Predicting residential radon concentrations in Finland : Model development, validation, and application to childhood leukemia

Downloaded from www.sjweh.fi on December 02, 2019 This work is licensed under a Creative Commons Attribution 4.0 International License. Print ISSN: 0355-3140 Electronic ISSN: 1795-990X Copyright (c) Scandinavian Journal of Work, Environment & Health Original article Scand J Work Environ Health Online-first -article doi:10.5271/sjweh.3867 Predicting residential radon concentrations in Finland: Model development, validation, and application to childhood leukemia by Nikkilä A, Arvela H, Mehtonen J, Raitanen J, Heinäniemi M, Lohi O, Auvinen A Indoor radon prediction models were created based on ~80 000 measurements and modern machine learning methods were used in modelling. The performance of the models was comparable to the previously published ones. We observed a non-significant risk of childhood leukemia from indoor radon. However, the modelling involves some uncertainties. Affiliation: Faculty of Medicine and Health Technology, Tampere University Arvo Ylpön katu 34, 33520 Tampere, Finland. [email protected] Key terms: cancer; carcinogen; childhood leukemia; epidemiology; etiology; Finland; Finland; indoor radon; leukemia; radon; radon concentration This article in PubMed: www.ncbi.nlm.nih.gov/pubmed/31763683 Additional material Please note that there is additional material available belonging to this article on the Scandinavian Journal of Work, Environment & Health -website. Scand J Work Environ Health – online first 1 Original article Scand J Work Environ Health – online first. doi:10.5271/sjweh.3867 Predicting residential radon concentrations in Finland: Model development, validation, and application to childhood leukemia by Atte Nikkilä, MD, PhD,1 Hannu Arvela, PhD,2 Juha Mehtonen, MS,3 Jani Raitanen, MS,4, 5 Merja Heinäniemi, PhD,3 Olli Lohi, MD, PhD,6 Anssi Auvinen, MD, PhD 2, 4, 6 Nikkilä A, Arvela H, Mehtonen J, Raitanen J, Heinäniemi M, Auvinen A. Predicting residential radon concentrations in Finland: Model development, validation, and application to childhood leukemia. Scand J Work Environ Health – online first. doi:10.5271/sjweh.3867 Objectives Inhaled radon gas is a known alpha-emitting carcinogen linked especially to lung cancer. Studies on higher concentrations of indoor radon and childhood leukemia have conflicting but largely negative results. In this study, we aimed to create a sophisticated statistical model to predict indoor radon concentrations and apply it to a Finnish childhood leukemia case–control dataset. Methods Prediction was based on ~80 000 indoor radon measurements, which were linked to national registries for potential indoor radon predictors based on the literature. In modelling, we used classical methods, random forests and deep neural networks. We had 1093 cases and 3279 controls from a nationwide case–control study. We estimated odds ratio (OR) for childhood leukemia using conditional logistic regression adjusted for potential confounders. Results The r2 of the final log-linear model was 0.21 for houses and 0.20 for apartments. Using random forest method, we were able to obtain slightly better fit for both houses (r2 = 0.28) and apartments (r2 = 0.23). In a risk analysis based on the case–control data with log-linear model, we observed a non-significant (P=0.54) increase with predicted radon concentrations [OR for the 2nd quartile 1.08, 95% confidence interval (CI) 0.77–1.50, OR 1.10 with 95% CI 0.79–1.53 for the 3rd, and 1.29 with 95% CI 0.93–1.77 for the highest quartile]. Conclusions Our modelling and the previously published models performed similarly but involves major uncertainties, and the results should be interpreted with caution. We observed a slight non-significant increase in risk of childhood leukemia related to higher average indoor radon concentrations. Key terms cancer; carcinogen; epidemiology; etiology; indoor radon. 1 Faculty of Medicine and Health Technology, Tampere University, Tampere, Finland. 2 STUK - Radiation and Nuclear Safety Authority, Helsinki, Finland. 3 Institute of Biomedicine, University of Eastern Finland, Kuopio, Finland. 4 Faculty of Social Sciences, Unit of Health Sciences, Tampere University, Finland. 5 UKK Institute for Health Promotion Research, Tampere, Finland. 6 Tampere Center for Child Health Research, Tampere, Finland. Correspondence to: Atte Nikkilä, Faculty of Medicine and Health Technology, Tampere University Arvo Ylpön katu 34, 33520 Tampere, Finland. [E-mail: [email protected]] Radon (Rn-222) is an alpha-radioactive element in the decay chain of uranium. It is generated in the ground from the decay of radium and, as a gas, it occurs in high concentration in soil pore air. A number of physical factors and processes are involved in the generation and transfer of radon from mineral grains to soil gas and in the movements of radon-bearing soil air. The entry of soil gas into built spaces is controlled by the flow dynamics of soil air in the porous soil media and through a large variety of gaps, air-permeable building elements and openings in the structures in contact with soil or in floor structures in crawl space houses. Indoor radon concentrations vary widely depending on the uranium concentration of the terrain, soil permeability, entry from the ground to buildings and ventilation ( 1–3 ). Finland has one of the highest average residential radon-222 concentrations in the world, 96 Bq/m3, resulting in a mean annual effective dose of 1.6 mSv based on ICRP-65 from 1993 and a dose of 4.5 mSv based on the ICRP-137 (4–6). The new ICRP-137 from 2017 is based on both dosimetric estimates and epidemiology. The radiation dose is largely due to the short-lived progeny rather than radon itself. Dose to the bone marrow from inhaled radon progeny is substantially lower than that to the lung ( 7, 8 ). 2 Scand J Work Environ Health – online first Indoor radon and childhood leukemia Health effects of indoor radon Uranium and hard-rock miners are exposed to very high concentrations of radon progeny and such occupational exposure has been shown to increase the risk of lung cancer ( 9 ). Lower residential radon concentrations have also been shown to increase the risk of lung cancer ( 10 ). The International Agency for Research on Cancer (IARC, World Health Organization) has classified radon as a recognized Group 1 human carcinogen ( 11 ). No excess of leukemia has been consistently associated with radon exposure in uranium miners ( 12–22 ). Results from previous studies on the possible effect of exposure of indoor radon on risk of childhood leukemia have been largely negative but still inconclusive. The potential dose pathway for the association, in addition to the exposure of the red bone marrow, has been suggested to be through the exposure of lymphocytes within the tracheobronchial epithelium ( 23 ). Studies in Norway, France, the UK, and Switzerland showed no association, but a Danish case–control study with complete residential histories and a statistical model with 40% r2 reported an elevated risk ( 24–28 ). In these studies, exposure estimates were derived from model-based predictions of radon exposure ( 29–31 ). Efforts to construct a good prediction models have been made in the UK and detailed information on soil has been essential ( 32, 33 ). Some smaller studies have used actual radon measurements but shown no materially elevated risks ( 34–37 ). In addition, several ecological studies have evaluated the association between incidence rates and regional average radon levels and have consistently reported positive risk estimates ( 38 ). Estimation of indoor radon When estimating the effects of indoor radon, or most other environmental exposures, a direct measurement would be the optimal way to define exposure. However, that is not always possible due to practical reasons. To study risk factors of small expected effect size with sufficient statistical power, a large number of subjects is needed, and performing thousands or even millions of repeated on-site radon measurements does not currently appear feasible. Further, participation bias in measurement program is likely to be a significant problem. However, robust results have been reported using statistical models for predicting radon concentrations in similar scenarios ( 29–31 ). Many country-specific models have been published with varying performance ( 29, 39 ). Low-rise residential buildings (single family houses, semi-detached houses and terraced houses) will be referred to as houses. Dwellings in multi-story block houses are called apartments. Indoor radon concentrations are determined by a variety of factors in a complex chain of processes. In low-rise residential houses, the soil-borne radon gas dominates, with regard to indoor radon concentration. The main processes include concentration of uranium in mineral grain, emanation of radon from mineral grains to soil gas, movements of radon-bearing soil air in the porous soil media, and entry of radon-bearing soil gas into living spaces. In foundation structures, gaps, airpermeable building blocks, and openings in the structures increase the soil air entry into indoor spaces. The entry rate is controlled by the flow dynamics of soil air in the porous soil media and physical modelling shows that the air permeability of the sub-soil is a much more important factor than the effective area of the air leakage routes ( 40 ). Therefore, the highest values are measured in houses situated in hilly areas with porous soil of coarse gravel, for example on eskers (a long ridge of gravel or other sediment, typically having a winding course, deposited by meltwater from a retreating glacier or ice sheet). The lowest values in low-rise buildings are found in areas of impermeable clay. Air exchange in the building is the process of diluting radon concentration in indoor air ( 41 ). In Finland, in apartment buildings soil-borne radon is not an important radon source, except for apartments on the lowest level and with floors in contact with soil. On upper floors, radon gas emanated from rock-based building materials, normally concrete elements, dominates. The national average indoor radon concentration caused by building materials in apartments is clearly lower (49 Bq/m3) than the average concentrations caused by soil-borne radon in low-rise residential houses (121 Bq/m3) ( 6 ). Also, the range of radon concentrations in apartments is narrower compared with houses as the percentage of measurements above 400 Bq/m3 in apartments was 0.7% in the national survey and 3.8% in houses. Uranium concentration of local gravel material can be utilized as a determinant for radon concentration in apartments because gravel has been used as concrete ballast material. Other important predictors of indoor radon are the dwellings age and the existence of cellars in detached houses ( 42, 43 ). Also, the story of the dwelling in blocks of flats has been shown to predict the concentration ( 44, 45 ). Seasonal variation has also been documented ( 43 ). Radon concentrations are highest during the heating season. Indoor radon measurements have been carried out in the period of November–April in Finland ( 46 ). Modelling indoor radon concentrations has been proven to be particularly difficult as limited or no data are available on several important determinants. For example, the type of building foundation correlates strongly with radon concentrations but is rarely available ( 3, 43 ). The same stands true also for the type or source of gravel used for the foundation. However, due to the importance of the data from the original building soil, Scand J Work Environ Health – online first 3 Nikkilä et al the effect of the lack in the knowledge of the transported layers of mineral material is decreased. Ventilation strategies, either natural or mechanical, are not included in the database of the Population Register Centre of Finland. However, history of the prevalence of ventilation strategies in Finnish low-rise residential buildings is well known based on national sample surveys ( 6, 41, 47 ). With regard to modelling, the effect of ventilation strategies seems to be limited compared with uranium concentration in soil or soil permeability ( 41 ). The building code for radon prevention and the associated practical guidelines were revised in Finland in 2003–2004. Thereafter, preventive measures have become more common and effective and, in houses completed since 2006, indoor radon concentrations have been markedly reduced. These data are of great importance when constructing a statistical model. The national radon prevention study in 2009 showed that in houses with preventive measures, the radon concentration was on average reduced by ≥50% compared with houses with no preventive measures ( 47 ). Furthermore, results from more than 200 000 individual radon measurements in Finnish dwellings are recorded in the database of the Radiation and Nuclear Safety Authority. In Finland, only regional indoor radon modelling has been conducted and no nationwide studies on modelling radon concentrations have been published. Aims of the study Using statistics, we modelled indoor radon concentration in a given dwelling using measurements from the nationwide database and internally validated its performance and robustness. Then we applied the model to examine potential association between residential radon and childhood leukemia using data from a nationwide childhood leukemia case–control study ( 48 ). Methods Radon measurements We obtained results of all indoor radon measurements (N=244 059) from the database compiled by STUK – the Radiation and Nuclear Safety Authority and linked them to the building database of the Population Register Center by address and postal code. We used the oldest available measurement from each dwelling to minimize the effect of potential radon protection renovations. If there were two or more measurements with the same start dates, the one with higher measured concentration was used to maximize the models’ ability to recognize the high concentrations. Combining databases The Population Register Centers building database contains data on a dwelling type (house versus apartment), year of completion, floor area (m2), total area (m2), total volume (m3), number of floors, area of the basement, main building material (rock-based materials, wood, others) and air-conditioning. All predictive variables were required to be available from nationwide registries (Population Registry Center) and, thus, not all important predictors, that were available only in STUK’s radon database (type of foundation, radon protection, the floor of the dwelling), could be utilized in modelling. The linkage of the measurements to the building database of the Population Register Center was based on street address and postal code as the key. This resulted in one-to-many linking problem due to multiple buildings in the same postal address. In such cases, we selected buildings with the best match in terms of building type, year of completion and coordinates. To deal with the remaining discrepancies between databases (STUK's and the Population Registry Center’s Building databases) after the primary selection, we created three sets of filtering criteria to acquire the best compromise between accuracy and sample size. We also aimed to explore whether there would be substantial differences in models with differently filtered radon datasets. The sample sizes of different filtering levels are represented in the figure 1. The first level required >100 m difference in Euclidean distance by coordinates, a >10-year difference in year of completion and no observable discrepancy in building type between the two databases. The second level required that there be no missing values in any of the filtering variables of the first level and thus all filters could be applied to every building (as the first level inhibited missing values from triggering the filter). The third level also allowed for no missing values and involved stricter criteria for >10 m Euclidean distance and identical year of completion. The numbers of buildings fulfilling the three sets of criteria are shown in the figure 1 (with other exclusions). Radon concentrations in houses and apartments were modelled separately as the major predictors differed based on the literature. Dwellings with missing or ambiguous building type were excluded. For the house model, the median postal-code-specific indoor radon concentration was derived from the 20% of the measurements sampled from the dataset left outside modelling to avoid using derivatives of the measurements as predictors. As the average number of dwellings per postal code area was relatively low and the total number of postal areas was relatively high, this resulted in some missing values (N=5697, 3.6%) and, also, some postal areas were represented by only few measurements. For the apartment model, we constructed a database of 4 Scand J Work Environ Health – online first Indoor radon and childhood leukemia county-specific median radon concentrations in apartments based on two nationwide representative surveys (conducted in 1991 and 2006) ( 6, 49 ). The measurements from the 1991 survey were calibrated to match the values from the more recent survey. Additional data for the models To complement the model, we obtained data on the soil type as vector maps and terrain elevation as a 100 × 100 m square map from Geological Survey of Finland (GTK). Regarding the soil type, for each area, the map with the highest resolution (1:20 000, 1:50 000 and 1:100 000) available was used. STUK also provided us with an 8 × 8 km square map of soil uranium concentration (Bq/kg) ( 50 ). The vector maps for dwellings were evaluated using QGIS (v. 3.2.1) and square maps were evaluated with a basic R script. Detailed soil types were classified into three categories by permeability. The classification was based on the grain size distribution of the soil type. Air permeability of soil types is closely related to grain size distribution. Soil air permeability is highest for coarse gravel (grain size 6–20 mm) and lowest for clay with a very low grain size (>0.002 mm). The database presents the soil type at the depth of 1 meter, which is representative of the depth of house foundations. Several terms were created to characterize year of construction: a categorical variable in 5-year intervals, as well as a separate indicator term for pre-1940 were used. For apartments, the latter term was defined with 1950 as the cut-off. The building material was classified as rock-based, wood, or other/ unknown. We also created a binary variable to estimate exhaust fan-based ventilation: any type of ventilation based on the building registry and building completed before year 2000 for houses and any type of ventilation with building completed between 1950 and 2006 for apartments. The presence of a basement was modelled as a three-step variable (no basement, basement and dwelling built before 1990, basement and dwelling built after 1990) due to new prevalent practice of hill-side houses instead of full basement houses. Modelling indoor radon We applied multiple approaches for developing the two final radon prediction models. The methods were used similarly for both models from predictor selection to validation. First, we started with a log-linear model with all the available predictors. All continuous potential predictors were log-transformed. We used a backward selection algorithm starting with the full model and used multiple imputation to deal with the missing data. The proportions of missing data for each potential predictor are presented in the supplementary material ( www. sjweh.fi/show_abstract.php?abstract_id=3867 ), table S1. We defined measured indoor radon outliers as values with z>3 and excluded them. We then created two categorical models with radon concentrations divided into quartiles: a polynomial and a multinomial. We also tested a model with a binary dependent variable by dividing the radon concentration by its 80th percentile. Finally, we experimented with modern machine learning algorithms (random forest and deep neural networks) as an alternative to the traditional methods ( 51, 52 ). For random forest models, we set the number of trees grown to 2000 and 560 for apartments and houses, respectively, based on the point, where the model errors started to converge. Deep neural network was specified as a 4-layer network with 256, 128, 64, and 1 nodes with rectified linear unit as activation funcTo optimize both accuracy and sample size, we decided on selecting the first level of filtering as the basis for our main analyses. a - Less than 100 m difference in Euclidean distance by coordinates, less than a 10-year difference in year of completion and no observable discrepancy in building type. b - Less than 100 m difference in Euclidean distance by coordinates, less than a 10-year difference in year of completion, no observable discrepancy in building type and no missing values in any of the filtering variables. c - Less than 10 m difference in Euclidean distance by coordinates, no difference in year of completion, no observable discrepancy in building type and no missing values in any of the filtering variables. Figure 1. Flow chart of the indoor radon measurements and the necessary exclusions. To optimize both accuracy and sample size, we selected the first level of filtering as the basis for our main analyses. a <100m difference in Euclidean distance by coordinates, <10-year difference in year of completion and no observable discrepancy in building type. b <100m difference in Euclidean distance by coordinates, <10-year difference in year of completion and no observable discrepancy in building type and no missing values in any of the filtering variables. c <10m difference in Euclidean distance by coordinates, a <10-year difference in year of completion and no observable discrepancy in building type and no missing values in any of the filtering variables. Scand J Work Environ Health – online first 5 Nikkilä et al tion in each except the output layer. The model was trained with 80% of the data with additional 20% used as validation for each epoch for 1000 epochs or until convergence according to mean squared error. We used five-fold cross validation to explore the robustness and potential over-fitting of the log-linear model. We also calculated the Spearman correlation between the measured and predicted indoor radon concentrations. Categorical models were evaluated with Cohen’s kappa. We performed sensitivity analyses on different levels of filtering regarding the slight discrepancies between databases. Childhood leukemia case–control study The indoor radon exposure was predicted with log-linear model for the cases and controls using our nationwide case-control dataset ( 48 ). Briefly, the cases included all Finnish children diagnosed with childhood leukemia during 1990–2011. The 1100 cases were identified from Finnish Cancer Registry (M9800 - M9948 in ICD-O-3). Three controls were individually matched on sex and year of birth to each case from the Finnish Population Register Center. Each control was assigned a reference date to match the diagnosis of the respective case. We assumed a two-year latency based on results summarized by UNSCEAR, which automatically results in null exposure for subjects less than two years of age at their reference date as well as their controls ( 53 ). These cases and their respective controls were excluded from the analyses. We obtained also complete residential histories which yielded, in total, 7334 residencies with the aforementioned latency period. As a sensitivity analysis, we experimented with a five-year latency period. The cases were classified by leukemia subtype into pre-B-ALL (precursor B-cell acute lymphoblastic leukemia), T-ALL (T-cell acute lympoblastic leukemia), unspecified ALL, AML (acute myeloid leukemia) and others. The genetic subtypes were obtained from the hospital records. We obtained data on gestational age, birth weight, maternal smoking from the Medical Birth Registry. Diagnoses of Down syndrome and other congenital malformations were obtained from the Congenital Malformation Registry. In addition, we obtained data on parental education, occupation, and socioeconomic status from Statistics Finland. When applying the model to the childhood leukemia dataset for subjects (3.0% for cases and 2.6% for controls) with only municipality of residence available (for at least one residence), we used municipalityspecific radon estimates. For residential periods abroad (1.4% for cases and 0.7% for controls), we used worldwide indoor radon average 39 Bq/m3 ( 54 ). In the rare cases where a dwelling could not be classified as either a house or an apartment, we also used the municipalityspecific median (1.2% for cases and 1.2% for controls). Otherwise, we applied the model after using multiple imputation for missing data on variables required for the prediction. As the dependent variable of the model was log-transformed before fitting the curve, the predictions represent geometric means of the estimated indoor radon concentrations when transformed back into Bq/m3. Radon exposure prediction We calculated cumulative radon exposure as Bq/m3 integrating over time to cover the whole residential history taking two-year latency period into account and divided it into quartiles for the conditional logistic regression analyses. We also calculated the average concentration of the exposure period by dividing the cumulative exposure with the total length of the exposure period. Cumulative exposure accumulates with age and, thus, is highly correlated with it. The analyses were adjusted for potential confounders: Down syndrome (yes or no), large birth weight (LGA) (exceeds 90th birth weight percentile in relation to gestational duration), terrestrial gamma radiation and Chernobyl fallout [cumulative red bone marrow equivalent dose (mSv)], cumulative red bone marrow dose from CT exposure (mGy), maternal smoking during pregnancy (yes or no), as well as parental socioeconomic status and education. Both socioeconomic status and education were known individually for each parent. Socioeconomic status was classified into five classes (self-employed, upper level employee, lower level employee, manual worker and other) and education into three levels (upper secondary, bachelor’s degree, master’s or doctor’s degree) ( 55 ). Statistical analysis All analyses were performed using R software version 3.4.0. For the modelling and visualization, the R libraries included: multiple imputation (Amelia, v. 1.7.5), k-fold cross validation (DAAG, v. 1.22), Cohen’s kappa (psych, v. 1.8.4), Bland-Altman plot (BlandAltmanLeh, v. 0.3.1; ggExtra, v. 0.8; ggplot2, v. 3.1.0), ordered logistic regression (MASS, v. 7.3-51), Brant’s test (brant, v. 0.2-0) multinomial logistic regression (nnet, v. 7.3-12), random forests (randomForest, v. 4.6-14), keras (keras, v. 2.2.0). The risk analyses after prediction were carried out with conditional logistic regression from the library survival (v. 2.43-1). Variance inflation was examined using car-library (v. 3.0-2). We used 5% as the significance threshold and all reported p-values are two-sided. For multiple testing corrections we used the BenjaminiHochberg method. Effect modification was investigated by including interaction terms into the model and evaluating improvement in model fit. 6 Scand J Work Environ Health – online first Indoor radon and childhood leukemia Ethical considerations No informed consent from the study subjects was needed according to the Finnish regulations as the study was carried out entirely through registers and databases, without any contact with the study subjects. Results Radon measurements The median indoor radon concentration in 93 219 unique linked dwellings from the STUK database was 137.3 Bq/m3 (IQR 68.0 Bq/m3, 267.4 Bq/m3), with the 95th percentile 732.7 Bq/m3, the 99th percentile 1913.0 Bq/m3 and the maximum 38,883 Bq/m3. The distribution was log-normal and after log-transformation, the distribution was normalized when evaluated using a Q-Q plot. After exclusions, the material included 73 903 (94.1%) houses and 3709 (4.7%) apartments, with median radon concentrations 143 Bq/m3 (IQR 71 Bq/m3, 276 Bq/m3) and 66 Bq/m3 (IQR 38 Bq/m3, 134 Bq/m3), respectively. The descriptive statistics and distributions of predictors are represented in tables 1a and b by indoor radon quartiles. Modelling indoor radon concentrations The final predictors, their estimates and confidence intervals (CI) with adjusted P-values for the log-linear model are reported in tables 2a, b and c. For the house model, most of the selected predictors had a highly statistically significant effect due to large sample size. Especially for the houses, the construction year displayed an inverted U-shaped curve relationship with indoor radon, with lower concentrations in newer buildings due to stricter radon protection regulation. Rockbased building materials were associated with higher residential radon than wood as a building material, and higher indoor radon concentrations were also associated with more porous soil. Uranium concentration in soil exerted a major influence in the house model. In general, we identified fewer predictors with mostly smaller coefficients for apartments. For both models (houses and apartments), the year of completion was an important predictor. It explained 10.6% and 4.61% of the variance and for the house and apartments, respectively. Soil permeability was also influential (houses 2.97% and apartments 7.05%). The other proportions of the variation explained by each predictor are reported in table 3. For the final log-linear house model, we observed Akaike’s information criterion (AIC) 157 739 and Bayesian information criterion (BIC) 158 036 and for the apartment model AIC 9993 and BIC 10 161. Performance of the models The final model of the log-transformed indoor radon concentration reached r2 of 0.21 for the house model and 0.20 the apartment model. The Spearman correlation between the measured and predicted values in the validation dataset was 0.45 for the houses and 0.44 for the apartments. The scatterplots of measured and predicted indoor radon concentrations also showed only a modest correlation with a narrower range of predicted than observed concentrations (figure 2), but both models were unable to accurately identify the lowest and highest radon concentrations (figure 3). In the five-fold crossvalidation with 80–20 split, the models appeared robust with no indication of substantial over-fitting for either model. The mean squared error was 0.84 for the houses and 0.88 for the apartments. We observed variance inflation due to multicollinearity of the predictors. For the apartments, the predictors with generalized varianceinflation (GVIF) >2 were soil permeability (5.1), formation by ice-age (4.3), year of completion (2.3) and soil uranium concentration (2.3). For the house model, five predictors showed GVIF >2: soil permeability (2.6), formation by ice-age (2.4), year of completion (2.6), floor area (2.3) and total volume (2.2). The weighted Cohen’s kappa for measured and predicted values by quartiles of measured indoor radon was 0.33 for houses and 0.38 for apartments. If only one split at 80th percentile was used, the weighted kappa was 0.10 for houses and 0.25 for apartments. Exploratory modelling attempts In exploratory analyses, the predictors of both dwelling types remained largely similar when an ordered logistic regression was used instead of the log-linear model to predict indoor radon in quartiles, but the assumption of parallel lines was not met for the categorized year of completion when evaluated with Brant’s test. This also applied to multinomial logistic regression. Ordinary logistic regression for binary radon split at p80 gave poor results. We did not observe major changes in r2 (0.21–0.24 for houses and 0.20–0.24 for apartments) or in the coefficients when different levels of measurement filtering were used. Using modern machine learning methods, we were able to markedly improve the coefficient of determination [random forest (apartments 0.23, houses 0.28), deep neural network (apartments 0.19, houses 0.18)]. We also observed lower coefficients of determination when using the newest available radon concentration for each dwelling. Scand J Work Environ Health – online first 7 Nikkilä et al Table 1a. Proportions and statistics of the predictors by measured indoor radon quartiles. [IQR=interquartile range.] 1st quartile 2nd quartile 3rd quartile 4th quartile Houses N=14 770 (43.3 Bq/m3) a Apartments N=932 (27.0 Bq/m3) a Houses N=14 767 (103 Bq/m3) a Apartments N=917 (50.6 Bq/m3) a Houses N=14 768 (196 Bq/m3) a Apartments N=924 (88.0 Bq/m3) a Houses N=14 768 (438 Bq/m3) a Apartments N=925 (250 Bq/m3) a N (%) N (%) N (%) N (%) N (%) N (%) N (%) N (%) Building material Rock-based 1712 (11.6%) 883 (94.6%) 2211 (15.0%) 867 (94.5%) 2445 (16.6%) 877 (94.9%) 2602 (17.6%) 805 (87.0%) Wood 12802 (86.7%) 44 (4.7%) 12286 (83.2%) 38 (4.1%) 12087 (81.8%) 36 (3.9%) 11973 (81.1%) 109 (11.8%) Other 256 (1.7%) 6 (0.6%) 270 (1.8%) 12 (1.3%) 236 (1.6%) 11 (1.2%) 193 (1.3%) 11 (1.2%) Soil permeability Impermeable 6166 (41.8%) 275 (29.5%) 6124 (41.5%) 273 (29.8%) 6094 (41.3%) 301 (34.7%) 5246 (35.5%) 253 (27.4%) Moderately permeable 6029 (40.8%) 186 (19.9%) 5989 (40.6%) 150 (16.4%) 5775 (39.1%) 175 (18.9%) 5230 (35.4%) 163 (17.7%) Highly permeable 1910 (12.9%) 97 (10.4%) 2097 (14.2%) 120 (13.1%) 2489 (16.9%) 127 (13.7%) 3996 (27.1%) 290 (31.4%) Formation by ice-age On a formation 1341 (9.1%) 82 (8.8%) 1630 (11.0%) 99 (10.8%) 1999 (13.5%) 108 (11.7%) 3365 (22.8%) 257 (27.8%) Mechanical ventilation Exhaust ventilation (approx.) b 1723 (11.7%) 459 (49.2%) 2701 (18.3%) 457 (49.8%) 3425 (23.2%) 476 (51.5%) 3622 (24.5%) 506 (54.7%) Basement Yes (built before 1990) 404 (2.7%) 85 (9.1%) 433 (2.9%) 52 (5.7%) 468 (3.2%) 35 (3.8%) 479 (3.2%) 46 (5.0%) Yes (built after 1990) 429 (2.9%) 57 (6.1%) 336 (2.3%) 50 (5.5%) 286 (1.9%) 49 (5.3%) 286 (1.9%) 58 (6.3%) Number of floors Median 1 4 1 4 1 4 1 3 IQR 1–2 3–6 1–2 3–4 1–2 4–5 1–2 2–4 p95 28282827 a Measured indoor radon concentration median. b For apartments, any ventilation reported and building completed between 1950 and 2006 was used as the definition; for houses, any ventilation and building competed before the year 2000 was used as the definition. Table 1b. Proportions and statistics of the predictors by measured indoor radon quartiles. [IQR=interquartile range.] I quartile II quartile III quartile IV quartile Houses (10.3–95.0) a Apartments (15.3–85.7) a Houses (10.3–95.0) a Apartments (11.0–85.8) a Houses (10.7–95.0) b Apartments (13.9–83.4) b Houses (10.7–95.0) b Apartments (13.9–87.1) b Median (IQR) Median (IQR) Median (IQR) Median (IQR) Median (IQR) Median (IQR) Median (IQR) Median (IQR) Number of floors 1 (1–2) 4 (3–6) 1 (1–2) 4 (3–4) 1 (1–2) 4 (4–5) 1 (1–2) 3 (2–4) Soil’s uranium (Bq/kg) 39.4 (29.6–51.6) 45.2 (31.5–62.5) 43.7 (32.3–53.7) 51.8 (34.5–66.3) 46.3 (36.9–56.6) 51.8 (41.4–66.3) 48.7 (40.2–58.1) 51.8 (45.0–59.7) Area of the floors (m2) 154 (116–200) 1973 (1322–3101) 160 (120–205) 2016 (1298–2928) 163 (126–206) 1784 (1187–2725) 160 (127–202) 1365 (751–2050) Total area (m2)184 (145–235) 2420 (1580–3724) 184 (149–244) 2396 (1538–3610) 184 (151–245) 2088 (1284–3320) 184 (151244) 1552 (822–2336) Total volume (m3)565 (455–750) 7382 (4941–11597) 570 (455–766) 7280 (4840–10738) 566 (459–766) 6482 (4250–9971) 560 (454–750) 4895 (2780–7340) Year of completion (year) 1979 (1956–2004) 1973 (1963–1986) 1981 (1963–1995) 1972 (1964–1984) 1983 (1970–1992) 1974 (1965–1985) 1983 (1973–1991) 1979 (1968–1988) Terrain elevation (m) 84 (31–106) 36 (13–101) 87 (43–108) 38 (17–96) 89 (46–109) 49 (17–101) 91 (50–112) 83 (25–114) Radon (area b-log) 6.67 (6.10–7.07) 3.67 (3.49–3.71) 6.63 (6.11–6.97) 3.72 (3.61–3.73) 6.68 (6.31–6.95) 3.78 (3.71–3.92) 6.82 (6.56–7.01) 3.85 (3.71–4.08) a Soil’s uranium (Bq/kg): min–max. b For houses and apartments the spatial units were postal areas and counties, respectively. Childhood leukemia case-control data After exclusions, we included 1093 (4 had prohibition of data use and 3 had incorrect identification codes) childhood leukemia cases diagnosed in 1990–2011. Of these, 826 (75.6%) were pre-B-ALL, 64 (5.9%) were T-ALL, 20 were unclassified ALL (1.8%), 146 were AML (13.6%), and 34 were other (3.1%). A majority of the cases were diagnosed at age 2–7 years, and the median age was 4.52 [interquartile range (IQR) 2.72, 8.23]. Down syndrome, intrauterine growth, and maternal smoking during pregnancy were associated with risk of childhood leukemia (table S2). In total, there were 7443 different dwellings (1839 for cases and 5604 for controls) in the subjects’ residential histories using the two-year latency period. The residential radon concentrations were estimated with either the house (56.1%, N=1032 for cases and 54.9%, N=3079 for 8 Scand J Work Environ Health – online first Indoor radon and childhood leukemia Table 2a. Coefficients and 95% confidence intervals of the predictors from the final model after backwards selection algorithm. [CI=confidence interval] Predictor Apartments Houses coefficent (95% CI) P-value acoefficent (95% CI) P-value a Other materials (ref) b0 Mainly built from rock-based materials - - 0.08 (0.02–0.14) 0.01 Mainly built from wood - - -0.07 (-0.13–0.01) 0.02 Unknown or other soil (ref) b0 0 Impermeable soil 0.04 (-0.02–0.12) 0.4 0.10 (0.06–0.14) <0.001 Moderate permeability soil 0.08 (-0.01–0.18) 0.10 0.16 (0.12–0.21) <0.001 Highly permeable soil 0.31 (0.13–0.48) 0.001 0.27 (0.22–0.32) <0.001 Not on a land formation by the ice-age (ref) b0 0 On a land formation by the ice-age 0.23 (0.06–0.41) 0.02 0.27 (0.24–0.3) <0.001 No basement (ref) b0 Basement, dwelling built before 1990 - - 0.33 (0.28–0.37) <0.001 Basement, dwelling built after 1990 - - 0.09 (0.04–0.14) 0.001 a Benjamini-Hochberg adjusted P-values. b The reference category for class variables. Table 2b. Coefficients and 95% confidence intervals of the predictors from the final model after backwards selection algorithm. Predictor Apartments Houses coefficent (95% CI) P-value acoefficent (95% CI) P-value a Before the first category (ref) b0 0 Built in 1940–1945 - - 0.11 (0.04–0.19) 0.005 1945–1950 - - 0.01 (-0.04–0.05) 0.8 1950–1955 -0.33 (-0.52– -0.13) 0.002 -0.14 (-0.19– -0.1) <0.001 1955–1960 -0.25 (-0.42– -0.09) 0.006 -0.20 (-0.25– -0.16) <0.001 1960–1965 -0.32 (-0.46– -0.18) <0.001 0.10 (0.05–0.15) <0.001 1965–1970 -0.25 (-0.38– -0.12) 0.001 0.22 (0.18–0.27) <0.001 1970–1975 -0.13 (-0.25– -0.01) 0.06 0.35 (0.31–0.39) <0.001 1975–1980 0.01 (-0.13–0.15) 0.90 0.43 (0.39–0.47) <0.001 1980–1985 0.26 (0.12–0.39) 0.001 0.61 (0.57–0.65) <0.001 1985–1990 0.22 (0.08–0.36) 0.005 0.57 (0.53–0.61) <0.001 1990–1995 0.14 (-0.02–0.31) 0.10 0.50 (0.45–0.55) <0.001 1995–2000 -0.04 (-0.23–0.15) 0.70 0.32 (0.27–0.37) <0.001 2000–2005 -0.06 (-0.27–0.14) 0.60 0.08 (0.04–0.13) <0.001 2005–2010 -0.43 (-0.66– -0.2) 0.001 -0.14 (-0.19– -0.1) <0.001 2010–2015 -0.45 (-0.72– -0.18) 0.003 -0.53 (-0.57– -0.48) <0.001 2015–2020 0.06 (-0.86–0.98) 0.90 -0.76 (-0.9– -0.61) <0.001 a Benjamini-Hochberg adjusted P-values. b The reference category for class variables. Table 2c. Coefficients and 95% confidence intervals of the predictors from the final model after backwards selection algorithm. Predictor Apartments Houses coefficent (95% CI) P-value acoefficent (95% CI) P-value a Number of floors -0.16 (-0.25– -0.07) 0.001 -0.15 (-0.18– -0.13) <0.001 Floor area -0.13 (-0.07– -0.19) <0.001 -0.05 (-0.07– -0.03) <0.001 Total area of the building - - - - Total volume of the building - - -0.01 (-0.03–0.00) 0.2 Elevation from sea level 0.07 (0.03–0.10) 0.001 0.15 (0.14–0.16) <0.001 Soil’s uranium concentration 0.17 (0.06–0.29) 0.007 0.70 (0.68–0.73) <0.001 County specific median radon 0.96 (0.79–1.13) <0.001 - - Postal area specific median radon - - 0.09 (0.08–0.10) <0.001 Intercept 0.77 (0.19–1.35) 0.02 1.09 (0.92–1.25) <0.001 a Benjamini-Hochberg adjusted P-values. b The reference category for class variables. Scand J Work Environ Health – online first 15 Nikkilä et al aa7922. 45 . Valmari T , Arvela H , Reisbacka H . Radon in Finnish apartment buildings . Radiat Prot Dosimetry 2012 Nov; 152 ( 1-3 ): 146 – 9 . https://doi.org/10.1093/rpd/ncs211. 46 . Arvela H , Holmgren O , Hänninen P . Effect of soil moisture on seasonal variation in indoor radon concentration: modelling and measurements in 326 Finnish houses . Radiat Prot Dosimetry 2016 Feb; 168 ( 2 ): 277 – 90 . https://doi. org/10.1093/rpd/ncv182. 47 . Arvela H , Holmgren O , Reisbacka H . Radon prevention in new construction in Finland: a nationwide sample survey in 2009 . Radiat Prot Dosimetry 2012 Mar; 148 ( 4 ): 465 – 74 . https://doi.org/10.1093/rpd/ncr192. 48 . Nikkilä A , Erme S , Arvela H , Holmgren O , Raitanen J , Lohi O et al. Background radiation and childhood leukemia: A nationwide register-based case-control study . Int J Cancer 2016 Nov; 139 ( 9 ): 1975 – 82 . https://doi.org/10.1002/ ijc.30264. 49 . Arvela H , Castrén O , Mäkeläinen I . Survey of residential radon in Finland, STUK-A108 (in Finnish). Helsinki: Säteilyturvakeskus; 1993 . 50 . Arvela H , Markkanen M , Lemmelä H . Mobile Survey of Environmental Gamma Radiation and Fall-Out Levels in Finland After the Chernobyl Accident . Radiat Prot Dosimetry 1990 ; 32 : 177 – 84 . https://doi.org/10.1093/ oxfordjournals.rpd.a080734. 51 . Breiman L . Random Forests . Mach Learn 2001 ; 45 : 5 – 32 . https://doi.org/10.1023/A:1010933404324. 52 . Schmidhuber J . Deep learning in neural networks: an overview . Neural Netw 2015 Jan; 61 : 85 – 117 . https://doi. org/10.1016/j.neunet.2014.09.003. 53 . UNSCEAR . Effects of radiation exposure on children. 2013 . 54 . UNSCEAR . Sources and effects of ionizing radiation. 2008 . 55 . Nikkilä A , Raitanen J , Lohi O , Auvinen A . Radiation exposure from computerized tomography and risk of childhood leukemia: finnish register-based case-control study of childhood leukemia (FRECCLE) . Haematologica 2018 Nov; 103 ( 11 ): 1873 – 80 . https://doi.org/10.3324/ haematol.2018.187716. 56 . Muikku M , Bly R , Kurttio P , Lahtinen J , Lehtinen M , Siiskonen T et al. Suomalaisten keskimääräinen efektiivinen annos: Annoskakku 2012. [Abstract available in English] Säteilyturvakeskus; 2014 . 57 . Verger P , Hubert P , Cheron S , Bonnefous S , Bottard S , Brenot J . Use of Field Measurements in Radon Mapping in France . Radiat Prot Dosimetry 1994 ; 56 : 225 – 9 . https://doi. org/10.1093/rpd/56.1-4.225. 58 . Verdi L , Weber A , Stoppa G . Indoor radon concentration forecasting in South Tyrol . Radiat Prot Dosimetry 2004 ; 111 ( 4 ): 435 – 8 . https://doi.org/10.1093/rpd/nch069. Received for publication : 20 March 2019