Predicting Monthly Community-Level Radon Concentrations with Spatial Random Forest in the Northeastern and Midwestern United States

In 1987, the United States Environmental Protection Agency recommended installing a mitigation system when the indoor concentration of radon, a well-known carcinogenic radioactive gas, is at or above 148 Bq/m3. In response, tens of millions of short-term radon measurements have been conducted in residential buildings over the past three decades either for disclosure or to initially evaluate the need for mitigation. These measurements, however, are currently underutilized to assess population radon exposure in epidemiological studies. Based on two relatively small radon surveys, Lawrence Berkeley National Laboratory developed a state-of-the-art national radon model. However, this model only provides coarse and invariant radon estimations, which limits the ability of epidemiological studies to accurately investigate the health effects of radon, particularly the effects of acute exposure. This study involved obtaining over 2.8 million historical short-term radon measurements from independent laboratories. With the use of these measurements, an innovative spatial random forest (SRF) model was developed based on geological, architectural, socioeconomical, and meteorological predictors. The model was used to estimate monthly community-level radon concentrations for ZIP Code Tabulation Areas (ZCTAs) in the northeastern and midwestern regions of the United States from 2001 to 2020. Via cross-validation, we found that our ZCTA-level predictions were highly correlated with observations. The prediction errors declined quickly as the number of radon measurements in a ZCTA increased. When ≥15 measurements existed, the mean absolute error was 24.6 Bq/m3, or 26.5% of the observed concentrations (R2 = 0.70). Our study demonstrates the potential of the large amount of historical short-term radon measurements that have been obtained to accurately estimate longitudinal ZCTA-level radon exposures at unprecedented levels of resolutions and accuracy.