Autism in gifted youth is associated with low processing speed and high verbal ability

Cohort 1: gifted clinical sample

Over a ten-year period (2009-2019), N=1,074 children and adolescents ages six to 18 were evaluated at a university-based clinic that specializes in the assessment and counseling of gifted and twice-exceptional students. Some of the students were evaluated more than one time (1,254 total visits). Although students in the sample had Full Scale IQs as high as 158, a few of the clients evaluated at the clinic had very low IQs (e.g., 55). The average IQ for the cohort was 116.9 (SD of 14.5), the median was 117. Some individuals in the cohort were diagnosed with ASD, whereas others received a different diagnosis such as ADHD or SLD, and many did not have a diagnosis. Thus, from the perspective of cognitive ability, this cohort provided a broader IQ distribution compared to previously cited twice-exceptional research[17–19] in which the samples only included autistic individuals with very high IQ (twice-exceptional), i.e., IQ ≥ 120. Consequently, the current study offers a more inclusive IQ and diagnostic sample: autistic individuals within the average range, or below, individuals with very high IQ and not diagnosed with ASD (i.e., gifted only), and non-autistic individuals with IQs in the average range. Demographic information concerning these participants is presented in Table 1.

Cohort 2: SPARK

SPARK[25] is an ongoing, US-based genetic study of autism of nearly 300,000 individuals. In the current study, we used SPARK data primarily to understand the relationship between age at diagnosis (provided during enrollment) and IQ. In SPARK, two indices of IQ are available for a subset of participants: 1) parent-reported IQ in bins of approximately 10 points (N=17,634) and 2) IQ scores as contained in the electronic health record of participants and uploaded into a central SPARK database by the clinical recruitment centers (N=3,339). For both of these IQ indices, we obtained data from the SPARK phenotype release V7. SPARK is approved by the Western IRB (IRB 20151664). All participants provided informed consent.

Cohort 3: ABCD

The Adolescent Brain Cognitive Development (ABCD) study [26] is a general population sample that does not exclude on the basis of neuropsychiatric diagnoses. The study includes longitudinal data collection across behavioral and neuroimaging modalities and also includes genetic data. For cognitive-behavioral analyses, release 2 data were used (N=10,601). For genetic analyses, the cohort was further filtered to participants in which genetic data were available that passed our quality control and clustered in the majority European cluster based on SNPs (genetic clustering detailed below). After this filtering, 6,559 ABCD participants remained. The ABCD cohort was genotyped on the Affymetrix NIDA SmokeScreen Array and was processed through standard QC steps before release, including removing SNPs with low call rate and individuals with potential contamination problems or high missing data. The SNP QC process was based on the recommendations by [27] using PLINK [28] and R [29] with the same parameters used for the SPARK genotyping QC. In total, 399,016 SNPs and 9,324 individuals passed QC. After this QC, the remaining cohort was merged and clustered with the 1,000 Genomes Phase 3. Clustering was based on the first 10 components from multi-dimensional scaling of the combined kinship matrix of the cohort and 1,000 Genomes. This combined cohort was clustered into 5 groups, representing the 5 distinct super-populations. For genotype imputation, ambiguous SNPs were also removed, leaving 372,694 SNPs. These remaining individuals and SNPs were imputed to the 1,000 Genomes Phase 3 ALL reference panel using the Genipe pipeline [30]. 6,659 individuals clustered in the European cluster. Genipe performed LD calculation and pruning with PLINK [28], genotype phasing with SHAPEIT [31], and genotype imputation by IMPUTE2 [32] using default parameters.

Psychoeducational and diagnostic instruments

Sattler [33] offers a thorough description of the Wechsler Intelligence Scale for Children (5th Edition; WISC-V). Multiple subtests comprise the WISC-V, including ten primary subtests: Block Design, Similarities, Matrix Reasoning, Digit Span, Coding, Vocabulary, Figure Weights, Visual Puzzles, Picture Span, and Symbol Search. From the ten primary subtests, the first seven are used to obtain the Full Scale IQ (FSIQ). In addition to the ten primary subtests, there are six secondary subtests as well as five complementary subtests. Scores from the Similarities and Vocabulary subtests comprise the Verbal Comprehension Index; Block Design and Visual Puzzles subtest scores comprise the Visual Spatial Index; Matrix Reasoning and Figure Weights subtest scores comprise the Fluid Reasoning Index; Digit Span and Picture Span comprise the Working Memory Index, and Coding and Symbol Search comprise the Processing Speed Index. The primary index scores as well as the FSIQ, are reported as standard scores with a mean of 100 and a standard deviation of 15. Standard scores for each of the scales have a mean score of 10 and a standard deviation of 3.

The WISC-V [34] is a widely used measure of intellectual ability with good psychometric properties (see the WISC-V Technical and Interpretive Manual; [35]). Confirmatory factor analysis shows that the WISC-V measures five interrelated but distinct general abilities; thus, empirical data match the hypothesized structure of the test, which is rooted in contemporary intelligence theory [34]. Criterion validity studies demonstrate that the WISC-V can validly be interpreted as a measure of intelligence in children based on intercorrelations with the Kaufman Assessment Battery for Children (KABC-II; the WISC-V FSIQ and KABC-II Fluid Crystallized Index and Mental Processing Index were 0.77 to 0.81, respectively), the Kaufman Test of Educational Achievement (KTEA-3; most correlations were found to be within the Moderate range), the Wechsler Individual Achievement Test (WIAT-III; moderate correlations ranging from 0.40 to 0.73 were found between the WISC-5 FSIQ and WIAT-3 Composite scores), and the Wechsler Abbreviated Scale of Intelligence (WASI-II; correlation coefficients are moderately high, ranging from 0.57 to 0.87). Test-retest reliability of the WISC-V is high, ranging from 0.88 (Processing Speed) to 0.96 (FSIQ and GAI). WISC-5 internal consistency estimates range from 0.81 to 0.94; for the FSIQ, internal consistency across the 11 age groups ranged from 0.96 to 0.97.

The Behavior Assessment System for Children, 3rd Edition (BASC-3; [36]) is described as “a multidimensional system used to evaluate the behavior and self-perceptions of children and young adults ages 2 through 25 years” (p.1). There are multiple components, including a self-report (SRP), a teacher-report (TRS), and a parent-report (PRS). Each of the three forms is further divided into Pre-school (age 2-5), Child (ages 6 through 11 years), Adolescent (ages 12 through 18 years), and College (18-25) forms that contain similar subscales to facilitate comparisons between forms; for 6-to 7-year-old students, the self-report form is administered as an interview (SRP-I). Scores are reported as T-scores with a mean of 50 and standard deviation of 10. There are two types of scales: clinical scales (measuring behavioral and emotional concerns) and adaptive scales (measuring strengths or protective factors). Among the clinical scales, scores of 60 through 69 are considered At-Risk and scores of 70 and above are considered Clinically Significant. Among the adaptive scales, scores of 31 through 40 are considered At-Risk, whereas scores of 30 and below are considered Clinically Significant.

There is extensive research supporting the validity and reliability of the BASC-3 in clinical settings [36]. Reliability measures across the TRS, PRS, and SRP suggest good internal consistency for both scale and composite scores. Median internal consistency for the BASC-3 TRS yielded alpha coefficients between .87 and .91 for the clinical and adaptive scales, and between .95 and .97 for the composite scales [36]. Median internal consistency for the BASC-3 PRS yielded alpha coefficients between 0.83 and 0.89 for the clinical and adaptive scales, and between 0.93 and 0.97 for the composite scales [36]. For the SRP, alpha coefficients were between 0.82 and 0.86 for the clinical and adaptive scales, and between 0.93 and 0.95 for the composite scales. Test-retest reliability was also high across the TRS, PRS, and SRP forms. Test-retest reliability for the TRS ranged from 0.87 to 0.91 for the clinical and adaptive scales, and from 0.95 to 0.97 for the composite scales. The PRS had test-retest reliabilities ranging from 0.83-0.89 for the clinical and adaptive scales, and between 0.93 and 0.97 for the composite scales. For the SRP, the clinical and adaptive scales had test-retest coefficients between 0.82 and 0.86, as well as composite scale coefficients between 0.93 and 0.95. The validity of the BASC-3 is substantially supported by the results of factor analyses consistent with scale composition, strong correlations with other instruments that assess emotional and behavioral symptoms in children (e.g., the Minnesota Multiphasic Personality Inventory and the Achenbach System of Empirically Based Assessment), and consistency between the results of this measure and the clinical diagnoses of the child being assessed. Additionally, there is good consistency between the BASC-3 and previous versions of the BASC, which have been used in hundreds of research studies in various settings, countries, languages, and clinical populations [36].

All BASC-3 clinical and adaptive skills were included in the present analysis. The BASC-3 scales of interest in this study included the TRS, PRS, and SRP Anxiety scale, the TRS School Problems scale, the PRS Functional Communication and Attention Problems Scales, and the SRP Sense of Inadequacy Scale and Hyperactivity scales.

Cognitive ability and age at diagnosis

Data were pulled from the SPARK version 7 phenotype release (N=17,634 in this sample). Parents reported the approximate full-scale IQ (in bins spanning approximately ten points each) of their children as obtained through clinical testing. Age at professional diagnosis of autism was also reported by parents. We grouped these SPARK participants into three groups according to the parent-reported IQ: high (IQ ≥ 120), average (IQ 70-119) and low (IQ < 70). See Figure 2. Using age of six years as a threshold for diagnosis (corresponding with typical entry into the public school system), we compared diagnosis and non-diagnosis counts in the low and high IQ groups using a 2×2 table and Fisher’s exact test. Similarly, low or high IQ group membership was stratified by a binary “diagnosed by age 18” variable, indicating whether the individual would have been diagnosed by a child/adolescent or adult clinic. Odds ratios, their 95% confidence intervals, and p-values were recorded to assess significance of these tests of association.

• Download figure
• Open in new tab

Figure 1. Overview of study design.

Our discovery sample consisted of N=1,074 clients seen at the Belin-Blank Center (B-BC), a specialty clinic at the University of Iowa that serves gifted individuals. We digitized and entered the psychoeducational data from this sample into a REDCap database, which we then used to conduct hypothesis-driven and unsupervised learning analyses with the aim of gaining insight specifically into the cognitive, behavioral, and daily living challenges of “twice-exceptional” (2e) individuals: those with exceptional cognitive ability and having a diagnosis relevant to mental health. For the purposes of this study, we focus on autism, though twice-exceptionality extends beyond autism. Finally, to test the robustness and generalization of our findings, we used data from the ABCD study (N=10,601). We also used the genetic data in ABCD to compute polygenic risk scores for autism, which were then tested for association with traits of interest.

• Download figure
• Open in new tab

Figure 2. High IQ delays autism diagnosis.

In a sample from the SPARK autism study, N=17,634 autistic individuals had a parent-reported IQ and age at autism diagnosis (A). To illustrate the relationship between IQ and age at diagnosis, we grouped these individuals into broad groups of low (< 70), average (70-119) and high (>= 120) IQ. Individuals in the high IQ group showed significantly delayed diagnosis compared to their low IQ counterparts, with high IQ individuals entering adulthood without a diagnosis at twice the rate of low IQ individuals, and entering school age (i.e., age 6 years) without a diagnosis at 5.2 times the rate of low IQ individuals. To address the sampling and reporting biases in this larger sample, we also examined the same comparisons in a smaller sample (N=3,339) with full scale IQ scores pulled from clinical records available to SPARK recruitment sites (B). These results follow the same overall trend as in A, but with much stronger effects: high IQ individuals enter adulthood without an autism diagnosis at 14.5 times the rate of their low IQ counterparts, and they enter school age (i.e., age 6 years) without a diagnosis at more than nine times the rate of their lower-IQ counterparts. Together, these results suggest significant delayed diagnosis and more importantly, crucial years of development spent without services and support for their disability.

• Download figure
• Open in new tab

Figure 2. Interactions between IQ and ASD diagnosis are associated with specific anxiety and attention-related problems.

Using psychoeducational data from clinical records, we tested scores from the BASC for association with IQ, ASD diagnosis, and their interaction (with sex and age as additional covariates). Most strikingly, the interaction between ASD and IQ leads to an increased sense of inadequacy (A) as IQ increases in those diagnosed with ASD (red points). Similarly, increasing IQ in autistic individuals is associated with increased anxiety according to self (B), parent (C), and teacher report (D). The remaining tests of association are summarized in (E), where t-statistics are shown, with FDR<0.05 indicated by a circle.

Statistical analysis

Both BASC and IQ scores were standardized (mean 0 and SD 1) within each decile of age. For each measure, Analysis of Variance (ANOVA; anova() and lm() functions in R version 3.3) was utilized to test the significance of main effects of ASD diagnosis and full-scale IQ, as well as their interaction. In the linear model underlying the ANOVA, individuals were re-weighted such that for individuals with FSIQ < 120, their contribution to the least-squares fit was proportional to the pointwise estimate (based on their FSIQ) of density in a Gaussian distribution with mean 100 and SD 15. Collectively, individuals with FSIQ >= 120 were each weighted the same, and the sum of these weights equaled the sum of the weights for individuals with FSIQ < 120. This reweighting has the effect of emphasizing individuals whose FSIQ is close to the mean of 100, thus making the resulting comparisons with individuals with FSIQ > 120 more in line with what would be expected when comparing with a typical population sample. Although this approach had the effect of boosting power and resulted in a greater number of significant associations, in general the direction of association was congruent with an unweighted approach. The individual-level weights, together with the age-standardized IQ and BASC scores used in these analyses are included in Supplemental Table 1. For all ANOVA tests, adjustment for multiple testing was accomplished using the Benjamini-Hochberg procedure [37] and is reported as false discovery rate (FDR) where applicable.

Unsupervised learning

We employed archetypal analysis [38] as implemented in the archetypes package for R. Input features were the Z-scaled index scores and subtest scores of the WISC-IV and WISC-V. These were imputed with the median value where missing (12% missingness). To minimize the impact of outliers, the archetypal analysis procedure was trained on 5,000 random mixtures of the original data, where the original data matrix was multiplied by a random mixing matrix following a Poisson distribution with rate=1. The values of this random mixing matrix were normalized such that the contribution from each of the 1,074 individuals summed to 1 for each mixture. The archetypal coefficients for the observed data were then inferred using the predict() function of the archetypes R package. The optimal number of archetypes was inferred using the elbow plot heuristic, which suggested diminishing returns after k=5 archetypes. Individuals were assigned to archetypes using the following heuristic: briefly, an individual is assigned to an archetype if the coefficient for that archetype is the maximal coefficient for that individual and it is more than twice the next-highest archetypal coefficient. Those individuals who did not meet this criterion were grouped together in a “mixed” group. Subsequent analyses used archetype coefficients as explanatory variables in linear models testing association with achievement and behavioral indices available in the database. To test association of the PS/VC discrepancy with BASC behavioral measures, we encoded a predictor variable as follows: 1 if the individual was assigned to A4, 3 if the individual was assigned to A5, and 2 for all others. This variable represents the ordinal progression from PS+/VC-(A4) to PS-/VC+ (A5). The models used for significance testing were therefore BASC ~ sex + FSIQ + PS_VC, where BASC is the measure in question from the BASC, FSIQ is the full-scale IQ, and PS_VC is the ordinal variable described above.

Replication analysis

To determine whether key behaviors were robustly associated with a PS/VC discrepancy in an independent sample, we utilized data from the large ABCD study, a US-based general population study of adolescents with extensive behavioral, genetic, and neuroimaging data [39]. Specifically, we constructed a PS/VC “discrepancy score” by multiplying participants’ NIH toolbox scores for the pattern task (processing speed), the picture vocabulary task (verbal), and the reading task (verbal), by −1, 0.5, and 0.5, respectively, and then summing. These values were then Z-scaled to mean of 0 and standard deviation of 1. Individuals with large positive values of this score would have a large PS/VC-type discrepancy similar to that observed in A5 in the B-BC sample (Fig. 3A). Likewise, ABCD participants with an extreme negative score would have a similar discrepancy to individuals belonging to the A4 archetype in the B-BC sample (Fig. 3A). We note that there would have been a number of ways to approach this analysis (e.g., factor analysis, canonical correlation analysis, etc.). We investigated these alternative approaches and found that they yielded similar findings to those described below. However, given that the PS/VC discrepancy score based on the simple difference in task scores is more straightforward and directly interpretable, we chose that in the end. To identify behavioral correlates of the PS/VC discrepancy in ABCD, we tested associations between the ABCD PS/VC discrepancy score and subscales of the Child Behavior Checklist (CBCL)[40]. In total, twenty subscales derived from CBCL items are available in the ABCD data release (see Figure 4C). We tested linear models of the form y ~ sex + total_comp + PS_VC, where y is the CBCL subscale score under investigation, total_comp is the NIH Toolbox total composite score, and PS_VC is the PS/VC discrepancy score described above.

• Download figure
• Open in new tab

Figure 3. Unsupervised learning uncovers a discrepancy between processing speed (PS) and verbal comprehension (VC) that is significantly associated with autism diagnosis.

We performed archetypal analysis, which yielded five distinct cognitive patterns or archetypes that best typify the patterns seen in the B-BC sample (A). In particular, archetype 4 (A4) and archetype 5 (A5) are nearly mirror-images of each other, with A5 showing a PS deficit and VC strength (PS-/VC+), and A4 showing a PS strength and VC deficit (PS+/VC-). A5 showed a significant over-representation of autistic individuals (OR=2.46, 95% CI: [1.5,4.1]), while A4 was significantly depleted (OR=0.39, 95% CI: [0.16,0.94]). Taken together, this suggests an axis of risk and resilience, defined by a PS/VC discrepancy. We tested subscales from the BASC for association with the PS/VC discrepancy (B) to identify specific problem areas linked to this cognitive pattern, highlighting attention and internalizing problems similar to those identified in the diagnosis-driven analysis (see Figure 2).

• Download figure
• Open in new tab

Figure 4. Generalization and polygenic association of the PS/VC discrepancy in the ABCD study.

We used scores from the NIH Toolbox available from ABCD to represent verbal comprehension (mean of vocabulary and reading tasks) and processing speed (pattern task) (A). The difference between verbal and processing speed scores (PS/VC discrepancy) was Z-scaled and is depicted in color in panel A. The PS/VC discrepancy score is positively correlated with the NIH Toolbox total composite score, but this effect is driven largely by individuals at the higher end of this score (B), suggesting that the PS/VC discrepancy is particularly relevant for individuals with high overall cognitive ability. Computed subscales from the CBCL are correlated with each other to varying degrees, and our clustering identified loose super-clusters of internalizing and externalizing subscales (C). These CBCL subscales were tested for association with the PS/VC discrepancy score in order to test the generalization of our observations from the B-BC sample (D). We found evidence suggesting that PS/VC is significantly associated with anxiety and other internalizing symptoms, as well as attention problems, in agreement with our initial findings. Finally, we found that PS/VC is associated with increasing polygenic risk for autism (panel E, p = 0.004), in agreement with our initial observation of over-representation of autism in A5 (the PS-/VC+ cognitive archetype, see Figure 3A).

Polygenic score analysis

Polygenic scores (PGS) for autism diagnostic status were computed in the ABCD sample (N=6,002) with LDPred2 [41] as described in [42]. Within ABCD, PGS were binned into three groups based on percentiles: depleted risk (20^th percentile and below), neutral risk (21^st-79^th percentile), and high risk (80^th percentile and above). These three groups were coded with integer values 1, 2, and 3, respectively and were used as a quantitative explanatory variable. Sex and total composite score were used as additional covariates in a linear model predicting PS/VC discrepancy: PS_VC ~ sex + total_comp + PGS, where PS_VC is the PS/VC discrepancy score described above, total_comp is the NIH Toolbox total composite score, and PGS is the integer-coded polygenic score for autism. Tests of significance were performed using the fitted slope and its standard error (i.e., the summary() function in R).

A discrepancy between processing speed (PS) and verbal comprehension (VC) is associated with autism in the gifted clinical cohort

The central goal of our archetypal analysis was to uncover recurrent and distinct patterns of cognitive ability, and then determine whether any of those patterns are associated with autism diagnosis and other behaviors as observed from the BASC. This analysis is complementary to our diagnosis-driven analysis primarily because it utilizes an unsupervised learning approach and is agnostic to the diagnosis, which is only considered in a post-hoc analysis. Our analysis supported the presence of five cognitive archetypes in our sample (Figure 3). These archetypes correspond, approximately, to uniformly elevated ability (archetype 1, A1), generally lower ability with a specific strength in perceptual reasoning (archetype 2, A2), uniformly decreased ability (archetype 3, A3), strength in processing speed (PS) coupled with weakness in verbal comprehension (VC) (archetype 4, A4), and weakness in processing speed coupled with strength in verbal comprehension (archetype 5, A5). A total of 42 individuals did not meet criteria for exclusive assignment to any archetype and were considered “mixed”. Summary statistics for each of these archetypes are provided in Table 3. Notably, A4 and A5 are of particular interest because they are mirror images of each other (A4: PS+/VC− and A5: PS−/VC+). Further, these two cognitive archetypes show differential depletion and enrichment for autistic individuals, respectively (OR=0.39, 95% CI: [0.16,0.94] for A4 and OR=2.46, 95% CI: [1.5,4.1] for A5).

Higher IQ delays autism diagnosis

In SPARK (N=17,634), individuals with parent-reported IQ of 120 or above were 5.2 times more likely than the low IQ group (IQ < 70) to be diagnosed after age six (OR=5.2 [4.7,5.2], p < 2.2×10⁻¹⁶), and twice as likely to be diagnosed in adulthood (i.e., age ≥ 18 years; OR=2.0 [1.2,3.5], p=0.01). See Figure 2. This parent-report data is limited by the lack of adults (most autistic adults who join SPARK do so without the involvement of their parents and also lack an option to self-report IQ in a way analogous to the parent-reported IQ), leading to over-convergence of diagnosis rates between the low and high groups in adulthood. To complement the analysis of the parent-reported IQ data, we also examined data from N=3,339 SPARK participants whose IQ scores were entered into their clinical records by medical providers in connection with a clinical IQ test, and subsequently gathered by SPARK recruitment centers who have consented access to these data. This analysis was in agreement with the overall trend observed in the parent-report data, but with much more pronounced effects: we found that individuals with IQ >=120 entered adulthood without a diagnosis at nearly 15 times the rate of their lower-IQ counterparts (OR=14.5 [2.8, 94.6], p=0.0005). Similarly, these high IQ individuals entered school age (i.e., age 6 years) without a diagnosis at more than nine times the rate of their lower-IQ peers (OR=9.2 [6.1,14.0], p < 2.2×10⁻¹⁶). Taken together, these converging results suggest that autistic individuals with exceptional ability spend proportionally less of their developing years without access to supports and services for their disabilities. Furthermore, the observation that an over-abundance of these high IQ autistic individuals do not receive a diagnosis until adulthood (and are therefore beyond the reach of autism experts in child psychiatry clinics) has ramifications for clinical practice in adult psychiatry, which often lacks specific expertise in neurodevelopmental conditions such as autism. This mismatch holds the potential for increased rates of misdiagnosis.

Anxiety and attention problems increase with FSIQ specifically in autistic individuals in the gifted clinical cohort

The primary motivating question for our diagnosis-driven analysis was to understand whether increasing IQ might in some way be a liability for those with an autism diagnosis, when compared both with their non-autistic, high IQ peers, as well as their autistic but average IQ peers. This goal was operationalized by examining the interaction term in a linear model of functional outcomes (from the BASC) predicted by autism diagnosis, full-scale IQ, and their interaction (see Supplemental Table 2). Sex and age were included as additional covariates. As shown in Figure 3, we found that self-report of a sense of inadequacy had the strongest ASD*IQ interaction (beta=0.3, 95% CI:[0.15,0.45], p=7.1×10⁻⁵), indicating that autistic individuals show a 0.3 SD increase in this measure for each SD increase in FSIQ. Furthermore, anxiety was a point of confluence between self (beta=0.15, 95% CI: [0.01,0.29]), parent (beta=0.17, 95% CI: [0.05,0.29]), and teacher report (beta=0.14, 95% CI: [0.02,0.27]), though only parent report survived multiple testing correction at FDR < 0.05 (see Figure 3 A-D). Other significantly associated (all FDR<0.05) outcomes include parent reported functional communication (beta=-0.16, 95% CI: [-0.27,-0.05]) and attention problems (beta=0.16, 95% CI: [0.05,0.27]), teacher-reported school problems (beta=0.22, 95% CI: [0.07,0.37]), and self-reported hyperactivity (beta=0.22, 95% CI: [0.06,0.38]). To determine whether any of these associations would be considered potentially clinically meaningful, we dichotomized the BASC scores using the suggested threshold of >60 (or < 40 for reverse-scored variables). We found that teacher-reported school problems were nearly four times as likely to be clinically meaningful (i.e., by the thresholds indicated above) for an autistic individual with full-scale IQ of 120 vs. 100 (OR=4.2, 95% CI: [2.1,8.4]). Similarly, self-reported sense of inadequacy was three times as likely to be clinically relevant for autistic individuals at FSIQ 120 vs. 100 (OR=2.9, 95% CI: [1.4,5.9]), and parent reported functional communication issues more than twice as likely (OR=2.3, 95% CI: [1.6,3.5]).

PS/VC discrepancy is predictive of internalizing, attention, and other problems in the gifted clinical cohort

Fourteen of 46 BASC measures were significantly associated with the PS/VC discrepancy at FDR < 0.05, and the full summary statistics of these tests are available in Supplemental Table 4. Briefly, parent and teacher reports of withdrawal, teacher-reported attention problems, parent- and teacher-reported atypicality, teacher-reported school problems, and teacher-reported depression were all significantly positively associated with PS/VC discrepancy (all FDR < 0.05, see Supplemental Table 4). Parent reports of activities of daily living, parent- and teacher-reported social skills, parent- and teacher-reported leadership skills, teacher-reported functional communication, and teacher-reported adaptability were all significantly negatively associated with PS/VC discrepancy all FDR < 0.05, see Supplemental Table 4). Together, these results paint a picture of the PS/VC discrepancy contributing to problems with attention, internalizing issues (withdrawal, depressive symptoms), poor leadership and social skills, and reduced adaptability, functional communication, and daily living skills. Notably absent are externalizing and somatization symptoms. Also noteworthy is the absence of association with any self-reported measures, which may suggest that those impacted by a PS/VC discrepancy may have little insight into the problems they face, while parents and teachers seem to be more attuned to these challenges.

PS/VC discrepancy is predictive of internalizing and attention problems in both the gifted clinical cohort and the ABCD cohort

Nine of 20 available CBCL subscale scores showed a significant association at a nominal p < 0.05. Summary statistics for the PS/VC discrepancy score association are found in Supplemental Table 5 and the corresponding results are presented in Figure 4. The most striking association of the PS/VC discrepancy is with the “sluggish cognitive tempo” (SCT) subscale of the CBCL [40] (beta=0.05, t=5.2, p=2.6×10⁻⁷). This association, together with other significant associations with attention and ADHD-related subscales (attention: beta=0.03, t=2.8, p=0.005, ADHD: beta=0.03, t=2.6, p=0.009), suggest that PS/VC discrepancy is strongly associated with attention issues in non-clinical samples, in agreement with the findings from our analysis of the clinical sample. Further, we found that ABCD PS/VC is associated with a variety of internalizing symptoms, captured in the anxious/depressed, internalizing, and anxiety disorder subscales of the CBCL (all p < 0.05). Again, this connection of internalizing symptoms with PS/VC discrepancy is consistent with our findings in the clinical sample. Finally, a number of other subscale associations align with observations seen in the clinical sample: associations between PS/VC and OCD and thought problem subscales of the CBCL are congruent with the increased atypicality linked to A5 in the clinical sample, and the association with total problems (p < 0.05) may reflect the functional and daily living associations found in the clinical sample. Overall, these analyses suggest a general convergence in both clinical and non-clinical samples of the PS/VC cognitive discrepancy onto a combination of attention problems, internalizing symptoms, and atypicality.

PS/VC discrepancy is associated with polygenic risk for autism in the ABCD cohort

Although the above analysis established the robustness of the behavioral associations with PS/VC discrepancy in an independent sample (ABCD), replicating the clinical sample’s observed enrichment (A5) and depletion (A4) of autism diagnosis is more challenging: ABCD does not have reliable diagnosis information that can be used to robustly identify participants with ASD. Instead, we appealed to the genetic data in ABCD to determine whether individuals with high polygenic risk for autism showed an elevated PS/VC. Indeed, autism PRS is significantly associated with PS/VC (beta=0.07, t=2.9, p=0.004), in agreement with our observation in the clinical sample of a significant excess of autism in A5 (and depletion in A4).