Ambient Air Pollution and Autism in Los Angeles County, California

Background: The prevalence of autistic disorder (AD), a serious developmental condition, has risen dramatically over the past two decades, but high-quality population-based research addressing etiology is limited. Objectives: We studied the influence of exposures to traffic-related air pollution during pregnancy on the development of autism using data from air monitoring stations and a land use regression (LUR) model to estimate exposures. Methods: Children of mothers who gave birth in Los Angeles, California, who were diagnosed with a primary AD diagnosis at 3–5 years of age during 1998–2009 were identified through the California Department of Developmental Services and linked to 1995–2006 California birth certificates. For 7,603 children with autism and 10 controls per case matched by sex, birth year, and minimum gestational age, birth addresses were mapped and linked to the nearest air monitoring station and a LUR model. We used conditional logistic regression, adjusting for maternal and perinatal characteristics including indicators of SES. Results: Per interquartile range (IQR) increase, we estimated a 12–15% relative increase in odds of autism for ozone [odds ratio (OR) = 1.12, 95% CI: 1.06, 1.19; per 11.54-ppb increase] and particulate matter ≤ 2.5 µm (OR = 1.15; 95% CI: 1.06, 1.24; per 4.68-μg/m3 increase) when mutually adjusting for both pollutants. Furthermore, we estimated 3–9% relative increases in odds per IQR increase for LUR-based nitric oxide and nitrogen dioxide exposure estimates. LUR-based associations were strongest for children of mothers with less than a high school education. Conclusion: Measured and estimated exposures from ambient pollutant monitors and LUR model suggest associations between autism and prenatal air pollution exposure, mostly related to traffic sources.

Research | Children's Health Autistic disorder (AD) is a serious develop mental condition characterized by impairments in social interaction, abnormalities in verbal and nonverbal communication, and restricted stereotyped behaviors thought to be attribut able to insults to the developing fetal and/or infant brain (American Psychiatric Association 2000; Geschwind and Levitt 2007). The preva lence of autism has risen for the past 20 years, partly due to changes in case definition and improved case recognition. HertzPicciotto and Delwiche (2009) suggested the observed rise in incidence in California between 1990 and 2001 may partially but not fully be explained by younger age at diagnosis (12% increase) and inclusion of milder cases (56% increase). Although evidence for genetic contributions is considered quite strong, twin concordance research recently suggested that environmen tal causes are also important (Hallmayer et al. 2011), and it is quite conceivable that mul tiple genes interact with environmental fac tors (Cederlund and Gillberg 2004;Glasson et al. 2004).
Few studies to date have examined the impact of air pollution on brain develop ment in general during pregnancy, although air pollution exposure during the prenatal period has been associated with a variety of adverse birth outcomes (Ritz and Yu 1999;Ritz et al. 2000;Srám et al. 2005;Williams et al. 1977) and neuropsychological effects later in childhood (CalderónGarcidueñas et al. 2008;Edwards et al. 2010;Perera et al. 2006Perera et al. , 2012Suglia et al. 2008;Tang et al. 2008;Wang et al. 2009). The biologi cal mechanisms by which air pollution may cause autism are largely unknown, although the immune system has been implicated as possibly playing a role (HertzPicciotto et al. 2008). Only three studies to date have exam ined associations between autism and air pol lution exposures during the prenatal period (Kalkbrenner et al. 2010;Volk et al. 2010;Windham et al. 2006). In one study, autism was associated with ambient air concentra tions of chlorinated solvents and heavy metals near birth residences (Windham et al. 2006). Another study of autism reported elevated odds ratios (ORs) for methylene chloride, quinoline, and styrene exposures in ambient air, but nearnull effect estimates for ambient air metals and other pollutants (Kalkbrenner et al. 2010). A third study reported that chil dren born to mothers living within 309 m of a freeway during pregnancy were more likely to be diagnosed with autism than children whose mothers lived > 1,419 m from a free way (Volk et al. 2010).
We derived air pollution exposure mea sures using data from government air moni toring stations that provide information on spatial and temporal variations in criteria pol lutants, and from a land use regression (LUR) model we developed for the Los Angeles Air Basin. The LUR model allowed us to greatly improve our spatial characterization of traffic related air pollution. Because heterogeneity of the autism phenotype and its severity may be attributable to influences on different critical gestational windows of brain development (Geschwind and Levitt 2007), we also season alized these traffic measures to investigate vul nerable trimesters of development. Here we examine associations between measured and modeled exposures to prenatal air pollution and autism in children born to mothers in Los Angeles County, California, since 1995.

Methods
In this populationbased case-control study, our source population consisted of chil dren born in 1995-2006 to mothers who resided in Los Angeles County at the time of giving birth.
Case ascertainment and definition. In Los Angeles, children with autism are identified through seven regional centers, contracted by the California Department of Developmental Services (DDS), whose staff determine eligi bility and coordinate services in their respec tive service areas. Cases are children given a primary diagnosis of AD, the most severe among the autism spectrum disorders (ASD) diagnoses, between 36 and 71 months of age at a Los Angeles Regional Center dur ing 1998-2009. During our study period, eligibility for DDS services did not depend on citizenship or financial status-services were available to all children regardless of socioeconomic, health insurance status, or racial/ethnic identification. Referrals to the regional centers are usually made by pediatri cians, other clinical providers, and schools, but parents may also selfrefer their children.  Services 1986Services , 2007.
Record linkage. We attempted to link 10,821 DDS records of children with autism to their respective birth records using the National Program of Cancer Registries Registry Plus TM Link Plus Software [Centers for Disease Control and Prevention (CDC) 2010a]. Given the child's first and last name, birth date, and sex; mother's first and last name and birth date; and father's last name and birth date, we probabilistically matched the two records and reviewed all high scor ing linkages (≥ 25), almost half of the link ages (9,120 of 22,806), only accepting those manually confirmed to be likely matches (see CDC for record linkage concepts) (CDC 2010b). The remaining lower scoring linkages were reviewed using SAS version 9.2 (SAS Institute Inc., Cary, NC) and accepted on the condition that the child's first and last name, and birth date matched perfectly. We correctly linked 8,600 DDS records (79.5% of all cases) to birth records. Of the 2,221 DDS records not linked to CA birth records, 35% were not born in Los Angeles County, 46% were missing birthplace information, and only 19% recorded the child as born in Los Angeles County. The most common rea son for nonlinkage was missing or incomplete linkage information on either of the records.
From among linked cases, we further excluded children whose mother's residency was outside of Los Angeles County during her pregnancy (n = 41), records with missing or implausible gestational ages (< 21 or > 46 weeks) or birth weights (< 500 g or > 6,800 g) (n = 508), and cases who did not have a pri mary diagnosis of AD (n = 448), leaving a final sample of 7,603 children with autism successfully linked to a birth certificate who met all inclusion criteria.
Control selection. We selected 10 con trols for each case from our source popula tion. Using birth certificates, each control was randomly selected without replacement and matched on birth year and sex. In addition, each control's gestational age at birth had to be equal to or greater than the gestational age at birth of their matched case to ensure prena tal exposures could be estimated for compa rable lengths of time. Children were eligible as controls if they had no documentation of autism-did not have a DDS record in Los Angeles County by 2009, had a plausible gestational age (21-46 weeks inclusive) and birth weight (500-6,800 g inclusive), and the mother resided in Los Angeles County at the time of birth.
Matching by birth year balanced the large increase in autism rates during the case ascer tainment period, 1998-2009. The matched control set included 76,030 children born dur ing 1995-2006. From among these, we further excluded 248 control children who died before 6 years of age (71 months) based on California death records, leaving 75,782 controls.
Residential locations at delivery that were reported on birth certificates were mapped using a custom geocoder (Goldberg et al. 2008), and further exclusions were necessary if residential addresses were not geocodable (9 cases, 147 controls) [see Supplemental Material, Exposure assessment. Using measurements for the criteria pollutants carbon monoxide (CO), nitrogen dioxide (NO 2 ), nitric oxide (NO), ozone (O 3 ), and particulate matter concentrations with an aerodynamic diameter ≤ 10 µm (PM 10 ) and ≤ 2.5 µm (PM 2.5 ) from nearest monitoring stations, we estimated average exposures for the entire pregnancy and for three specific periods during pregnancy based on the birth date and gestational age reported on the birth certificate: first trimester (estimated first day of last menstrual period through day 92), second trimester (days 93-185), and third trimester (day 186 to date of birth). The length of each pregnancy aver aging period for controls was the same as for their matched case: Averaging periods for each autistic risk set were truncated at the gesta tional age of the matched case at birth. Hourly measurements for CO, NO 2 , NO, and O 3 (1000-1800 hours) were first averaged for each day if sufficient data were available [for details, see Supplemental Material, Table S2 (http://dx.doi.org/10.1289/ehp.1205827)]. Daily averages for the gaseous pollutants and 24hr measurements of PM 10 and PM 2.5 (col lected every 6 and 3 days, respectively) were then averaged over the different pregnancy periods when data were sufficient to do so (see Supplemental Material, Table S2).
To classify prenatal exposures to traffic related pollutants on a more spatiallyresolved scale, we extracted NO and NO 2 concen tration estimates at each residential location from the LUR model surfaces we developed for the Los Angeles Air Basin (Su et al. 2009).

This LUR model was based on approximately 200 measurements of outdoor air pollution taken during 2006-2007 in locations across
Los Angeles County, in addition to predic tors of traffic exhaust concentrations (such as traffic counts, truck routes, and roadways). The model explained 81% and 86% of the variance in measured NO and NO 2 concen trations, respectively (Su et al. 2009).
The LUR models most closely approxi mate annual average concentrations. Thus, in addition to using the LUR annual average ("unseasonalized") estimates, we also gener ated "seasonalized" estimates to incorporate yearly and monthly air pollution variations. Specifically, using ambient air monitoring data for NO and NO 2 at the closest monitor ing station, the LUR estimates were adjusted to represent pregnancy month-specific LUR values by multiplying the LUR (unseasonal ized) estimates for NO and NO 2 by the ratio of average ambient NO and NO 2 during each pregnancy month to annual average ambient NO andNO 2 (2006-2007). These seasonal ized monthly LUR values were then averaged over each pregnancy period. We applied the same exclusion criteria for missing values as described above when generating the preg nancy month scaling factors using the govern ment monitoring data.
Statistical analysis. We calculated Pearson's correlation coefficients to examine relations between the various pollutant mea sures. Associations between air pollution expo sure and odds of AD diagnosis were examined using one and twopollutant models. We adjusted for LUR estimates of trafficrelated exposures in our monitorbased pollutant models and assessed particles and the gaseous pollutant ozone together in the same model. We calculated ORs and 95% CIs using condi tional logistic regression to estimate increases in odds of AD per interquartile range (IQR) increase in pregnancy exposures, based on exposure distributions in the controls.
We adjusted for potential confounders for which data were available on birth certificates based on prior knowledge (see Table 1 for cat egories used in models): maternal age, maternal place of birth, race/ethnicity, and education; type of birth (single, multiple), parity; insurance type (public, private, or other, a proxy for socio economic status); and gestational age at birth (weeks). In addition, we estimated pollutant effects without adjustment for gestational age to allow for the possibility that this factor might be an intermediate and thus on the causal pathway between air pollution and autism.
We expected maternal education to cor relate with estimates of air pollution and autism (Ponce et al. 2005), so we also used unconditional logistic regression models to estimate associations stratified by maternal education (less than high school, high school, volume 121 | number 3 | March 2013 • Environmental Health Perspectives more than high school) controlling for the matching variables (birth year, sex, and gesta tional weeks at birth) in addition to the other covariates noted above.

Results
Both mothers and fathers of children with autism were older and more educated than parents of control children, and mothers were more often nonHispanic white but less often Hispanic, especially foreignborn Hispanic (Table 1). A higher percentage of mothers of case children were primiparous and had mul tiple gestations. As expected, children with autism had a lower mean gestational age at birth and birth weight than control children. Of the children with autism not linked to a Los Angeles County birth record, parental characteristics were undetermined because of frequent missing information-50-60% missing maternal and paternal age/birthday (results not shown). However, of these non linked DDS records, 42% of families were Hispanic (results not shown), comparable to the 41.9% of Hispanic mothers of case chil dren included in this study (Table 1).
We estimated 4-7% relative increases in odds of an AD diagnosis per IQR increase in unseasonalized LUR measures of NO and NO 2 in adjusted models (Table 2). These OR estimates remained similar (1.03 to 1.09) in twopollutant adjusted models (Table 3). ORs for autism per IQR increase in monitor based estimates of entire pregnancy exposure to NO and NO 2 were slightly smaller than associations with IQR increases in LUR based estimates (Table 2). We also estimated increases in odds of AD diagnosis per IQR increase in entire pregnancy exposure to ozone (OR = 1.06; 95% CI: 1.01, 1.12) and PM 2.5 (OR = 1.07; 95% CI: 1.00, 1.15) ( when we mutually adjusted for both pollutants (Table 3). In addition, without adjustment for gestational weeks at birth, associations increased further or remained the same; for the twopollutant models including ozone and PM 2.5 (O 3 OR = 1.14; 95% CI: Adjusting for maternal education changed air pollution effect estimates most strongly, likely because socioeconomic status is strongly associated both with air pollution exposure and autism diagnosis. We also investigated potential effect measure modification of the air pollution and autism association: We exam ined whether air pollution effect estimates vary according to strata of maternal education pos sibly due to differences in vulnerability, in actual exposure, or exposure and outcome mis classification. Generally, LURbased trafficre lated pollutant estimates showed the strongest association with autism in children of the least educated mothers, compared with mothers in the highest educational stratum (Table 4).

Discussion
We estimated an approximately 3-9% rel ative increase in the odds of AD per IQR increase in entire pregnancy exposure to NO (9.40 ppb) and NO 2 (5.41 ppb) as estimated by our twopollutant LUR models. Our LUR model was built upon neighborhoodlevel measures of nitrogen oxides (NO x ) and rep resents smallerscale variability in exhaust pol lutants, compared with estimates based on air monitoring station measurements (Zhou and Levy 2007). We also estimated a 5-15% relative increase in the odds of AD per IQR increase in entire pregnancy exposure to PM 2.5 (4.68 µg/m 3 ) (Table 3), a pollutant whose concentrations are driven partly by fossil fuel combustion in motor vehicles. In addition, an 11.54ppb increase in O 3 expo sures during pregnancy was associated with a 6-12% relative increase in the odds of having a child diagnosed with autism.
Few studies have previously examined associations between air pollution-related exposures during the prenatal period and later development of autism, and none used ambient air monitoring data or LUR mod els to estimate risk in a large population. A relatively small study (284 cases, 657 con trols) in the San Francisco Bay, California, area used studyspecific census tract pollution    Volk et al. 2010). Trimesterspecific addresses were geocoded, and measures of distance to freeways and major roads were calculated using geographic information sys tem software. This small study was the first to suggest that trafficrelated exposures might increase the risk of autism. In our study, we observed weaker associations with monitor based and modeled air pollution exposure estimates in a much larger study population. Gestational toxicity may plausibly result from maternal exposure to NO 2 , which has been shown to disturb early neuromotor development in animals, causing coordina tion deficits and reduced activity and reactivity in rats (Tabacova et al. 1985); specifically, NO 2 exposure at low (0.05-0.10 mg/m 3 ) and high (1 and 10 mg/m 3 ) concentrations for 6 hr each day throughout gestation affected neuromotor development in offspring. The mean NO 2 level in our study (30.8 ppb) [see Supplemental Material, Table S3 (http:// dx.doi.org/10.1289/ehp.1205827)] falls within the exposure range classified as "low" in this animal study (0.05-0.10 mg/m 3 or 26.6-53.2 ppb). Beckerman et al. (2008) suggested that NO may be a proxy measure for ultra fine particle (UFP; < 0.1 µm in aerodynamic diameter) exposures from traffic exhaust and reported strong correlations between 1week average concentrations of NO, NO 2 , and NO x and shortterm (10 min) measures of UFP (r = 0.8-0.9) at varying distances from a major expressway in Toronto, Canada. Fine particles (PM 2.5 ) can cause oxidative stress, and in vitro animal and human postmortem brain stud ies showed they can trigger cellular toxicity and brain cell pathology (Lai et al. 2005;Li et al. 2003, Peters et al. 2006). HertzPicciotto et al. (2005 found that maternal PM 2.5 expo sures 2 weeks before birth were associated with altered lymphocyte immunophenotypes, and suggested that this might mediate effects of air pollution on childhood morbidity. Developmental immune system disruption has been hypothesized to play a role in neurobe havioral disorders such as autism, considering the close connection between the development of the immune system and the central nervous system (HertzPicciotto et al. 2008).
To our knowledge, this is the first study to suggest associations between ozone and AD. Although O 3 levels have dropped over the last decade, the Los Angeles region still often has the highest levels of O 3 nationwide, violating federal health standards an average of 137 days/year (averages from 2007 through 2009) (Roosevelt 2011). In contrast with the trafficrelated and particle associations that became positive only when we adjusted for maternal education, O 3 effect estimates moved closer toward the null after adjustment for covariates. This is consistent with expectations, because trafficrelated pollution is higher in lowerSES (socioeconomic status) neighbor hoods, whereas O 3 levels are higher in subur ban highSES areas, and autism is more likely to be diagnosed earlier in children of mothers with higher SES. Specifically, O 3 and NO fol low opposite distribution patterns across the Los Angeles Air Basin. O 3 is formed by photo chemical reactions in the presence of precursor pollutants from exhaust, and concentrations are low near freeways/roadways (due to pres ence of strong NO emission sources) and higher in suburban neighborhoods ). Controlled animal studies suggest that O 3 may cause adverse neurobehavioral effects after gestational exposure (Kavlock et al. 1980;Petruzzi et al. 1995;Sorace et al. 2001).
We relied on information recorded on California birth certificates to adjust for potential confounding by prenatal risk factors for autism reported in the literature (Gardener et. al. 2009(Gardener et. al. , 2011-parental age at birth, parity, maternal place of birth, and multiple births. However, we were unable to control for potential confounding due to maternal physical and mental health history, or mater nal active or passive smoking. Women giv ing birth in Los Angeles are predominantly Hispanic, and our survey of 2,543 women giving birth in Los Angeles County in 2003 found that only 1% of foreignborn Hispanic, 5% of U.S.born Hispanics, and 7% of non Hispanic whites were active smokers dur ing pregnancy (Hoggatt et al. 2012). Also, a recent study found no association [prev alence ratio = 0.88 (95% CI: 0.72, 1.08)] of maternal smoking during pregnancy with AD (Kalkbrenner et al. 2012). Confounding by other SESrelated factors potentially cor related with air pollution is also a concern. Families of lower SES are more likely exposed to air pollution, and less likely represented in the autism case group, possibly due to under ascertainment (Durkin et al. 2010;Grineski et al. 2007; Institute of Medicine 1999), which could have potentially biased our effect estimates toward the null. However, we esti mated stronger associations among those with the lowest maternal education for LURbased estimates of NO and NO 2 . We adjusted for type of insurance (public vs. private pay), as well as other SES indicators important in the Los Angeles community (i.e., maternal place of birth and education) because we previ ously showed that these factors were sufficient to adjust adequately for SES in Los Angeles County birth outcome and air pollution stud ies; effect estimates for air pollution and birth outcomes were very similar when we adjusted for maternal occupation, income, and educa tion or simply for birth certificate-derived SES measures (Hoggatt et al. 2012).
In addition to being a confounder, ges tational age at birth may also be a mediator between air pollution and autism. In analyses not adjusting for gestational weeks at birth we estimated larger or similar effect sizes. However, not adjusting for gestational age at birth may also result in biased estimates because of our matching design. Specifically, because controls were sampled from among children who at birth had reached at mini mum the gestational age of the matched case, gestational age as a matching variable required that we analytically control for it. Thus the magnitude and direction of any potential bias from adjusting or not adjusting for gestational age at birth is not easily quantifiable.
A source of exposure measurement error is the reliance on address information reported on birth certificates, which does not account for women who worked far from home or res idential mobility during pregnancy. Previous U.S.based studies (1997)(1998)(1999)(2000)(2001)(2002)(2003)(2004) indicate that 15-30% of women change residence dur ing pregnancy Lupo et al. 2010). In our previous populationbased survey of 2,543 women residing in 111 ZIP codes in Los Angeles County and deliver ing in 2003, 22% reported moving during pregnancy (Ritz et al. 2007). Our survey also found pregnant women of lower SES less likely to be employed and more likely to spend time near their residence, suggesting exposure is less misclassified for lower com pared with higherSES women.
Distance from a monitoring station likely introduced some nondifferential mis classification of exposure, especially for pol lutants such as CO and NO 2 that are more heterogeneously distributed. On average, the distance between home addresses and the nearest monitoring station was 6.7 miles in our study, and monitorbased estimates of CO, NO, and NO 2 are questionable in their validity if air pollution measurements are more accurate representations of actual exposures for women living closer to a station (Ghosh et al. 2012;Wilhelm et al. 2011). Ambient station measures for PM 2.5 and O 3 , however, are less likely to misrepresent actual exposures, because these pollutants are generally considered more homogeneously distributed over larger regions.
LURderived NO and NO 2 estimates are much more spatially resolved than monitor based estimates, and were previously associ ated with adverse pregnancy outcomes in the same Los Angeles population (Ghosh et al. 2012;Wilhelm et al. 2011). Our LUR model not only represents local trafficrelated pol lution well, it reduces possible confounding by spatial SES factors. For example, autism diagnoses have been reported to vary spatially in California due to SES (Van Meter et al. 2010), but measures of air pollution are not inherently influenced by these spatial factors related to SES ). For pollutants that are more homogeneous over larger regional areas, such as PM 2.5 and O 3 , confounding due to SES is possible; never theless, associations were stronger when we mutually adjusted for both pollutants.
A major strength of our study was the use of our novel LUR exposure measures for trafficrelated pollution in addition to routine, government monitoring station data for crite ria pollutants to help identify specific emis sions of concern for autism. Furthermore, selection bias due to participation is unlikely to have occurred.

Conclusions
The observed association with the LUR model estimates and monitoring station-based O 3 and PM 2.5 measures suggest a link between AD and trafficrelated exposures during preg nancy. Ideally, future autism and air pollu tion studies should use neighborhoodlevel monitoring or modeling of air toxins such as polycyclic aromatic hydrocarbons and pos sibly speciated PM 2.5 to determine whether these results are reproducible with improved air pollution assessment.