Derivation of European exposure values of internal exposure to environmental pollutants using human biomonitoring data

Abstract

Human biomonitoring, measuring chemicals or their metabolites directly in tissues and fluids, can, in principle, reveal EU-wide exposure patterns. The pooled HBM4EU data (2014–2021) were assembled from national and regional cohorts that each followed different, often not clearly documented, sampling schemes. The resulting dataset lacks a unified probabilistic design, and any uneven coverage against geographic and socio-demographic aspects, as well as the absence of sampling weights make the derivation of ”European” reference exposure values a challenge. This thesis focuses on the HBM4EU children’s age-group (6–12 yrs) and phthalates/mono-benzyl phthalate (MBzP) as a test-case. Exploratory analysis of the children dataset confirmed a North–East bias, an excess of high-education households and urban–rural mismatches; sampling year shadows cohort, magnifying site heterogeneity. Pronounced MBzP gradients by region, DEGURBA, sampling season and education affirmed the need for weights and cluster-robust inference, potentially providing a transferable template for other future initiatives. A population–standardisation grid for EU-27 children was built crossing one-way Eurostat margins for region (North, South, West, East), sex (male, female), season (each pre-weighted at 0.25), DEGURBA (urban, towns/suburbs, rural) and household-education (ISCED0–2, 3–4, ≥5). Age was fixed at 9 years, considering also the regional Eurostat data showing that uniform single-year counts across the 6–12-yr span. The Cartesian product yields 288 cells; each cell weight equals the product of its five marginal proportions and the set is normalised to 1. This construction assumes the five dimensions are mutually independent; in the absence of joint tabulations. These grid weights served a dual role: for model–based routes: each regression was fitted to the HBM4EU children data, after which its fitted values were projected onto the 288 cell profiles and post-stratified with the grid weights to represent an average EU child. Two specifications were considered: (i) an ordinary-least-squares model, with and without region–specific interaction blocks, evaluated with delta-method SEs; and (ii) a random-intercept mixed model, with and without interactions, propagating uncertainty via the analytic Delta-method, as well a “MCf ixed” Monte-Carlo SE (resampling) only the fixed-effect coefficients but also a “MC-full” MonteCarlo SE (resampling both the fixed effects and a new cohort-level intercept on every replicate). For design–based routes: the same weights were merged back to the HBM4EU data; dividing each cell weight by the number of sampled children in that cell, yielding observation–level probabilities that drove direct post-stratification, survey-design analysis and marginal raking, with weight trimming explored as sensitivity checks. The EU-27 standardised geometric mean estimates yielded narrower ranges, both with the model-based and design-based approaches, once cohort clustering was not considered. Declaring cohorts as PSUs inflated the confidence bands. Weight-trimming and marginal raking reduce design effects and sharpen intervals with negligible impact on the central estimate, whereas extending the calibration to a region×age margin lowered the geometric mean notably while substantially cutting the effective sample size. Next steps could focus on: (a) future initiatives collecting study-specific design weights before pooling to keep all downstream estimates design-consistent; (b) reporting both an efficient mixed-effects projection (MC-fixed SE) and a raked, cluster-robust survey estimate to bracket uncertainty; (c) test on the effect of replacing regional margins with finer, e.g. country-level or joint margins (if available); (d) include single-year age calibration and extend the framework to adolescents, adults and further biomarkers; and (e) explore weight- and cluster-aware design-based regression (svyglm) or GEE/GEE2, providing population-average estimates with sandwich-robust SEs.