Estimating Formative Measurement Models in IS Research – Analysis of the Past and Recommendations for the Future

Abstract

While debates on the appropriateness of formative measurement within structural equation models continue, such models are frequently found in IS research. IS researchers faced with such a model must identify the best method to estimate the model parameters, and have at their disposal covariance-based structural equation modeling (CBSEM), Partial Least Squares path modeling (PLS), and regression with summed scales, among other techniques.While all these methods can estimate models with formatively-specified latent variables, IS researchers frequently cite the presence of formative measurement as the reason for choosing PLS for model estimation over alternatives. Intuitively, a composite-based method such as PLS would appear to have an advantage in this particular scenario. In fact, some PLS researchers argue that PLS should only be used for such models. However, there is a dearth of empirical studies showing whether such an advantage does indeed exist.In this research, we discuss the statistical problems posed by models that include formatively-specified latent variables, and present a large-scale simulation study to investigate the relative performance of different estimation methods when faced with formative measurement, using models from studies published in MIS Quarterly. Based on our simulation results, we present recommendations for IS researchers interested in the estimation of models that include formatively-specified latent variables.