I have many colleagues who advocate the use of SEM-PLS.
However, an editor from a top tier journal suggested that "modern ML SEM procedures can account for non-normal distributions and small sample size without the drawbacks of PLS".
I searched for some critical perspectives on PLS. I am attaching one of these contributions for your perusal. What are your views about PLS?
I look forward receiving your remarks and suggestions (on the use of PLS. Should I use other statistical programs - e.g. AMOS, LISREL, MPLUS, R ?).
Best Regards,
Mark
Take a look at this:
https://www.smartpls.com/documentation/learn-pls-sem-and-smartpls/choosing-between-pls-sem-and-cb-sem
I like this Editorial in one of our top journals very much:
https://misq.org/misq/downloads/download/editorial/556/
Dear Mark,
first I would like to distinguihs between models and estimators (I find the expressions SEM as well as PLS-SEM not very helpful as it covers nothing and everything):
Currently there are two kinds of constructs, the representations of the theoretical concept in SEM: (i) composite and (ii) latent variable.
A composite is fully composed by its observable variables, while a latent variable (common factor) is the cause of the covariance between its observable variables. these two different types of constructs can be linked via a structural model.
As estimators you have for example ML, GLS, but also PLS-PM, and GSCA.
While the ML consistently estimates common factor models, it is also asymptotcially efficient, i.e., if the sample size goes to infinty, there is no other consistent estimator with a smaller standard error. So it is the prefered estimator for latent variable/common factor models. However, it cannot estimate composite models (it depends on the position of the composite). On the other hand, PLS-PM (mode B) can consistently estimate composite models, like GSCA. So they are prefered to estimate pure composite models.
Over the last years, PLS and GSCA were enhanced, i.e., PLSc and GSCAm, to consistenly estimate structural models containing latent variables. However, they cannot be as assymptotically efficient as the ML estimator, but if your model contains both composites and latent variables, or composites only, they might be a good option.
Considering the use of composites in structural equation models, and thus the application of PLS-PM, you can find several strands in the literature: (i) they regard the composite as a good way to represent artifacts (design concepts) by composites, while behavioral concepts are well represented by a latent variable. Against the background of confirmatory research, they consider the model fit, i.e., how well fits the model the data. I would say this approach is most similar to the viewpoint of SEM (ii) They employ the composite to model all theoretical concepts. They argue that between the theoretical concept and its representation, i.e., the construct, is actually the validity gap. So nobody knows if the latent variable is a good representation of it, therefore, the composite can be an equally representation as well. (iii) Sometimes, people want to do out-of-sample prediction, Since the ML estimator and its relatives do not provide construct scores, PLS-PM is employed as it creates scores to estimate the model parameters. These can also be used for predictions. In general, in predictive modeling consistency of an estimator is not a big issue.
(iv) there's a group who combines the ideas (ii) and (iii), called causal predictive modeling. Typically they do not care about model fit in the sense of SEM, instead they consider the out-of-sample predictive power and different kind of construct validity for composites.
So now it depends on your type of research and on the viewpoints described above, which arguments for PLS-PM/PLS-SEM you choose.
I agree with the editor's opinion that the ML estimator can properly deal with non-normality and small sample size.
Regardsless of which type of estimator you want to use, R is a good option. First it is for free, second it contains packages that provide the ML estimator (e.g., lavaan ) and PLS-PM (e.g., cSEM or matrixpls). Of course, if you want to estimate more complex models, commerical softwares like MPlus might be advantageous.
HTH
Best regards,
Florian
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling