American Journal of Epidemiology Advance Access originally published online on May 15, 2008
American Journal of Epidemiology 2008 167(12):1504-1510; doi:10.1093/aje/kwn086
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
PRACTICE OF EPIDEMIOLOGY |
Assessment of Selection Bias in the Canadian Case-Control Study of Residential Magnetic Field Exposure and Childhood Leukemia
1 Electric Power Research Institute, Palo Alto, CA
2 British Columbia Cancer Agency, Vancouver, British Columbia, Canada
3 PW International, Inc., Burnaby, British Columbia, Canada
Correspondence to Dr. Gabor Mezei, 3420 Hillview Avenue, Palo Alto, CA 94304 (e-mail: gmezei{at}epri.com).
Received for publication November 12, 2007. Accepted for publication March 11, 2008.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
The authors evaluated the role of selection bias in the 1999 Canadian case-control study of residential magnetic field exposure and childhood leukemia. They included cases, participating controls, and first-choice nonparticipating controls in their analyses. Exposure was assessed by wire coding, a classification system based on the distribution line characteristics near homes. Although an imperfect measure of magnetic field exposure, wire coding is the only method applicable to nonparticipating subjects. First-choice nonparticipant controls tended to be of lower socioeconomic status than their replacements (non-first-choice participant controls), and lower socioeconomic status was related to higher wire code categories. The odds ratios for developing childhood leukemia in the highest exposure category were 1.6 (95% confidence interval: 1.0, 2.6) when the actual participating controls were used and 1.3 (95% confidence interval: 0.8, 2.1) when the first-choice ideal controls were used, regardless of their participation. Overall, the authors conclude that, although there is some evidence for control selection or participation bias in the Canadian study, it is unlikely to explain entirely the observed association between magnetic field exposure and childhood leukemia. Inherent problems in exposure assessment for nonparticipating subjects, however, limit the interpretations of these results, and the role of selection bias cannot entirely be dismissed on the basis of these results alone.
case-control studies; child; electromagnetic fields; epidemiologic methods; leukemia; selection bias
Abbreviations: CI, confidence interval; IPPE, income per person equivalent; OR, odds ratio
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
On the basis of an epidemiologic association observed in childhood leukemia case-control studies, power-frequency magnetic field exposure has been classified as a possible human carcinogen (group 2B) by the International Agency for Research on Cancer (1). This classification was heavily influenced by the results of two pooled analyses showing an approximately twofold increase in childhood leukemia risk at exposure levels above 0.3–0.4 µT (2, 3). Because neither conclusive laboratory evidence nor a known biophysical mechanism is available to explain a causal relation between magnetic field exposure and childhood leukemia, control selection bias in case-control studies is considered to be a possible noncausal explanation for the observed epidemiologic association (4). Although some evidence in support of the selection bias hypothesis exists, it is far from conclusive. The main difficulty in assessing selection bias in previously conducted case-control epidemiologic studies of magnetic field exposure and childhood leukemia is that no information, especially no exposure information, is available on the potential control subjects, who could not be identified or contacted or who had declined to participate in the study. Hatch et al. (5), in a reanalysis of the 1997 National Cancer Institute study (6), were able to compare results with and without including partial participant controls in the analysis. Partial participant controls were study subjects for whom exposure was assessed by way of magnetic field measurements at the front door or by wire coding, but for whom no indoor measurements were available. Hatch et al. found that inclusion of partial participant controls in the analyses attenuated the association between childhood leukemia and measures of magnetic field exposures (measurements at the front door and wire codes). However, no information was available on nonparticipant controls, and the full effect of selection bias could not be assessed.
Among the studies included in the pooled analyses (2, 3), only in the study by McBride et al. (7) could the authors identify nonparticipant controls by virtue of selecting controls from provincial health insurance rolls. Although the original Canadian study was interpreted as providing no evidence of an association between magnetic field exposure measures and childhood leukemia, the Canadian data—when combined with other studies and included in the pooled analyses with uniform cutpoints and methods—also contributed to the observed association.
Taking advantage of available information on nonparticipant controls, we attempted to evaluate the effect of potential selection bias on the observed effect estimates in the Canadian study (7). Although wire codes are a far-from-perfect measure of magnetic field exposure, we used wire coding to estimate exposure because this method could be applied to any known address and required no subject participation or property access.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
The current analyses examined data from the 1999 study by McBride et al. (7). Details of subject recruitment, control selection, and data collection are described in the original publication. Briefly, the study was a case-control study of incident childhood leukemia cases diagnosed between 1990 and 1995 in five Canadian provinces. Controls were randomly selected from the provincial health insurance rolls and matched on age, sex, and province of residence. In Canada, the province-sponsored insurance plans cover an estimated 95 percent of the population (8). If the first selected control for a given case could not be identified or declined to participate, then a second control was selected. If the second-choice control did not participate, then a third control was selected. The process was repeated until a matched control could be selected for each case. Overall, 399 participating cases and 399 participating controls were included in the 1999 analysis. The overall participation rates were 89 percent and 59 percent for cases and controls, respectively. In Alberta, it could not be determined from the records whether a selected control was first choice or later choice; therefore, 59 participating case-control pairs and eight nonparticipating controls from that province were excluded from our current analyses. From the remaining four provinces, 340 cases, 187 first-choice participating controls, 153 non-first-choice participating controls, and 153 first-choice nonparticipating controls, a total of 833 subjects, were included in the current analysis (table 1).
|
New wire coding was completed for the address at diagnosis (or corresponding time) of the included cases and participating and nonparticipating controls between November 2004 and June 2006. Wire coding was conducted following the protocols used by Wertheimer and Leeper (9) and Kaune and Savitz (10), and the team collecting the data was blinded to the case/participant control/nonparticipant control status of the residences. The Wertheimer-Leeper wire code classification system has five categories (underground and very low, ordinary low, ordinary high, and very high current configurations) from lowest to highest. The Kaune-Savitz wire code classification system has three categories (low, medium, and high). As in the original 1999 study, where the residence was part of an apartment block that was greater than five stories above ground level or four suites per floor, the residence was classified in the "APT" category, and no wire coding was assigned as both wire coding and perimeter magnetic field measurements (in the 1999 study only) were not considered to reflect accurately the magnetic field environment of the residence itself.
In the original 1999 study, wire coding was attempted for residences of the mother from 1 year before the birth of the subject to the time of birth and for residences of the child from birth to the diagnosis or reference date. For those residences for which consent could be obtained, the team was able to enter the subject's property. The distance measurements required for the wire coding (usually for measuring distance of the power line from the side of the house) were completed by use of a distance-measuring wheel.
In the new study, wire coding was carried out without entering the subjects' property, for the addresses at the time of diagnosis for the cases or at the corresponding time for controls. Either a distance-measuring wheel or a handheld laser distance meter (Disto Classic 5a; Leica Geosystems AG, Heerbrugg, Sankt Gallen, Switzerland), with a range of 0.2–200 m and typical accuracy of ±1.5 mm, was used to determine the separation distance between the power line and the residence. The distance meter has an integrated telescopic viewer that allows the measurement target to be pinpointed accurately. Where the power line was at the back of the house and it was impossible to see whether there was any protrusion at the back of the house that might decrease the separation distance between the house and the power line, the rear corner of the house as viewed from the front of the house was used as the reference point for distance measurement. This limited access to the rear of the property could result in an error in the distance measurement and misclassification of the first and other span secondary in the wire codes.
In our analyses, the main comparison was between the first-choice nonparticipant controls and the non-first-choice participant controls selected for the same individual cases. In the original study, the latter group was used as a replacement for the former group, and any difference in exposure distribution between these groups may directly affect the exposure effect estimates. Effect estimates (odds ratios) were calculated by both the "ideal" control group (all first-choice controls regardless of whether they were participants or not) and the "actual" control group (all participating controls regardless of whether they were first- or later-choice controls); results were then compared.
Comparisons were made between wire coding for the subjects with wire codes measured in both the original and current studies. Comparisons were also made for wire code categories, income, and urban/rural status between first-choice nonparticipant and non-first-choice participant controls. (Rural status was assigned to municipalities with a population size of less than 100,000.) Odds ratios and 95 percent confidence intervals were estimated by matched-pair conditional logistic regression. Associations between wire code category and income were assessed by use of linear regression.
Estimates for the risk of childhood leukemia for each of the wire code metrics were also computed. To be comparable with the original study and the pooled analyses, odds ratios and 95 percent confidence intervals were computed by unconditional logistic regression with adjustment for age, sex, and province of residence. Results were similar when we used matched-pair conditional logistic regression (not reported). Tests for trend were computed by entering consecutive integers for the wire code categories.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
In the original 1999 study, the homes of 630 subjects (participants only) in the four included provinces were wire coded. For the current analyses, wire coding was completed for the homes of 814 subjects. Of these subjects, the homes of 626 subjects were wire coded both times. The overall agreement between the old and new wire-coding results was generally good (table 2). For 122 subjects, the two assessments resulted in different wire code categories, with a tendency for higher categories according to the more recent assessment (ptrend < 0.001) (table 2). The reasons for the discrepancies in wire coding between the two assessments were misclassification due to limited access (42 subjects), change in wiring (33 subjects), inaccuracies in wire size estimation (23 subjects), new constructions (12 cases), inaccuracies in distance measurement (10 subjects), and error in the original classification (two subjects).
|
Of the 153 first-choice nonparticipant–non-first-choice participant control pairs, we had wire code information for both subjects in 143 pairs. (No wire codes were available for 10 first-choice nonparticipant controls.) Neither the Wertheimer-Leeper (ptrend = 0.84) (table 3) nor the Kaune-Savitz (ptrend = 0.82) (results not shown) wire code category distributions showed major departure from symmetry in pairwise comparisons. For the Wertheimer-Leeper wire code categories, the first-choice nonparticipant controls were, however, slightly more likely to be in the very high wire code category (odds ratio (OR) = 1.5, 95 percent confidence interval (CI): 0.7, 3.4) compared with non-first-choice participant subjects. For the Kaune-Savitz wire code categories, this tendency was weaker (OR = 1.2, 95 percent CI: 0.6, 2.2).
|
Socioeconomic status, as measured by the income per person equivalent (IPPE) of the residence area, was clearly related to wire code categories; higher wire code categories were associated with lower IPPE (table 4). Socioeconomic status, as measured by IPPE, was also significantly higher among non-first-choice participant controls than among first-choice nonparticipant controls (table 5). In the rural areas, we observed substantially less apartments and homes with underground wire code categories (table 4), while urban subjects were somewhat less likely to participate (table 5). Adjustment for IPPE or urban/rural status, however, did not substantially change the observed relation between participation status and wire code categories when we compared first-choice nonparticipant and non-first-choice participant controls (table 6).
|
|
|
In table 7, we present the main results of our analyses. Using the original Wertheimer-Leeper wire code classification from 1999, we found that case-control comparisons (with actual controls) resulted in odds ratios of 0.8 (95 percent CI: 0.6, 1.11) and 1.7 (95 percent CI: 1.0, 3.0) for ordinary high and very high current configuration categories, respectively, when the combined group of underground and very low and ordinary low current configurations was used as the reference category. A similar analysis, which used the original classification but excluded the Albertan subjects (for whom wire coding was not repeated in the new study), resulted in similar but somewhat lower odds ratios. Repeating the same analysis with actual controls but using the new wire code classification from 2006, we observed odds ratios of 1.1 (95 percent CI: 0.7, 1.5) and 1.6 (95 percent CI: 1.0, 2.6) for ordinary high and very high current configurations, respectively. Using the ideal control group instead of the actual control group resulted in no change in the odds ratio of the ordinary high current configuration (OR = 1.1, 95 percent CI: 0.7, 1.5) but a slight attenuation of the odd ratio for the very high current configuration (OR = 1.3, 95 percent CI: 0.8, 2.1). For the Kaune-Savitz classification, we observed a similar pattern, although the risk increases in the high exposure category were smaller. For the Wertheimer-Leeper wire code classification, when all five categories were kept separate and the underground category was used as the reference category, we observed no clearly identifiable trend.
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
Our results showed that first-choice nonparticipant controls tended to be of lower socioeconomic status than their replacements (non-first-choice participant controls) and that lower socioeconomic status was also related to higher wire code categories, clearly indicating differential participation of controls. Similar tendencies in participation by socioeconomic status and an inverse relation between socioeconomic status and magnetic field exposure have been reported previously (4). Although the nature of the relation between various measures of socioeconomic status and childhood leukemia is far from settled (11), it is unlikely that this relation is a major source of confounding or that it significantly contributes to the development of selection bias. The overall distribution of wire code categories in our study, however, was not significantly different between the first-choice nonparticipant and the non-first-choice participant control groups. Nevertheless, we observed some excess of very high current configuration (19 vs. 12 from a total of 143) among the first-choice nonparticipant controls compared with their replacements (non-first-choice participant controls).
Because of a lower prevalence of the very high current configuration among replacement controls, some attenuation of the effect estimate was observed when we used the ideal control group instead of the actual control group in a case-control analysis. This attenuation, however, was observed only when the lower three categories were used as a combined reference group. We did not observe an attenuation of effect estimates when the underground category alone was used as the reference group.
Our results, overall, are suggestive of a small upward bias in the highest exposure category. This bias, however, is unlikely to explain entirely the observed association. Our results, taken together with the results of the two pooled analyses, which showed no systematic differences between the results of the measured field studies (where selection bias is possible) and the Nordic studies (where selection bias is not likely) (2, 3), weaken the argument that selection bias alone is responsible for the observed epidemiologic association between magnetic field exposure and childhood leukemia.
The small differences between the effect estimates when using the actual and ideal controls in our study may also be due to chance, especially since this observed difference was largely dependent on the choice of reference category. Because of the virtually indistinguishable distribution of measured fields in the lowest three wire code categories (7), the combination of these three categories may be the most appropriate reference group as also suggested by Greenland et al. (2). Our findings, nevertheless, appear to be consistent with previous results. Some evidence of similar selection bias has been shown in the United Kingdom Childhood Cancer Study, as well as in the National Cancer Institute and German studies (12–15).
Our study is the first to directly examine the effect of nonparticipation in a previously published electromagnetic field–childhood leukemia epidemiologic study. Other studies examined the effect of including partial participants in the analyses, but they were not able to include nonparticipants and could not evaluate the full extent of selection bias in the study.
Because of some limitations of our study, we cannot entirely dismiss the role of selection bias as a possible explanation for the observed association between measured fields and childhood leukemia on the basis of our results alone. Actual magnetic fields in the homes were not measured for nonparticipants, and wire codes are highly imperfect measures of measured fields. This exposure misclassification may attenuate or mask any real differences, if there are any, between the analyses using actual and ideal controls. The lack of information on measured fields also makes it impossible to assess any potential bias at the higher end of the exposure distribution. Although 12 percent of the study subjects were in the very high current configuration wire code category according to the new classification, in the original study only about 7 percent and 4 percent of the study subjects had average measured fields above 3 µT and 4 µT, respectively, above which exposure levels an association was observed in the two pooled analyses. If a potential bias is restricted to the higher fields only or the bias becomes progressively stronger with increasing exposure, then our current study may have underestimated the potential for bias. In addition, according to the two pooled analyses, the observed association was weaker in the Canadian study than in most of the other studies. Therefore, the potential for bias was also more limited in the Canadian study.
Apartment dwellers tended to have lower socioeconomic status and were also less likely to participate in the original study. The mean measured field in the bedroom of 11 measured apartments was 0.120 µT, which is close to the mean fields of the ordinary low wire code categories (0.111 µT). Exclusion of apartment dwellers (only about 10 percent of the study population) from the analyses, therefore, is unlikely to result in large over- or underestimation of the magnitude of the observed selection bias.
Overall, we conclude that, although there is some evidence for control selection or participation bias in the Canadian study, it is unlikely to explain entirely the observed association between magnetic field measures and childhood leukemia. Inherent problems in exposure assessment for nonparticipant subjects, however, limit the interpretations of these results.
|
ACKNOWLEDGMENTS |
|---|
The study was funded by the Electric Power Research Institute and Electricite de France.
Conflict of interest: none declared.
|
References |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTSDISCUSSIONReferences |
|---|
- IARC Working Group on the Evaluation of Carcinogenic Risks to Humans. Non-ionizing radiation, part 1: static and extremely low-frequency (ELF) electric and magnetic fields. IARC Monogr Eval Carcinog Risks Hum (2002) 80:1–395.[Medline]
- Greenland S, Sheppard AR, Kaune WT, et al. A pooled analysis of magnetic fields, wire codes, and childhood leukemia. Epidemiology (2000) 11:624–34.[CrossRef][Web of Science][Medline]
- Ahlbom A, Day N, Feychting M, et al. A pooled analysis of magnetic fields and childhood leukaemia. Br J Cancer (2000) 83:692–8.[CrossRef][Web of Science][Medline]
- Mezei G, Kheifets L. Selection bias and its implications for case-control studies: a case study of magnetic field exposure and childhood leukemia. Int J Epidemiol (2006) 35:397–406.
[Abstract/Free Full Text] - Hatch EE, Kleinerman RA, Linet MS, et al. Do confounding or selection factors of residential wiring codes and magnetic fields distort findings of electromagnetic fields studies? Epidemiology (2000) 11:189–98.[CrossRef][Web of Science][Medline]
- Linet MS, Hatch EE, Kleinermann RA, et al. Residential exposure to magnetic fields and acute lymphoblastic leukemia in children. N Engl J Med (1997) 337:1–7.
[Abstract/Free Full Text] - McBride ML, Gallagher RP, Theriault HG, et al. Power-frequency electric and magnetic fields and risk of childhood leukemia in Canada. Am J Epidemiol (1999) 149:831–42.
[Abstract/Free Full Text] - Cancer incidence in five continents. IARC Sci Publ (1997) Vol 7(143):i–xxxiv. 1–1240.
- Wertheimer N, Leeper E. Electrical wiring configurations and childhood cancer. Am J Epidemiol (1979) 109:273–84.
[Abstract/Free Full Text] - Kaune WT, Savitz DA. Simplification of the Wertheimer-Leeper wire code. Bioelectromagnetics (1994) 15:275–82.[CrossRef][Web of Science][Medline]
- Poole C, Greenland S, Luetters C, et al. Socioeconomic status and childhood leukemia: a review. Int J Epidemiol (2006) 35:370–84.
[Abstract/Free Full Text] - Exposure to power-frequency magnetic fields and the risk of childhood cancer. UK Childhood Cancer Study Investigators (UKCCS). Lancet (1999) 354:1925–31.[CrossRef][Web of Science][Medline]
- Childhood cancer and residential proximity to power lines. UK Childhood Cancer Study Investigators (UKCCS). Br J Cancer (2000) 82:1073–102.[CrossRef][Web of Science][Medline]
- Schuz J, Grigat JP, Brinkmann K, et al. Residential magnetic fields as a risk factor for childhood acute leukaemia: results from a German population-based case-control study. Int J Cancer (2001) 91:728–35.[CrossRef][Web of Science][Medline]
- Selection bias in epidemiologic studies of EMF and childhood leukemia. In: Proceedings of the 2001 Whistler workshop (2003) Palo Alto, CA: Electric Power Research Institute, Inc. (http://www.epriweb.com/public/RS_1008149.pdf).
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  What Happens to Your Manuscript: Characteristics of Papers Published in Volumes 167 and 168 Am. J. Epidemiol., May 15, 2009; 169(10): 1275 - 1276. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||