IQ trajectories in autistic children through preadolescence

Abstract Background We extended our study of trajectories of intellectual development of autistic individuals in early (mean age 3 years; T1), and middle childhood (mean age 5 years, 7 months; T2) into later middle childhood/preadolescence (mean age 11 years, 6 months; T3) in the longitudinal Autism Phenome Project cohort. Participants included 373 autistic children (115 females). Methods Multivariate latent class growth analysis was used to identify distinct IQ trajectory subgroups. Baseline and developmental course group differences and predictors of trajectory membership were assessed using linear mixed effects models for repeated measures with pairwise testing, multinomial logistic regression models, and sensitivity analyses. Results We isolated three IQ trajectory groups between T1 and T3 for autistic youth that were similar to those found in our prior work. These included a group with persistent intellectual disability (ID; 45%), a group with substantial increases in IQ (CHG; 39%), and a group with persistently average or above IQs (P‐High; 16%). By T3, the groups did not differ in ADOS‐2 calibrated severity scores (CSS), and there were no group differences between Vineland (VABS) communication scores in CHG and P‐High. T1‐T3 externalizing behaviors declined significantly for CHG, however, there were no significant T3 group differences between internalizing or externalizing symptoms. T1 correlates for CHG and P‐High versus ID group membership included higher VABS communication and lower ADOS‐2 CSS. A T1 to T2 increase in VABS communication scores and a decline in externalizing predicted CHG versus ID group membership, while T1 to T2 improvement in VABS communication and reduction in ADOS‐2 CSS predicted P‐High versus ID group membership. Conclusions Autistic youth exhibit consistent IQ developmental trajectories from early childhood through preadolescence. Factors associated with trajectory group membership may provide clues about prognosis, and the need for treatments that improve adaptive communication and externalizing symptoms.


INTRODUCTION
Given the heterogeneity of autism (Geschwind & Levitt, 2007), it remains difficult to provide reliable answers about what the future holds for young autistic children. Some never acquire functional spoken language, sustain close interpersonal relationships outside of family members or caregivers, or live independently. Others develop meaningful reciprocal friendships, obtain post-secondary education, and work and live in the community (Mason et al., 2021). Some even "lose" their autism diagnoses (Fein et al., 2013). Intellectual ability level, as assessed using IQ or a developmental quotient (DQ) (both referred to as IQ) is perhaps the most significant predictor of outcomes across key life domains for autistic individuals (Miller & Ozonoff, 2000;Munson et al., 2008). Early IQ also is the strongest predictor of adult outcomes in autistic individuals .
While there have been multiple studies examining the association between intellectual functioning in childhood and later outcomes, few have been longitudinal and fewer still have investigated IQ-based subgroups/phenotypes using data-driven or clinicallybased clustering strategies. A first study to isolate IQ-based subgroups using data-driven methods idenified four unique groups based on IQ level and relative strength of verbal versus non-verbal abilities in 2-5 1/2 year olds (Munson et al., 2008). Two subsequent studies employed clinical grouping methods. The first examined a prospective longitudinal cohort of 85 children assessed at 2, 3, and 19 years (n = 85). They used age-19 IQ to group participants into VIQ<70 and VIQ>70 sub-groups who did and did not retain their diagnoses.
Eighty five percent of the group remaining intellectually disabled could be identified from early IQ scores. Participants losing their autism diagnosis received more early intervention and exhibited early reduction in restricted and repetitive behaviors . The second study using a clinical grouping approach assessed participants at ages 2 and 13 years, assigned children to a best outcomes (IQ > 80 with no diagnosis of autism by the second assessment; 16%), more able (IQ > 80 throughout, 20%), and more challenged (IQ<80%; 63%) groups (Zachor & Ben-Itzchak, 2020). The more challenged group showed decreased cognitive ability and increased social and repetitive behavior severity over time.
To the best of our knowledge, a study by our group has been the only prospective longitudinal study to use an empirical data-driven approach to isolate developmental trajectories of intellectual functioning in children as young as ages 2-8 years old (Solomon et al., 2018). Four distinct groups were identified. Two had persistent intellectual disability (ID) (43% of the sample), 1 had IQs starting in the intellectual disability range that then increased by at least 2 standard deviations (35%), and 1 had IQs remaining in the average or better range over time (22%). Communication and social adaptive functioning lagged IQ in all autism but not non-autistic groups. While internalizing symptoms decreased over time for all groups, externalizing symptoms declined only for the group experiencing substantial increases in IQ.
The current study aims to extend our past investigation of trajectories of IQ development in one of the few relatively large, cognitively heterogeneous, and recent longitudinal cohorts-the Autism Phenome Project (APP)-by adding a third data point from our middle childhood assessment and by investigating additional developmental issues pertinent to the preadolescent developmental period. We again isolate phenotype groups based on IQ and characterize them based on autism symptoms, communication adaptive functioning, and problem behavior symptoms including internalizing and externalizing. To gain insight on predictors of later childhood/pre-adolescent outcomes, we then investigate variables assessed at or before T1 and changes between variables assessed at T1 and T2. These analyses focus on group differences in potential predictors for children who remained in the ID group versus those who did not by T3.

METHOD Participants
Participants were members of the longitudinal APP cohort, which began recruiting both autistic and typically developing children

Key points
� In this study of the intellectual development of autistic individuals from early childhood through age 12, we found there were three IQ trajectories-a group with intellectual disability from early childhood through preadolescence (ID; 45%), a group whose IQs increased at least 1 standard deviation referred to as Changers (CHG; 39%) and a group whose IQs were in the average or above range through the period (P-High; 16%).
� Although autistic youth exhibited lower adaptive functioning than would be expected based on IQ. By preadolescence, there were no significant group differences between adaptive communication in the CHG or P-High groups.
� Early correlates for being in the CHG or P-High groups versus the ID group, included stronger early VABS communication scores and lower ADOS CSS.
� Improved communication adaptive functioning and decreased externalizing between T1 and T2 was a marker of becoming a member of CHG versus ID, while reduced ADOS-2 CSS and improved adaptive communication were predictive of being in P-High versus ID.
� Findings suggest that early communication adaptive functioning and may be a stronger prognostic marker than IQ scores, and that communication adaptive functioning and externalizing symptoms may be treatment targets that are associated with later improvements in intellectual ability levels.
through an internal data base and advertisements placed with local providers and other organizations and groups known to be involved with young autistic children and their families starting in 2006.
Baseline assessments were conducted in children at 2-5 years of age, followed by longitudinal assessments across childhood. Four total assessments have been completed. A fifth is in progress and the cohort has been expanded. To increase female representation within the APP cohort, we initiated the Girls with Autism-Imaging of Neurodevelopment (GAIN) study in 2014. All participants in the GAIN study are automatically included in the APP dataset. This explains why the gender ratio in new participants is enriched for females. Inclusion criteria for autism were based on the NIH Collaborative Programs of Excellence in Autism as described in our prior study (Solomon et al., 2018). Although the full cohort included TD children, we examined IQ trajectory classes within the autistic group and thus excluded TD participants from analyses. IQ/DQ assessments were completed at three of these assessment points, which we refer to as T1 (mean age = 3.0 years, SD = 0.5, n = 373); T2 (mean age = 5.6 years, SD = 0.9, n = 154); and T3 (mean age = 11.5 years, SD = 0.9, n = 116). One hundred and eighty-two autistic participants had IQ data only at T1, 112 participants had data at two timepoints (T1 and T2: 75, T1 and T3: 37), and 79 participants had data at all three timepoints. See Table 1 for a summary of demographic and clinical characteristics including the IQ scores of the entire sample across the three assessments. We included all autistic APP participants with IQ data at T1 in our analyses. Supplementary Table S1 compares the demographic and clinical characteristics of the participants with complete data versus those with only 1 follow-up visit and those with only baseline data to illustrate their similarity to the entire sample. We did not find a systematic pattern of IQ differences for children having fewer visits as compared to those with complete data. The only other observed characteristic significantly related to missingness was sex, because the most recent participants were from the GAIN cohort. Thus, sex was included as a covariate in all models.

Statistical analyses
We first identified distinct IQ-based subgroups and their differential developmental trajectories by conducting a latent class growth analysis (LCGA) of autistic participants' full-scale IQ scores using Mplus 8 (Muthen, 2017). All participants with at least one timepoint were included (n = 373), and both linear and quadratic age-based models were evaluated for best fit. Models were estimated using full-information maximum likelihood, which permitted us to include the participants with missing data, under the missing-at-random assumption. Information-heuristic (e.g., information criterion values) and inferential (e.g., likelihood ratio tests) relative fit comparisons were used to select the best-fitting solution. Information-heuristic indices include the Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC), and sample size-adjusted BIC (SBIC), for which lower values indicate better fit, as well as the approximate Bayes Factor (BF) (Wasserman, 2000). BF compares a larger model with a smaller one and a higher score indicates the larger model is the more probable correct model (values between 1-3 represent weak, 3-10 moderate, and >10 strong evidence for the larger model). As an inferential index, we used the approximate correct model probability (CMP) (Schwarz, 1978), which compares a single model versus all other models under consideration; models with a CMP >0.10 should be considered as candidate models. We used the highest posterior probability from the best fitting model to assign each participant to their most likely subgroup.
Next, we examined differences in trajectories of clinical characteristics for the identified subgroups using linear mixed effects models (Laird & Ware, 1982)   performed 100 times (i.e., for each draw) and results were combined across draws using standard methods for multiple imputation for missing data (Rubin, 1987). The same strategy was employed to examine the robustness of the predictors of trajectory membership.

RESULTS
At T1, a significant proportion of autistic participants achieved the lowest possible MSEL standard score so verbal, nonverbal, and full- "Persistently High IQ" [P-High]) presented a trajectory that demonstrated relative stability with a gradual increase during childhood.
See Figure 1. The average assignment probabilities for the subgroup classes were 0.80, 0.85, and 0.86, respectively. Group membership was very similar to that identified in our previous manuscript using data from T1 and T2 only (Table S3). Demographic and clinical characteristics for all subgroups across the three timepoints are presented in Table S4.
To affirm that the IQ increases of the CHG and other groups did not simply reflect language acquisition and the consequent increase in VIQ, we examined those participants with FSIQ changes of 15 points or more (1 standard deviation) from T1 to T3. Notably, for CHG, 89% also showed increases in NVIQ, while 84% experienced changes in both VIQ and NVIQ (Table S5). We also completed trajectory analyses using NVIQ and VIQ. Here we found that 87.8% of those categorized in CHG in the current analysis would continue to be if NVIQ were used. These percentages were 55.9% for P-High and 82.6% for ID. Values were all over 80% when VIQ was used (Table S6).
(2) ADOS-2 Calibrated Severity Score (CSS): Parameter estimates for all mixed-effects models fitted to clinical variables and adjusted for sex are summarized in Table S7. In the CHG and P-High groups, ADOS-2 CSS decreased from T1 to T2 although it returned to the T1 levels by late middle childhood/ preadolescence (T1 vs. T3, CHG: p = 0.95; P-High: p = 0.94). For the ID group, ADOS-2 CSS scores remained consistent from T1 to T3 (p = 0.98). By T3, the three groups did not differ in ADOS-2 CSS. See Figure 2A.
The developmental pattern of autism symptom severity change has been studied previously by our group in a smaller sample not including all data points (Waizbard-Bartov et al., 2022). The current results did not overlap with this other study given that there were no significant associations between IQ trajectory membership and their three groups (defined by increasing, decreasing, and stable calibrated ADOS-2 CSS). See Table S8.
(3) Communication Adaptive Functioning: From T1 to T3, CHG significantly increased in VABS communication score (p = 0.03), while ID decreased and P-High remained relatively stable (ID: p < 0.001, P-High: p = 0.11; Figure 2B). Thus, while differences between the three subgroups were present at T1, CHG and P-High showed no communication score differences by T3 (p = 0.66), despite their being significantly higher than ID (both p < 0.001).
(4) Internalizing and Externalizing Symptoms: The three autistic subgroups had similar CBCL internalizing subscale scores at T1.
By T3, the score for the ID group decreased, although this reduction was not significantly different than that found in the other groups, and there were no group differences in scores at T3 (after adjusting for multiple comparisons, all p > 0.06, Figure 2C). On the externalizing subscale, the three groups also had comparable scores at T1. The CHG group showed a significant externalizing score decline from T1 to T3 (p < 0.001), however, here too, none of the groups differed on this variable at F I G U R E 1 IQ trajectories of the three full-scale IQ subgroups: Changers (CHG), persistently high IQ (P-High) and persistent intellectual disability (ID).
(5) Demographic Characteristics, and Loss of Diagnosis: The three autism subgroups did not differ in sex composition or maternal and paternal age at childbirth (  Sensitivity analysis results (Supplementary Tables S9 and S10) supported the primary analyses. While the magnitude of the estimates generally slightly decreased after accounting for uncertainty in group assignment, all primary analysis findings remained significant.

DISCUSSION
We extended the study of the trajectories of intellectual development of autistic individuals into late middle childhood/preadolescence in the cognitively heterogeneous APP cohort. Consistent with our prior work, autistic participants were assigned a group with intellectual disability from early childhood through preadolescence (ID; 45%), a group whose IQs increased substantially during early childhood referred to as Changers (CHG; 39%) or a group whose IQs were in the average or above range through the period (P-High; 16%). Unlike our prior study where P-High ADOS-2 CSS scores declined, the new groups did not differ with respect to autism severity at T3. Between middle childhood and preadolescence, VABS communication scores increased in CHG, decreased in ID, and stayed the same in P-High, such that there were no T3 group differences between CHG and P-High. T1-T3 externalizing declined significantly for CHG, although, there were no T3 group differences for internalizing or externalizing.
T1 correlates for CHG and P-High versus ID group membership at T3 included higher VABS communication and lower ADOS-2 CSS. A T1 to T2 increase in VABS communication scores and a decline in externalizing predicted CHG versus ID group membership at T3, while a T1 to T2 improvement in VABS communication and a reduction in ADOS-2 CSS predicted P-High versus ID group membership.
The rapid IQ gains in the CHG group we found in prior work slowed after middle childhood. While this is not consistent with two recent studies that report average mean IQ improvements through adolescence (Prigge et al., 2021;Simonoff et al., 2020), these studies examined mean differences versus trajectories, and Prigge et al.
investigated only intellectually able participants. Also noteworthy is that the positive T1-T2 autism symptom severity and communication adaptive functioning changes in CHG and P-High also slowed between T2 and T3. Waizbard and colleagues observed a similar pattern when they focused on autism symptom severity (Waizbard-Bartov et al., 2022). While we cannot entirely rule out that the reversion back to original scores was a statistical artifact, this pattern was not present for all measures or groups, providing support for a true reversion. Perhaps the complexity of social and cognitive developmental tasks of early adolescence expose more autism related traits, resulting in relative skill declines. In fact there is a growing consensus that the period of transition to school may be a turning point in autistic development (Georgiades et al., 2022) with age 6 representing a time of plateauing in early symptom improvement.
Only the CHG group experienced significant reductions in externalizing symptoms between T1 and T3. While internalizing scores in P-High and CHG did not increase with the beginning of adolescence as might be expected (Solomon et al., 2012), the ID group experienced some reduction in these symptoms, as has been found by others (Edirisooriya et al., 2021). However, it is not clear that internalizing symptoms, and especially anxiety, can be well measured for children with intellectual disability (Kerns et al., 2021), so these findings must be interpreted with caution.  (Duncan & Bishop, 2015). In fact, in our sample, although T1 IQ and the VABS communication were highly correlated overall (r = 0.70), correlations between the VABS and IQ differed substantially for the trajectory groups, ranging from r = 0.5 for ID; 0.46 for P-High; to 0.25 for CHG.
Another clinically interesting observation with prognostic implications was that, contrary to popular clinical belief, language acquisition and VIQ change were not the sole drivers of overall intellectual development. Instead, we found that for CHG, 89% showed increases in NVIQ, while 84% experienced changes in both VIQ and NVIQ (Table S4). Thus, NVIQ did not become stable by age 3 in most participants, and even individuals with moderate mental disability could become members of CHG.
A second set of clinically and potentially intervention-relevant markers were those associated with T1-T2 changes. Here we found that T1-T3 increases in VABS communication scores rendered the CHG and P-High groups equivalent by T3 and distinguished both from ID. We also found that T1-T2 increases in externalizing symptoms were more characteristic of the CHG versus the ID group, and that T1-T2 decreases in ADOS CSS were more characteristic of participants in the P-High group. Although the precise cause and effect associations between IQ, adaptive functioning, externalizing, and autism symptoms remain unclear, our results suggest that each of these areas can improve. Furthermore, they may be critical treatment targets that drive the development of intellectual functioning, and fortunately, effective interventions in these areas have been developed (Kenworthy et al., 2014;Kim et al., 2021;Solomon et al., 2008).
This study had several limitations. First, some have shown that the DAS and MSEL are not entirely comparable, especially in the middle IQ ranges (Farmer et al., 2016). Although, others find no systematic differences (Bishop et al., 2011). Additionally, neither measure does a good job of assessing profound intellectual disability, requiring us to use floor scores for 28 participants. Second, while the LCGA and linear mixed-effects models used were able to handle missing data and produce valid results in the presence of data missing at random, their results may be biased if missingness depends on the missing values themselves. While formally testing whether the assumption of missingness at random (MAR) holds would require data from non-responders, our examination of missingness thus far suggests that MAR may be a plausible assumption here. Finally, by focusing on IQ, we adopted a very narrow definition of future outcomes. Recent studies have rightly encouraged broadening the meaning of outcomes Mason et al., 2021).
In conclusion, we showed that autistic youth from our middle childhood assessment continued to display IQ trajectories that were similar to those we observed earlier in childhood. We identified early and ongoing correlates of late middle childhood/preadolescence outcome which hold the potential to provide critical information related to prognosis and treatment development.