Because of the significant symptom heterogeneity found in autism, it is often conceptualized as ‘autism spectrum disorder’ (ASD). Variability in IQ is one of the most salient dimensions of this heterogeneity. Both DSM-IV (APA, 1994) and ICD-10 (WHO, 1992) definitions of Asperger syndrome include cognitive developmental level as one of the key features distinguishing it from autism. It is estimated that 70% of individuals with autism have IQs in the mentally retarded range (Fombonne, 2003), yet some individuals have above average intellectual ability (Miller & Ozonoff, 2000). Moreover, the IQ profiles of individuals with and without mental retardation tend to differ, with higher IQ individuals typically having a higher Verbal IQ on average (Ghaziuddin & Mountain-Kimchi, 2004; Gilchrist, Green, Cox, Burton, Rutter, & Couteur, 2001). However, there is extreme individual variability in IQ, making it unlikely that a specific cognitive profile can be used for differential diagnostic purposes (Filipek et al., 1999; Siegel, Minshew, & Goldstein, 1996). Nevertheless, researchers have used various strategies to subtype individuals with autism. Some have focused on medical conditions or known biological etiologies contributing to the disorder. Miles et al. (2005) defined subtypes of autism based on whether individuals had features that were stable from birth, suggesting an organic factor, including all of the syndromes that are currently acknowledged as causes of autism (e.g., Fragile X). These individuals, comprising what the authors call the ‘complex’ autism group, also tend to have more seizures, dysmorphic physical features, microcephaly, lower IQs, and a tendency toward poorer outcome than the ‘essential’ autism group. The ‘essential’ autism group is characterized by higher incidence of sibling recurrence and a family history of autism, higher male to female ratio, higher likelihood of regression and macrocephaly, and overall higher IQs. According to Miles et al., assigning individuals to the complex and essential groups allows for the first stage of characterizing the etiologic heterogeneity of those with ASDs. This separation might be especially useful for genetic analyses because it provides a more homogenous group of individuals (essentials). However, until the etiological substrates of autism are identified, it is impossible to know how truly homogenous this group is.

Additional research has focused on defining ASD subgroups according to behavioral patterns of social interaction. Wing and Gould (1979) first characterized autism according to three subtypes: aloof, passive, or active-but-odd. The aloof subtype, which includes children who tend to reject contact and avoid gaze, is typically the most impaired and severely autistic (e.g., Castelloe & Dawson, 1993; Sevin, Matson, Coe, Love, Matese, & Benavidez, 1995). Levels of IQ tend to correspond to social typology, with the aloof group having the lowest IQ, followed by the passive and then active-but-odd groups (Borden and Ollendick, 1994). The aloof group also tends to have the lowest levels of adaptive behavior, worse language and communication skills, and higher ratings of stereotyped behavior/restricted interests. Intellectual functioning likely accounts for a large proportion of the variance in predicting language and communication skills, the presence of stereotyped behaviors, and other prototypically autistic behaviors, which may partly contribute to group assignment. Indeed, Volkmar et al. (1998) found that IQ is often a predictor of social subtype assignment; however, it may not fully account for it.

Because level of intellectual functioning may be among the strongest indicators of subtype, investigators have often attempted to divide the ASD group by choosing an a priori IQ cutoff in order to designate and then characterize the resulting high- and low-functioning groups (Bartak and Rutter, 1976; Allen et al., 2001). The lower- functioning cognitive subgroup, defined as having an IQ below 70 or 80, tends to exhibit more self-injury, stereotypies, and prototypical autism behaviors. Yet such cutoff points are somewhat arbitrary, making it likely that there is diagnostic overlap between the cognitive subgroups generated. Furthermore, distinctions between verbal and nonverbal information processing abilities are often not explored, but may be important in identifying subtypes in autism. Tager-Flusberg and Joseph (2003) investigated discrepancies between verbal and nonverbal IQ in children with autism and found children with discrepantly high nonverbal skills relative to verbal skills had greater social impairment independent of absolute level of verbal ability and overall ability.

Attempts have been made to investigate subgroups within a dimensional construct on the basis of non-unimodal distributions. Meehl (1995) notes that bimodality and marked skewness may be suggestive of latent groups, however, the presence of bimodality is neither a necessary nor sufficient condition for the existence of latent subgroups. For example, when two latent distributions have a mean difference of 2 SDs and equal variances, bimodality may not even be apparent. On the other had, Grayson (1987) has noted that even when bimodality is observed in measured variables the underlying structure may still be a continuous dimension.

Statistical strategies may provide a more empirical basis for characterizing individuals within possible ASD subgroups. A review of the literature indicates that most cluster analytic studies yield 2, 3, or 4 subgroups based on degree of impairment. Sevin et al. (1995), for example, used cluster analysis to classify 34 children with autism or PDD-NOS into 4 groups, described as ranging from high-functioning to low-functioning (severe) autism, with IQ decreasing with severity, and differing significantly between groups. Similarly, Eaves, Ho, and Eaves (1994) used a standard clustering algorithm and principal components analysis of variables to parse 166 children into 4 groups ranging from ‘typically autistic’ and lower-functioning to a higher-functioning group that more closely resembled Asperger syndrome. Again, severity of autism was related to intellectual impairment in that the most impaired subtype had the lowest average IQ. In a longitudinal examination of 138 school age children with autism, Stevens et al. (2000) employed hierarchical agglomerative cluster analysis to validate a 2 group solution, in which cognitive level was the largest separating variable. Children who were lower-functioning as defined by nonverbal IQ at pre-school tended to show poorer outcome at school age, suggesting that nonverbal IQ is an extremely potent predictor membership among school-age children.

Often, cluster analytic techniques have been used to determine which behavioral features of autism tend to correlate or account for the majority of variance – or which factors ‘cluster’ together. Once a cluster solution is determined and individuals are assigned to groups, the subtypes are characterized using various descriptors, including level of intellectual functioning. In using IQ as a descriptor only after the groups have been defined, however, these analyses make it difficult to determine the true role of intellectual capacity in the formation of subgroups, and the actual distribution of IQ in the samples. Few investigators have focused exclusively on cognitive functioning as the empirical indicator of subgroup classification. Those who have specifically investigated the role of intellectual capacity in differentiating ASD subtypes have often found that IQ is the most significant contributor in discriminating between groups and the basis of differences between subtypes (e.g., Miller and Ozonoff, 2000).

Although the goal of cluster analysis is to determine the categories underlying autism spectrum disorders, these methods often yield groups with considerable diagnostic overlap. Unfortunately, under such conditions, cluster analysis often (a) fails to identify the correct number of clusters in datasets where group membership is known, and (b) performs poorly in sorting individuals into subgroups (e.g., Krieger & Green, 1999; Tonidandel & Overall, 2004). Furthermore, it has long been recognized among statisticians that clustering algorithms partition datasets into subgroups, even if the distributions are known to be continuous (see Beauchaine, 2003). Thus, results derived solely from cluster analysis do not provide strong evidence for subgroups of autism, and do not eliminate the possibility of a spectrum of autistic-like disorders (Prior et al., 1998). In fact, data from eight cluster analytic studies suggest that children with PDD-NOS may fit into one of two overlapping groups, and that the subtypes resemble each other, existing along a continuum, and differing only by degree of impairment (Myhr, 1998). In a review of subtyping studies of autism, Beglinger and Smith (2001) posit their ‘best guess’ that symptom heterogeneity can be represented by three continua (developmental delay, social impairment, and repetitive behaviors) and rough divisions can be drawn along these continua yielding four subgroups. The authors also note the weaknesses associated with cluster analytic techniques, including the dependence on the investigators’ choice of variables and characteristics of the sample. This conclusion of the presence of a "continuum containing subgroups" highlights the continued difficulty researchers in this area have in determining whether true differences between subgroups in autism can be reliably distinguished.

In part as a result of the limitations of cluster analysis, additional classification techniques, including latent class analysis (LCA) and taxometrics, have been developed. Although rarely used to evaluate whether subgroups of autism exist, these techniques offer several advantages over clustering algorithms (Beauchaine & Marsh, in press). For example, LCA provides objective measures of fit for comparing alternative subgroupings, and taxometric analyses are far less prone to identify spurious subgroups within continuous distributions. The lone example of taxometric analysis (based on an adaptation of the regression-mixture model, Golden & Mayer, 1995) in autism is the Autism and Language Disorders Nosology project (Rapin, 1996) which found evidence for 2 discrete subgroups, or taxa, in a sample of children with PDD (Fein, et al., 1999) with the nonverbal IQ of about 65 optimally dividing the groups. In the present paper, both LCA and maximum covariance (MAXCOV), the most widely studied taxometric algorithm, were used to address the question of whether subgroups of ASD can be identified from the verbal and nonverbal IQ scores of probands.

To summarize, although it is unclear whether distinct subtypes of autism exist, a recurring pattern emerges in which IQ strongly predicts social functioning, adaptive behavior, severity of symptoms, and prognosis (Coplan & Jawad, 2005; Howlin, Goode, Hutton, & Rutter, 2004; Bolte & Poustka, 2002; Liss et al., 2001; Carpentieri & Morgan, 1996). We used both MAXCOV and LCA to analyze verbal and nonverbal IQ scores obtained from a large sample of preschool-aged children diagnosed with ASD, who were evaluated through the NICHD Collaborative Program of Excellence in Autism (CPEA). Although cluster analysis offers no proven means of choosing among models with different numbers of classes and tends to over-extract classes when defining subtypes, LCA and MAXCOV offer an alternative and more conservative approach to determine whether there is a bimodal or multimodal distribution of intellectual functioning among individuals with autism. By using young children in this analysis, we hoped to minimize individual difference related to experience and treatment.

This evidence in hand, we used latent class analyses to explore this question and to assess whether 2 or more latent groups based on IQ could be identified. The LCA analysis revealed a 4 group solution as providing the best fit for these data. These four groups can be described as a mixture of the overall level of functioning in the group coupled with the presence or absence of a large discrepancy between verbal and nonverbal functioning. The largest group, comprised children with low nonverbal and very low verbal scores and reflected those with the most severe cognitive impairments. A second group also showed very low verbal abilities but had nonverbal scores over 40 points higher on average. The third and fourth groups reflected commensurate verbal and nonverbal abilities with the third group showing mild to moderate impairments in their cognitive functioning and the fourth, highest functioning group scoring in the low average range.

Groups with identical patterns of Mullen scores were observed when contrasting the 2 group latent class analysis results with those from the MAXCOV method which yields, by definition 2 groups when evidence of taxonicity is present. Children who were classified in a similar manner with these two methods matched the lowest and highest functioning groups identified in the LCA four class model. Children classified differently across the two methods seem to comprise 2 distinct IQ-based subgroups that are different both from a simple low- or high-functioning group, as well as different from each other. The verbal versus nonverbal dimension of intellectual functioning in addition to the absolute level of overall functioning is therefore important to evaluate when considering subgroups of children with autism.

Though no sex differences were present in the likelihood of group classification, there was an intriguing mean difference in age among the groups. Children in group 2, which showed an extreme discrepancy between verbal and nonverbal scores was on average nearly 1 year younger than the children in the other groups. Beside the low verbal scores and large verbal-nonverbal discrepancy, this group also showed the lowest level of repetitive and stereotyped behaviors on the ADI. One interpretation of these findings is that group 2 may reflect a developmental or maturational effect rather than a uniquely different IQ subgroup. As children in this group age, those whose language and communication ability improve would look like those in group 3 (with more commensurate verbal and nonverbal abilities), whereas those who make slow progress would tend to resemble those in group 1 (with more profound verbal impairments). Perhaps as well, levels of repetitive behaviors may increase to levels similar to the other groups as these behaviors often follow the social and communication impairments in young children with autism. Alternatively, this group may signify a specific language impairment (SLI) in these children as described by Kjelgaard and Tager-Flusberg (2001). Whether language deficits relative to nonverbal IQ in children with autism compared with children with SLI represents the same or different underlying mechanisms is not known. However, clarifying areas of similarity and difference between children with ASD, children with other developmental disabilities, and typically developing children continues to be an important line of inquiry for understanding the genetics and underlying neurobiologic and neurocognitive mechanisms involved in autism.

This age difference in group 2 also raises an important measurement issue when working with children with moderate to severe cognitive impairments. Ratio-based IQ scores were used in this study to avoid the problem of a floor effect on low-end scores, differences in mean age among the groups (which serves as the denominator in the scores) may come in to play. For example, there may be specific items on the Mullen that tend to be particularly difficult (e.g., those requiring imitation?) or easy (e.g., those clearly requiring no spoken instruction?) for children with autism in comparison to children in the normative sample functioning at roughly the same developmental level. Simply examining age-equivalence scores rather than the ratio-based scores does not address this issue as functioning at a 20-month-old level will mean vastly different things for a 2- versus 4-year-old child.

The highest functioning group identified by the latent class analysis (group 4) had much higher mean Mullen scores (verbal IQ = 88, nonverbal IQ = 97) compared to group 3 (verbal IQ = 63, nonverbal IQ = 70). Despite this large difference in intellectual functioning these groups did not differ on the Vineland socialization, daily living skills, or motor skills domains. The average standard score across the four Vineland subdomains was 73 for group 4 and 67 for group 3 showing a much bigger relative weakness in adaptive functioning for group 4. In addition, groups 3 and 4 did not differ on any ADOS or ADI scores. Despite the striking difference in levels of cognitive functioning between groups 3 and 4, this did not translate into systematic differences in adaptive functioning or level of autism symptoms as measured by the ADOS and ADI.

We also found that LCA group membership accounted for significant proportion of the variability of Vineland socialization, communication, daily living skills, and ADI social scores beyond that accounted for by the Mullen verbal IQ and nonverbal IQ scores. This provides additional evidence of the importance of considering IQ in young children with ASD more than simply a single dimension. Indeed, the results presented here suggest that there is more than a simple linear relationship between intellectual functioning and adaptive behavior and autism symptoms, even when independently measuring both verbal and nonverbal intellectual abilities.

This study comprises a very large sample of young children who have been carefully diagnosed as being on the autism spectrum, and thus, we believe the findings here should generalize to the broader population of children with ASD in this age range. However, this sample was not collected as a population-based epidemiologic sample. Replication is clearly needed. One should be particularly cautious in making assumptions regarding the relative size of each of the IQ groups described. Of the 1,824 Mullen subscales (4 subscales X 456 children) only 11% represent scores in the “average” range (T-scores of 40 or greater, 16^th percentile or greater). Many studies of older children with ASD report a much higher proportion in the average range of intellectual functioning. There are several issues that may bear on the representativeness of this sample. First, this sample may truly contain fewer high-functioning children with ASD than are present in the population as families who seek to participate in university studies with their preschooler with ASD may not be representative of all preschoolers with ASD. Second, these children were tested primarily between in the late 1998 and 2001 and perhaps this represents a cohort effect compared to more recent samples. Third, by assessing preschoolers and using a developmental instrument designed for use from infancy through age 5, this sample may have included a greater proportion of more impaired children than is often found in samples of older children. It is often a fine line between getting a valid estimate of a nonverbal child’s level of intellectual functioning and simply being unable to collect valid, interpretable quantitative data. Some of these children will continue to fall further and further behind their same-age peers in terms of their cognitive development. Many well-developed instruments for older children are limited in their ability to measure the low end of intellectual functioning as normative data is simply not available. Thus, there may be portion of the children in the present sample, who by virtue of the difficulties in obtaining meaningful quantitative data, may simply never take part in studies of older children with ASD.

Whether the methods used here or similar statistical methods will provide evidence for latent classes based on IQ in older children remains an open question and one that needs further study. However, careful attention needs to be paid to how the use of a given measurement instrument will impact the scores it yields (e.g., how much variability on the low end of the scale is there?) and even the resulting sample the study produces. This is particularly true when studying verbal information processing and language development in children with autism (Tager-Flusberg, 2000). The likelihood that a sample of preschoolers with autism, when assessed five or ten years later, will present an even wider range of functioning than they did at the outset makes the identification and measurement of the underlying neurocognitive mechanisms at work in autism that much more difficult.. In this pursuit, we must be aware of how our measurement tools, presumably impartial and objective, may also become construction tools that actively shape and influence the phenomenon we wish to study. Any seasoned child assessor knows, a la Heisenberg, that “observing” a child at a table with an open briefcase full of booklets and blocks can have a profound influence on what we end up measuring.