Journal of Rehabilitation Medicine 51-8 | Page 34

L. Turner-Stokes et al. full dataset, and the Rasch sub-sample of n = 320 were similar on all parameters. The results of EFA are summarized in Table II. All items loaded strongly onto a single first component (all loadings > 0.58) with Cronbach’s alpha 0.976, accounting for 60% of the variance. Four factors had eigenvalues > 1.0, but the fourth contained only one item (”Community mobility”), which is rarely assessed and so rated 1 (”not assessed”) by scoring convention. The remaining 3 factors (Motor (16 items), Psychosocial (9 items) and Communication (5 items)) together accoun- ted for 73% of the variance. As suggested by a previous factor analysis in TBI patients, a 2-factor model was also explored, which accounted for 70% of the variance. CFA for a 3-factor model showed a marginal fit as shown in Figure 2. This model included correlated of error terms for nine pairs of items with the largest modification indices. The fit indices did not meet the criteria for an excellent fit, but were all in the border- line acceptable range (RMSEA 0.094, CFI = 0.926, TLI = 0.918, NFI = 0.922) While the χ 2 /df ratio was large (18.103), giving a highly significant p-value (0.000), this is more likely to be a reflection of the large sample size than the model fit (30). The fit was only marginally improved by further correlation of error terms and thus we proceeded to Rasch analysis. Rasch analysis The suitability of the Partial Credit Model for Rasch analysis was confirmed by the significant likelihood- ratio test (χ 2 (df86) = 7,325.0, p < 0.001) Table III includes fit statistics for individual items together with response frequencies for each category within the main dataset (n = 1,956) and for the domain scores presented for both the full dataset and the sub-sample (n = 320). A preponderance of scores at the ends of the range is expected for a dataset that includes both admission and discharge data, and indeed there was a greater preponderance of response category scores at the lowest end of the scale for the admission sample, and upper end in the discharge sample. Table IV summarizes the overall fit statistics from the main stages of the Rasch analysis. First analytical pathway (all 30 items). The initial analysis of the full 30-item scale is marked by satisfactory reliability (PSI = 0.94). However, the overall model fit was poor with significant item- trait interaction. At the individual item level, 14 out of 30 items showed significant misfit to the Rasch model on the n = 320 sub-sample (Table III). 570 Fig. 2. Summary of findings from confirmatory factor analysis. CFA confirmed a 3-factor solution. Correlations between the factors ranged from 0.74 to 0.86. www.medicaljournals.se/jrm Second analytical pathway (3 super-items). The second analytical pathway was conducted using 3 super-items created by combining the items within the Motor, Communication and Psychosocial do- mains identified from the factor analyses. Prior to combining domain items into super-items residual correlation matrix was examined. Residual cor- relations exceeding the cut-off point of 0.20 above the mean of all residual correlations were found for 59 pairs of items in Motor, 5 pairs in Com- munication, and 26 pairs in Psychosocial domain. Rasch analysis of 3 super-items resulted in accep- table overall model fit (χ 2 (df24) = 36.72, p = 0.05), strict uni-dimensionality and no local dependency when tested on the sub-sample (n = 320) (Table IV, Pathway 2). The proportion of common error-free variance A=0.88 was marginally below 0.90 sug- gesting that use of transformation table based on this 3 super-item solution may be preferable to the ordinal scale for calculation of change scores. This analysis was replicated with the full dataset (n = 1,956) showing good reliability PSI of 0.81,