Message/Author 


We have a client who would like to use this technique to do analysis. Before we can recommend it or not, we would like to learn about it. Does MPlus do it? Has anyone ever heard of it? Thanks. 

bmuthen posted on Thursday, June 28, 2001  1:19 am



No Mplus does not do Partial Least Squares. PLS is a technique championed by Herman Wold as an alternative to the LISREL ML approach by Joreskog (Wold's student). There is a literature on PLS  early writings include a 1982 North Holland 2part volume called Systems Under Indirect Observation (Joreskog & Wold eds.). SEMNET members probably can be asked about more recent contributions. 

Anonymous posted on Sunday, March 06, 2005  4:42 am



If Mplus Version 3.11 connot handle PLS estimation what estimation method is recomended for SEM models with nonnormal data (high skewed data)? 


We have a few estimators that are robust to nonnormality. I would recommend what we refer to as MLR. With this estimate, you obtain maximum likelihood parameter estimates with standard errors and a chisquare test statistic that is robust to nonnormality. The standard errors are computed using a sandwich estimator. The chisquare test statistic is also referred to as the YuanBentler T2 test statistic. If your variables are censored, Mplus also has an option for dealing with censored data. 


Since this thread seems to have ended quite some time ago, I was wondering if the current version of MPlus is capable of fitting path analysis models with latent variables? More specifically, I had someone ask me about fitting a path model that includes a single latent variable (Youth Human Capital in their words), a mediator (program participation), and several covariates (i.e., race, native language, etc...). Is is possible to fit models like this now? 


Mplus does not do Partial Least Squares but has always been able to estimate path models with both observed and latent variables. 


Again, because this thread has not been updated in a while, is the newest version able to do PLSSEM and/or able to deal with formative and/or reflective first and secondorder constructs? The reason I ask is that I have read information along the lines of the following: Traditional SEM programs (e.g., MPlus) are able to compute models including formative secondorder constructs, however, they cannot handle formative firstorder constructs. Therefore, PLSSEM (e.g., using software like Smart PLS) is the only option to calculate models containing both formative firstorder and secondorder constructs. The problem is that there is so much conflicting information in research articles (and related materials) published by authors who seemingly have formed into a number of different groups with potential financial conflicts of interest. In your opinion, can Mplus be used to test SEM models with mediating constructs using bootstrapping procedures when formative measures/constructs exist at both the first and secondorder levels? Thank you in advance for your time and consideration. 


Can you give me an example of a secondorder formative model? 


If we have a standard mediated model with three constructs (C1, C2 and C3), where C1 > C2 C2 > C3 C1 > C3 And, C1, C2 and C3 are all measured with formative indicators. To give a little more background to ensure we are on the same page, Hair and colleagues have authored a growing number of articles in which they have put forward an argument that PLSSEM methods (and alternative software) should be used for SEM research with (1) small sample size, (2) nonnormal data, and (3) formative measures. It is this third point which I am most interested in trying to understand. Traditional SEM is based on the covariances between constructs, which PLSSEM seems to be more focused on trying to maximize the amount of variance which can be explained. Thus, there is a mathematical difference between how a traditional SEM model's results are calculated compared to a PLSSEM model's results. I am trying to understand if this difference in calculations would (really) affect a model if formative indicators were used (instead of reflective indicators). I hope this information is more clear. Thanks again in advance. 


I think regular SEM would be ok with formative indicators for c1 and c2, but I am not sure about c3 because c3 doesn't influence anything try it out to see if the model is identified. I liked this article and debate in that issue: Reflections on Partial Least quares Path Modeling Cameron N. McIntosh1, Jeffrey R. Edwards2, nd John Antonakis3 Organizational Research Methods 2014, Vol. 17(2) 210251 Bollen has a recent paper on a defense of formative indicator modeling. 


Thank you for your response and the McIntosh et al. citation. The article provides some good information about PLSSEM (aka PLSPM) techniques. I was unable to find any recent (last few years) paper by Bollen about formative indicators and how they are best handled in SEM and/or PLS methods and techniques, however, Bollen has published some commentary on the issues in the last decade and beyond. Thank you for your consideration in responding so quickly to my queries. 


You may want to ask Bollen about his latest paper in this series, which might not be published yet. The title says something like "In defense of formative....". 

Back to top 