Factor analysis of aphasic syntactic comprehension disorders

This research was supported by grants from NIDCD (DC00942 to David Caplan and DC007564 to Gayle DeDe). We are grateful to Sue Kemper for comments on an earlier draft and suggestions regarding CFAs.

Abstract

Background: Many people with aphasia have impaired sentence comprehension. Previous studies using factor analysis have suggested that a single factor accounts for performance on measures of sentence comprehension (e.g., Caplan, Baker, & Dehaut, 1985). However, this work has been limited to measures of accuracy on single sentence comprehension tasks.

This research was supported by grants from NIDCD (DC00942 to David Caplan and DC007564 to Gayle DeDe). We are grateful to Sue Kemper for comments on an earlier draft and suggestions regarding CFAs.

Aims: The purpose of this study is to further examine the factor structure underlying aphasic sentence comprehension using accuracy, reaction time (RT), and on‐line measures using both exploratory and confirmatory factor analyses.

Methods & Procedures: A total of 42 people with aphasia and 40 non‐brain‐damaged controls were tested on 11 sentence types and five tasks. Accuracy and RT data are reported for the whole sentence presentation version of sentence–picture matching, and accuracy data are reported for object manipulation. Confirmatory factor analyses examining measurement invariance across groups and tasks are presented. Exploratory factor analyses of on‐line syntactic processing are also presented.

Outcomes & Results: Results indicated that one‐factor models best account for accuracy and RT data. Measurement of factors was partially invariant across groups and tasks. Factor structures suggestive of syntactic processes emerged in the analyses of on‐line measures.

Conclusions: This study suggests that syntactic processes may load on separate factors during on‐line parsing and that syntactic processes do not dissociate when the parser's output is used in the service of a task at the end of the sentence.

Notes

This research was supported by grants from NIDCD (DC00942 to David Caplan and DC007564 to Gayle DeDe). We are grateful to Sue Kemper for comments on an earlier draft and suggestions regarding CFAs.

1. Note that this control group differs from the one presented in Caplan et al. (2005).

2. The least squares method of EQS (Bentler, 1989) was used to estimate the standardised parameters of the model, which provide a measure of the strength between the observed, or manifest, variables and the underlying, or latent, factors. Standardised residuals between the predicted and observed covariance matrices were inspected to ensure that they are centred on zero, with few exceeding .20. Model fit was first assessed by chi‐square measures of the discrepancy between the observed and predicted models (Tabachnick & Fidell, 1996). Because of chi‐square sensitivity to sample size, the standardised root mean square residual (SRMR) of the difference between the predicted and observed variances and covariances, and the root mean square error of approximation (RMSEA), which includes a penalty for lack of parsimony were also calculated (Jaccard & Wan, 1996; Tabachnick & Fidell, 1996). Finally, the comparative fit index (CFI), which compares the observed model to one in which all covariances are set to zero and is a highly stable measure, especially for small sample sizes (Jaccard & Wan, 1996), and the non‐normed fit index (NNFI), a similar measure that is adjusted for the number of parameters in the model, were calculated. Non‐significant chi‐square values, SRMRs and RMSEAs less than .10, and CFIs and NNFIs above .90 are considered indications of good model fits. One‐ and two‐factor models were compared with chi‐square difference tests, in which the one‐factor model’s chi‐square and DF were subtracted from the same measures associated with the two‐factor model (Byrne, 1994). A significant result indicates that the two‐factor model accounts for more of the total variance in the dataset and should be retained, while a non‐significant result indicates that the models are statistically equivalent and favours the one‐factor model.