Psychometric Analysis of the Autism Spectrum Quotient Using Diagnostic Classification Modeling

Background:

The Autism Spectrum Quotient (AQ), a widely-used self-report questionnaire measuring autism traits in adults, has reported good predictive validity and moderate to high internal consistency for its five subscales (Baron-Cohen et al., 2001). Factor analyses have not shown the AQ to be unidimensional; its use as a raw score measuring a single latent variable has been criticized (Lundqvist & Lindner, 2017). Items demonstrated a need for revision due to redundancies and negative point-measure correlations. To facilitate greater understanding of the AQ, examination of how each item contributes to its estimation of the latent variables it measures is needed.

To evaluate the psychometric properties of the AQ items and subscales, we used a diagnostic classification model (DCM), psychometric models designed to classify respondents according to specified categorical latent traits (Ravand & Robitzch, 2015). The categorical classifications of each latent trait provide maximum separation between groups based on item responses. While classical test theory models weight each item equally in raw scores, DCMs account for variable test item quality.

Objectives:

We aim to characterize psychometric properties of the AQ items and subscales using a DCM. We also aim to evaluate predictive validity of subscale classifications in comparison to raw subscale scores.

Methods:

200 undergraduates completed questionnaires, including the AQ, Perseverative Thinking Questionnaire (PTQ), and Patient Health Questionnaire (PHQ; a depression measure).

First, a unidimensional DCM was estimated with AQ as the single latent trait. A second DCM was estimated using the five subscales as latent variables. The relation between AQ classifications and other variables were assessed using independent-samples t-tests. A moderated regression was constructed to determine the impact of classifications on the relation between PTQ and PHQ.

Results:

The unidimensional model characterized 31.5% of participants as high traits (HT). HT participants had significantly higher PTQ and PHQ compared to those with lower traits (LT; both p = .03). The mean difference in probability for HT and LT participants endorsing an item was .16 (see Table 1 for details of the first 20 items).

In the five-attribute model, correlations between attributes varied, from negative correlations with Attention to Detail and all others, to positive correlations with Social Skills and Communication/Imagination (both r > .9), suggesting the AQ is not a unidimensional measure. Comparing the relative fits of unidimensional and five-attribute models using a chi-squared test, the five-attribute model had better fit (p < .0001).

While neither overall raw score (p=.17) nor HT grouping (p=.55) was significant in the moderated regression, classifications (but not raw scores) of Attention Switching and Attention to Detail subscales were significant (p=.004 and .02, respectively). Figure 1 illustrates Attention Switching as a moderator.

Conclusions:

Due to varied specificity in AQ items – “I tend to notice details that others do not” versus “I usually notice car number plates or similar strings of information” – the DCM model showed greater predictive validity in our moderated regression (illustrating attention as a moderator in the relation between perseveration to depression) than raw scores. Future research should examine the impact of item parameters on classical test theory.