The reduction of skewness and kurtosis of observed variables by data transformation: Effect on factor structure

The present paper examines the effect of skewness and kurtosis reduction through data transformation on the factor structure obtained by exploratory and confirmatory factor analyses. Data are from a 16-item scale that measures students' opinions about university teaching, each item measured on a 5-point Likert format. Observed distributions do not comply with the assumption of normality, so different variable transformations were performed to reduce the skewness and kurtosis of the data. Our results suggest that goodness-of-fit indices and factor loadings are more sensitive to the estimation method employed (for a given transformation) than to the transformation procedure (for a given estimation method). For moderate sample sizes and correctly specified models maximum likelihood estimation method is the one that performs the best, even when assumption of multivariate normality is violated, provided the Mardia coefficient falls in the range up to 70. Neither the complexity of factor structure nor the theoretical commonality among variables were taken into account, thus limiting the present results.