Maternal Intake of Supplemental Iron and Risk of Autism Spectrum Disorder

Rebecca J. Schmidt; Daniel J. Tancredi; Paula Krakowiak; Robin L. Hansen; Sally Ozonoff


Am J Epidemiol. 2014;180(9):890-900. 

In This Article


Study Population

Participants in the Childhood Autism Risks from Genetics and the Environment (CHARGE) Study, a population-based case-control study,[25] who had undergone interviews from the start of the study in 2003 until the questionnaire was revised in September 2011 were included in these analyses. Eligible children included those who were aged 24–60 months, had been born in California, were living with at least 1 biological parent who spoke English or Spanish, and resided in one of the catchment areas on a specified list of California Department of Developmental Services regional centers that coordinate services for children with autism and developmental delay. Children were excluded if they had impairments that would preclude a valid developmental assessment. Children with genetic syndromes were not excluded if they met other inclusion criteria.

Children with autism and developmental delays were identified through the Department of Developmental Services' regional centers, clinics and providers, self-referrals by parents, and public outreach. A stratified random sample of children from the general population, identified from state birth files, was generated by frequency-matching to the projected distribution of autism cases on age, sex, and regional center catchment area. The CHARGE Study protocol was approved by institutional review boards at the University of California, Davis, and the University of California, Los Angeles, and by the State of California Committee for the Protection of Human Subjects. Written informed consent was obtained before participation.

Diagnostic Confirmation

All children were assessed at the Medical Investigation of Neurodevelopmental Disorders (MIND) Institute clinic (Sacramento, California) for confirmation of their diagnosis. Children were assessed for cognitive function using the Mullen Scales of Early Learning[26] and for adaptive function using the Vineland Adaptive Behavior Scales.[27] The Autism Diagnostic Interview–Revised[28,29] and the Autism Diagnostic Observation Schedule–Generic[30,31] were used to confirm autism diagnoses. The children of families recruited from the general population or with developmental delays were screened for evidence of ASD using the Social Communication Questionnaire;[32] if children scored above 14, they were evaluated for autism. Autism case status was defined as meeting criteria in the communication, social, and repetitive-behavior domains of the Autism Diagnostic Interview–Revised and scoring at or above the total cutoff point for autistic disorder on the Autism Diagnostic Observation Schedule–Generic, module 1, 2, or 3. A broader definition of impairment encompasses ASD as defined by Risi et al..[33] Because autism and ASD represent different symptom severities along the continuum of the disorder,[33] we present results for the combined ASD group in accordance with the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition.[34] All interviews and clinical assessments were conducted in English or Spanish by bilingual staff.

Maternal Supplemental Iron Intake

As previously described,[35] trained telephone interviewers collected information from the mother on her intake of multivitamins, prenatal vitamins, iron-specific vitamins, cereals, and other supplements (including whether or not each item had been consumed and, if so, what brand and dose had been consumed, how frequently, and in which months) during a defined period (the index period) beginning 3 months before pregnancy and continuing throughout each month of pregnancy and while breastfeeding. From this information, we calculated average daily intake of iron (and other nutrients) for each product and summed these values to a total value for each month for each woman. Iron amounts were assigned to each brand/product based on information obtained from the manufacturer; if this information was not available, a standard amount was assigned based on the amount most commonly found in similar products.

Statistical Analyses

Data were reviewed for outliers using univariate descriptive analyses. Logistic regression was used to calculate odds ratios (with 95% confidence intervals) as measures of association between iron intake categories and case status, using SAS software, version 9.3 (SAS Institute Inc., Cary, North Carolina). Stratified analyses and interaction terms were used to examine effect modification of maternal iron intake by child sex, interpregnancy interval (defined as elapsed time from the last livebirth or previous pregnancy with a gestational age of ≤20 weeks to the start of the current pregnancy), maternal and child race/ethnicity (non-Hispanic white, Hispanic, or other), maternal age, and maternal metabolic conditions (prepregnancy obesity, defined as body mass index (weight (kg)/height (m)2 ≥30), hypertension, and/or diabetes), as defined previously.[36] The above variables were also examined as potentially confounding factors because of their relationship with iron status, as were the following variables: child's birth year, paternal age, home ownership, month in which prenatal care began, number of prenatal-care visits, parity, maternal birthplace, education, folic acid intake during the first month of pregnancy (<600 µg/day or ≥600 µg/day), cigarette smoking, residing with a smoker, and alcohol consumption. Our analyses started with a full model containing potential confounders identified in the bivariate analyses as being broadly associated (P < 0.2) with both ASD and quintile categories of iron intake (based on control intake). Variables were then excluded using backward selection, retaining in the model any variables that caused at least a 10% change in the exposure parameter estimates. Maternal folic acid intake, home ownership, and child's birth year were the only variables meeting the confounder criteria.

In sensitivity analyses, we assessed the impact of missing data using multiple imputation via the Markov chain Monte Carlo algorithm.[37] To ensure that the results represented the study base, we used survey research methods to fit the logistic regression models, with participants being assigned weights equal to the inverse of the estimated participation probability in strata defined by the entry case group and demographic factors.[38] Estimates were also stratified by maternal folic acid intake during the first month of pregnancy to further control for folic acid's correlated association. To assess the effect of recall bias using the length of time elapsed before mothers were asked to recall their intake, associations were also examined for children who were under the median age of controls at the start of the interview (completion of the interview could require multiple sessions) compared with those who were older at the start of the interview.