Monday, June 3, 2013
Multi-group analyses with WarpPLS: The pooled standard error and Satterthwaite methods (and more)
The WarpPLS menu options “Explore multi-group analyses” and “Explore measurement invariance” allow you to conduct analyses where the data is segmented in various groups, all possible combinations of pairs of groups are generated, and each pair of groups is compared. In multi-group analyses normally path coefficients are compared, whereas in measurement invariance assessment the foci of comparison are loadings and/or weights. The grouping variables can be unstandardized indicators, standardized indicators, and labels. These types of analyzes can also be conducted via the menu option “Explore full latent growth”, which presents several advantages (as discussed in the WarpPLS User Manual).
Related YouTube videos:
Explore Multi-Group Analyses in WarpPLS
https://youtu.be/m2VKQGET-K8
Explore Measurement Invariance in WarpPLS
https://youtu.be/29VqsAjhzqQ
I have also been asked how the standard errors reported by WarpPLS can be used in manual multi-group analyses (i.e., not automated) using the pooled standard error and Satterthwaite methods. These issues, as well as other related issues, are now addressed in one single publication:
Kock, N. (2014). Advanced mediating effects tests, multi-group analyses, and measurement model assessments in PLS-based SEM. International Journal of e-Collaboration, 10(3), 1-13.
http://www.scriptwarp.com/warppls/pubs/Kock_2014_UseSEsESsLoadsWeightsSEM.pdf
This publication includes all of the equations used, and also addresses other tests, such as a test of mediating effects using the Sobel method. It is also a good reference for the automated approach employed in WarpPLS.
While using the WarpPLS menu options “Explore multi-group analyses” and “Explore measurement invariance” is recommended (and much less time-consuming), revised Excel spreadsheets are available from www.warppls.com to partially automate the calculations for mediating effects tests and multi-group analyses, respectively:
http://www.scriptwarp.com/warppls/rscs/Kock_2013_MediationSobel.xls
http://www.scriptwarp.com/warppls/rscs/Kock_2013_MultiGroup.xls
Subscribe to:
Post Comments (Atom)
13 comments:
Hi,
I am using WarpPls for research.
A reviewer of my paper suggests me to use "mean structure approach". Can this be done in WarpPls?
Thanks in advance,
Kevin
Hi Kevin.
I suggest you double-check with the editor, but I believe that what the reviewer is asking you to do is the strong version of the test discussed in the section “Use in multi-group analyses” of the publication linked to this post. That is supported also by the second Excel sheet that is linked.
The test is a mean comparison equivalent, where what one compares are the structural paths (betas) across models (which are very close to the mean paths in the resample sets). The strong version that I am referring to is discussed on page 6, starting with the following text: “The procedure just discussed focuses on path coefficients, which are structural model coefficients. This procedure should also be employed with weights …”
The reviewer may have meant something else, so I strongly suggest you double-check with the editor (who can contact the reviewer). Please post an update here if you can, so others know how to address a request like this in the future.
The formula in your pdf for multigroup is incorrect. The Keil et al paper you used incorrectly used the formula provided to them from Chin which they footnoted in the paper. The correct formula used for PLS analysis was orginated by Chin in 2000 and can be found at:
http://disc-nt.cba.uh.edu/chin/plsfaq/multigroup.htm
A more thorough discussion of Multi-group comparisons with good references - including newer procedures can be found in the following paper
Qureshi & Compeau, ASSESSING BETWEEN-GROUP DIFFERENCES IN INFORMATION SYSTEMS RESEARCH: A COMPARISON OF COVARIANCE- AND COMPONENT-BASED SEM, MIS Quarterly Vol. 33 No. 1, pp. 197-214/March 2009
Hi Anon. What is incorrect with the formula?
Dear Ned
I am undertaking group-analysis, where some of the latent variables show significant mean differences. However the paths show no differences.
How do I interpret the results?
Thank you for the excel.
Regards
Mona
Hi Mona.
The LV scores are standardized. Thus, the LV scores' means are zero, for all LVs.
How can they be different?
If you are interested in conducting a comparison of means (CM)-like test using WarpPLS, check the article below. Just keep in mind that you will be comparing the means of unstandardized variables (which could be the unstandardized versions of LVs). LVs themselves have a mean of zero and standard deviation of 1.
Kock, N. (2013). Using WarpPLS in e-collaboration studies: What if I have only one group and one condition? International Journal of e-Collaboration, 9(3), 1-12.
http://www.scriptwarp.com/warppls/pubs/Kock_2013_IJeC_OneGroupOneCondition.pdf
Hello Ned
I was comparing the unstandardised LV mean values across groups.
Hi Ned,
Thanks for this. I am wondering if there is an issue with comparing multiple path coefficients using this method and Type I errors.
Cheers,
Kim
Hi Kim. Since this is a procedure that does not include the multivariate adjustments one gets when including the grouping variable as a moderator in a model, one could argue that there is a greater likelihood of false positives (type I errors). However, my experience is that when moderators are included in models, their true “amplification” effect tends to be underestimated – in some cases you can see clearly that in the 3D graphs in WarpPLS (version 4.0 on). So, the issue is still open, and I am not aware of research that has addressed it explicitly.
Hi Ned,
Thanks for that. I have run the Satterthwaite and PSE methods using your excel tool, and also specified a model with which sample (i.e., 1 or 2) moderated all path coefficients. The significantly different path coefficients were the same paths that were significantly moderated by sample. I was wondering what your thoughts on this approach were.
Cheers,
Kim
A multigroup analysis is indeed analogous to a moderation analysis. However, if the interaction variable (used to implement the moderation analysis) is strongly correlated with one or more of the other LVs in the model, then the results may diverge considerably.
Post a Comment