Eligible participants for this multisite longitudinal study included those who were between 6 and 26 years at Visit 1, and who had a diagnosis of, or suspected IDD. During Visit 1, ID or borderline ID criteria were based on the DSM-5 [1], with adaptive behavior deficits measured by the VABS-3 [2] and IQ < 80 on the Stanford-Binet Intelligence Scales, 5th Edition (SB5). Three groups were recruited: FXS (full mutation, with genetic confirmation), DS (with genetic confirmation if possible), and those with other or idiopathic intellectual disability (OID; with genetic confirmation of negative fragile X mutation). The DS and FXS groups were chosen specifically for this project because of the active translational research programs for these conditions and the urgent need for performance-based outcome measures. The OID group was recruited as an ID comparison group for previously published work [7, 12]. Though we have analyzed NIHTB-CB performance at the group level in these previous investigations, we chose to combine analyses across all groups in the present study given general developmental improvements observed in the NIHTB-CB across all groups previously [12], and to limit the number of analytic models (with multiple reference groups).

A mental age equivalence of at least 3.0 years as measured by the SB5 was required for inclusion, in concordance with NIHTB-CB age limits. Participants were required to be stable with usual treatment for at least 4 weeks before each visit. Exclusion criteria consisted of uncorrectable or uncorrected vision impairment, significant motor impairment preventing touch screen or keypad responses, or history of head injury, brain infection, stroke, or other neurological problems such as uncontrolled daily seizures or excessive sedation from medication. Recruitment sources consisted of research registries, flyers at local clinics, announcements through parent support foundation websites, and mailings to families registered with state departments that provide services to individuals with IDD. A total of 318 participants with IDD were recruited for Visit 1, and of those recruited, 55 individuals were ineligible: 21 with IQ > 79 and 34 with mental age below 3.0 years, leaving a final sample of 263. Participant retention at visit 2 was 81.36% (n = 214, meanage = 17.90 years). Full protocol, details of the NIHTB-CB, and its performance at baseline in the present IDD samples has been reported previously [7, 12, 26].

The NIHTB-CB, VABS-3 interview and SB5 were completed at Visit 1. For some participants, assessments were conducted over two days. After completion of the SB5, participants completed the NIHTB-CB while their parent/caregiver completed the VABS-3 with a psychologist or trained personnel. The same procedure was conducted again approximately 2 years later at Visit 2.

The NIHTB-CB [29] (Version 2.0) is an iPad-based battery that assesses Fluid Cognition, Crystalized Cognition, and a Total Cognition Composite through 7 tests as described above in the Introduction. A published manual of standardized NIHTB-CB administration procedures for IDD is available [30]. Unadjusted standard scores (USSs; non age-adjusted) were used for all NIHTB-CB tests. USSs have a mean of 100 and SD of 15. The USSs are recommended for longitudinal measurement because, like change sensitive or growth scale scores, they scale performance across all individuals on the measure based on the difficulty of items they received and their performance relative to everyone else in the original norming sample, allowing scores to be compared across time [31].

The VABS-3 [2] interview form was used to measure adaptive behavior (AB) domains including Communication (consisting of Expressive Language, Receptive Language, and Written Language), Daily Living Skills (DLS; consisting of Personal, and Community skills), and Socialization (consisting of Interpersonal Relationships, Play and Leisure, and Coping Skills). VABS-3 growth scale values (GSVs) were used for all analyses as they have been shown to be sensitive in individuals with IDD, particularly given their robust performance longitudinally and being less susceptible to floor effects [32, 33].

The SB5, which is standardized for individuals between 2–85 years, provides an overall index of intellectual ability reported as the Full-Scale IQ (FSIQ). In part, due to its broad developmental range, the SB5 has performed well in our prior studies of IDD [7, 26, 34]. Our protocol utilizes mental (rather than chronologic) age to select some NIHTB-CB test versions (e.g., PSM) and VABS-3 start points, which were derived from the SB5 FSIQ [30] for each participant. Deviation IQ scores were used to avoid inaccurate assessments of intelligence that can occur with standard score flooring in persons with IDD [34].

Our previous work has identified some limitations using the NIHTB-CB in individuals with IDDs, particularly for those with lower mental ages (i.e., < 5 years) [7]. A specific limitation relates to composite scores in the NIHTB-CB. For instance, the Fluid Cognition Composite is comprised of five subtests, all of which are required to generate a composite score. Many individuals in our sample are unable to pass practice and thus complete all five subtests. Therefore, they do not have composite scores available, thus limiting our power to test hypotheses at the construct level. Structural equation modeling (SEM) provides an analytic framework to help combat this issue, given that this modeling approach is robust to missing data, aiding our ability to retain the full sample in our models. In the present study, bivariate latent change score (BLCS) models provided two-year estimates of both cognitive and adaptive behavior change in individuals with FXS, DS, and OID. This modeling framework allowed us to examine the association between latent change for cognition and adaptive behavior across construct levels of each assessment.

In order to examine the association of developmental change between cognitive and adaptive behavior domains, permutations of BLCS models were used to compare change in the three cognitive domains measured by the NIHTB-CB (Fluid, Crystallized, Total Cognition Composites) and the three AB domains measured by the VABS-3 (Communication, DLS, and Socialization), resulting in the evaluation of nine models plus one full model (including all cognitive domains and all AB domains). Latent change score models are a type of structural equation modeling that provide estimates of change as latent variables based on two or more time points. In the BLCS framework, each model can assess the association between the latent change estimated for two constructs of interest [35]. We have utilized latent change scores previously to characterize developmental change in this sample across individual NIHTB-CB subtests [12]. Missing data were handled with full information maximum likelihood estimation, which is a standard recommendation to provide accurate parameter estimates in the presence of missing data [36].

Generally, each model contained latent scores for cognition and AB at Visit 1 and Visit 2 that were derived from the observed scores from each domain’s respective subtests [35, 37,38,39,40]. Latent change scores for cognition (ΔCOG) and adaptive behavior (ΔAB) were included to model change from Visit 1 to Visit 2. Furthermore, we included an estimate of the correlated change between ΔAB and ΔCOG in each model to assess cross-domain coupling of cognition and adaptive behavior. Time between visits and participant age were each used as covariates at the latent level in all models to control for any differences in cognitive and AB change due to variations in timing between Visit 1 and Visit 2, as well as age-related changes – modeled as those between 6 and 16 years at Visit 1, and those 16 years or older at Visit 1, which we have previously demonstrated in this population [12]. Analyses included all participants with a valid NIHTB-CB score, even without completion of visit 2 as BLCS models are robust to missingness [35]. Supplementary Figure 1 graphically presents a generic representation of these models, and Table 3 provides details of each model’s specification. For model fit we first specified base models in which nothing was correlated, and each variable received an equated intercept and variance across time [41]. We then assessed each model’s fit by comparing to its base model (utilizing robust fit parameters including CFI, TLI, and RMSEA) using methods from Savalei et al. [42], and indices of fit based on Little [43] [i.e., CFI > 0.85; TLI > 0.85; RMSEA < 0.08]. In order to control for false discovery rates, Benjamini–Hochberg procedures [44] were conducted for every class of analysis for the models with adequate fit.