular cancer incidence was derived from hospital discharge records for orchiectomies performed between 1997 and 2014 on males resident in the study area. The spatial pattern of testicular cancer incidence could also depend on measured/unmeasured confounders. Municipal deprivation index (DI) components, extracted from the 2011 census, were included. Two initial ecological regressions - frequentist and Bayesian with spatial random effects on the response - estimated exposure effects without confounders. Subsequently, causal inference models adjusted for unobserved spatial confounders and DI were implemented: a bivariate shared model (two spatial random effects: on response and exposure); a two-stage model incorporating the spatial random effect of the exposure model into the disease model; a model incorporating the propensity scores from the exposure model into the disease model. All models were fitted using separately the two exposure variables (assumed to depend on different selection mechanisms), both including and excluding the DI.
Results: Seventy orchiectomies were recorded between 1997 and 2014 in the 21 municipalities of the red area of Veneto. The highest excess risk was observed in Lonigo: 16 observed cases against 8.71 expected ones (SIR 1.84, 95% CI [1.05, 2.98]). The hidden spatial structures of the two exposures differed, suggesting the hypothesis of different selection mechanisms. The response spatial structure suggested a north-south gradient for unmeasured confounding. Estimated coefficients associated with the two exposures from the various models varied in precision, but were all positive (on the log-scale). After rescaling by the corresponding SD to be comparable, they ranged between 0.187 (90% CrI [0.009, 0.385]) and 0.382 (90% CrI [0.141, 0.622]) for the blood concentration, and between 0.146 (90% CrI [0.011, 0.282]) and 0.226 (90% CrI [0.043, 0.409]) for the water concentration, representing the change associated with a 1 SD increase in exposure.
Conclusions: Despite differences in precision, results were consistent, indicating at the ecological level a dose–response relationship between PFOA exposure and testicular cancer incidence.