The Study to Explore Early Development (SEED) identified children with autism aged 30–68 months by end of data collection through multiple sources, including health, education, early intervention, and other service program settings at study sites within areas of six states (California, Colorado, Georgia, Maryland, North Carolina, Pennsylvania) (Wiggins et al., 2021). Caregivers were asked questions about sociodemographic characteristics, medical history, services and treatments, and other information via telephone interview and mailed questionnaires; caregivers also filled out various standard instruments such as the Child Behavior Checklist (CBCL) (Achenbach, 1992). All participating families completed the Social Communication Questionnaire (SCQ) (Rutter et al., 2003a), and children who had an autism spectrum disorder (ASD) diagnosis upon enrollment or had ASD risk noted on the SCQ were evaluated with the Mullen Scales of Early Learning (MSEL) (Mullen, 1995) and ADOS (Lord et al., 1999; Lord et al., 2012), and mothers completed the VABS, Second Edition (VABS-II) (Sparrow et al., 2005) and ADI-R (Lord et al., 1994; Rutter et al., 2003b). Children who did not meet these criteria were evaluated only with the MSEL; if ASD behaviors were witnessed during MSEL administration, the family was asked to also complete the ADOS, ADI-R, and VABS.

Final classification was based on diagnostic criteria for ASD from results of a comprehensive developmental evaluation, including the ADOS and ADI-R, performed by clinicians with established research reliability. Clinical judgment was also considered for children with an early learning age equivalent <24 months who met SEED ASD criteria on the ADOS and ADI-R since the ADI-R is not appropriate for children in this mental age range (Schendel et al., 2012; Wiggins et al., 2015). This population-based ascertainment protocol supported the identification and rigorous assessment of children with autism with and without a previous diagnosis from clinical and community-based sources. Written informed consent was obtained from all participating families. For this research, analysis was restricted to children in the SEED ASD group (n=1,480). Sociodemographic characteristics of the sample are provided in Table 1.

For Hypothesis 1, we computed VABS-II summary statistics. Mean scores were compared between VABS-II domains using Friedman tests and post-hoc Wilcoxon signed rank sum tests, overall and within each ELC stratum. To estimate the amount of variance in VABS-II scores explained by cognitive ability and autism symptom severity (Hypothesis 2), hierarchical linear regression analysis was performed. In univariate analysis, child’s sex, race, ethnicity, age at administration of the MSEL, and birth order; family’s current income relative to the federal poverty level; primary language spoken in the home; maternal age at enrollment, education, and place of birth; and study site were associated with VABS-II scores and included as predictors in Model 1. In subsequent separate steps, different predictors were added to the regression model, including MSEL composite raw score (Model 2); to account for the potential of a curvilinear relationship in which cognitive and adaptive scores were similar in the lower MSEL score range while cognitive development outpaced adaptive scores in the higher MSEL score range, we also included the quadratic term. In Model 3, ADOS CSS was added, followed by the interaction between ADOS CSS and the two cognitive variables (Model 4) in order to explore the potential effect modification of cognitive development on the relationship between adaptive scores and autism severity. Finally, we added CBCL Total T-score in Model 5 to estimate the additional variance in AB scores explained by overall behavioral and emotional problems (Green & Carter, 2014; Kraper et al., 2017; Franchini et al., 2018). We did not examine externalizing and internalizing problems separately due to significant correlations between these two domains. Nested models were then compared using Wald tests incorporating robust standard errors to determine if the additional variance explained by a new predictor was significant. Separate regressions were run for the VABS-II Composite and each domain score. Regression analysis was limited to complete cases (n=1,278 [86.4% of ASD cases] for the Composite regression; n=1,291 [87.2% of ASD cases] for domain regressions), and we also performed sensitivity analyses using multiple imputation to account for children missing data.

Finally, we described cognitive ability/AB discrepancies among children with ASD with both MSEL ELC and VABS-II scores recorded (Hypothesis 3), overall and within three age cohorts (2–3 years, 4 years, and 5 years of age). We estimated the percent of children with autism with either, both, or neither ELC or VABS-II score indicating significant challenges and calculated Cohen’s kappa coefficients to assess the degree of agreement between the two measures in categorization of functioning (McHugh, 2012). Additionally, after converting both ELC and VABS-II scores into SD units rounded to two decimal places, we calculated the magnitude of cognitive ability/AB discrepancies and determined the percent of children within each cognitive stratum with one and two SD discrepancies between the two measures.

Regression analysis was conducted using R Version 4.2.2 (R Core Team, 2022); the R packages ‘lmtest’ (Zeileis & Hothorn, 2002) and ‘sandwich’ (Zeileis, 2004; Zeileis et al., 2020) were used for nested model comparisons. All other statistical analyses were conducted using SAS Version 9.4 (SAS Institute Inc., 2010). The design and implementation of the SEED study have benefited from community stakeholder input from autistic self-advocates, family members, and service providers, although there was no direct community involvement in the reported study.

On average, the SEED ASD group exhibited significant AB delays. As hypothesized, the Socialization mean was lowest and Motor Skills highest (1.9 and 1.4 SDs below the mean, respectively). In children with ELC≤70, mean VABS-II Composite and domain scores were all 1.7 SD or more below the mean. In children with ELC>70, all mean domain scores were below the population mean of 100, with mean Composite, Daily Living Skills, and Socialization greater than one SD below the mean.

Contrary to our hypothesis, Communication scores were higher than Daily Living Skills on average (Table 2). Between-domain analysis revealed that all mean domain scores significantly differed from one another (p<0.0001 for all comparisons except Communication versus Daily Living Skills in children with ELC≤70 [p=0.03]), overall and within each ELC stratum. Overall and for children with ELC≤70, mean Motor Skills scores were highest, followed by Communication, Daily Living Skills, and Socialization, while in children with ELC>70, mean Communication scores were highest, followed by Motor Skills, Daily Living Skills, and finally Socialization.

There was wide variation in VABS-II scores in the overall ASD group and within cognitive strata. VABS-II Composite scores ranged from 41–114 (3.9 SDs below the mean to 0.9 above the mean) (Table 2). Overall and in children with ELC>70, Communication scores had the largest range of any domain, while in children with ELC≤70, Daily Living Skills scores had the largest range; Socialization had the smallest range of any domain overall and in both strata (Table 2).

In regression analysis, sociodemographic factors explained between 6.0% (Motor Skills) to 17.2% (Communication) of variance in VABS-II scores (Table 3). Addition of the MSEL composite raw score explained significant and substantial variance in VABS-II scores (ΔR²=51.1% for the VABS-II Composite score), with domain-specific ΔR² values ranging from 58.0% in Communication to 31.9% in Motor Skills scores (Table 3, Model 2a). Addition of the ADOS CSS (Model 3a) explained a significant but small amount of additional variance in the Communication (ΔR²=0.1%), Daily Living Skills (ΔR²=0.2%), and Socialization (ΔR²=0.4%) domains. Addition of the MSEL/ADOS CSS interaction only explained significant additional variance for Motor Skills (ΔR²=0.5%) (Model 4a). Finally, addition of the CBCL Total T-score explained significant additional variance for all VABS-II scores, explaining most in Socialization (ΔR²=4.9%) and least in Motor Skills (ΔR²=1.0%) (Model 5a). Altogether, these predictors explained 67.2% of the variance in composite AB scores, with most explained variance in VABS-II Communication scores (R²=76.4%), followed by Daily Living Skills (R²=57.3%), Socialization (R²=48.9%), and Motor Skills (R²=39.5%) (Table 3).

In children with MSEL ELC≤70, the full model explained between 72.1% (Communication) and 31.9% (Motor Skills) of AB score variance (Table 3). The MSEL composite raw score explained significant additional variance for every VABS-II score, ranging from 60.2% for Communication to 24.1% for Motor Skills. The ADOS CSS explained significant additional variance for Communication (ΔR²=0.1%), Socialization (ΔR²=0.8%), and Motor Skills (ΔR²=0.4%), while the additional variance explained by the MSEL/ADOS CSS interaction was not significant for any score. CBCL Total T-score explained significant additional variance for the Composite and every domain, ranging from 3.9% for Socialization to 0.6% for Motor Skills.

In comparison, in children with ELC>70, these predictors explained less of the VABS-II score variance with R² values ranging from 39.0% for Communication to 16.9% for Motor Skills (Table 3). The MSEL composite raw score explained significant additional variance for every VABS-II score, ranging from 13.4% for Communication to 1.2% for Socialization. Addition of the ADOS CSS explained significant additional variance in Daily Living Skills (ΔR²=0.9%) and Motor Skills (ΔR²=1.1%), while the MSEL/ADOS CSS interaction did not explain significant additional variance for any VABS-II score. CBCL Total T-scores explained significant additional variance for every VABS-II scores, ranging from 9.0% for Socialization to 2.5% for Motor Skills. For all scores except Communication, the additional variance explained by CBCL scores was similar to or greater than that explained by the MSEL composite raw score, particularly in the Socialization domain in which the CBCL Total T-score explained 9.0% significant additional variance compared to only 1.2% significant additional variance explained by the MSEL composite raw score (Table 3).

Analysis of multiply imputed data resulted in similar R² and ΔR² values. Additionally, patterns were generally similar when the Visual Reception raw score was used as our measure of cognitive ability, except that, generally, the Visual Reception raw score explained a smaller amount of additional variance in VABS-II scores than the composite raw score (see Table S2).

Figure 1 shows the wide range of composite AB scores among children with ELC scores below and above the two SD below the mean cutoff (Figure 1). Cohen’s kappa coefficients indicated weak (Composite [0.45]) to minimal (Communication [0.36], Daily Living Skills [0.37], Socialization [0.38], and Motor Skills [0.26]) agreement between MSEL ELC and VABS-II scores regarding whether or not both scores fell above or below the two SD cutoff for significant functional impairment or delay. In our sample, 25.3% of children had ELC≤70 but VABS-II Composite scores indicating no significant adaptive delays, while 3.3% and 8.1% had ELC>70 but VABS-II Composite and Socialization scores, respectively, indicating significant delay (Figure 1, Table S5). Over a third (39.7%) of children with ASD and ELC≤70 did not have a composite AB score indicating significant adaptive delay (Figure 1, Quadrant I). In addition, among children with ELC>70, 9.0% and 22.2% had VABS-II Composite and Socialization scores, respectively, indicating significant adaptive delay (Figure 1, Quadrant IV; Table S5). Similar results for other VABS-II domain scores are provided in supplementary Table S5. Patterns were similar when the relationship between VABS-II scores and the MSEL Visual Reception T-score were examined, but a smaller percentage of children had cognitive scores indicating significant delay when measured by the Visual Reception T-score rather than ELC (49.5% versus 63.7%) (see Figure S1).

Consistent with our third hypothesis, we found discrepancies between MSEL ELC and VABS-II scores. Most children with ELC>70 had cognitive scores exceeding their VABS-II scores (62.1% for the Composite). The largest discrepancies were seen in the Socialization domain, with 36.2% of children having ELC>AB with at least a one SD discrepancy and 12.3% having a two SD or more discrepancy between the two scores (Figure 2). Discrepancies were also large for other domains (Figure 2). In children with ELC≤70, adaptive scores often exceeded ELC (91.9% for the Composite), and the proportion with AB scores at least one SD greater than their ELC ranged from 73.8% for Motor Skills to 43.1% for Socialization (Figure 3).

When stratified by age, the proportion of children with a two SD or more discrepancy between their ELC and VABS-II scores decreased with increasing age (Figures 2 and 3), with this trend generally more pronounced in children with ELC>70. There were substantial proportions of children with at least a one or more SD difference in all age cohorts.

The large, population-based SEED sample includes children with ASD from diverse racial, ethnic, and socioeconomic backgrounds. Participants all underwent rigorous, research-reliable, professional assessments and had nearly complete data on instruments measuring cognitive development, adaptive functioning, autism symptom severity, and behavioral and emotional health. Additionally, recruitment procedures allowed for identification of ASD cases from a variety of community sources that may otherwise be missed when participants are recruited primarily from clinical populations, as is the case for much of the previous literature on adaptive functioning in children with autism. The broad diagnostic net for case identification in SEED also made possible the inclusion of children without a previous ASD diagnosis, mitigating the risk of a sample biased towards children with access to healthcare or intervention services.

One limitation of our study was the lack of a measure of social demands and support, particularly given our finding that sociodemographic characteristics explained variation in adaptive functioning (Mandell et al., 2005; Constantino et al., 2020). Future research could examine the association between AB and measures of social determinants of health, such as access to resources like early diagnosis and early intervention programs and family and neighborhood environment. Future research could also examine domain-specific associations between cognitive ability, autism symptom severity, and behavioral and emotional problems and AB scores.

Another consideration is the restricted age range of children in our sample. It is important to note that ID is not typically diagnosed in preschool-aged children due to the instability of IQ scores in this age group. However, early childhood cognitive scores have been shown to be predictive of adult outcomes, including IQ (Bishop et al., 2015; Magiati et al, 2014; Pickles et al., 2020), and this study provides evidence that the gap between cognitive and AB scores found in previous ASD research emerges early and may have important implications for research on ASD with co-occurring ID and interpretation and of past research that classified children with autism as “high-“ or “low-functioning” based solely on cognitive scores. Finally, this analysis focused on the MSEL and VABS-II, and the findings may not generalize to all measures of cognitive and adaptive development.