IJE Advance Access originally published online on November 24, 2005
International Journal of Epidemiology 2006 35(2):370-384; doi:10.1093/ije/dyi248
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Article |
Socioeconomic status and childhood leukaemia: a review
1 Department of Epidemiology, University of North Carolina School of Public Health, Chapel Hill, NC, USA
2 Department of Epidemiology, University of California, Los Angeles, CA, USA
3 Department of Statistics, University of California, Los Angeles, CA, USA
4 Division of Geriatrics, School of Medicine, University of California, Los Angeles, CA, USA
5 Division of Epidemiology, Department of Health Research and Policy, Stanford University School of Medicine, Stanford, CA, USA
6 Division of Preventive and Behavioral Medicine, University of Massachusetts Medical School, Worcester, MA, USA
7 Environment Department, Electric Power Research Institute, Palo Alto, CA, USA
* Corresponding author. Department of Epidemiology (CB 7435), University of North Carolina School of Public Health, Chapel Hill, NC 27599–7435, USA. E-mail: cpoole{at}unc.edu
 |
Abstract |
|---|
 Top Abstract MethodsResultsDiscussionConclusionsReferences |
|---|
Background A long-held view links higher socioeconomic status (SES) to higher rates of childhood leukaemia. Some recent studies exhibit associations in the opposite direction.
Methods We reviewed journal literature through August 2002 for associations between childhood leukaemia and socioeconomic measures. We determined the direction of each association and its P-value. We described the results with regard to study design, calendar period, geographic locale, and level of the socioeconomic measures (individual or ecological). For measures with sufficient number of results, we computed summary P-values across studies.
Results Case–control studies conducted in North America since 1980 have involved subject interviews or self-administered questionnaires and have consistently reported inverse (negative) associations of childhood leukaemia with individual-level measures of family income, mother's education, and father's education. In contrast, associations have been consistently positive with father's occupational class in record-based case–control studies and with average occupational class in ecological studies.
Conclusions Connections of SES measures to childhood leukaemia are likely to vary with place and time. Validation studies are needed to estimate SES-related selection and participation in case–control studies. Because different socioeconomic measures (such as income and education) and individual-level and ecological-level measures may represent different risk factors, we advise researchers to report these measures separately rather than in summary indices of social class.
Keywords Education, income, socioeconomic status, leukaemia
Accepted 14 October 2005
In a 1985 review of the epidemiology of childhood cancers, Greenberg and Shuster classified five of the six studies published through 1982 as showing increased childhood leukaemia risk in association with high socioeconomic status (SES).1 The view has persisted that SES and childhood leukaemia are positively related. A 1999 National Cancer Institute publication listed high SES as a ‘known risk factor’ for acute lymphoblastic leukaemia (ALL),2 the most common type of leukaemia in children; it cited three previous reviews,3–5 but those reviews cited only each other, Greenberg and Shuster,1 and the studies Greenberg and Shuster cited. Many studies published since 1982 have reported associations in the opposite (inverse or negative) direction,6 with lower rates associated with higher levels of SES.
It would be difficult to compare these studies quantitatively because of the vast differences in SES measures used and in the social implications of measures (e.g. years of education) over places and times. Nonetheless, some directional patterns might be expected. These observations prompted us to undertake a review of the literature on this topic, to see if studies of childhood leukaemia and SES display any directional patterns in relation to geographic locale, calendar time, study design features, and type of SES measure.
|
Methods |
|---|
|
TopAbstractMethodsResultsDiscussionConclusionsReferences |
|---|
Literature search
Three databases, PubMed, PsycInfo, and ERIC, were searched through August 31, 2002 for original research in English, using the keywords ‘childhood’, ‘leukemia’, and ‘leukaemia’. PubMed was searched from January 1, 1965; PsycInfo from January 1, 1960; and ERIC from January 1, 1966. The PsycInfo and ERIC searches returned few citations, so those searches were broadened to ‘childhood’ and ‘cancer’. A manual search of Index Medicus from 1945 through 1964 was conducted as well, using the keywords ‘leukemia’ and ‘leukaemia’.
Case reports, case series, articles on diagnostic or prognostic characteristics or treatment outcomes, letters, editorials, and news briefs were excluded. One author (C.L.) reviewed the remaining manuscripts. Additional relevant reports referenced in these articles were also collected.
Inclusion criteria
To be included, a study had to compare leukaemia incidence or mortality for children or young adults (ages 0–24 years), and had to provide enough data on an SES measure to determine at least the direction of the association and the P-value. In many studies, disparate types of cancer were combined, the most extreme combination being total childhood cancer, which we excluded. In the few studies for which SES results were separated by leukaemia subtype, we combined all leukaemias for our main analyses, but also compared results by subtype within those studies to check comparability. We also included the few studies that combined leukaemias with non-Hodgkins lymphoma (NHL); we did this because the International Classification of Diseases had only one code for the leukaemias and NHL prior to 1957, and in childhood NHL incidence is less than a third of that of leukaemias. Several included studies were restricted to acute lymphocytic (lymphoblastic) leukaemia (ALL, which composes roughly three-quarters of all childhood leukaemia), and a few studies were restricted to acute non-lymphocytic leukaemias (ANLs).
Measures of SES that we considered were family income, parental education (separately for mothers and fathers), paternal occupational class (professional, manual, etc.), paternal employment status (employed or unemployed), household density (persons per room), other selected home characteristics (owned or rented, rental cost, apartment or single family, etc.), and clinic type (public or private). Degree of urbanization (urban, rural, suburban, etc.), residential mobility (e.g. the number of homes occupied by a childhood leukaemia patient between birth and diagnosis), and household size were not considered SES measures. Race and ethnicity were considered beyond the scope of the present review.
Abstraction
From the collected reports, one author (C.L.) abstracted the following information onto a standardized form: study design, study location, dates of leukaemia diagnosis or death, leukaemia type (if given), case ascertainment and inclusion criteria, control or cohort ascertainment and inclusion criteria, primary exposure of interest, descriptions of SES variables, distributions of SES variables and leukaemia, and estimated associations between SES variables and leukaemia. For studies with multiple reports, we chose the most recent or comprehensive one as the primary information source, referring to the others as needed for supplemental information. The abstracted information was summarized in tables, which were checked against the original articles by two other authors (S.G. and C.P.) before and during analysis.
We recorded whether each SES variable was measured at the individual level (e.g. family income of each subject) or at the ecological level (e.g. median household income for the census tract of each subject) and whether the analyses were conducted at the individual or ecological level. We also noted whether interviews or questionnaires were required of parents (not whether they were the source of the SES information) or instead records alone were used; the aim was not to subclassify quality of SES ascertainment (which is limited and unvalidated regardless of source) but rather to classify according to potential for refusal bias based on need for active parental participation. For the case–control studies we further noted the source of controls, using the terms ‘registry’ for control selection from an official population database (mostly birth registers but in a few cases government insurance or supplement rolls) and ‘area’ for selection by direct area survey methods (such as neighbourhood walking).
Analysis methods
Because childhood leukaemia rates are very low, we ignored the exclusion or inclusion of cases from the control sampling frames in the case–control studies and distinctions among different types of relative risks (RRs). All studies involved incidence.
Some studies did not give RR estimates, while others gave multiple estimates with varying degrees of adjustment or for varying subsets of subjects. When indicated and permitted by the data presented in an article, we computed our own estimates using methods previously described (in Ref.7, Ch. 32). Among the indications was the need to produce a summary estimate across age and sex groups or the need to use correct analytical methods when published estimates were produced by inappropriate methods (e.g. using the entire study population to generate expected numbers of cases for subsets of that population).8 Household density and some scales of occupational class increase as SES declines. These measures were rescaled to make increasing values correspond to increasing SES.
In view of the pronounced heterogeneity among the SES measures used, we did not attempt to compare or combine the quantitative values of the estimated RRs or of the RR trends from the different studies. Even within a given SES measure, there were strong grounds for doubting comparability from study to study. Twelve years of maternal education or $20 000 of annual family income could not reasonably be expected to mean the same thing for childhood leukaemia patients diagnosed in 1965 and 1995, even in the same locale and after adjustment for inflation. Inter-regional and international variability made comparability of SES measures across studies even more suspect. Hence, we limited our formal meta-analyses to non-parametric comparisons of study results.
We determined whether each RR estimate or trend was positive (higher rates in upper SES levels) or negative (higher rates in lower SES levels) and analysed the prevailing direction of association and the P-values for the studies. For studies employing binary SES scales, these determinations were usually made by noting whether the estimated RR was greater than or less than the null value of 1.0 and by inferring the P-value from the 95% confidence interval (in Ref.7, Ch. 32).
For SES scales with more than two levels or categories, we looked for trend analyses in the publications. In the absence of reported trend estimates (usually only the ‘P for trend’ was reported), we estimated the trends from the categorical RR estimates and their confidence intervals.9 These analyses required the assignment of a score to each category of each SES measure. Because the score assignments were uncertain, we conducted sensitivity analyses with different sets of scores. The sets differed in the distance between scores for adjacent categories: evenly spaced scores, scores farther apart at the lower end of the scale than at the higher end, and scores farther apart at the higher end than at the lower end. Supplemental tables, available from IJE online, give details of these computations.
Because of our strong interest in the evolution over time of associations between childhood leukaemia rates and SES measures, we list the results in each table in ascending order of the study's start date (i.e. the first year of the calendar period during which cases were ascertained). We also list selected study characteristics of importance, particularly geographic locale and study-design features.
Within each grouping of studies, we conducted a global test for associations based on the study P-values. Let Ps be the upper one-sided P-value from study s. We used the approximate 
2 statistic 
S 
–1 (PS)2, where –1(PS) is the 100Ps percentile of the standard normal distribution (i.e. the quantile function); e.g. –1(0.025) = –1.96. Because –1(1 – PS)2 = –1 (PS)2 the same result is obtained using the lower P-value. This test is non-specific in that it does not assume homogeneity of direction among the study results and has some sensitivity (power) for any pattern of departure from the null (including those induced by publication bias).
We also used ‘vote-counting’ 10,11 methods, which employ an exact binomial test (in Ref.7, Ch. 14) of the proportion of P-values at or below a specified level 
. If there were no association underlying any study and no publication bias, we should expect the proportion of studies with P 
to be . Using one-sided P-values, the resulting test of the observed proportion with P is sensitive only to departures from the null in that direction. For counts of positive or negative associations, = 0.5, and the test is the ordinary sign test. For separate counts of positive and negative associations having a two-sided P 0.05, one uses = 0.025 with the one-sided P-values (in Ref.7, Ch. 12); this cut-off reflects the reporting practices of some studies. These tests are less powerful (less sensitive) than the 2 test for detecting the presence of associations but are more specific in that they test only directional tendencies. To compare results from different groups of studies, we used Fisher's exact test (in Ref.7, Ch. 14) for a between-group difference in the proportion of positive associations.
We computed these statistics only for counts of all available results for a given SES measure and not across measures or subsets suggested by a visual search for patterns in the results. We further confined the formal analyses to individual-level measures of family income, mother's education, father's education, and father's occupational class. For other SES measures, there were too few studies or the scales were too non-comparable to warrant the use of summary methods.
|
Results |
|---|
|
TopAbstractMethodsResultsDiscussionConclusionsReferences |
|---|
Search
The PubMed search returned 5452 citations. The broadened searches of the other two databases returned an additional 269 citations from PsycInfo and 34 from ERIC. The Index Medicus search added 11 citations. After the initial exclusions (case reports, etc.), 595 manuscripts remained for review resulting in 116 reports representing 73 distinct studies after the remaining inclusion criteria were applied.12–127 Upon the second examination, 11 papers lacked information to determine the direction of SES–leukaemia associations, their P-values, or both.16,24,58,59,71,79,82,90,113,115–117 Another 17 papers reported results only for urbanization,35,43,60 residential mobility,14,15,21,52,77,78 race, or ethnicity.23,29,37–39,62,63,105,123 These exclusions reduced the number of distinct studies in the review to those listed in Table 1; because of multiple publications from single studies, several appear in later tables under other primary authors.
|
Under the Design column, Table 1 shows basic subclassifications of the case–control (CC) studies by source of controls and whether interviews or questionnaires were required of subjects All cohort and ecological studies were based on records alone. In no analysis did the summary tests change meaningfully when restriction was made to studies using only registry controls.
Family income
All the individual-level measures of annual or monthly family income (Table 2) were case–control studies. Most were from North America, all but one included interviews or a questionnaire, and most started in the 1980s. Associations were present (2 test, P < 0.001), with 8 of the 11 associations negative (sign test P = 0.1) and 5 negative with P 0.05 (vote-count P < 0.0001). The three positive associations were from the three earliest studies, and all eight studies starting in 1980 or later produced negative associations.
|
Ten income scales had more than two categories, thus, requiring a sensitivity analysis of category score assignments. With every set of alternative scores, the number of negative associations increased from eight (sign test P = 0.1) to nine (sign test P = 0.03) and the number of negative associations with P
0.05 increased from five to six (vote-count P < 0.0001).
Mother's education
All studies that examined mother's education (Table 3) were case–control studies. About half the studies were in North America, all but one included interviews or a questionnaire, about half started in the 1980s, and one-third of the studies were in the 1990s. The education measures were about equally divided between years of school and attained educational level. Associations were present (2 test, P < 0.001) with 11 of the 18 associations negative (sign test P = 0.2), 7 of the negative associations having P 0.05 (vote count P < 0.0001), and one of the positive associations having P 0.05 (vote count P = 0.07). The positive associations were not confined to the earliest studies; however, of the nine North American associations, only the earliest was positive.
|
Thirteen of the maternal education scales had more than two categories. In the sensitivity analysis of score assignments, the number of negative associations ranged from 11 (sign-test P = 0.2) to 12 (sign-test P = 0.1), the number of negative associations with P
0.05 ranged from seven to eight (both vote-count P < 0.0001), and the number of positive associations with P 0.05 remained constant at two (vote-count P = 0.07).
Father's education
All 13 of the studies with results for father's education (Table 4) were case–control studies, and all but one included interviews or a questionnaire, As with mother's education, the scales were about equally divided between years of school and level of educational attainment. Associations were present (2 test, P < 0.001) with 11 of the 13 associations negative (sign test P = 0.01) and four associations negative with P 0.05 (vote count P = 0.0002). Nine of the studies used scales of father's education with more than two categories. In the sensitivity analyses of category score assignments, the only change was in the number of negative associations with P 0.05, which ranged from three (vote count P = 0.004) to four as observed (vote count P = 0.0002).
|
Father's occupational class
Results for measures of father's occupational class are summarized in Table 5. All but three of the studies were from Europe, all but two were case–control studies, and half involved only records. The direction of results from Gardner33,34 varied by choice of control group and information source, reflecting instability due to small numbers, but the
2 test gave P < 0.001 regardless of choice from or exclusion of Gardner. Nine of the remaining 12 associations were positive (sign test P = 0.09), including both cohort studies. Three of these 12 associations were positive with P 0.05 (vote count P = 0.003) and one was negative with P 0.05 (vote count P = 0.3).
|
Four studies had scales requiring category score assignments. Sensitivity analyses used alternative scores for these scales and varied the results from one study that had two control groups and two sources of information on fathers' occupations.33,34 The number of positive associations ranged from nine (sign test P = 0.2) to 11 (sign test P = 0.03); the numbers of positive and negative associations with P
0.05 remained unchanged. The Fisher's exact P-value comparing the proportion of positive associations seen here to the proportion positive for father's education was 0.004. The striking difference of father's occupational class from the income and education tables arises from the six record-only studies, all of which show positive associations, whereas four of the seven interview or questionnaire studies are negative. However, the record-only studies were conducted in much earlier calendar times.
Leukaemia subtypes
Most SES results we found were for all leukaemias combined, and most of the exceptions were confined to one subtype (usually ALL). Of those studies that separated SES data by subtype, only two had enough ANL cases to allow reliable internal comparisons. The series of studies by the Children's Cancer Group obtained very similar results for acute myeloid leukaemia (AML) and ALL (shown under the Brondum and successive Shu reports in Tables 1
–4; the larger P-values under Brondum reflect the smaller number of AML cases); e.g. for maternal education 
16 years vs 12, odds ratios were 0.65 for AML and 0.78 for ALL.19 Likewise, the study by Reynolds95 obtained very similar estimates for ANL an ALL, e.g. for highest parental education 16 years vs <12, odds ratios were 1.24 for ANL and 1.34 for ALL. Consequently, excluding or including non-ALL cases where a choice was possible made no difference to our results (which were dominated by ALL cases, regardless).
Other individual-level SES measures
Table 6 summarizes results for other SES indices measured at the individual level. Most were from case–control studies in Europe. There was a mild tendency towards positive associations. None of these results changed in the sensitivity analyses of category score assignments.
|
Ecological SES measures
Table 7 summarizes results from studies using ecological measures of SES. These include the results from all the ecological designs as well as one cohort study and one case–control study. All but one of the studies using ecological SES measures had start dates in the 1970s or earlier. Half the studies were conducted in the UK. Almost all the associations were positive.
|
|
Discussion |
|---|
|
TopAbstractMethodsResultsDiscussionConclusionsReferences |
|---|
Although it has been stated that the association of childhood cancer with SES ‘has been examined only to a limited extent,’131 we found four dozen studies reporting quantitative results on childhood leukaemia alone. A sharp contrast is present in this literature. Individual-level measures of family income, mother's education, and father's education are consistently associated with childhood leukaemia in the negative direction, with higher rates associated with lower SES levels. Occupational class, whether measured at the ecological or individual level, is associated with childhood leukaemia just as consistently in the opposite direction, with higher rates associated with higher SES.
The meta-confounding problem
Unfortunately, the types of study designs and SES measures in studies conducted in different parts of the world and at different time periods are so highly correlated with each other as to be intractably confounded. No study in North America with a start date more recent than the 1970s used father's occupational class as a measure of SES. No European study used family income and very few used parental education. Almost none of the case–control and cohort studies used ecological SES measures. The preponderance of the ecological studies had start dates before the 1980s and most of the case–control studies had start dates in the 1980s or later. Finally, within most of the study groupings used to form Tables 2–7, almost all studies were of one design (usually case–control). This homogeneity within groupings precluded a meaningful comparison of results based on whether subject cooperation was required or not and allowed no examination of study-design effects. With methodological and population characteristics so strongly associated with each other, it is impossible to disentangle their separate influences.
It would be valuable for future European studies to examine family income, mother's education, and father's education, and for more North American studies to investigate father's occupational class. New ecological analyses would be inexpensive and could help illuminate whether the shift in findings over time has been due to changes in study designs or changes in underlying relations. For example, Borugian et al.132 observed a positive association of neighbourhood income with childhood leukaemia incidence in Canada. This contrasts with the negative associations observed using parental income and education in the Canadian case–control study by McBride et al.70 The conflict suggests that differences in study design or measure are responsible for the apparent time trend in the association.
Different SES measures, including individual-level and ecological measures, may represent different risk factors not only across time periods and geographic locales but also at the same place and time. From this perspective, it is more valuable to keep SES measures separate than to combine them into composite measures. This view contrasts with one in which SES is considered a single construct measured more or less accurately by various indices at the individual and ecological levels. It would be simple and inexpensive for researchers conducting case–control and cohort studies to add community-level SES measures to the individual-level measures in their data analyses. It would even be possible to retrofit datasets for the studies reviewed here with ecological SES variables.
Other statistical limitations
A major limitation of the present review was the focus on the direction of associations and not on their magnitude, owing to the lack of comparability of SES measures across place and time as well as among studies. A much more extensive project (for which we lacked resources) would develop transformation of quantitative results to a common scale to allow pooling. This project, however, would require extensive outside information, such as information on the time-and-place variations in social meaning of such factors as income (in various currencies), and graduation from elementary school, secondary school, and college.
The 2 test (like the P-value) is influenced by both association and study size and so has more power to detect associations than does vote counting. The vote-counting approach is very crude; for example, it classifies RR estimates of 0.20 and 0.98 as ‘negative associations’,53 and classifies P-values of 0.001 and 0.049 as ‘P 0.05’. Vote counting is also inefficient because it weights equally all studies, regardless of size, given that the probability of obtaining an association in one direction or the other, like the probability of obtaining P 0.05, is independent of study size under the null hypothesis. Because of the high relative weighting vote counting gives to small studies, it might also increase concern about bias against publishing ‘non-significant’ results (wherein results with P > 0.05 are more likely to go unpublished than are other results). If this bias is not directional, however, the expected number of positive and negative studies under the null remains 1/2 even if the study P-values do affect publication, and hence (unlike the 2 test or a vote count with 0.05) the sign test will not reflect this bias. We judge the likelihood of important publication bias of this type to be small, however, because of the incidental manner in which the SES results were provided in most of the reports. With the exception of some of the ecological studies, SES variables were almost always reported as descriptive characteristics, often in Table 1 of a published paper.
We have little doubt that detailed reporting by studies of separate SES components (household income, maternal and paternal education and occupation, plus neighbourhood and other ecological SES measures) would greatly aid future quantitative meta-analyses. We would encourage editors to ask for such details in reports.
Methodological problems
In some studies, inappropriate analysis methods were used, which we attempted to correct whenever possible. For this and other reasons, we were obliged to rely in many instances on unadjusted estimates. Sometimes the categories of SES scales were described ambiguously. This was particularly true of parental education scales in which it was impossible to tell whether a given category (e.g. ‘college’) represented the highest level begun or the highest level completed. Our vote counts appeared insensitive to the use of alternative category scores. We suspect that quantitative estimates of summary RRs and of RR trends would have been much more sensitive to this and other elements of the analysis as well as to publication bias. Thus, the lack of resolution in the vote counts may have been accompanied by increased robustness.
In most of the reports we reviewed, SES measures were incidental variables, typically reported for descriptive purposes only. As a consequence, many of the associations we examined were not adjusted for other variables, including the matching variables used in many case–control studies. This may lead to bias, which for the matching factors would usually be towards the null. Another consequence of SES reports being incidental is that our search results were probably incomplete and those reports found were often sketchy. We were unable to include more than a dozen papers because of the incomplete reporting of SES results. We are also aware of at least one study133 for which SES results were reported outside the journal literature.134
Several of the case–control studies that produced negative associations used telephone-based sampling (such as random digit dialling), which tends to produce control groups of higher SES than the target population135–138; area selection (e.g. neighbourhood walking) may be biased in the same direction.139 Nonetheless, we noted no systematic difference in results between these studies and those that selected from government registries of births or children. On the other hand, nearly all the studies of family income, mother's education, and father's education required subject interviews and so suffered meaningful degrees of incomplete participation. This problem might account in part or whole for the negative associations seen in those studies. This possibility is consonant with our observation that the associations for father's occupational class tended to be positive, an excess entirely attributable to the record-only studies for that measure. Nonetheless, there were other differences between the study groupings that might explain the difference such as time period (the record-only studies tended to use earlier data) as well as the measure itself.
The estimates contrasting extreme SES categories in some of the case–control studies were on the order of RR = 0.50 and even lower, especially in North America.19,41,66,70,113 It may seem unlikely that such strong associations were produced entirely by selection and response bias. Unfortunately, relevant quantitative information is limited. Results reported by Hatch et al.140 from the study by Linet et al.61 yield, for a binary scale of family income with a $20 000 per year cut-point, an estimate of RR = 0.72 based on all interviewed participants and RR = 0.56 based on those participants for whom in-home magnetic field measurements were obtained. The estimated bias factor of 0.56/0.72 = 0.78, however, pertains to the final act of participation among families who had already been selected, contacted, recruited, and interviewed. Moreover, the very low RR estimates for childhood leukaemia and SES measures in this and other studies were not for binary SES scales but for contrasts between the extreme ends of scales with four, five, or six categories.
Voigt et al.141 compared controls who participated right away and who delayed their participation (i.e. who initially refused but ultimately participated) by annual household income and occupational class in studies of women in Washington State. The proportions of initial refusers and early responders with incomes under $45 000 were essentially identical (70%), but the proportion with managerial or professional positions was substantially lower among the initial refusers (13%) than among the early responders (22%).
The results of Hatch et al. and Voigt et al. indicate bias towards controls with spuriously high SES distributions. Nonetheless, being confined to families and individuals who were selected and who did give interviews, these results probably understate the total magnitude of control-selection and control-participation bias. There is a need for validation data to be obtained by simulating alternative control recruitment methods in cohorts of individuals and families whose SES is known but who have not already demonstrated themselves to be willing research participants. Without such data, the possibility of large selection and participation bias will call into question many case–control study results.
Participation rates among cases in case–control studies are usually quite high, often reaching 90% or more. Nevertheless, the selection of cases, as opposed to their participation, could potentially be biased in so-called hospital-based studies. Suppose, for instance, that hospitals, clinics, referral centres and oncology practices that tend to treat patients at one end or the other of SES scales are more likely to be interested in collaborating in epidemiological research. In such an event, even a study with a control group perfectly representative of the population of the geographic area in which the medical care facilities are located would be biased.
The cases in several studies in this review were identified because they were already participants in therapeutic trials. The results reported by Brondum et al.,19 Shu et al.,113 and Linet et al.61 were from related case–control studies with cases from the Children's Cancer Group, a consortium of physicians and centres conducting clinical trials, including about half of the childhood leukaemia patients in the US.142 If patients' families, physicians, and treatment facilities differ by SES in their willingness to take part in studies and the refusal rate was high enough, this difference could be an important source of case selection bias. Hatch et al.140 did in fact report enough case non-participation in the Linet et al.61 study to be of concern (although it was about half the rate as that of controls).
Results for mother's education from Greek case–control studies merit special attention. In the first, conducted in Attica and Crete in 1987–91,86 the association was strongly negative with estimates of RR = 0.54 for 7–14 years of education and RR = 0.38 for >14 years (trend P = 0.004; Table 3). In the next study,87 conducted throughout Greece with hospitalized controls in 1993–94, the association was strongly positive, with estimates of RR = 2.0 for 10–12 years and RR = 2.9 for >12 years (trend P = 0.0001; Table 3). An analysis of this second study, extended through 1997 but confined to the 24% of cases for whom blood samples could be obtained and analysed,89 produced a weakly negative association with estimates of RR = 0.68 for 7–9 years, RR = 0.92 for 10–12 years, and RR = 0.77 for >12 years (trend P = 0.7, ineligible for the present review because of overlap with the second study). The variation across these estimates is from RR = 0.38 to RR = 2.9 or 
8-fold. This is well beyond chance variation, and it seems unlikely that the true association of childhood leukaemia with maternal education in Greece vacillated so wildly over a single decade. Thus, we would not dismiss methodological explanations for any of the associations from case–control studies in this literature, including the inverse associations involving family income, mother's education, and father's education.
One could test the hypothesis that individual-level SES measures and ecological SES measures represent different risk factors by adding ecological measures to interview-based case–control studies (an addition that could be done retrospectively to existing studies). True validation studies of control-selection methods in cohorts of known SES but unknown willingness to participate in epidemiological research will be difficult to conduct but may be essential to resolving doubts about the validity of case–control results.
|
Conclusions |
|---|
|
TopAbstractMethodsResultsDiscussionConclusionsReferences |
|---|
The review by Greenberg and Shuster1 accurately characterized research through 1982 as showing a positive association of SES with childhood leukaemia. The vast majority of research has appeared since then, however, and the overall picture has changed considerably. Positive associations are still seen in ecological analyses of average occupational class and in record-based case–control studies of father's occupational class. But negative associations predominate in studies that use interviews or questionnaires to obtain individual-level measures of family income, mother's education, or father's education.
This experience points to the desirability of updating reviews frequently, and of encouraging studies to measure and report SES in more detail than is the norm. It also argues for the writing of review articles at varying levels of generality. Typical reviews cover the entire epidemiology of childhood leukaemia or of all childhood cancers. These overviews have value, but they are too broad to cover any particular aspect in any depth. If there were more focused reviews of narrowly defined topics, broader reviews would have a more reliable foundation upon which to depend.
KEY MESSAGES
|
|
Acknowledgments |
|---|
We are grateful to David Savitz and the referees for their helpful suggestions. This research was supported by the Electric Power Research Institute. Dr. Poole also received partial support from the National Institute of Environmental Health Sciences grant P30ES10126.
|
References |
|---|
|
TopAbstractMethodsResultsDiscussionConclusionsReferences |
|---|
1 Greenberg RS, Shuster JL Jr. Epidemiology of cancer in children. Epidemiol Rev 1985;7:22–48.
2 Smith MA, Ries LAG, Gurney JG, Ross JA. Chapter I. Leukemia. In: Ries LAG, Smith MA, Gurney JG, Linet M, Tamra T, Young JL, Bunin GR (eds). Cancer incidence and survival among children and adolescents: United States SEER Program 1975–1995, National Cancer Institute, SEER Program. NIH Pub. No. 99-4649. Bethesda, MD, 1999, pp. 17–34.
3 Ross JA, Davies SM, Potter JD et al. Epidemiology of childhood leukemia, with a focus on infants. Epidemiol Rev 1994;16:243–72.
4 Robison L, Ross J. Epidemiology of leukaemias and lymphomas in childhood. In: Chessells J, Hann I (eds). Bailliere's Clinical Paediatrics. London: WB Saunders Co., 1995, pp. 639–57.
5 Sandler DP, Ross JA. Epidemiology of acute leukemia in children and adults. Semin Oncol 1997;24:3–16.[ISI][Medline]
6 Electric Power Research Institute. Selection bias in epidemiologic studies of EMF and childhood leukemia. EPRI Report 1008149. Palo Alto, CA: EPRI and Geneva: World Health Organization, 2003.
7 Rothman KJ, Greenland S. Modern Epidemiology. 2nd edn. Philadelphia: Lippincott-Raven, 1998.
8 Greenland S. Bias in indirectly adjusted comparisons due to taking the total study population as the reference group. Stat Med 1987;6:193–95.[Medline]
9 Greenland S, Longnecker MP. Methods for trend estimation from summarized dose-response data, with applications to meta-analysis. Am J Epidemiol 1992;135:1301–09.
10 Hedges LV, Olkin I. Vote-counting methods in research synthesis. Psychol Bull 1980;88:359–69.[CrossRef][ISI]
11 Bushman BJ. Vote-counting procedures in meta-analysis. In: Cooper H, Hedges L (eds). The Handbook of Research Synthesis. New York: Russell Sage Foundation, 1994, pp. 194–213.
12 Alexander FE, Ricketts TJ, McKinney PA, Cartwright RA. Community lifestyle characteristics and risk of acute lymphoblastic leukaemia in children. Lancet 1990;336:1461–65.[CrossRef][ISI][Medline]
13 Alexander FE, Cartwright RA, McKinney PA, Ricketts TJ. Leukaemia incidence, social class and estuaries: an ecological analysis. J Public Health Med 1990;12:109–17.
14 Alexander FE, McKinney PM, Cartwright RA. Migration patterns of children with leukaemia and non-Hodgkin's lymphoma in three areas of northern England. J Public Health Med 1993;15:9–15.
15 Alexander FE, Leon DA, Cartwright RA. Isolation, car ownership, and small area variation in incidence of acute lymphoblastic leukaemia in children. Paediatr Perinat Epidemiol 1996;10:411–17.[ISI][Medline]
16 Birch JM, Swindell R, Marsden HB, Morris Jones PH. Childhood leukaemia in North West England 1954–1977: epidemiology, incidence and survival. Br J Cancer 1981;43:324–29.[Medline]
17 Bithell JF, Dutton SJ, Draper GJ, Neary NM. Distribution of childhood leukaemias and non-Hodgkin's lymphomas near nuclear installations in England and Wales. BMJ 1994;309:501–05.
18 Browning D, Gross S. Epidemiological studies of acute childhood leukemia. A survey of Cuyahoga county, Ohio. Am J Dis Child 1968;16:576–85.
19 Brondum J, Shu XO, Steinbuch M, Severson RK, Potter JD, Robison LL. Parental cigarette smoking and the risk of acute leukemia in children. Cancer 1999;85:1380–88.[CrossRef][ISI][Medline]
20 Buckley JD, Robison LL, Swotinsky R et al. Occupational exposures of parents of children with acute nonlymphocytic leukemia: a report from the Childrens Cancer Study Group. Cancer Res 1989;49:4030–37.
21 Chan LC, Lam TH, Li CK et al. Is the timing of exposure to infection a major determinant of acute lymphoblastic leukaemia in Hong Kong? Paediatr Perinat Epidemiol 2002;16:154–65.[CrossRef][ISI][Medline]
22 Cocco P, Rapallo M, Targhetta R, Biddau PF, Fadda D. Analysis of risk factors in a cluster of childhood acute lymphoblastic leukemia. Arch Environ Health 1996;51:242–44.[Medline]
23 Coebergh JW, van der Does-van den Berg A, van Wering ER et al. Childhood leukaemia in The Netherlands, 1973–1986: temporary variation of the incidence of acute lymphocytic leukaemia in young children. Br J Cancer 1989;59:100–05.[ISI][Medline]
24 Diamond EL, Schmerler H, Lilienfeld AM. The relationship of intra-uterine radiation to subsequent mortality and development of leukemia in children: a prospective study. Am J Epidemiol 1973;97:283–313.
25 Dickinson HO, Parker L. Quantifying the effect of population mixing on childhood leukaemia risk: the Seascale cluster. Br J Cancer 1999;81:144–51.[CrossRef][ISI][Medline]
26 Dockerty JD, Elwood JM, Skegg DC, Herbison GP. Electromagnetic field exposures and childhood cancers in New Zealand. Cancer Causes Control 1998;9:299–309.[CrossRef][ISI][Medline]
27 Dockerty JD, Skegg DC, Elwood JM, Herbison GP, Becroft DM, Lewis ME. Infections, vaccinations, and the risk of childhood leukaemia. Br J Cancer 1999;80:1483–89.[CrossRef][ISI][Medline]
28 Dockerty JD, Draper G, Vincent T, Rowan SD, Bunch KJ. Case–control study of parental age, parity and socioeconomic level in relation to childhood cancers. Int J Epidemiol 2001;30:1428–37.
29 Duncan MH, Wiggins CL, Samet JM, Key CR. Childhood cancer epidemiology in New Mexico's American Indians, Hispanic whites, and non-Hispanic whites, 1970–82. J Natl Cancer Inst 1986;76:1013–18.[Medline]
30 Fasal E, Jackson EW, Klauber MR. Birth characteristics and leukemia in childhood. J Natl Cancer Inst 1971;47:501–09.[ISI][Medline]
31 Feychting M, Ahlbom A. Magnetic fields and cancer in children residing near Swedish high-voltage power lines. Am J Epidemiol 1993;138:467–81.
32 Fulton JP, Cobb S, Preble L, Leone L, Forman E. Electrical wiring configurations and childhood leukemia in Rhode Island. Am J Epidemiol 1980;111:292–96.
33 Gardner MJ, Hall AJ, Snee MP, Downes S, Powell CA, Terrell JD. Methods and basic data of case–control study of leukaemia and lymphoma among young people near Sellafield nuclear plant in West Cumbria. BMJ 1990;300:429–34.[ISI][Medline]
34 Gardner MJ, Snee MP, Hall AJ, Powell CA, Downes S, Terrell, JD. Results of case–control study of leukaemia and lymphoma among young people near Sellafield nuclear plant in West Cumbria. BMJ 1990;300:423–29.[ISI][Medline]
35 Gilman EA, Knox EG. Geographical distribution of birth places of children with cancer in the UK. Br J Cancer 1998;77:842–49.[Medline]
36 Githens JH, Elliot FE, Saunders LH. The relation of socioeconomic factors to incidence of childhood leukemia. Public Health Rep 1965;80:573–78.[ISI][Medline]
37 Glazer ER, Perkins CI, Young JL Jr, Schlag RD, Campleman SL, Wright WE. Cancer among Hispanic children in California, 1988–1994: comparison with non-Hispanic white children. Cancer 1999;86:1070–79.[CrossRef][ISI][Medline]
38 Goodman MT, Yoshizawa CN, Kolonel LN. Incidence trends and ethnic patterns for childhood leukaemia in Hawaii: 1960–1984. Br J Cancer 1989;60:93–97.[Medline]
39 Goodman MT, Yoshizawa CN, Kolonel LN. Ethnic patterns of childhood cancer in Hawaii between 1960 and 1984. Cancer 1989;64:1758–63.[CrossRef][ISI][Medline]
40 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]
41 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]
42 Groves FD, Gridley G, Wacholder S et al. Infant vaccinations and risk of childhood acute lymphoblastic leukaemia in the USA. Br J Cancer 1999;81:175–78.[CrossRef][ISI][Medline]
43 Hems G, Stuart A. Childhood leukaemia in Scotland, 1939–68. Scott Med J 1972;17:13–17.[Medline]
44 Infante-Rivard C, Labuda D, Krajinovic M, Sinnett D. Risk of childhood leukemia associated with exposure to pesticides and with gene polymorphisms. Epidemiology 1999;10:481–87.[CrossRef][ISI][Medline]
45 Infante-Rivard C, Fortier I, Olson E. Markers of infection, breast-feeding and childhood acute lymphoblastic leukaemia. Br J Cancer 2000;83:1559–64.[CrossRef][ISI][Medline]
46 Infante-Rivard C, Krajinovic M, Labuda D, Sinnett D. Parental smoking, CYP1A1 genetic polymorphisms and childhood leukemia (Quebec, Canada). Cancer Causes Control 2000;11:547–53.[CrossRef][ISI][Medline]
47 Infante-Rivard C, Mathonnet G, Sinnett D. Risk of childhood leukemia associated with diagnostic irradiation and polymorphisms in DNA repair genes. Environ Health Perspect 2000;108:495–98.[ISI][Medline]
48 Infante-Rivard C, Olson E, Jacques L, Ayotte P. Drinking water contaminants and childhood leukemia. Epidemiology 2001;12:13–19.[CrossRef][ISI][Medline]
49 Infante-Rivard C, Krajinovic M, Labuda D, Sinnett D. Childhood acute lymphoblastic leukemia associated with parental alcohol consumption and polymorphisms of carcinogen-metabolizing genes. Epidemiology 2002;13:277–81.[CrossRef][ISI][Medline]
50 Infante-Rivard C, Amre D, Sinnett D. GSTT1 and CYP2E1 polymorphisms and trihalomethanes in drinking water: effect on childhood leukemia. Environ Health Perspect 2002;110:591–93.
51 Ji BT, Shu XO, Linet MS et al. Paternal cigarette smoking and the risk of childhood cancer among offspring of nonsmoking mothers. J Natl Cancer Inst 1997;89:238–44.
52 Kaletsch U, Kaatsch P, Meinert R, Schuz J, Czarwinski R, Michaelis J. Childhood cancer and residential radon exposure—results of a population-based case–control study in Lower Saxony (Germany). Radiat Environ Biophys 1999;38:211–15.[CrossRef][ISI][Medline]
53 Kaye SA, Robison LL, Smithson WA, Gunderson P, King FL, Neglia JP. Maternal reproductive history and birth characteristics in childhood acute lymphoblastic leukemia. Cancer 1991;68:1351–55.[CrossRef][ISI][Medline]
54 Kinlen LJ, O'Brien F, Clarke K, Balkwill A, Matthews F. Rural population mixing and childhood leukaemia: effects of the North Sea oil industry in Scotland, including the area near Dounreay nuclear site. BMJ 1993;306:743–48.[ISI][Medline]
55 Kinlen LJ, Hudson CM, Stiller CA. Contacts between adults as evidence for an infective origin of childhood leukaemia: an explanation for the excess near nuclear establishments in west Berkshire? Br J Cancer 1991;64:549–54.[ISI][Medline]
56 Kleinerman RA, Linet MS, Hatch EE et al. Magnetic field exposure assessment in a case–control study of childhood leukemia. Epidemiology 1997;8:575–83.[CrossRef][ISI][Medline]
57 Kleinerman RA, Kaune WT, Hatch EE et al. Are children living near high-voltage power lines at increased risk of acute lymphoblastic leukemia? Am J Epidemiol 2000;151:512–15.
58 Lauriault CD, Jim RT. Childhood leukemia in Hawaii: experience at Kauikeolani Children's Hospital. Hawaii Med J 1996;26:125–26.
59 Laval G, Tuyns AJ. Environmental factors in childhood leukaemia. Br J Industr Med 1988;45:843–44.[Medline]
60 Li CY, Lin RS, Lin CH. Urbanization and childhood leukaemia in Taiwan. Int J Epidemiol 1998;27:587–91.
61 Linet MS, Hatch EE, Kleinerman RA et al. Residential exposure to magnetic fields and acute lymphoblastic leukemia in children. N Engl J Med 1997;337:1–7.
62 London SJ, Thomas DC, Bowman JD, Sobel E, Cheng TC, Peters JM. Exposure to residential electric and magnetic fields and risk of childhood leukemia. Am J Epidemiol 1991;134:923–37.
63 Lowengart RA, Peters JM, Cicioni C et al. Childhood leukemia and parents' occupational and home exposures. J Natl Cancer Inst 1987;79:39–46.[ISI][Medline]
64 Lubin JH, Linet MS, Boice JD Jr et al. Case–control study of childhood acute lymphoblastic leukemia and residential radon exposure. J Natl Cancer Inst 1998;90:294–300.
65 MacMahon B. Prenatal X-ray exposure and childhood cancer. J Natl Cancer Inst 1962;28:1173–91.[ISI][Medline]
66 Ma X, Buffler PA, Selvin S et al. Daycare attendance and risk of childhood acute lymphoblastic leukaemia. Br J Cancer 2002;86:1419–24.[CrossRef][ISI][Medline]
67 MacMahon B, Newill VA. Birth characteristics of children dying of malignant neoplasms. J Natl Cancer Inst 1962;28:231–44.[ISI][Medline]
68 Magnani C, Pastore G, Luzzatto L, Terracini B. Parental occupation and other environmental factors in the etiology of leukemias and non-Hodgkin's lymphomas in childhood: a case–control study. Tumori 1990;76:413–19.[ISI][Medline]
69 Maskarinec G, Cooper J, Swygert L. Investigation of increased incidence in childhood leukemia near radio towers in Hawaii: preliminary observations. J Environ Pathol Toxicol Oncol 1994;13:33–37.[Medline]
70 McBride ML, Gallagher RP, Theriault G et al. Power-frequency electric and magnetic fields and risk of childhood leukemia in Canada. Am J Epidemiol 1999;149:831–42.
71 McKinney PA, Cartwright RA, Saiu JM et al. The inter-regional epidemiological study of childhood cancer (IRESCC): a case control study of aetiological factors in leukaemia and lymphoma. Arch Dis Child 1987;62:279–87.[Abstract]
72 McKinney PA, Alexander FE, Cartwright RA, Ricketts TJ. The Leukaemia Research Fund Data Collection Study: descriptive epidemiology of acute lymphoblastic leukemia. Leukemia 1989;3:880–85.[Medline]
73 McKinney PA, Smith K, Findlay E. The chief scientist reports. The Scottish case control study of childhood leukaemia and cancer: methodology and environmental measures of exposure. Health Bull (Edinb) 1995;53:222–29.
74 McKinney PA, Juszczak E, Findlay E, Smith K. Case–control study of childhood leukaemia and cancer in Scotland: findings for neonatal intramuscular vitamin K. BMJ 1998;316:173–77.
75 McKinney PA, Juszczak E, Findlay E, Smith K, Thomson CS. Pre- and perinatal risk factors for childhood leukaemia and other malignancies: a Scottish case control study. Br J Cancer 1999;80:1844–51.[CrossRef][ISI][Medline]
76 McWhirter WR. The relationship of incidence of childhood lymphoblastic leukaemia to social class. Br J Cancer 1982;46:640–45.[ISI][Medline]
77 Meinert R, Kaatsch P, Kaletsch U, Krummenauer F, Miesner A, Michaelis J. Childhood leukaemia and exposure to pesticides: results of a case–control study in northern Germany. Eur J Cancer 1996;32A:1943–48.[CrossRef]
78 Michaelis J, Schuz 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]
79 Michelozzi P, Capon A, Kirchmayer U et al. Adult and childhood leukemia near a high-power radio station in Rome, Italy. Am J Epidemiol 2002;155:1096–103.
80 Muirhead CR. Childhood leukemia in metropolitan regions in the United States: a possible relation to population density? Cancer Causes Control 1995;6:383–88.[CrossRef][ISI][Medline]
81 Murray L, McCarron P, Bailie K et al. Association of early life factors and acute lymphoblastic leukaemia in childhood: historical cohort study. Br J Cancer 2003;86:356–61.
82 Olsen JH, Nielsen A, Schulgen G. Residence near high voltage facilities and risk of cancer in children. BMJ 1993;307:891–95.[ISI][Medline]
83 Ou SX, Han D, Severson RK et al. Birth characteristics, maternal reproductive history, hormone use during pregnancy, and risk of childhood acute lymphocytic leukemia by immunophenotype (United States). Cancer Causes Control 2002;13:15–25.[CrossRef][ISI][Medline]
84 Perrillat F, Clavel J, Jaussent I et al. Family cancer history and risk of childhood acute leukemia (France). Cancer Causes Control 2001;12:935–41.[CrossRef][Medline]
85 Perrillat F, Clavel J, Auclerc MF et al. Day-care, early common infections and childhood acute leukaemia: a multicentre French case–control study. Br J Cancer 2002;86:1064–69.