Please take look at here: Article Predictive validity and formative measurement in structural ...

You may also find this article usefulArticle Predictive model assessment in PLS-SEM: guidelines for using...

Dear Asma Malik ,

please also have a look at the following papers:

G. S. Hubona, F. Schuberth, and J. Henseler, “A clarification of confirmatory composite analysis (CCA),” International Journal of Information Management, vol. 61, p. 102399, 2021, doi: 10.1016/j.ijinfomgt.2021.102399.

F. Schuberth, “Confirmatory composite analysis using partial least squares: Setting the record straight,” Review of Managerial Science, vol. 15, pp. 1311–1345, 2021.

J. Henseler and F. Schuberth, “Using confirmatory composite analysis to assess emergent variables in business research,” Journal of Business Research, vol. 120, pp. 147–156, 2020, doi: 10.1016/j.jbusres.2020.07.026.

Best regards,

Florian

Whilst it is ok to say... Have a look at this paper and that... Necessary reading you might say BUT establishing formative vs reflective item measures at the outset is a conceptual issue.

For this... You provide little guidance to advise. Regards Brad.

Hey Bradley,

I agree that deciding about reflective or causal-formative measurement is a conceptual issue. However, the point that I wanted to make with my proposed literature was that confirmatory composite analysis does not contain the assessment of predictive validity. As explained by Schuberth (2021), Hair et al (2020) just labeled the assessment steps of PLS-SEM as confirmatory composite analysis which have nothing to do with CCA as originally developed by Henseler et al. (2014), Schuberth et al. (2018) and introduced to business research by Henseler & Schuberth (2020).

Second, as shown in Schuberth (2021) (and others) the assessment steps of PLS-SEM are not able to identify misspecified models. Hence, researchers relying on these heuristic rules face a high risk that their models are misspecified.

Third, as known in the literature as well (e.g., Dijkstra, 1985) PLS-SEM does not estimate reflective or causal-formative measurement models. PLS-SEM cannot deal with latent variable models.

Against this background, I would not evaluate the predictive power of a model estimated by PLS. This is in line with Evermann & Rönkkö (2021) who mention that usually researchers do not face research questions where predictive power of a model is of interest.

Considering when to apply which type of model, you can have a look at Bollen & Diamantopoulos (2017) and Yu et al. (in press).

Best regards,

Florian

References

T. K. Dijkstra, Latent Variables in Linear Stochastic Models: Reflections on “Maximum Likelihood” and “Partial Least Squares” Methods, vol. 1. Amsterdam: Sociometric Research Foundation, 1985.

J. Evermann and M. Rönkkö, “Recent developments in PLS,” Communications of the Association for Information Systems, vol. 44, 2021.

K. A. Bollen and A. Diamantopoulos, “In defense of causal–formative indicators: A Minority Report.,” Psychological Methods, vol. 22, no. 3, Art. no. 3, 2017.

X. Yu, S. Zaza, F. Schuberth, and J. Henseler, “Representing forged concepts as emergent variables using composite-based structural equation modeling,” Data Base for Advances in Information Systems.

Thanks Florian. Your reply concerning predictive validity is now very clear. This is as complete a reply to some of the most pressing issues and work around PLS and philosophy of modeling that I have seen. You have just added to my reading list!

A lot got bent, shaped and adapted since T. K. Dijkstra (1985) and others work.... not always in a good way.

The misspecified section of your response is a caveat and necessary warning for all researchers.

All the best. Brad

I disagree on the statement "the assessment steps of PLS-SEM are not able to identify misspecified models. Hence, researchers relying on these heuristic rules face a high risk that their models are misspecified." by respected Florian Schuberth , as it is a an always on scholarly debate with established literature in favour and against this statement. In general and for the benefit of the question asking researcher Asma Malik , I agree with respected Bradley Wilson . Tq.

Dear Muhammad Zia Aslam ,

that the PLS-SEM rules are not able to identify misspecified models is well-known in the literature (e.g., Rönkkö & Evermann, 2013; Rönkkö et al., 2016). Moreover, I have shown based on population data, that they are not capable of identifying misspecified models that could have been identified by overall model fit assessment. Could you please point me to some literature that demonstrate the opposite? Moreover, could explain to me why it should be a good idea to calculate, for example, the composite reliability based on inconsistent PLS loadings? See also the Appendix of Schuberth (2021) for an example why it is not reasonable to do so.

I agree that PLS-SEM literature, argues against model fit assessment, however, their arguments are rather weak and in fact most of them are unfounded (Schuberth et al. forthcoming).

Best regards,

Florian

M. Rönkkö and J. Evermann, “A critical examination of common beliefs about partial least squares path modeling,” Organizational Research Methods, vol. 16, no. 13, Art. no. 13, 2013.

M. Rönkkö, C. N. McIntosh, J. Antonakis, and J. R. Edwards, “Partial Least Squares Path modeling: Time for some serious second thoughts,” Journal of Operations Management, vol. 47–48, pp. 9–27, 2016.

F. Schuberth, M. E. Rademaker, and J. Henseler, “Assessing the overall fit of composite models estimated by partial least squares path modeling,” European Journal of Marketing.

Muhammed, I am not sure whether you have interpreted it right.... AND the answer is IT DEPENDS. Even in the 1970s and 1980s Joereskog et al. talked about 3 modeling approaches in CBSEM. SC = strictly confirmatory (no change- not modifications), AM = Alternative Models (this might allow some modification of the model or to test a competing model) and then MG = Model Generating (where it might be very exploratory or data driven or even AI in modern applications)....

SO... I say it depends... but you are right in that the majority of research fit probably in the AM realm. There are model modifications.

What bothers me is the reviewer and editorial bias that rewards or publishes work under the guises of backwards integrated presentations of SC when we all know in reality it may have been AM or MG (even though you do not see the alternatives or the models generated).

Thanks for continuing the debate. Do not call me respected.. and not Bradley either (I think I am getting told off)... please call me Brad.

Hey Bradley Wilson,

I fully agree, but this is not a specific PLS-SEM problem and in fact affects empirical research in general. Similarly, can be said about judging research based on p-values, which got also heavily criticized (see, e.g., Wasserstein et al. (2019).

Best regards,

Florian

Wasserstein, Ronald L., Allen L. Schirm, and Nicole A. Lazar. "Moving to a world beyond “p< 0.05”." (2019): 1-19.

Dear Florian Schuberth as I am not a statistics expert, I was just trying to mention that putting an argument totally against PLS ability in response to a new researcher's question might not be useful. I totally agree with you that there is all that literature that establish a case against PLS abilities from many aspects. However, abundance of published research based on PLS analysis is the support for researchers who want to use PLS for their research and, "however, their arguments are rather weak and in fact most of them are unfounded" is not a universally accepted fact but is concluded by few researchers just like the following articles reply to the PLS critiques. Tq.

Henseler, J., Dijkstra, T. K., Sarstedt, M., Ringle, C. M., Diamantopoulos, A., Straub, D. W., et al. (2014). Common beliefs and reality about pls: Comments on rönkkö and evermann (2013). Organizational Research Methods, 17(2), 182-209.

Hair, J. F., Sarstedt, M., & Ringle, C. M. (2019). Rethinking some of the rethinking of partial least squares. European Journal of Marketing, 53(4), 566-584.

It is ok to debate and argue... it is how we do it given the purpose or problem we solve that matters! B

Dear Muhammad Zia Aslam,

I disagree. The quote of mine refers to overall model fit assessment. Henseler et al. (2014) show that fit measures such as SRMR can detect such misspecification. Hence, their paper provides not a source against my statement. Similarly, Hair et al. (2019) provide no arguments against overall model fit assessment. They state that threshold values for fit measures are not derived for composite models. This is a different problem. They write PLS-SEM does not minimize a discrepancy and therefore, it is not clear whether the fit should be assessed in this way. With the same argument, you should not employ the SRMR to assess models estimated by ML because they have not minimized the SRMR. I addressed this issue in Schuberth et al (in press). Moreover, they wrote if the goal is not explanation, fit assessment is not important. Agreed. However, if you are not conducting explanatory research you are also not really concerned about misspecified models. For example if you want a model that shows high out-of-sample predictive power, usually the structure of the model is not so interesting as long as the predictive power is large. Consequently, predictive models are much more complex and comprise various layers.

I can only recommend to read the recent paper of Evermann & Rönkkö (2021), they provide a nice overview on PLS.

Best regards,

Florian

Dear Florian Schuberth your explanations have been beneficial for my understanding on certain SEM issues. Thank you.

Hi, Dr. Christian M. Ringle and Dr Florian Schuberth. Thank you very much for coming to this post. I ended up developing a conceptual model with reflective dimensions. Causal formative is not widely accepted because of limited understanding in academic literature and academic community so far. If anyone in your team can work on the project of causal formative models in PLS-SEM in a series of papers, that will help in future. Also, Hair et al. (2020) propose a step-wise approach to CCA, which seems relatively concise to those, not in the statistics domain. Because it is new, some exceptions are there. For example, predictive validity must have some separate papers. How to start the process? Usually, at the data collection stage, we are not very clear about collecting criterion variable data and statistical techniques. Statistical technique decision sometimes starts when we collect the whole data set. In that case, it becomes difficult to confirm predictive validity.

I would appreciate it if you could link up any seminar organised by SMART PLS that includes a particular discussion of the CCA approach recommended by Hair et al. (2020) and the comparative one, with a specific focus on papers with procedures and cut-off values not yet accomplished. But, I think for the purpose of solution, the step-wise measurement model quality can be refined with the existing contributions in CCA literature. If there are still contradictions, we heartily expect that both the CCA approach and PLS-CCA (Hair et al., 2020) will bring some viable solutions to the academic community within SMART PLS. Another suggestion is to upload all your seminars on the youtube SMART PLS channel. Renowned researchers on youtube have made the life of newbies easy. Although SMART PLS is not very expensive for competing with other solutions such as WARPLS and ADANCO, if academic researchers get one year or six months free of cost license, this will help them explore more out of it. If it is cost-effective, you may see more output from researchers in future.

With all best wishes and regards,