Evidence for Implicit Bias in the Implementation of Autism Screening Tools in Primary Care

Background: Health disparities associated with age of ASD diagnosis have been observed for racial/ethnic minorities (Mandell et al., 2002), females (Shattuck et al., 2009) and economically disadvantaged children (Mazurek et al. 2014). While ASD screening models and tools have been discussed as possible contributors to these disparities, biased primary care screening/referral patterns have not been directly investigated.

Objectives: The objective of this study was to investigate error patterns in primary care screening and referral that may contribute to later delays in ASD diagnosis and early intervention for minority children.

Methods: A retrospective review of 18- and 24/30-month well visits (n = 4886) at an urban academic medical center yielded data on primary care provider’s interpretation of the Modified Checklist for Autism in Toddlers (M-CHAT); manual re-scoring of M-CHATs allowed for the provider’s interpretation to be classified as a True Positive (TP), True Negative (TN), False Positive (FP) or False Negative (FN). Demographic data and referrals for diagnostic/intervention services were also extracted from the medical record.

Results: Children in the study sample were representative of the racial/ethnic diversity of the urban area (40.1% African American, 30.6% Hispanic; 17.3% Spanish-speaking), and 80.9% of the children were publically insured. Among children determined to have a positive M-CHAT upon manual re-scoring, unweighted logistic regression analyses revealed that females (OR = 1.98. p = .004) and children of Spanish-speaking parents (OR = 2.10, p = .047) were more likely be misclassified as a negative screen by their providers, but no other demographic variables were significant predictors. When the model was propensity weighted to control for associations between demographics and the availability of an original M-CHAT to re-score, the gender bias remained significant (OR = 1.82, p = .014) but the language bias did not (OR = 1.83, p = .23). Additional biases emerged when considering any kind of misclassification (FN and FP): weighted models indicated that African American children (OR = 1.47, p = .028), speakers of languages besides English or Spanish (OR = 1.62, p = .054), and children with public insurance (OR = 1.54, p = .046) were all more likely to be misclassified. Children with correctly identified (vs. overlooked) positive screens were more likely to be referred for diagnostic/intervention services, irrespective of demographics (p < .001).

Conclusions: Patterns of primary care provider errors in interpreting autism screening tools disproportionately impacted minority and disadvantaged children in this sample, impeding access to diagnostic and early intervention services. These results strongly implicate implicit provider biases as a factor in the downstream health disparities observed for minority children who receive ASD diagnoses later than Caucasian children, and are less likely to access quality intervention services. Beyond the development of enhanced tools and screening models, unconscious and unintentional biases must be monitored and mitigated by health care professionals and researchers to adequately serve all children and families.