Latent Classes by ADOS Severity Score Trajectory
A linear model of 5 latent trajectory classes had the most parsimonious fit to longitudinal ADOS CSS data, as suggested by the lowest Bayesian Information Criteria (BIC) in comparison with other models (see ). Greater numbers of dimensions or classes led to models with higher BIC.
Latent Severity Class Model Comparison
The linear fixed part coefficients of the 5-class model showed no evidence of a significant relationship between ADOS severity and chronological age (0.2 = 0.33, P = .8), suggesting no significant overall age trend masked by the grouping into latent classes.
One of the 5 classes in this best-fit model included only 6 participants (final diagnoses: n = 1 autism, n = 2 pervasive developmental disorder-not otherwise specified (PDD-NOS), n = 3 intellectual disability/language delay) with 22 assessments. These children had stable mild-severity scores from 1 to 3 over time, with 1 outlying score of 6. Because of the small size of this class, these participants were dropped from further analyses. The 4 remaining latent trajectory classes are shown in . Participant chronological age was restricted to a maximum of 10 years for graphical representation, because data for the 11 to 15 age span were more sparse.
ADOS severity score latent trajectory classes. y axis depicts ADOS CSS.
The 4 classes included a persistent high-severity class (class 1: persistent high; 46% of 339 remaining participants), a moderately severe class (class 2: persistent moderate; 38%), a class that tended to increase in ASD severity over time (class 3: worsening; 9%), and a class that decreased in ASD severity over time (class 4: improving; 7%). Trajectory class was not associated with the number of assessments per individual (F[3,335] = 1.3, P = .27). The average probability with which children were assigned to their best class was high for classes 1, 3, and 4 (P = .82, .79, and .81 respectively) but lower (P = .68) for class 2 (persistent moderate). The average probability that children assigned to this class might have belonged to class 3 (worsening) was not small (P = .21). The worsening class was marked by score variability: 70% of children assigned to this class had most recent severity scores higher than previous scores, whereas the remaining 30% showed wide variability across time, some of them “ending” on an improved score. In contrast, all children assigned to the improving group had most recent scores milder than previous scores.
describes initial and final diagnostic measures and demographic variables in the 4 latent classes. ADI-R domain totals are reported as sums of “current” scores of only those algorithm items comparable across age groups at both initial and final assessment, to compare stability or change over time by latent class. Trends in raw scores were observed to fall (ie, improve) slightly over time in Current Social-Communication on the ADI-R and Social Affect scores on the ADOS, and to rise (ie, worsen) slightly over time on ADOS Restricted Repetitive Behavior scores across the first 3 classes. The worsening class was the only group to exhibit greater severity over time in any ADI-R Current domain mean score (Verbal Communication and Restricted Repetitive Behavior).
Latent Severity Classes: Descriptives and Predictors
Covariates as Predictors of Latent Class Membership
As shown in , gender, race, and initial NVIQ did not significantly predict latent class membership. Higher initial VIQ significantly predicted membership in the improving, worsening, and moderate classes in comparison with the persistent high reference class. Relative risk ratios were generated from multinomial logistic regression; race and gender were entered as binary predictors (0 = white or male; 1 = other race or female), and VIQ and NVIQ scores were standardized before being entered into the model. Relative risk ratios indicate the multiple of odds for specific class membership (eg, improving) in a particular group (eg, females) in comparison with membership in the persistent high class (eg, 1 SD difference in VIQ increased the odds of being in the improving class, relative to the persistent high class, by 383%).
Diagnosis, Regression Status, and Treatment Variables by Latent Severity Class
Almost all children with a final diagnosis of autism were assigned to persistent high (60%) or moderate (36%). Participants with PDD-NOS most commonly were assigned to persistent moderate (45%), worsening, and improving classes (17.3% each). Three children in the worsening severity class ultimately received nonspectrum diagnoses (n = 1 language disorder, n = 1 disruptive behavior disorder, n = 1 intellectual disability), as did 4 children in the improving class (n = 1 Tourette’s syndrome, n = 1 mood disorder, n = 2 language disorders).
Language regression scores did not differ significantly across the 4 classes, F(3,439) = 2.3, P = .08. We found no significant class difference between participants with the highest levels of either MPST or ABA therapy hours in comparison with children who had received less or none of either type of intervention, χ2(4) = 4.3, P = .36 for MPST; χ2(4) = 3.5, P = .48 for ABA.
IQ and Adaptive Behavior Trajectories Within ADOS Latent Severity Classes
The pattern of change in sample mean VIQ and Vineland Daily Living V-scale (standard scores) in each trajectory group is shown in and .
VIQ trajectories by latent severity class.
Vineland Adaptive Behavior Scales “Daily Living” V scores by latent class.
All classes showed an increasing (improving) trend over time in VIQ but with marked between-class differences (generalized estimating equations Wald test over intercept, linear and quadratic terms; 0.2 = 219.60, P < .001). The improving class means exhibited a much steeper curve indicating progress that occurred earlier and was greater overall than in the other 3 classes. At baseline, the improving class had significantly higher VIQ than the persistent high and moderate classes (F(3,322)=18.21, P < .001), although it did not differ significantly from the worsening class (P = .642). Tests at age 6, when IQ appeared to stabilize, indicated the improving class had significantly higher mean VIQ than the remaining 3 classes (P < .001); persistent moderate and worsening were similar (P = .164), although both were above the persistent high class (P < .001).
For Vineland Daily Living Skills (including such skills as toileting, dressing, and chores), the classes showed similar and relatively unimpaired scores at age 2 but diverged thereafter (0.2 = 103.16, P < .001). Modest gains were made by the improving class, with marked declines noted in the other groups. By age 6, the improving class was significantly better than the other classes (at P = .006 or smaller), with no significant differences among those 3 (P = .243 or greater).