Maternal Prenatal Smoking and Autism Spectrum Disorder in Offspring

A California Statewide Cohort and Sibling Study

Ondine S. von Ehrenstein; Xin Cui; Qi Yan; Hilary Aralis; Beate Ritz


Am J Epidemiol. 2021;190(5):728-737. 

In This Article


We used a cohort registry linkage design including all births during 2007–2010 in California; data were retrieved from Office of Vital Statistics birth rolls. ASD cases were identified using records maintained at the California Department of Developmental Services (DDS) and collected by 21 regional centers through December 31, 2013. We included only cases with a primary diagnosis of "autistic disorder" based on the Diagnostic and Statistical Manual of Mental Disorders (code 299.00 in DSM IV-TR—the standard diagnostic instrument until December 31, 2013),[28] as reported on the DDS Client Development Evaluation Report. Validation studies have established the reliability and validity of the Client Development Evaluation Report in California.[29] Eligibility for DDS services does not depend on citizenship or financial status (i.e., services are available to all children). We linked DDS case records to California birth records using a probabilistic linkage based on child and parental identifiers (e.g., first/last name, birth date, and sex) to estimate the probability/likelihood that 2 records are for the same person, assigning total scores for a linkage as the sum of scores generated from matching individual fields using the National Program of Cancer Registries Link Plus Software[30] (linkage rate: 86.3%). We manually checked pairs with borderline scores; nonlinkage was mainly due to missing information on records. We excluded 455 case and 82,100 noncase records with missing or implausible/nonviable gestational ages (included: 147–322 days) or birth weights (included: 500–7,000 g); after restricting to singleton births, this yielded a final sample of 11,722 cases and 2,003,382 noncases. Among these, siblings (born 2007–2010) were identified using a similar approach based on child and maternal identifiers (mandatory) and paternal when available, to estimate the probability/likelihood that 2 birth records relate to the same mother/parents.

We classified ASD cases based on DDS evaluation records by "intellectual disability" status (recorded as "mental retardation" at the time of study, diagnosed according to Diagnostic and Statistical Manual of Mental Disorders–IV criteria corresponding to International Classification of Diseases, Ninth Revision) applying the last evaluation record for each child. We determined age at diagnosis in the CDER; the later recorded diagnosis date was used for 44 cases with 2 recorded dates.

This research was approved by the University of California, Los Angeles, Office of the Human Research Protection Program and the California Committee for the Protection of Human Subjects, and was exempted from informed consent requirements because there was no contact with human subjects.

Smoking Exposure and Pregnancy Variables

Information on pregnancy characteristics such as gestational age, birth weight, and pregnancy complications, as well as sociodemographic data, were retrieved from birth records. Smoking information was recorded (since 2007) on California birth certificates as number of cigarettes smoked per day 3 months prior to pregnancy and in each trimester.

Statistical Analysis

Odds ratios and 95% confidence intervals were estimated for maternal smoking and ASD using logistic regression. Smoking was assessed according to developmental period. All models adjusted for year of birth and sex, and additionally for potential confounders selected based on prior knowledge,[31–33] including maternal age, education, race/ethnicity, parity, and pregnancy complications (definitions displayed in Table 1). We conducted sensitivity analyses adjusting for additional covariates, including preterm birth, birthweight, type of insurance (as proxy for socioeconomic status (SES)[34]), maternal birth place (United States/non–United States), and DDS Regional Center catchment area (in a subsample); because none of these variables changed the estimates of interest by >5%, they were not retained in final models. Further we stratified by ASD with and without intellectual disability comorbidity to examine phenotypes by severity. We also conducted analyses restricted to term births and to diagnosis received at or before age 3 years, and not adjusting for pregnancy complications.

To examine the role of confounding by unmeasured shared familial factors, we implemented a sibling comparison restricting to the group of cases with siblings ("case families"). That is, from the original full cohort, we selected all ASD cases with a sibling (n = 2,705) and the nonaffected siblings of ASD cases (born 2007–2010; n = 2,639); these subjects were considered to belong to the same families (case families, n = 2,611) (Figure 1). Conditional logistic regression clustering on maternal identifier as the family indicator for sets of siblings was applied. We conducted sensitivity analyses to compare the findings from the case-family only analyses with those using all siblings (Figure 1).

Figure 1.

Flow chart of sample derivation for the maternal smoking and autism spectrum disorder study, among births in California, 2007–2010.

If associations seen in the full cohort persist in the sibling comparison subcohort, this supports the hypothesis that maternal smoking in pregnancy increases offspring's ASD risk even after controlling for family-specific factors. To the extent that associations are attenuated, at least 2 explanations should be considered; first, the full cohort associations did at least in part result from familial confounding; second, nondifferential measurement error of the exposure (i.e., random misclassification of maternal smoking on the birth record). While the sibling comparison should adjust for shared familial factors by design,[35] the approach has several limitations; it relies on within-family discordance, is prone to exposure misclassification and to unmeasured nonfamilial confounding, and usually has less power and generalizability.[35,36] Analyses were conducted using SAS, version 9.4 (SAS Institute Inc., Cary, North Carolina).