Racial and Ethnic Differences in Maternal Health and Autism · Ben Messaoud K, Khachadourian V, Anderson M et al.

Methods

The study cohort consisted of children born within the KPNC health care system between January 1, 2010, and December 31, 2017. Children were excluded if they had no KPNC enrollment after birth, if maternal age at delivery was younger than 13 or older than 55 years, or if key covariate data were missing. Mothers were required to be KPNC members for at least 3 months in each of the 2 years prior to delivery.

The follow-up period for each child started at birth and ended at the earliest occurrence of any of the following: February 29, 2020, a first diagnosis of autism spectrum disorder (ASD) after 12 months of age, disenrollment from KPNC, or death. This report follows the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) reporting guideline for cohort studies. Study procedures were approved by the KPNC institutional review board. The requirement for informed consent was waived because only deidentified administrative data were used.

KPNC is a large, integrated health care system that serves over 4.5 million individuals. The sociodemographic characteristics of its members are broadly representative of California’s general population in the counties served by KPNC, with the exception of underrepresentation among individuals at the extremes of the income distribution.12,13

Exposure

Exposures were the International Classification of Diseases, Ninth Revision (ICD-9) and Tenth Revision (ICD-10) diagnoses recorded during the perinatal period in the mother, shown to be significantly associated with autism in our previous work, including in the KPNC system.3,4 All maternal diagnoses were identified via electronic health records (EHRs) from inpatient and outpatient encounters with a KPNC provider (see eTable 1 in Supplement 1). Maternal diagnoses were not mutually exclusive; a mother could have more than 1 condition, with our previous work finding limited impact of comorbidity on the estimates associated with most diagnoses.4 The Chronic Condition Indicator, developed by the Agency for Healthcare Research and Quality,14 was used to classify diagnoses into chronic or nonchronic. For nonchronic conditions, we classified children as exposed if their mother received a given diagnosis within 12 months before childbirth. For chronic conditions, a longer detection window of 48 months was used, reflecting the enduring nature of these diagnoses and the difficulty of accurately identifying their onset using EHRs. This broader time frame enhances the sensitivity for capturing maternal chronic conditions that may not be actively managed during pregnancy but may still affect fetal development. Sensitivity analyses exploring the impact of varying detection windows for chronic conditions are reported elsewhere.3

Outcome

The outcome was an ASD diagnosis after the child reached age 12 months, as recorded in the child’s EHR and based on the Diagnostic and Statistical Manual of Mental Disorders (Fourth Edition) or Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition) criteria. Among children diagnosed with ASD, 85% underwent evaluation at a specialized KPNC ASD assessment center, where standardized procedures, including the Autism Diagnostic Observation Schedule, were used by a multidisciplinary team. The remaining 15% received diagnoses from qualified clinicians such as child psychiatrists, developmental-behavioral pediatricians, pediatric neurologists, or general pediatricians. ASD diagnoses received before age 12 months were not included due to the lack of specificity of diagnosis at such an early age.

Covariates

Covariates were extracted from the KPNC EHR and included maternal age at delivery, type of insurance at delivery (categorized as KPNC insurance vs government insurance), child’s year of birth, and sex. Insurance type served as a proxy for socioeconomic status (SES). Additionally, the number of days the mother utilized health care services in the year before childbirth was used as a proxy for health care utilization, encompassing both primary and specialist care visits. Race and ethnicity categories were defined in the health system and included Asian and Pacific Islander, Black, Hispanic, White, and other or unknown. The other or unknown category was heterogeneous (see eTable 2 in Supplement 1 for exact composition, derived from the KPNC Virtual Data Warehouse demographic table).

Statistical Analysis

Analytical procedures were aligned with those described before.3,15 Each maternal diagnosis was independently examined for its association with autism using Cox proportional hazards models, with stepwise adjustment for covariates.16 To account for the clustering of siblings within families, we calculated robust sandwich estimators. The heterogeneity of associations by race and ethnicity was evaluated by entering an interaction term between maternal diagnosis and maternal race and ethnicity in 1 set of regression models. Because 25 conditions were tested across racial and ethnic strata, we controlled the false discovery rate using the Benjamini-Hochberg procedure. Additionally, we conducted analyses stratified by maternal race and ethnicity to evaluate the association of each diagnosis with autism in each race and ethnicity group. Diagnoses reported in fewer than 5 mothers of children with or without autism in each race and ethnicity group were excluded to reduce the likelihood of sparse data bias.17 The only exceptions were diagnoses where the only sparse group were children in the other race or ethnicity category; these diagnoses were included in the analyses, but the individuals whose mothers reported other were excluded from the analytical sample. Diagnoses for which the interaction term was statistically significant (Q value <.05) were considered to have differential associations with autism across racial and ethnic subgroups. Two-sided Wald tests with robust (sandwich) standard errors were used to calculate P values for the Cox models, and significance was set at less than .05. Data were analyzed using RStudio software versions 2024.09.0 and 2026.07.1 (R Project for Statistical Computing) from September 2024 to July 2026.