IJE Advance Access originally published online on November 22, 2005
International Journal of Epidemiology 2006 35(2):397-406; doi:10.1093/ije/dyi245
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Methodology |
Selection bias and its implications for case–control studies: a case study of magnetic field exposure and childhood leukaemia
1 Environment Department, Electric Power Research Institute, Palo Alto, CA, USA.
2 Department of Epidemiology, UCLA School of Public Health, Los Angeles, CA, USA.
* Corresponding author. Electric Power Research Institute, 3412 Hillview Avenue, Palo Alto, CA 94303, USA. E-mail: gmezei{at}epri.com
 |
Abstract |
|---|
 Top Abstract Magnetic field--childhood...Possible interpretations of the...ConclusionsReferences |
|---|
Based on the epidemiological association between residential exposure to extremely low frequency-magnetic fields (ELF-MF) and childhood leukaemia, the International Agency for Research on Cancer classified ELF-MF as a possible human carcinogen. Since clear supportive laboratory evidence is lacking and biophysical plausibility of carcinogenicity of MFs is questioned, a causal relationship between childhood leukaemia and magnetic field exposure is not established. Among the alternative explanations, selection bias in epidemiological studies of MFs seems to be the most plausible hypothesis. In reviewing the epidemiological literature on ELF-MF exposure and childhood leukaemia, we found evidence both for and against the existence of selection bias. To evaluate the potential for selection bias, we examined the relationship of socioeconomic status to subject participation and exposure to MFs. We find that, often, reporting of selection processes in itself is biased and incomplete, making the interpretation and evaluation of a potential for bias difficult. However, if present, such a bias would have wide implications for case–control studies in general. We call for better reporting and for evaluation of the potential for selection bias in all case–control studies, as well as, for the development of novel methods in control selection and recruitment.
Keywords Epidemiological methods, selection bias, childhood leukaemia, extremely low frequency-magnetic fields
Accepted 13 October 2005
|
Magnetic field—childhood leukaemia association |
|---|
|
TopAbstractMagnetic field--childhood...Possible interpretations of the...ConclusionsReferences |
|---|
Numerous epidemiological studies have examined the potential association between various measures of extremely low frequency (ELF)-magnetic field (MF) exposure and development of childhood leukaemia and other childhood cancers. All recent expert evaluations concluded that there is an association between childhood leukaemia development and exposure to ELF-MF. The National Institute of Environmental Health Sciences Working Group reported that there is limited evidence that residential exposure to ELF-MF is carcinogenic in children.1 The National Radiological Protection Board (NRPB) in the UK stated that relatively high average exposure to ELF-MF (0.4 µT or more) is associated with a doubling of the risk of childhood leukaemia.2 The International Agency for Research on Cancer (IARC) classified ELF-MF as a possible human carcinogen in June 2001.3 The International Commission for Non-Ionizing Radiation Protection (ICNIRP) Standing Committee on Epidemiology concluded that among all the health outcomes evaluated in epidemiological studies of ELF-MF, the strongest evidence for an association exists between childhood leukaemia and post-natal exposure to MFs > 0.4 µT.4
Expert reviews completed after 2000 were strongly influenced by the results of two pooled analyses of epidemiological studies of MFs and childhood leukaemia.5,6 One of the pooled analyses, by Greenland et al.,5 included original data from 15 epidemiological studies of MFs and childhood leukaemia. Of the included studies, 12 studies had data on measured or calculated MFs. Based on these 12 studies, there was no association between childhood leukaemia and MFs < 0.3 µT. However, the summary odds ratio for MF exposure >0.3 µT as compared with exposure <0.1 µT was 1.7 with 95% confidence interval (95% CI) at 1.2 and 2.3. Results from the individual studies were consistent with the pooled results. The association between childhood leukaemia and wire codes, based on eight studies, was less consistent across studies and the odds ratios for very high current configuration (VHCC) vs low current configuration (LCC) varied between 0.7 and 3.0. Based on four studies with both wire code and MF data, however, the summary effect estimate for VHCC vs LCC was elevated and it remained elevated and almost unchanged even after adjusting for MF levels [odds ratio = 1.6; 95% CI 1.2–2.3].
The other pooled analysis, by Ahlbom et al.,6 used more restrictive inclusion criteria, based on which they included four studies with calculated MFs and five studies with 24 or 48 h measured fields. There was no apparent association between MFs and childhood leukaemia below MF exposure level of 0.4 µT. However, the summary odds ratio for exposure >0.4 µT as compared with exposure <0.1 µT was 2.1, with 95% CI at 1.3 and 3.3. Results were very similar when calculated field studies and measured field studies were analysed separately. Sensitivity analyses showed that results were not dependent on any one study, however, the largest number of cases in the highest analysed MF exposure category was contributed by the study conducted by the National Cancer Institute in the US.7 In both pooled analyses, there was sufficient homogeneity of individual effect estimates to allow appropriate pooling of the results.
|
Possible interpretations of the association |
|---|
|
TopAbstractMagnetic field--childhood...Possible interpretations of the...ConclusionsReferences |
|---|
In spite of the consistent epidemiological findings, it remains uncertain as to whether or not this association is causal. Following the publication of the two pooled analyses, chance was largely discounted as a likely explanation for the observed epidemiological association. Publication bias is very unlikely to occur, especially in the pooled analysis by Ahlbom, which included only recent and high quality studies in a time period when this research field received much scrutiny and when all ongoing studies were known. A potential causal relationship between some physical characteristics of ELF-MF exposure and childhood leukaemia remains one of the possible explanations for the consistently found association between ELF-MF and childhood leukaemia in epidemiological studies. However, the lack of convincing experimental evidence in either cellular or animal studies showing that environmental ELF-MF exposure may affect biological processes and increase risk of cancer development has been frequently interpreted as a major argument against a causal explanation.1
Among potential alternative explanations, the role of confounding has been examined extensively.1,8,9 In spite of this extensive research, no single confounder or set of confounders has been identified that could explain the association. Although the lack of an identified confounder should not strengthen one's belief in causality, it has been repeatedly used as an argument for a causal relationship between ELF-MF and leukaemia. Recently, exposure to contact current has been hypothesized to explain the ELF-MF–childhood leukaemia association.10,11 Although some recent work has been focused on contact current exposure, more research needs to be completed before contact current exposure could be considered as the culprit for the epidemiological association between ELF-MF and childhood leukaemia.
Measurement error and exposure misclassification are also considered among the potential major sources of error in the ELF-MF–childhood leukaemia epidemiological studies. It is mostly agreed, however, that exposure misclassification is likely to be non-differential. Although the possibility of differential misclassification in selected studies has also been raised, the direction and magnitude of these types of errors remain speculative.12,13
Although there are numerous reviews of this literature, no attempt has been made to consistently examine the potential role of selection bias in the association between MF exposure and childhood leukaemia. In this paper we examine all relevant published literature and evaluate the potential magnitude and direction of bias. We also develop suggestions for further research that might elucidate this.
Potential role of selection bias in the EMF–childhood leukaemia association
Among the most likely non-causal explanations for the apparent epidemiological association between ELF-MF and childhood leukaemia, selection bias has been discussed repeatedly.2–4,14 Selection bias is a common and potentially serious problem of all case–control studies. Since most of the MF–childhood leukaemia epidemiological studies were case–control studies it has been proposed that selection bias may be fully or, at least, partially responsible for the described epidemiological association.
Selection bias has been characterized in many ways and epidemiology textbooks usually differentiate many types based on their mechanism and manifestation. Hernan et al.,15 however, recently showed that all types of selection bias have the same underlying causal ‘common effects’ structure. Hernan et al. argued that an apparent association will develop between an exposure (E) and an outcome (D) within strata of a common effect (C) of the exposure (or a cause of exposure) and the outcome (or a cause of outcome).
In a case–control study, factor C could also be interpreted as an indicator (perhaps an indirect one) of the subjects' selection/inclusion into the study. The probability of being selected is, by nature of the case–control studies, clearly dependent on the subjects' outcome status. For example, in a childhood leukaemia case–control study, cases of childhood leukaemia are much more likely to be included than healthy non-cases. In an epidemiological study of MFs, exposure may not directly determine the subjects' selection or inclusion probabilities, but the study subjects' socioeconomic status may influence both the subjects' selection or inclusion probability and their exposure status. In other words, in order for selection bias to occur through socioeconomic status, socioeconomic status needs to be related to subject participation and exposure to MFs, and participation needs to be differential by case–control status.
Hypothetical examples
In the following simple examples, we give an illustration of how differences in control inclusion probability and exposure prevalence across strata of a covariate (e.g. socioeconomic status) could distort the result of a hypothetical case–control study. Note, however, that our choices of parameters are not completely hypothetical, but are the best available estimates for the incidence of childhood leukaemia and prevalence of MF exposure.
Assumptions. Exposure is not causing the disease (i.e. the causal odds ratio is 1.0). Disease has an annual incidence of 0.00005 (5 per 100 000 per year). All incident cases occurring during the study period of 1 year could be included into the case–control study. Of the population, 20% has high (H20 population), 60% has intermediate (M60 population), and 20% has low (L20 population) socioeconomic status. Disease incidence is constant across all levels of socioeconomic status. In the M60 population, the prevalence of the exposure is 0.04 and the non-cases are included as controls in the study with a probability of 0.00005 (i.e. we select one control for each case).
In Example A (Figure 1), H20 population has the same exposure prevalence and control inclusion probability as M60 population. Figure 1 shows how changing control inclusion probability and exposure prevalence in the L20 population alone could distort the observed association. For example, if in the L20 population, the exposure prevalence is 0.12 (3 times the exposure prevalence in the M60 population), and the non-cases could be included as controls in the study with a probability of 0.00001 (0.2 times the probability of control inclusion in the M60 population) then the observed odds ratio is 1.29.
|
In Example B (Figure 2), L20 population has an exposure prevalence of 0.08 and the control inclusion probability is 0.00001. Figure 2 shows how changing exposure prevalence and control inclusion probability in the H20 population could result in biased odds ratios. For example, if in the H20 population, the exposure prevalence is 0.01 and the control inclusion probability is 0.0002 (4 times the probability of control inclusion in the M60) then the observed odds ratio is 1.75.
|
As one can see, realistic scenarios can easily result in biased effect estimates in the magnitude of 1.2–1.7. Relatively extreme selection bias needs to be present, however, to produce biased effect estimates in the order of 2, or beyond. In real life, only limited data are available from a small number of studies to assess these associations.
Case and control participation
In Figure 3 we present the hypothesized causal relationship between residential MF exposure, childhood leukaemia, socioeconomic status, residential mobility, and subject selection and participation in a case–control study. In the following, we discuss each association shown in Figure 3.
|
Low participation rate in some of the studies, in theory, provide the potential for control selection bias to occur. We plotted observed effect estimates for measured or calculated MF levels >0.3 µT against case and control participation rates (Figure 4). We observed no decrease in the magnitude of effect estimates with more complete case or control participation, which may argue against the role of selection bias. Low participation rate in case–control studies, however, does not necessarily indicate the presence of selection bias, and a study may be completely unbiased in spite of poor participation. Even if selection and/or response rates are different for cases and controls and/or for exposed and unexposed subjects, it does not necessarily indicate selection bias; selection bias develops only if the selection/inclusion probabilities are differential for cases and controls based on their exposure status; i.e. when the ratio of selection probabilities of exposed and unexposed cases is different from the ratio of the selection probabilities of exposed and unexposed controls. Low overall participation rate, however, may allow for a wider range of differential selection.
|
Subject participation rates varied widely in past ELF-MF–childhood leukaemia epidemiological studies (Table 1).7,16–30 The participation (or inclusion) rates in individual studies largely depended on the level of involvement the study required from its participants—each level of involvement usually further decreasing the number of participants in the studies.
|
The four studies from the Nordic countries—calculating MF exposure levels based on the subjects' addresses and, therefore, requiring no contact with the study subjects—included 94–100% of the eligible cases and the eligible and selected controls.18–20,22 The other—all case–control—studies requiring personal contact and at least the completion of an interview tended to achieve lower participation rates of all eligible subjects (participation in the interview phase of the studies could be as low as 68% among eligible cases and 37% among eligible controls).7,16,17,21,23–30 Also, as the level of invasiveness increased for exposure assessment (from wire coding to spot measurements and 24/48 h measurements) the participation rates tended to be lower for both cases and controls. In studies with matched analytical design, further reduction in the effective sample size could occur (to as low as 31 and 9% of all eligible cases and controls, respectively) owing to frequently incomplete measurement data on one of the members of the pair.23,24
Some of the selection and self-selection processes are reported in some epidemiological studies, therefore, the numbers of participants and participation rates at various levels could be tracked. However, owing to incomplete reporting, tracking of the selection is difficult in many studies. In addition to observable processes of selection and participation, there is also an inherent, unobservable or not always clearly reported selection process when the eligibility criteria and sampling frame for the study are determined. In most studies, there are no reliable estimates on these unobservable selection rates. For example, screening of phone calls makes identification of the entire eligible pool virtually impossible with random digit dialling (RDD). An additional problem arises from the inconsistent ways studies describe their eligibility criteria, selection procedures and participation rates, making it difficult or impossible to compare participation rates across various studies.31 In addition, in most instances only the marginal rates of participation could be observed (i.e. for cases and controls, or for exposed and unexposed subjects) but not the joint participation rates (i.e. for exposed cases or unexposed cases). Therefore, it is frequently impossible to determine without additional information from outside the study whether the selection process resulted in a biased effect estimate in the study.
Although selection bias most frequently refers to control selection bias (i.e. biased selection of controls), case selection bias (i.e. biased selection of cases) may also be a potential problem if non-participation occurs among cases, as well. For example, cases in the Linet et al. study were ascertained through treatment study groups that covered more than half of the cases occurring in the US. This could lead to bias if there were systematic differences between children included in these treatment study groups and children who were not. Potentially more problematic is the inclusion of only childhood leukaemia cases who were in remission in the Canadian studies where personal monitoring was used to assess exposure. In all of theses studies, however, case participation was higher than that of controls.
Participation and socioeconomic status
Table 2 shows how overall participation in ELF-MF epidemiological studies was related to the subjects' socioeconomic status. Spinelli et al.32 analysed data from the 1999 Canadian study of ELF-MF and childhood leukaemia by McBride et al.,27 and found that participating controls tended to live in census tracts with higher average income than did non-participant controls.33 In the UK study, a marked under-representation of children with lower socioeconomic status can be observed among the participating ELF-MF study controls in comparison with the first-choice controls.28,29
|
Hatch et al.8 reanalysed data from the Linet et al. 1997 study7 and observed that, compared with partial participants (i.e. subjects who answered the questionnaires and allowed measurements at the front door but did not agree to have measurements done in their homes), full participants (i.e. subjects with both questionnaire and in-home measurement data) tended to have higher socioeconomic status (higher education, higher income, and more likely to own residence). The study had, unfortunately, no information on non-participants. In a similar analysis of the German study, Michaelis et al. showed that those who fully participated in the German ELF-MF epidemiological study had, on the average, higher socioeconomic status (e.g. higher family income) and were more likely to live in urban settings than partial participants (i.e. subjects who participated in the interview phase of the study but who did not allow in-home MF measurements).33
Although children with higher socioeconomic status were thought to have somewhat higher risk of childhood leukaemia, a recent careful examination of the literature showed that there is no consistent association between socioeconomic status and childhood leukaemia.34,35 In most ELF-MF epidemiological studies, however, cases tended to have lower socioeconomic status indices than controls (Table 3).33 If socioeconomic status does not affect the risk of childhood leukaemia then the socioeconomic status disparities between cases and controls may probably indicate differences in participation between cases and controls based on their socioeconomic status.
|
Participation and MF exposure
No study was able to determine exposure among non-participants. A few studies compared, however, exposure distribution between subjects who partially participated to those who fully participated in a study. Hatch et al.8 found that full participants were less likely to live in homes with high wire codes (VHCC) or high measured fields (measurements at front door >0.2 µT). Savitz et al.,16 in his 1988 study, compared subjects with and without MF measurements among those with wire code classification. They found that subjects without field measurement were more likely to live in homes with higher wire codes.
Socioeconomic status and MF exposure
There is also limited data available on a potential association between socioeconomic status and MF exposure. Gurney et al.36 assessed the relationship between family income and wire codes and found that lower family income tended to be associated with higher wire codes. Hatch et al. 8 presented data on the association between some socioeconomic measures and wire codes and measured MFs. Subjects with lower annual family income and those who rented (not owned) their homes were more likely to live in homes with the highest exposure categories (VHCC or measured MF > 0.2 µT). An analysis of the German study also showed that lower income tended to be associated with higher MF exposure.33
Residential mobility
It was also hypothesized that selection bias may be occurring through differential participation of study subjects based on residential mobility, similarly to socioeconomic status. Some evidence supporting this hypothesis could be found in the literature. In several studies, cases tended to be more residentially mobile than controls even prior to diagnosis.21,25–27 At the same time, Jones et al.37 reported that people who changed addresses more frequently (high residential mobility) were more likely to live at an address with higher wire codes. Note, however, that the study might have overestimated mobility, because it was not based on families with children and included in and out mobility, as well. Reanalysis of the German study also showed that partial participants (who gave interviews but did not allow measurements) were more residentially mobile than subjects who fully participated in the study.33
Examination of selection bias in ELF-MF–childhood leukaemia studies
A limited number of studies tried to address directly the issue of selection bias in ELF-MF–childhood leukaemia studies and attempted to identify the direction of bias and to quantify the potential magnitude of bias from this source. Gurney et al.36 estimated that differential participation of cases and controls by their income status could result in an upward bias of the high wire code—childhood leukaemia association in a case–control study; the odds ratio would be inflated by 1.03 to 1.24-fold. Using GIS-based automatic wire coding of residences in the Denver area—the study area of the Savitz et al.16 study—Ebi et al.38 also examined the possible role of selection bias in the Savitz et al. study. They found that when random samples of hypothetically eligible residences in the study area were included as controls, the effect estimate for HCC vs LCC comparison was lower (odds ratio = 1.32) than when the original study controls were used (odds ratio = 1.55).
The most comprehensive evaluation of the possible role of selection bias in an existing study was an already mentioned analysis by Hatch et al.8 They demonstrated that the association between measured MF exposure >0.3 µT and childhood leukaemia development was stronger when only full participants were included in the analysis (odds ratio = 1.9) as compared with an analysis that included all subjects (partial and full participants) (odds ratio = 1.6). Similar reduction was also seen in the effect estimates for wire codes. This was, however, highly limited by the fact that the study assessed what one would have to suspect as being only the tip of the selection bias iceberg. Every single person in Hatch's analysis had already been successfully identified, selected, contacted, and had at least some field strengths measured. Thus the distinction between partial and full participation would be slight.
Most recently in a simulation study using the German data, Schuz and Feychting39 estimated the likely magnitude of selection bias by imputing MF exposure levels for non-participants based on characteristics of their residence. They found that inclusion of non-participants in the analysis resulted in a substantial decrease in the observed relative risk estimates (from 2.8 to 1.6).
The results of these studies support the argument that selection bias is, in fact, likely to occur in ELF-MF–childhood leukaemia epidemiological studies, and this bias is likely to result in bias away from the null, i.e. in overestimation of the potential effect of ELF-MF on childhood leukaemia incidence.
The pooled analysis by Ahlbom et al.,6 provides a strong argument against the role of control selection bias as the explanation for the association between childhood leukaemia and MFs. The Nordic studies, requiring no direct contact with study subjects, achieved high participation rate, and were less prone to selection bias than other case–control studies. If the entire association was due to selection bias, we would expect an association only among those studies with the potential for selection bias and not in the Nordic studies where selection bias is unlikely. In fact, the pooled analysis showed similar risk increases in the Nordic studies as in the rest of the studies. Note, however, that of the 44 case children with average exposure >0.4 µT in the pooled analysis, only eight were from the Nordic studies and five of these eight children came from one single study from Sweden.6 Also arguing against selection bias is the fact that adjustment for socioeconomic status in both pooled analyses had little effect on the relative risk estimates.5,6
Another argument against selection bias is that there is an apparent lack of a consistent association in studies of childhood brain tumours and residential MFs.40 Many of the leukaemia studies included in the pooled analysis examined brain tumours as well and there is no reason to think that selection bias will affect one outcome and not the other. Note, however, that the conclusion regarding childhood brain tumours is tentative, as brain tumour studies have generally been smaller and fewer and a pooled analysis of brain tumour studies is yet to be conducted.
|
Conclusions |
|---|
|
TopAbstractMagnetic field--childhood...Possible interpretations of the...ConclusionsReferences |
|---|
Owing to low participation among cases and controls in past studies, the potential for selection bias is large in all studies of ELF-MF exposure and childhood leukaemia, except for the ones that were based on existing records and thus did not require subject participation. This seems to be particularly true for studies that used RDD and also for studies that used medical rolls or birth certificates.
In reviewing the epidemiological literature on ELF-MF exposure and childhood leukaemia, we found evidence both for and against the existence of selection bias. The available evidence that may indicate that control selection bias is a potential explanation for the epidemiological association between ELF-MF and childhood leukaemia is yet rudimentary. Most of the evidence relates to wire codes, which are imperfect measures of MFs, or they examine only partial participants and not non-participants. The biggest challenge will be to establish how the distribution of measured MFs among non-participants differs from the distribution among participants. To understand the exact role of selection bias in the association between MF and childhood leukaemia, it will also be important to determine how socioeconomic status relates to the exposure of interest and to subject selection and participation.
Even if present, selection bias may not necessarily explain the entire association between MF exposure and childhood leukaemia. If, however, control selection bias proved to be the explanatory factor for the MF—childhood leukaemia association, this would have tremendous implications for epidemiology in general, and especially for case–control studies of any environmental exposures or other factors related to socioeconomic status and/or residential mobility. It is particularly true for studies examining weak associations. For rare diseases, such as cancers, case–control design is now the accepted and most common approach. A large percentage of these studies might be also susceptible to similar selection bias.
Future case–control studies should pay particular attention to the entire process of subject (case and control) selection, inclusion, and participation. Attempt should be made by the authors to identify and report every step along the course of subject selection. It should include the thorough and exact identification (with all the inclusion and exclusion criteria) of the base population, the determination of how the actual sampling pool relates to the base population, and the recording of all levels of selection, inclusion, participation, and non-participation. Reporting should not be limited to a statement, such as, ‘X (usually high) percentage of the subjects participated in the study’, which typically refers to the very last step of the subject inclusion process. Investigators should also collect the maximum amount of information possible regarding all non-participant and partial-participant subjects, and conduct simulations to evaluate the magnitude and direction of any potential bias introduced by the subject selection process.
The difficulties of proper control selection and case recruitment are likely to get only worse in the future. Changing society and technology, such as telemarketing, call screening, and mobile phones will make RDD obsolete.41 However, the advent of an Internet might provide opportunities for new study designs and subject recruitment and engagement techniques.
KEY MESSAGES
|
|
Acknowledgments |
|---|
This work was funded by the Electric Power Research Institute.
|
References |
|---|
|
TopAbstractMagnetic field--childhood...Possible interpretations of the...ConclusionsReferences |
|---|
1 Portier CJ, Wolfe MS. Assessment of health effects form exposure to power-line frequency electric and magnetic fields. NIEHS Working Group Report. NIH Publication No. 98–3981. Research Triangle Park, NC: NIEHS, 1998.
2 NRPB. ELF Electromagnetic fields and the risk of cancer. Documents of the NRPB. Vol. 12 (No. 1). Chilton, UK; National Radiological Protection Board, 2001.
3 IARC. Static and extremely low-frequency electric and magnetic fields. IARC Monographs on the evaluation of the carcinogenic risks to humans. Vol. 80. Lyon, France: International Agency for Research on Cancer, 2002.
4 Ahlbom A, Cardis E, Green A, Linet M, Savitz D, Swerdlow A. Review of the epidemiologic literature on EMF and health. Environ Health Perspect 2001;109 (Suppl 6):911–33.
5 Greenland S, Sheppard AR, Kaune WT, Poole C, Kelsh MA. A pooled analysis of magnetic fields, wire codes, and childhood leukemia. Epidemiology 2000;11:624–34.[CrossRef][ISI][Medline]
6 Ahlbom A, Day N, Feychting M et al. A pooled analysis of magnetic fields and childhood leukaemia. Br J Cancer 2000;83:692–98.[CrossRef][ISI][Medline]
7 Linet MS, Hatch EE, Kleinermann RA et al. Residential exposure to magnetic fields and acute lymphoblastic leukemia in children. N Eng J Med 1997;337:1–7.
8 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][ISI][Medline]
9 Langholz B. Factors that explain the power line configuration wiring code-childhood leukemia association: what would they look like? Bioelectromagnetics 2001;(Suppl. 5):S19–31.
10 Kavet R, Zaffanella LE, Pearson RL, Dallapiazza J. Association of residential magnetic fields with contact voltage. Bioelectromagnetics 2004;25:530.[Medline]
11 Kavet R, Zaffanella LE. Contact voltage measured in residences. Implications to the association between magnetic fields and childhood leukemia. Bioelectromagnetics 2002;23:464–74.[CrossRef][ISI][Medline]
12 Jaffa KC, Kim H, Aldrich TE. The relative merits of contemporary measurements and historical calculated fields in the Swedish childhood cancer study. Epidemiology 2000;11:353–56.[Medline]
13 Mezei G, Kheifets L. Is there any evidence for differential misclassification or for bias away from the null in the Swedish childhood cancer study? Epidemiology 2001;12:750–52. (Erratum in: Epidemiology 2002;13:116.)[Medline]
14 Wartenberg D. The potential impact of bias in studies of residential exposure to magnetic fields and childhood leukemia. Bioelectromagnetics 2001;(Suppl. 5):S32–47.
15 Hernan MA, Hernandez-Diaz S, Robins JM. A structural approach to selection bias. Epidemiology 2004;15:615–25.[CrossRef][ISI][Medline]
16 Savitz D, Wachtel H, Barnes FA, John EM, Tvrdik JG. Case–control study of childhood cancer and exposure to 60-Hz magnetic fields. Am J Epidemiol 1988;128:21–38.
17 London SJ, Thomas DC, Bowman JD, Sobel E, Cheng T-C, Peters JM. Exposure to residential electric and magnetic fields and risk of childhood leukemia. Am J Epidemiol 1991;134:923–37.
18 Feychting M, Ahlbom A. Magnetic fields and cancer in children residing near Swedish high-voltage power lines. Am J Epidemiol 1993;138:467–81.
19 Olsen JH, Nielsen A, Schulgen G. Residence near high voltage facilities and risk of cancer in children. BMJ 1993;307:891–95.[ISI][Medline]
20 Verkasalo PK, Pukkala E, Hongisto MY et al. Risk of cancer in Finnish children living close to power lines. BMJ 1993;307:895–99.[ISI][Medline]
21 Michaelis J, Schüz J, Meinert R et al. Childhood leukemia and electromagnetic fields: Results of a population-based case–control study in Germany. Cancer Causes Control 1997;8:167–74.[CrossRef][ISI][Medline]
22 Tynes T, Haldorsen T. Electromagnetic fields and cancer in children residing near Norwegian high-voltage power lines. Am J Epidemiol 1997;145:219–26.
23 Dockerty JD, Elwood JM, Skegg DC, Herbison GP. Electromagnetic field exposures and childhood cancers in New Zealand. Cancer Causes Control 1998;9:299–309. (Erratum in: Cancer Causes Control 1999;10:641.)[CrossRef][ISI][Medline]
24 Dockerty JD, Elwood JM, Skegg DC, Herbison GP. Electromagnetic field exposures and childhood leukaemia in New Zealand. Lancet 1999;354:1967–68.[CrossRef][Medline]
25 Green LM, Miller AB, Agnew DA et al. Childhood leukemia and personal monitoring of residential exposures to electric and magnetic fields in Ontario, Canada. Cancer Causes Control 1999;10:233–43.[CrossRef][ISI][Medline]
26 Green LM, Miller AB, Villeneuve PJ et al. A case–control study of childhood leukemia in southern Ontario, Canada, and exposure to magnetic fields in residences. Int J Cancer 1999;82:161–70.[CrossRef][ISI][Medline]
27 McBride ML, Gallagher RP, Thériault HG et al. Power-frequency electric and magnetic fields and risk of childhood leukemia in Canada. Am J Epidemiol 1999;149:831–42.
28 UK Childhood Cancer Study Investigators (UKCCS). Exposure to power-frequency magnetic fields and the risk of childhood cancer. Lancet 1999;354:1925–31.[CrossRef][ISI][Medline]
29 UK Childhood Cancer Study Investigators (UKCCS). Childhood cancer and residential proximity to power lines. Br J Cancer 2000;82:1073–102.[CrossRef][ISI][Medline]
30 Schuz J, Grigat JP, Brinkmann K, Michaelis J. 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][ISI][Medline]
31 Olson SH, Voigt LF, Begg CB, Weiss NS. Reporting participation in case–control studies. Epidemiology 2002;13:123–26.[CrossRef][ISI][Medline]
32 Spinelli JJ, McBride ML, Abanto ZU, Tamaro S, Bajdik CD, Gallagher RP. Assessing response bias as an explanation for the observation of low socioeconomic status (SES) as a risk factor for childhood leukemia. (Abstract) Am J Epidemol 2001;153:S254 (948).
33 EPRI. Selection bias in epidemiologic studies of EMF and childhood leukemia. Proceedings of the 2001 Whistler workshop. Technical Report. Palo Alto, CA: Electric Power Research Institute, 2003.
34 Poole C, Greenland S, Luetters C, Kelsey JL, Mezei G. Socioeconomic status and childhood leukaemia: a review. Int J Epidemiol 2006;35:370–84.
35 Greenland S, Luetters C, Kelsey JL, Mezei G. Is socioeconomic status related to childhood leukemia incidence? (Abstract) Am J Epidemiol 2004;159:S95 (378).
36 Gurney JG, Davis S, Schwartz SM, Mueller BA, Kaune WT, Stevens RG. Childhood cancer occurrence in relation to power line configurations: a study of potential selection bias in case–control studies. Epidemiology 1995;6:31–35.[ISI][Medline]
37 Jones TL, Shih CH, Thurston DH, Ware BJ, Cole P. Selection bias from differential residential mobility as an explanation for associations of wire codes with childhood cancer. J Clin Epidemiol 1993;46:545–48.[Medline]
38 Ebi KL, Kheifets LI, Pearson RL, Wachtel H. Description of a new computer wire coding method and its application to evaluate potential control selection bias in the Savitz et al. childhood cancer study. Bioelectromagnetics 2000;21:346–53.[Medline]
39 Schuz J, Feychting M. An evaluation of potential biases in epidemiological studies on childhood leukemia and exposure to extremely low-frequency magnetic fields. Abstract P-A-73. Proceedings of the Annual Meeting of the Bioelectromagnetics Society, Dublin, Ireland, 2005.
40 Kheifets LI. Electric and magnetic field exposure and brain cancer: a review. Bioelectromagnetics 2001;(Suppl. 5):S120–131.
41 Ross JA, Spector LG, Olshan AF, Bunin GR. Invited commentary: Birth certificates—A best control scenario? Am J Epidemiol 2004;159:922–24.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  G. Mezei, J. J. Spinelli, P. Wong, M. Borugian, and M. L. McBride Assessment of Selection Bias in the Canadian Case-Control Study of Residential Magnetic Field Exposure and Childhood Leukemia Am. J. Epidemiol., June 15, 2008; 167(12): 1504 - 1510. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. Geneletti, S. Richardson, and N. Best Adjusting for selection bias in retrospective, case-control studies Biostat., May 14, 2008; (2008) kxn010v1. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
F. Mensah, E. Willett, J Simpson, A. Smith, and E Roman Birth Order and Sibship Size: Evaluation of the Role of Selection Bias in a Case-Control Study of Non-Hodgkin's Lymphoma Am. J. Epidemiol., September 15, 2007; 166(6): 717 - 723. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. C Brown, T. Dwyer, C. Kasten, D. Krotoski, Z. Li, M. S Linet, J. Olsen, P. Scheidt, D. M Winn, and for the International Childhood Cancer Cohort Cons Cohort Profile: The International Childhood Cancer Cohort Consortium (I4C) Int. J. Epidemiol., August 1, 2007; 36(4): 724 - 730. [Full Text] [PDF]
|
|
|
|
|
|
J. Schuz, A. L. Svendsen, M. S. Linet, M. L. McBride, E. Roman, M. Feychting, L. Kheifets, T. Lightfoot, G. Mezei, J. Simpson, et al. Nighttime Exposure to Electromagnetic Fields and Childhood Leukemia: An Extended Pooled Analysis Am. J. Epidemiol., August 1, 2007; 166(3): 263 - 269. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||