Diagnostic shifts in autism spectrum disorder can be linked to the fuzzy nature of the diagnostic boundary: A data-driven approach

Abstract

Background:

Diagnostic shifts at early ages may provide invaluable insights into the nature of separation between autism spectrum disorder (ASD) and typical development. Recent conceptualizations of ASD suggest the condition is only fuzzily separated from non-ASD, with intermediate cases between the two. These intermediate cases may shift along a transition region over time, leading to apparent instability of diagnosis.

Methods:

We used a cohort of children with high ASD risk, by virtue of having an older sibling with ASD, assessed at 24 months (N=212) and 36 months (N=191). We applied machine learning to empirically characterize the classification boundary between ASD and non-ASD, using variables quantifying developmental and adaptive skills. We computed the distance of children to the classification boundary.

Results:

Children who switched diagnostic labels from 24 to 36 months, in both directions, (dynamic group) had intermediate phenotypic profiles. They were closer to the classification boundary compared to children who had stable diagnoses, both at 24 months (Cohen’s d = 0.52) and at 36 months (d = 0.75). The magnitude of change in distance between the two time points was similar for the dynamic and stable groups (Cohen’s d = 0.06), and diagnostic shifts were not associated with a large change. At the individual level, a few children in the dynamic group showed substantial change.

Conclusions:

Our results suggested that a diagnostic shift was largely due to a slight movement within a transition region between ASD and non-ASD. This fact highlights the need for more vigilant surveillance and intervention strategies. Young children with intermediate phenotypes may have an increased susceptibility to gain or lose their diagnosis at later ages, calling attention to the inherently dynamic nature of early ASD diagnoses.

INTRODUCTION

Children with ASD have heterogeneous developmental trajectories [1–4]. While most individuals retain their diagnostic categorization assessed at early ages [5–7], others have more dynamic presentations and may show substantial change over development [2, 6, 8]. For some, this process of change may eventually lead to losing or gaining diagnoses [5–7, 9].

The dynamic nature of ASD diagnosis provides valuable insights into the separation between ASD and non-ASD, introducing the notion of intermediate cases [6, 10, 11]. Intermediate cases correspond to children who are neither clearly affected nor clearly unaffected. These children may shift within a transition region between ASD and non-ASD over time, rendering a fuzzy boundary [12] between the two putative categories.

Elaborating the nature of boundaries that separate ASD from typical development and other psychiatric conditions may have significant clinical consequences. If the possibility of shifts within a transition region between ASD and non-ASD is indeed a core feature of the condition rather than being an artifact of misclassification, ASD may be better conceptualized as a dynamic condition. An ASD or a non-ASD diagnosis at a given time should then be seen as a “current” state of a child within the phenotypic and developmental continuum, instead of as a final label. The inherently dynamic nature of diagnosis highlights the importance of early treatment practices and close developmental surveillance [10, 11], with repeated screening that goes beyond the first few years of life [6, 9].

Although the idea of intermediate cases and possible diagnostic shifts has been described previously [5, 6, 9, 10, 13], there has not yet been a direct quantitative evaluation of a transition region between ASD and non-ASD or its relationship to diagnostic shifts. In this work, we introduce a data-driven and mathematically principled approach to study the links between early diagnostic shifts in ASD and the fuzzy nature of the diagnostic boundary. We define a measure that quantifies the distance of a child to the classification boundary using machine learning classification. We hypothesize that as a child gets closer to the diagnostic boundary at a given age, the likelihood of changing diagnosis at later ages increases. We study a sample of young children at high-risk by virtue of having an older sibling with ASD, assessed for ASD at 24 and again at 36 months, where some children gained or lost ASD diagnosis over this period.

METHODS

Participants

Phenotypic data were collected using the LORIS platform [14] as a part of the Infant Brain Imaging Study (IBIS), a longitudinal, multisite imaging study of children at high familial risk for ASD by virtue of having an older sibling with ASD [15]. Details on recruitment and exclusion criteria have been reported elsewhere [16]. In this study, we only included children with an older sibling with ASD. Informed consent approved by each site’s Institutional Review Board was obtained for all families.

Measures

The Autism Diagnostic Observation Schedule, Second Edition (ADOS-2) [17] is a diagnostic instrument for directly measuring ASD symptoms. We used the calibrated severity score (CSS) [18] as an overall measure of symptom severity, ranging between 1 to 10.

Autism Diagnostic Interview, Revised (ADI-R) [19] is a standardized parent interview for the assessment of ASD. We used three summary scores measuring domains of reciprocal social interactions (A Total), language and communication (B Total), and repetitive behaviors and interests (C Total).

Mullen Scales of Early Learning (MSEL) [20] is an examiner administered standardized developmental instrument. We used five t scores (Mean 50, SD 10) for gross motor, fine motor, visual reception, expressive language, and receptive language.

Vineland Adaptive Behavior Scales, Second Edition (VABS) [21] is a parent-report instrument assessing adaptive behavior for children. We used four composite standardized scores (Mean 100, SD 15) for communication, daily living, socialization, and motor skills.

Diagnostic Confidence was rated by clinicians during the diagnostic assessment. The clinicians coded their confidence levels as “Not Confident”, “Somewhat Confident”, and “Very Confident”.

Diagnostic Classification

A clinical best estimate procedure [16] determined whether criteria for ASD were met using the DSM-IV-TR [22] at 24 and 36 months. DSM-IV-TR (rather than DSM-5) was used for the whole sample to standardize the diagnostic framework, as data collection had started using DSM-IV-TR. We defined four diagnostic groups based on the categorical diagnosis at the two time points: the stable non-ASD group (NN), the stable ASD group (AA), the first dynamic group (NA; from non-ASD to ASD), and the second dynamic group (AN; from ASD to non-ASD). The stable groups retained their diagnosis between the two time points. Dynamic groups either gained (NA) or lost (AN) a diagnosis at 36 months.

In order to study the relationship between dynamic groups and atypical development consistent with the broader autism phenotype (BAP), we defined another diagnostic classification as ATYP [10, 23], at both time points independently. At a given time point, the ATYP classification did not have an ASD diagnosis and met at least one of the following conditions: ADOS-2 CSS ≥ 3, two or more MSEL t scores ≤ 35, or one or more MSEL t scores ≤ 30.

Sample

The initial sample included 222 participants with a diagnosis at both time points. We excluded participants with missing variables, as explained in Supplementary Section S1. The final dataset (see Table 1) included 212 participants at 24 months and 191 participants at 36 months. Demographics of participants are given in Supplementary Table S1.

Table 1.

Participant characteristic in four groups defined by diagnostic (Dx) categories at 24 and 36 months. Test statistics for Kruskal-Wallis H test, comparing two stable groups (NN, AA) and the combined dynamic group (NA&AN), are provided in the bottom row.

	Groups			Sex	MSEL Mean (SD)					VABS Mean (SD)				ADI-R Mean (SD)			ADOS-2 Mean (SD)
	Label	Size	Dx	M:F Ratio	Fine Motor	Gross Motor	Exp. Lang.	Recep. Lang.	Visual Recep.	Comm.	Daily Living	Motor Skills	Soc.	A total	B total	C total	CSS
24 Months	NN	140	N	1.4	49.7 (9.5)	48.6 (8.7)	47.6 (11.8)	51.9 (10.7)	53.8 (10.8)	101.3 (8.6)	102.2 (8.8)	100.2 (10.2)	100.2 (9.0)	3.6 (3.1)	3.5 (2.5)	0.6 (1.1)	1.7 (1.1)
	NA	11	N	2.7	45.3 (5.1)	50.4 (5.5)	45.6 (6.1)	45.5 (11.3)	51.2 (7.9)	99.8 (8.8)	98.5 (9.0)	102.7 (11.8)	101.8 (11.9)	5.0 (4.3)	1.7 (2.0)	0.7 (1.1)	1.8 (1.5)
	AN	16	A	2.2	40.0 (11.1)	45.1 (10.0)	43.5 (11.8)	43.4 (15.0)	43.7 (10.0)	95.1 (8.4)	95.7 (6.5)	94.1 (6.9)	90.9 (6.0)	6.8 (3.8)	6.7 (4.3)	1.1 (1.1)	5.2 (2.0)
	AA	45	A	4.0	41.3 (9.4)	40.4 (9.7)	35.6 (11.6)	32.0 (14.2)	42.3 (10.3)	85.0 (13.1)	90.3 (9.8)	94.1 (9.2)	86.2 (10.1)	10.0 (5.8)	6.7 (4.0)	2.0 (1.6)	6.1 (1.8)
	Test Statistics for NN vs. (NA&AN) vs. AA H-stats (p-value)				31.57 (1×10−7)	21.00 (3×10−5)	30.31 (3×10−7)	55.47 (9×10−13)	35.33 (2×10−8)	55.55 (9×10−13)	47.71 (4×10−11)	13.02 (1×10−3)	57.64 (3×10−13)	54.36 (2×10−12)	22.88 (1×10−5)	40.37 (2×10−9)	107.65 (4×10−24)
	Groups			Sex	MSEL Mean (SD)					VABS Mean (SD)				ADI-R Mean (SD)			ADOS-2 Mean (SD)
	Label	Size	Dx	M:F Ratio	Fine Motor	Gross Motor	Exp. Lang.	Recep. Lang.	Visual Recep.	Comm.	Daily Living	Motor Skills	Soc.	A total	B total	C total	CSS
36 Months	NN	127	N	1.4	49.9 (13.1)	--	50.8 (9.4)	50.2 (9.9)	59.3 (13.4)	100.5 (9.7)	97.7 (9.5)	95.1 (9.6)	99.9 (9.9)	2.7 (2.8)	2.8 (2.5)	0.7 (1.6)	1.5 (0.9)
	NA	8	A	3.0	33.1 (10.1)	--	43.7 (6.9)	39.1 (8.2)	51.8 (13.7)	93.0 (6.3)	86.5 (5.3)	86.0 (7.4)	91.9 (8.6)	4.1 (2.6)	3.8 (4.8)	1.2 (1.5)	5.6 (2.3)
	AN	17	N	1.8	42.3 (14.7)	--	46.1 (7.1)	46.9 (8.9)	49.9 (12.4)	94.9 (9.2)	91.4 (10.8)	90.8 (8.8)	90.0 (9.8)	5.9 (4.6)	4.9 (3.6)	2.6 (2.6)	2.1 (1.6)
	AA	39	A	3.3	34.1 (12.7)	--	37.4 (11.1)	35.0 (10.6)	39.9 (16.0)	84.6 (15.3)	78.7 (10.7)	84.1 (9.9)	80.3 (11.5)	10.8 (6.0)	9.6 (4.7)	3.8 (2.5)	6.5 (2.6)
	Test Statistics for NN vs. (NA&AN) vs. AA H-stats (p-value)				40.11 (2×10−9)	--	43.49 (4×10−10)	48.19 (3×10−11)	41.54 (1×10−9)	40.59 (2×10−9)	66.54 (4×10−15)	34.16 (4×10−8)	67.51 (2×10−15)	62.86 (2×10−14)	58.46 (2×10−13)	60.12 (9×10−14)	91.42 (1×10−20)

N: non-ASD, A: ASD, M: Male, F: Female

Note that the comparisons for all variables, except for Motor Skills at 24 months, have significant results after Bonferroni multiple comparison correction for 25 tests.

Statistical Approaches

Individual Variables:

We used the Mann–Whitney U test and Kruskal-Wallis H test to compare two and three groups, respectively. These non-parametric tests were chosen as several variables had non-normal distributions. We reported the effect size using Cohen’s d and the common language effect size (CLEF) [24]. CLEF, ranging between 0 and 1, is computed by checking all possible pairwise combinations of two groups to arrive at the probability that a randomly selected score from one group will be higher than a randomly selected score from the other group. With statistical group comparisons, we combined the two dynamic groups (NA, AN) into a single dynamic group due to the small sample size.

Machine Learning:

We used machine learning to predict diagnostic categories (ASD, non-ASD). The details are provided in Supplementary Section S2. Specifically, we used Support Vector Machine (SVM) classification [25]. Given a set of training examples, each labeled as belonging to one of the two categories, SVM learns a classification boundary between ASD and non-ASD that can be used to predict the diagnostic label of a novel child. We calculated the distance of each child to this boundary. A distance of 0 means being just at the boundary. A negative distance places the child into the non-ASD category and a positive distance to the ASD category. Children with higher absolute distances belong to their categories with higher certainty.

We combined 24- and 36-month data into a single sample for the machine learning analysis. Very similar results were achieved when analyzing 24- and 36-month data separately (see Supplementary Section S3). We used eight MSEL and VABS variables as the input to SVM. The Gross Motor variable from MSEL was excluded as it was not assessed at 36 months. We did not include ADOS-2 and ADI-R because those instruments directly assess diagnostic features, thus their inclusion might cause circularity in predicting diagnostic labels. We provided results with these two clinical instruments in the supplementary materials.

Overall classification performance was reported using accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). For each participant, we computed (a) average classification accuracy, ranging between 0 and 1, and (b) absolute distance to the classification boundary, having values between 0 and ∞.

Longitudinal Analysis:

We compared three variables between the groups, namely distance to classification boundary, classification accuracy, and diagnostic confidence rating of clinicians. We used generalized mixed-effect modeling for longitudinal analysis. For a given outcome variable (distance, accuracy, or confidence), a generalized mixed-effect model captured its relationship with sex (male, female), time (24, 36), and diagnostic stability (stable, dynamic). Details are provided in Supplementary Section S4.

Clustering:

We used clustering techniques to test whether the dynamic groups constituted separate subcategories. We used multiple clustering methods (k-means, spectral clustering, and hierarchical clustering) and two complementary validity indices (Silhouette [26] and Davies-Bouldin [27]) for robust interpretation of the results.

RESULTS

Phenotypic Profiles of Diagnostic Groups

We observed a statistically significant mean score hierarchy between the four groups (Figure 1 and Table 1). The two dynamic groups (NA, AN) had intermediate profiles between the two stable groups (NN, AA). The dynamic groups switched their profiles between 24 and 36 months. The NA group showed increasing impairments over time, but less impairment than the AA group. The reverse was true for the AN group, showing improvement over time, but not reaching the level of the NN group.

Figure 1. Phenotypic profiles of the groups.

Developmental trajectories between 24 and 36 months are shown for all diagnostic groups. The solid lines show group means, and the dashed lines show individual trajectories. With MSEL and VABS, higher scores indicate better developmental and adaptive skills. With the clinical instruments (ADI-R and ADOS-2) this is reversed. The two dynamic groups (NA, AN), on average, had scores between the two stable groups (NN, AA). With individual trajectories, heterogeneity within all groups is evident.

Separation between ASD and non-ASD

The machine learning classifier achieved high classification performance (Accuracy = 84%, Sensitivity = 69%, Specificity = 90%, PPV = 71%, NPV = 89%). The contribution of Socialization score (VABS) to classification, as seen in Supplementary Figure S1, was the highest, consistent with its face validity as a core feature of ASD. Receptive Language (MSEL), more at 24 months, and Daily Living (VABS), more at 36 months, also had high contributions.

Distance to Classification Boundary, Classification Accuracy, and Diagnostic Confidence

Distance, accuracy, and confidence values are listed in Table 2. Figure 2 illustrates the distance of individuals to the classification boundary, as well as how much they moved between 24 and 36 months. At 24 months, between the two groups with a non-ASD diagnosis (NN, NA), the NA group was closer to the boundary compared to the NN group. Similarly, between the two groups with an ASD diagnosis (AA, AN), the AN group was closer to the boundary. Comparison between the combined stable group (NN & AA) and the combined dynamic group (NA & AN) yielded small-to-moderate effect sizes both for absolute distance (Cohen’s d = 0.52, CLEF = 0.65, Mann-Whitney U test p = 0.0099) and accuracy (Cohen’s d = 0.94, CLEF = 0.69, p = 0.0001). Similar results were observed at 36 months with distance (Cohen’s d = 0.75, CLEF = 0.73, p = 0.0004) and accuracy (Cohen’s d = 0.80, CLEF = 0.68, p = 0.0001). The NA group was closer to the boundary compared to the AA group, and the AN group was closer compared to the NN group.

Table 2.

The distance to classification boundary, classification accuracy, and diagnostic confidence for the groups. The distance value is unitless. Higher values indicate better separation by the machine learning classifier. The accuracy takes values between 0 and 1, indicating average classification accuracy by the classifier. The confidence is given as the percentage of “Very Confident” diagnosis within each group. Test statistics for Mann-Whitney U test, comparing the combined stable group (NN&AA) and the combined dynamic group (NA&AN) in terms of absolute distance and accuracy, are provided in the bottom row.

	24 Months			36 Months
Groups	Confidence (%)	Distance Mean (SD)	Accuracy Mean (SD)	Confidence (%)	Distance Mean (SD)	Accuracy Mean (SD)
NN	70.5	−0.31 (0.20)	0.93 (0.25)	88.9	−0.28 (0.22)	0.90 (0.28)
NA	80	−0.22 (0.21)	0.82 (0.39)	16.7	0.04 (0.16)	0.61 (0.47)
AN	15.4	−0.05 (0.20)	0.34 (0.43)	53.3	−0.08 (0.20)	0.65 (0.45)
AA	53.5	0.13 (0.25)	0.68 (0.46)	60.5	0.26 (0.29)	0.85 (0.33)
Test Statistics for (NN&AA) vs. (NA&AN) U-stats (p-value)		3266.00 (0.0099)	379.00 (0.0001)		2994.00 (0.0004)	2805.00 (0.0001)

All comparisons were significant after Bonferroni multiple comparison correction for 4 tests.

Figure 2. Distance to classification boundary.

The distance of individuals to the classification boundary (gray dashed line) is illustrated. Each point corresponds to an individual child. Box plots show group level stats. (a, b) The location of a child (value on the y-axis) indicates how far she/he is from the classification boundary. The classification boundary divides the space into two regions, namely the ASD region (top) and the non-ASD region (bottom). Anyone in the ASD region is classified as ASD by the machine learning model. Children with higher absolute distances from the boundary are classified with higher certainty. The dynamic groups, on average, are closer to the boundary compared to stable groups. (c) The difference between the distance at 36 months and the distance at 24 months is given. A positive difference indicates a movement towards the ASD region, and a negative difference towards the non-ASD region. On average, the NA group moved towards the ASD region, the AN group towards the non-ASD region. (d) Magnitude (i.e., absolute value) of differences. All groups, on average, moved similar amounts between the two time points.

Between the two time points, the NA group moved towards the ASD region of the classifier (i.e., having more ASD-like phenotypes at 36 months, see Figure 2c), and the AN group moved towards the non-ASD region. Notably, the magnitude of these movements, on average, was not significantly larger than the movements in the two stables groups (Cohen’s d = 0.06, CLEF = 0.55, p = 0.5065), suggesting that diagnostic shifts were not associated with substantial phenotypic changes. Nevertheless, there were a few children in the dynamic groups who covered long distances between the two time points. The NA group moved slightly more than the AN group, although this was not significant (Cohen’s d = 0.73, CLEF = 0.63, p = 0.3123).

Clinicians rated all assessments either as Somewhat Confident or Very Confident. The distribution of confidence levels across the groups followed very similar patterns with those of distance and accuracy (see Table 2 and Supplementary Figure S2). Overall, an ASD diagnosis was associated with lower confidence, lower distance, and lower accuracy compared to a non-ASD diagnosis (AA < NN at both time points, AN < NA at 24m, and NA < AN at 36m).

The results of longitudinal models are given in Supplementary Table S2. The effect of time on the distance was not significant. Sex and diagnostic stability had significant effects on the distance (higher distance with females and stable groups), but the effect of sex was not significant after multiple comparison correction. Accuracy and confidence, on average, increased with time and stability. The effect of sex was significant only for confidence (higher in females) but did not survive multiple comparison correction.

Dynamic Group as a Separate Subcategory

We tested whether the dynamic groups (NA, AN) constituted a separate subcategory, rather than being within a transition region that is shared by ASD and non-ASD. All clustering methods and validity indices identified, at most, two clusters in the data (Supplementary Figure S3). The dynamic groups were distributed across these two clusters (Supplementary Table S3) and did not constitute another cluster. The dynamic groups (NA, AN) were assigned to their clusters with lower membership probabilities compared to the stable groups at both time points. This difference was significant only at 36 months (Cohen’s d = 0.60, CLEF = 0.71, p = 0.0012).

Dynamic Group and Atypical Development

The numbers of ATYP cases in the diagnostic groups are listed in Supplementary Table S4. At 24 months, only 18% of the NA group was classified as ATYP. The ATYP rate was 30% in the NN group. At 36 months, the rate was 35% in the AN group and 19% in the NN group.

Results with Clinical Instruments

The results of the machine learning analysis using the two clinical instruments (ADOS-2, ADI-R) are given in Supplementary Section S5, Supplementary Table S5, and Supplementary Figures S4, S5. The clinical instruments depicted a sharper classification boundary. The dynamic groups, as compared to the stable groups, had similar distances to boundary and accuracy at 24 months, made larger movements between 24 and 36 months, but had lower distance and accuracy at 36 months. In other words, the dynamic groups had an intermediate phenotype only at 36 months.

DISCUSSION

In this study, we treated the uncertainty of diagnosis at age two as a valid feature of ASD, rather than as a sign of poor procedural reliability. We showed that diagnostic changes between 24 and 36 months, when investigated in a mathematically principled way, can provide valuable insights into the nature of separation between ASD and non-ASD.

Our results suggested that children who changed diagnosis were, on average, close to the classification boundary and had cognitive, adaptive, and clinical phenotypes intermediary to the two stable groups, spanning a transition region in the diagnostic space. A diagnostic change between 24 and 36 months was, on average, a result of a slight shift within the transition region rather than a jump over a chasm between the categories.

The results with clinical instruments (ADOS-2, ADI-R) depicted a sharper classification boundary and larger movements between the time points for the dynamic groups. These results are indeed expected since the clinical design of these instruments (i.e., designed to capture the presence of symptoms) amplifies differences between ASD and non-ASD, causes loss of individual differences, and eventually conflates the dynamic and stable groups (e.g., NA and NN at 24 months).

The presence of an intermediate phenotype was suggested by previous studies [5, 6, 9, 10, 13], and sometimes associated with atypical development (ATYP) consistent with BAP [10, 13, 28]. Notably, the dynamic groups included some children with ATYP classification, yet the majority were not classified as ATYP at either time point. The difference between the intermediate phenotype as defined in this study and ATYP may seem surprising at first. Nevertheless, ATYP classification was defined based on ADOS-2 and MSEL [10, 23], while our machine learning model was trained using MSEL and VABS. The thresholds on the two instruments were learned by the machine learning model based on the characteristics of the current sample, unlike predefined thresholds of the ATYP classification. Thus, the intermediacy of the dynamic groups as defined in this study refers to their relative placement in the phenotypic space with respect to both ASD and non-ASD and does not necessarily correspond to subclinical symptoms of ASD.

The dynamic groups (NA, AN) showed notable heterogeneity. For example, not all children were closer to the boundary compared to stable groups. Several participants in the NA group had greater distance to boundary and phenotypic profiles comparable to that of the NN group. They showed substantial change between 24 and 36 months. Substantial improvements in children with severe impairments were also reported before [2]. This observation suggests that not all cases of diagnostic shifts are due to having an intermediate phenotype. Developmental regression, delayed onset of symptoms after 24 months, or dramatic improvement due to interventions could be among other factors.

Several children in the NA and AN groups had similar distances to the classification boundary, yet were placed in the opposite categories by clinicians. Moreover, many children were close to the boundary yet had stable diagnoses. This could be due to a decoupling between ASD severity and developmental/adaptive skills captured by MSEL and VABS [29]. Future studies should explore the underlying factors to these varying outcomes, including different etiologies, varying timing of ASD onset [3, 4], or different treatment histories. Unfortunately, this study did not collect historical data such as important life events or treatment records. Analysis of such data, especially to identify positive effects of interventions for the AN group is a promising future direction.

Children with intermediate phenotypes, even if not diagnosed with ASD at 24 months, may have an increased likelihood of an ASD diagnosis one year later. Similarly, children who are diagnosed with ASD but close to the classification boundary may not meet the clinical criteria for ASD one year later. Future studies are still needed to determine whether this phenomenon extends across later childhood, adolescence, and adulthood. Nevertheless, our findings underline the need for more vigilant surveillance and faster initiation of intervention rather than waiting for a categorical diagnosis to begin treatment. Such prudent practices become especially important for children with high familial risk [10] to maximize access to intervention and improve the quality of life for those individuals and their families [30, 31]. Identifying predictive behavioral or neurobiological patterns of children who may change diagnosis could provide improvements to current clinical practices [11, 32]. This could also lead to more informed intervention strategies tailored based on each child’s particular profile of strengths and weaknesses [10, 33].

Limitations

The current work has several limitations. First, our sample size and reported moderate effect sizes highlight the need for replication studies. Second, the study sample was limited to children at high-familial risk for ASD and does not represent the entire population of children with ASD. Thus, our findings may not generalize to populations without a family history of ASD. Third, our primary variables (MSEL, VABS) assessed developmental and adaptive skills, and do not fully capture symptom patterns that are linked to core ASD features. Thus, our results characterize the nature of separation between ASD and non-ASD within a rather limited scope defined by these variables. Fourth, some variables were from parental reports, which may raise concerns, especially in a high-risk sample. Parents, who already raised a child with ASD, may be influenced by severity differences between the older and younger siblings or may be better informed than parents who are unfamiliar with ASD in the general population. Nevertheless, parents in this sample rated behaviors consistent with trained clinicians [34].

Availability of materials:

The source code of all computer programs used during this study are available from the corresponding author on reasonable request.