Estimation Methods for the Structural Equation Models: Maximum Likelihood, Partial Least Squares and Generalized Maximum Entropy

摘要

The concept of Latent Variables (LVs or latent constructs) is, probably, one of the most charming and discussed of the last fifty years, although, even today, it is only possible to give a negative definition of it: what is not observable, lacking both of origin and of measurement unit. One of the difficulties for a researcher in the economic-social sciences in the specification of a statistical model describing the casual-effect relationships between the variables derives from the fact that the variables which are object of the analysis are not directly observable (i.e. latent), for example, the performance, the customer satisfaction, the social status etc. Although such latent variables cannot be directly observable, the use of proper indicators (i.e. manifest variables, MVs) can make the measurement of such constructs easy. Thanks to the SEM, it is possible to analyze simultaneously, both the relations of dependence between the LVs (i.e, Structural Model), and the links between the LVs and their indicators, that is, between the corresponding observed variables (i.e, Measurement Model). The different and proper methodologies of estimate of the dependence are topics of this work. In particular, the aim of this work is to analyze Structural Equation Models (SEM) and, in particular, some of the different estimation methods mostly adopted: the Maximum Likelihood-ML, the Partial Least Squares- PLS and the Generalized Maximum Entropy - GME, by illustrating their main differences and similarities.