I am more familiar with SEM tools than PLS. On the other hand, one student I am supervising wanted to use PLS on account that it has significantly smaller sample size requirements than SEM. However, I have my own misgiving with the lack of goodness of Fit statistics in PLS (to my knowledge). Is there any goodness of fit statistics (or its equivalent) in PLS?
I would like opinion from both PLS and SEM users on this matter.
The answers above can be regarded as outdated. The linked paper explains how to assess the goodness-of-fit of PLS path models.
Article Using PLS Path Modeling in New Technology Research: Updated Guidelines
The R package PLSPM provide a pseudo goodness of fit index for PLS models. Recently some efforts have also been made to introduce some other goodness of fit indices. But they are still in experimental stage.
The best way will be to use bootstrap re-sampling to test the model.
Dear Prof. Pandey,
Thank you for your excellent response.
Regarding the pseudo GoF index, my student is using SmartPLS and unfortunately he is not familiar with the R software. I am very interested in the bootstrap re-sampling method you mentioned. Could you kindly elaborate this process or provide the link to this procedure.
Hello!
In the recently released book from Hair et al ( 2013), they have mentioned that there is no global goodness of fit in PLS
http://www.sagepub.com/books/Book237345
You can use different technique such as blindfolding and bootstrapping and many other index to test validity and reliability in PLS,
For example, Bootstrapping is used for significant testing of coefficients in formative measurement models
Blindfolding is used for significant testing of path model coefficients in structural models
Thank you Dr. Zolfaghari, the link you provide is indeed insightful.
While browsing for similar topic, recently I discover this paper by Rönkkö and Evermann (2013) criticizing the use of PLS in causal modelling. This paper made me reconsider using PLS at all. I would like to hear response from PLS users, since I am not too familiar with PLS.
http://orm.sagepub.com/content/16/3/425.full
Here is the link that guide you to PLS goodness of fit
http://old.sis-statistica.org/files/pdf/atti/RSBa2004p739-742.pdf
Despite the criticizes of PLS ( component-based SEM) from some researchers, I used it and I got very nice results, so I think it is very good statistical technique especially for small samples. The CVB/SEM requirements are very difficult to achieve It needs at least 300-400 cases to give reasonable results.
Dear Salehudin
Sorry for delay in responding to your query. In SmartPls you can get bootstrap resampling based estimated. for this you can find the relevant menu 'bootstraping' in the 'calculate' menu shown on the main bar (at the top). It is third in the drop down menu of the 'calculate'.
A 'shiny' based R program is also available to use many of the functionality of the R. Information about it can be obtained from 'sem-n-r' site.
Hope this helps
Thank you for the link Dr.Salehudin it was a great read.
Sorry that I answer late.
I don't consider those criticizes as strong enough to avoid using PLS-SEM. I think that there is more need to researches that address the limitation of PLS-SEM.
I think that the biggest myths about PLS-SEM are two things:
1- Researchers often think that PLS-SEM is an alternative to CB-SEM that can be used when we have small sample size or complex model. (that article address this issue in some extent)
2- Researchers consider validity and reliability testing criterias as a rigorous statistical tool that must be hold true in PLS (this is the scheme that you can see in the Rönkkö and Evermann's article) although this is not true in PLS context . for example, according to the book that I've been mentioned earlier, although we have some cut off point (for example 0.7 for factor loading) in reliability and validity of indicators, no one must eliminate an indicator only based on statistical consideration.
I think when somebody wants to use PLS, must be very cautious in applying that method, and must not rely only on statistical tools. Content validity considerations play great roles in PLS-SEM.
I really recommend the recent book that Hair et al. have written on this topic.
Best regards
Hello Doctor !
Here is the recent article that have been published in response to the article that you have been mentioned before.
https://www.researchgate.net/publication/260037058_Common_beliefs_and_reality_about_PLS_Comments_on_Rnkk__Evermann_(2013)?pli=1&loginT=fDrwuLO-OhmviQTKCSzd0c1ZHU8axRyCK2-dpj1dPa47zcuFu-q9kg*
Article Common beliefs and reality about PLS: Comments on Rönkkö & E...
As I know, the answer is No
Lisrel is a CB-SEM software.
CB= covariance base
Hi Imam and all
1.0 To clarify: SEM has two distinctive paths (base algorithm):
a. Component based SEM: PLS such as SmartPLS and WarpPLS which is based on ordinary least square (OLS) algorithm.
b. Covariance based SEM: AMOS and others.
2.0 Apart from sample size and data distribution (normality), PLS SEM is commonly used for predictive purposes. Thus, it's prediction oriented.
3.0 Covariance-based SEM such AMOS, is normally used to confirm a given theory. Thus, it's theory driven.
4.0 In the case of PLS-SEM, goodness of fit indices are only useful and applicable if you are comparing two or more competing models. However, if your interest is to test for predictive efficacy of your research model, these indices are not of concern. It's beta weights and R-square values in the models should attract most attention.
5.0 WarpPls gives you a very good set of those indices.
Although I am late to answer, I want to add these to the discussion:
Hair et al. (2014) states in his recent book that: "Tenenhaus et al. (2004, 2005) proposed a PLS goodness-of-fit index(GoF) as 'an operational solution to this problem as it may be meant as an index for validating the PLS model globally (Tenenhaus et al., 2005,p.173)'. Henseler and Sarstedt (2012) recently challenged the usefulness of the GoF both conceptually and empirically. Their research shows that the GoF does not represent a goodness-of-fit criterion for PLS-SEM. In particular, the GoF is, unlike fit measures in CB-SEM, not able to seperate valid models models from invalid ones." (p.185)
Hair, J.F., Hult, G.T.M., Ringle, C.M. and Sarstedt, M. (2014) A Primer on Partial Least Squares .structural Equation Modeling (PLS-SEM). Sage Publication, Los Angeles.
And for a comparison of CB-SEM and PLS-SEM, you can see the link below:
https://www.researchgate.net/post/How_to_decide_which_package_is_best_for_my_research_Amos_or_Smart_PLS?_tpcectx=qa_overview_following&_trid=547862dfd685cc5b508b459e_
Well, there is one goodness of fit measure for PLS-SEM but many researchers have advised not to use it. For more info, check http://www.sagepub.com/books/Book237345
Since the objective of PLS is to maximize variance explained rather than fit, therefore prediction oriented measure such as R squares are used to evaluate PLS models.
The answers above can be regarded as outdated. The linked paper explains how to assess the goodness-of-fit of PLS path models.
Article Using PLS Path Modeling in New Technology Research: Updated Guidelines
The Goodness of fit (GoF) index is fundamentally is an index which measures the projection and reliability of the measurement model. Specifically, it can be understood as the geometric mean of the average communality and the average R² of the endogenous latent variables (Henseler & Sarstedt, 2013).
In my research i am using this formula
(GoF = SQRT ( (average AVE) * (average R2) )
{Source: Wetzels, Odekerken-Schröder, and Oppen, (2009), cited in: Lee et al., (2015)}-----> Please Read
The exposure of whether the GoF is small, medium or large, it is mentioned by Wetzels et al. (2009)
1) GoF small = 0.1,
2) GoF medium = 0.25,
3) GoF large = 0.36
On the other hand Lee, Park, and Lee, (2015), mentioned that it also depending on the effect size of R² which has the cutoff values as
1) small: 0.02
2) medium: 0.13
3) large: 0.26)
The GoF does not do what it claims. It is not a goodness of fit index. It does not indicate the reliability of the measurement model. It does not say anything about the model (Henseler & Sarstedt 2013)
Article Goodness-of-Fit Indices for Partial Least Squares Path Modeling
Greetings,
Check this out:
" When using PLS-SEM, it is important to recognize that the term fit has different meanings in the contexts of CB-SEM and PLS-SEM. Fit statistics for CB-SEM are derived from the discrepancy between the empirical and the model-implied (theoretical) covariance matrix, whereas PLS-SEM focuses on the discrepancy between the observed (in the case of manifest variables) or approximated (in the case of latent variables) values of the dependent variables and the values predicted by the model in question (Hair et al., 2012a). While a global goodness-of-fit measure for PLS-SEM has been proposed (Tenenhaus et al., 2004), research shows that the measure is unsuitable for identifying misspecified models (Henseler and Sarstedt, 2012; see Chapter 6 for a discussion of the measure and its limitations). As a consequence, researchers using PLS-SEM rely on measures indicating the model’s predictive capabilities to judge the model’s quality." (Henseler et al., 2014). "
https://www.smartpls.com/documentation/functionalities/goodness-of-fit
Hello,
I clearly disagree with you. The different meaning of the term fit does not depend on whether one uses covariance-based SEM or variance-based SEM, but whether one conducts confirmatory or explanatory research (see http://rdcu.be/FlPa). In explanatory research, one would like to explains as much variation in a dependent variable as possible. The R² is thus a natural measure of fit. In contrast, confirmatory research is interested in the discrepancy between the empirical and the model-implied variance-covariance matrix. This discrepancy can be assessed by a bootstrap-based test of model fit as proposed by Theo Dijkstra and me (Article Consistent and asymptotically normal PLS estimators for line...
) or criteria such as the SRMR.In general, please note that your answer only contains papers from 2014 or older. Since then, the world (at least as far as it relates to PLS) has changed a lot. I strongly recommend reading the following paper to get an updated view on what PLS is and how it can be used: http://www.composite-modeling.com/app/download/12407051124/2016_IMDS_Henseler_Hubona_Ray.pdf?t=1452187001
Should be seen also:
".....In this context, fit (as put into effect by SRMS), RMStheta, and the exact fit test offer little value. In fact, their use can even be harmful as researchers may be tempted to sacrifice predictive power to achieve better “fit” (Hair et al., 2017).
Hair, J. F., Hult, G. T. M., Ringle, C. M., and Sarstedt, M. (2017). A Primer on Partial Least Squares Structural Equation Modeling (PLS-SEM), 2nd Ed., Sage: Thousand Oaks.
If one conducts predictive research (as indicated by a concern for predictive power), fit as indicated by SRMR is indeed of little value. However, Hair et al.‘s statement does not make any sense for structural equation modeling, whose central pillar is fit as determined by means of goodness of fit tests.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling