Does the Factor Structure of IQ Differ between the DAS-II Normative Sample and Autistic Children?

Background: Intellectual disability commonly co-occurs with ASD, and the DAS-II (Differential Ability Scales, 2^nd Edition) is frequently used as part of an intellectual disability assessment in school-age children with ASD. Like most IQ assessments, the DAS-II was developed with a nationally representative normative sample. Despite widespread use, it remains unknown whether the normative DAS-II measurement model (e.g., factor structure, loadings) holds for the autistic population (Wicherts, 2016). In other words, it is not known if the DAS-II measures the same construct in the same way for autistic and neurotypical children. If not, then the often discussed “IQ splits” in ASD (e.g., high matrix but low verbal similarities subtest scores, Nowell et al., 2015; Klin et al., 2005; Seigel et al., 1996) might be due to measurement bias and artifacts rather than intelligence (Thompson, 2016). Thus, measurement invariance requires that the relationship between DAS-II scores and the latent construct of intelligence should not be affected by autism (Thompson, 2016; Wicherts & Dolan, 2010).

Objectives: To determine whether the normative structure of the DAS-II, which influences score calculation and interpretation, holds in individuals with ASD.

Methods: We obtained the raw DAS-II normative data for 2,400 individuals, as well as the raw DAS-II data from the Simons Simplex Consortium (SSC) for 1,317 autistic individuals and a replication sample of 416 autistic individuals from The Children’s Hospital of Philadelphia’s Center for Autism Research (CAR). Using the three-factor structure specified in the DAS-II technical manual as a baseline model, we combined normative and SSC datasets for multigroup confirmatory factor analyses to assess how well that measurement model fit both samples (Chen, Sousa, & West, 2005). We analyzed configural invariance (subtests loading on the same factors for each group), metric invariance (subtests loading at equal levels between groups), and scalar invariance (mean subtest scores equal across groups). Additional analyses including subtest correlations, exploratory structural equation modeling, and replication with the CAR ASD sample will be completed by May 2019.

Results: Both normative and ASD sample data showed excellent fit with the 3-factor model (normative: CFI=1.000; TLI=1.000; RMSEA=0.000; ASD: CFI=0.998; TLI=0.995; RMSEA=0.034). The configural and metric models showed excellent fit (CFI>0.995, TLI>0.995, RMSEA<0.025), but the scalar model showed a decrease in fit statistics (ΔCFI=0.010; ΔTLI=0.016; ΔRMSEA=0.038) that exceeded recommended thresholds (ΔCFI<0.010 or ΔRMSEA>0.010; Chen, 2007), suggesting measurement bias. A likelihood ratio test confirmed significantly worse fit of the scalar model compared to metric and configural models (metric: χ²(5)=107.7, p<0.00).

Conclusions: Preliminary analyses suggest only weak factorial invariance (i.e., configural and metric, not scalar invariance) and measurement bias in autistic DAS-II scores. Although the factor structure and loadings do not vary significantly between groups, the pattern of subtest means does; consequently, the lower mean DAS-II subtest scores of the autistic population cannot be attributed only to lower levels on the latent constructs of verbal, nonverbal, or spatial intelligence. These results have important clinical implications, namely that artifacts may influence DAS-II scores for autistic patients. Recommendations for appropriate interpretation of DAS-II scores from children with ASD will be discussed.