Structural Equation Modeling in Language Testing and Learning Research: A Review

Despite the recent increase of structural equation modeling (SEM) in language testing and learning research and Kunnan's (1998) Kunnan, A. J. 1998. An introduction to structural equation modelling for language assessment research. Language Testing, 15: 295–332. [Crossref] , [Google Scholar] call for the proper use of SEM to produce useful findings, there seem to be no reviews about how SEM is applied in these areas or about the extent to which the current application accords with appropriate practices. To narrow these gaps, we investigated the characteristics of the use of SEM in language testing and learning research. Electronic and manual searches of 20 journals revealed 50 articles containing a total of 360 models analyzed using SEM. We discovered that SEM was most often used to investigate learners' strategy use and trait/test structure. Maximum likelihood methods were most often used to estimate parameters of a model; model fit indices of chi-squares, comparative fit index, root mean square error of approximation, and Tucker–Lewis index were often reported, but standardized root mean square residual rarely was. Univariate and multivariate normality checks were infrequently reported, as was missing data treatment. Sample sizes, when judged according to Kline's (2005 Kline, R. B. 2005. Principles and practice of structural equation modeling, 2nd, New York, NY: Guilford. [Google Scholar]) and Raykov and Marcoulides's (2006 Kline, R. B. 2005. Principles and practice of structural equation modeling, 2nd, New York, NY: Guilford. [Google Scholar]) guidelines, were in most cases adequate, and LISREL was the most widely used program. Two recommendations are provided for the better practice of using and reporting SEM for language testing and learning research.