Methods
Participants
The study population included offspring of participants in NHS II, a prospective cohort of 116,430 U.S. female nurses 25–43 years of age when recruited in 1989, followed biennially (Solomon et al. 1997). NHS II participants originally were recruited from 14 states in all regions of the continental United States, but they now reside in all 50 states. The study was approved by the Partners Health Care Institutional Review Board and complied with all applicable U.S. regulations; return of completed questionnaires constituted consent to participate.
In 2005, NHS II participants were asked whether any of their children had been diagnosed with autism, Asperger's syndrome, or "other autism spectrum," and 839 women replied affirmatively. In 2007, we initiated a pilot follow-up study, shortly followed by a full-scale follow-up as described previously (Lyall et al. 2012). The follow-up questionnaire included questions about the pregnancy and birth, child's sex, and diagnosis. NHS II protocol allows re-contacting only the nurses who responded to the most recent biennial questionnaire. Thus, this follow-up was attempted with the 756 mothers of ASD cases for whom this was the case. Mothers who reported having more than one child with ASD were directed to report about the youngest one. Controls were selected from among parous women not reporting a child with ASD in 2005. For each case mother, controls were randomly selected from among those women who gave birth to a child in a matching birth year, to yield a total of 3,000 controls. Six hundred thirty-six (84%) mothers of cases and 2,747 (92%) mothers of controls responded; 164 women (including 51 case mothers) declined to participate.
For the current study, only children whose estimated conception month was June 1989 or later were included because nurses' addresses before this month were unknown. Of the 265 children reported to have an ASD diagnosis who met this criterion we excluded 4 for whom ASD was not confirmed by the mother on the follow-up questionnaire, and another 2 with genetic syndromes associated with ASD (n = 1 Down syndrome; n = 1 Rett syndrome). The remaining 259 children were classified as ASD cases. There were 1,640 control children who met the conception month criterion. We further excluded participants missing PM data because their addresses could not be geocoded (8 cases and 30 controls), controls who were reported to have ASD on the 2009 questionnaire (n = 9), and children missing data on birth month (6 cases and 79 controls). The final study sample included 245 cases and 1,522 controls born 1990 through 2002. The average (± SD) year of diagnosis of the ASD cases was 1999 ± 3.3. None of these children were reported to have been adopted. Of 188 ASD cases with data on ASD in siblings, 7.4% were reported to have a sibling with ASD. Analyses excluding those 7.4% were similar to analyses including all children and are therefore not reported.
Case Validation
SD diagnosis was validated by telephone administration of the Autism Diagnostic Interview–Revised (ADI-R) (Lord et al. 1994) in a subsample of 50 cases randomly selected from mothers who indicated on our follow-up questionnaire willingness to be contacted (81% of all case mothers). In this sample, 43 children (86%) met full ADI-R criteria for autistic disorder [which is stricter than the broader "autism spectrum disorder" of the current DSM-V (Diagnostic and Statistical Manual of Mental Disorders, 5th Edition) criteria, or other autism spectrum disorders including PDD-NOS (pervasive developmental disorder not otherwise specified) or Asperger syndrome of DSM-IV criteria], defined by meeting cutoff scores in all three domains (social interaction, communication and language, restricted and repetitive behaviors) and having onset by 3 years of age. The remaining individuals met the onset criterion and communication domain cutoff and missed the autistic disorder cutoff by one point in one domain (n = 5; 10%), or met cutoffs in one or two domains only (n = 2; 4%), thus indicating presence of ASD traits [for further details on scoring of ADI-R, see Lord et al. (1994)]. In addition, Social Responsiveness Scale (SRS) scores (Constantino et al. 2000), obtained for approximately 90% of eligible cases, also indicated accuracy of case ascertainment. Although it is not a clinical diagnostic instrument, the SRS is a widely used measure of social functioning and autistic traits, and has been shown to have excellent validity as compared to ADI-R and ADOS (Autism Diagnostic Observation Schedule) (Constantino et al. 2003). Among our ASD cases, 93% met the SRS cutoff for ASD. In contrast, 93% of controls completing the same measure fell within the normative range. Therefore, both ADI-R and SRS scores support reliable ASD case ascertainment in our population. For all analyses only the maternal reports were used for determination of ASD status.
Exposure Assessment
Residential locations of the nurses were determined from the mailing addresses used for the biennial NHS II questionnaire. Monthly ambient exposure predictions of airborne particulate matter with an aerodynamic diameter ≤ 10 μm (PM10) and ≤ 2.5 μm (PM2.5) were generated from nationwide expansions of previously validated spatiotemporal models (Yanosky et al. 2008, 2009, 2014). The models use monthly average PM10 and/or PM2.5 data from the U.S. EPA's Air Quality System (http://www.epa.gov/ttn/airs/airsaqs/), a nationwide network of continuous and filter-based monitors, as well as monitoring data from various other sources. The models also incorporated information on several geospatial predictors including distance to road, population density, point sources (e.g., power-generating utilities, waste combustors), elevation, and meteorology. All data were used in generalized additive statistical models (Yanosky et al. 2008) with smoothing terms of space and time to create separate PM prediction surfaces for each month. Because monitoring data on PM2.5 are limited before 1999, PM2.5 in the period before 1999 was modeled using data on PM10 and visibility data at airports (Yanosky et al. 2009, 2014). PM10–2.5 predictions were calculated as the difference between monthly PM10 and PM2.5 predictions. These models provide estimates for any geolocation in the conterminous United States by monthly intervals. The models also have been shown to have low bias and high precision: The normalized mean bias factor for PM2.5 is –1.6%, and the absolute value of the prediction errors is 1.61. For PM2.5–10 these values are –3.2% and 4.18, respectively (Yanosky et al. 2014).
For each child, we estimated exposures to PM2.5 and PM10–2.5 before, during, and after pregnancy by averaging monthly concentrations for the mother's residential address during the relevant months. The months of pregnancy were determined from the child's birth month and gestational age at birth, as reported by the mother. Exposures to PM during each pregnancy trimester were calculated similarly.
Covariates
The following covariates, all associated with autism in previous studies, were included in multivariable models: child's birth year, birth month, and sex, maternal age at birth, paternal age at birth, and median census tract income in the birth year. Among these variables, only census tract income (1.5%) and paternal age (10.6%) had missing data. We used the missing indicator method for missing data. We conducted sensitivity analyses to evaluate the influence of adjusting for gestational factors (premature birth, birth weight, gestational diabetes, preeclampsia), smoking during pregnancy, state, marital status, median census house value, paternal education, and maternal grandparents' education. All covariate data except for census variables were from maternal self-report.
Statistical Analyses
Logistic regression models were used to estimate odds ratios (OR) and 95% confidence intervals (CI) of ASD by PM exposures modeled both using PM quartiles and as continuous variables, in separate models. Exposures to different PM size fractions were examined in separate models, and also together in a single model.
For nurses who moved residence between two questionnaires straddling pregnancy, we did not know the exact date of moving. Therefore, we conducted separate analyses for exposures assigned assuming the nurse was at the earlier address during the whole intervening period (prepregnancy address) or at the later address during the whole period (postpregnancy address). In addition, to reduce misclassification of exposure, we conducted analyses that were limited to those mothers for whom the pre- and postpregnancy addresses were identical [160 cases (65%) and 986 controls (65%), referred to here as "nonmovers"].
To examine temporal specificity of any associations between PM and ASD, we considered the association with PM2.5 exposure during the 9 months before pregnancy, the pregnancy period, and the 9 months after birth. These examinations were restricted to nonmovers with complete data for all exposure periods, and each time period was considered independently, and then also in a single model that included all three time periods simultaneously. Because of differences in ASD rates by sex and prior suggestions that air pollution effects may be specific to boys, we a priori decided to also examine associations stratified by sex of the child. For simplicity, we did this only among the children whose mothers did not move during pregnancy. We used SAS version 9.3 (SAS Institute Inc., Cary, NC) for data extraction, and R version 3.0.1 (http://www.r-project.org/foundation/) for Linux-gnu for analyses. All analyses were conducted at the 0.05 alpha level.