Language disorder typologies: Clustering and Principal Components Analysis in the EpiSLI database

Abstract

There is substantial heterogeneity within the population of children with language impairment (Bishop, 1998; Leonard, 2014). This heterogeneity has led the research community to consider the possibility of subtypes of language impairment (Bishop & Rosenbloom, 1987; Dollaghan, 2011; Rapin & Allen, 1983; Tomblin et al., 1997). But, previous research on language impairment subtypes is inconclusive (Dollaghan, 2011; Leonard, 2014). This study utilized the EpiSLI database (Tomblin, 2010) to empirically examine the possibility of subtypes by applying statistical analyses designed to detect and classify clusters. Further, this study addressed three key gaps from prior work by: (a) using continuous variables without a priori number of clusters, (b) including cognitive variables, and (c) evaluating subtype models. The EpiSLI database was spilt into two datasets based on clinical status: Typically Developing and Language Impaired. Language and cognitive measures within the datasets were used for clustering analyses. The study included two methods of cluster analysis: Ward’s and K-means. The results indicated that both the Typically Developing and the Language Impaired datasets did meet the a priori criteria for detecting structure, but this structure did not aggregate into interpretable clusters. Principal components analysis was then applied and reduced the variables to eight components, but did not improve results in terms of yielding interpretable clusters. These results support previous research that argues against subtyping in language impairment (Dollaghan, 2011; Leonard, 2009) while replicating a previous cluster analysis using the EpiSLI data (Tomblin & Zhang, 1999). Because consistent subtypes cannot be demonstrated empirically despite theoretical justification for subtyping patients with SLI, clinicians should focus on individual strength and weaknesses in assessment rather than attempting to subtype patients. Future work should seek to replicate these methods in other databases and with additional measures in order to further explore the possibility of language impairment subtypes.