How is GSCA different from partial least squares (PLS)?

PLS, also called PLS path modeling or PLS-SEM, is a well-known method for component-based structural modeling. Main theoretical differences between PLS and GSCA can be described as follows.

  • PLS is limited-information, whereas GSCA is full-information

PLS is a limited-information method in that it estimates its two sub-models (measurement and structural) separately, each being linked to a separate (local) optimization criterion. This is mainly because PLS does not combine the sub-models into one, which facilitates the derivation of a single optimization criterion. Conversely, GSCA is a full-information method, which optimizes a single criterion to estimate all parameters simultaneously. Both limited- and full-information methods tend to provide unbiased parameter estimates, whereas a full-information method generally tends to provide more reliable estimates (i.e., smaller standard errors) than a limited-information method.

 

  • The concept of overall model fit is not well applicable to PLS

As a limited-information method, PLS divides the parameters into two sets and estimates each set sequentially. This characteristic complicates the assessment of model fit, as no unique criterion is optimized according to which the degree of optimization can be assessed. While several model fit indexes have been proposed in PLS, researchers have long complained that these indexes are based on parameters that are not explicitly optimized as part of the algorithm [21,22]. Thus, any statement about a model’s quality and decisions regarding potential model modifications based on existing model fit indexes are questionable when using PLS [21].

 

  • GSCA is more flexible than PLS

For example, PLS’s measurement model is typically assumed to be unidimensional, where each indicator is assigned to only one component [23]. GSCA allows for multidimensional measurement models, which link an indicator to multiple components (e.g., growth curve models or models with cross loadings). Also, PLS’s structural model is considered recursive in that the method does not allow for reciprocal component relationships, whereas GSCA’s structural model can be non-recursive accommodating such reciprocal relationships. Furthermore, GSCA can constrain parameters to be certain values or to be equal in a single group or across multiple groups and compare the adequacy of unconstrained and constrained models. PLS is not capable of such constrained analyses.

 

Despite their theoretical differences, PLS and GSCA will likely produce similar estimates of parameters in unidimensional, unidirectional and unconstrained component models [24,25], although GSCA still tends to provide smaller standard errors and confidence intervals [12].