|Home | About | Journals | Submit | Contact Us | Français|
Pooled analyses may provide etiologic insight about associations between exposure and disease. In contrast to childhood leukemia, no pooled analyses of childhood brain tumors and exposure to extremely low-frequency magnetic fields (ELF-MFs) have been conducted. The authors carried out a pooled analysis based on primary data (1960–2001) from 10 studies of ELF-MF exposure and childhood brain tumors to assess whether the combined results, adjusted for potential confounding, indicated an association. The odds ratios for childhood brain tumors in ELF-MF exposure categories of 0.1–<0.2 μT, 0.2–<0.4 μT, and ≥0.4 μT were 0.95 (95% confidence interval: 0.65, 1.41), 0.70 (95% CI: 0.40, 1.22), and 1.14 (95% CI: 0.61, 2.13), respectively, in comparison with exposure of <0.1 μT. Other analyses employing alternate cutpoints, further adjustment for confounders, exclusion of particular studies, stratification by type of measurement or type of residence, and a nonparametric estimate of the exposure-response relation did not reveal consistent evidence of increased childhood brain tumor risk associated with ELF-MF exposure. These results provide little evidence for an association between ELF-MF exposure and childhood brain tumors.
Since 1979, numerous residential studies have examined the potential association between exposure to extremely low-frequency magnetic fields (ELF-MFs) and childhood cancer (1, 2). The initial research examined all childhood cancers as an endpoint, but later focus shifted to a potential association between residential ELF-MF exposure and childhood leukemia (3). Two pooled analyses (4, 5) have provided a basis for concluding that an association exists between residential exposure to ELF-MFs above 0.3 μT/0.4 μT and the risk of childhood leukemia. Based largely on this epidemiologic association, the International Agency for Research on Cancer has classified ELF-MF exposure as possibly carcinogenic to humans (Group 2B) (6).
Evidence linking childhood brain tumors to residential ELF-MF exposure appears to be weaker. Individual studies of childhood brain tumors, usually suffering from even smaller numbers of exposed cases than leukemia studies, have generally been unable to examine the potential association with elevated ELF-MF exposure with satisfactory statistical precision. In a recent meta-analysis, Mezei et al. (7) found no overall increase in childhood brain tumor risk, with the exception of high-cutpoint analyses (0.3 μT or 0.4 μT), where the possibility of a moderate risk increase could not be excluded. Unlike studies of childhood leukemia, however, to our knowledge no pooled analyses of brain tumors and ELF-MFs have been conducted. Such an analysis was identified as a high research priority in the World Health Organization research agenda issued in 2007 (8).
We carried out a pooled analysis based on primary data from 10 studies of ELF-MFs and childhood brain tumors to assess whether the combined results, adjusted for potential confounding, indicated an association between ELF-MF exposure and childhood brain tumors.
We searched the published literature through PubMed, as well as references of identified papers, and conducted an informal survey of epidemiologists involved in ELF-MF research to identify relevant studies on residential ELF-MF exposure and childhood brain tumors. To be included, studies had to provide data for children, provide separate data for cancers of the brain or central nervous system, and provide measured or calculated values for residential exposure to ELF-MFs.
We identified 16 studies published between 1979 and 2010, of which 10 could be included in the pooled analyses (Table 1). Four studies (2, 9–11) were not included because they did not have measurements or calculated fields for childhood brain tumors. One study (12) was not included because it used dwellings rather than persons as units of analysis, included only outside spot measurements, and overlapped with another study (13). A sixth study (14) was not included because its cases were a small subset of those from another study (15). Appendix Table 1 summarizes the methods and findings of the studies that did not meet our inclusion criteria. Two included studies had a large overlap, with perhaps 90% of the cases in the United Kingdom Childhood Cancer Study (UKCCS) (16) also being included in the Kroll et al. (15) study, but they differed in terms of type of exposure surrogates and timing of exposure. To maintain independence of observations, we included only 1 of these studies in any given analysis. Investigators in the UKCCS (16) used a 2-phase measurement strategy, in which 48-hour measurements were conducted when either a shorter measurement (108 minutes) or a characteristic of the residence indicated that ELF-MF exposure was elevated. In this pooled analysis, we used only second-phase measurements, which were taken for all subjects (with their matched cases or controls) who had potential sources of high exposure (>0.1 μT) in the first phase.
Investigators utilized stratified sampling of controls in all of the studies, although the stratification variables were not the same in all studies. In Finland, the authors of the original publication reported findings from a cohort study (17), but in preparation for this pooled analysis, a control group was selected and the data were evaluated using a matched case-control design with 6 additional years of follow-up. For some studies, the same controls were used for both brain tumors and leukemia cases in the original publications, and we maintained the same approach in these instances. Since we wanted to use as many of the cases and controls as possible to increase the flexibility of the analysis, we ignored the matching and instead included adjustment for age at diagnosis, gender, and study.
To make the data as consistent as possible across studies, we limited the age of diagnosis to 0–15 years inclusive and converted all measurements to microteslas. One Finnish patient with 3 tumors was included only once. One Japanese patient with cavernous angioma and 2 corresponding controls were excluded. However, germ-cell tumors and corresponding controls were included in the analysis, although these tumors were not consistently included in all individual studies.
We focused on ELF-MFs present in the general area of the home; that is, we excluded exposure occurring in schools, data on which would have been available for only 1 study (UKCCS) (16), and did not include short-duration exposures close to appliances. In all studies, investigators took long-term measurements or spot measurements and/or calculated the strengths of ELF-MFs. Only 2 of the included studies, both from the United States (18, 19), had information on wire codes, but they also had measurements. We did not use wire codes in the analyses. With regard to long-term measurements, measurements were taken for 24 hours in 2 studies (18, 20), for 48 hours (for highly exposed subjects) in 1 study (16), and for a 1-week period in 1 study (21); these measurements are referred to as “long-term measurements” throughout this paper. Arithmetic mean values rather than geometric means were used, since these were available for all studies.
Our analyses included separate analyses for long-term measurements, calculated fields, and spot measurements. Investigators had collected data from the child's birth home, the home in which the child had lived longest prior to diagnosis, the latest home in which the child had lived prior to diagnosis that was near a power line, and/or the home in which the child was living at diagnosis. For subjects with data from more than 1 home, we used the following hierarchy to select a single exposure proxy for the analysis: Diagnosis home was used if data were available; if not, then the latest home lived in before diagnosis; if not, then the home lived in the longest; and if not, then birth home. We also performed separate analyses for diagnosis homes, longest-lived-in homes, and birth homes. For these analyses, for subjects with more than 1 type of exposure proxy, we used long-term measurements if they were available; if not, then we used calculated fields; and if neither measure was available, then we used spot measurements. In addition, in an attempt to maximize the available sample size, we conducted a “best measure” pooled analysis in which we selected a single best exposure measurement for each subject using both the exposure metric hierarchy (long-term over calculated fields over spot) and the home hierarchy (diagnosis over latest over longest-lived-in over birth). The choice between selection of type of exposure metric before residence and selection of residence before type of exposure metric was immaterial, since both approaches yielded identical measurements for each subject.
Additional potential confounders for which data were available included type of dwelling, mobility, urbanization, socioeconomic status, and exposure to traffic exhaust. The number, type, and coding of potential confounders differed among the studies (see Table 1). We examined socioeconomic status (standardized to a 3-level ordinal variable), urbanization (dichotomized as urban/rural), dwelling type (dichotomized as single-family/multiple-unit), and mobility (dichotomized as number of residences before diagnosis: 1/>1).
An analysis plan that included hypotheses, hierarchy of measurements, and cutpoints was developed and agreed upon prior to analysis. The data were analyzed using both ordinary logistic regression, with fixed intercepts to adjust for study, and mixed-effects logistic regression, with random intercepts and exposure effect coefficients for study. Separate intercepts were used for East Germany and West Germany. Ordinary and mixed-effects logistic regression gave similar results; we present results for ordinary logistic regression.
We conducted an analysis using best measures with continuous exposure as a linear predictor, reporting results as the odds ratio for an increase of 0.2 μT. We also used this analysis for a likelihood ratio test of homogeneity of effects across studies, in which models with and without study-specific coefficients for exposure were compared. In addition, we estimated the trend in the log odds of being a case using a generalized additive model (22), using a nonparametric curve (natural cubic smoothing spline with interior and boundary knots at the unique values of exposure) to estimate the risk associated with exposure, while controlling for study, age, and gender. The amount of smoothing is determined by the degrees of freedom (df), with higher df corresponding to less smoothing. For this analysis, we transformed exposures using an inverse cubic transformation to reduce the influence of outliers at high exposure levels. These analyses were conducted using the gam package in R, version 2.9.2 (23).
In other analyses, we used increasing exposure categories of 0.1–<0.2 μT, 0.2–<0.4 μT, and ≥0.4 μT to examine the exposure-response relation, with a reference category of <0.1 μT. Single cutpoints at 0.3 μT and 0.4 μT were also explored. We also obtained odds ratios using a moving window of exposure. These analyses used exposure categories of 0.1–<0.2 μT, 0.15–<0.25 μT, 0.20–<0.30 μT, 0.25–<0.35 μT, ≥0.30 μT, ≥0.35 μT, and ≥0.40 μT, with a reference category of <0.1 μT, and results were adjusted for age, gender, and study. Treating age at diagnosis as continuous or categorical gave similar results; results obtained with age included as a continuous variable are presented. The influence of individual studies was examined by omitting 1 study at a time. These analyses were conducted using Stata 10 (24).
Among the included studies, 7 were conducted in Europe (13, 15–17, 20, 25, 26), 2 in the United States (18, 19), and 1 in Japan (21). In 4 studies (16, 18, 20, 21), investigators examined childhood brain tumor risk in relation to long-term exposure to ELF-MFs, and in 5 studies (13, 15, 17, 25, 26) they used calculated ELF-MFs. Spot ELF-MF measurements were available in 6 studies (13, 16, 18–21); spot measurements were the only available exposure metric in 1 study (19) (Table 1).
Table 1 also shows the numbers of cases and controls for each study, along with data on variables supplied in those publications. There were a total of 8,372 cases and 11,494 controls with 1 or more of the 3 types of exposure metrics, and 79% of cases were contributed by the large registry-based United Kingdom case-control study (15). However, numbers in the high-exposure categories remained small even for this large set of data.
Table 2 shows the distribution of subjects by exposure level for long-term, spot, and calculated field studies. The number of subjects with values below 0.3 μT varied by study. Values above 0.3 μT or 0.4 μT were relatively infrequent in all studies, even in studies that focused on the population living next to power lines (ranging from 0.03% to 6.53% for fields above 0.4 μT).
Table 3 presents results by type of exposure metric. The pattern was not consistent: Risk did not increase with increasing exposure. Notable features were dips in the odds ratios for 0.2–<0.4 μT versus <0.1 μT for calculated fields and ≥0.4 μT versus <0.1 μT for spot measurements.
Table 3 also presents results by residence for which the estimate of exposure was obtained. These analyses included calculated fields for some studies and long-term measurements for others; the analysis for home at diagnosis included all 3 types of measurements. The highest odds ratio for ≥0.4 μT versus <0.1 μT was for the longest-lived-in home; however, this was based on only 3 exposed cases. Again, there was no clear exposure-response pattern.
The continuous best-measure exposure analysis gave an odds ratio estimate of 0.96 for each 0.2-μT increase in exposure (95% confidence interval (CI): 0.86, 1.07). The homogeneity test based on this analysis yielded a χ2 statistic (8 df) of 5.78, corresponding to a P value of 0.672, which supports the appropriateness of pooling.
Table 4 presents results from the best-measure analysis with categorical exposure levels, along with adjustment for a number of potential confounders. Because the 2 United Kingdom studies (15, 16) included a large overlap of subjects, we present results with only 1 of those studies included in any given analysis. Not all potential confounders were available in all studies. Analyses adjusting for confounding were carried out in the subset of studies and subjects for whom data on the confounder were available. While risks for a given subset varied, adjustment did not change the odds ratio estimates. All confidence intervals included the null value. Results shown here and elsewhere were similar when a cutpoint of 0.3 μT was used (data not shown).
Examination of the residentially stable subgroup also did not show an increase in risk with exposure (Table 4). We evaluated age as a possible effect modifier by stratifying the sample at age 8 years at diagnosis (bottom of Table 4). While the odds ratio for the ≥0.4-μT category was higher in the older age group, the odds ratios for the lower exposure categories were lower in the older age group, and the results did not provide consistent evidence that the effect of exposure varied by age.
Odds ratio estimates using categorical cutpoints and involving relatively small numbers of subjects are vulnerable to unstable results. To address this concern, we also calculated odds ratios using a moving window of exposure levels (Figure 1). These results did not suggest a trend of increasing risk with increasing exposure: There was a dip in the intermediate exposure category, after which the risk returned to 1.
Figure 2 presents nonparametric estimates of the trend in the log odds of being a case from a generalized additive model, with adjustment for study, age, and gender. The confidence intervals widen as exposure increases, reflecting the smaller number of subjects at high exposure levels. The curves suggest only variation around the null value of zero. The spikes are artifacts created when multiple subjects have coinciding case/control status and exposure values.
Influence analyses omitting studies one at a time yielded odds ratios for ≥0.4 μT versus <0.1 μT ranging from 0.99 to 1.41, with adjustment for gender and age at diagnosis. Omission of any individual study from the analysis did not produce important changes in the summary odds ratio (Figure 3). The most influential study was the study by Preston-Martin et al. (18). Although this study was relatively small in terms of the overall number of subjects, it had the highest number of exposed cases (7 cases with exposure of ≥0.4 μT); the other studies had 0–2 cases in this exposure category. The odds ratio for the 4 methodologically similar Nordic studies (13, 17, 25, 26) (combined odds ratio (OR) = 1.53, 95% CI: 0.65, 3.59), which were record-based and thus thought to be less susceptible to selection bias, was higher than the odds ratio for the other studies (15, 16, 18–21) (combined OR = 0.78, 95% CI: 0.36, 1.70).
We conducted a pooled analysis of 10 epidemiologic studies on the association of residential ELF-MF exposure and childhood brain cancer. Pooled analysis, considered the gold standard for synthesizing results from multiple studies, allows for comparison across different studies and metrics, free of artifacts introduced by analytic differences, and for derivation of statistically more stable results. Pooled analysis uses raw data from previous studies and thus can apply identical analyses to all included studies. The choices of cutpoints, reference groups, metrics, etc., in a pooled analysis may differ from the choices made in the original studies and may result in changes in the study-specific effect estimates.
Results from pooled analyses, however, are prone to the same biases that might have been operating in the original studies. The studies using measurements generally had low participation rates (40%–80% as reported in the original studies) and thus had large potential for selection bias (27, 28). Studies estimating risks for calculated fields do not require participation and thus are less vulnerable to selection bias, but they neglect sources of ELF-MFs other than high-voltage power lines and thus are likely to introduce exposure misclassification and loss of statistical power. In the United Kingdom, for example, it has been shown that high-voltage sources account for only 23% of exposures above 0.2 μT and 43% of those above 0.4 μT (29). In this context, note that in order to include as many subjects as possible, we included calculated fields from the Kroll et al. (15) study in preference to the UKCCS measured fields (16) in some analyses. Since the original studies did not follow a common protocol, we were forced to make analytical choices, particularly for exposure assessment, when we combined data from individual studies. However, results were not sensitive to the choices we made and were consistent regardless of type of exposure metric used and with omission of single studies.
None of our analyses showed statistically significant increases or a consistent pattern suggestive of an association between childhood brain cancer risk and various measures of residential ELF-MF exposure. Some of the odds ratios for the highest exposure category (ELF-MFs above 0.3 μT or 0.4 μT) were elevated but had 95% confidence intervals that included 1. The pattern of odds ratios across increasing exposure categories was not consistent with an exposure-response relation; odds ratios for the middle category were mostly below 1, again with wide confidence intervals.
Our results were robust and insensitive to cutpoints, age and confounder adjustment, and various model assumptions and specification. The numbers and available information did not allow for separate analysis of different histologic types. Although there was some heterogeneity between studies, our results were not driven by any single study.
In the previous meta-analysis, results from 13 epidemiologic studies of childhood brain cancer provided no strong evidence for an increase in risk with residential exposure to ELF-MFs (7). The suggestion of a moderate risk increase at higher exposure levels (>0.3 μT or 0.4 μT), however, could not be excluded (OR = 1.68, 95% CI: 0.83, 3.43) (7).
Although our results contained hints of a risk increase in some subanalyses, as is expected when numerous analyses are performed, these increases were small, highly dependent on particular studies included in the subset, and inconsistent with regard to increasing exposure for all models chosen. Taken as a whole, our results provide little evidence for an association between ELF-MF exposure and childhood brain tumors.
Author affiliations: Department of Epidemiology, School of Public Health, University of California, Los Angeles, Los Angeles, California (L. Kheifets, J. Monroe, S. Oksuzyan); Division of Epidemiology, Institute of Environmental Medicine, Karolinska Institute, Stockholm, Sweden (A. Ahlbom, M. Feychting); Department of Biostatistics, School of Public Health, University of California, Los Angeles, Los Angeles, California (C. M. Crespi); Danish Cancer Society, Copenhagen, Denmark (C. Johansen, J. Schüz); Institute of Medical Biostatistics, Epidemiology and Informatics, University of Mainz, Mainz, Germany (J. Schüz); Childhood Cancer Research Group, University Department of Paediatrics, Medical Sciences Division, University of Oxford, Oxford, United Kingdom (M. F. G. Murphy); Department of Preventive Medicine, School of Medicine, University of Southern California, Los Angeles, California (S. Preston-Martin); Epidemiology and Genetics Unit, Department of Health Sciences, University of York, York, United Kingdom (E. Roman, J. Simpson); Department of Health Policy, National Research Institute for Child Health and Development, Tokyo, Japan (T. Saito); Disease Prevention and Public Health Institute, Mount Sinai School of Medicine, New York, New York (D. Savitz); National Grid plc, London, United Kingdom (J. Swanson); Cancer Registry of Norway, Institute of Epidemiological Cancer Research, Oslo, Norway (T. Tynes); Environmental Epidemiology Unit, Department of Environmental Health, National Public Health Institute, Kuopio, Finland (P. Verkasalo); and Environment Department, Electric Power Research Institute, Palo Alto, California (G. Mezei).
The authors thank Dr. M. Kelsh for providing data for one of the studies. The authors acknowledge funding from the Electric Power Research Institute, Palo Alto, California (grant P19836C9738) and Southern California Edison, Rosemead, California (grant POQ1159217). Dr. C. M. Crespi was also supported by National Cancer Institute grant CA016042.
Dr. J. Swanson worked on this paper with permission from National Grid plc, London, United Kingdom, but no results were shared with National Grid nor was approval sought from anyone in National Grid. The views expressed are solely those of the authors.