SmartPLS Formative 2nd order Constructs
HTML-код
- Опубликовано: 21 авг 2024
- In this video I demonstrate how to model 2nd order formative factors and how to do a 2-step analysis using latent variable scores. This is sometimes called the "repeated indicator" approach.
I now have an article published that cites this video.
Paul Benjamin Lowry and James Gaskin (2014). "Partial least squares (PLS)structural equation modeling (SEM) for building and testing behavioral causal theory: When to choose it and how to use it," IEEE Transactions on Professional Communication (57:2), pp. 123-146.
www.kolobkreati...
Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RUclips channel where we post a new video almost three times per week: ruclips.net/channel/UCiujxblFduQz8V4xHjMzyzQ
Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074
And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en
Check it out! Thanks!
Great thanks to you for your helpful videos. I can NOT finish my thesis without your videos. You "are" my statistical professor!
+Kathleen VanBenthem for some reason I can't reply to you directly, so I will do it here. The answer is yes. It should be nearly, if not exactly, identical.
The datasets are available on the statwiki. See the url at the beginning of the video, or go to my channel homepage for a direct link.
Mr James Gaskin, thank you very, Sir, for all the videos you uploaded on RUclips concerning SPSS, AMOS and SMART-PLS... they are all very helpful and useful. I listened and follow up most of them. I actually had no idea of SMART- PLS or AMOS, Now I got a few ideas on them... I'm really struggling with my analysis in my PhD thesis. I hope I can get in touch with you for consultancy in your free time. thanks again... Barrie.
I’m happy to answer short/concise questions if you get stuck. I answer a dozen or so questions from people all around the world each day. If your question requires understanding a model, please send a picture of the model instead of trying to describe the relationships.
Hi, James. I switched from AMOS to SmartPLS, because I have some formative factors which (as far as I know) cannot be handled in AMOS. I have a hard quesition: In my model I have a third order construct. It is formatively formed by 3 2nd order constructs. Each of 2nd order constructs is formed by 3 1st order constructs. If I follow the logic of your video, I see two opportunities: 1. Make a fake model without the third order construct and just connect my second order constructs (IVs) with the DV. Then I can do what you proposed: add newly calculated data to my initial data file and run the same model but this time (a) substituting all items with the newly calculated 2nd order data. (b) introduce the 3rd order construct and repeat the procedure one more time. An alternative approach would be to introduce in the initial model (!) the 3rd level construct and again follow your steps - add items (a)to the 1st order construct, (b) add all items from all 1st order constructs to the 2nd order construct. (c) add all items from all underlying 1st order constructs to the 3rd order construct. (It sounds weird I think, but seems to be logical). I made both ways and the results are slightly (very slightly) different. Which approach would be more correct?
+Eugen Bogodistov You are correct that both will work. The difference in results is likely due to the different way error is measured. I would recommend using whichever approach you can most simply describe. Both are fine, so either will work. You just need to be able to write it up succinctly.
Formative 2nd order is for a multi-dimensional construct (something that is made up of a few different things), where each dimension is captured as a latent variable (i.e., has many indicators or items that are used to capture it). Hope this helps. There is some literature on it as well. Just google it.
I have four first order constructs: X, Y, Z, U;;;; C is the second order construct of those first order X, Y, Z, U. I run second order construct test relationship between construct A and construct C. Result show there is insignificant test the relationship between A and C (second order construct).
You only need to do the two step approach with the Latent Variable Scores when the 2nd order formative factor is also endogenous. If it is just exogenous, then you are fine.
I invert the model back to reflective because this is what you do in a repeated indicator approach. See the middle model in figure 2 of this paper in the Journal of Applied Statistics: "Using the Hybrid Two-Step estimation approach for the identification of second-order latent variable models"
As for the second part, it is just a single score, so it is not really "reflective" or "formative" despite my comment that it is "reflective I suppose". Hope this helps.
Hi. I'm currently doing my analysis for reflective-formative type II by following this video and your Lowry & Gaskin (2014) article. The analysis worked fine.
My problem is how to report the measurement model result for the 2nd order formative constructs(weight, t-stat,VIF)? The PLS report only showed the weight, t-stat & VIF for the 1st order constructs.
I have never done this before, so the answer I give is just an educated guess. I would say that it can be done either way and the results will not substantially differ (they will differ, but not much). The more accurate approach would be to do it before the second step.
So my guess is that it is because when all those first order variables are aggregated into the 2nd order, their aggregate variance (represented by C) is not correlated with A, even though many of the individual 1st order factors do significantly correlate with A. The aggregate is something fairly different than the sum of parts. You might try breaking C up into two separate 2nd order factors.
Dear James,
I have got a model which is a bit different of what you demonstrate in this video. Actually I have a type 2 construct as the dependent constructs which has 4 dimensions. and the independent is a reflective unidimensional construct. the point is how we should run the analysis (in SmartPLS or any other) to be able to see the effect of the IV on each dimension of the our type 2 construct (first-order). looking forward to your kind advice.
Tony
I created three first order reflective constructs, and these are the three dimensions of the second order formative construct called burnout. I hope that clarifies.
I got it from the smartpls forum. They might have a reference mentioned somewhere in there. Best of luck!
I'm 95% positive that you can just use the repeated indicator approach, rather than having to create new constructs using the Latent Variable Scores. So, just do the first part of this video (create 2nd order constructs that include all the items from their first order constructs), but don't worry about producing a new model after that with Latent Variable Scores.
Dear Dr. James Gaskin, in your video you apply a mixture of the repeated indicator approach and the use of latent variable scores in a two-stage approach, which is appropriate in case of an endogenous HOC (in this case the 2nd order construct BurnM that is of the reflective-formative type). In the second step of the two-stage approach, however, you use the latent variable scores of the latent variables in your structural model (BurnM, EthicalCon, UnsupBos, Satw), while Hair et al. (2014: 233-234) suggest to use the latent variable scores for the LOCs (the first-order constructs BMa, BMb, BMc) and embed the HOC (BurnM) in the nomological net (so leaving the setup of the other latent variables in the model unchanged). So, now I'm in doubt of what to do in case of a mixture of the repeated indicator approach and the use of latent variable scores in a two-stage approach, which is appropriate in case of an endogenous HOC: should I take the latent variable scores of the first/lower-order constructs, or the latent variable scores of the latent variabels in my structural model?
Either approach should yield similar results. And there seems to be literature support for both approaches. Try them both and see what happens :)
James Gaskin Hi James, I have been under the same confusion as Jos. I have gone through a number of academic articles that deal with HOC including your one with Paul Lowry. I haven't found article to support your approach. Would you kindly refer one?
leenmorrison The approach from this video uses latent variable scores. A quick search on scholar.google.com for: "latent variable scores" "second order factor" formative
seems to return several promising results. The one by Becker 2012 looks particularly promising.
Jos Schijns James Gaskin
I'd say James' method would work fine. But indeed, Hair et al (2014, p233) explain it as follows (an edited quote from the book):
"In many formative-formative and reflective-formative Higher Order Component (HCM) applications, [...] (i)n this kind of situation, a mixture of the repeated indicator approach and the use of latent variable scores in a two-stage approach [..] is appropriate. In the first stage, the repeated indicator approach is used to obtain the latent variable scores for the LOCs, which, in the second stage, serve as manifest variables in the HOC measurement model (Ex. 7.25). Thereby the HOC is embedded in the nomological net in such a way that it allows other latent variables as precedecessors to explain some of its variance, which may result in significant path relationships."
Problem with this method in SmartPLS: in the second stage, you would again do a repeated indicators now repeating the newly created LOC LVS again in the HOC. Or am I missing something?
***** In the second stage, you would only include the LVS values. At least, this is how I understand it.
Just means that there are probably other variables that you are not considering that better explain the variance in the dependent variable. This amount of variance is okay if it is a financial model or predicting any sort of firm performance. It is also okay if you are doing a very exploratory study in a new research area. It is also okay if the DV is something new and untested. I'm not sure about which literature to read.
As always a great tutorial video in SmartPLS. It would be great if you could upload the datasets you use in order to follow your examples. Thank you
You only need to use the Latent Variable Scores if you are using endogenous 2nd order factors. So, if you are predicting the 2nd order factor with some other latent construct then use latent variable scores. Otherwise you are fine.
Try Becker et al 2012 "Hierarchical latent variable models in PLS-SEM: guidelines for using reflective-formative type models". I think this is the right one.
If the indicators are all for a single factor, then they should just be made formative. You may need to end up trimming a couple of the reflective ones. Or, you can create a 2nd order satisfaction factor with 2 first order factors - one reflective and one formative.
thanks a lot James! Have a great weekend!
Hello James. Thanks for your help with this video. My question is: I have a second order formative construct and have followed your 2-step approach of using the latent variable scores and then creating another model. How do I also include the reflective sub-constructs (in your example, BMa, BMb, and BMc) into this newly created model? In your IEEE piece, I see that you have beta weights for your 3 sub-constructs that make up your second-order formative construct communication quality.
I think I figured out an answer to my question, but would still like your opinion. My second order formative construct is exogenous (predictor only), so simply using a repeated indicator approach works in SmartPLS? I simply made all items from each sub-construct reflective of my second order construct. Then, I created 5 sub-constructs with the same items (divided among the constructs of course) and made them formative of the second order construct. Is this method correct? Thanks again!
Yes, it should work. Nice work.
Dear James,
I have second order construct as a moderator, and I use variable score to do interaction.. My question is, as the PLS variable scores are standardized, in interaction do I need to click mean-centralized? My own suggestion is, to copy and paste variable score (unstandardized) and later click that, which later for interaction graph may be more useful... The way that you do, lead to standardized the DV also.. Please let me know, your helpful suggestions..
I appreciate your contributions to my research:)
Dear James,
So, in both cases when we have Formative-Formative and Reflective-Formative construct, the repeated items on the 2nd order should be reflective?
SmartPLS calculates it all for you, so I'm not sure I understand the question. Still report the loadings for the measurement model and the path coefficients and R-square for the structural model.
Question for James Gaskin... I assume your final coefficients from predictors of burnout ("ethicalCon" and "unsupBoss" ) to Burnout would be identical if you had not used the lv scores for those predictors (the new scores you added in the second dataset) and had just created the same reflective constructs for these two predictors as in the first time you ran the model. They are approx .3 and .4 each time.
Dear Dr Gaskin,
Thanks, Great videos. It would be great if you could update them according to new version of SmartPLS as small additions have been done there.
I would love to do that sometime. I just need to find the time... :)
Saying they are not good enough does not depend on the R-square, but on their loadings, reliability, and validity.
Sir Gaskin, you are the best researcher i saw since you share all your knowledge ! Thanks a lot! Following you... I wanna ask what kinds of threats I can have if i apply repeated indicators approach for third order LV (with different number of indicators).. OR only can i apply for 2nd order LV in the same model? TIA
I've never considered third order variables... If you must do it this way, I would recommend creating LVS for the 2nd order first. Then use those to create your third order variables.
James Gaskin Thank you very much!
In my research, one of my formative construct (i got 3 formative construct in my model) have 3 indicator, but each of them is reflected by >=1 questions in my questionnaire, so there are 8 blocks shown in the smartPLS construct. Does this means they are already a 2nd order formative construct in the first place?
I happen to learn about PLS-SEM (well, basically multivariate analysis) in a short period of time (2 weeks to be exact :D) so please forgive my lack of knowledge
You're too cool mister!
Dear Professor Gaskin,
thanks for your great tutorial on how to analyze formative second-order constructs with PLS. As part of my PhD thesis I'm introducing a new reflective first-order, formative second-order construct and your instructions are very helpful. Nevertheless I wonder if it is somehow possible to use the latent variable scores from PLS output in AMOS with the purpose of assessing model fit? I don't know if it is possible at all. Maybe you can help me? I'm also interested in the reasons if it does not work.
Thanks in advance for your help.
As long as the model in AMOS is reflective, then you can assess model fit. However, your second order factor is formative, and so it won't work.
Please, could you help me???
Hello, the 2dn order model is the regular step for the PLS Method? .
All results from my first order model are ok.
So, is necessary that I test it, In the 2dn orden model?.
All constructs are reflexives. So.. If I test it in the 2dn order I need use only the second step from your video?
Thank you very much
Please update video on 2nd order reflective factor using repeated indicator approach...thanks in advance....
Dear mister Gaskin,
Thank you very much for this helpful video.
I am currently working on a paper that includes a model with a second order reflective variable, and I was wondering if there is a paper or book chapter I can refer to when using the approach you show in your video?
The SmartPLS handbook does include a chapter on second order reflective models, but it fails to provide the instructions for how to execute such a analysis in SmartPLS.
Mr. Gaskin! thanks your reply, I have another question.
if all my model include the following relationships:
1.A(1st order construct) to B(2nd order construct)
2.A(1st order construct) to C(1st order construct)
3.B(2nd order construct) to C(1st order construct)
4.B(2nd order construct) to D(2nd order construct)
5.C(1st order construct) to D(2nd order construct)
all constructs and indicator variables are reflective.
Dear James,
Can u plz introduce a reference for your analysis method for 2nd order formative construct?
Thank you
Dear James, thank you for this very helpful video! If I am right, your higher order construct is reflective firstorder, formative second-order.
I want to model a formative-formative construct. Now my question. When I "drop" all items on the second order construct (in order to measure it), do I model them formative or reflective. (This question might sound weired, however, I really get messy results, if I model them formative). Thanks in advance.
Hi Dr James,
Does this method work for formative-formative type models (Becker et al., 2012)? If so, does that mean I just need to reverse the arrows instead of following what you did in the video? And if not, then what method should I use? Thanks so much!
Thanks for your work! Could you simply explain how can I determine whether to use a formative second order or first order?
I create another project. I run test between construct A with construct X, Y, Z, U ( first order contructs of C). Result reveal that there is significant tested between construct A and the four constructs! Why A significant with X,Y,Z,U, but insignificant with C?!
James Gaskin Dear Prof. Gaskin, I read your recent publication (i.e. Lowry & Gaskin, 2014). But can you illustrated the validation for 1st order with formative indicators? Thanks.
Thank you James, your reply is very helpful, I really appreciate. Indeed, my research is about purchasing maturity's impact on purchasing performance, about this result (small R square), can I also say the indicators I designed for the purchasing performance(reflective construct) are not good enough? In other words, if I test more and better indicators for performance, then the result of R square would be higher?
Thank you very much!
Hello Mr. Gaskin! Thanks a lot for uploading this, your video is the answer i've looking for. I have been frustrated with the t-stat result of my formative construct for some time. I just want to ask about why did you create 3 first order construct? What do you mean by the burnout have 3 dimensions? Thanks! :)
do you have any video showing how to do the hierarchical latent with repeated indicator approach? perhaps having unequal numbers of indicators. even though repeated indicator is advisable to be used with with constructs of equal number of indicators (Chin et al, 2003; Lohmoller, 1989; Ringle et al, 2012) but this still can not be proved in the current literature (Becker, Klein and Wetzels, 2012).
This paper will guide you to all the references you need "Hierarchical Latent Variable Models in PLS-SEM: Guidelines for Using Reflective-Formative Type Models"
Thank you for sharing your knowledge... your videos and wiki are a Godsend for those of us trying to improve our understanding of SEM. My question regarding this video is related to reporting results for tests of convergent and discriminant validity. I presume those measures would come out of the first model rather than the second with its single indicator given that all constructs will then have an AVE of 1. However, would you report AVEs and perform your Fornell-Larcker for the first order constructs or the second?
Both. For first order factors that do not belong to a second order factor, use the first model. For all second order models, use the second model.
Dear Prof. Gaskin:
I applied your approach for SmartPLS Reflective-formative 2nd order constructs. Everything was fine and variance defined by the lower order constructs was one; however, when I executed bootstrap, the t test and P value which must be significant was showing insignificant value. I used your approach to form first order construct model and executed the bootstrap, the same construct was showing significant value. Could help me sort out the problem.
These two statistics are with respect effect of Relationship Marketing Orientation on Business Performance Effectiveness
T Statistics (|O/STDEV|) P Values
1.566 0.118
1.323 0.186
T Statistics (|O/STDEV|) P Values
7.757 0.000
13.519 0.000
1.698 0.090
I will be very grateful to you
In the first model, the lower order dimensions will completely explain the variance in the 2nd order factor, which makes it so that any other variables cannot explain the variance. That is why we use the two-step approach. The LVS model allows for variance to be explained by other variables.
Also, another questions if you can help me:
I have a SATISFACTION variable which is made up of two types of model: a reflective,for measurement of overall satisfaction, and formative, for measurement of satisfaction. How can I make the same variable (satisfaction) have both formative and reflective indicators using smart PLS?
So, coming out of the same construct, make a formative AND a reflective model.
Thanks
Thank you for helping us.
Rabee Shurafa commented , Dear If the model is reflective and not formative can I do the same? Because I have second order model but all the items are reflective and not formative.
You replied : If they are reflective, then there is no need to do this two stage approach. Still use the repeated indicator approach, but you do not need to extract latent variable scores.
Thank you so much for this comment. This is what I did. However, I did it, but discriminant validity for all items of second order was not met. In this case, what should I do? if I drop items, the number of items become few, 7 from 33 items, removing some of the constructs. please, I m waiting your advice.
Perhaps these are really part of a higher order construct. This is usually my first suspicion if the correlations are too high.
If it is an endogenous 2nd order formative factor, then you need to use latent variable scores (2-step approach). Otherwise you can just model it all at once.
Dear James,
I have a Type II (Reflective-first order-Formative Second Order) model. I have developed indicators for the Formative second order construct, so I do not have the problem of the swamping effect. I do not have an exogenous variable affecting my first order formative latent variables. Do I still need to manually do the cut/paste of the latent variable scores, or in this case you do not need to do this procedure ?
Thanks for your answer,
Juan Carlos.
Hi Dr. Gaskin,
In your video you used the scores of the second order (BurnM) obtained from the first stage of Reflective-Formative type to be used as single indicator in your second stage model (ignoring LOCs scores). However, Hair et al, 2014 book page 233-234 describes the procedure differently. In the first stage of the Reflective-Formative type we obtain the latent variable scores for the LOCs then these scores serves as manifest variables for HOC (BurnM in your example) in formative way. Obviously, if we apply their procedure to your example, BurnM scores in the second stage will be different form it scores at first stage as we will have normal formative model in the second stage using LOCs scores as indicators.
Could you please let me know which one is right or regarded to render more accurate results. Much appreciated
In a covariance-based program (like AMOS or Lisrel), Hair's is the approach I would take because we can model just the measurement model while not modeling the causal model. However, in a PLS program (like SmartPLS), we have to model both models simultaneously. In this case, it makes more sense to model it as intended (with higher order factors), and then use the latent variable scores for that factor as shown in this video.
Thanks for replying. Wouldn't the HOC scores be overestimated when using the repeated approach in the reflective-formative type model?(because of the swamping out effect). If so, treating the scores of the first order latent variables as manifest variables for HOC may be a solution!
Please let me know what you think. Many thanks
Try it both ways and see if you observe any meaningful difference. I would be surprised if the difference was great. If so though, you are right that using the first order LVS as manifest variables of the HOC would be the best approach.
Thanks James!!
Hello Mr. Gaskin, first of all let me thank you for your videos. They are amazingly helpful. Regarding this specific video I've got a few questions left regarding the evaluation of a model.
I've got a model with a formative second order construct. The first order constructs (to be seen as dimensions) are reflective and so are all other constructs in the model. The second order construct is solely endogenous. When it comes to evaluation of the model now, three things are unclear to me as I am now dealing with formative first order constructs and not indicators as I am used to:
1. Do I need to test for multicollinearity between the first order constructs? If yes, how? I am a little bit confused as the construct is formative with reflective indicators.
2. Are the T-Values listed under "Total Effects" in the Bootstrapping Report those I am looking for when evaluating significance?
3. Do I include the second order construct and therewith the indicators of the first order constructs when assessing discriminant validity with the help of cross loadings and the Latent Variable Correlations?
I know these are a lot of questions, but I searched the internet for several hours now and I found just a few sources (e.g. Jarvis 2003) on formative-reflective second order constructs that couldn't help me. That "mixture" confuses me a lot.
Thanks a lot for your help in advance!
1. To calculate the VIF (which you probably should), you would have to use the latent variable scores. So, create LVS in SmartPLS, then test for VIF in SPSS.
2. Yes. Pretty much anything greater than 1.96 will be significant at 0.05. SmartPLS 3.0 calculates the p-value for you.
3. Usually we assess discriminant validity only for the highest level factors. So, if a factor is first order, but part of a 2nd order factor, then we would not assess discriminant validity for that first order factor; just for the 2nd order one.
James Gaskin
Hello Mr. Gaskin, thank you very much for the quick response! That helps a lot. Keep up the great work!
Hi Gaskinatin,
Great and helpful video! Thanks for the clear advice. I've got a question. At 2:28 in the video you reverse all items for BurnM so they become reflective (Mode A) instead of formative Mode B. Why do you do invert the model? It is a reflective-formative model shouldn't the indicators stay formative? The same question for the second stage when the latent variable score for BurnM is used. Shouldn't that be also in Mode B? Thanks!
Hi Gaskination,
Many thanks for your illustration. Do you happen to know how to deal with dummy variables using SmartPLS?
More specifically I want to test to the effect of student orientation on student satisfaction and I want to control for the mission of the university over student satisfaction. I have five groups of missions, so should I do like this:
Group1: 1.0.0.0.0
Group2: 0.1.0.0.0
Group3: 0.0.1.0.0
Group4: 0.0.0.1.0
Group5: 0.0.0.0.1
Does this sound right?
Many thanks in advance
Dear Professor Gaskin, is it possible to make a 3rd order formative construct like this? I am interested in testing two factors affects the second order constructs and want to use this second order constructs with another second order construct to form a latent construct. Thanks a lot!
+Liyuan Wang Sorry for the delay in responding. RUclips has stopped notifying content creators when comments are made on their videos... So I just thought no one was commenting on my videos... It should work..., but I've never tried it.
Yes, I mean all 7 indicators. Thanks a lot James. BTW, I read an article by Becker et al. 2012: Hierarchical Latent Var Model in PLS-SEM ..., is the approach you shown in this video similar to the two-stage approach they referred to? (thanks again, Prap - PhD candidate, Netherlands)
Hi James! I have a question.... I have a formative 2nd order dependent construct and 2 reflective independent constructs. I have also hypothesized an interaction in my model. At which point in the two-step process you describe in your video should I compute the interaction term in the model? Should I do it in the first step, before I get the latent variable scores or should it be done after in the second step? Thanks for your videos. They have been very helpful.
Most studies that develop and validate instrument do not categorically tell us whether they are second order, or not. How could one identify that
It is second order if it has multiple dimensions or underlying factors.
thank you very much, it is very helpful.
thanks, Mr. Gaskin
can that means I can only just do an one-step analysis if my model have only endogenous 2nd order reflective factors and haven't any endogenous 2nd order formative factor, but if my model includes an endogenous 2nd order formative factor, then I need to use latent variable scores (2-step approach). So, It depends on whether my model include an endogenous 2nd order formative factor or include only endogenous 2nd order reflective factor?
thank you very much!
would you please explaining to us how to do exactly with a second order reflective construct
Gaskin, r u sure of reversing items back to be reflective (in the 2nd order construct) while building its 1st order formative componants? Moreover, the arrows of the 1st order indicators still reflective even the construct are formative. Is there a logic behind this procedure?
For the model involved, the items for the first order were definitely reflective, while at the second order they were definitely formative. Here is a video where I explain this further: ruclips.net/video/gw0xvvJw-AM/видео.html
Hi James! Thank you for your videos. I followed this method in testing the structural model in my dissertation papers since I had second-order formative constructs in each model. One of my committee members has asked me why I also had single indicator measures for the constructs that are not second order constructs e.g. like the unsupportive boss construct in your video in the second order model, and suggested I might be loosing something doing it this way. Do you have a reference for what you do in the video? Thanks.
Search google scholar for "latent variable scores" and "repeated indicator". The Becker et al 2012 paper will suffice I believe.
i did made second order contructs C in the project 1. and i built second project 2. in the project 1: relationship between construct A and C are insignificant. but in project 2, construct A significant with some first order of construct C. How can i explain this???Please help.
Hi James,
Thanks again for your endless effort in making us educated on precarious statistical topics
I have a quick question for you; after forming the 2nd order formative constructs in SmartPls, how about using the new data file as a feed for AMOS to do co-variance based analysis on these newly formed formative constructs? is that a viable strategy?
Sameer Sheikh hmm... that's an interesting thought. There is no reason why that shouldn't work. It would just be path modeling at that point, and AMOS can handle that just as well as PLS. Good thinking :)
James Gaskin
Okay thanks! I have a model consisting of two formative constructs- Form A & Form B; to be precise they are second-order formative, first-order reflective. Up to this point it’s complicated but manageable. But next I have to show complementarity of Form A and Form B.
In the field from where I come from complementarity is usually shown either as (i) an interaction of the two constructs or (ii) a 2nd-order formative construct formed out of both constructs.
Now I want to go formative way of complementarity, but the dilemma is, should I
1) Make a third-order ‘complementarity’ construct to show complementarity (1st order reflective- 2nd order formative- 3rd order formative)? if i go this way then i basically make a new third-order formative construct by combining Form A and Form B
Or
2) Make a second-order ‘complementarity’ construct (by combining all the 1st order reflective constructs that made up Form A and Form B into one second-order formative construct)? if i go this way i am in a way dissolving Form A and Form B and combining their first-order constructs into one unitary second-order constructs
Kindly let me know what's your thought on this; which one of the above makes more sense conceptually and statistically?
Sameer Sheikh I would go with the first option you propose, although I am not very familiar with complementarity.
+Sameer Sheikh oops. I replied as gaskin stories (a different youtube account I have).
James Gaskin no worries, as long as it's a Gaskin brand :)
I am guessing the process of forming 3rd formative construct from 2nd order formative construct is the same as making 2nd order formative construct from 1st order reflective construct?
Second, do you think it might need an effort from me to convince the reviewers about the viability and soundness of 3rd order construct, considering you hardly find any paper that uses one
That looks correct.
Thank you very much for posting your tutorials! I have a question concerning the indicators when using a second order construct. If I use the repeated indicators approach and run the PLS algorithm, I get two values for each indicator. I wondered, what happens with those indicators I use for the HOC? Do they also have to meet the criteria for the measurement model? Or how do I assess the measurement model of the HOC? Thanks a lot for an answer.
+MJ Student For the HOC, look at the regression weights between the HOC and the first order constructs.
Thank you very much for posting tutorials! It walked me though my analysis. Just to clarify my concepts, please tell me if I have to check mediation/moderation for (say) each of the 3 dimensions (1st order formative) of Burnout, I will copy the "latent variable scores" of each of the three dimensions the same way you did for Burnout (2nd order formative) construct?
You can do it all at once for the 2nd order factor.
@@Gaskination Thank you.
@@Gaskination Thank you very much
Hi Prof. James
firstly, thank you for your shared video. I have (IQ) construct in my model, this construct can measured by four dimensions with 14 items. in literature review, some articles used it as second-order, while others used as first-order (as reflective construct). please advice me which method is the best.
Salih Hajem It really depends on the items. If the items are all interchangeable and mean essentially the same thing, then it can be a reflective first order. However, this is unlikely since there are 14 items. It is more likely that there are multiple underlying dimensions (or first order factors). You can probably model it as a 2nd order reflective construct. If the loadings from first order to second order are all strong enough (roughly greater than 0.700 on average), then you are fine to model it reflectively as a 2nd order factor.
James Gaskin thanks to answer me, Please I need more explication. IQ construct can measured with 4 dimensions (IQC,IQF, IQCU and IQA). each dimension has 3 Items. Can I measure IQ construct with combined all 12 Items together without dimensions.
Salih Hajem You can try, but it is probably better to measure them in a 2nd order factor with four dimensions.
Thanks so much
James Gaskin Please Prof., can you support me by sending any Guide, book or any articles for analyzing 2nd-order formative construct.
my email: salihhjm74@yahoo.com
because I'm predicting the 2nd order factor with some other latent construct(no. 1,4 and 5) in my model. So, I need to do a 2-step analysis using Latent Variable Scores. no matter whether the path coefficient is 0.
is that right?
Thank you!!
Dear Dr. Gaskin,
If the dimensions are already pre-determined, can we use this approach? For example for e-service quality the dimensions are efficiency, responsiveness, system availability, compensation, contact, privacy and fulfilment. Should we use the two-stage approach or will the repeated indicators approach be better?
Also, does the approach works on reflective-reflective second order constructs?
Thank You.
The approach here with latent variable scores is only necessary if the formative 2nd order factor is being predicted by some other factor. If the factor is instead reflective/reflective, or if it is not being predicted by other factors, then you don't need to generate the factor scores.
Hi Dr. James Gaskin,
Do you have any video on handling formative constructs in AMOS? If yes please share. Thank you very much
+Amjad Shamim Nope. It is not meant to be done. AMOS is designed for reflective constructs only.
Thanks for this video, I have a second order model in my study: Commitment, which has 3 types: normative, calculative and affective, each have 5 indicators: 15 in total.
In my model I made a construct for "commitment" with three second order constructs (one for each type), with their 5 indicators. According to what I saw in this video, I have to also make those 15 indicators come out of the "Commitment" variable, not only out of the second order constructs.Is that right? (otherwise I get error)
Hello Dr Gaskin
I have a query. I am constructing a tool and after efa i have 5 sub dimensions. The sub dimensions are first order reflective measurement model. And the subdimensions further add up to the construct which is second order formative. I hd to delete few first oder indicators. After deleting those scores do i have to omit those items from the second order model as well?
Yes. If the indicators do not appear in the first order, then remove them from the 2nd order.
Thanks for the video! Could you please reccomend literature that supports this? A recent article (Becker, J.-M., Klein, K., & Wetzels, M. (2012). Hierarchical Latent Variable Models in PLS-SEM: Guidelines for Using Reflective-Formative Type Models) questions the use of this two-stage approach for formative hierarchical latent variables. They state that the repeated indicator approach with mode B on the higher-order construct should be used for reflective-formative hierarchical latent variables.
Thank you so much
Hi Gaskin
Thanks for the tutorials, they are really useful.
I need some clarification for some of the steps you followed.
Which PLS version you are using? I downloaded a demo version 3. I did all the first steps of the second order construct, but I couldn't calculate the Latent variables scores because there is no option for default report. So please advise me how to do it.
I'm using version 2. I haven't used version 3 in a long time, so I'm not positive, but the latent variable scores (sometimes called factor scores) should be hiding around the report somewhere.
Dear Dr. Gaskin,
Is it a problem when my reflective constructs have a different amount of questions? The result of the test is different when I take the average of each formative construct (the formative constructs are becoming a single-item construct) instead of all the questions separate. For example, does it make sense when “BMa” consists out of two questions and “BMb” out of three questions?
Erwin Dreschler
No, not a problem, as long as they have at least two indicators (four is best). Removing items or collapsing formative factors will always change the results because the items are all interrelated and error is not accounted for as well when averages are used.
Thanks James for your valuable advise. But I am confused about a point again. In my case, for example, BurnM has 16 indicators and BMa has 10 of them. So, BMa divided to two sub-constructs includes BMa1 and BMa2. Now my question is exactly here. Should I repeat the indicators for BMa and again for BMa1/BMa2 ? I mean should I repeat the indicators (three times) for BurnM , repeat for BurunMa and repeat for BurnMa1?
I really appreciate your advise
+Azadeh Tabesh I think I would have to see a picture of the model. My email is james.gaskin@byu.edu
Hi James, I realize this is a 10 year old video - do we still have to do it like that ? I suspect we can avoid this step in latest version of Smart PLS ?
Here is the latest video for SmartPLS 4:
Formative 2nd order: ruclips.net/video/LIz9QQ-3g6I/видео.html
Reflective 2nd order: ruclips.net/video/sGqp6hRktI8/видео.html
Dear Mr. Gaskin, I do have the second order formative construct organizational performance. I assume that this second order formative construct is measured by three first order constructs. Now, I am not sure how I should create the direction of causality, cause in your video it seems that it does not matters in which direction the arrows are constructed. I would create the model with arrows from indicators to first order constructs and as well from indicators to second order constructs. I had runned the pls algorithm in smartpls for each possibility and it creates different types of latent scores. I would be pleased for any recommendation. Thank you in advanced.
Usually the first order constructs that are dimensions of formative second order constructs are reflective. So, have first order indicators as reflective. Have the arrows from first order latent factors point toward the 2nd order latent factor. Then have the repeated indicators for the 2nd order factor be reflective. Hope that makes sense...
James Gaskin
Thanks alot for your very fast answer. Yes, that makes sense to me. However, how should I handle it if the dimensions/components (first order constructs) are measured formative as well? Should I then create both, the first and second order construct with its repeated indicators as formative? I would be very thankful if you could answer this question.
Dear James, I have a question regarding with the R square. My research model ,including 5 reflective constructs and 1 second-order formative constructs are very significant in terms of validity, discriminant or reliability by using PLS, the t-statistics is also very significant. However, the R square of two endogenous latent variables are very low, around 0.1, what can I evaluate on this finding? does that mean this model has no meaning? do you know some literature about this?
Thank you!
Hello, Mr. Gaskin! Thank you for this video, I have a question,
Should I have to do a 2-Step analysis, can I just do a analysis before 3:06 if the path coefficient of "from UnsupBoss to Burnout" and "from EthicaiCon to Burnout" isn't 0?
Thanks
Thanks Sir, We are having 2nd order reflective model. So what we need to calculate in this case. Do we need to calculate the path cofficients and r square for the same..?and if no than how we can calculate the loadings.
Professor Gaskin,
If I have a reflective.reflective type of 2nd order construct wit for lower order components with different number of indicators among them. Should I use a two-step aproach in order to assing the indicators to the higher-order component? The same way you do with the formative type?
No need when it is reflective, reflective. Just model it without any repeated indicators. (unless smartPLS won't let you do that... I can't remember)
Dear James, How do I model a second order latent (D) that is formative construct of two first order reflective latent variables (A and B) each with 2 indicators and one first order formative latent variable (C) of 3 indicators. Will your approach apply to this situation? By modelling all 5 indicators reflective to D then make A, B, and C formative links to D?
Can you also make some videos of interpreting the data and writing for an academic paper, too? I don't know how to use it to interpret the result that i have.
I've often thought about making such a video, but then I think that we can just use existing published articles as a guide.
Sorry for the noob question- but is there a way to "check" whether it's a 2nd order formative construct? The method above seems to be that we know it is theoretically, so we designate it as formative statistically. What I want to check for is that I have these 4 first order reflective constructs but are they part of a 2nd order formative construct...Thanks!
+Val_C Usually you will find that they are strongly correlated statistically (like >0.800 correlation). This is an indication that they are part of a higher order construct.
Loadings are for reflective factors, weights are for formative factors.
what is Latent variable score(unstandardised) which is in Index Value result? Is it ok to use this instead of latent variable scores you used in this video.
My model has reflective-reflective 2nd construct and I'd like to two-step approach. But I am very confused that which latent variable scores should be used to transform multi indicators into uni-indicator.
If all factors are reflective, then you do not need to use latent variable scores.
thanks for reply. but i heard it is possible to do it in smartpls too.
Dear James, thank you for your great tutorials.
I want to combine three different latent variables (each with six reflective indicators) into one second order construct. This "new" construct, together with another latent variable then point to the endogenous variable. Can I use the same procedure as you show in this tutorial for building and measuring my model?
Johannes Gml Yes. But if the new second order factor is reflective, then you should actually be able to skip the LVS step.
Use the same approach, but no need to do the latent variable scores.
Hi James. If the second order construct has its overall items which are different with its dimensions' items. Does it still follow your steps or will be simplier>? I.e. I have information quality and its dimensions (completenss, relevance, timeliness). IQ has its own overall measures, each of dimension has itsn own measures. All of them are reflective construct at first order. but second order, IQ are formative.
+Tran Phuong If it has its own items, then it is not a 2nd order construct.