Association of childhood cancer with residential traffic density

SAVITZ FEINGOLD L. Association of childhood cancer with residential traffic density. Scand J Work Environ Health 1989;15:360-363. Data from a recently completed case-referent study of child hood cancer were used to explore a possible role of environmental exposures from traffic exhaust. The street addresses of 328 cancer patients and 262 population-based referents were used to assign traffic den sity (vehicles per day) as a marker of potential exposure to motor vehicle exhaust. An odds ratio of 1.7 [95 "10 confidence interval (95 "10 CI) I .0-2.8J was found for the total number of childhood cancers and 2. I (95 "10 CI 1.1-4.0) for leukemias in a contrast of high and low traffic density addresses (~500 versus < 500 vehicles per day). Stronger associations were found with a traffic density cutoff score of ~ 10 000 vehicles per day, with imprecise odds ratios of 3.1 (95"10 CI 1.2-8.0) and 4.7 (95"10 CI 1.6---13.5) for the total number of cancers and leukemias, respectively. Adjustment for suspected risk factors for child hood cancer did not substantially change these results. Though the results are inconclusive, the identified association warrants further evaluation.


Materials and methods
The details of the study methods have been provided elsewhere (10) and are therefore outlined only briefly in this report. The eligible cases consisted of all incident childhood cancers (age of child 0-14 years) diagnosed between I January 1976 and 31 December 1983 among residents of the 1970 Standard Metropolitan Statistical Area (Adams, Arapahoe, Boulder, Denver, and Jefferson counties) of Denver, Colorado. The Colorado Central Cancer Register was complete for the period 1979-I982, and area hospitals provided access to additional cases, which allowed comprehensive ascertainment throughout the study period. Of the cases, 95.2 (l!o were histologically confirmed, and an additional 3. I (l!o were confirmed by direct visualization or radiography (I I).
The referents were selected by random digit dialing, matched to cases by the age (plus or minus 3 years), sex, and telephone exchange area of the patient. Although referents were sought for each case, the inclusion of a given case or referent in the analysis was not affected by our ability to produce a matched pair. The intent was to produce a group balanced on the matching criteria, and the data were analyzed through stratified rather than matched analysis (12).
In order to ensure that the referents occupied their homes in the patient's telephone exchange area at the time of the case's diagnosis, only potential referents who had stayed in their home from the time of the case's diagnosis to the time of the interview (when they were determined to reside in the correct telephone exchange area) were eligible. Although it would have been preferable to include potential referents who previously resided in the area but subsequently moved, there was no mechanism for identifying and contacting such families in their new locations. Thus referents were restricted to have been occupants of their homes at the time the case was diagnosed (whereas cases were not so restricted); therefo re more residentia lly sta ble referents than cases were produced (10).
In this an alysis, the residence occ upied at th e time of the case's diagnosis (or at the referent ' s age when the mat ched ca se was dia gnosed) was characterized with respect to tra ffic den sity. The location of each reported address was ident ified on a detai led street map o f the Denver Standard Metropolitan Statistical Area . On the basis of the precise location, th e approximate loca tion of the homes was identified on traffic den sity maps provided by the Denver Division of Highway Planning and Research, and the specific segment of the street to which the density figure s app lied was car efull y not ed. On th e ba sis of the maps and lists of the streets and traffic vo lume s, the home was assigned a traffic den sity score of < 500 vehicles per da y or the recorded number of vehicles per day (ranging from 500 to over 100 000 per day).
Data on potential confounders were collected principall y thro ugh an interview with the parent (mother preferred) regarding numerou s potential risk factors such as fami ly demography, cancer history, expo sur e to X ra ys and medi cat ions during pregnan cy, parents' occupat ional history, and mother's and child 's illnesses . Wire configuration codes were obtained as a ma rker of long-term magnetic field exposure (10). On the basis of associations with childhood ca ncer (eith er risk factors or artifacts of th e study design), the following variables were examined in detail: sex, age, year of dia gno sis, single-family versus multiple occupancy residence, residence in or out of Denver at birth, residential mobility fro m birth to dia gn osis, mother's age, father's ed ucation, per cap ita income, mo ther' s smoking during pre gnancy, and wire code at the time of diagnosis. (Not e that the markers of residential sta bility would address the possible bias from the restric tion of referents ba sed on long-term occupancy of their homes.) Odds ratios and 95 070 confidence intervals were calculated wit h test-b ased meth ods (13) . Stratified analyses produced Mantel-Haenszel adj usted odds ratios and confidence intervals (14) . Finally, unconditional logistic regression anal ysis was used to control for multiple potential confounders (15). in telepho ne screening (10) to pro duce an overall 75 0J0 response. Nonrespondent cases were mo re likely to have been diagnosed earlier an d be Hispanic or Black; no informat ion was available concern ing no nrespondent referents.
The crude results relating traffic density to childhood can cer inc idence are provided in table 1. T he association for total childhood cancers [odds ratio (OR) 1.7] was slightly greater for leukemia (OR 2. I) though not acu te lymphocytic leuk emia (OR 1.6). Similar effect estimates (O R 1.6-1.8) were ob tained for brai n cancer and other cancers, with a dim inished association for so ft-tissue tumors and the absence of an association for lymphomas . Because the referent selection procedures failed to provide a referent for each case, th e po ssibilit y of bias due to incomplete coverage was exam ined in a matched analysis. Though the preci sion was great ly reduced, the matched odds ratio for all childhood canc ers was 2.0 (bas ed on 36 discordant pair s), similar to the results of the unmatched analysis.
Th e more highl y exposed gro up (::: 500 vehicles/ d) was subdivided for an examination of the pos sible expo sure-response gradients (ta ble 2), although small numbers limited thi s analysis to the tota l number of cancers and leukemias. A cutoff sco re of 5000 vehicles/d yielded odds ratios of 1.8 [95 (J/o confidence interval (95 0J0 C I) 0.9 -3.3] for th e total nu m ber of can cers an d 2.7 (95 0J0 CI 1.3-5.9) for leukemias in the highest exposure group. An even more pronounced gradient was found with a high er cutoff score of 10 000 vehicle s/d, with odds ratios of 3. 1 (95 0J0 C I 1.2-8.0) for total cancers and 4.7 (95 0J0 CI 1.6-13.5) for leukemias in the highe st exposu re group.
As note d earlier, da ta on pote ntial confounders were obta ined in an inte rview with th e subjects ' parents , with adjusted results restricted to the subset of interviewed cases and referents (71 OJ(J of eligible cases and 80 070 of identified referents). Thi s restriction alon e (in the absence of an y co nfou nding) rai sed the odds ratios slightly, especially for leukemia (OR 2.3) and brain cancer (OR 2.6). Adj ustm ent for sex, age, year of diag-  nosis, type of residence (single family, other), location at birth (Denver, other), mother's age, father's education, per capita income, and wire configuration code at diagnosis had little effect except for a tendency of residence type to diminish the odds ratios and mother's age and mother's smoking to elevate the odds ratios. The logistic regression analysis for the total number of cancers, leukemias, and brain tumors with these three variables in the model (plus wire code for leukemia) produced adjusted odds ratios of 1.7 for the total number of cancers (95 070 CI 0.8-3.6), 1.9 for leukemias (95 070 CI 0.7-5. I), and 2.0 for brain tumors (95070 CI 0.7-6.1). Overall, these results provide evidence against substantial confounding by the measured potential risk factors. The sex-and age-specific results for the total number of cancers, leukemias, and brain tumors are presented in table 3. The imprecision within strata is apparent, but there is a suggestion of enhanced associations for the females (especially for brain tumors) and a consistently stronger effect for the 0-to 4-yearold than for the 5-to I4-year-old children. Though only crude results are reported in table 3, logistic regression with adjustment for residence type, mother's age, and mother's smoking did not materially change the effect estimates, although the precision was further reduced.

Discussion
These results indicate an association between traffic density near the home occupied at the time of diagnosis and childhood cancer which was not accounted for by potential confounders. The odds ratios were on the order of 1.6-2.0 for all cancers except lymphomas and soft-tissue tumors. Stronger but less precise associations were observed for females, younger children (age 0-4 years), and residentially stable children. Evidence of increasing risk with increasing traffic density was found for the total number of cancers and leukemias.
Potential sources of bias include nonresponse, differential mobility of cases and referents, unmeasured confounders, and nondifferential exposure misclassification. The precision of many of the odds ratio estimates was less than would be desirable, and random variation must be included among the possible sources of error. In order for nonresponse to have biased the odds ratios, there would have to have been a differential loss of high and low traffic density cases and referents (16). In spite of the differential response for cases and referents, such a pattern is unlikely.
The same reasoning is applicable to the concern for the differential mobility of the cases and referents. The cases were unrestricted by patterns of movement, but only referents who remained in their home from th e tim e of the case 's diagnosis to the time of interview were included . The relationship of residential mobility to traffic density is unknown, although one might specula te that more sta ble residents live in more desirable, lower traffic den sity homes. Referents ca nnot be evaluated directly for th eir representativene ss of th e expos ur e distribution in the study base that generated the cases, but a fter adj ustme nt for a number of potential sources of selectio n bia s, they sho uld pr ovide valid results.
Unmeasured confounders a re always a possibility, especia lly when so little is known about the determinants of child hood ca ncer. For example, a variety of env ironmental a nd social factors is ass oc iated with living in rural versus urban settings . An y co m po nent of living in an urban area th at is predicti ve of childhood cancer would confound the results for traffic density. Limited data on childhood leukemia do not support strong gradients by rural/urban status, with a pproximately a 10 070 increase in childhood leukemia mortality in urban areas found in on e study (17) and a 10 % decrease in childhood leukem ia mortality repo rted in another (18). These data ar gue against substa ntial confounding by some unmeasured correlate of urban residence.
Finally, the inherent limitations in the marker of traffic density should be noted . Whether thi s is a surrogate for ambient benzene or so me other co m po nent or a correlate of tra ffic exhaust, it contains substa ntial mi scla ssification when co m pared with so me more direct determinant of child hood cancer . Since onl y the street address was coded , nearby high-traffic st ree ts were not included. The result ing mi scla ssification is almost certain to be nondifferential with respect to the case-referent status a nd would thus bia s the odds rat ios toward s the null (19) .
Becau se of th ese limitations and the no velt y of thi s observat ion , the study results mu st be interpreted very cautiou sly. Spe cificall y, th ese data do not stro ngly implicate tra ffic-related air pollution in general (or benzene in particular) in the et iology of childhood ca ncer. Nonetheless, th e data do suggest that a n association is present bet ween high tr affic den sit y and childhood cance r. There is no ob viou s meth odological flaw which indicates that these result s are spurio us. Given th e limited knowledge of risk factors fo r child ho od ca ncer, furt her research is warranted to wa rds identifying a n environmental factor for whi ch traffic density might serve as a marker . If the observed results a re actually reflective of an underlying cau sative relation , then mor e precise mea sures of hypothesized etiologic age nts (eg, benzene) sho uld produce mu ch stro nger associations.