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!
Dr Gaskin, your video is awesome. Can I please ask why is that you didn't apply LVS to the second order variables (content and spirit)? Thank you very much.
Dear Dr Gaskin, Thank you so much for your awesome tutorial videos. I am using the repeated indicator approach for model consist of Reflective-Formative-Formative, (the first order construct is reflective, the second is formative and the third is formative). At the second and third order level, there are no new indicator has been added to the model. All the indicators came from survey results are used in the first dimension. As indicators, I use 51 barriers and 37 drivers derived from 85 survey questionnaire ( 5 scale likert). In other words, 88 indicators planned to be used with 85 responses. Therefore 4 indicators are eliminated at the beginning in order to be able to run the model in Smart PLS. However, when I use consistent PLS for validation of measurement model, I have to remove 64 indicators of 87 indicators in order to reach ≥0.5 AVE values for the first order. All first-order factors are reflective so consistent PLS Algortihm is used for validating the measurement model. But when I delete the items having factor loading 0.4 -0.7 according to the constructs AVE values (less than 0.5), I only have 23 indicators left. When I test the structural model with these 23 indicators P-Values of all the constructs (except 1 contruct) are calculated above 0.05. I try regular PLS Algorithm for measurement model but I also get higher P values. What do you recommend for this situation where the number of indicators are almost equal to the number of data at the beginning? Does this problem occur since I hypotized the model somehow an incorrect way? If so, How can I fix it?
As a quick question, how should we use hybrid model approach for 3rd and 4th models if no new indicators are included to the model except for the first order? In other words half of the indicators are used at the first order and the other half is used for the 2nd order but if we have 3rd and 4th order in our model, should we divide the data based on the number of dimensions we have at the model and use them?
The criteria you are using to eliminate indicators is too strict. I would recommend not removing indicators unless keeping them completely undermines the factor validity. As for your second question, you can test them altogether.
Dr Gaskin, you are saving lives of desperate PhD students like me. I've been looking for examples and ways to deal with variables without indicators or items (in my case of study illustrations). I would like to ask you please, in case the steps you suggested are also acceptable in a model where levels are informative /reflective /reflective? Also for those variables without items, theoretically can't we just suggest new items for it. Or is it better to proceed as you did in the video. Thank you🙏
If the higher order factor is formative and endogenous (being predicted by other factors that are not its dimensions), then make sure to use LVS. Otherwise, it is optional. As for factors without measures, I'm not sure what those could be. If you want to include a factor in your model, it must have measures (items) or else it cannot be predicted or do any predicting (because it has no numbers associated with it).
@@Gaskination Thank you for your feedback. I mean do you recommend, in case we can, suggesting items for the ASU from theory even though they aren't solid or proven, you think its better than applying the repeated approach?
@@Tabsit_Academy If the higher order factor has its own indicators, then it is first order, not higher order. When possible, it is usually best to measure constructs directly (rather than indirectly through multiple dimensions).
As a rough guideline, the minimum sample size in a PLS-SEM analysis should be equal to the larger of the following (10 times rule): (1) 10 times the largest number of formative indicators used to measure one construct or (2) 10 times the largest number of structural paths directed at a particular construct in the structural model. Researchers should, however, follow more elaborate recommendations such as those provided by Cohen (1992) that also take statistical power and effect sizes into account. Alternatively, researchers should run individual power analyses, using programs such as G*Power. PROF. I want an explanation in this paragraph.thanks alot
Great video, but I guess the number of items must be equal across all lower order constructs (subdimensions). This is crucial because otherwise the number of repeated items for the higher order construct are not equal. This may bias your result by the fact that lower order constructs with a higher number of items predict more variance of the higher order construct...
Hi James, this video just made it so "easy" to understand what's really happening if you build a model with second or third order latent variables. I do have a question: If I were to predict on a second-order formative variable but would also like to know the distinct effect of my IV on the first-order (reflective) variables (dimensions), could I just "eliminate" the second order and treat my first-order reflective variables as a set of independent variables? All the best
Hi Dr. Gaskin, this video is so very helpful to my thesis! Thank you so much! I do have a question: when do you add control variable in SmartPLS models when you have second-order constructs? Do you add them and obtain standardized control variable values in step one?
Hi Dr. Gaskin, thank you for the great video. I do have a question regarding the repeated indicator approach. Do we need to set all the 1st order indicators on the 2nd or 3rd order construct as formative if the the 2nd or 3rd order construct is formative? Thanks!
Hi Dr James,,,thanks so much for explaining the 3rd Order. Anyway, I seem to realize the things that you may missed after adding the new constructs of PIIT and CSE at 9:37. Note that you put respectives indicators to each of the construct but you didnt include them also in ASU like what you did with the others. So, maybe thats why its resulted in negative effect for PIIT and CSE.....again, thats my theory, not sure if its right or wrong
The reason I did not add the indicators from PIIT and CSE into ASU is because they are antecedents (predictors), rather than dimensions. In other words, they predict ASU, but they are not part of what comprises ASU. Hope this makes sense.
Dear Mr. Gaskin, I have a few questions regarding this topic. 1. My second-order construct variables also have their own indicators. So, my question is do I also put these indicators of the second-order constructs alongside the indicators from the previous dimensions, and then calculate it as the video suggests? 2. Moreover, I also have a moderating variable that moderates one of the second-order construct variables to the third-order construct, do I need to calculate this differently? or is it enough by just following your video tutorial, but only adding the moderating variable? I hope you understand my questions. Thank you for your attention.
2nd order factors should only be comprised of indicators from their dimensions. They should not have their own indicators. If they do, then they are not 2nd order factors. As for the moderator, you can do this in the 2nd stage, after computing factor scores.
@@Gaskination James Gaskin you said that 2nd order factors should only be comprised of indicators from their dimensions. They should not have their own indicators. If they do, then they are not 2nd order factors. What do we call them then if we have construct with dimensions formative indicators + also have its own formative indicators?
@@ilkersahin9387 I've never heard of such a thing. In this case, I would guess that the first order dimensions are instead predictors of the factor, rather than dimensions of it.
Thanks for your explanation, but when I check this approach I found that the f square has given wrong results, where the obtained results based on this approach above the 100,000. As I know the f square results should be between 0.02 to 0.35, even though the results may exceeding these value but should be not around a thousand. Looking forward to hearing from any experience on this issue.
Hmm... There must be a calculation error. I don't have access to the backend code for the application, so I don't know what might be going wrong. Perhaps someone on the SmartPLS forum might know.
The factor analysis was to validate the outer model for the latent factors. It is best to validate factors (such as reliability, convergent and discriminant validity) before testing structural hypotheses.
Hi James, if we had the following relation 'Construct ABC' -> CSE -> ASU -> PIIT. I have added the 'Construct ABC' myself to your model, where 'Construct ABC' is the exogenous independent variable, CSE, ASU, and PIIT are the endogenous variables and ASU is the 3rd order construct. Would we calculate the Latent Variable Score for 'Construct ABC' like we would do for CSE, ASU and PIIT even though 'Construct ABC' isn't directly connected to ASU (the 3rd order construct)?
Because ABC is exogenous only, it does not need to be reduced to a single variable (factor score). However, it can be and it shouldn't make a huge difference.
Hello Dr. Gaskin, thank you for this useful video. Could I please ask you whether the analysis process is similar for the reflective-reflective second-order model? Hope to receive your reply! Thank you!
If it is reflective all the way down, then there is no need to do the 2 stage approach. You can just model it with all the items and it will be fine. The problem only arises when you have a higher order factor that is formative and that is predicted by other factors.
@@hibaelbirou4038 The two stage approach only needs to be used when you have a higher order factor that is formative and that is predicted by other factors. Otherwise, the process is just normal without having to produce latent variable scores.
@@Gaskination Dear James do you mean if, for example, we have a 2 order construct that the first order is formative and the second order is reflective and endogeneous we do not need to produce a latent variable score? or whenever in one of the levels we have formative and predicting that multidimensional construct we should produce latent variable scores? Looking forward to your response.
Hello Dr Gaskin - I have a question for this model: if we wanted to test for mediation, with a mediator variable, what would be the steps? Step 1: validating lower measurement constructs, step 2: validating higher order constructs (such as in video); step3: latent variables scores (such as in video) and step 4: mediation (that is looking at specific indirect effects after doing bootstrapping? I have a model with second order reflective - formative factors and I wanted to check if the order of the steps is correct. Thank you very much - but for your videos, I would go bazzinga with all the stat books.
If the higher order factor is not formative, or if it is not being predicted by another factor (excluding its lower order dimensions), then you do not need to do the step with latent variable scores.
@@Gaskination Thank you for the quick reply. In my model I do have a formative-reflective factor being predicted, so your clarification was very helpful. Loved the tutorials on SmartPLS4 - really looking forward to the official release.
Dr. Gaskin - For the 3rd order model shown in Figure 1, how would you go about assessing model fit? Is there a formal test, such as SRMR, that can be applied? Can we assume good fit by an anlysis of the measurement and structural models?
IF I were to assess model fit for a formative model, and if it was higher order, I would rely probably exclusively on SRMR. You can check the smartpls site about model fit.
Hello from Germany - dear James Gaskin - and thank you for the usefull video. So did I get it right: You only need to export (14:01) the modell, when you have a 3rd order factors? I am working on my master thesis and I have talked to my Prof. I have created a second order factors - so its not necessary?
You have to export the LVS when you have a formative endogenous 2nd or 3rd ordrer factor. If the higher order factor is not endogenous (endogenous in this case means that some other factor predicts it), or if it is not formative (formative in this case means that it has multiple dimensions or other factors pointing to it as its indicators), then you can just run the model normally (without LVS).
Hello again Prof, In your video, at 8:15, where you are interpreting t-statistics, and you mention these are weights. However, in Smart PLS 3.0 report section, the second order weights do not show in outer weights, but in path coefficients. So does that mean you are interpreting the path coefficients as weights? or is there any other way to calculate second order weights?
Thank you very much for your video!. It is really helpful for my study. Could I please ask you whether the analysis process is similar for the formative-formative second-order model. I mean the first -order constructs are formative, not reflective as in your example. Hope to receive your reply!. Tks so much.
Dr Gaskin please is the extraction of these latent variable scores and the use of it in a subsequent path model only relevant when we use the repeated indicator approach or can also be used in a normal path analysis. Thank you very much
It is often used for normal path analysis to reduce complexity. However, if your model is not too complex, then just keep the original measurement model and don't use LVS. LVS is slightly less accurate.
what if path coefficients between first order construct and second order construct are not significant? I am confused about checking measurement model between construct and its dimensions. Can you please explain what to look for?
Hello Dr. Gaskin, Thanks for the super useful video! I have two question: 1. If I want to test the hierarchical structure of a construct with 3rd order factor as it consist of different dimensions which weighing is the right to use: factor or path? When should I use which one in general? 2. Can I do the testing also along a multiple group analysis? Thanks in advance!
@@Gaskination Thanks for the quick answer! I have one additional question, it would be great if you could help me one more time here! If I test the hierarchical structure of a construct and I have a treatment and a control group in my data set, do I have to test the hierarchical structure per group or can I test for both groups together? Thank you again!
Hi Dr. Gaskin. Your smart PLS videos, really saved my work. If you are available; i just need a confirmation of my steps for my study; i'm not good at statistics. I have second order effects for A( consist of A1 and A2) and B (consist of B1 and B2); first order affect for C which is mediator between A->B. For factor analysis; is it proper to do the analysis through this model? I will make mediation model control after exporting variables to excel. After these steps, I also want to do a mediation analysis on A1->B1 again C mediator effect; with the same latent variables of first model. Would it be proper to use the same latent variables or should I make a factor analysis again. As i watched i didn't need it but just wanted to be sure. Sorry if my questions are too basic; i would be appreciated if you will answer them. Thank you...
You can analyze the measurement model (factor analysis) with all these variables in the position you proposed for the structural model. As long as B is not formatively measured, then you can analyze the structural model with the full latent factors - no need to export to Excel for factor scores.
@@Gaskination Hi Dr. Gaskin, thank you for your consideration. Actually, i think i give missing information: A1 consists of 43 latent variables; A2 consists of 6 latent variables; B1 consists of 23 latent variables and B2 consists of 8 latent variables. In that case to validate my structural model; i should export my scores to excel and then create a new path model to assess second order model right? Note: I don't have a formative measurement. Thank you.
@@gizakgul You can still do it latently in SmartPLS. However, reducing it to factor scores (which is what I think you're implying when you say to export it to Excel) is one way to simplify it.
James, are the repeated indicators always reflective? It just seemed that if say Spirit is composed of two 1st order variables, the indicators added directly to Spirit would be formative (?). If you could point to any specific rules about the repeated indicators, that'd be great. Thanks!
I always treat them as reflective, although I am not positive of the reference for it. Most likely you will find the answer in the Joseph Hair book called "A primer on PLS analysis"
Dear Sir, I have a model with 2 independent variables (each have 3 constructs) and one dependent variable and 3 moderators variables. • My hypotheses are as follow : 1- To test the first orders (in this case the 6 constructs) with the dependent variable under the existence of 3 moderators variables. 2- To test the second orders (in this case the 2 constructs) with the dependent variable under the existence of 3 moderators variables. questions is : Can I do the measurement analysis , structural model analysis and multigroup analysis for the first order model separately in the first step, and then do these analysis again (measurement analysis , structural model analysis and multigroup analysis) for the second order model. Waiting for your feedback and thanks in advance for your co-operation
Dear James thanks for your video. I have a question and I appreciate it If you can please answer it...If in a two order construct the first one is formative and the second one is reflective, do we still need to produce a latent variable score? Or this latent variable score is only when the highest level construct, in this case, the second one, is formative and endogenous?
Thank you James for this video, it is a lifesaver here.. haha. I have a small problem here. 1. After loading the second-order (formative), I realized that the AVE of the second-order (formative) is below 0.5 when the lower order (reflective is above 0.5). Do I need to delete some items on both sides? 2. The R squared is above 1, something around 1.1 and 1.5, is this normal? or What do I do to normalize the situation? Thank you.
Formative factors shouldn't have AVE, as they are not required to have correlated indicators. Rsquare greater than 1 is not correct. This might be due to teh repeated indicator approach. In such a case, report the Rsquare after using latent variable scores.
Hi James, do you have to also ensure that HTMT ratios are fine. My 2nd order t-values are all good including the factor loadings, however, I have a problem with HTMT ratios from the 2nd order to the 1st order and between the 2nd order constructs.
That makes sense to have a bit of trouble between the constituent factors of a higher order factor, as these are intended to be somewhat overlapping. The real problem is when you have HTMT issues between factors that are not part of a higher order factor.
Dear Dr Gaskin, first of all, thank you for the amazing video as it helps me to understand more about HOC. Second, may I ask for your advice regarding the repeated indicator approach? As my research is about evaluating the users acceptance and satisfaction towards a developed application, so I am using Technology Acceptance Model 3 right now. Basically my first order construct consists of 6 latent variables (A, B, C, D, E and F) in which will be influencing the second order construct consisting of perceived usefulness (A, B and C) and perceived ease of use (D, E, and F) and lastly the perceived usefulness and perceived ease of use will be influencing the third order which is behavioural intention. This looks quite similar to your model excpet I have extra two latent variables at the FOC and 2nd and 3rd order constructs both have their own indicators My question is, my 2nd and 3rd order constructs both have their own indicators ( reflective), so does it mean that the repeated indicator approach is not suitable for them in this case? If it is not suitable, how should I perform the analysis of this model? Because my first order construct is reflective and second and third order is formative. And does it need to perform two-stage analysis?
In this case, PU and PEOU, as well as BIU are not higher order factors. They are just being predicted. So, no need to follow this two stage approach. HOC are constructs with lower order dimensions defining them (measurement claims), rather than predicting them (structural claims).
@@Gaskination Thank you so much for your prompt reply Dr Gaskin and I truly appreciate that. Now I have got a clearer picture. However my results are quite strange even though the measurement model has been established well construct reliability and validity, the structural model doesn't seem fine because the t statistics doesn't exceed 1.96 and p value is close to one. Some of the paths even show n/a result which confusing me. May I know what should I do to sort it out? ( I have removed a few constructs that have low loading)
@@limunng1841 If N/A is showing up, then there is some error happening. If you are using PLS_consistent, try using the regular PLS algorithm instead. I have seen lots of trouble with PLSc.
@@Gaskination Thank you for your reply Dr. Yes, I have just tried using the regular PLS and all the t-statistics and p-values showed up. Although some of the values obtained did not fulfill the criteria and I need to justify for that. Also, I just want to confirm if I apply regular PLS instead of PLS_consistent to my reflective model, is it okay with that? Since I have read many references and it is suggested that PLS-consistent should be used for reflective model.
@@limunng1841 PLS consistent is better for reflective models when it works. However, when it does not work, there is not much that can be done. In this case, if you have access to AMOS or Mplus or another covariance-based application, that might be more justifiable than using SmartPLS.
Dear Dr. James, will the process and interpretation be the same for a formative DV, formative IDV and Formative Mediating construct? Shall be grateful for your reply.
The process is the same for any endogenous 2nd order formative factor. If it is not endogenous (being predicted by other factors) and 2nd order (has first order latent dimensions) and formative, then the process is simpler (no need for using latent variable scores).
Hi Dr. Gaskin - thank you so much for this video. If I may ask a question: I have a construct, health information privacy concerns, that has been modeled as a reflective-reflective 2nd order factor in prior literature. In my study, health information privacy concerns is both endogenous and exogenous (essentially a mediator). I'm a bit confused as to what approach (repeated indicator or LVS) I should be using
If it is reflective all the way down, then you can model it without multiple stages using LVS. You only need to use LVS if you have a formative higher order factor that is in an endogenous position.
Hi Dr Gaskin, thank you so much for yr video. It really helps me a lot. I would like to ask you about my model. Fyi I have three exogenous contructs, where one is second order factor (reflective-formative), and another two is first order factors (reflective). My endogenous second order construct is (reflective-formative) and I also have one moderating factor and one mediating factor, both are first order and reflective. Can you advise which approach should I use for my model? I am quite confuse as I have combination of first order n second order factors as exogenous & also moderation/ mediating factor. Thanks Dr!
If you have a higher order factor that is formative and endogenous (being predicted by other latent factors that are not just its lower order dimensions), then you need to use a multistage approach like in the video above. In all other contexts, you can just do it in one stage without having to use latent variable scores.
Hello, Prof. James. I am just wondering why is there very little literature available on PLS use in HR/employees behavior. Most of the literature referenced the use of PLS in marketing research only. I know, small size, data normality and use of formative variables are basic reasons to use PLS, which is not specific to marketing but I think most of the social sciences research meet those basic requirements, but still very less work. Any good reason you know, why?
If we had the following relation CSE -> ASU -> PIIT, where CSE is the exogenous independent variable and PIIT is the endogenous dependent variable and ASU is the 3rd order construct , would we use the approach of one-indicator (shown till the first 18 min)?
Excellent video! You are doing a great job. Could you please help me with a question? I am searching for hours to find the correct answer. I have a formative second order endogenous construct in my model. I have first-order formative indicators linking to the formative second order endogenous variable. Can I follow the same method followed here. Any difference with arrows in second order should be considered? Any differences from what described in the video? Thanks for the time.
Doctor Gaskin reciba un cordial saludo, mi pregunta es: ¿Se podría obtener las puntuaciones en varios pasos? Por ejemplo, en tu modelo tienes un constructo de tercer orden, entonces, en el paso 1 se extraen las puntuaciones del primer orden y se reemplaza en el segundo orden, en el paso dos se obtiene las puntuaciones del constructo de segundo orden y se reemplaza en el de tercer orden. Como el constructo es de tercer orden ya no habría mas pasos que realizar. ¿Sería adecuado?. gracias su excelente explicación.
Dear James, I was wondering in the case of having a higher-order construct, how should we report R2, F2 and Q2. Do we need to report them only for the highest order or for all! I appreciate your help here.
Hello Dr. Gaskin, Thank you for the video. what happened if we have 1 exo and 1 end. variable composed as reflective formative model? end. variable is multidimensional construct with 2nd and 3rd order, meanwhile exo.variable is first and second model. In addition, the exo.variable also serves as the mediating variable. Also, for the HTMT values, when should we assess it? Is it during the first stage when we lay all the the first order constructs Thank you!
I think you'll need to email me a picture of your model and the question. You can find my email if you google me. HTMT is for assessing discriminant validity. You can assess it for all levels. The most important level to assess it is the level at which you will be testing your hypotheses.
Hi dear JAMES, ın the final model in this video can we set such a hyphothesis that : Repurpose has significant effect on Spirit formation/creation or for example: can we set a hyphothesis that Spirit is positively predicted by recombine. I try to ask that how can we hyphothesis that second order is formed-created-predicted by first order construct in FORMATIVE MODEL? Is it same with LISREL models, I just try to understand.
Usually stated hypotheses are about causal (structural) relationships, rather than measurement relationships. Since Repurpose and Recombine are dimensions of Spirit, we would not make a hypothesis about it.
Dear prof. Gaskin, I generally use PLS-SEM for my analyses. Specifically, the models I test (Human Resource Management filed) require reflective - formative higher order components. Recently, the model I tested was similar to the one presented in this video. The relevant paper was sent to a top tier journal. The editor marked the manuscript as a major revision, although he / she commented that the readers of the journal might be unfamiliar with PLS-SEM. Hence, he / she suggested to use a more familiar to readers statistical approach. Hence, my question is this. Could I run the same model with AMOS? In Human Resource Management literature, researchers tend to use the single-index approach (they simply calculate a mean score in SPSS, e.g., for the adaptive system use) and run the model in AMOS. However, it is difficult (if not impossible) to run a CFA. Do you have any ideas? Put it simply, how would you run the model in the video if PLS-SEM was not an option? Thank you in advance for your time in reading my comment :) Kind regards, Panos
Dear Dr. Gaskin I am very appreciated to your help , also i have question is it allowed to build relationships among the sub-factors of first-order or second-order factors like your example if we supposed their relation between repurposes and recombine or between content and spirit
If you do that, it would be better to remove the higher order factor for that model. Causal relationships should be drawn only between the highest order factors in the model.
If i keep it , there are any statistical problem could affect the result because i dropped the second order factor and that did not affect the relationship of IV factors .
Hi James, the video is really helpful. But how to do a moderation analysis with 2nd order construct? I followed the steps exactly shown in this video. But the "add moderation effect" is disabled for every 2nd order as well as normal constructs.
Hi,James, thanks for your videos here. Your demonstration here helps me a lot. but I still have two questions: first, my second order construct is a mediator (both being predicted and predicting), i want to confirm that is this also the situation that needs to apply the latent variable score approach. second, is there any reference talking about what is a good weight of these formative indicators/ dimensions (just like the factor loadings in reflective factors , such as factor loadings above 0.6 or 0.5 will be considered as good and acceptable). I understand that the weight of formative indicators means the importance and contributions of these indicators to the constructs, it is just my reviewers want to me to answer : based on what standard or reference can we call the formative indicators contribute substantially. thank you again for your attention
1. Yes. Anytime the 2nd order formative factor is being predicted (even if it is predicting something else), it needs to use the 2 stage approach with latent variable scores. 2. Probably Hair or Ringle (or both) have talked about good formative weights. I think Gefen and Straub talk about it as well. See some references here: statwiki.kolobkreations.com/index.php?title=References#Partial_Least_Squares
@@Gaskination Dear Prof. James, Thanks for the useful video, just need to clarify one thing, according to Becker et. al. (2012), we may use the scores of the first order constructs of the repeated indicators method in stage two of the Two stages method , but in your case you have used the scores of second order instead of using the scores of first order scores. Could you please justify it or guide me to a reference, because I am in need to follow your approach.
Hi Dr. Gaskin, thanks for the helpful video. May I ask a question, how to calculate CR and AVE for second and third order construct in reflective-reflective-reflective model. I think Marko or Hair mentioned that for second order construct, researcher should assess the reliability and validity manually. If u support this, Is manually calculation way for both second and third order variables the same or there is difference? please your advice in this matter with many thanks in advance.
The calculation at every level is the same. If for first order, use the indicator regression weights. If for 2nd order or higher, use the regression weights from that order to the lower order (2nd to 1st, or 3rd to 2nd).
@@Gaskination Thank u so much for ur reply, I have two more questions please; 1- Is it possible to use the repeated indicator approach for model consist of Reflective-Reflective-Formative, (the first order construct is reflective, the second is reflective and the third is formative). 2- As u mentioned in your video that when the second or third order construct is not endogenous, no need to use two stage approach, so could u please guide me to any book or article to use it as reference? Many thanks in advance.
@@abdohasan9140 Repeated indicators should be used whenever there is a higher order formative factor. So, in this case, yes. The logic for not needing to use the two stage approach is simply mathematical. No reference necessary. When that formative factor with repeated indicators is endogenous, then the predictors cannot predict even 1% of the variance in the factor because the repeated indicators are identical to the measures from the formative dimensions. Therefore, the formative dimensions predict 100% of the variance, leaving none for any other predictors.
hi.. thank you so much for your videos they are very helpful. Can you please distinguish between the factor and path analysis and when to use them when running the PLS Algorithm, Thank you
Factor is for when you are examining the outer (measurement) model. Path is for when you are examining the inner (structural) model. However, you'll likely find the results nearly identical except in a few weird edge cases.
@@Gaskination I have some follow-up questions regarding second-order constructs within the context of my research framework. To ensure I provide all relevant details, including my research framework and objectives, I would appreciate the opportunity to discuss this via email due to confidentiality concerns. Could you kindly share an appropriate email address for this purpose? Thank you very much for your assistance
i have just would like to discuss further the repeat indicator procedure that as reminder of Henseler. Is that correct that with reflective construct will be followed by repeat indicators of second construct and all of the indicator repeat of all indicators in primary construct. thanks
Hi Dr. Gaskin, thanks for the informative video. Apparently, I did read on Hair et al. (2017) and Becker et al. (2012) that in performing the repeated indicator approach, the number of indicators in the LOC should be equal to avoid biases. Can you assist to confirm this? Also, how to deal with HOC when there is an unequal number of indicators in the LOC? Thanks for your kind guidance!
It is true that the best approach is to have equal numbers of indicators in the first order factors. Otherwise, the factor with the most indicators will naturally explain the most variance in the HOC (which is comprised of all those repeated indicators). So, the model I have here is probably not perfect. If you do have an unequal number, you could try to trim redundant ones. Theoretically, all reflective indicators are redundant. The optimal number of indicators for a reflective construct is four.
@@unconsciouscognition7449 thanks and well noted on the response! I just got another latest reference that unequal number of indicators does not impact the result (Cheah et al., 2018). However, do you have any idea if 5th order constructs is possible with repeated idicators approach? Thanks again!
Hi Miss Nabilah . I read the same article of Becker,et.al (2012) and in the conclusion after running different types of simulations , it seems that they concluded that the unequal number of indicators did not lead to the biased result. The biased result was due to other factors.
Dear Professor, I struggled to run my model as my dependent variable is a higher-order variable. Your video helped me to solve that issue, and thanks. I got a question regarding the second-order reflective formative model. When I evaluate the validity and reliability of the reflective items, Composite reliability and AVE meet the threshold levels. But Cronbach alpha and Rho values are less than 0.7. Should I remove those items before calculating the repeated latent variable score for the formative model? Please reply. Thanks
@@fazfaz2605 This one is probably fine: Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of marketing research, 39-50.
Dr. Gaskin - I'm sorry to bother you again. My model contains a HOC made up of 4 LOCs - privacy concerns as the HOC; collection, errors, secondary use, and unauthorized access as the LOCs. Unfortunately after removing a few items based on low outer loadings, I'm finding that the AVE for one of the LOCs does not increase beyond 0.479, which is below the 0.5 threshold for AVEs. Cronbach's alpha, rho A, and Composite reliability for this LOC are all at acceptable levels - just the AVE is the problem. In fact no matter what I do the AVE will not increase beyond the 0.479 value. I'm also seeing some discriminant validity concerns via the Fornell-Larcker criterion with 2 of the LOCs. How would you suggest I handle these concerns? On another note - I have both Hair PLS books (primer and advanced) and have a few papers on how to handle reflective-reflective HOCs in SmartPLS.. However, the HOC I'm using is a bit more complex than the examples used in any of the books/papers that I have. Is there a resource you would recommend that walks the researcher step-by-step on how to handle validity and reliability testing for more complex HOCs? For example, when using the repeated indicator approach, how to handle outer loadings on the reflective-reflective HOC (do the same rules apply as when dealing with non-HOC items? Finally, do you have a video (or a recommended source) for testing for heteroscedasticity using SmartPLS? I'm so sorry for all of the questions. Thank you so much for these videos - they're a lifesaver.
1. Regarding reliability and convergent validity, AVE is a strict measure of convergent validity. Malhotra and Dash (2011) note that "AVE is a more conservative measure than CR. On the basis of CR alone, the researcher may conclude that the convergent validity of the construct is adequate, even though more than 50% of the variance is due to error.” (Malhotra and Dash, 2011, p.702). Malhotra N. K., Dash S. (2011). Marketing Research an Applied Orientation. London: Pearson Publishing. 2. If HOC is reflective all the way down, then you can evaluate it the same way you evaluate a first order reflective construct. If the HOC is formative, and using the repeated indicator approach, then follow the guidelines in this video, based on this paper: Gaskin, James; Godfrey, Stephen; and Vance, Alex (2018) "Successful System Use: It’s Not Just Who You Are, But What You Do," AIS Transactions on Human-Computer Interaction (10) 2, pp. 57-81 3. Hengky has responded to this issue here: forum.smartpls.com/viewtopic.php?t=2006
Hi Dr James. Thanks for this video. I am inspired by this video and am writing a paper on multilevel analysis using Smart PLS. Does the procedure hold same when the 2nd order is formative with one variable and 3rd order has 3 variables and so on?. I am working in social sciences. Since iit is an integrated psychological model of two theories, formative in nature I have this doubt. Thank you.
I don't think I understand the situation. You have a 2nd order factor that only has one indicator? That doesn't make sense to me. Here is a paper that uses this approach: Gaskin, James; Godfrey, Stephen; and Vance, Alex (2018) "Successful System Use: It’s Not Just Who You Are, But What You Do," AIS Transactions on Human-Computer Interaction (10) 2, pp. 57-81
If the third and second order factors are perfectly predicted, often with big T-score, and R-square value above 0.90 (just as you stated around 9:22 in the video); some of the reviewer ask me, how do I justify with such large R-square value? Are they suspicious and unlikely to be true? Is there citations that support such big R-square values? Thanks for such an awesome video, learned a lot.
Hi Prof Gaskin The significance of t values is composlory during developing 2nd order construts? Secondly, loading is like -0.553 about one form two dimensions during developing 2nd order construt then it alarming or what?
1. For a reflective factor, yes. For a formative factor, hopefully, but not compulsory. 2. A negative loading from a 1st order dimension of a 2nd order factor is not good. This means it is inversely related to the other dimensions. Check the scale direction (if it was reversed).
First of all thanks for your videos and sorry for my English :) I have a question about my model. I dont know if it is reflexive or formative. My latent variables are Innovativeness, Creativity, and Innovative Performance. Innovativeness is composed of 10 unobservable dimensions and each contains 3 observable indicators. The relationship between Innovativeness and innovative performance, I believe to be reflexive. But the relationship between the dimensions of innovation and innovation I think its formative, am I right?
You are probably right. The way to know if it is formative or reflective is whether the dimensions are interchangeable (i.e., essentially the same thing) or if they are different parts of the higher order factor. If interchangeable, then reflective, if not, then formative.
Hi Dr. Gaskin, thank you for your video. I have a question. I used the repeated indicators approach in the 1st and 2nd order factors. Nevertheless, I used a multi-actor sample so that the data for 1st and 2nd order indicators are different. I have an issue with HTMT or Fornell Larcker and VIF for the factors with repeated indicators. On the other hand, convergent validity meets the criteria. Do you have any suggestions on how I can fix the issue?
The factors with repeated indicators do not need to be discriminant with their lower order dimensions. They are expected to be highly correlated (since they use the same indicators).
@@Gaskination Thank you for your reply, Dr. Gaskin. Do you have a paper or any references that I can refer to? I have read papers related to PLS-SEM, I still have not found the statement of discriminant validity issue can be ignored when we use the same indicators in the model. I need this for supported my data analysis in the thesis.
@@Abd-mv1mr Probably the reason you have not found such a statement is because it is mathematically apparent. There is no need to declare it because it is self-evident. I don't think you need any reference for this. You can simply state the obvious that repeated indicators imply significant overlap in shared variance.
Thanks for the video, it's very helpful. But there's one thing I wanna ask. What if the indicator that you put in second stage or first stage latent variable has loading factor value below 0.7 but those indicators are resulting above 0.7 in 3rd stage latent variable? Thanks
For reflective factors, loadings below 0.700 are fine, as long as the AVE or CR are above recommended thresholds. For formative factors, there is no minimum indicator weight.
@@GaskinationThank you Dr. James. I want to ask again. What if the first order construct has loading score less than 0.7 towards the 2nd order construct? Do I have to delete that 1st order construct?
@@ayunsaskhiyah7044 The same rules apply here, but we can be somewhat more lenient since it is at the 2nd order. So, unless the loading is truly low (e.g.,
I am evaluating a third-order model just like this one but it stops at ASU level, where first-order indicators are reflective, second and third-order are formative repeated indicators. I have done individual analysis but now I want to evaluate the impact of first-order factors on third-order through the second-order factors for two different product categories, how should I do the comparison in SmartPLS3? (2nd order factors will act as moderators) MGA doesn't seem appropriate for this or maybe I don't know how to use it. Variables in the model are the same and the number of data sets for both the categories is the same as well 449. Can you please guide.
If the third order is repeated indicators, then it will be useless to test the effect of first order on third order since the third order is comprised of the first order indicators already. This is not a mediation (or any kind of structural relationship). It is a measurement relationship.
Hi Dr. Gaskin, your videos are very helpful, I have a question, I am getting half of variables' HTMT criterion (discriminant validity) greater than .90. and Cronbach's Alpha smaller than .70. however composite reliability , average variance extracted (AVE), Fornell-Larker Criterion are good. in this case how do I establish discriminant validity ? please help!!! TIA.
If AVE, CR, and FL criterion is fine, then this is good enough. SEM is not about optimizing and satisfying all criteria, but instead about using validation criteria to show that your model and variables are probably sufficiently valid to place confidence in the results.
Here are two excellent sources for this question: forum.smartpls.com/viewtopic.php?t=16249 www.smartpls.com/documentation/algorithms-and-techniques/consistent-pls
Thank for the video, it is very useful and exact what i need. However, i am confused between consistent Bootstrap & normal Bootstrap. I have taken workshop from SmartPls and they were also running the model either reflective or formative with normal PLS/Bootstrap method. Now, after looking at this video i am not sure shall I use consistent bootstrapping even for normal reflective model? or the normal bootstrapping/PLS. For formative also. Your reply would be really helpful.
Consistent is only for models with all reflective factors. The reason they didn't use it during the workshop is probably because it is occasionally unable to produce estimates.
Hi Dr. James, Really need your help: I've built a 2nd order reflective factor but I'm wondering what is the meaning of the path coefficient from the 1st to the 2nd order? In my case, the 1st order is Consumer Brand Engagement that is reflected in Cognitive (0,55), Affectional (0,95), and Activation (0,8) dimension. What do the values mean? That Cognitive dimension only impacts 55% of a variance in Consumer Brand Engagement? Thank you so much for the help, Diogo
The estimates between 2nd order factors and their reflective first order dimensions are the extent to which that dimension is correlated with the other first order dimensions.
@@Gaskination thank you so much. So there is no way to find out how each dimensions affects the main construct (Consumer Brand Engagement)? I have another short question: is it possible that a direct path coefficient is -0.458 (with significance) but indirectly is positive? Kind regards, Diogo
@@diogodorey4486 In a reflective design, the indicators don't affect the factor. They are a manifestation of the factor. As for the indirect question, yes, that is perfectly logical. This is what is called competitive mediation. You can read more about it here: Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of consumer research, 37(2), 197-206.
Hi Mr Gaskin, first of all, many thanks for all your great videos - including this one! They are a great help! :) Regarding the models i would have two questions. Maybe i have missed it, but do I actually also have to correct the the second order construct's outer loadings? Especially when they are all ok (>0,7) for the first order constructs? In my model, the loadings of all my first order construct items look great, but on the second order they are mostly below 0,7. My second question would be: Do i also have to do the two stage approach when i am using my second order construct only as a moderator and/or exogenous variable? Thanks to a lot in advance to everyone who could help me here! BR, Konstantin
The loadings between the 1st order dimensions and their 2nd order factor are to be evaluated in the same way as the loadings from the observed indicators. So, if formative, then they should be significant. If reflective, they should be above 0.700 on average. Often scholars are more lenient with these loadings though, since they are more loosely reflective than a unidimensional factor. Also, you only have to use the 2 stage approach when the 2nd order factor is formative and endogenous. Otherwise, there is no need.
Dear prof thanks for this video. I have a question. In second order reflective-reflective mearment model, CR value between .60 to .70 is acceptable? I have just two dimensions. If yes than can you share any citation with me? Thanks.
Yes, when there are only two estimates used in the calculation of the CR value, then less than 0.7 is acceptable. This is stated in Joseph Hair's book: Hair, J. F., Anderson, R. E., Babin, B. J., and Black, W. C. 2010. Multivariate Data Analysis. Upper Saddle River, NJ: Pearson.
Hello Boss, thank you for this video. I have a question. Does this second order means that the middle two variables are formative in nature and model is mixed model? please reply
@@Gaskination I am bit confused, because according to my understanding , one variable cannot be formative and reflective at the same time. For ASU Variable: Spirit and Content are Dimensions right? How come one variable can be reflective and formative at the same time? Need a bit explanation kindly.
@@dr.mudassir ASU is formative, but the repeated indicators are set reflectively. They are only repeated indicators though. So, it is a formative factor.
@@Gaskination Thank you for your explanation. One last question, can a higher model variables be reflective or all higher model variables are considered formative? i can see some videos that put directions outside instead inside.
hello sir, i am following the same pattern of second order formative construct model in SmartPLS 2. however, for using the results for further calculations, i need to use the latent variable scores from PLS and convert them to SPSS for calculation purposes like for multicollinearity. so, which of the latent variable scores will be used for further testing in SPSS or any other software? (i mean the first order model's latent variables or the second order model's latent variables?) Please help, i am stuck up.
You can use both or either. They are all valid scores. I would recommend using the 2nd order when available, and 1st order when they don't belong to a 2nd order factor.
Hi Dr. Gaskin, thank you for your video, it's really helpful. I have a question, if all my second order constructs in the research model are reflective-reflective, is repeated indicator approach appropriate in this case? And what will happen if the two-stage approach was used in the reflective-reflective second order construct? Thank you.
Repeated indicator approach is not necessary unless the 2nd order factor is both formative and endogenous. In the case of reflective-reflective, it is unnecessary. If you do it anyway (perhaps to simplify a model), there is little harm. The model should be roughly equivalent.
James in your example if we were to attach some new reflective indicators linked directly to the latent variable ASU what would happen to the carried over indicators from the first order level added to the second order latent variables Spirit and Content. Do they still get carried to the third order ASU or not.
@@Gaskination HI James, Ok thanks for your reply, for the latent factor are you referring to the Blue Circles in Smart PLS. If so would I add a new latent variable called "New1" and connect it to ASU. The indicator would then connect to New1. Is that correct. Do you have an example showing this arrangement you could share with me.
Hi James, I have a 3-construct formative model and there is a comment from the reviewer saying that "The dimensions related to XXX service include many aspects of E-service process, so that the authors should consider second-CFA approach.". 1) In fact that "E-service process" is a formative construct with several items named as E-quality, should I understand as that the reviewer wants me to change that the E-quality to reflective on the 2nd order and run the results of "PLS outer loading, CR Cronbach alpha & AVE"? 2) Still I can use SmartPLS for such analysis? 3) Is the clip here is the best for my case or any other (like "SmartPLS Formative 2nd order Constructs" of yours) you would recommend? Thanks!
1. Yes, that's what it sounds like. 2. Yes, SmartPLS can be used for 2nd order factors, as shown in the video above. If you are only having reflective factors, you could use AMOS instead to measure model fit. 3. This one above should be fine.
@@Gaskination Thanks a lot, James! So, should I run the whole model including both 2 formative constructs and one changed reflective construct for the result? Or just the individual model of that reflective construct? Because the CFA includes Chi-square, RMSEA, CFI, NFI also, do I need to report those or just "PLS outer loading, CR Cronbach alpha & AVE" are enough?
@@YY-vy2bq I would guess you want to run the whole model with all factors or else you won't be able to effectively test your theory. If you still have formative factors, then you'll need to use SmartPLS.
@@Gaskination Thanks James! May I know should I report "Chi-square, RMSEA, CFI, NFI" also, or just "PLS outer loading, CR Cronbach alpha & AVE" are enough for that model with one reflective constructs while the rest constructs are formative?
@@YY-vy2bq SmartPLS does not generate many model fit metrics. I think it only generates SRMR and a couple of others: ruclips.net/video/Lp3r5F-eVCs/видео.html
Hi Mr Gaskin, I have a second-order formative construct in my paper, does it necessary CFA calculating for that construct? I mean, in the part of data analysis, factor loads and fit indices have been reported but I did not enter formative construct into the CFA calculation. In other words, I didn't mention that in measurement model, and now I am going to calculate CFA for formative construct separately, does it accurate ? I hope that you got my words and help me to improve my paper. thanks alot.
Can you help me by telling that, from above video in second order. its formative model right/? because i have same model. Variable E-HRM, sub variable 5, and these 5 sub variables have items 3 each.
I use SmartPLS 3 using student license. Unfortunately It doesn't support copy function in Excel format, even Ctrl+c function as well. In this case, how can I copy LVS?
Depends on the DV. If it is a perceptual variable like satisfaction or behavioral intention, then it should be a higher R-squared, like 30% or more. If it is an observed variable like revenue growth or price-per-share, then it is fine to be lower, even as low as 2%.
Dear Dr.@@Gaskination , Can you pls provide us with a reference to study more about R2 value. I heard the threshold must be more than 1 for human behaviour related studies. One of my models got something like 24%, with high t statistics. Reviewers had no questions about R square. Pls recommend a reading on R2. Thank you in advance.
@@tbasou R2 must be less than 1.00 because it is a percent of explained variance in the endogenous variable. R2 increases as you add more predictors. In the social sciences, it is best to have R2 above about 20%, but if it was as high as 70%, we would be suspicious that there might be a tautological relationship in the model.
@@tbasou Yes, because it is a percent, it must be less than (or equal to) 1.00 because that would be 100%. You cannot explain more than 100% of the variance in an outcome variable.
So we don't need to check AVE in 2nd order reflective-formative model? Only reliability and discriminant validity? And what about the threshold for removing factor loadings?
AVE is relevant for all reflective factors (regardless of 1st or 2nd order). Reliability is also only applicable to reflective factors. Discriminant validity is applicable to all. As for removing items, if they are on reflective factors, then you can remove them if they are less than 0.500 (if you still have at least two items). For formative factors, it is hard to justify removing an item unless the VIF is really high.
@@Gaskination alright, thank you so much, I wasn't expecting such a prompt response. I got confused because PLS itself higglight all the factor loadings below .70, however you only removed the negative. Everything clear now. Thanks once again. You are doing a great job of facilitating beginners.😊😊
@@Gaskination what if the first-order construct is formative, do I need to remove those indicators with high VIF from the first, second and third orders?
Hi, I think your video is not yet completed ..... I remember you said earlier that R square should not be more than "1'; yet after all the explanation, your result shows that the some of the R square is still more than "1" ...... please elaborate more....... thank you in advance!!
The R-square in a structural model should not exceed 1.00. However, the R-square in the measurement model shown is not relevant. You are correct that it appears I did not finish this video to proceed to the structural model. The next steps would be to extract the latent variable scores and then create a new model with those. We would then find that the R-squares for those variables are less than 1.00
@@Gaskination Thank you for your reply, Mr. Gaskin. Yes, I saw your other videos explaining the next step for the 2nd order (higher order); both the one with SmartPLS v.3 and Smart PLS v.2 - they help my project a lot. Thank you for sharing and I hope you keep making videos about Smart PLS with more details and depth ....... , cheers!!
This is the most popular reference for it: Jarvis, C. B., MacKenzie, S. B., & Podsakoff, P. M. (2003). A critical review of construct indicators and measurement model misspecification in marketing and consumer research. Journal of consumer research, 30(2), 199-218. Here is a video to explain it a bit: ruclips.net/video/gw0xvvJw-AM/видео.html
Because this is a measurement model, rather than a causal model, the predictive measures (R2) are not relevant. When we move on to the causal model, the R2 becomes normal again.
@@Gaskination Hi, and thank you for all the great videos. In my model the 2nd order construct is exogenous, meaning it is not being predicted. However, like in your video, the R2 values are >1 when running the PLS algorithm using the "path" weighting scheme. How can I achieve R2
@@MeineUndDeineOma This is fine and normal. The R-squared for an exogenous construct is irrelevant. It should be 1.00 because it is being "predicted" by its own indicators.
@@Gaskination Thank you for your quick reply, this is awesome! Since I modeled a 2nd order construct using the reflective formative approach I assume I have to use the simple PLS algorithm? The results when running the Consistent PLS algorithm deviate a lot, especially R2 and significance (t-values).
I did not. Jorg and I could never work out a time to discuss it. However, after discussing it with other PLS experts, we determined that the results are very similar using this method.
I understand the example used in Chp 5 of A primer on partial least squares structural equation modeling. Does it mean we cannot do a test on convergent validity if we dont have a single item global measure that is reflective for our formative construct. Say my formative construct is safety concern. However this is no single item global measure for safety concern available in the literature (in fact safety perception is a scale I have developed, lets say) then what do I do?
Correct. If you are using a formative scale, then you cannot check its reliability or convergent validity unless you have a target global measure for that construct. In such a case, you'll need to rely on basic formative validation, such as shown here: ruclips.net/video/Qx_7pN7eAes/видео.html
what about the model fit after run your model? the SRMR, NFI, and Chi-square value when we run a double layer latent construct? it does not give the SRMR, NFI, and Chi-square values. Please help.
Model fit and PLS are not really compatible. I would not worry about model fit when running a PLS model. If you want model fit, switch to CB-SEM model (now available in SmartPLS as well).
hi dr. Gaskin, i have few questions regarding assessing higher order constructs. so i have reflective formative type model, LOC consists 4 LV and the HOC is predicting another LV & not endogenous, what should i do to validate this model? im also using the repeated indicator approach & Since few of my variables on the LOC has low AVE, composite reliability and cronbachs alpha, i had to eliminate some of the indicators so that i can meets the criteria. I did follow your guide on when you eliminate an indicator from the LOC, you also have to remove the one in the HOC and vice versa. After i eliminate some of the indiators all of my LOC has already meet the AVE minimum value, but somehow my HOC still has a very low AVE value (0.28) So my question is do i have to keep eliminating the indicators until my HOC's AVE value acceptable (but then id lost quite a lot of indicators on LOC too)? Or am i doing it wrong? Is there any specific guides on validating this type of model? Thankn you!
1. As long as the HOC is not endogenous, then you don't need to do the two stage approach. 2. Correct, remove from both. 3. Don't worry about the HOC AVE. Instead look at the loadings from the LOCs to the HOC. They should be significant and moderately strong (ideally >0.500) if the HOC is reflective. If the HOC is formative, then AVE is completely irrelevant for that construct.
@@Gaskination thank you so much for answering my questions! yes, the HOC is formative. last question, so after i'm done evaluating the reflective measurement model, what criteria do i need to fulfill to validate the formative HOC? or i could just run the hypothesis test right away? thanks again, Dr.
Hi James, I am currently working on exogenous second order construct with four dimensions having unbalanced number of items, pointing towards that exogenous which is linking towards dependent variable i tried to conduct two stage approach taking those dimensions as items formative towards second order construct but what I get is rho_A=1 only and nothing else furthermore p value of second order is 0.000 however if I try to go with two stage approach taking first order items as reflective from second order results are quite okay. What one is better? Any theoretical suggestions? Kindly guide me
Hi, this video have really helped me build my study model. But my model is a second order one that connects each of the four 1st order variables to one 2nd order variable, then all the five variables are connected to one endogenous variable. Attempt to run bootstrapping kept returning "Singular Matrix Problem. Please, how do I resolve this?
Hello Professor, I am studying mediation effect of ambidexterity. Ambidexterity is the interaction of explorative (3 items) and exploitative (2 items) learnings. I am unable to create an interaction effect of both learnings where the product will be a mediator. If I do it the conventional way, that will moderated mediation or mediated moderation. I am using SmartPLS 3. Your guidance will be appreciated. Thank you
You want to create an interaction and use that interaction in the mediator position? I’ve never considered (or heard of) this. If we had to label it, I would call it mediating moderation (not mediated moderation).
@@Gaskination Yes, I am studying mediation effect of ambidexterity which is the interaction of exploitative and explorative learning. I searched but couldn't find a solution. Analyses can be performed through Amos but that is too cumbersome that's why I want to do it through SmartPLS. Another method is to make it a higher-order construct and I will proceed with that. Thank you for considering my query.
good evening, by conducting an analysis by SMART PLS, on a trial version of 30 days, on a sample of 40 students, and knowing that my model and third degree and the response rate was almost in favor, I have not could get good results what is the problem. thank you
Dr. james Gaskin , warm greetings from Jordan please i have question about hierarchical component model can the higher-order components play role of a mediating between its lower-order components and the dependent construct? I mean if we hypothesis that X as high-order component mediate the relation with X1(low-order component) and Y as a dependent construct. please, if you have reference for this point
Theoretically, this doesn't make any sense. The higher order component represents its lower order dimensions, so it can't mediate their effects. It is comprised of the same indicators.
James Gaskin thanks for your reply Then can I state the hypothesis from low-order to dependent construct without to mention the higher-order factor directly for example in your tutorial if I said trying new features affecting adaptive system use (I am saying like this because I want to know is it belongs to the higher-order factor or not and at the same time I want to know is that low-order significant to the dependent construct or not? Simply , I want to know which of low-order factors significant to the dependent factor ?
If you want to know the direct effects from lower order dimensions to DV, then remove the higher order factor for this model. You can always put the higher order factor back in for another model.
I wanted to run my pilot results for 30 sample size but it rejected it saying sample size is too small. It should be at least 36. Any advise you can give me on this?
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!
thanks a lot prof.
OMG! I have this kind of model and your video is very helpful, thank you so sooo much!
Dr Gaskin, your video is awesome. Can I please ask why is that you didn't apply LVS to the second order variables (content and spirit)? Thank you very much.
Because they were already captured in the 3rd order factor (ASU). Including them in the second stage model would have been redundant.
Dear Dr Gaskin,
Thank you so much for your awesome tutorial videos. I am using the repeated indicator approach for model consist of Reflective-Formative-Formative, (the first order construct is reflective, the second is formative and the third is formative). At the second and third order level, there are no new indicator has been added to the model. All the indicators came from survey results are used in the first dimension. As indicators, I use 51 barriers and 37 drivers derived from 85 survey questionnaire ( 5 scale likert). In other words, 88 indicators planned to be used with 85 responses. Therefore 4 indicators are eliminated at the beginning in order to be able to run the model in Smart PLS.
However, when I use consistent PLS for validation of measurement model, I have to remove 64 indicators of 87 indicators in order to reach ≥0.5 AVE values for the first order. All first-order factors are reflective so consistent PLS Algortihm is used for validating the measurement model. But when I delete the items having factor loading 0.4 -0.7 according to the constructs AVE values (less than 0.5), I only have 23 indicators left. When I test the structural model with these 23 indicators P-Values of all the constructs (except 1 contruct) are calculated above 0.05. I try regular PLS Algorithm for measurement model but I also get higher P values. What do you recommend for this situation where the number of indicators are almost equal to the number of data at the beginning? Does this problem occur since I hypotized the model somehow an incorrect way? If so, How can I fix it?
As a quick question, how should we use hybrid model approach for 3rd and 4th models if no new indicators are included to the model except for the first order? In other words half of the indicators are used at the first order and the other half is used for the 2nd order but if we have 3rd and 4th order in our model, should we divide the data based on the number of dimensions we have at the model and use them?
The criteria you are using to eliminate indicators is too strict. I would recommend not removing indicators unless keeping them completely undermines the factor validity. As for your second question, you can test them altogether.
Dr Gaskin, you are saving lives of desperate PhD students like me. I've been looking for examples and ways to deal with variables without indicators or items (in my case of study illustrations). I would like to ask you please, in case the steps you suggested are also acceptable in a model where levels are informative /reflective /reflective?
Also for those variables without items, theoretically can't we just suggest new items for it. Or is it better to proceed as you did in the video.
Thank you🙏
If the higher order factor is formative and endogenous (being predicted by other factors that are not its dimensions), then make sure to use LVS. Otherwise, it is optional. As for factors without measures, I'm not sure what those could be. If you want to include a factor in your model, it must have measures (items) or else it cannot be predicted or do any predicting (because it has no numbers associated with it).
@@Gaskination Thank you for your feedback. I mean do you recommend, in case we can, suggesting items for the ASU from theory even though they aren't solid or proven, you think its better than applying the repeated approach?
@@Tabsit_Academy If the higher order factor has its own indicators, then it is first order, not higher order. When possible, it is usually best to measure constructs directly (rather than indirectly through multiple dimensions).
As a rough guideline, the minimum sample size in a PLS-SEM analysis should be equal to the larger of the following
(10 times rule): (1) 10 times the largest number of formative indicators used to measure one construct or (2) 10 times
the largest number of structural paths directed at a particular construct in the structural model. Researchers should,
however, follow more elaborate recommendations such as those provided by Cohen (1992) that also take statistical
power and effect sizes into account. Alternatively, researchers should run individual power analyses, using programs
such as G*Power.
PROF. I want an explanation in this paragraph.thanks alot
Thank you for this video, it clarifies many doubts
Great video, but I guess the number of items must be equal across all lower order constructs (subdimensions). This is crucial because otherwise the number of repeated items for the higher order construct are not equal. This may bias your result by the fact that lower order constructs with a higher number of items predict more variance of the higher order construct...
Yes, I have read that somewhere, although I can't remember where.
Hi James, this video just made it so "easy" to understand what's really happening if you build a model with second or third order latent variables. I do have a question: If I were to predict on a second-order formative variable but would also like to know the distinct effect of my IV on the first-order (reflective) variables (dimensions), could I just "eliminate" the second order and treat my first-order reflective variables as a set of independent variables?
All the best
Correct. You would just test an alternative model without the 2nd order factor.
Hi James, I hv questions
1) Why you used PLSc instead of PLS algorithm (minute 5.41)
2) Why you chose “factor” not “path” (minute 5.50)
1. Because my first order factors were all reflective.
2. Because I was doing a factor analysis, rather than a path analysis.
Hi Dr. Gaskin, this video is so very helpful to my thesis! Thank you so much!
I do have a question: when do you add control variable in SmartPLS models when you have second-order constructs? Do you add them and obtain standardized control variable values in step one?
Hi Dr. Gaskin, thank you for the great video. I do have a question regarding the repeated indicator approach. Do we need to set all the 1st order indicators on the 2nd or 3rd order construct as formative if the the 2nd or 3rd order construct is formative? Thanks!
No. Set the indicators as reflective on the 2nd order.
thanks boss...finally done my manuscript...
Excellent video!
Hi Dr James,,,thanks so much for explaining the 3rd Order. Anyway, I seem to realize the things that you may missed after adding the new constructs of PIIT and CSE at 9:37. Note that you put respectives indicators to each of the construct but you didnt include them also in ASU like what you did with the others. So, maybe thats why its resulted in negative effect for PIIT and CSE.....again, thats my theory, not sure if its right or wrong
The reason I did not add the indicators from PIIT and CSE into ASU is because they are antecedents (predictors), rather than dimensions. In other words, they predict ASU, but they are not part of what comprises ASU. Hope this makes sense.
Thank you so much! Your video is so helpful!
This is the thing i ve needed more than Middle East needs democracy
u saved my life.. omg thank youuuu amillion
Dear Mr. Gaskin, I have a few questions regarding this topic.
1. My second-order construct variables also have their own indicators. So, my question is do I also put these indicators of the second-order constructs alongside the indicators from the previous dimensions, and then calculate it as the video suggests?
2. Moreover, I also have a moderating variable that moderates one of the second-order construct variables to the third-order construct, do I need to calculate this differently? or is it enough by just following your video tutorial, but only adding the moderating variable?
I hope you understand my questions.
Thank you for your attention.
2nd order factors should only be comprised of indicators from their dimensions. They should not have their own indicators. If they do, then they are not 2nd order factors. As for the moderator, you can do this in the 2nd stage, after computing factor scores.
@@Gaskination
James Gaskin
you said that 2nd order factors should only be comprised of indicators from their dimensions. They should not have their own indicators. If they do, then they are not 2nd order factors. What do we call them then if we have construct with dimensions formative indicators + also have its own formative indicators?
@@ilkersahin9387 I've never heard of such a thing. In this case, I would guess that the first order dimensions are instead predictors of the factor, rather than dimensions of it.
Thanks for your explanation, but when I check this approach I found that the f square has given wrong results, where the obtained results based on this approach above the 100,000. As I know the f square results should be between 0.02 to 0.35, even though the results may exceeding these value but should be not around a thousand.
Looking forward to hearing from any experience on this issue.
Hmm... There must be a calculation error. I don't have access to the backend code for the application, so I don't know what might be going wrong. Perhaps someone on the SmartPLS forum might know.
Thanks a lot James for the video. The reason you used factor analysis is because of having 2nd & 3rd orders?
The factor analysis was to validate the outer model for the latent factors. It is best to validate factors (such as reliability, convergent and discriminant validity) before testing structural hypotheses.
Ahan, thanks 😊
Hi James, if we had the following relation 'Construct ABC' -> CSE -> ASU -> PIIT. I have added the 'Construct ABC' myself to your model, where 'Construct ABC' is the exogenous independent variable, CSE, ASU, and PIIT are the endogenous variables and ASU is the 3rd order construct. Would we calculate the Latent Variable Score for 'Construct ABC' like we would do for CSE, ASU and PIIT even though 'Construct ABC' isn't directly connected to ASU (the 3rd order construct)?
Because ABC is exogenous only, it does not need to be reduced to a single variable (factor score). However, it can be and it shouldn't make a huge difference.
Hello Dr. Gaskin, thank you for this useful video.
Could I please ask you whether the analysis process is similar for the reflective-reflective second-order model?
Hope to receive your reply!
Thank you!
If it is reflective all the way down, then there is no need to do the 2 stage approach. You can just model it with all the items and it will be fine. The problem only arises when you have a higher order factor that is formative and that is predicted by other factors.
@@Gaskination please Dr. Gaskin ! what is the analysis process when we had formative -reflective second-order model?
@@hibaelbirou4038 The two stage approach only needs to be used when you have a higher order factor that is formative and that is predicted by other factors. Otherwise, the process is just normal without having to produce latent variable scores.
@@Gaskination Dear James do you mean if, for example, we have a 2 order construct that the first order is formative and the second order is reflective and endogeneous we do not need to produce a latent variable score? or whenever in one of the levels we have formative and predicting that multidimensional construct we should produce latent variable scores? Looking forward to your response.
one more question, how to run the above-explained model, if the first order was not reflective, I mean if all were formative?
Hello Dr Gaskin - I have a question for this model: if we wanted to test for mediation, with a mediator variable, what would be the steps? Step 1: validating lower measurement constructs, step 2: validating higher order constructs (such as in video); step3: latent variables scores (such as in video) and step 4: mediation (that is looking at specific indirect effects after doing bootstrapping? I have a model with second order reflective - formative factors and I wanted to check if the order of the steps is correct. Thank you very much - but for your videos, I would go bazzinga with all the stat books.
If the higher order factor is not formative, or if it is not being predicted by another factor (excluding its lower order dimensions), then you do not need to do the step with latent variable scores.
@@Gaskination Thank you for the quick reply. In my model I do have a formative-reflective factor being predicted, so your clarification was very helpful. Loved the tutorials on SmartPLS4 - really looking forward to the official release.
Dr. Gaskin - For the 3rd order model shown in Figure 1, how would you go about assessing model fit? Is there a formal test, such as SRMR, that can be applied? Can we assume good fit by an anlysis of the measurement and structural models?
IF I were to assess model fit for a formative model, and if it was higher order, I would rely probably exclusively on SRMR. You can check the smartpls site about model fit.
Hello from Germany - dear James Gaskin - and thank you for the usefull video. So did I get it right: You only need to export (14:01) the modell, when you have a 3rd order factors? I am working on my master thesis and I have talked to my Prof. I have created a second order factors - so its not
necessary?
You have to export the LVS when you have a formative endogenous 2nd or 3rd ordrer factor. If the higher order factor is not endogenous (endogenous in this case means that some other factor predicts it), or if it is not formative (formative in this case means that it has multiple dimensions or other factors pointing to it as its indicators), then you can just run the model normally (without LVS).
Hello again Prof,
In your video, at 8:15, where you are interpreting t-statistics, and you mention these are weights. However, in Smart PLS 3.0 report section, the second order weights do not show in outer weights, but in path coefficients. So does that mean you are interpreting the path coefficients as weights? or is there any other way to calculate second order weights?
Yes. The regression line between 2nd and 1st order factors can be interpreted as a weight, even though the output treats it like a path.
Hi dr.Gaskin
your video is very helpful
please share more new training about pls
Here is a full playslist about SmartPLS 3: ruclips.net/p/PLnMJlbz3sefKTL7KGy_JIYTSpFXizxW1X
Thank you very much for your video!. It is really helpful for my study. Could I please ask you whether the analysis process is similar for the formative-formative second-order model. I mean the first -order constructs are formative, not reflective as in your example. Hope to receive your reply!. Tks so much.
Sorry. I was logged into one of my other accounts (udreamed).
Hi Ngoc. Just checking to see if you found the right way for formative-formative model?
Thanks for Awesome Sharing
Dr Gaskin please is the extraction of these latent variable scores and the use of it in a subsequent path model only relevant when we use the repeated indicator approach or can also be used in a normal path analysis.
Thank you very much
It is often used for normal path analysis to reduce complexity. However, if your model is not too complex, then just keep the original measurement model and don't use LVS. LVS is slightly less accurate.
what if path coefficients between first order construct and second order construct are not significant? I am confused about checking measurement model between construct and its dimensions. Can you please explain what to look for?
Those paths should definitely be significant, or else it is hard to justify the first order dimensions are truly part of this higher order factor.
Hello Dr. Gaskin, Thanks for the super useful video! I have two question:
1. If I want to test the hierarchical structure of a construct with 3rd order factor as it consist of different dimensions which weighing is the right to use: factor or path? When should I use which one in general? 2. Can I do the testing also along a multiple group analysis? Thanks in advance!
1. Use factor for validating the factors, and path for validating the paths. (the results will probably be nearly identical though)
2. Yes.
@@Gaskination Thanks for the quick answer! I have one additional question, it would be great if you could help me one more time here! If I test the hierarchical structure of a construct and I have a treatment and a control group in my data set, do I have to test the hierarchical structure per group or can I test for both groups together? Thank you again!
Hi Dr. Gaskin. Your smart PLS videos, really saved my work. If you are available; i just need a confirmation of my steps for my study; i'm not good at statistics. I have second order effects for A( consist of A1 and A2) and B (consist of B1 and B2); first order affect for C which is mediator between A->B. For factor analysis; is it proper to do the analysis through this model? I will make mediation model control after exporting variables to excel. After these steps, I also want to do a mediation analysis on A1->B1 again C mediator effect; with the same latent variables of first model. Would it be proper to use the same latent variables or should I make a factor analysis again. As i watched i didn't need it but just wanted to be sure. Sorry if my questions are too basic; i would be appreciated if you will answer them. Thank you...
You can analyze the measurement model (factor analysis) with all these variables in the position you proposed for the structural model. As long as B is not formatively measured, then you can analyze the structural model with the full latent factors - no need to export to Excel for factor scores.
@@Gaskination Hi Dr. Gaskin, thank you for your consideration. Actually, i think i give missing information: A1 consists of 43 latent variables; A2 consists of 6 latent variables; B1 consists of 23 latent variables and B2 consists of 8 latent variables. In that case to validate my structural model; i should export my scores to excel and then create a new path model to assess second order model right? Note: I don't have a formative measurement. Thank you.
@@gizakgul You can still do it latently in SmartPLS. However, reducing it to factor scores (which is what I think you're implying when you say to export it to Excel) is one way to simplify it.
James, are the repeated indicators always reflective? It just seemed that if say Spirit is composed of two 1st order variables, the indicators added directly to Spirit would be formative (?). If you could point to any specific rules about the repeated indicators, that'd be great. Thanks!
I always treat them as reflective, although I am not positive of the reference for it. Most likely you will find the answer in the Joseph Hair book called "A primer on PLS analysis"
Thanks!
Dear Sir,
I have a model with 2 independent variables (each have 3 constructs) and one dependent variable and 3 moderators variables.
• My hypotheses are as follow :
1- To test the first orders (in this case the 6 constructs) with the dependent variable under the existence of 3 moderators variables.
2- To test the second orders (in this case the 2 constructs) with the dependent variable under the existence of 3 moderators variables.
questions is :
Can I do the measurement analysis , structural model analysis and multigroup analysis for the first order model separately in the first step, and then do these analysis again (measurement analysis , structural model analysis and multigroup analysis) for the second order model.
Waiting for your feedback and thanks in advance for your co-operation
Yes, that is a fine approach.
Dear James thanks for your video. I have a question and I appreciate it If you can please answer it...If in a two order construct the first one is formative and the second one is reflective, do we still need to produce a latent variable score?
Or this latent variable score is only when the highest level construct, in this case, the second one, is formative and endogenous?
Correct. LVS are only needed when the highest order factors are formative and endogenous.
@@Gaskination thank you so much...I really appreciate that.
Thank you James for this video, it is a lifesaver here.. haha. I have a small problem here. 1. After loading the second-order (formative), I realized that the AVE of the second-order (formative) is below 0.5 when the lower order (reflective is above 0.5). Do I need to delete some items on both sides? 2. The R squared is above 1, something around 1.1 and 1.5, is this normal? or What do I do to normalize the situation? Thank you.
Formative factors shouldn't have AVE, as they are not required to have correlated indicators. Rsquare greater than 1 is not correct. This might be due to teh repeated indicator approach. In such a case, report the Rsquare after using latent variable scores.
Hi James, do you have to also ensure that HTMT ratios are fine. My 2nd order t-values are all good including the factor loadings, however, I have a problem with HTMT ratios from the 2nd order to the 1st order and between the 2nd order constructs.
That makes sense to have a bit of trouble between the constituent factors of a higher order factor, as these are intended to be somewhat overlapping. The real problem is when you have HTMT issues between factors that are not part of a higher order factor.
Dear Dr Gaskin, first of all, thank you for the amazing video as it helps me to understand more about HOC. Second, may I ask for your advice regarding the repeated indicator approach? As my research is about evaluating the users acceptance and satisfaction towards a developed application, so I am using Technology Acceptance Model 3 right now. Basically my first order construct consists of 6 latent variables (A, B, C, D, E and F) in which will be influencing the second order construct consisting of perceived usefulness (A, B and C) and perceived ease of use (D, E, and F) and lastly the perceived usefulness and perceived ease of use will be influencing the third order which is behavioural intention. This looks quite similar to your model excpet I have extra two latent variables at the FOC and 2nd and 3rd order constructs both have their own indicators
My question is, my 2nd and 3rd order constructs both have their own indicators ( reflective), so does it mean that the repeated indicator approach is not suitable for them in this case?
If it is not suitable, how should I perform the analysis of this model? Because my first order construct is reflective and second and third order is formative. And does it need to perform two-stage analysis?
In this case, PU and PEOU, as well as BIU are not higher order factors. They are just being predicted. So, no need to follow this two stage approach. HOC are constructs with lower order dimensions defining them (measurement claims), rather than predicting them (structural claims).
@@Gaskination Thank you so much for your prompt reply Dr Gaskin and I truly appreciate that. Now I have got a clearer picture. However my results are quite strange even though the measurement model has been established well construct reliability and validity, the structural model doesn't seem fine because the t statistics doesn't exceed 1.96 and p value is close to one. Some of the paths even show n/a result which confusing me. May I know what should I do to sort it out? ( I have removed a few constructs that have low loading)
@@limunng1841 If N/A is showing up, then there is some error happening. If you are using PLS_consistent, try using the regular PLS algorithm instead. I have seen lots of trouble with PLSc.
@@Gaskination Thank you for your reply Dr. Yes, I have just tried using the regular PLS and all the t-statistics and p-values showed up. Although some of the values obtained did not fulfill the criteria and I need to justify for that. Also, I just want to confirm if I apply regular PLS instead of PLS_consistent to my reflective model, is it okay with that? Since I have read many references and it is suggested that PLS-consistent should be used for reflective model.
@@limunng1841 PLS consistent is better for reflective models when it works. However, when it does not work, there is not much that can be done. In this case, if you have access to AMOS or Mplus or another covariance-based application, that might be more justifiable than using SmartPLS.
Dear Dr. James, will the process and interpretation be the same for a formative DV, formative IDV and Formative Mediating construct?
Shall be grateful for your reply.
The process is the same for any endogenous 2nd order formative factor. If it is not endogenous (being predicted by other factors) and 2nd order (has first order latent dimensions) and formative, then the process is simpler (no need for using latent variable scores).
Hi Dr. Gaskin - thank you so much for this video. If I may ask a question: I have a construct, health information privacy concerns, that has been modeled as a reflective-reflective 2nd order factor in prior literature. In my study, health information privacy concerns is both endogenous and exogenous (essentially a mediator). I'm a bit confused as to what approach (repeated indicator or LVS) I should be using
If it is reflective all the way down, then you can model it without multiple stages using LVS. You only need to use LVS if you have a formative higher order factor that is in an endogenous position.
@@Gaskination Thank you Dr. Gaskin!
Hi Dr Gaskin, thank you so much for yr video. It really helps me a lot. I would like to ask you about my model. Fyi I have three exogenous contructs, where one is second order factor (reflective-formative), and another two is first order factors (reflective). My endogenous second order construct is (reflective-formative) and I also have one moderating factor and one mediating factor, both are first order and reflective.
Can you advise which approach should I use for my model? I am quite confuse as I have combination of first order n second order factors as exogenous & also moderation/ mediating factor. Thanks Dr!
If you have a higher order factor that is formative and endogenous (being predicted by other latent factors that are not just its lower order dimensions), then you need to use a multistage approach like in the video above. In all other contexts, you can just do it in one stage without having to use latent variable scores.
Hello, Prof. James. I am just wondering why is there very little literature available on PLS use in HR/employees behavior. Most of the literature referenced the use of PLS in marketing research only. I know, small size, data normality and use of formative variables are basic reasons to use PLS, which is not specific to marketing but I think most of the social sciences research meet those basic requirements, but still very less work. Any good reason you know, why?
It is most likely because PLS is not taught in most universities, but CB-SEM is sometimes taught. So, people use what they know.
If we had the following relation CSE -> ASU -> PIIT, where CSE is the exogenous independent variable and PIIT is the endogenous dependent variable and ASU is the 3rd order construct , would we use the approach of one-indicator (shown till the first 18 min)?
I guess this answers my query, ruclips.net/video/kPeUTKjMF7o/видео.html
Excellent video! You are doing a great job. Could you please help me with a question? I am searching for hours to find the correct answer. I have a formative second order endogenous construct in my model. I have first-order formative indicators linking to the formative second order endogenous variable. Can I follow the same method followed here. Any difference with arrows in second order should be considered? Any differences from what described in the video? Thanks for the time.
You can follow this approach for that type of model.
Doctor Gaskin reciba un cordial saludo, mi pregunta es: ¿Se podría obtener las puntuaciones en varios pasos? Por ejemplo, en tu modelo tienes un constructo de tercer orden, entonces, en el paso 1 se extraen las puntuaciones del primer orden y se reemplaza en el segundo orden, en el paso dos se obtiene las puntuaciones del constructo de segundo orden y se reemplaza en el de tercer orden. Como el constructo es de tercer orden ya no habría mas pasos que realizar. ¿Sería adecuado?. gracias su excelente explicación.
si eso estará bien.
Dear James, I was wondering in the case of having a higher-order construct, how should we report R2, F2 and Q2. Do we need to report them only for the highest order or for all! I appreciate your help here.
R2 will be for the highest order endogenous factors. F2 is for all individual predictors' effects. Q2 is for the model as a whole.
@@Gaskination Dear james, thank you so much.
Hello Dr. Gaskin,
Thank you for the video. what happened if we have 1 exo and 1 end. variable composed as reflective formative model? end. variable is multidimensional construct with 2nd and 3rd order, meanwhile exo.variable is first and second model. In addition, the exo.variable also serves as the mediating variable.
Also, for the HTMT values, when should we assess it? Is it during the first stage when we lay all the the first order constructs
Thank you!
I think you'll need to email me a picture of your model and the question. You can find my email if you google me. HTMT is for assessing discriminant validity. You can assess it for all levels. The most important level to assess it is the level at which you will be testing your hypotheses.
@@Gaskination Thank you for the kind help Sir. I will contact you further via email.
Hi dear JAMES, ın the final model in this video can we set such a hyphothesis that : Repurpose has significant effect on Spirit formation/creation or for example: can we set a hyphothesis that Spirit is positively predicted by recombine. I try to ask that how can we hyphothesis that second order is formed-created-predicted by first order construct in FORMATIVE MODEL? Is it same with LISREL models, I just try to understand.
Usually stated hypotheses are about causal (structural) relationships, rather than measurement relationships. Since Repurpose and Recombine are dimensions of Spirit, we would not make a hypothesis about it.
Dear prof. Gaskin,
I generally use PLS-SEM for my analyses. Specifically, the models I test (Human Resource Management filed) require reflective - formative higher order components. Recently, the model I tested was similar to the one presented in this video. The relevant paper was sent to a top tier journal. The editor marked the manuscript as a major revision, although he / she commented that the readers of the journal might be unfamiliar with PLS-SEM. Hence, he / she suggested to use a more familiar to readers statistical approach. Hence, my question is this.
Could I run the same model with AMOS? In Human Resource Management literature, researchers tend to use the single-index approach (they simply calculate a mean score in SPSS, e.g., for the adaptive system use) and run the model in AMOS. However, it is difficult (if not impossible) to run a CFA. Do you have any ideas? Put it simply, how would you run the model in the video if PLS-SEM was not an option?
Thank you in advance for your time in reading my comment :)
Kind regards,
Panos
Mistral you would have to generate factor scores or averages and then just build a path model in Amos.
Thank you so much for the reply. I'll give it a try!
Dear Dr. Gaskin I am very appreciated to your help ,
also i have question is it allowed to build relationships among the sub-factors of first-order or second-order factors like your example if we supposed their relation between repurposes and recombine or between content and spirit
If you do that, it would be better to remove the higher order factor for that model. Causal relationships should be drawn only between the highest order factors in the model.
If i keep it , there are any statistical problem could affect the result because i dropped the second order factor and that did not affect the relationship of IV factors .
I do not see any problem with it.
Hi James, the video is really helpful. But how to do a moderation analysis with 2nd order construct? I followed the steps exactly shown in this video. But the "add moderation effect" is disabled for every 2nd order as well as normal constructs.
The easier approach would be to get to stage two (obtain latent variable scores) and then add the moderating effect with the updated model.
Hi,James, thanks for your videos here. Your demonstration here helps me a lot. but I still have two questions:
first, my second order construct is a mediator (both being predicted and predicting), i want to confirm that is this also the situation that needs to apply the latent variable score approach.
second, is there any reference talking about what is a good weight of these formative indicators/ dimensions (just like the factor loadings in reflective factors , such as factor loadings above 0.6 or 0.5 will be considered as good and acceptable). I understand that the weight of formative indicators means the importance and contributions of these indicators to the constructs, it is just my reviewers want to me to answer : based on what standard or reference can we call the formative indicators contribute substantially.
thank you again for your attention
1. Yes. Anytime the 2nd order formative factor is being predicted (even if it is predicting something else), it needs to use the 2 stage approach with latent variable scores.
2. Probably Hair or Ringle (or both) have talked about good formative weights. I think Gefen and Straub talk about it as well. See some references here: statwiki.kolobkreations.com/index.php?title=References#Partial_Least_Squares
@@Gaskination Dear Prof. James,
Thanks for the useful video, just need to clarify one thing, according to Becker et. al. (2012), we may use the scores of the first order constructs of the repeated indicators method in stage two of the Two stages method , but in your case you have used the scores of second order instead of using the scores of first order scores. Could you please justify it or guide me to a reference, because I am in need to follow your approach.
@@mohamedshekar9207 The scores of the first order are subsumed by the scores of the 2nd order.
Hi Dr. Gaskin, thanks for the helpful video. May I ask a question, how to calculate CR and AVE for second and third order construct in reflective-reflective-reflective model. I think Marko or Hair mentioned that for second order construct, researcher should assess the reliability and validity manually. If u support this, Is manually calculation way for both second and third order variables the same or there is difference? please your advice in this matter with many thanks in advance.
The calculation at every level is the same. If for first order, use the indicator regression weights. If for 2nd order or higher, use the regression weights from that order to the lower order (2nd to 1st, or 3rd to 2nd).
@@Gaskination Thank u so much for ur reply, I have two more questions please;
1- Is it possible to use the repeated indicator approach for model consist of Reflective-Reflective-Formative, (the first order construct is reflective, the second is reflective and the third is formative).
2- As u mentioned in your video that when the second or third order construct is not endogenous, no need to use two stage approach, so could u please guide me to any book or article to use it as reference?
Many thanks in advance.
@@abdohasan9140 Repeated indicators should be used whenever there is a higher order formative factor. So, in this case, yes. The logic for not needing to use the two stage approach is simply mathematical. No reference necessary. When that formative factor with repeated indicators is endogenous, then the predictors cannot predict even 1% of the variance in the factor because the repeated indicators are identical to the measures from the formative dimensions. Therefore, the formative dimensions predict 100% of the variance, leaving none for any other predictors.
hi.. thank you so much for your videos they are very helpful. Can you please distinguish between the factor and path analysis and when to use them when running the PLS Algorithm, Thank you
Factor is for when you are examining the outer (measurement) model. Path is for when you are examining the inner (structural) model. However, you'll likely find the results nearly identical except in a few weird edge cases.
@@Gaskination thank you so much for the clarification
@@Gaskination I have some follow-up questions regarding second-order constructs within the context of my research framework. To ensure I provide all relevant details, including my research framework and objectives, I would appreciate the opportunity to discuss this via email due to confidentiality concerns. Could you kindly share an appropriate email address for this purpose?
Thank you very much for your assistance
@@hannahaherdemla5364 You can find my email by googling me and then completing a captcha.
i have just would like to discuss further the repeat indicator procedure that as reminder of Henseler. Is that correct that with reflective construct will be followed by repeat indicators of second construct and all of the indicator repeat of all indicators in primary construct. thanks
Hi Dr. Gaskin, thanks for the informative video. Apparently, I did read on Hair et al. (2017) and Becker et al. (2012) that in performing the repeated indicator approach, the number of indicators in the LOC should be equal to avoid biases.
Can you assist to confirm this? Also, how to deal with HOC when there is an unequal number of indicators in the LOC?
Thanks for your kind guidance!
It is true that the best approach is to have equal numbers of indicators in the first order factors. Otherwise, the factor with the most indicators will naturally explain the most variance in the HOC (which is comprised of all those repeated indicators). So, the model I have here is probably not perfect. If you do have an unequal number, you could try to trim redundant ones. Theoretically, all reflective indicators are redundant. The optimal number of indicators for a reflective construct is four.
@@unconsciouscognition7449 thanks and well noted on the response! I just got another latest reference that unequal number of indicators does not impact the result (Cheah et al., 2018).
However, do you have any idea if 5th order constructs is possible with repeated idicators approach?
Thanks again!
Hi Miss Nabilah .
I read the same article of Becker,et.al (2012) and in the conclusion after running different types of simulations , it seems that they concluded that the unequal number of indicators did not lead to the biased result. The biased result was due to other factors.
Dear Professor, I struggled to run my model as my dependent variable is a higher-order variable. Your video helped me to solve that issue, and thanks. I got a question regarding the second-order reflective formative model. When I evaluate the validity and reliability of the reflective items, Composite reliability and AVE meet the threshold levels. But Cronbach alpha and Rho values are less than 0.7. Should I remove those items before calculating the repeated latent variable score for the formative model? Please reply. Thanks
If CR and AVE are fine, then you can proceed.
Thanks a lot for the reply.Can I have any reference for this as I am writing a paper from my PhD study?Thanks
@@fazfaz2605 This one is probably fine: Fornell, C., & Larcker, D. F. (1981). Evaluating structural equation models with unobservable variables and measurement error. Journal of marketing research, 39-50.
@@Gaskination Thanks a lot professor.
Dr. Gaskin - I'm sorry to bother you again. My model contains a HOC made up of 4 LOCs - privacy concerns as the HOC; collection, errors, secondary use, and unauthorized access as the LOCs. Unfortunately after removing a few items based on low outer loadings, I'm finding that the AVE for one of the LOCs does not increase beyond 0.479, which is below the 0.5 threshold for AVEs. Cronbach's alpha, rho A, and Composite reliability for this LOC are all at acceptable levels - just the AVE is the problem. In fact no matter what I do the AVE will not increase beyond the 0.479 value. I'm also seeing some discriminant validity concerns via the Fornell-Larcker criterion with 2 of the LOCs. How would you suggest I handle these concerns?
On another note - I have both Hair PLS books (primer and advanced) and have a few papers on how to handle reflective-reflective HOCs in SmartPLS.. However, the HOC I'm using is a bit more complex than the examples used in any of the books/papers that I have. Is there a resource you would recommend that walks the researcher step-by-step on how to handle validity and reliability testing for more complex HOCs? For example, when using the repeated indicator approach, how to handle outer loadings on the reflective-reflective HOC (do the same rules apply as when dealing with non-HOC items?
Finally, do you have a video (or a recommended source) for testing for heteroscedasticity using SmartPLS? I'm so sorry for all of the questions. Thank you so much for these videos - they're a lifesaver.
1. Regarding reliability and convergent validity, AVE is a strict measure of convergent validity. Malhotra and Dash (2011) note that "AVE is a more conservative measure than CR. On the basis of CR alone, the researcher may conclude that the convergent validity of the construct is adequate, even though more than 50% of the variance is due to error.” (Malhotra and Dash, 2011, p.702). Malhotra N. K., Dash S. (2011). Marketing Research an Applied Orientation. London: Pearson Publishing.
2. If HOC is reflective all the way down, then you can evaluate it the same way you evaluate a first order reflective construct. If the HOC is formative, and using the repeated indicator approach, then follow the guidelines in this video, based on this paper: Gaskin, James; Godfrey, Stephen; and Vance, Alex (2018) "Successful System Use: It’s Not Just Who You Are, But What You Do," AIS Transactions on Human-Computer Interaction (10) 2, pp. 57-81
3. Hengky has responded to this issue here: forum.smartpls.com/viewtopic.php?t=2006
@@Gaskination Thank you so much Dr. Gaskin for your advice. I will do my research on the resources you suggested. Thank you again
Hi Dr James. Thanks for this video. I am inspired by this video and am writing a paper on multilevel analysis using Smart PLS. Does the procedure hold same when the 2nd order is formative with one variable and 3rd order has 3 variables and so on?. I am working in social sciences. Since iit is an integrated psychological model of two theories, formative in nature I have this doubt. Thank you.
I don't think I understand the situation. You have a 2nd order factor that only has one indicator? That doesn't make sense to me. Here is a paper that uses this approach:
Gaskin, James; Godfrey, Stephen; and Vance, Alex (2018) "Successful System Use: It’s Not Just Who You Are, But What You Do," AIS Transactions on Human-Computer Interaction (10) 2, pp. 57-81
If the third and second order factors are perfectly predicted, often with big T-score, and R-square value above 0.90 (just as you stated around 9:22 in the video); some of the reviewer ask me, how do I justify with such large R-square value? Are they suspicious and unlikely to be true? Is there citations that support such big R-square values? Thanks for such an awesome video, learned a lot.
I would not report these r-squares because they are irrelevant for the model. They are just a step to get to the correct model.
can you explain me more please
Hi Prof Gaskin
The significance of t values is composlory during developing 2nd order construts?
Secondly, loading is like -0.553 about one form two dimensions during developing 2nd order construt then it alarming or what?
1. For a reflective factor, yes. For a formative factor, hopefully, but not compulsory.
2. A negative loading from a 1st order dimension of a 2nd order factor is not good. This means it is inversely related to the other dimensions. Check the scale direction (if it was reversed).
First of all thanks for your videos and sorry for my English :) I have a question about my model. I dont know if it is reflexive or formative. My latent variables are Innovativeness, Creativity, and Innovative Performance. Innovativeness is composed of 10 unobservable dimensions and each contains 3 observable indicators. The relationship between Innovativeness and innovative performance, I believe to be reflexive. But the relationship between the dimensions of innovation and innovation I think its formative, am I right?
You are probably right. The way to know if it is formative or reflective is whether the dimensions are interchangeable (i.e., essentially the same thing) or if they are different parts of the higher order factor. If interchangeable, then reflective, if not, then formative.
Hi Dr. Gaskin, thank you for your video.
I have a question. I used the repeated indicators approach in the 1st and 2nd order factors. Nevertheless, I used a multi-actor sample so that the data for 1st and 2nd order indicators are different. I have an issue with HTMT or Fornell Larcker and VIF for the factors with repeated indicators. On the other hand, convergent validity meets the criteria. Do you have any suggestions on how I can fix the issue?
The factors with repeated indicators do not need to be discriminant with their lower order dimensions. They are expected to be highly correlated (since they use the same indicators).
@@Gaskination Thank you for your reply, Dr. Gaskin. Do you have a paper or any references that I can refer to? I have read papers related to PLS-SEM, I still have not found the statement of discriminant validity issue can be ignored when we use the same indicators in the model. I need this for supported my data analysis in the thesis.
@@Abd-mv1mr Probably the reason you have not found such a statement is because it is mathematically apparent. There is no need to declare it because it is self-evident. I don't think you need any reference for this. You can simply state the obvious that repeated indicators imply significant overlap in shared variance.
@@Gaskination Thank you Dr. Gaskin. I really get help from you.
Thanks for the video, it's very helpful. But there's one thing I wanna ask. What if the indicator that you put in second stage or first stage latent variable has loading factor value below 0.7 but those indicators are resulting above 0.7 in 3rd stage latent variable? Thanks
For reflective factors, loadings below 0.700 are fine, as long as the AVE or CR are above recommended thresholds. For formative factors, there is no minimum indicator weight.
@@GaskinationThank you Dr. James. I want to ask again. What if the first order construct has loading score less than 0.7 towards the 2nd order construct? Do I have to delete that 1st order construct?
@@ayunsaskhiyah7044 The same rules apply here, but we can be somewhat more lenient since it is at the 2nd order. So, unless the loading is truly low (e.g.,
I am evaluating a third-order model just like this one but it stops at ASU level, where first-order indicators are reflective, second and third-order are formative repeated indicators. I have done individual analysis but now I want to evaluate the impact of first-order factors on third-order through the second-order factors for two different product categories, how should I do the comparison in SmartPLS3? (2nd order factors will act as moderators) MGA doesn't seem appropriate for this or maybe I don't know how to use it. Variables in the model are the same and the number of data sets for both the categories is the same as well 449. Can you please guide.
If the third order is repeated indicators, then it will be useless to test the effect of first order on third order since the third order is comprised of the first order indicators already. This is not a mediation (or any kind of structural relationship). It is a measurement relationship.
Hi Dr. Gaskin,
your videos are very helpful, I have a question, I am getting half of variables' HTMT criterion (discriminant validity) greater than .90. and Cronbach's Alpha smaller than .70. however composite reliability , average variance extracted (AVE), Fornell-Larker Criterion are good. in this case how do I establish discriminant validity ? please help!!!
TIA.
If AVE, CR, and FL criterion is fine, then this is good enough. SEM is not about optimizing and satisfying all criteria, but instead about using validation criteria to show that your model and variables are probably sufficiently valid to place confidence in the results.
Dear James, What is the difference between using PLS Algorithm and Consistent PLS Algorithm in SmartPLS 3?
Here are two excellent sources for this question:
forum.smartpls.com/viewtopic.php?t=16249
www.smartpls.com/documentation/algorithms-and-techniques/consistent-pls
Thank for the video, it is very useful and exact what i need. However, i am confused between consistent Bootstrap & normal Bootstrap. I have taken workshop from SmartPls and they were also running the model either reflective or formative with normal PLS/Bootstrap method. Now, after looking at this video i am not sure shall I use consistent bootstrapping even for normal reflective model? or the normal bootstrapping/PLS. For formative also. Your reply would be really helpful.
Consistent is only for models with all reflective factors. The reason they didn't use it during the workshop is probably because it is occasionally unable to produce estimates.
@@Gaskination thank You for clarification.
Hi Dr. James,
Really need your help: I've built a 2nd order reflective factor but I'm wondering what is the meaning of the path coefficient from the 1st to the 2nd order? In my case, the 1st order is Consumer Brand Engagement that is reflected in Cognitive (0,55), Affectional (0,95), and Activation (0,8) dimension. What do the values mean? That Cognitive dimension only impacts 55% of a variance in Consumer Brand Engagement? Thank you so much for the help,
Diogo
The estimates between 2nd order factors and their reflective first order dimensions are the extent to which that dimension is correlated with the other first order dimensions.
@@Gaskination thank you so much. So there is no way to find out how each dimensions affects the main construct (Consumer Brand Engagement)?
I have another short question: is it possible that a direct path coefficient is -0.458 (with significance) but indirectly is positive?
Kind regards,
Diogo
@@diogodorey4486 In a reflective design, the indicators don't affect the factor. They are a manifestation of the factor. As for the indirect question, yes, that is perfectly logical. This is what is called competitive mediation. You can read more about it here:
Zhao, X., Lynch, J. G., & Chen, Q. (2010). Reconsidering Baron and Kenny: Myths and truths about mediation analysis. Journal of consumer research, 37(2), 197-206.
Hi Mr Gaskin,
first of all, many thanks for all your great videos - including this one! They are a great help! :)
Regarding the models i would have two questions. Maybe i have missed it, but do I actually also have to correct the the second order construct's outer loadings? Especially when they are all ok (>0,7) for the first order constructs? In my model, the loadings of all my first order construct items look great, but on the second order they are mostly below 0,7.
My second question would be: Do i also have to do the two stage approach when i am using my second order construct only as a moderator and/or exogenous variable?
Thanks to a lot in advance to everyone who could help me here!
BR, Konstantin
The loadings between the 1st order dimensions and their 2nd order factor are to be evaluated in the same way as the loadings from the observed indicators. So, if formative, then they should be significant. If reflective, they should be above 0.700 on average. Often scholars are more lenient with these loadings though, since they are more loosely reflective than a unidimensional factor. Also, you only have to use the 2 stage approach when the 2nd order factor is formative and endogenous. Otherwise, there is no need.
@@Gaskination Thank you very much for your detailed answer! That really helped a lot!! Kind regards from Munich
Dear prof thanks for this video. I have a question. In second order reflective-reflective mearment model, CR value between .60 to .70 is acceptable? I have just two dimensions. If yes than can you share any citation with me?
Thanks.
Yes, when there are only two estimates used in the calculation of the CR value, then less than 0.7 is acceptable. This is stated in Joseph Hair's book:
Hair, J. F., Anderson, R. E., Babin, B. J., and Black, W. C. 2010. Multivariate Data Analysis. Upper Saddle River, NJ: Pearson.
Hello Boss, thank you for this video. I have a question.
Does this second order means that the middle two variables are formative in nature and model is mixed model? please reply
Spirit and Content (the middle two variables) are both formative. Variables like PIIT are reflective. So, yes, this is a mixed model.
@@Gaskination I am bit confused, because according to my understanding , one variable cannot be formative and reflective at the same time. For ASU Variable: Spirit and Content are Dimensions right? How come one variable can be reflective and formative at the same time? Need a bit explanation kindly.
@@dr.mudassir ASU is formative, but the repeated indicators are set reflectively. They are only repeated indicators though. So, it is a formative factor.
@@Gaskination Thank you for your explanation. One last question, can a higher model variables be reflective or all higher model variables are considered formative?
i can see some videos that put directions outside instead inside.
@@dr.mudassir Yes, a higher order factor can be reflective. It is uncommon, but it is possible.
hello sir, i am following the same pattern of second order formative construct model in SmartPLS 2. however, for using the results for further calculations, i need to use the latent variable scores from PLS and convert them to SPSS for calculation purposes like for multicollinearity. so, which of the latent variable scores will be used for further testing in SPSS or any other software? (i mean the first order model's latent variables or the second order model's latent variables?)
Please help, i am stuck up.
You can use both or either. They are all valid scores. I would recommend using the 2nd order when available, and 1st order when they don't belong to a 2nd order factor.
Hi Dr. Gaskin, thank you for your video, it's really helpful. I have a question, if all my second order constructs in the research model are reflective-reflective, is repeated indicator approach appropriate in this case? And what will happen if the two-stage approach was used in the reflective-reflective second order construct? Thank you.
Repeated indicator approach is not necessary unless the 2nd order factor is both formative and endogenous. In the case of reflective-reflective, it is unnecessary. If you do it anyway (perhaps to simplify a model), there is little harm. The model should be roughly equivalent.
James in your example if we were to attach some new reflective indicators linked directly to the latent variable ASU what would happen to the carried over indicators from the first order level added to the second order latent variables Spirit and Content. Do they still get carried to the third order ASU or not.
There should be no reason to do that. Instead, you would include a new first-order latent factor as a dimension of ASU.
@@Gaskination HI James, Ok thanks for your reply, for the latent factor are you referring to the Blue Circles in Smart PLS. If so would I add a new latent variable called "New1" and connect it to ASU. The indicator would then connect to New1. Is that correct. Do you have an example showing this arrangement you could share with me.
@@martinemmerson2809 Yes, a blue circle is a latent factor in SmartPLS. Your stated approach is correct. I don't have a video for that though.
HI, thank you for the video. just one question, is it possible to have two-way connection between two latent variables?
Not in SmartPLS. However, when running the algorithm, you can select the option to connect all LVs. This simulates a two-way connection.
Hi James, I have a 3-construct formative model and there is a comment from the reviewer saying that "The dimensions related to XXX service include many aspects of E-service process, so that the authors should consider second-CFA approach.".
1) In fact that "E-service process" is a formative construct with several items named as E-quality, should I understand as that the reviewer wants me to change that the E-quality to reflective on the 2nd order and run the results of "PLS outer loading, CR Cronbach alpha & AVE"?
2) Still I can use SmartPLS for such analysis?
3) Is the clip here is the best for my case or any other (like "SmartPLS Formative 2nd order Constructs" of yours) you would recommend?
Thanks!
1. Yes, that's what it sounds like.
2. Yes, SmartPLS can be used for 2nd order factors, as shown in the video above. If you are only having reflective factors, you could use AMOS instead to measure model fit.
3. This one above should be fine.
@@Gaskination Thanks a lot, James! So, should I run the whole model including both 2 formative constructs and one changed reflective construct for the result? Or just the individual model of that reflective construct? Because the CFA includes Chi-square, RMSEA, CFI, NFI also, do I need to report those or just "PLS outer loading, CR Cronbach alpha & AVE" are enough?
@@YY-vy2bq I would guess you want to run the whole model with all factors or else you won't be able to effectively test your theory. If you still have formative factors, then you'll need to use SmartPLS.
@@Gaskination Thanks James! May I know should I report "Chi-square, RMSEA, CFI, NFI" also, or just "PLS outer loading, CR Cronbach alpha & AVE" are enough for that model with one reflective constructs while the rest constructs are formative?
@@YY-vy2bq SmartPLS does not generate many model fit metrics. I think it only generates SRMR and a couple of others: ruclips.net/video/Lp3r5F-eVCs/видео.html
can we get the excel data to be used in this model?
Hi Mr Gaskin,
I have a second-order formative construct in my paper, does it necessary CFA calculating for that construct? I mean, in the part of data analysis, factor loads and fit indices have been reported but I did not enter formative construct into the CFA calculation.
In other words, I didn't mention that in measurement model, and now I am going to calculate CFA for formative construct separately, does it accurate ? I hope that you got my words and help me to improve my paper.
thanks alot.
It is best to include all the factors together in the factor analysis. This will help to demonstrate discriminant validity,
Thank you so much dear Mr Gaskin .
Can you help me by telling that, from above video in second order. its formative model right/? because i have same model.
Variable E-HRM, sub variable 5, and these 5 sub variables have items 3 each.
I use SmartPLS 3 using student license. Unfortunately It doesn't support copy function in Excel format, even Ctrl+c function as well. In this case, how can I copy LVS?
Not sure. If you can't export to excel or copy to clipboard. Sorry about that.
Dear Dr. James Gasking: Thanks for the knowledge sharing. I have one question regarding R Square, what is the minimum threshold for it?
Depends on the DV. If it is a perceptual variable like satisfaction or behavioral intention, then it should be a higher R-squared, like 30% or more. If it is an observed variable like revenue growth or price-per-share, then it is fine to be lower, even as low as 2%.
Dear Dr.@@Gaskination , Can you pls provide us with a reference to study more about R2 value. I heard the threshold must be more than 1 for human behaviour related studies. One of my models got something like 24%, with high t statistics. Reviewers had no questions about R square. Pls recommend a reading on R2. Thank you in advance.
@@tbasou R2 must be less than 1.00 because it is a percent of explained variance in the endogenous variable. R2 increases as you add more predictors. In the social sciences, it is best to have R2 above about 20%, but if it was as high as 70%, we would be suspicious that there might be a tautological relationship in the model.
Dear Dr. Gaskin, It is correct I got here "R2 must be less than 1.00"?
@@tbasou Yes, because it is a percent, it must be less than (or equal to) 1.00 because that would be 100%. You cannot explain more than 100% of the variance in an outcome variable.
So we don't need to check AVE in 2nd order reflective-formative model? Only reliability and discriminant validity? And what about the threshold for removing factor loadings?
AVE is relevant for all reflective factors (regardless of 1st or 2nd order). Reliability is also only applicable to reflective factors. Discriminant validity is applicable to all. As for removing items, if they are on reflective factors, then you can remove them if they are less than 0.500 (if you still have at least two items). For formative factors, it is hard to justify removing an item unless the VIF is really high.
@@Gaskination alright, thank you so much, I wasn't expecting such a prompt response. I got confused because PLS itself higglight all the factor loadings below .70, however you only removed the negative. Everything clear now. Thanks once again. You are doing a great job of facilitating beginners.😊😊
@@Gaskination what if the first-order construct is formative, do I need to remove those indicators with high VIF from the first, second and third orders?
Hi, I think your video is not yet completed ..... I remember you said earlier that R square should not be more than "1'; yet after all the explanation, your result shows that the some of the R square is still more than "1" ...... please elaborate more....... thank you in advance!!
The R-square in a structural model should not exceed 1.00. However, the R-square in the measurement model shown is not relevant. You are correct that it appears I did not finish this video to proceed to the structural model. The next steps would be to extract the latent variable scores and then create a new model with those. We would then find that the R-squares for those variables are less than 1.00
@@Gaskination Thank you for your reply, Mr. Gaskin. Yes, I saw your other videos explaining the next step for the 2nd order (higher order); both the one with SmartPLS v.3 and Smart PLS v.2 - they help my project a lot. Thank you for sharing and I hope you keep making videos about Smart PLS with more details and depth ....... , cheers!!
@@Wawan-wx5jm can you please share the links to those videos? I have the same model, need to learn assessing structural model
Please how can we know if the second order factor should be formative or reflective? Is there a helping reference? Thanks for all Dr.
This is the most popular reference for it: Jarvis, C. B., MacKenzie, S. B., & Podsakoff, P. M. (2003). A critical review of construct indicators and measurement model misspecification in marketing and consumer research. Journal of consumer research, 30(2), 199-218.
Here is a video to explain it a bit: ruclips.net/video/gw0xvvJw-AM/видео.html
Thank you for very useful video. Could I ask you a question about how to interpret R square value of ASU which exceeds 1.
Because this is a measurement model, rather than a causal model, the predictive measures (R2) are not relevant. When we move on to the causal model, the R2 becomes normal again.
thank you very much ^^
@@Gaskination Hi, and thank you for all the great videos. In my model the 2nd order construct is exogenous, meaning it is not being predicted. However, like in your video, the R2 values are >1 when running the PLS algorithm using the "path" weighting scheme. How can I achieve R2
@@MeineUndDeineOma This is fine and normal. The R-squared for an exogenous construct is irrelevant. It should be 1.00 because it is being "predicted" by its own indicators.
@@Gaskination Thank you for your quick reply, this is awesome! Since I modeled a 2nd order construct using the reflective formative approach I assume I have to use the simple PLS algorithm? The results when running the Consistent PLS algorithm deviate a lot, especially R2 and significance (t-values).
@James Gaskin did you upload the updated video for this?
I did not. Jorg and I could never work out a time to discuss it. However, after discussing it with other PLS experts, we determined that the results are very similar using this method.
@@Gaskination Thanks a lot sir. Your videos, reviews and replies are really very much helpful for me.
I understand the example used in Chp 5 of A primer on partial least squares structural equation modeling.
Does it mean we cannot do a test on convergent validity if we dont have a single item global measure that is reflective for our formative construct.
Say my formative construct is safety concern.
However this is no single item global measure for safety concern available in the literature (in fact safety perception is a scale I have developed, lets say) then what do I do?
Correct. If you are using a formative scale, then you cannot check its reliability or convergent validity unless you have a target global measure for that construct. In such a case, you'll need to rely on basic formative validation, such as shown here: ruclips.net/video/Qx_7pN7eAes/видео.html
what about the model fit after run your model? the SRMR, NFI, and Chi-square value when we run a double layer latent construct? it does not give the SRMR, NFI, and Chi-square values. Please help.
Model fit and PLS are not really compatible. I would not worry about model fit when running a PLS model. If you want model fit, switch to CB-SEM model (now available in SmartPLS as well).
hi dr. Gaskin, i have few questions regarding assessing higher order constructs.
so i have reflective formative type model, LOC consists 4 LV and the HOC is predicting another LV & not endogenous, what should i do to validate this model?
im also using the repeated indicator approach & Since few of my variables on the LOC has low AVE, composite reliability and cronbachs alpha, i had to eliminate some of the indicators so that i can meets the criteria. I did follow your guide on when you eliminate an indicator from the LOC, you also have to remove the one in the HOC and vice versa.
After i eliminate some of the indiators all of my LOC has already meet the AVE minimum value, but somehow my HOC still has a very low AVE value (0.28)
So my question is do i have to keep eliminating the indicators until my HOC's AVE value acceptable (but then id lost quite a lot of indicators on LOC too)? Or am i doing it wrong? Is there any specific guides on validating this type of model?
Thankn you!
1. As long as the HOC is not endogenous, then you don't need to do the two stage approach.
2. Correct, remove from both.
3. Don't worry about the HOC AVE. Instead look at the loadings from the LOCs to the HOC. They should be significant and moderately strong (ideally >0.500) if the HOC is reflective. If the HOC is formative, then AVE is completely irrelevant for that construct.
@@Gaskination thank you so much for answering my questions! yes, the HOC is formative. last question, so after i'm done evaluating the reflective measurement model, what criteria do i need to fulfill to validate the formative HOC? or i could just run the hypothesis test right away? thanks again, Dr.
@@paradaisyp2711 Here is a video that should help: ruclips.net/video/Qx_7pN7eAes/видео.html
thanks a lot, dr🙏🏻
Hi James, I am currently working on exogenous second order construct with four dimensions having unbalanced number of items, pointing towards that exogenous which is linking towards dependent variable i tried to conduct two stage approach taking those dimensions as items formative towards second order construct but what I get is rho_A=1 only and nothing else furthermore p value of second order is 0.000 however if I try to go with two stage approach taking first order items as reflective from second order results are quite okay. What one is better? Any theoretical suggestions? Kindly guide me
Two stage is fine in this case. Always do two stage when any endogenous variable is 2nd order formative.
Hi, this video have really helped me build my study model. But my model is a second order one that connects each of the four 1st order variables to one 2nd order variable, then all the five variables are connected to one endogenous variable. Attempt to run bootstrapping kept returning "Singular Matrix Problem. Please, how do I resolve this?
Try removing the paths between the first-order dimensions and the other endogenous variable. Their effects happen through the second-order factor.
Hello Professor, I am studying mediation effect of ambidexterity. Ambidexterity is the interaction of explorative (3 items) and exploitative (2 items) learnings. I am unable to create an interaction effect of both learnings where the product will be a mediator. If I do it the conventional way, that will moderated mediation or mediated moderation. I am using SmartPLS 3. Your guidance will be appreciated. Thank you
You want to create an interaction and use that interaction in the mediator position? I’ve never considered (or heard of) this. If we had to label it, I would call it mediating moderation (not mediated moderation).
@@Gaskination Yes, I am studying mediation effect of ambidexterity which is the interaction of exploitative and explorative learning. I searched but couldn't find a solution. Analyses can be performed through Amos but that is too cumbersome that's why I want to do it through SmartPLS. Another method is to make it a higher-order construct and I will proceed with that. Thank you for considering my query.
good evening, by conducting an analysis by SMART PLS, on a trial version of 30 days, on a sample of 40 students, and knowing that my model and third degree and the response rate was almost in favor, I have not could get good results what is the problem. thank you
The problem is probably too much error due to such a low sample size.
Hi, " There are question marks coming up everywhere on loading , R2, on running the model" Please suggest how to correct them.
This sometimes happens with PLSc. Try using the regular PLS algorithm.
Dr. james Gaskin , warm greetings from Jordan
please i have question about hierarchical component model
can the higher-order components play role of a mediating between its lower-order components and the dependent construct?
I mean if we hypothesis that X as high-order component mediate the relation with X1(low-order component) and Y as a dependent construct. please, if you have reference for this point
Theoretically, this doesn't make any sense. The higher order component represents its lower order dimensions, so it can't mediate their effects. It is comprised of the same indicators.
James Gaskin thanks for your reply
Then can I state the hypothesis from low-order to dependent construct without to mention the higher-order factor directly for example in your tutorial if I said trying new features affecting adaptive system use (I am saying like this because I want to know is it belongs to the higher-order factor or not and at the same time I want to know is that low-order significant to the dependent construct or not?
Simply , I want to know which of low-order factors significant to the dependent factor ?
If you want to know the direct effects from lower order dimensions to DV, then remove the higher order factor for this model. You can always put the higher order factor back in for another model.
Thank you very much. for this moment i will rely on the indirect effect results from IV to DV
Hi , James. could you tell me when we use the second order, what test we need to care ?
Same tests as for regular models.
I wanted to run my pilot results for 30 sample size but it rejected it saying sample size is too small. It should be at least 36. Any advise you can give me on this?
reduce the number of variables or the number of predictors. 30 is a very small sample size...
Should we use consistent PLS? Is there a rule for it? I mean when to use it etc.
ruclips.net/video/YfuMCPCdpJQ/видео.html
I just found it before you reply :) thanks.