Individual items, or in the case of the current study, individual questions from the RBS-R, were independently assessed for conceptual fit on the factor they most strongly loaded on to determine if it is an appropriate factor fit considering the other items that loaded strongly on the respective factor. The overall goal of EFA was to identify factors, based on a given dataset, and maximize the amount of variance explained by the model . Once a model has been theoretically and/or statistically established and hypotheses have been made, a confirmatory factor analysis can inform the likelihood of the hypothesized results.A Confirmatory Factor Analysis was conducted once the relationships among variables were established through statistical analyses and a theoretical model was evaluated . While the EFA allows for all items to load on any factor, the CFA restricts the factors on which items load. Each item was permitted to load on only one factor. Model fit was determined using recommended indices of model fit including Chi-Squared test, RMSEA, RMR, CFI and TLI. Additionally, the CFA model produced a weighted root mean square residual that is an empirically supported measure of model fit comparable to the other fit indices and is suggested to be highly useful for data that isn’t normally distributed . A WRMR value above 1.0 is considered good model fit. Factor loadings from the CFA were reported as the standardized model estimate loadings and associated standard errors.Cluster analysis provides a unique approach to examining which results in the identification of patterns that organize variables into taxonomies, grouping cases with similar patterns together . For the current study, blueberry grow pot the K-means cluster analysis was run to systematically and conceptually group participants with similar RRB patterns together.
The newly established factors from the CFA of the RBS-R were used to examine the various patterns of RRB presentation for this population. The goal of a k-means clustering is to partition individuals into clusters where every participant belongs to a cluster with others presenting with similar patterns . The optimal number of clusters must strike a balance between successfully compressing the data as a single cluster would, while maintaining maximum accuracy where every participant is assigned to its own cluster. The optimal number of clusters for the data was determined using both theoretical and empirical considerations. Previous research exploring RRBs have defined between two and six distinct types of RRBs; yet, there hasn’t been a clustering of those RRBs into distinct profiles to serve as a comparison or as an empirical rationale to test the fit of a specific number of clusters. Therefore, comparisons of three, four, five and six cluster solutions were conducted. One approach that was used to determine model fit for each cluster was to examine the number of iterations it took to satisfy the convergence criterion . There is no guarantee that data will cluster and iterate to convergence quickly, if at all. Therefore this is a reasonable justification for this approach in determining the fit between the number of clusters and the data being analyzed. Statisticians have concluded that it is acceptable to institute a maximum criterion of between 15 and 20 iterations for the data to reach convergence criterion where the clusters optimally fit the data. Cluster statistics were explored after running three, four, five and six cluster solutions; results are described below.The final research aim was to determine the ability of several behavioral and developmental characteristics to predict cluster membership. Correlation analyses among all predictors were conducted prior to running the MLR to determine presence of collinearity.
A multinomial logistic regression was run with individual cluster assignment as the outcome variable and participants’ standardized scores of ASD severity, nonverbal IQ, hyperactivity, anxiety, and coping skills as predictors. Age differences across clusters was independently examined by running a one-way ANOVA prior to running the MLR to determine if age significantly differed among clusters. The MLR provides a unique approach to determine the odds ratio of an individual being in one cluster relative to the odds of them being in the comparison cluster based on several characteristics . Therefore, it is important to choose a comparison cluster that will provide the most robust information in the analysis of these comparison solutions. Prior to exploring the individual cluster phenotypes to decide on a comparison cluster, the options were carefully considered and a conceptual decision was made. The comparison group should be the one that differs the most from the others, or the group that could be considered the “optimal outcome” group that possesses characteristics that researchers would want to test and discover what makes that group of participants different . Therefore, the cluster with the lowest levels across all RRBs was used as the baseline comparison cluster. Goodness of fit of the MLR model was assessed using the log-likelihood , which sums the probabilities of predicted outcomes and actual outcomes, analogous to the residual sum of squares in typical multiple regression. That is, the LL variable indicates how much unexplained data remains after the model is fit; where large values of the LL statistic tends to describe a poor fit for the model . Results of the multinomial logistic regression produced significance statistical values, which indicated the extent to which individual characteristics were able to significantly predict membership to one cluster over another. The individual parameter estimates for each comparison between the optimal profile group vs. the other profiles were individually examined to determine the significant and non significant results across predictor variables and interactions. The significance values were used to determine which of the characteristics were significant in predicting profile membership, with the odds ratio statistic indicating the odds of a participant being in a cluster when compared to the odds of them being a member of the optimal outcome profile group. Overall model fit statistics as well as individual parameter estimates of the multinomial logistic regression were examined.The recent changes to the DSM have created a more comprehensive list of RRB subtypes than were previously included and set a more stringent benchmark to meet criteria in the RRB domain. Such changes reflect the progression of research supporting the importance and independence of RRBs as an integral component of diagnosis, rather than a by-product of the “core” social communication impairments . From its earliest conception, ASD has been characterized by the presence of frequently and highly repetitious behaviors, with a marked desire for environmental sameness and consistency . Yet, this complex behavioral domain is historically under-represented in research efforts and falls secondary to social communication deficits in ASD research. Reviews of past studies on RRB presentation have highlighted issues including a lack of methodological consistency, with varying approaches to defining, organizing, and measuring RRBs. These discrepancies have led to splintered advancements in understanding the etiology, early behavioral manifestations and longitudinal developmental implications of RRBs . The primary aims of this study were to characterize RRB phenotypes of individuals with ASD and to determine the influence of developmental and behavioral characteristics on RRB profiles. This study revealed that there were five distinct RRB subtypes captured by the RBS-R, with five distinct phenotypic profiles generated from those subtypes. Hyperactivity, hydroponic bucket anxiety and coping skills significantly predicted participants’ RRB phenotype, while IQ and symptom severity had little effect.
The findings in this study provide a unique perspective when conceptualizing ASD symptomology and the influence of non-ASD specific traits on this core domain.The five-factor model result from the factor analyses of the RBS-R exhibits substantial consistency with previous studies examining the factor structure of the RBS-R . Comparisons between factor results of the RBS-R can be seen in Figure 3. Most notably, the current study excluded 3 items from the original compulsive scale as the item factor loadings were above .4 on more than two newly calculated factors. These results indicated that there wasn’t a single factor that accounted for the variability of each item, forcing those items to be excluded. Similarly, five items on the original ritualistic scale were excluded which included items regarding eating, sleeping, travel, play and self-care as they were highly loaded on multiple factors. These findings indicate they may not be sufficiently differentiating types of RRBs measured by each question, which leads investigators to wonder if the questions are adequately differentiating between RRB subtypes. Bishop, et al. investigated RRB data from both the ADI-R and the RBS-R and found that the ADI-R items resulted in a two-factor model, whereas the RBS-R resulted in a five-factor model as the best fit. When examined in conjunction with findings from the Lam & Aman study as well as the current results, it is evident that using a measure with a wider range of questions such as the RBS-R provides more in-depth and informative results when examining the specific types and severities of RRBs. Despite the utility of a measure dedicated to specific RRB types and severity, factor results from previous studies fail to be substantiated with each study, leading to the conclusion that a final set of RRB subtypes have yet to be established unequivocally across studies. Further, each analytic result has not been entirely consistent with the six conceptually derivedsub-scales that Bodfish, et al. originally established. Discrepancies between the subtypes and the original sub-scales, as well as between the previously proposed models can be seen below in Figure 3. As previously discussed, RRBs comprise a complex and heterogeneous set of behaviors that vary greatly depending on the population being measured; therefore, it is not a complete surprise that each factor analytic study has resulted in slightly altered structures. However, given the vast age range include in the current study and largest number of participants to date for an RBS-R factor analysis, the resulting factor structure warrants consideration as an organizational RRB factor structure to be analyzed for confirmatory analyses in future studies using the RBS-R. As seen in Figure 3, researchers who organized and defined more than two categories of RRBs had one striking consistency, the inclusion of an independent category of self-injurious behavior . Further, self- injury is arguably the most recognizable and disruptive RRB consistently found to be related to greater impairment with significantly lower IQ and higher severity of ASD symptoms . In fact, the most recent study examining RRB subtypes concluded that SI behaviors create significant difficulty in dichotomizing RRBs, as the SI items fail to load with the repetitive sensory motor category or with the insistence on sameness supporting the existence of additional subcategories . Further, SI is the only RRB subtype to consistently load identically as an entire sub-scale in every factor analytic study of the RBS-R, which was also true in the current study, indicating its distinctiveness .The RRB subtype found most consistently across studies has been self-injurious behaviors. As seen in Figure 3, researchers who organized and defined more than two categories of RRBs had one striking consistency, the inclusion of an independent category of self-injurious behavior . Further, self- injury is arguably the most recognizable and disruptive RRB consistently found to be related to greater impairment with significantly lower IQ and higher severity of ASD symptoms . In fact, the most recent study examining RRB subtypes concluded that SI behaviors create significant difficulty in dichotomizing RRBs, as the SI items fail to load with the repetitive sensory motor category or with the insistence on sameness supporting the existence of additional subcategories . Further, SI is the only RRB subtype to consistently load identically as an entiresub-scale in every factor analytic study of the RBS-R, which was also true in the current study, indicating its distinctiveness .Cluster analysis provides a novel approach to statistically explore phenotypic profiles and the co-occurrence of RRB types and severity across individuals with ASD. This is the first study of its kind to statistically generate clustered phenotypes, each consisting of multiple RRB subtypes. RRBs don’t occur in isolation; the pattern of behavior is fluid with minimal evidence to explain the variations seen across and within individuals. By studying RRBs in a way that allows for multiple RRBs to co-occur at varying levels, researchers may gain a more accurate and informative picture of how these behaviors manifest across individuals with ASD. However, when researchers rely solely on parent report measures, there is a limited scope of distinct behaviors from which combination or cluster phenotypes can be derived.