Results of this meta-analysis indicated an inverse association between atopy/allergies and childhood ALL; the pooled odds of any reported atopy/allergy were 31% lower among cases than among controls and were 21% lower for asthma, 26% lower for eczema, and 45% lower for hay fever. No association was observed for leukemia overall or AML or for any of the specific atopic conditions and AML examined separately, although there were fewer studies of AML and fewer AML cases. The divergent results observed for ALL and AML are of interest, given that atopic status should have been captured similarly across the 2 case groups.
Two conflicting hypotheses have been proposed to explain a causal relation between atopy and cancer. The immune surveillance hypothesis asserts that the immune system recognizes antigens of malignant cells as foreign and mounts a response to them, preventing a majority of potential cancers from developing (
24). The presence of an atopic condition is thought to increase the vigilance of the immune system in monitoring for, identifying, and eliminating malignant cells (
3,
25). In support of this hypothesis is the observation that immunocompromised persons have a higher incidence of specific malignancies than persons with intact immune systems (
26). Notably, immunosupressed persons have an increased risk of non-Hodgkin lymphoma (
27), a malignancy with histology identical to that of T-cell ALL (
2).
The second hypothesis is that chronic stimulation of the immune system by allergens increases the risk of carcinogenesis (
25,
26). A greater number of proliferating cells increases the probability of genetic errors, such as pro-oncogenic mutations, that may not be repaired prior to subsequent divisions (
25). This mechanism has been proposed as an explanation for the positive association between autoimmune disease and malignancy (
28), for example.
It is not clear whether either of these mechanisms is applicable to childhood/adolescent cancers; however, investigators in the majority of childhood/adolescent leukemia studies reported inverse associations, potentially supporting the immune surveillance hypothesis. Other possible explanations include reverse causality, where the leukemic process induces atopic manifestations, and a common etiology for both atopy and leukemia. There are case reports exemplifying the former explanation (
29–
31). As an example of the latter, Smith et al. (
32) demonstrated that the hygiene hypothesis may also apply to childhood leukemia. In that ecologic study, decreases in the prevalence of hepatitis A infection, a marker for hygiene due to the fecal-oral mode of transmission, were associated with increased ALL risk in the United States and Japan (
32). Neither of these is consistent with the observed inverse association, however.
If an inverse relation existed, ALL incidence would be expected to decrease with increasing atopy prevalence. The observed increase in ALL incidence (
1) is not inconsistent, however, since ALL is multifactorial, requiring 2 or more “hits” (
33), and atopy would constitute a single “hit” (i.e., it is neither sufficient nor necessary). Assuming that the odds ratio for atopy/allergies is 0.69 and the prevalence of atopy in cases is 30%–40%, the attributable fraction for atopy is −14% to −18% (
34). It is possible that other putative risk factors with increasing secular trends (e.g., birth weight (
35)) may contribute to the increase in ALL and offset any decrease attributable to atopy.
Although no biologic mechanism has been established, the principal factor linking childhood/adolescent leukemia and atopic disease is the rate at which the immune system matures (
10). Hypotheses by Greaves (
36) and Kinlen (
37) suggest an etiologic role for the immune system in the development of childhood leukemia via delayed exposure and abnormal response to early-life infections; no responsible virus or other infectious agent has been identified to date (
38,
39). Predominance of T helper 2 (Th2) cells versus T helper 1 (Th1) cells is often described with respect to atopy (
40), where each cell type has a specific cytokine profile and associated sequelea. Infants are born with a Th2-dominated immune profile; nonatopic infants gradually migrate to a Th1-dominant process by age 2 years, while infants with a family history of atopy fail to make the Th2-to-Th1 transition (
41). Early exposure to infectious agents is thought to stimulate this transition. Two additional T-cell types, T-regulatory and Th17, may also play a role in the complex interplay between atopy, infections, and autoimmune diseases (reviewed by Chang et al. (
42)).
There were several notable sources of heterogeneity across the 10 included studies. Results of the stratified analysis indicated that the most important of these were case-control study design, source of exposure data, and/or inclusion of a latency period. Only 3 studies were nested within a larger cohort (
10,
12,
14); these types of studies have the advantage of temporality over other case-control designs, in that exposures are measured prior to leukemia diagnosis.
Two aspects of exposure measurement—which data were collected and how they were collected—probably contributed to the observed heterogeneity. As shown in , definitions for composite atopy/allergy variables differed across studies in that authors included different permutations of asthma, eczema, hay fever, and hives. One study included neurodermatitis instead of eczema (
22). In 2 studies, composite variables incorporated food, drug, bee-sting, pollen, dust, or pet-dander allergies, accounting for a large proportion of exposed persons and potentially driving the observed associations (
21,
23). Söderberg et al. (
14) used the Swedish hospital discharge registry to identify patients discharged for asthma, which is qualitatively different than exposure classifications used in other studies and probably involves substantial underascertainment. Notably, some authors presented definitions that included the presence of symptoms (
10) or the use of medications (
11). Although they were not used in the current meta-analysis, these definitions, in concert with physician diagnosis, may increase sensitivity and could be considered in future studies.
Three studies used subject medical records for exposure assessment (
10,
12,
14), which also has the notable advantage of temporality. Medical records are typically considered the gold standard for medical history data; however, this is only applicable if medical records from all sources are obtainable and complete, which is not always the case (
43).
Atopic disease was measured via parental report in the remaining studies (
11,
13,
19–
23), introducing several potential sources of misclassification and/or recall bias. Subclinical atopy may not be diagnosed by physicians or recognized by parents. For example, it has been estimated that 30%–40% of “healthy” persons are atopic (
44), meaning that they produce an immunoglobulin E response to an environmental stimulus. A related issue is that atopic persons experience changes in symptoms over time (
45). Conversely, not all cases of asthma, eczema, or hay fever are atopic (
46). Parents of cases may be more motivated and more likely to recall exposures than parents of controls (
47). It is also possible that early symptoms of leukemia may be mistaken for atopic disease (
3). Alternatively, parents of cases may be underreporting allergic diseases if immunosuppressive therapy administered for leukemia reduces allergic symptoms and parents report allergy status after initiation of treatment (
22). Parents of controls may report allergic disease arising after a specified reference date, since the reference date does not carry the personal significance of a diagnosis date (
22). A sensitivity analysis by Schüz et al. (
22), wherein reference dates for controls were reclassified to interview dates, revealed that the observed associations were attenuated somewhat, but they could not be wholly attributed to recall bias.
Results of 1 validation study indicated that parental recall of an asthma diagnosis is high (87% agreement with medical records) (
48). A second validation study by Hughes et al. (
10) demonstrated that accuracy of maternal recall of asthma was high and approximately equal between cases and controls (sensitivity was 81% in cases vs. 83% in controls), although the level of agreement between maternal recall and medical records was marginally higher among cases (in cases, κ = 0.69; in controls, κ = 0.60). Recall of eczema was somewhat lower (sensitivity was 51% in cases vs. 57% in controls), but agreement with medical records was equivalent between cases and controls (κ = 0.46). Therefore, parental recall of atopic diagnosis is predicted to be similar across cases and controls, although degree of recall will probably vary by condition.
To minimize concerns about misclassification, a latency period should be incorporated between onset of atopic symptoms and leukemia diagnosis. Importantly, only 4 included studies incorporated a latency period, and among those, the length of the latency period varied (3 months (
10), 6 months (
20), or 1 year (
12,
14)).
The age range of included subjects is another potential source of heterogeneity. Most studies examined children/adolescents under age 15 years; however, Bross and Natarajan (
20) did not include infants, Spector et al. (
12) limited their analysis to children aged 6 years or less, and Söderberg et al. (
14) included children and adolescents aged 18 years or less (by request). Age is an important consideration, because the prevalence of atopy increases with age (
45). Further, the age-standardized incidence of leukemia is greatest among children under age 5 years, with lower rates among infants and children/adolescents aged 5–19 years (
1). Importantly, in 3 studies investigators presented odds ratios by age group; however, categories differed across studies and data could not be reasonably pooled (
13,
20,
21).
The principal limitation of any meta-analysis is the great potential for selection bias, encompassing bias related to publication, language, citation, and multiple publication (
49). Failure to include all studies of a given association may produce a summary odds ratio that overestimates the true effect (
50). To minimize publication bias, we contacted international experts on childhood cancer etiology to inquire about unpublished or unidentified studies in addition to those located in the systematic electronic database search. Additional limitations include the small number of studies for each exposure-disease association and the relatively high level of heterogeneity detected across studies, which may restrict the generalizability of the summary odds ratios produced.
There were also limitations within the individual case-control studies. Recall bias, as discussed above, is a primary concern with retrospective study designs; selection bias is also of concern. Selection bias was probably absent in the studies based on registry or health maintenance organization data (
12,
14), but it cannot be ruled out in other studies with lower participation rates, especially since odds ratios for ALL were closer to the null for both cases and controls for response rates of ≥80% versus <80% in the stratified analyses. Also noteworthy is the exclusion of very sick or deceased cases from 1 study, potentially introducing survival bias (
11). Further, participants in case-control studies of childhood cancer tend to be of higher SES (
51), and SES is associated with both atopy (
52) and leukemia (
53). Adjustment for SES did not result in disparate stratum-specific odds ratios for ALL for atopy/allergies or asthma, but there was a possible effect for eczema. The potential effects of misclassification and other sources of bias are best evaluated by conducting a formal uncertainty analysis (
54).
Although the results of this meta-analysis indicate inverse associations between atopy/allergies and childhood/adolescent ALL, causes of the observed statistical associations must be investigated thoroughly to rule out explanations other than a direct causal relation, such as reverse causality or selection or recall bias. Ideally, future studies would include a prospective design with analysis of biologic specimens to avoid the pitfalls of temporality and misclassification that plague existing studies.
This article has been 
