... continued from previous pg.
More information is available at: http://www.bgs.ac.uk/radon/hpa-bgs.html. To assess measurement error, we also linked the ARIES dataset to estimates from the freely available radon ‘indicative atlas’31, which
is based on the ‘potential dataset’, but provides the estimates in 1-km-size squares.
Residential histories of mothers and children were geocoded to postcode centroid level, and were linked to average potential radon exposure using ArcGIS software (version 10.6)32 within the ALSPAC Data Safe Haven. This resulted in at least one address match for 986 mothers and 1001 young people (including two sets of twins). Once each residential address had a radon potential exposure class assigned, time spent at each address was calculated. This was merged with ARIES sample prevision dates, allowing time- weighted average potential radon exposures to be calculated up to the ‘mothers at middle age’, ‘children at 7’ and ‘children at 15/17’ sample extraction time points. For the cord and antenatal sample extractions, radon exposure potential of address at date of birth or closest address (temporally) to sample time point were assigned respectively.
These data were then linked to ALSPAC self-reported data selected to test for potential confounding (described below). After the linked exposure data were processed to minimise the risk of participant disclosure, the linked methylation-radon data were used for statistical analyses.
Statistical methods
For these analyses we only use participants with complete data. In the primary analyses average potential radon exposure was analysed as a continuous variable (range 1–6) to assess linear exposure-response associations. In addition, we also analysed associations based on binary exposure classifications (≤5% vs >5%).
Associations were tested using linear models using the limma R package (version 3.32.10)33. Associations were tested in (1)*** univariate analyses but with adjust- ment for the surrogate variables34 to handle batch effects, sex differences, cell count heterogeneity and possible unknown confounders35, and (2) additionally with adjustment for potential confounding factors maternal age at birth, maternal
BMI, smoking during pregnancy, partner smoking during pregnancy, AHRR CpG site that detects own smoking nearly as accurately as self-report36, mother alcohol intake in early pregnancy, equivalized income, parental occupation, and parental education, and (3) all factors of models 1 and 2 and additionally for damp problems, central heating, boiler location, gas cooking, time windows open in the summer/winter day/night, and heavy traffic.
Because of the exploratory nature of this study, associations at false discovery rate (FDR) less than 20% calculated using the method37 are reported instead of a more traditional 10% or 5% threshold. Where associations were positive, these sites
were defined as “hypermethylated” and conversely when for inverse associations, these were defined as “hypomethylated”.
Results
Participants and location
The results were based on 786 to 980 participants with complete information, depending on the time analysed. A graphical overview of the geographical study area and the distribution of potential radon exposure classes, as well as the distribution of addresses in each class, is shown graphically in Figure 1, and indicates that 79% of addresses are in areas with low probability (class 1 and 2) of exposure
>200 Bq m-3.
###
61