This surveillance-based study showed increasing ASD prevalence associated with increasing SES in a dose-response manner, with a stronger SES gradient in ASD prevalence in children with versus without a pre-existing ASD diagnosis. The main results of this study were consistent with the only study larger than this to examine the association between ASD risk and an indicator of SES. That study, published in 2002 by Croen and colleagues, looked at more than 5000 children with autism receiving services coordinated by the California Department of Developmental Services and found a stepwise increase in autism risk with increasing maternal education 
. Our results were somewhat consistent, but also contrasted somewhat, with Bhasin and Schendel's case-control study based on surveillance data collected in 1996 in Atlanta, Georgia. That study found a positive association between SES and risk of ASD based on ascertainment through health care providers, but not based on ascertainment only from school records 
. Bhasin and Schendel suggested that this difference by the type of information source might indicate selection bias because in the U.S. access to school-based services is universal whereas access to healthcare is not. In contrast to the Bhasin and Schendel study, our study included a larger number of children with autism identified only from school records (635 vs. 246), was restricted to 8-year-old children (an age at which children with autism are more likely to have been identified, whereas the age range of the Bhasin and Schendel study was 3 through 10 years), and covered the 2002 and 2004 study years (versus 1996, a time when schools were just beginning to use the autism category). Our finding of an SES gradient in autism prevalence regardless of source of information (health vs. school) was not consistent with the hypothesis that the frequency of children with autism identified only through school sources is constant across SES categories. This finding suggests that the observed SES gradient in autism prevalence may not be due entirely to ascertainment bias.
Epidemiologists long have suspected that associations between autism and SES are a result of ascertainment bias, on the assumption that as parental education and wealth increase, the chance that a child with autism will receive an accurate diagnosis also increases 
. A number of investigators and recent reviews of the epidemiology of autism have concluded that any association observed between autism risk and SES has been due to such bias 
. The present population-based study of U.S. surveillance data provides some support for this conclusion by showing a stronger SES gradient in prevalence among children with ASD with than without a pre-existing ASD diagnosis. In a previous analysis of ADDM Network data for children identified by the surveillance system as meeting diagnostic criteria for ASD, Mandell and colleagues found non-Hispanic white and Asian children to be more likely than non-Hispanic black and Hispanic children to have a pre-existing ASD diagnosis 
. In addition to biased ascertainment resulting from those with higher SES having greater access to diagnostic services, it is possible that “diagnostic bias” on the part of clinicians might contribute to ascertainment bias. In a study designed to identify possible diagnostic bias, Cuccaro and colleagues found evidence that clinicians might be more likely to assign autism diagnoses to case vignettes of children with developmental disabilities if the children's backgrounds were described as higher SES rather than lower SES 
. At the same time, our observation of a significant, if weaker, SES gradient in ASD prevalence when the results are restricted to cases without a pre-existing diagnosis points to the possibility that factors other than ascertainment bias might also contribute to the positive association between ASD prevalence and SES.
A possible reason for the lack of consistency between our findings and those of epidemiologic studies conducted in Denmark 
and Sweden 
, and which found no association between autism risk and SES, might be that the Scandinavian countries have less socioeconomic diversity and more equitable access to services than the U.S. population. The lack of consistency also could be due to the small number of cases and limited statistical power in the Scandinavian studies, and differences in study designs.
An important advantage of this study was that it was large enough to allow stratified analyses of the association between autism risk and SES among demographic and patient subgroups. It is notable that the SES gradient is observed in all four racial/ethnic strata. Also notable is that, although the overall ASD prevalence was higher among non-Hispanic White and Asian children than among non-Hispanic Black or African-America and Hispanic children, when the results were stratified by SES, we saw that the racial/ethnic differences in prevalence varied by SES (). The lower prevalence among non-Hispanic Black or African-American and Hispanic children was seen only in the low SES category, and the fact that more non-Hispanic Black or African-American and Hispanic children live in poverty contributed to the lower overall prevalence among these groups.
The only subgroup in which the SES gradient was not observed was the subgroup with co-occurring autism and intellectual disability (). The lack of an SES association among this subgroup might have been due to counter-associations because intellectual disabilities among children overall are inversely associated with SES 
. It could also be an indication of ascertainment bias if children with intellectual disabilities are more likely than other children to be evaluated for developmental disorders including autism.
An important limitation of this study was that the ADDM Network surveillance system relies on information for children who have access to diagnostic services for developmental disabilities. We could not rule out the possibility that the quality and quantity of evaluations and information available for case ascertainment might have varied by SES. We looked for evidence of this by examining the number of evaluations per child with ASD recorded in the ADDM Network surveillance system, reasoning that if the higher prevalence of ASD among children of higher SES was due to increased access to diagnostic services, high SES might be associated with a higher number of diagnostic evaluations per child. However, we found no association between the number of evaluations per child and SES. We also examined the mean ages at diagnosis by SES and found that children of high SES received an ASD diagnosis at an average age of 58.0 months, 1.1 month earlier than those of middle SES (p
0.2838) and 2.7 months earlier than those of low SES (p<0.0272). This modest difference in age at identification may indicate that diagnostic bias contributes to the SES gradient in ASD prevalence in some studies, though not necessarily in the present study which relied on surveillance at the age of eight years and included cases with and without a pre-existing ASD diagnosis.
Another limitation of this study was the reliance on area-level measures of SES that might not have served as accurate proxies for the SES of individuals or specific families or households. Though perhaps not ideal, these measures have been shown to be reasonable proxies for individual-level SES and have the advantage of serving as indicators of the social and economic contexts in which children live but without introducing ecological fallacy 
. Another limitation of the SES indicators used in this study is that they were based on residential address at the age of eight years rather than at the age of first diagnosis (for children with a pre-existing ASD diagnosis) or other time points, which may have allowed evaluation of whether families of children with ASD migrate to higher SES neighborhoods to improve their access to services, as suggested by Palmer and colleagues 
. A further limitation of our use of aggregate census data for denominator or comparison group data in this study was that we were unable to perform multivariable analyses to evaluate and control for confounding effects of variables such as parental age and other perinatal risk factors 
If the SES gradient found in this study is due only to ascertainment bias, this would imply that there are significant SES disparities in access to diagnostic and other services for children with autism in communities across the United States. It also would imply that the current estimate of ASD prevalence might be substantially undercounted, with children of low and medium SES being under-identified and underserved relative to those with high SES.
The presence of an attenuated but still statistically significant SES gradient when the analysis was restricted to children with no pre-existing ASD diagnosis suggests the overall SES gradient may not be entirely due to ascertainment bias and points to the possibility that factors associated with socioeconomic advantage might be causally associated with the risk for developing autism. The types of exposures that might merit consideration in future research could include a wide range of factors, from physical or social environmental factors to which children living in more advantaged environments might have higher exposures, to immunological factors (such as that suggested by the “hygiene hypothesis” 
) or other biological factors (for example, those associated with parental age). It is also possible that the SES association demonstrated in this study was a result of confounding by unknown factors associated with both high SES and susceptibility to ASD, or to effect modification. Further research to identify such factors could lead to advances in our understanding of the etiology and identification of autism and to possible interventions.