Parallel analysis and MBI-HSS

Mr. Editor: It has been only recently possible to validate the Maslach Burnout Inventory-Human Services Survey (MBI-HSS) 1 among health professionals of Cali 2 , an important step for using this instrument with local empirical support in regard to its reliability of scoring and internal structure. However, two aspects of this analysis can be considered as methodological weaknesses. First, the Cronbach alpha coefficient was calculated for the total group of items, and this is absolutely inappropriate because: a)the authors did not demonstrate empirical support for accomplishing this (e.g., a hierarchical factor analysis), b) the literature indicates that factors in the MBI-HSS are generally independent, a characteristic also reported by Córdoba et al. 2 , and the same authors of the MBI-HSS 1) the authors did not report the inter-factor correlations with which an appreciation could have been obtained, at least an heuristic one of the common degree of variance among the factors. Secondly, the authors obtained seven factors in their exploratory factor analysis; this large number of actors seems to be a product of applying a factor retention method that is now consensually seen as inaccurate and little recommended 3,4. Specifically, it is known as Kaiser' s rule, Guttman´s rule or simply K1 4. The problem identified with this method is its over-estimation of the number of factors to be retained 3,4 , a situation that clearly occurs in the results reported by Cordoba et al. 2 , as reported in their Table 2. A more accurate method which has gained a scientific consensus for good practices for retaining the number of factors is called parallel analysis 3,4. This procedure is based on the work of Horn 4 , which consists of randomly creating the same number of variables as the number of items analyzed (in the case of MBI-HSS, 22 items),correlating them and extracting eigenvalues against which the eigenvalues derived from the empirical data under analysis are compared. This procedure was applied to the eigenvalues reported by Cordoba et al. 2 , in Table 2 by means of the Monte Carlo PCA 5 program (100 replications). Our results are shown in Table 1. The appropriate number of factors to retain is achieved by making a one to one comparisons of the eigenvalues, keeping the empirical eigenvalue that is less than the random eigenvalue. Conceptually, this indicates that the significant eigenvalues must be greater than those generated randomly. In Table …