I have several videos on MGA. The process is essentially the same if there are latent factors. Just examine the structural paths. You can use chi-square difference test for those paths. If you want to do due diligence, you can also do a measurement invariance test. I have a few videos on that as well.

@@Gaskination Thank you very much for your detailed response. Which plugin is suitable for group comparison when there are latent variables? (I noticed on your website that the Multigroup plugin is not appropriate when the model includes latent variables).

Try these: chatgpt.com/share/678da343-f924-800b-9fc4-09ae72e67b22

If you have access to actual performance values (rather than bins), use it as a continuous variable. More variance in measurement is better. If you don't have those original values, then just use one or the other, rather than both.

Here is the link to the full playlist of the newer (2016) series kzbin.info/aero/PLnMJlbz3sefJaVv8rBL2_G85HoUko5I--

Yes, that sounds fine as long as the provided evidence can be used to validate all latent variables.

I just like to be liberal in my moderation assessments. p-values are just guidelines, not cutoffs.

@@Gaskination That's really helpful thanks I just worried I was missing something unique to X2 tests in this context

There is no automated way to do this in SmartPLS. In Mplus or AMOS you could write syntax to subtract one from the other and then bootstrap it and see if the confidence intervals include zero. Another way to do it, while still using SmartPLS is to use a heterogeneity test. I have a tool for it in the Excel Stats Tools file on statwiki.gaskination.com/

Thank you for your quick reply. Learned a lot by following you!

@@Gaskination Professor, I have downloaded the excel file and located the sheet named Heterogeneity Test. I still have two more questions about the tool: 1. It captions group 1 and group 2, should it be used for group difference analysis between different groups, in which circumstance, don't we have to consider the sample size difference? Or it also applies to the difference test between two paralled serial mediation paths in an SEM model (kinda like Hayes' Model 82 in Process Templates, but latent variables )? 2. Is using unstandardized coefficients a must? Can we use standardized coefficients while using the sheet. Looking forward to your reply.

@@李年林-e2g It is okay to use on two paths with the same sample, as in your case. But it does expect unstandardized estimates. I don't have one for standardized estimates.

@@Gaskination Got it. Thank you, Professor Gaskination.

We check VIFs for predicted (endogenous) variables. So, this can be for a formative variable (using its indicators) or for any variable that has several structural paths (from other variables) entering it.

check whether the original correlations are positive or negative, before creating the model. Like in a CFA.

@@Gaskination the original correlation is negative.

@@muhammadumairnazir1673 So then, do you mean that the indirect effect comes out positive or negative and were you anticipating a different direction? If so, just remember that an indirect effect is the multiplication of the two paths from IV to M and M to DV. So, if both of those are negative, the indirect effect will come out positive.

@@Gaskination okay, yes. got the answer. thank you very much

Yes, these two approaches can produce quite different estimates, especially when all factors are reflective. PLS tends to overestimate loadings and underestimate correlations. If all of your factors are reflective, I'd recommend CB-SEM, rather than PLS. SmartPLS now does CB-SEM: kzbin.info/www/bejne/fISUdWeBotJ0eLcsi=gPplnV4n-Lq9ekaK

​@@Gaskination thank you, and i also want to know, my model has the moderator, when i do the measurement reliability and validity in Smart PLS, should i draw the moderating path?

@@Theaxuanan If you are using the PLS side, then yes. If you are using the CBSEM side, then just create a CFA and include the moderator variable if it is latent.

@@Gaskination thank you very much! it's clear.

Usually this means there is a group in one of the predictors that is related to a value in the DV with zero variance. For example, if one of the predictors is autonomy and the DV is job satisfaction, perhaps anyone who responded with 5/5 for autonomy also responded 5/5 for job satisfaction, without any exception (variance). I would recommend omitting one predictor at a time (probably one that doesn't directly predict the DV in the regular model) until you find the variable causing the issue.

The published threshold in Kock's paper is 3.3. If you would like to hedge this with other approaches to CMB, that would be prudent. For example, you can look at the largest correlations or the HTMT values.

yes. it is correct to omit the binary variable from the CFA.

@@Gaskination but my dependent variable is binary

@@raseenarasin2296 Yes. So, the binary variable is excluded from the CFA and then included during the structural modeling.

Try the solutions offered here: statwiki.gaskination.com/index.php?title=Plugins#Troubleshooting

@@Gaskination thank you now its working

It only creates groups that have at least ten members. If one of your age groups does not meet this criteria, then it will not create that group.

@@Gaskination Thank you so much. your guidance means so much in my journey. god bless.

I would not worry too much about indicator error terms unless they are wildly irregular.

@@Gaskination Thank you!

The estimates are just the coefficients. So, in the case of the structural model, that would be the regression weight or beta.

email me at [email protected] and I will help you resolve it.

Yes, if you have a good theoretical or rational reason for it and you provide that justification.

@Gaskination Thankyou sir

I stopped updating the AMOS plugins after version 26 because IBM changes the codebase every time they update the software, and their changes always make my plugins inoperable. Now that SmartPLS can do covariance-based SEM, I've dropped AMOS completely. If you are using 26 or lower, you could try the solutions here: statwiki.gaskination.com/index.php?title=Plugins#Troubleshooting If you need to use AMOS versions above 26, the plugins are available on github. Here is a video: kzbin.info/www/bejne/hnTGhYuVotWLg5Isi=x6SgFW3Few3SwYq0You can also use the Excel Stats Tools Package on the homepage of statwiki statwiki.gaskination.com/Hope this helps!James

This video you are commenting on explains how to do it. Did you have a specific question?

Yes, your observation is correct. In a configural model, each group’s factor variances or loadings need their own reference point. Typically, this means fixing one loading or variance to 1 in each group independently. If you fix the factor variance at 1 for both groups, you effectively impose equality constraints too early, potentially confounding differences in factor structure with differences in variance. This explains why your fit results changed. For configural invariance, it’s standard to let the comparison group’s factor variance be freely estimated, ensuring each group’s scale is set independently before testing further invariance steps.

@@Gaskination Thanks for your reply! In metric and scalar invariance analysis, I did set there the factor variance for the reference group at 1 and that for the comparison group free to estimate. Otherwise, it would be confounding as we discussed. On the other hand, I am confused here. You said that you set factor variance at 1 for BOTH groups in configural invariance analysis in your video. You also said the it was standard to let the comparison group’s factor variance be freely estimated. Which one is more appropriate? The point is that a configural model serves as an anchor for further measurement invariance analyses. May I make it clear: 1. Setting factor variance at 1 across groups Each model would then be standardized with respect to their OWN scaling which means there would be two anchor points. The subsequent comparisons would be confounded by two different set of scaling of the groups. 2. Setting factor variance at 1 for the reference group but free for the comparison group In configural model, the parameters of the comparison group would then be estimated with respect to the standardization of the reference group. Hence, there would be only one set of scaling. The subsequent comparisons would no long be confounded. Am I right? Thank you very much for your reply!

@@francischui2442 I'm glad you're thinking about this. Yes, I think you are correct. Though I suspect the difference would be nominal.

Having zeros in that column just means there are zero missing values. If you get a 1, that means you have one empty cell.

@Gaskination ok thanks 👍

You can, but it is very messy because you have to cross-multiply every pair of indicators. So, it is much simpler to take a two stage approach by saving factor scores first.

@@Gaskination My mediator is well-being and moderator is organizational support. I can take all variables as observed or latent. I wonder is it superior to use all as latent variables? Thanks

@@ibrahimcolak7779 Latent is certainly more precise (as it accounts for measurement error). But if you are doing interactions, a two step approach is much more common and feasible.

@@Gaskination That's OK. Many thanks for your valuable information :)

If you have 750, then you have more than enough. No need to cite anything if you have so much sample. The trouble would be if your sample was too small.

@Gaskination i need citation for that refrence plz include any authors name year whivh i can give as supporting refrence

@@hameemislam753 I don't think I understand what you're trying to support. Are you trying to justify 750 sample size as sufficient? To do this, just do a power analysis and then cite something on power analysis, like: -MacCallum RC, Browne MW, Sugawara HM (1996) Power analysis and determination of sample size for covariance structure modeling. Psychological Methods 1(2):130-149. -Tanaka JS (1987) 'How big is big enough?': Sample size and goodness of fit in structural equation models with latent variables. Child Development 58(1):134-146. -Wolf EJ, Harrington KM, Clark SL, Miller MW (2013) Sample size requirements for structural equation models: An evaluation of power, bias, and solution propriety. Educational and Psychological Measurement 73(6):913-934.

It is usually considered a bad idea to remove items from a formative latent construct unless those items are completely redundant.

@Gaskination sir can we remove 1 complete dimension as I want to work on 2nd order construct only and not make my construct as 3rd order

@@kanupriyathakur-r1j If that makes sense theoretically and it still measures what you want to measure, then you can justify it. There is no citation for that action. You just need to explain why it is the best course of action.

I think I just use the sum of squared loadings divided by the number of indicators.

I would use the one with direct paths and indirect paths, as that will be more complete information. If none of the direct paths are significant, there may be justification to omit them. Or, if you model it in the CB-SEM side and the model fit is good enough without the direct paths, then you could omit them.

@ thank youuu 🙏🏻

Make sure when creating the dataset, you check the boxes for all the variables you want, and not just for LVS. See 1:21

0.700 is not a firm cutoff. It is just a recommended target for reflective indicator coefficients. Formative indicator coefficients do not have the same target.

@@Gaskination Thanks so much prof

This channel is a good start. Also the statwiki.