Auditory Evoked Response Data Reduction by PCA: Development of Variables Sensitive to Reading Disability

Abstract

Long latency auditory evoked responses (AER) were formed on 232 healthy normal and learning impaired subjects to tone pairs of 50 msec inter-stimulus interval (TALAER) and also to the words “tight” and “tyke” (TTAER). Both evoked potential (EP) types have been used to demonstrate differences between good readers (WIAT Basic Reading score > 115, N=42) and poor readers (Reading score < 85, N=42). A largely automated, hands off approach was used to reduce artifact contamination, to develop canonical measures for discriminating good from poor readers, and to predict reading scores across the entire population including intermediate (average) readers. Eye and muscle artifact were diminished by multiple regression. Substantial EP data reduction was enabled by an unrestricted use of Principal Components Analysis (PCA). For each EP type, 40 factors encompassed 70–80% of initial variance, a meaningful data reduction of about 90:1. Factor interpretation was enhanced by mapping of the factor loadings. By discriminant analysis, resulting factors predicted reading group membership with over 80% jackknifed and also split-half replication accuracy. By multiple regression, they produced a canonical variate correlating significantly (p<0.001) with the Basic Reading score (r=0.39). The TTAER factors were more useful than the TALAER factors. The relevance of rapid auditory processing and phonemic discrimination measurements to dyslexia is discussed.