The search strategy was designed to identify previous observational epidemiologic studies examining the relationship between residential pesticide exposures during critical exposure time windows (preconception, pregnancy, childhood) and childhood leukemia. Preliminary searches using Ovid MEDLINE were conducted to inform the design of the final search strategy detailed below. An information specialist at the University of Ottawa was also consulted in finalizing the search strategy.

The search strategy was first developed to search the Ovid MEDLINE (1950?March week 3, 2009) and Ovid MEDLINE database of in process and other nonindexed citations (1950 to 31 March 2009) and then adapted to search the Ovid EMBASE (Excerpta Medica Database; 1980 to week 13 2009) (^{Ovid 2009}), TOXNET (Toxicology Data Network; U.S. ^{National Library of Medicine 2009}) (through 31 March 2009), OpenSigle (2009) (through 31 March 2009), and ^{ProQuest Digital Dissertations and Theses (2009)} (through 31 March 2009). The following medical subject headings (MeSH) and key words were used:

• Exposure: exp Environmental Exposure/, exp Environmental Pollutants/, exp Pest Control/, exp Pesticides/, pesticid$.tw, herbicid$.tw, insecticid$.tw, fungicid$.tw
• Population: exp Child/, exp Adolescent/, exp Infant/, child$.tw, adolescen$.tw, infant?.tw, newborn?.tw, youth.tw, teenage$.tw
• Outcome: exp Hematologic Neoplasms/, exp Leukemia/, leuk?emia$.tw

Search terms were grouped according to the Boolean operators OR and AND. A complete depiction of the Ovid MEDLINE search strategy is given in Supplemental Material, Table 1 (available online at doi:10.1289/ehp.0900966.S1 via http://dx.doi.org).

All titles and abstracts identified were independently examined by two of us (M.C.T. and D.T.W.) in order to determine their potential suitability for inclusion in the systematic review. After this primary screen, the complete articles were obtained and the inclusion/exclusion criteria applied. Discrepancies were resolved by consensus. No language criteria were applied. Where abstracts were identified or further details required, particularly relating to the designation of pesticide exposure as residential or occupational, the corresponding author was contacted to ascertain further details of the study. In addition to searching the databases listed above, the reference lists of all included studies and journal Web sites were also hand searched; studies identified manually were evaluated in the same manner as above.

Original epidemiologic studies of childhood leukemia using a case?control or cohort design with an assessment of at least one index of residential/household pesticide exposure/use were included here. Reports were excluded if they were review articles, ecologic studies, case reports, cluster investigations, or studies of adults or if they examined residential exposure or proximity to agricultural pesticides. Where there were multiple publications, the most relevant report was retained (usually the most recent).

After identification of all relevant studies, data abstraction was performed independently by the same two reviewers (M.C.T. and D.T.W.). A standard data abstraction form was prepared and piloted to collect relevant data related to referencing, study design, subject selection, exposure assessment, statistical analysis, and results. A single exposure index was identified for each original study and, where data were available, each combination of exposure time window and pesticide type (unspecified, insecticides, herbicides).

All included studies underwent independent quality assessment by the same two reviewers (M.C.T. and D.T.W.). We used a modified version of the ^{Downs and Black (1998)} checklist for the assessment of the methodological quality of randomized and nonrandomized studies of health care interventions [Supplemental Material, Table 2 (doi:10.1289/ehp.0900966.S1)]. Before conducting the quality assessment, the two reviewers discussed the individual items on the checklist to clarify their interpretation. No attempt was made to blind the reviewers of the authorship or publication status of the original studies. Differences in quality assessment were resolved by consensus.

We conducted meta-analyses using Review Manager (RevMan) version 5.0 (Nordic Cochrane Centre, Cochrane Collaboration, Copenhagen, Denmark). Generic inverse variance data were combined using random effects models to obtain a summary odds ratio (OR) and 95% confidence interval (CI) for the relationship between residential pesticide exposures (unspecified, insecticides, herbicides) and childhood leukemia by exposure time window (preconception, pregnancy, childhood). Heterogeneity across individual studies was quantified by the I² statistic (^{Higgins et al. 2003}). Low, moderate, or high degrees of heterogeneity may be approximated by I² values of 25%, 50%, and 75%, respectively (^{Higgins et al. 2003}). We conducted subgroup analyses according to total quality score (? median) and individual quality score components (> median for external validity and exposure measurement), study design (hospital-based or population-based case?control study), cell type (ALL, AML), location (indoor, outdoor), maternal residential pesticide use (vs. household use or exposure) only, year of publication (studies published in 2000 or later only), and publication status (studies published in the peer-reviewed literature only). Where multiple exposure indices were reported per exposure time window, pesticide type, and study, sensitivity analyses were undertaken using exposure or time window definitions different from those used in the main analysis. Finally, we also examined the impact of removing studies with extreme ORs or the highest weight in analysis, as well as removing individual studies in a sequential manner. Because of the small number of included studies, we assessed publication bias by visual inspection of inverted funnel plots, based on the main finding from all studies (^{Ioannidis and Trikalinos 2007}).

The results of the search strategy and study selection process are detailed in Figure 1. Of the 1,776 studies identified using our search algorithm, 112 were retained from the primary screening process. Most studies were excluded during primary screening because they were irrelevant (n = 1,178), a duplicate record (n = 380), or a review article (n = 93). After the secondary screening process, 17 studies were retained (listed in Table 1). Major reasons for exclusion during secondary screening were irrelevance (n = 36), examination of occupational or residential exposure to agricultural pesticides exposure only (or unclear) (n = 27), or a letter or editorial with no results presented (n = 14).

Of the 17 identified case?control studies, 6 were hospital based, 10 were population based, and 1 reported results separately for both hospital and population controls (Table 1). Studies were conducted in the United States, Canada, Mexico, Japan, France, Brazil, and Germany. Most of the studies were published in the peer-reviewed literature; however, three doctoral dissertations presented results not published elsewhere (^{Davis 1991}; ^{Dell 2004}; ^{Steinbuch 1994}). Although most studies examined both ALL and acute nonlymphoblastic leukemia cases among children and adolescents up to a maximum age of 19 years, one study examined infantile acute leukemia (^{Pombo-de-Oliveira et al. 2006}) and another examined both ALL and AML in children with Down syndrome (^{Alderton et al. 2006}).

Studies varied in size, ranging from a total of 49 leukemia cases (with 7?25 cases ever exposed, depending on exposure category and time window) in the dissertation by ^{Dell (2004)} up to 1,184 cases (with 25?164 exposed) in the German study by ^{Meinert et al. (2000)}. All studies conducted to date relied on parental reports of residential pesticide exposure or use inside or outside of the home, either by themselves or by professional exterminators [Supplemental Material, Appendix 1 (doi:10.1289/ehp.0900966.S1)]. Although most studies assessed use of, or exposure to, pesticides or specific pesticide subgroups (insecticides, herbicides, fungicides), some studies also attempted to collect information on pesticide names and formulations or on target organism (^{Davis 1991}; ^{Dell 2004}; ^{Infante-Rivard et al. 1999}; ^{Ma et al. 2002}; ^{Steinbuch 1994}). Nine studies considered both residential and occupational pesticide exposures (^{Buckley et al. 1989}; ^{Dell 2004}; ^{Kishi et al. 1993}; ^{Lowengart et al. 1987}; ^{Meinert et al. 1996}, ²⁰⁰⁰; ^{Menegaux et al. 2006}; ^{Rudant et al. 2007}; ^{Steinbuch 1994}), and the remaining eight studies were exclusively residential. Five studies clearly specified (or explicitly assumed) whether residential pesticide exposure during pregnancy was attributable to maternal use (^{Davis 1991}; ^{Lowengart et al. 1987}; ^{Ma et al. 2002}; ^{Menegaux et al. 2006}; ^{Rudant et al. 2007}) as opposed to household use or exposure.

Virtually all studies assessed pesticide exposures during separate preconception, pregnancy, and childhood time windows; however, time window definitions differed somewhat by study (Appendix 1). ^{Leiss and Savitz (1995)} considered only the last 3 months of pregnancy. ^{Ma et al. (2002)} considered the first 3 years of age in a separate manner. ^{Davis (1991)} considered the first 6 months of age separately from the remainder of the childhood period. Some studies also combined results from different exposure time windows in analysis and reporting (^{Meinert et al. 1996}, ²⁰⁰⁰; ^{Rudant et al. 2007}; ^{Steinbuch 1994}).

Quality scores are presented in Supplemental Material, Table 3 (doi:10.1289/ehp.0900966.S1). For hospital-based studies, total scores ranged from 7 to 12, with a median value of 9, of a possible maximum score of 20. For population-based studies, quality scores were higher, with a range of 9?14 and a median of 11. More recent studies tended to have higher quality scores. In assessing external validity, questions remained regarding the representativeness of subjects (both selected and participating), particularly for earlier hospital-based studies. Only ^{Buckley et al. (1989)} reported that interviewers were blind to case/control status; however, the ethics of such practices have also been questioned (^{Infante-Rivard et al. 1999}). Because of the self-reported nature of exposure data, no study received a point for avoidance of bias from misclassification, since the possibility for differential misclassification remained. Only ^{Dell (2004)} and ^{Ma et al. (2002)} reported results for a clearly defined preconception exposure time window. There were few data regarding frequency or duration of pesticide use, with most studies reporting only ?ever/never? use of/exposure to the pesticide of interest. Six studies attempted to examine potential exposure?response relationships (^{Alderton et al. 2006}; ^{Buckley et al. 1989}; ^{Dell 2004}; ^{Infante-Rivard et al. 1999}; ^{Ma et al. 2002}; ^{Meinert et al. 2000}). Although confounding is difficult to assess because there are few established risk factors for childhood leukemia, most studies examined or adjusted for at least a range of sociodemographic and maternal characteristics. Four studies, however, explicitly assessed the potential confounding influence of maternal or childhood X-ray exposure (^{Dell 2004}; ^{Fajardo-Gutierrez et al. 1993}; ^{Kishi et al. 1993}; ^{Lowengart et al. 1987}).

To assess the possibility of publication bias, we examined the main findings from all included studies in an inverse funnel plot [Supplemental Material, Figure 1 (doi:10.1289/ehp.0900966.S1)]. Although limited by the small number of individual studies, there was some evidence for asymmetry, with a lack of small studies found with effect sizes smaller than those from larger studies. Asymmetry may also be due to a range of other factors, including study quality, methodological differences, or the study populations per se. We attempted to identify all relevant original studies possible, including three doctoral dissertations (^{Davis 1991}; ^{Dell 2004}; ^{Steinbuch 1994}) and two studies published in a language other than English (^{Fajardo-Gutierrez et al. 1993}; ^{Kishi et al. 1993}).

Of the 17 identified studies, we excluded two from the quantitative data synthesis due to a lack of CIs (^{Schwartzbaum et al. 1991}) or a unique study population (Down syndrome cases only) (^{Alderton et al. 2006}). Supplemental Material, Appendix 2 (doi:10.1289/ehp.0900966.S1) lists the individual studies included in each overall and subgroup analysis by exposure time window and pesticide type.

Preconceptional household use of unspecified residential indoor (summary OR = 1.53; 95% CI, 0.98?2.39; I² = 0%) and outdoor (OR = 1.69; 95% CI, 1.03?2.77; I² = 0%) pesticides was positively associated with childhood leukemia based on the two available studies (^{Dell 2004}; ^{Ma et al. 2002}). We also found a significant positive association with preconceptional residential insecticide use (OR = 1.92; 95% CI, 1.34?2.74; I² = 0%) in the same two studies.

Exposure to unspecified residential pesticides during pregnancy had a significant and positive association with childhood leukemia when combining results from 11 studies (OR = 1.54; 95% CI, 1.13?2.11; I² = 66%), although the combined estimate had substantial heterogeneity (Figure 2A, Table 2). The magnitude of the positive association increased somewhat and heterogeneity was reduced when we examined ALL only (OR = 2.04; 95% CI, 1.54?2.68; I² = 19%), indoor use of unspecified pesticides (OR = 1.86; 95% CI, 1.25?2.77; I² = 9%), maternal use of unspecified pesticides (OR = 2.07; 95% CI, 1.62?2.64; I² = 19%), and studies published in the peer-reviewed literature only (OR = 1.81; 95% CI, 1.37?2.39; I² = 36%). We observed the largest OR for studies published in or since the year 2000 (OR = 2.17; 95% CI, 1.85?2.53; I² = 0%).

Exposure to residential insecticides during pregnancy was associated with a significant increase in risk of childhood leukemia, when combining the results from eight studies (OR = 2.05; 95% CI, 1.80?2.32; I² = 0%), with little evidence of heterogeneity (Figure 2B, Table 3). The summary OR changed little according to total study quality, individual quality components, hospital- versus population-based design, or publication status. The association was somewhat stronger for studies of ALL (OR = 2.14; 95% CI, 1.83?2.50; I² = 0%) compared with AML (OR = 1.85; 95% CI, 1.29?2.64; I² = 0%) and for indoor use (OR = 1.90; 95% CI, 0.61?2.23; I² = 0%) compared with outdoor use (OR = 1.54; 95% CI, 0.86?2.74; I² = 36%), although they are based on fewer studies.

Exposure to residential herbicides during pregnancy also had a significant positive association with childhood leukemia (OR = 1.61; 95% CI, 1.20?2.16; I² = 0%) when combining the results from five studies (Figure 2C, Table 4). Again, we observed little difference in the summary OR according to study quality, study design, or publication status. The combined relative risk estimate increased somewhat for ALL (OR = 1.73; 95% CI, 1.28?2.35; I² = 0).

Results for the pregnancy exposure time window were fairly robust to sensitivity analyses: removing studies with extreme ORs or with the highest weight or including additional studies with wide or ill-defined exposure time windows. However, removing the study of AML by ^{Steinbuch (1994)} from the unspecified pesticide analysis did result in a somewhat stronger association (OR = 1.74; 95% CI, 1.36?2.24; I² = 31%).

Results for residential pesticide exposure during childhood were somewhat similar but were attenuated compared with those for pregnancy (Figure 3A, Table 2). We found a significant positive association between residential unspecified pesticide exposure during childhood and childhood leukemia (OR = 1.38; 95% CI, 1.12?1.70; I² = 4%), with little heterogeneity. The magnitude of the association was somewhat stronger for indoor use (OR = 1.56; 95% CI, 1.02?2.39; I² = 7%), studies published in or since the year 2000 (OR = 1.55; 95% CI, 1.14?2.12; I² = 0%), and studies published in the peer-reviewed literature (OR = 1.56; 95% CI, 1.19?2.04; I² = 0%).

Exposure to residential insecticides during childhood was also positively associated with childhood leukemia (OR = 1.61; 95% CI, 1.33?1.95; I² = 0%) when combining results from seven original studies (Figure 3B, Table 3). With restriction to studies of higher total methodological quality, the magnitude of the association was reduced and was no longer significant (OR = 1.36; 95% CI, 0.84?2.21; I² = 30%). Our findings were similar when evaluating population-based studies (OR = 1.48; 95% CI, 1.03?2.11; I² = 40%), studies of ALL (OR = 1.35; 95% CI, 0.76?2.38; I² = 51%), and studies with results for outdoor insecticide use only (OR = 1.43; 95% CI, 0.71?2.86; I² = 78%). Sensitivity analysis using an alternate exposure metric for ^{Leiss and Savitz (1995)} (using home extermination for insects as opposed to pest strips for insects in the home) decreased the magnitude of the association and also increased heterogeneity (OR = 1.29; 95% CI, 0.84?1.97; I² = 62%). ORs were elevated, however, with recent year of publication (OR = 1.70; 95% CI, 1.28?2.27; I² = 0%) and among studies published in the peer-reviewed literature (OR = 1.73; 95% CI, 1.41?2.12; I² = 0%).

Finally, we observed no association between exposure to residential herbicides during childhood and childhood leukemia when combining results for four studies overall (OR = 0.96; 95% CI, 0.59?1.58; I² = 72%) (Figure 3C, Table 4). We also observed substantial heterogeneity. Sensitivity analysis using alternate exposure indices for ^{Ma et al. (2002)} (using year 2 of childhood as opposed to year 1) and for ^{Davis (1991)} (using garden herbicide use between 7 months of age and age at diagnosis as opposed to yard herbicides from 0 to 6 months of age) resulted in a summary OR for childhood leukemia of 1.38 (95% CI, 1.10?1.72; I² = 0%).