Air pollution and childhood cancer: A review of the epidemiological literature

Abstract

The authors evaluated support in the literature for the hypothesis that ambient air pollution causes childhood cancer. The PubMed database was searched for original articles, which were reviewed for evidence of a relation with the main types of childhood cancer, using criteria including sample size, magnitude and precision of relative risk estimates, presence of a dose–response pattern and potential for bias. The hypothesis has been studied almost entirely with respect to traffic-related air pollution. Since derivation of the hypothesis from 2 case–control studies in Denver, USA, two further case–control studies have provided new positive evidence and 4 case–control and 7 ecological studies mainly negative evidence. The 4 case–control studies providing positive evidence were relatively small and tended to have more methodological limitations than those showing no association. Publication bias is possible. The weight of the epidemiological evidence indicates no increased risk for childhood cancer associated with exposure to traffic-related residential air pollution. Nevertheless, the limited number of studies, the methodological limitations of both positive and negative studies and the absence of consistency in the results obviate a firm conclusion of no effect. In particular, nondifferential misclassification of exposure might have masked true, weak associations. © 2006 Wiley-Liss, Inc.

Cancer is the leading cause of mortality from disease in childhood in the western world. The incidence rates among white children are 120–150 per year per million boys and 110–140 per year per million girls in the countries of Europe and North and South America and in Australia and New Zealand, where incident cancer cases are registered routinely. In these areas, leukemias, central nervous system tumors and malignant lymphomas constitute about two-thirds of all childhood cancers. Significant differences in incidence rates are registered in populations of other ethnic origins.1

The etiology of childhood cancers is largely unknown. Few risk factors have been established, but these factors explain only a small proportion of childhood cancers.2, 3 Suspected or suggested risk factors include electromagnetic fields and air pollution from traffic,3 which are among the best studied environmental factors.

In many parts of the world, road traffic is the main source of ambient air pollution in urban areas and there is a growing body of literature on risks associated with proximity to high-density traffic. Air pollution from traffic is a complex mixture of many chemicals, of which many are known or suspected carcinogens; in 1987, the International Agency for Research on Cancer classified diesel and gasoline exhaust, respectively, as probably (Group 2A) and possibly (Group 2B) carcinogenic to humans.4

In this review, we evaluate the available evidence from the epidemiological literature on the hypothesis that ambient air pollution, in particular that from traffic, causes childhood cancer.

Materials and methods

The PubMed database was searched through September 2005 by entering the MeSH terms “air pollution,” “neoplasms” and “child,” yielding 277 hits. From these, 10 articles were selected in which the results of original epidemiological research were reported (which were written in English), in which ambient air pollution was considered as the exposure and exposure of children was considered and which dealt primarily with cancers diagnosed in persons under 15 years of age. The reference lists of these articles were reviewed, resulting in identification of another 6 articles fulfilling the same criteria. Yet another article was identified by reviewing recent issues of relevant journals. One article5 was excluded because it reported replication of a previous study,6 and another7 was excluded because the study population was included in a subsequent, larger study of the same design.8 Thus, 15 articles were used in this review.

Each article was evaluated for information on the aims, design, study population, setting, exposure assessment methods, results and potential for bias to facilitate assessment of the degree of support for the hypothesis that air pollution, in particular that from traffic, causes childhood cancer. The criteria used in this assessment included results for the main types of childhood cancer, the magnitude and precision of the relative risk estimates, presence of a dose–response pattern and potential for bias.

Results

The evidence from the 15 epidemiological studies is summarized in Table I and is discussed later.

Table I. Fifteen Studies of Air Pollution and Childhood Cancer

Reference, location, year of publication	Design	Cases: number and definition	Period and study area	Air pollution exposure assessment method	Indicator for exposure	Exposure difference evaluated	Relative risk estimate (95% CI)	Commentary
CNS, central nervous system. 1 Seventy-four of 491 case addresses and 48 of 472 control addresses. Some children contributed two addresses: that at birth and that at diagnosis.
Case–control studies
Wertheimer et al.,9 Colorado, USA, 1979	Case–control	n = 344, all cancers, 0–18 years, mortality	1950–73, Denver area	Traffic counts at home addresses at birth or at death	Home within 40 m of a street with >5000 vehicles per day	Yes vs No	Not estimated1	Traffic considered only as a potential confounder
Savitz et al.,6 Colorado, USA, 1989	Case–control	n = 328, all cancers, 0–14 years, incidence	1976–83, Denver area	Traffic counts at home address at time of diagnosis	Number of vehicles per day	>500 vs <500	All cancers: 1.7 (1.0–2.8); leukemia: 2.1 (1.1–4.0); CNS tumor: 1.7 (0.8–3.9); lymphoma: 0.7 (0.2–3.0)	Matched on age, sex and area. Indication of dose-response for all cancers and leukemia. Adjustment in a subset of cases and controls for mother's age, father's education, per capita income, wire configuration and other variables had little effect on the risk estimates.
						>10 000 vs <500	All cancers: 3.1 (1.2–8.0); leukemia 4.7 (1.6–13.5)
Feychtinget al.,10 Sweden, 1998	Case–control	n = 142, all cancers, 0–15 years, incidence	1976–83, children living within 300 m of high-voltage power lines in Sweden	Modelled peak concentrations of NO2. Based on latest address within power-line corridor	99th percentile of1-h means of NO2 over 1 year	>50 (75th percentile) vs <40 (50th percentile)	All cancers: 2.7 (0.9–8.5); leukemia: 2.7 (0.3–20.6); CNS tumor: 5.1 (0.4–61.2)	Matched for calendar time, geographical area and residence near same power line. Adjustment for electromagnetic fields and socioeconomic status did not materially change the results. Indication of dose–response relation.
Raaschou-Nielsen et al.,11 Denmark, 2000	Case–control	n = 1989, leukemia, CNS tumor and lymphoma, 0–14 years, incidence	1968–91, Denmark	Traffic density and modelled air pollution concentrations at residence. Based on all home addresses from time of conception to time of diagnosis	Number of vehicles per day during childhood	>10,000 vs <500	All cancers: 1.0 (0.7–1.6); leukemia 1.1 (0.6–2.2); CNS tumor: 0.9 (0.4–1.8); lymphoma: 1.3 (0.4–4.8)	Matched on sex, age and calendar time. Adjusted for urban development, geographical region, type of residence, electromagnetic fields, mother's age and birth order.
					Cumulated NO2 concentrations during childhood	Highest 1% vs lowest 50%:	All cancers: 1.2 (0.6–2.3); leukaemia: 0.4 (0.1–1.3); CNS tumor: 1.0 (0.3–3.1); lymphoma: 4.7 (1.2–17.6)	No significant dose–response relation for any of these results.
Langholz et al.,12 California, USA, 2002	Case–control	n = 212, leukemia, 0–10 years, incidence	1978–84, Los Angeles County	Sum of traffic counts at all streets within 457 m (1500 feet) of home address at which the child had resided the longest. A distance-weighted metric was used.	Number of vehicles per day (distance-weighted)	>28,497 (upper 20%) vs < 2301 (lower 20%)	Leukemia:1.4 (0.7–3.0)	Matched on sex and age. Adjustment for wire coding. Further adjustment for other variables changed the results very little. No evidence of dose–response relation. Highest risk (RR = 1.6) in second quintile.
Crosignani et al.,13 Italy, 2004	Case–control	n = 120, leukemia, 0–14 years, incidence	1978–97, province of Varese	Modelled concentration of benzene outside residence at timeof diagnosis	Yearly mean	>10 vs <0.1 μg/m^3	Leukemia:3.9 (1.4–11.3)	Matched by age and sex. Adjustment for socioeconomic status of the area had minimal influence on the risk estimates. Significant dose–response relation across three exposure groups.
Reynolds et al.,8 California, USA, 2004	Case–control	n = 4369, all cancers, 0–4 years, incidence	1988–97	Road density and traffic densitybased on traffic counts and lengthof roads within152 m (500 feet)of address at time of birth	Road density in miles per square mile. Traffic density in vehicle miles travelled per square mile	Traffic density: highest 10% vs lowest 25%	All cancers: 0.92 (0.80–1.06); leukaemia:0.92 (0.73–1.15); CNStumor: 1.22(0.87–1.70)	Matched for age and sex. Adjustment for race and ethnicity. Little effect of further adjustment for maternal age, birth weight, neighborhood income or county-level benzene emissions. No indication of dose–response relation.
Steffenet al.,14 France, 2004	Case–control	n = 280, acute leukemia, 0–14 years, incidence	1995–99, Nancy, Lille, Lyon, Paris	Face-to-face interviews with mothers. Standardized questionnaire on different types of exposure	Heavy traffic roads within 50 m of residence.Neighbouringrepair garage or petrol station	Heavy traffic roads during childhood (yes vs no).	0.9 (0.7–1.3)	Matched on age, sex, ethnic origin, hospital center, rural or urban setting. Adjustment for a number of potential confounders did not modify the association for neighbouring business.
						Neighbouringbusiness during childhood (yes vs no)	4.0 (1.5–10.3)
Ecological studies
Alexander et al.,15 United Kingdom, 1996	Ecological. Observed-to-expected ratios correlated with prevalence of car ownership in 3270 small-area units	n = 438, acute lymphocytic leukemia, 0–14 years, incidence	1984–1989, England and Wales	Information on car ownership derived from 1981 census	Proportion of households with no car	>44% (least cars)vs <17%(most cars)	Acutelymphocytic leukaemia: 1.4 (0.8–2.5)	Adjusted for socioeconomic status and degree of isolation at area level with little effect on the risk estimates. No dose–response relation.
Knox et al.,16 United Kingdom, 1997	Ecological. Observed-to-expected numbers of deaths compared for areas at different distances from potential hazard sites	n = 22 458, all cancers, 0–15 years, mortality	1953–80, all of England, Scotland and Wales	Identification of location of major industrial sites, transport routesand airborne drift, as well as home addresses at death and birth ofcancer cases	Distance from potential environmental hazard sites	Residence within 5 km of ahazard site	Significantlyelevated case density ratios observed inareas close toabout 30different typesof potentialhazards. Ratios were mostly between 1.10and 1.20.Ratios werenot highernear benzene works.	No adjustment. Number of postcodes used to calculate expected number of cases; actual population at risk unknown.
Nordlinder et al.,17 Sweden, 1997	Ecological. Incidence rates compared among municipalities	n = 982, leukemia and lymphoma, 0–24 years, incidence	1975–85, Sweden	Car density in municipality of residence	Cars per km^2	Acute lymphocytic leukemia^2:	Incidence rate (per 10^6 person–year)	No adjustment. Difference in incidence rates significant only for acute myeloid leukemia.
						<5 cars/km	21.1
						≥20 cars/km^2	18.6
						Acute myeloblastic leukemia:
						<5 cars/km^2	3.4
						≥20 cars/km^2	5.5
						Chronic myeloid leukemia:
						<5 cars/km^2	0.5
						≥20 cars/km^2	1.3
						Non-Hodgkin lymphoma:
						<5 cars/km^2	3.2
						≥20 cars/km^2	3.5
Harrison et al.,18 United Kingdom, 1999	Observed compared with expected numbers of cases in areas close to benzene sources	n = 381, solid tumor and leukemia, 0–15 years, incidence	1990–94, West Midlands	Distance from home to main roads and petrol stations inferred from GIS operations. Based on post code of home address at time of diagnosis	Residence within 100 m of a main road or petrol station (Yes/No)	Main roads:		No adjustment. Uncertainties in population numbers and age and sex distribution used to calculate expected numbers of cases
						Leukemia:	1.2 (0.7–1.7)
						CNS tumor:	0.8 (CI not given)
						Petrol stations:
						Leukemia:	1.5 (0.7–2.9)
						CNS tumor:	0.8 (CI not given)
Reynolds et al.,19 California, USA, 2002	Ecological. Incidence rates compared between areas with different traffic loads	n = 7143, all cancers, 0–14 years, incidence	1988–94, California	Different traffic measuresallocated to eachof 21 519 small-area units. Allocation ofcases to blocks based on home address at time of diagnosis	Vehicle miles travelled, per day, per square mile	>320,701 (upper 10%) vs <33,291 (lower 25%)	All cancers: 1.08 (0.98–1.20); leukemia: 1.15 (0.97–1.37);glioma (CNS):1.14 (0.90–1.45)	Adjusted for age, race, ethnicity and sex on ecological level. Uncertainties in population numbers. No indication of increased risk for two other measures of traffic load: vehicles per square mile and miles of road per square mile.
Reynolds et al.,20 California, USA, 2003	Ecological. Incidence rates compared for areas with different air pollution levels	n = 6989, all cancers, 0–14 years, incidence	1988–94, California	Mean concentrations of carcinogenic air pollutants modelled for each census tract and weighted by cancer potency in an exposure score index. Based on 1990 census data and address at time of diagnosis	Exposure scores in relation to different emission sources	Mobile source exposure scores: highest 10% vs lowest 25%	All cancers: 1.04 (0.95–1.14); leukaemia: 1.18 (0.98–1.41); gliomas (CNS): 1.02 (0.83–1.26)	Adjusted for age, race, ethnicity and sex on ecological level. Uncertainties in population numbers. Little effect of further adjustment for socioeconomic status. No indication of dose–response relation.
Visser et al.,21 Netherlands, 2004	Standardized incidence rates for exposed vs unexposed population	Five cases of acute lymphocytic leukemia, 0–14 years, in exposed population. Incidence.	1989–97, Amsterdam	Traffic counts at home address in 1998	Home within 50 m of a street with > 10 000 vehicles per day	Yes vs No	1.35(0.44–3.15)	Incidence rates age-and sex-standardized

Case–control studies

Initial interest in the traffic-related hypothesis was prompted by 2 case–control studies of wire configuration and electromagnetic fields near the homes of children with cancer, which considered proximity to high-density traffic as a potential confounder. The first study noted an excess of cancer deaths among children with more than 5,000 vehicles per day9 and Savitz et al. also found an apparent association with traffic in their study of electromagnetic fields.6, 22 It is interesting that the findings of these 2 independent studies from the Denver area, USA, were similar, and the suggested associations have been explored further in a number of case–control studies in various countries.8, 10, 11, 12, 13, 14 Most have found no association.

In a Swedish nested case–control study, information on traffic density, street type, speed limit, street width and distance between the house and the street was used to calculate the NO₂ concentration at home addresses.10 The study noted elevated, but mainly statistically insignificant, risk for childhood cancers among children living at addresses with high concentrations of NO₂.

In a large Danish population-based case–control study, the residential history of each child was traced from 9 months before birth to the time of diagnosis.11 The concentrations of benzene and NO₂ at each address were calculated from a model based on data on traffic and the configuration of the street and buildings at the address, emission factors for the Danish car fleet, meteorological variables and the background air pollution concentration. Overall, the results indicated no association between childhood cancer and traffic density, NO₂ or benzene exposures either in utero or during childhood.

Two case–control studies were conducted in California, USA.8, 12 Langholz et al.12 evaluated traffic density near the residences of children with leukemia and controls from an earlier study of electromagnetic fields23 in the area of Los Angeles, which has some of the heaviest traffic in the USA. A large statewide California study evaluated risk relations for childhood cancers among young children on the basis of traffic patterns near the maternal residence at the time of the child's birth.8 Neither study found evidence for traffic-associated risks for childhood cancers.

A small case–control study of leukemia was conducted in the province of Varese, Italy.13 Main roads within 300 m of the home address at the time of diagnosis were considered in the exposure assessment. Benzene concentration was calculated on the basis of traffic density on surrounding roads and the distances from the home address to roads with heavy traffic. The study showed a statistically significant association with a trend across the 3 exposure categories.

A hospital-based French study evaluated proximity of children's residence to roads with heavy traffic and to repair shops and gasoline stations.14 Information on a wide range of exposures was obtained from interviews with the mothers of children with acute leukemia and control children hospitalized for noncancer diseases. Although the study showed no risk association for proximity to roads with heavy traffic, the results suggested increased risks associated with living near a repair garage or a gasoline station in utero (odds ratio: 2.2; 95% confidence interval (CI): 0.9, 5.7) or during childhood (odds ratio: 4.0; 95% CI: 1.5, 10.3).

Figure 1 presents the key results for leukemia from the case–control studies. It shows that the highest relative risk estimates have the widest confidence intervals. The most precise relative risk estimates were all close to 1. Figure 2 shows a similar pattern for the 4 studies that provided relative risk estimates for all childhood cancers combined.

Figure 1. Eight case–control studies on traffic-related exposure and childhood leukemia (VPD, vehicles per day).

Figure 2. Four case–control studies on traffic-related exposure and all childhood cancers combined (VPD, vehicles per day).

Ecological studies

Patterns of childhood cancer in association with various proxy measures of area air quality have been evaluated in studies that offer another perspective on the potential association. The weight of the evidence from these studies also tends to be null.

In 2 early studies, leukemia rates associated with area-specific density of car ownership were investigated. In 3,270 electoral wards of England and Wales, Alexander et al. estimated socioeconomic status, degree of isolation and proportion of households with no car from census data and digital maps.15 The results showed a 30–40% higher risk for acute lymphocytic leukemia in areas with the fewest cars and an even higher risk for the age group 1–7 years (relative risk: 2.1; 95% CI: 1.1, 4.6). A study of the density of car ownership for each of the municipalities of Sweden found only one elevation among several comparisons.17 The incidence rate of acute myeloid leukemia was significantly higher in the municipality group with highest car density, although no dose–response pattern was evident. For all leukemias combined, the rates were almost identical for the lowest and highest exposure groups.

Two studies in the United Kingdom investigated the risk for childhood cancer associated with proximity to emission sources of air pollutants. Knox and Gilman identified potential environmental hazards, including factories and transport routes, near the addresses of children who died of cancer.16 The observed cases were enumerated in a series of areas defined by distance from each type of hazard, and the expected numbers were calculated on the basis of the number of post codes in the same areas. Childhood cancers were reported to be geographically associated with atmospheric pollution from petroleum-derived volatiles and from kiln and furnace smoke and gases. No cancer-specific associations were seen with the different sources, and no association was found with proximity to benzene-emitting plants. Increasing case density ratios were found 300–1,000 m from freeways and decreasing case density ratios 1,000–7,000 m from freeways. These results are inconsistent with a causal interpretation as air pollution from ground sources, such as traffic, is diluted, to reach local background concentrations 100–200 m from the source.18, 24 In a recent study based on the same case-series, Knox found 2–3 times higher migration outward than inward across a 1 km circle around hot spots emitting gaseous air pollutants.25

Harrison et al. investigated childhood leukemia in association with exposure to benzene on main roads and gasoline stations.18 Among the total of 130 children with leukemias and 251 with solid tumors identified, 24 children with leukemia and 31 children with solid tumors had lived near a main road, and 8 children with leukemia and 8 with solid tumors had lived near a gasoline station. The incidence rates for the District Health Authority as a whole and the estimated childhood population at risk in the much smaller post-code districts near a benzene source were used to calculate the expected numbers of cases near the benzene sources. The incidence ratios were slightly increased for leukemia and slightly decreased for solid tumors in relation to proximity to both main roads and gasoline stations.

In a large population-based study, rate ratios were estimated for vehicle density, road density and traffic density for block groups of California, USA.19 There was little or no evidence of rate differences in areas characterized by heavy traffic or vehicle or road density. In a subsequent exploratory study covering almost the same study population, Reynolds et al. studied airborne toxic agents from multiple sources using census tracts as the area unit.20 The rate ratios for all types of cancer, leukemias and gliomas were all close to 1 and showed no dose–response pattern in association with hazardous air pollutants from mobile and area sources. The rate ratio was 1.32 (95% CI: 1.11, 1.57) for leukemia and 1.13 (95% CI: 1.03, 1.23) for all cancers combined when the 10% of census tracts with the highest exposure from point sources was compared with the 25% of census tracts with the lowest exposure. A significant trend was found for these results for point sources.

In a study of cancer incidence and residential traffic intensity in Amsterdam, Netherlands, population-based registers were used to identify cancer cases and residences along roads.21 On the basis of only 5 cases of acute lymphocytic leukemia, the authors reported an elevated incidence ratio for this cancer among children residing along the busiest main roads. Because of the small number of cases and the large number of statistical tests, this result is largely uninterpretable.

Indices of quality

Table II summarizes the studies with respect to key methodological issues. The choice and the relative weight given to categories are subjective, but nevertheless, the scores allow an overview of important methodological aspects. Positive results were found predominantly in studies scoring lower on these criteria and negative results mainly in those scoring higher. Later we discuss each of the methodological issues summarized in Table II and systematic differences between positive and negative studies.

Table II. Characteristics of The 15 Studies

Study	Study mainly supports (+) or refutes (−) an association	More than 200 cases of childhood leukemia	Low potential for selection bias1	Exposure assessment method				Match or adjustment for age and sex	Further match or adjustment of reported results
				Based on total address history	Based on precise address (as opposed to area)	Estimate at address more sophisticated than traffic counts (e.g. distance-weights)	Validation of method reported
1 For ecological studies, zero score refers to uncertainty about population denominator in small-area units or movement patterns among noncancer children. 2 Adjustment for a number of variables in a subset of data had little effect on the risk estimates. 3 Matched on year of birth and stratum-specific analyses for age and sex. 4 No association with traffic at residence but association with repair garage or petrol station near the dwelling of the child. 5 Not specified whether expected numbers based on age- and sex-specific rates. They probably were. 6 Adjustment at ecological level.
Case–control studies
Wertheimer and Leeper (1979)9	+	0	1	0	1	0	0	0	0
Savitz and Feingold (1989)M6	+	0	0	0	1	0	0	1	02
Feychting et al. (1998)10	+	0	1	0	1	1	0	13	1
Raaschou-Nielsen et al. (2000)11	−	1	1	1	1	1	1	1	1
Langholz et al. (2002)12	−	1	0	0	1	1	0	1	1
Crosignani et al. (2004)13	+	0	1	0	1	1	0	1	0
Reynolds et al. (2004)8	−	1	1	0	1	0	1	1	1
Steffen et al. (2004)14	−4	1	0	1	1	0	0	1	1
Ecological studies
Alexander et al. (1996)15	−	1	1	0	0	0	0	15	16
Knox and Gilman (1997)16	−	1	0	0	0	0	0	0	0
Nordlinder and Jarvholm (1997)17	−	1	1	0	0	0	0	1	0
Harrison et al. (1999)18	−	0	0	0	0	0	0	0	0
Reynolds et al. (2002)19	−	1	0	0	0	0	1	16	16
Reynolds et al. (2003)20	−	1	0	0	0	0	1	16	16
Visser et al. (2004)21	−	0	1	0	1	0	0	1	0

Discussion

Of the 15 studies included, 9 were from various European countries, and 6 were from the USA. Eight were case–control studies and 7 were ecological. The first 2 studies to show associations between road traffic and childhood cancer were both in Denver, USA, and both associations were noted as subsidiary findings in studies of electromagnetic fields.6, 9 One of the 13 later studies provided clear support for the hypothesis,13 one study gave high but unstable point estimates10 and the remaining 11 studies gave mainly negative results. The 2 studies providing new support for the hypothesis were both relatively small, whereas the 5 largest studies all provided negative results with respect to traffic-related air pollution.8, 11, 16, 19, 20 Two of the ecological studies indicated weakly increased risks for childhood cancer among children living near point sources,16, 20 and one case–control study indicated gasoline stations and repair garages as potential risk factors.14

Selection bias

In all 8 case–control studies, the cases appear to have been properly ascertained: from population-based mortality registers,9 cancer registers6, 8, 10, 11, 12, 13 or hospital centers.6, 14 In 5 of the case–control studies, population, birth or health service registers were used for random selection of controls from the same study base that gave rise to the cases,8, 9, 10, 11, 13 and exposure was assessed without contacting the families. The remaining 3 case–control studies included contact with case and control families and selection of controls by random-digit dialing,6, 12 from among friends of cases12 and from among other hospital patients,14 implying potential selection bias. In particular, the use of random digit dialing for control selection tends to result in control subjects with higher socioeconomic status, especially in children.26, 27 This bias was suggested for one of the positive case-control studies,28 as higher socioeconomic neighbourhoods have been shown to have lower traffic density.29 Nonparticipation might have introduced selection bias in one of the studies (8% of cases, 25% of controls)6 and to a lesser degree in another (8% of cases, 10% of controls)12 but not in the third study (1% of cases, 2% of controls).14 In the first of these 3 studies,6 there appears to have been a deficit of controls of low socioeconomic status, which might have resulted in a spurious association with a marker of low socioeconomic status like residential traffic density.3 In the other 2 studies, cases and controls were similar with respect to a number of demographic variables.12, 14 Thus, selection bias is a concern in one of the studies with positive results,6 to a lesser degree in another 2 studies12, 14 and not in the remaining 5 case–control studies.8, 9, 10, 11, 13

Ecological design

While ecological studies offer another perspective by virtue of their focus on general area exposure, they also suffer from methodological challenges. Small area studies are subject to uncertainty in the magnitude and composition of populations at risk. For 2 of the British ecological studies,16, 18 the actual number of person-years at risk within the small-area units was unknown providing uncertain expected numbers of cases and for one of the Californian studies,20 the precise population at risk is uncertain because growth rate factors, 1988–1994, for the Californian population as whole was used to adjust the 1990 population data for each census tract.

Exposure assessment

The main methodological challenges in studying this hypothesis are in assessment of exposure. Benzene is one of the most prevalent traffic-related air pollutants and occurs in the urban atmosphere due to evaporation from and incomplete combustion in gasoline engines and to evaporation in gasoline stations during refueling of cars. The importance of ambient concentrations and proximity to gasoline engines in the exposure of children to benzene has been documented.30 Studies of occupational exposure have shown that benzene causes acute myeloid leukemia31 and other subtypes of leukemia in adults32 and may, therefore, cause leukemia in children. The fact that the concentrations of benzene in children's environments are much lower than those in the environment of workers33 is counterbalanced by the possibly greater susceptibility of children.34 Although benzene has been the focus of many studies, air pollution is a complex mixture of many chemicals, and other pollutants might contribute to cancer in children.

Exposure was assessed in relation to residential address in all the studies we included, but the timing of assessment differed. In 2 of the 4 negative case–control studies, exposure was assessed on the basis of all residences occupied between time of conception and time of diagnosis, and risk was analyzed in relation to exposure both in utero and during childhood.11, 14 In the other case–control studies with negative results, exposure at the address at the time of birth8 and the longest residential address during childhood were considered.12 Three of the 4 case–control studies with positive results considered only the residential address at the time of diagnosis or the latest address.6, 10, 13 This difference between the positive and negative studies might be a result of chance or be explained by biological relevance only of exposure close to the time of diagnosis. The age-dependent incidence rates and other evidence indicate that factors early in life contribute to the development of a large proportion of childhood cancers,35, 36, 37 which argues against the last-mentioned interpretation.

The geographical precision of the exposure assessment differed among the studies. In the ecological studies, air pollution was measured in relation to geopolitically defined areas, whereas in the case–control studies air pollution was estimated within a short distance of the child's address. The latter approach may be preferable, because air pollution concentrations can differ substantially within even a small area. None of the ecological studies gave positive results, perhaps due to geographical imprecision in the exposure assessment. Estimated concentrations outside the front door of the home may also not be the best proxy measure for actual personal exposure of children, who move about in many microenvironments within the home, daycare centers, schools and outdoors at many locations around the home. To our knowledge, it is not known which exposure assessment approach best reflects the long-term personal exposure of children to traffic-related air pollution.

In the studies in which exposure at an exact address was assessed, the methods differed, and the crudest being a simple traffic count. More sophisticated approaches included distance-weighted traffic counts and extensive air pollution models with comprehensive input data. Although models rely on the quality of the input data and numerous assumptions and are not necessarily superior to indicator data, they incorporated more extensive information on exposure potential. Of the 5 studies based on air pollution models, 2 gave positive results10, 13 and 3 negative results.11, 12, 20

Although it is not clear which proxy measure best reflects the personal exposure of children, it is clear that more weight should be accorded to results based on a validated exposure assessment method. The 4 studies in which a successfully validated method was used all gave negative results.8, 11, 19, 20

Differential misclassification of exposure

In most of the case–control studies, residential addresses were obtained from registers, municipalities and other objective sources, without regard to case or control status.8, 9, 10, 11, 13 Although the participants themselves reported their residential addresses in 3 case–control studies,6, 12, 14 differential reporting is unlikely. Exposure at addresses was assessed independently of case or control status in all but one of the case–control studies,14 in which information about roads and point sources was obtained by interviews with mothers, implying potential for recall bias. Such bias would be expected to result in spurious associations. This might account for the apparently higher risk for leukemia among children who lived close to gasoline stations or repair shops but would not explain the null result for proximity to roads with heavy traffic.14

Nondifferential misclassification of exposure

Traffic density at the exact address was used as a proxy for air pollution from traffic in many of the studies. Although this proxy is simple, it correlated with half-yearly NO₂ measurements (r = 0.66) in a validation study. The correlation with measurements was stronger for concentrations calculated from an air pollution model (r, 0.75–0.80),38 and application of this model in one of the case–control studies showed a relatively small degree of misclassification.11 In studies in which traffic was assessed in a small area around the residence,8, 12, 19 a correlation of similar magnitude was found between benzene measurements and traffic density (r = 0.69).8, 19

In spite of successful validation of exposure assessment methods, nondifferential misclassification of exposure was inevitable in all the studies. Although under certain conditions such misclassification can result in bias against the null hypothesis,39, 40, 41 we would expect a bias towards a null association.42 Even 50% misclassification between adjacent exposure categories would not be expected to mask a substantial risk increment.43 The studies with successful validation of the exposure assessment method showed no association between air pollution and childhood cancer,8, 11, 19, 20 and, given the expected degree of nondifferential misclassification in one of the studies with negative results,11 we would not expect true relative risks of the magnitude indicated by the studies with positive results6, 10, 13 to have been masked.

Exposure levels and variation

If traffic-related air pollution is a cause of childhood cancer, the positive and negative results might be due to differences in the composition or concentration of air pollution; however, road traffic was the main source studied, and pollution from traffic comprises the same pollutants throughout the western world. An indirect comparison of several study locations suggests that differences in traffic levels are unlikely to explain the different directions of results.8 A direct comparison of benzene concentrations led to the same conclusion: in one of the studies with positive results, conducted in Varese, Italy, 81% of the study participants lived at addresses with outdoor concentrations below 0.1 μg/m³, 16% between 0.1 and 10 μg/m³ and 3% above 10 μg/m³.13 In the United Kingdom, typical concentrations have been estimated to be 1.3 μg/m³ in rural areas, 4 μg/m³ in background urban air and 33 μg/m³ adjacent to roads with heavy traffic.33 In Denmark, the median concentration measured outside children's residences was 9 μg/m³ in Copenhagen and 2 μg/m³ in rural areas.30 Measurements of background levels in urban California between 1989 and 1997 showed that 50% of the measurements were below 4 μg/m³, 40% between 4 and 10 μg/m³ and 10% above 10 μg/m³.44 Altogether, neither the variation in nor the level of exposure to benzene appears to have been higher in Varese than in the settings of several studies with negative results.

Confounding

Aside from sex and age, few risk factors have been established for childhood cancer. These include some known inherited genetic alterations, intrauterine and postnatal exposure to ionizing radiation, treatment of pregnant women with diethylstilbestrol, Epstein-Barr virus and probably hepatitis B virus. Many factors have been suggested including, among others, other infections, parental occupational exposures (in particular to hydrocarbons and infections), paternal preconceptional smoking, maternal smoking, birth order, maternal age, birth weight, (unknown) factors related to socioeconomic status of the family and the residential community, (unknown) factors related to degree of urbanization, diet, medication, electromagnetic fields and radon.3 Among the established risk factors, age might be the most important potential confounder in studies of air pollution because of the strong age-association for incidence rates for several common childhood cancers and because children's age is probably associated with residential mobility and possibly with traffic measures at the residence. Results for associations between the suggested risk factors and childhood cancer are inconsistent and tend to be weak. Socioeconomic status is likely to be associated with traffic density at the residence29 and probably also with several other suggested risk factor such as parental smoking and occupational exposures, maternal age, diet, and radon in the dwelling. If socioeconomic status were also associated with childhood cancer, it might confound estimated associations between traffic-related air pollution and childhood cancer. However, most studies have shown no association between socioeconomic status of the family and childhood leukemia and studies using markers of the socioeconomic status of the residential area are inconsistent, although most have found an increased risk of childhood leukemia among children living in areas of high socioeconomic status.45 Diagnostic bias has been suggested as the explanation for a possible association between high socioeconomic status and childhood cancer.46

Childhood cancer incidence rates differ markedly by age and sex, but all the case–control studies except one with positive results9 included control for potential confounding by these factors. Furthermore, study results were adjusted for a number of suspected risk factors in all but 3 case–control studies reporting positive results.6, 9, 13 Although it is possible that unmeasured confounders might have contributed to the positive findings, this seems very unlikely because the studies included in the present review very consistently reported that adjustment for a number of potential confounders, including socioeconomic status, had no or minor effect on the risk estimates for air pollution.6, 8, 10, 11, 12, 13, 14, 15, 20

Subtypes of childhood cancer

Childhood cancers are many different illnesses, which cannot be expected to have a common set of causes. Most of the studies on air pollution and childhood cancer report on all childhood cancers combined, leukemias combined or central nervous system tumors combined, whereas the more rare types and subtypes are usually not reported. Among the few studies on air pollution exploring possible associations with relatively rare subtypes of childhood cancer, no significant associations were found for non-Hodgkin's lymphomas, astrocytomas, ependymomas, primitive neuroectodermal tumors and soft tissue sarcomas.6, 11, 17 Two studies indicated an association with acute nonlymphoblastic leukemia or acute myeloid leukemia,14, 17 whereas 2 other studies did not,8, 11 and finally, one study reported an association with Hodgkin's disease,11 which was not confirmed in another large study.19

Few studies report separate results for air pollution and rare types of childhood cancer and a true association, if it exists for these tumor types, may not be detectable if combined with other types of childhood cancer, for which no association exists.

Publication bias

Publication bias is expected to occur if small studies with positive results have a better chance of being published than small studies with negative results. The largest studies on this topic gave consistently negative results. The fact that all the studies with positive results are small is consistent with publication bias.

Power of studies with negative results

In the studies in this review with negative results and the greatest power,8, 11, 14 the upper 95% confidence limit was about or below 1.5, indicating that a higher true relative risk is unlikely in view of the statistical uncertainty.

Parental tobacco smoking

Many pollutants are common to air pollution from traffic and tobacco smoke. The mixed evidence for parental tobacco smoking and childhood cancer3, 47 is consistent with the largely null epidemiological evidence for traffic-related air pollution.

Conclusion

In conclusion, the evidence for an association between traffic-related air pollution and childhood cancer is weak. Few positive findings were obtained in studies with small numbers of cases, unvalidated exposure assessment methods and other methodological limitations. Two case–control studies with high statistical power, successfully validated exposure assessment methods, a substantial exposure gradient and low potential for selection bias and confounding8, 11 provide the strongest evidence against an association. In these studies, however, nondifferential misclassification of exposure might have masked a true but weak positive association. The weight of the epidemiological evidence to date indicates that the risk for childhood cancer is not associated with traffic-related air pollution at the residence. Nevertheless, the limited number of studies, methodological limitations to the studies and the absence of consistency in the results obviate a firm conclusion of no effect.

Acknowledgements

The authors are indebted to Theresa Saunders for assistance with the figures and to Bob Gunier (California, USA) for calculations of benzene concentrations.