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!
As I show in this video, you really only need the one constraint. But if yours does not work like that, then do what you can to make it work (remove both constraints and add a variance constraint?).
Hi James, what should Immers match with? I just jumped from the video CFA part 1 to this video. I do not understand how to name the second order variable (le.g. Immerse). Thanks
Hi James, how do I come up with the name of the second order construct (e.g. immerse in this example)? It seems like there is 1 step that I missed because there is an f before constructs, e.g. fFI, fTD My data contain Credibility and Informativeness which belong to Utilitarian. I have the data of Credibility and Informativeness only. And when I put the name Utilitarian as a higher order of Credibility and Informativeness, Amos failed to do the imputation of factor score as Utilitarian is missing from my the Spss file. Thanks again.
The naming is completely up to you. It doesn't require an f before the name. I just put that there because I had another variable already in the dataset named TD. So, just name it something that isn't already taken in the dataset. Then, when you impute the file, make sure to check the new file (usually named like the old file, but with _C at the end).
Good question. As far as I understand it, covariance based SEM doesn't need to distinguish between them. They are simply correlated. So, since the 2nd order factor is comprised of the two lower order factors, we are simply trying to discover to what extent the DV is correlated with the other two first order factors. This is accomplished just fine the way it is modeled. I hope that clarifies.
@catsfancyful You are correct. AMOS does not handle formative constructs very well. In fact, I've never successfully built a second order formative construct in AMOS. I recommend using partial least squares software for that. Something like pls-graph, or SmartPLS.
1. yes 2. I would call it 7-factor 3. I suppose so if immerse was included as a subfactor of the new latent factor. 4. You can analyze the correlation using the double arrow. It is just the standardized value. Or you can use a single arrow if you want to also get the r-square (squared multiple correlation in the output tab), which is more often used for evidence of predictability.
I can't find words to express my thoughts. Thank you! Thank You!. THANK YOU! Hope to take part in your researches (as I am very fond of doing academic researches)
Yes, you can try constraining the two factor loadings to be equal (name both of them "a" or something like that). That would be my first approach. I think you may also be required to constrain the variance on the 2nd order factor to 1 in a case like this. I can't remember.
I'm not sure if there is a standard practice one way or the other. If it were me though, I would do both. That way I could show that the first order factors are valid and reliable before using them in the 2nd order factor. And that way no reviewer can ding you for it.
Dear James, I hope to have your answers to the following: 1. Is it right that the first two (i.e. tFI and fTD) are subfactors of Immerse? 2. Do we call the whoe model as a 8-factor or a 7-factor model? 3. If we draw another latent variable to capture all these factors, will it be called a 3rd-factor model? 4. To prove predictive validity, I wish to analyze the correlation of a latent variable and another latent variable, do I use the single-arrow or double-arrow? Thanks a lot!
Honestly it just depends on everything else in the model. You might try a reliability test (Cronbach's alpha) to see if you have reliable factors. If not, then you might be able to identify where the problem is.
Hi James, I am looking at your videos recently and it is immensely helpful. Thank you. A quick question:- how do you measure higher (second order) level construct? Is it mean of all items in the first order?
For reflective 2nd order, there is no direct measurement. Instead, the "indicators" of the 2nd order factor are the first order latent factors. If formative, then you can use a repeated indicator approach, or an instrumental variable (single proxy capturing the main idea).
@@Gaskination its reflective, and I tried creating a new variable by calculating mean of all items collectively and using it . Also just assigning all items collectively to second order variable. Both didnt work. Can you possibly help?
@@ankitatibrewal6842 If it is refective, then it does not need any indicators. The 1st order latent dimensions are the "indicators" in this case. Just as in the video above.
Hi James, thank you for your great videos! I'm wondering that when we decide to combine certain factors into a 2nd order factor, is there any criterion we need to follow? Or just making this decision based on theory and our hypotheses? And, after combining 2 factors into a new factor in this video, what should we call the rest 6 factors when we report? Are they 2nd order factors as well (because they covary with the 2nd order factor) or just normal factors? Really appreciate it!
Terrie Chen 1. Decide 2nd order based on theory. You can also decide it based on statistics, if it also makes sense conceptually. For example, if the two factors are highly correlated, and are also clearly conceptually correlated. 2. The remaining factors are still 1st order.
Thanks James! If I make decision based on statistics, should I check the correlation matrix of final solution in EFA step or check the factor correlations of the initial measurement model in CFA?
+James Gaskin 2nd order factor's items are known as the first order factors isn't? Since the first order factors have their own items, should we compute composites so that we can conduct the EFA for the second order?
Thanks so much for your answer. yes it is exactly true If you have four or more dimensions as 2nd order factor you need one constrain, but if you have two dimensions as 2nd order factor with one constrain, it won't run. I have three way: 1. put constrain for each dimensions. 2. put two constrain and a variance constrain. 3. a variance constrain and a constrain for one dimension. which way is true? Thanks
Hi James. First off I want to say thanks for taking the time to make these incredibly helpful video tutorials! Several have proved invaluable for conducting my analysis thus far. I wanted to check something with you, though, if that's ok. I initially created a three-factor model - three indicators per factor - and found something approaching acceptable model fit. Then, based on the moderate to high intercorrelations of the three factors, I created a second order factor to which the original three factors are subordinate. However, the analysis on this new, second-order model yields *identical* values on all model fit indices to the original single-order three-factor model. Is this to be expected or must I have done something wrong? I was originally intending to see wether the first order or second order model has superior fit but this unexpected outcome has put the brakes on this ambition for the time being! Any advice you can give would be greatly appreciated! Thank you.
+Joshua Lord This is perfectly normal. When you deleted the three covariance arrows you increased degrees of freedom by 3. But then you replaced them with regression arrows, decreasing degrees of freedom by 3. So which model is better? Well, if the three covariances among the three first order variables are too high (greater than 0.800), then it is very likely these are simply three different manifestations of the same thing (a higher order factor) and should be modeled together. However, if the covariances are less than 0.800, then we can probably make an argument that they are sufficiently distinct to be modeled as first order factors.
Hi James, I'm sorry to bother you so much but this is exactly the structure I was thinking of for my model. The only difference is that in my case I would add three more 2nd order latent variables. Is this possible? If it is, the covariances are supposed to be drawn only among the 2nd order latent variables? Is this correct? And last, three of my four 2nd order latent variables are the independent variables in the regression analysis while the fourth one is the dependent variable. Is it okay to treat them all equally in there structure of the model? Thank you very much for your understanding and support. Sorry again, Valerio
correct. And you can covary all the 2nd order variables (including DV) together for the CFA. Then, when testing the relationships between IVs and DV, you should uncovary the DV, and then draw regression lines from IVs to DV.
James Gaskin Excellent! Thank you very much! I really appreciate it! Your support has been fundamental to me. I have just found this interesting paper that explains why higher-order CFA models may not make good conceptual sense. I found it quite interesting. It's a good food for thought. Here's the link to the paper (if you want to give it a look): www.researchgate.net/publication/236004024_Problems_with_Formative_and_Higher-Order_Reflective_Variables Thank you again James.
Hello James Gaskin, I have one last question. I built my model, so I have: 1) three 2nd order independent variable, pointing to their correspondent 1st order variables which in turn point to the observed variable. 2) the three 2nd order factors point also to the 2nd order dependent variable, which in turn points to 1st order variables, pointing to observed variables. 3) I put error terms on the 2nd order dependent and not on the 2nd order independent factors. Having all this set, I get no output by Amos. Where am I wrong? Is a structure like mine feasible? Thank you very much again!
James, Brown (2006) and Byrne (2001) raise concerns regarding a second-order factor accounting for only two factors. This is because although the overall model may be overidentified, the second-order factor model itself would be underidentified. Thus, there may be an identification issue with your model, even if this is not flagged in the output. Is this something you considered and, if so, why is this not a problem here? *Please note that I believe your approach to be correct if the second-order factor accounted for three factors.
Dear James, Two questions: 1. It is a 7-factor model in your video, with the first factor being 2nd-order. Mine is very similar. If I delete the other 6 factors and just run the second-order model for first factor (e.g. Immerse in yours), it won't run? 2. Comparing unconstrained model and constrained models (i.e. covraince between 2 factors = 1) is a way to test discriminant validity. But when I constrained a second-order (e.g. Immerse in yours) and another (e.g. fJoy), it didn't run? Thanks!
@Tomahawk1999 AMOS does not handle formative constructs well. The principles behind moderation do not change whether you have formative, reflective, first or second order constructs.
Dear Professor Gaskin I see in the comments below that someone inquired about formative second-order factors in AMOS, in which you recommended using SmartPLS. However, as my research is covariance-based, I was wondering if a reflective model in AMOS is appropriate when the first-order factors are hypothesised as having an effect on the second-order construct? If not, I was wondering if you know of any books or papers that explain why AMOS cannot deal with formative second-order models? Looking forward to hearing from you.
Modeling a formative factor as reflective will have bad consequences. The Barbara Byrne AMOS manual probably says something about formative models in AMOS. I know my article with Paul Lowry about PLS talks about it briefly.
There could be several reasons why it is not running. Amos is very picky. You are welcome to send it to my email address and I will take a look: james. gaskin@ byu. edu
Hi James, first of all: a big thank you to your excellent tutorials. You actually made me switch from Stata to SPSS/Amos :) I followed your tutorial building a 2nd order construct in AMOS which by itself runs nicely regarding all indices. The model I am trying to test for AVE, CR etc. has 5 latent variables with 6 items each and one latent 2nd order factor. At this point, there are no other latent variables or other variables linking into the model. When I want to test for AVE,CR etc. with your Excel Tool, I run into a problem (have already ruled out any violations against your assumptions regarding naming): After copying both tables into your validation tool, i get "Run time error 91: object variable or With Block Variable not set" Could you tell me what the reason might be? Testing the 1st order model with the same variables (excl. the new 2nd order variable) worked well in your excel. I found it strange that my correlations table of the 2nd order analysis only contains correlations of the error terms - could this be the reason and if so, what can I do to resolve the problem? Any help is much appreciated! Your videos alone were already super helpful!
+Spss Phd You have identified the issue. If there is only one highest order construct (in this case, a 2nd order construct with five underlying dimensions), then it treats it as if there is only one latent variable in my tool. To calculate the validity of this type of model you will need my old stats tools package. You can email me at james.gaskin@byu.edu to get that.
I would have to see your model, but my guess is that you might have neglected to constrain one of the regression weights between the 2nd order and 3rd order variable, or between the 2nd order and 1st order variables.
Hi, James. I have a question about 2 order CFA. I have a data for FLOW construct, which was validated by its author using European data. It was posited as a 3-dimensional single factor with (1) IWM (5 indicators), (2)WE (4 indicators), and (3) AB (4 indicators). Here I have collected Chinese data for it (N=869). And I did EFA for it before CFA, as I think maybe the pattern metrics are a bit different in Chinese context. And I arrived to a clean pattern after I delete IWM4 (very poor loading) and IWM3(cross loading). But then I checked the factor correlation where the discriminant validity problem popped up (correlation >.70). Therefore, I tried to model the construct FLOW as a second order factor to see if doing so would solve some problems. Surprisingly, for the first time of my experience, I got exactly same model fit indices (CFI, TLI, RMSEA, etc...) for FLOW as a second order factor VS a 3-dimensional single factor...I assume that a second order factor is better than the single factor as correlations between IWM and AB, AB and WE are above .70...So, I would like to know if there is anything wrong with my procedure? How can we potentially explain it? Thank you very much! :-)
The model fit won't have changed because you simply replaced the correlation arrows with regression arrows. Therefore you have the same degrees of freedom and you are still accounting for the same relationships. I agree that if you have discriminant validity issues between first order dimensions, then if there is theoretical reasoning to connect them, just make a 2nd order factor. So, you have done it correctly.
Hi Dr James, in the article "Operationalizing Multidimensional Constructs in Structural Equation Modeling: Recommendations for IS Research"' the authors have suggested that for higher order reflective dimensions, a separate CFA should be run where as Hair et al (2007) recommends a "'pooled CFA"' approach even with higher order constructs having dimensions. The video above indicates you have been using pooled CFA concept. if so can you offer some references?
Hair et al 2007 I suppose is a good start. I can't think of any others off the top of my head, but a quick search of google scholar for higher order factors will probably turn up some good results.
Hello! thank you for your quick answer. No the regression weights were fine. I may have found the cause though. I am covarying one 3rd order variable with one first order variable. Am I not supposed to do that? I tried to add a variable, which I covaried with the other two and it ran without a problem. But when I took it off it didn't.
Dear Prof. Gaskin I have some questions about second order CFA and CFA At first CFA: 1) Is there any normality Prerequisite for CFA? 2) If yes How could we do a CFA or SEM with non-normal data with Amos? Second order CFA: should we apply a separate CFA for first order components or not ? if yes should we calculate validity and reliability for them ?
Yes. Your data should be normally distributed if you are using maximum likelihood estimation. Bayesian doesn't require normality. I would do 2nd order CFA with all other constructs as well. Calculate validity and reliability at the highest level.
Many thanks for you response . Can I use boot-strap instead of bayesian for non-normal data? because there is no Bayesian test in Amos and boot-strap doesn"t require normal distribution too.
Dear Prof.Gaskin Thanks for your response . I have some more questions. 1)How should I do CFA with Bootstrap. 2)How we should do Moderating with Bootstrap? Best regards
The tutorial is interesting. May I please request you to describe how to conduct a Moderation or moderated mediation in case of a second order factor(s) as the exogeneous variables.
In this case, I would recommend using a software that makes these combinations for you, such as SmartPLS. Doing latent interactions in AMOS is very difficult and unwieldy.
Hello! Thank you for the video and the Excel document. This is so helpful! I tried to test the validity of a model including a 3rd order factor but it fell. Would you have any recommendation? Thank you in advance.
Hi James, How many constrain (I mean put “1”in regression weight) I need when I have 2nd order factors for every construct. Also the other question is I should put “1” in variance (through object properties) for every construct which have dimension or not? because if you have one construct with two dimension, it cannot run if you put just one constrain. you should put constrain for every dimension or put for variance. Is it true? I know I am in first step to learn Amos. Thanks so much.
Dr. regarding the last 10 secs of the video, It would be brilliant if run CFA with second-order factor solving teaching public how to solve the second order factor validity problems
Hello James, Thank you for your amazing videos. Please, I have a question regarding the reliability test (Cronbach's alpha), I have a 2nd order factor in my model which contains 2 sub-factors. How can I calculate the Cronbach's alpha for the 2nd order factor in SPSS? Should I put all the items of the 2 sub-factors together in the box to calculate Cronbach's alpha for 2nd order factor as a whole? or calculate Cronbach's alpha for each of the 2 sub-factors separately? (p.s. I downloaded your plugin and I found CR value for the 2nd order factor, but I could not find Cronbach's alpha value) Thank you.
watch my video called: "Iteration Limit Reached in AMOS" This shows how to find problematic items (ones with really high standard errors, or ones that load over 1.0) and how to fix them. Hope this helps.
James, thanks for sharing this information. I have one question about second-order models. When you have two or three first orders factors under a second order one, this part is not overidentified (as recommended). I saw one solution on Byrne's book, but it is not working. Do you know how to setup model to get overidentified with two or three first order under a second order factor? Thanks, Gustavo
Thanks very much, as always. What about a second-order formative construct? I understand AMOS doesn't handle this very well? Any thoughts would be appreciated.
Hi James Thanks for another excellent video. If one was interested in the relationship between the lower (1st) order facets and the endogenous variable but still wanted to account for the 2nd order structure of the model (e.g. looking at how extraversion facets loading on a global extraversion factor predicted salespeoples' job performance) is it possible to have the regression arrows going from the 1st order unobserved variables to the outcome...or will the only variance left in these lower variables be the variance unexplained by the 2nd order factor? If not, could you also do a comparison between 1st order and 2nd order direct effects on the DV using this approach? Regards Tom
Tom Bailey If the first order factor is part of a 2nd order factor, then you should not draw arrows between that first order factor and any other factor (aside from its 2nd order factor). Instead, create two models: one with the 2nd order factor and one without.
Hello James, thank you for the video. I want to ask you a question. I used 2nd order Why is the result of model fit in my measurement model and structural model same?
Thank you for answering my question. However, my model doesn't show any negative error term or any standard error that is greater than 1. But I noticed that when i correlate the M.I., the regression loadings as well as the correlation loading will be exceed 1. Is there any method that can allow me to adjust the M.I. while the loadings can stay below 1?
Hi, Your videos are awesome! I'm using them to conduct a second order SEM and I havequestions/clarifications. Can you calculate the composites to be used for the structural model? If so, should you include the first order composites in the structural model since no regression lines are linked to them? Grateful for any help! Thanks again!
Sabeeha Mohamed The composites will be created for whatever latent constructs are covaried during the CFA. This means the higher order constructs. In this case, in the structural model, include the higher order constructs, but not the dimensions (lower order constructs) that were part of it.
Hi James. Thank you for these videos. They are super helpful. I do have a question. I have a second order factor with 4 first order factors (as suggested by previous literature). When I check the loadings of the first order factors on the second order, there is one first order factor that have very low loading (0.1..) while the other 3 have 0.7-0.8. My question is what I can do. Is there any other way other than removing that first order factor? If I remove it, I feel it is hard to justify why mine is different from literature. Thanks so much.
+nganhavy It begs the question whether the 2nd order factor is actually formative instead of reflective. If formative, then the difference in loading is not a problem, but it should be modeled formatively and in PLS, not AMOS. If actually reflective, then there may be underlying issues with the items from that factor. If you plan to model it reflectively, I would perhaps recommend changing this to keep only the three strong dimensions, but then retain the other dimension outside this factor. It may work well as a predictor or covariate.
Hi Dr James, Please i need your help on this : i have a second order contruct and i find good regression weights between 1st and second order factors but before i found low correlation between 1st factors less than .5 (from .102 to .41) to go for a second order ! i'm confused since in the theory the 2nd order exist and reflect 1st order factors. Any advice ? thanks a lot
Dr Gaskin, I have a quick question. In addition to constraining one of the paths between the first order factors and the second order factor do we have to also constrain the variance of the second order factor (in this case immers) to 1? Some books state that both should be applied. Thank you in advance!
AMOS will run without that additional constraint applied. So, you can use either the variance constraint or the path constraint. Either will work and the results should be about the same.
Thank you so much Dr Gaskin. I have one more question and would be grateful if you could help. If we have a latent model, how do we compute the composite factors without doing the CMB test and imputing the factors? Let's say I have a latent model and there is no CMB, so I move on with my original data set but I still need the composite factors. Is there a way to do this? Thank you so much for your time!
I think I found the video for it. "Imputing factor scores in AMOS". It does not say anything about CMB, so I guess that must be the right one. Thank you!
Dear James, how can i perform chi-square variance test of the limited and unlimited measurement patterns. Its necessary to construct discriminant validity. I need chi-square and “d. f.” results. Thanks too much your helps.
James - Over the last several months, I've come across a few published CFAs that seem to be using higher-order factors to hide problems with discriminant validity. In one case, the correlation between two of the latent factors below the higher-order factor was .93 (!), but of course this wasn't reported, because the higher-order factor allowed this to be hidden. Interestingly, adding a higher-order factor didn't change the fit of the model at all - all it did was allow the reporting of the model to omit the correlations between latent factors. To me, this seems like statistical obfuscation to hide genuine problems with the model. Is it ever appropriate to use a higher-order factor when these discriminant validity problems exist? Do you have any recommended reading that might shed some light on this? Thanks!
This is absolutely fine if the two underlying constructs truly are subdimensions of a higher order construct. For example, I've worked with several students who want to assess job satisfaction and life satisfaction. However, these are so highly correlated, that we always run into discriminant validity issues. What we usually end up doing is placing them together into a second order factor called "overall satisfaction". Now, if we were combining something like satisfaction with productivity, that would be much different and would take some really good wordsmithing to convince readers that you really have a 2nd order factor.
James Gaskin Hi James: What are the practical implications of that move? Say I'm running a linear regression: Do I just include one measure (the second order factor) or do I include three measures (the first order factors)? In Rob's example, the two first order factors are correlated at .93. I would have concerns about multicollinearity.
Jake Jensen Multicollinearity would not be an issue if these were modeled as a reflective 2nd order factor. Then, when doing the regression, the causal path is only between the 2nd order factor and some DV.
Hi James, I have followed your guidance and 2nd order factor analysis seems to fit to mine. Thank you. I have one more question. If I want to test the impact of fIMMERS on DV, second factor analysis is the way to do right? but what if I want to see as well the impact of fTD which is subconstruct of fIMMERS on DV? then should i eliminate fIMMERS and validate measurement model and conduct structural model analysis?
Ok. Thanks .. Its same as Burnout has three dimensions i.e. Emotional Exhaustion, Depersonalization and Lack of Personal Accomplishment .. BO= EE+DP+LOPA ... Do you think AMOS is smart enough to understand this equation ... I have seen one of you tutorials on smartPLS (if I am not wrong?) ... Where you explained how a multiple dimension variable can be represented as a single dimension ...It was quite complex but it seemed some how the same as 2nd order factor?
Hi, thank you very much. Your tutorial really helps for my academic studies. However, in this video, i dont understand how you manage to run the last model with amos? Because there are both formative and reflective structure as i see, we cannot run formative structure in amos. Thank you.
Dear James, I would like to know about the correlation between 2 second order factors (both are exogenous in the model). If it is 1.007, does it indicate discriminant validity is not achieved? Or can it be interpreted as (1) even though they are independent variables, they still have linear relationship (1 of the factors serves as independent and dependent at the same time) or (2) may have causal effect. So, discriminant validity is not affected where there will be no issues on multicollinearity.
+Nitha A A correlation greater than 1.00 indicates an estimation error. It is also highly likely that they are strongly similar and may be part of a higher order factor.
+Nitha A yes. that is one way to do it. Or you can do a multicollinearity test to see which ones have overlapping variance. Or you can just look at the wording of the variables and see if they are very similar.
Thanks for your video!can i ask you something? I want to do CFA at a Second Order Construct with 2 factors, but as you know in this case we dont have identification. What should i do?Constrain the 2 factor loadings?constrain the variances?something else?thanks!
Dear James,Thanks for the videos! They are great. Currently I have a problem running SEM with panel data. I searched but cannot find any useful resources. May I please ask you to recommend some literature on SEM with panel/ pseudo panel? Thank you~
Dr. James Gaskin, Thank you for the You tube video however when I try to run my data I do not seem to find / see the excel sheet where you paste the correlation and standardized regression weights you copy from the measurement model
Hi Sir, I'm receiving this problem while running my model "The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint"
This means that one of the factors does not have a parameter constraint. Notice that all of the factors have one of their indicator regression weights constrained to equal 1.00. So, make sure even your second order factors have one regression weight equal 1.00. The other option is to constrain the variance of the factor to 1.00.
The model consist of 5 factors ( 4 first-order) and (1 second-order) the loading of 2nd order was -.56 and .78. when i compute the value of CR and AVE for all factors the value of 2nd order was CR lees than 0.3 and AVE lees than 0.5 because i have negative and positive value of loading. in this case what i should do.
This means that the two dimensions (lower order constructs) of the 2nd order construct are inverses of each other. You may consider reversing the negative one so that it is in the same direction as the other dimension. To do this, subtract all indicator values for that construct from m+1 where m=the maximum possible response. So, if you are on a 5-point Likert scale, then you would subtract from 6. Also, you'll need to reverse the construct label. So, if before it was "sad" now it needs to be "happy", because an increase in the scale would now represent a decrease in the original construct.
I received a notification of another reply, but I don't see it here. I will reply anyway. The first order dimensions need to be in the same direction. So, if it is confidence and anxiety. Perhaps change anxiety to calmness or serenity.
So I collected data where one of variables loaded on 2 factors -which was fine as theoretically there are 2 dimensions to it. However I don't intended to use it that way so I used it as 1 Amos. everything worked fine. but as a check, a professor asked me to do a 2nd order to make sure the 2 dimensions are equally useful/contributing enough.... I did the initial steps as per the video and 1 dimension has a value of .60 and the other .89. does that sound acceptable? Thanks!
Hello James, does a second order factor mean that all measurement items should be merged together to represent only one factor? How do I determine if I should choose a second order factor model? should I look at the modification indices of the two models (the model with first order factors only and the other with second order factor) and judge on the validity of both models? will this be enough to say that a second order factor is a plausible choice?
hishonline Usually the choice of a 2nd order factor is done based on the conceptual level. If the two factors are really just two dimensions or manifestations of a higher order construct, then they are possibly part of a 2nd order factor. Often these first order factors are highly correlated and may have difficulty achieving discriminant validity.
Hello Dr. Gaskin While running the CFA you didn't add tthe error term or unobserved variable to the construct immerse. While in SEM error term was added. Don't we need to add it while doing the CFA also. Kindly see if you can address my confusion. I will be more than obliged for your help. Regards, Dr. Namrata Chatterjee
Hi Prof. James, I need to prove that a construct is second order reflective construct having two first order factors but when i run it, the factor loadings doesn't show for both first order constructs (even if i constraint both of the two factors' path to 1) but the factor loading for each of first order construct's items is ok (above 0.7, items were deleted below 0.7). Squared multiple correlations for both the first order factors are both negative (-0.028 and -0.059) and also in notes for group/model "The following variances are negative" error occurs which shows that second order reflective construct also has a negative variance of -.0.036. How to fix this frustrating issue and how to prove it a 2nd order reflective construct? Plzz help.
AMOS does not like to run properly if there are no covariances between latent variables. This means that in addition to your 2nd order factor, you would need another latent variable to covary with it. This should fix the issue.
Dear Dr. Gaskin, Thanks for your wonderful videos. I am still confused about when to use reflective second order construct vs formative second order construct? Could you shed some light?
***** It all depends on the underlying factors. If the underlying factors are all the same sort of thing, then it is reflective. If the underlying dimensions are unique dimensions of the 2nd order factor, then it is formative.
James Gaskin thanks for your input.do you suggest any threshold correlation between the dimensions to determine " the underlying factors are all the same sort of thing" ?
***** Really it should be based on the construct definitions and conceptual correlations. If I had to put a statistical threshold on it, I would recommend going with a strong AVE and CR, so that all the reflective loadings averaged out above 0.700.
Dear Sir, Thank you for your very useful video! I have a problem in my thesis about second order factor and can not find an answer to it so far. There is a second-order factor including four first-order factors, and can I test its impacts on other factors directly but without any factors influencing on it? (which means this second-order factor is exogenous that no arrows at it). If so, how can I differentiate the measurement model and structure model??Thank you very much for your help!!
The way SEM works, the loadings on other factors are essentially the same in meaning as the loadings on the dimensions of that factor. So, your interpretation will be normal.
Hi..Sir..Thank you so much ...your videos are so helpful. I had a doubt..what if in case of model fit , it suggests us to covariate between error terms of two different dimensions of the same factor...in..modification indices..Can we do that? Thanks.
Hi James... Many thanks for your videos...I appreciate it Question...I did a second order factor for my independent variables (3 independent variables under one main independent variable).....after that, I did a data imputation...got the results (composite variables)....should I include the 3 independent composite variables with the main composite independent variable...or should I add only the main composite independent variable in amos....Thanks :)
+James Gaskin heey :) many thanks...I have a final question please....my composite variables are interval (scale) and I am trying to use the cross tab on spss to measure the strength of association....I see that on the cross tab there are options available for only nominal and ordinal...should I change to ordinal? or is there another way? thanks again :)
Thank you so much for all your videos. These are great. I have a model with second order formative independent construct( composed of four variables), a mediating variable and a dependent variable. However the DV is dichotomous variable. Can smartpls handle dependent variable that is dichotomous ? If not do you know how to run regression model with formative construct.
Hi James. Thank you for awesome lesson. Your videos are really helpful. Can i get the full reference of this video as I have to put it on my dissertation. Thank you.
Here you go: How to cite Gaskination resources : IEEE TPC PLS article: Paul Benjamin Lowry & James Gaskin (2014). “Partial Least Squares (PLS) Structural Equation Modeling (SEM) for Building and Testing Behavioral Causal Theory: When to Choose It and How to Use It,” IEEE TPC (57:2), pp. 123-146. Wiki: Gaskin, J., (2016), "Name of section", Gaskination's StatWiki. statwiki.kolobkreations.com RUclips videos: Gaskin, J., (Year video uploaded), "Name of video", Gaskination's Statistics. ruclips.net/user/Gaskination Stats Tools Package: Gaskin, J., (2016), "Name of tab", Stats Tools Package. statwiki.kolobkreations.com Plugin or Estimand: Gaskin, J., (2016), "Name of Plugin or Estimand", Gaskination's Statistics. statwiki.kolobkreations.com
hey how do u handle formative constructs in AMOS? how abt moderation between formative and reflective constructs, and moderation with second order constructs? any videos or links wud be really helpful, thanks for the videos, they are helpful and nice!
Hello Doctor Gaskin, For some reason, when I run the test with my second order factor (or first order), only the correlations between factors appear :( Did I omit something? (I use factors and their observed variables, everything runs great) Thank you for your wiki and all your comments in Q&A! Nico
Yes, the difference is that the structural model includes regressions between latent constructs, whereas the measurement model only includes covariances. The only exception to this is from the 2nd order factor to the first order factors.
Sir, I have concern about the demographic variable that you have added with covariance like age, se, yrs. Are they Moderator in your Conceptual Model? I have a categorical moderator so confused how to include it in SEM Model or Path Analysis although I have conducted moderation analysis in SPSS and AMOS (tested significance and z-score with stats too package). I will be really grateful to get your reply.
The ones I added are just control variables. Not moderators. Categorical moderators should be included in multigroup analysis: ruclips.net/video/w5ikoIgTIc0/видео.htmlsi=6TsM_zCCNBCvyJHG
If using a covariance-based software (such as AMOS), the exogenous variables (independent variables) should all be covaried. You cannot covary endogenous variables. However, you can covary their errors. In some software, such as Mplus, the errors of same-level (e.g., mediators) variables are covaried. This is optional in AMOS, but justified.
Hello sir, thank you for the videos. It is really helpful. Sir, i am stuck in my model. Actually i am having 4 high order constructs and i am trying to examining the relationship between them. I have taken the standardized scale for all these 4 constructs. When i run cfa, my convergent validity is perfectly fine but there is no discriminant validity as there is coming high correlation more than .80 between these constructs. Can you please suggest me how to rectify this problem? Thank you sir
This might be an indication of method bias. You can try to add a common latent factor (as shown in the CMB videos) to extract the variance common to all indicators. That should improve discriminant validity if it doesn't break your model. You can also try to reduce your model by one level. You can do this by first modeling all the lower order dimensions without the 2nd order factors. Then impute factor scores (see video called "impute factor scores in amos") and use these new factor scores as the indicators for the 2nd order factors (which will now be modeled as first order factors with factor scores as indicators). Hope this helps.
Hi James. Do you have a video for a handling third order construct? My IV is a third order construct with three subdimensions. One of the subdimensions has further five subdimensions. I want to validate the construct and study its impact on the DV. Please refer me to a helpful video. Another question. Should the subdimensions have equal number of items? If the items measuring one subdimension are four and for another eight, is that a problem?
Doing third order is same as doing 2nd order, but with just one more level. As for the even number of dimensions, it is recommended somewhere (not sure where) that equal/balanced is best, but I can't remember why.
James, when I run my CFA, the solution comes out just fine. My advisor asked me to try running with a second order factor, as well. When I do so, however, I get a negative error variance and a regression weight >1. I have no clue what could be causing this, especially given the original CFA worked fine. All I did was add in the second order latent factor exactly as you've described. Any thoughts?
This can sometimes happen when you only have two first order factors for the second order factor. If this is the case, you might need to either: 1. constrain the two regression weights (between 1st and 2nd order) to be equal and then constrain the variance of the 2nd order to be 1. Or, 2. constrain the error variance to be a small positive number (like, 0.05). I recommend the 2nd option first.
James Gaskin I have three first order factors. Wondering if maybe something is up with the data? Or does it just mean that the model doesn't fit? Thanks for your help!
Dear Dr. James, i have small query regarding loading values (Path estimate values) from first order factors to second order factor, what is the ideal value of the coefficient? is there any standard acceptable limit.
If we are talking about standardized loadings, then it is the same as for a first order indicator. Ideally you want it averaging out above 0.700, but they can be as low as 0.500 and still sometimes result in good enough quality criteria.
Hi James, Many thanks for the video on 2nd-order factor. I am having a little trouble with running the reliability/validity tests in my model as all my first-order factors are related to a 2nd order factor and I only have two 2nd order factors in the entire model. As there is only 1 covariance line that can be drawn between the two, I cannot use the excel package to extract reliability. Do you have any pointers for me as to what I can do?
Hi James, Thanks for all your wonderful videos, which enhances our knowledge. I have a question regarding second order factor in SEM (hypothesis testing): How to take care of Heywood case in second-order factor. I can't set a parameter to the new construct I created. Whereas a first order factor shows Heywood case on one of its items
Thank you for your reply. I had checked the link you have provided. For the step of aaa (same string constraint) for the 1st level construct (factor). There is no box for variance in it. If possible, can you please help me with it. I uploaded the SEM model image on google photos and the link is: photos.app.goo.gl/bb51HolmmYlaqupR2
Hi James, aside from my other issue, I have a question. can I have error term for 1st order construct. Only this model do not give Heywood case. Model fit like others is good. I haven't checked CR, and AVE for this model. Please model in the link photos.app.goo.gl/WHA4i3nI1WZTbbR83
Click the down arrow (next to the red up arrow) before double-clicking the variable. The parameters box is disabled when the model is displaying estimates.
The way you have drawn this one is essentially correct, except that F1 does not need an error because it is not endogenous (i.e., it doesn't have arrows pointing into it).
Hi James, I run second order factors on a construct that has five latent variables. It turned out that 2 out of 5 variables have factor loading less than 0.5 (but the p values are significant). What does it mean? Does it mean that those variables which factor loading less than 0.5 should not belong to the 'parent' construct and therefore should be removed? Thx
This means that the 2nd order factor is probably formative, not reflective. In this case, it cannot be modeled accurately in AMOS. You would need to use something like SmartPLS. Here is a video about how to do such a model in SmartPLS 3: ruclips.net/video/LRND-H-hQQw/видео.html
Hi James, I followed your steps and the path diagram seems to be working. Nevertheless, I see no p-values and SE in regression weights table for the second order constructs... They are just missing. Any ideas? Eugene
+Eugen Bogodistov Oh, for the 2nd order factors? If you have constrained those paths, they won't receive any values. That would be the only reason that I can think of.
Hi Eugen, I am facing the same problem and would highly appreciate it if you could share whether you found a solution. Furthermore, AMOS refuses to display standardized estimates for the second order factor. Thanks a lot in advance!
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!
As I show in this video, you really only need the one constraint. But if yours does not work like that, then do what you can to make it work (remove both constraints and add a variance constraint?).
Maestro James, muchas gracias por el video, me ha sido muy útil. Siga compartiendo su conocimiento. Un abrazo. Soy de Perú.
Great video again from you
Thank you for building young and old researchers alike
Hi James, what should Immers match with?
I just jumped from the video CFA part 1 to this video. I do not understand how to name the second order variable (le.g. Immerse). Thanks
I can do it now. I forgot to constrain 1 in the arrow
I just wanted to say thank you for all of your video's and unbelievably helpful resources :)
Hi James, how do I come up with the name of the second order construct (e.g. immerse in this example)?
It seems like there is 1 step that I missed because there is an f before constructs, e.g. fFI, fTD
My data contain Credibility and Informativeness which belong to Utilitarian. I have the data of Credibility and Informativeness only. And when I put the name Utilitarian as a higher order of Credibility and Informativeness, Amos failed to do the imputation of factor score as Utilitarian is missing from my the Spss file.
Thanks again.
The naming is completely up to you. It doesn't require an f before the name. I just put that there because I had another variable already in the dataset named TD. So, just name it something that isn't already taken in the dataset. Then, when you impute the file, make sure to check the new file (usually named like the old file, but with _C at the end).
Good question. As far as I understand it, covariance based SEM doesn't need to distinguish between them. They are simply correlated. So, since the 2nd order factor is comprised of the two lower order factors, we are simply trying to discover to what extent the DV is correlated with the other two first order factors. This is accomplished just fine the way it is modeled. I hope that clarifies.
@catsfancyful
You are correct. AMOS does not handle formative constructs very well. In fact, I've never successfully built a second order formative construct in AMOS. I recommend using partial least squares software for that. Something like pls-graph, or SmartPLS.
Thank you is very helpful, especially the 2nd model where the covariances between the 1st and 2nd order factors were eliminated.
1. yes
2. I would call it 7-factor
3. I suppose so if immerse was included as a subfactor of the new latent factor.
4. You can analyze the correlation using the double arrow. It is just the standardized value. Or you can use a single arrow if you want to also get the r-square (squared multiple correlation in the output tab), which is more often used for evidence of predictability.
I can't find words to express my thoughts. Thank you! Thank You!. THANK YOU! Hope to take part in your researches (as I am very fond of doing academic researches)
Dr. James thank you very much for your cooperation.
Yes, you can try constraining the two factor loadings to be equal (name both of them "a" or something like that). That would be my first approach. I think you may also be required to constrain the variance on the 2nd order factor to 1 in a case like this. I can't remember.
very helpful. Thanks!
I'm not sure if there is a standard practice one way or the other. If it were me though, I would do both. That way I could show that the first order factors are valid and reliable before using them in the 2nd order factor. And that way no reviewer can ding you for it.
Dear James,
I hope to have your answers to the following:
1. Is it right that the first two (i.e. tFI and fTD) are subfactors of Immerse?
2. Do we call the whoe model as a 8-factor or a 7-factor model?
3. If we draw another latent variable to capture all these factors, will it be called a 3rd-factor model?
4. To prove predictive validity, I wish to analyze the correlation of a latent variable and another latent variable, do I use the single-arrow or double-arrow?
Thanks a lot!
Honestly it just depends on everything else in the model. You might try a reliability test (Cronbach's alpha) to see if you have reliable factors. If not, then you might be able to identify where the problem is.
Thank you very much James! It is so helpful.
It is very helpful, I like it so much. Thank you.
My tool doesn't support SmartPLS. It is mostly for AMOS, although it has a couple tools for PLS (a group differences tool come to mind).
Not sure. Mine runs just fine with only two first order indicators on the second order factor. Try each way and see which one works.
Hi James,
I am looking at your videos recently and it is immensely helpful. Thank you.
A quick question:- how do you measure higher (second order) level construct? Is it mean of all items in the first order?
For reflective 2nd order, there is no direct measurement. Instead, the "indicators" of the 2nd order factor are the first order latent factors. If formative, then you can use a repeated indicator approach, or an instrumental variable (single proxy capturing the main idea).
@@Gaskination thank you, do you mean all items in first order will collectively be assigned to second order construct? It's reflective.
@@ankitatibrewal6842 Yes, if you are using the repeated indicator approach (which is used for formative 2nd order factors)
@@Gaskination its reflective, and I tried creating a new variable by calculating mean of all items collectively and using it . Also just assigning all items collectively to second order variable. Both didnt work. Can you possibly help?
@@ankitatibrewal6842 If it is refective, then it does not need any indicators. The 1st order latent dimensions are the "indicators" in this case. Just as in the video above.
Oh. I use a single constraint on only one path from 2nd order to first order factors. This mimics the first order structure.
Hi James, thank you for your great videos! I'm wondering that when we decide to combine certain factors into a 2nd order factor, is there any criterion we need to follow? Or just making this decision based on theory and our hypotheses? And, after combining 2 factors into a new factor in this video, what should we call the rest 6 factors when we report? Are they 2nd order factors as well (because they covary with the 2nd order factor) or just normal factors? Really appreciate it!
Terrie Chen
1. Decide 2nd order based on theory. You can also decide it based on statistics, if it also makes sense conceptually. For example, if the two factors are highly correlated, and are also clearly conceptually correlated.
2. The remaining factors are still 1st order.
Thanks James! If I make decision based on statistics, should I check the correlation matrix of final solution in EFA step or check the factor correlations of the initial measurement model in CFA?
If it is correlation in measurement model, should this model get a good fit before building 2nd order?
Terrie Chen Check it in 2nd order CFA. During the EFA, you might need to conduct a separate EFA for the 2nd order factor's items.
+James Gaskin 2nd order factor's items are known as the first order factors isn't? Since the first order factors have their own items, should we compute composites so that we can conduct the EFA for the second order?
Thanks so much for your answer. yes it is exactly true If you have four or more dimensions as 2nd order factor you need one constrain, but if you have two dimensions as 2nd order factor with one constrain, it won't run. I have three way: 1. put constrain for each dimensions. 2. put two constrain and a variance constrain. 3. a variance constrain and a constrain for one dimension. which way is true?
Thanks
Hi James. First off I want to say thanks for taking the time to make these incredibly helpful video tutorials! Several have proved invaluable for conducting my analysis thus far. I wanted to check something with you, though, if that's ok.
I initially created a three-factor model - three indicators per factor - and found something approaching acceptable model fit. Then, based on the moderate to high intercorrelations of the three factors, I created a second order factor to which the original three factors are subordinate. However, the analysis on this new, second-order model yields *identical* values on all model fit indices to the original single-order three-factor model. Is this to be expected or must I have done something wrong?
I was originally intending to see wether the first order or second order model has superior fit but this unexpected outcome has put the brakes on this ambition for the time being! Any advice you can give would be greatly appreciated! Thank you.
+Joshua Lord This is perfectly normal. When you deleted the three covariance arrows you increased degrees of freedom by 3. But then you replaced them with regression arrows, decreasing degrees of freedom by 3. So which model is better? Well, if the three covariances among the three first order variables are too high (greater than 0.800), then it is very likely these are simply three different manifestations of the same thing (a higher order factor) and should be modeled together. However, if the covariances are less than 0.800, then we can probably make an argument that they are sufficiently distinct to be modeled as first order factors.
Thanks a lot for your response. This is very helpful. Keep up the good work! :)
You shouldn't be allowed to covary a 1st order construct with its third order parent. Glad you got it to work.
thankyou soo much for this video . its help alot
You are great....
James, thanks for reply.
Regards,
Gustavo
Hi James,
I'm sorry to bother you so much but this is exactly the structure I was thinking of for my model.
The only difference is that in my case I would add three more 2nd order latent variables. Is this possible? If it is, the covariances are supposed to be drawn only among the 2nd order latent variables? Is this correct?
And last, three of my four 2nd order latent variables are the independent variables in the regression analysis while the fourth one is the dependent variable. Is it okay to treat them all equally in there structure of the model?
Thank you very much for your understanding and support.
Sorry again,
Valerio
correct. And you can covary all the 2nd order variables (including DV) together for the CFA. Then, when testing the relationships between IVs and DV, you should uncovary the DV, and then draw regression lines from IVs to DV.
James Gaskin Excellent! Thank you very much! I really appreciate it! Your support has been fundamental to me.
I have just found this interesting paper that explains why higher-order CFA models may not make good conceptual sense. I found it quite interesting. It's a good food for thought.
Here's the link to the paper (if you want to give it a look): www.researchgate.net/publication/236004024_Problems_with_Formative_and_Higher-Order_Reflective_Variables
Thank you again James.
Hello James Gaskin,
I have one last question. I built my model, so I have:
1) three 2nd order independent variable, pointing to their correspondent 1st order variables which in turn point to the observed variable.
2) the three 2nd order factors point also to the 2nd order dependent variable, which in turn points to 1st order variables, pointing to observed variables.
3) I put error terms on the 2nd order dependent and not on the 2nd order independent factors.
Having all this set, I get no output by Amos. Where am I wrong? Is a structure like mine feasible?
Thank you very much again!
Vale Ma All variables (whether latent or observed) that have arrows pointing at them, should have error terms.
James, Brown (2006) and Byrne (2001) raise concerns regarding a second-order factor accounting for only two factors. This is because although the overall model may be overidentified, the second-order factor model itself would be underidentified. Thus, there may be an identification issue with your model, even if this is not flagged in the output. Is this something you considered and, if so, why is this not a problem here? *Please note that I believe your approach to be correct if the second-order factor accounted for three factors.
+Nathan Hartmann I had not considered that issue in this video.
Thanks a lot. These three ways worked but give me different result. my question is which one is true? :)
not off the top of my head. I'd have to go play around with it. best of luck!
Dear James,
Two questions:
1. It is a 7-factor model in your video, with the first factor being 2nd-order. Mine is very similar. If I delete the other 6 factors and just run the second-order model for first factor (e.g. Immerse in yours), it won't run?
2. Comparing unconstrained model and constrained models (i.e. covraince between 2 factors = 1) is a way to test discriminant validity. But when I constrained a second-order (e.g. Immerse in yours) and another (e.g. fJoy), it didn't run?
Thanks!
@Tomahawk1999 AMOS does not handle formative constructs well. The principles behind moderation do not change whether you have formative, reflective, first or second order constructs.
Dear Professor Gaskin
I see in the comments below that someone inquired about formative second-order factors in AMOS, in which you recommended using SmartPLS. However, as my research is covariance-based, I was wondering if a reflective model in AMOS is appropriate when the first-order factors are hypothesised as having an effect on the second-order construct? If not, I was wondering if you know of any books or papers that explain why AMOS cannot deal with formative second-order models?
Looking forward to hearing from you.
Modeling a formative factor as reflective will have bad consequences. The Barbara Byrne AMOS manual probably says something about formative models in AMOS. I know my article with Paul Lowry about PLS talks about it briefly.
There could be several reasons why it is not running. Amos is very picky. You are welcome to send it to my email address and I will take a look: james. gaskin@ byu. edu
Hi James,
first of all: a big thank you to your excellent tutorials. You actually made me switch from Stata to SPSS/Amos :)
I followed your tutorial building a 2nd order construct in AMOS which by itself runs nicely regarding all indices.
The model I am trying to test for AVE, CR etc. has 5 latent variables with 6 items each and one latent 2nd order factor.
At this point, there are no other latent variables or other variables linking into the model.
When I want to test for AVE,CR etc. with your Excel Tool, I run into a problem (have already ruled out any violations against your assumptions regarding naming): After copying both tables into your validation tool, i get "Run time error 91: object variable or With Block Variable not set"
Could you tell me what the reason might be? Testing the 1st order model with the same variables (excl. the new 2nd order variable) worked well in your excel.
I found it strange that my correlations table of the 2nd order analysis only contains correlations of the error terms - could this be the reason and if so, what can I do to resolve the problem?
Any help is much appreciated! Your videos alone were already super helpful!
+Spss Phd You have identified the issue. If there is only one highest order construct (in this case, a 2nd order construct with five underlying dimensions), then it treats it as if there is only one latent variable in my tool. To calculate the validity of this type of model you will need my old stats tools package. You can email me at james.gaskin@byu.edu to get that.
I would have to see your model, but my guess is that you might have neglected to constrain one of the regression weights between the 2nd order and 3rd order variable, or between the 2nd order and 1st order variables.
Nice video. Thank you so much. Could you explain how we should draw models with formative second order constructs in AMOS?
Hi, James. I have a question about 2 order CFA. I have a data for FLOW construct, which was validated by its author using European data. It was posited as a 3-dimensional single factor with (1) IWM (5 indicators), (2)WE (4 indicators), and (3) AB (4 indicators). Here I have collected Chinese data for it (N=869). And I did EFA for it before CFA, as I think maybe the pattern metrics are a bit different in Chinese context. And I arrived to a clean pattern after I delete IWM4 (very poor loading) and IWM3(cross loading). But then I checked the factor correlation where the discriminant validity problem popped up (correlation >.70). Therefore, I tried to model the construct FLOW as a second order factor to see if doing so would solve some problems. Surprisingly, for the first time of my experience, I got exactly same model fit indices (CFI, TLI, RMSEA, etc...) for FLOW as a second order factor VS a 3-dimensional single factor...I assume that a second order factor is better than the single factor as correlations between IWM and AB, AB and WE are above .70...So, I would like to know if there is anything wrong with my procedure? How can we potentially explain it? Thank you very much! :-)
The model fit won't have changed because you simply replaced the correlation arrows with regression arrows. Therefore you have the same degrees of freedom and you are still accounting for the same relationships. I agree that if you have discriminant validity issues between first order dimensions, then if there is theoretical reasoning to connect them, just make a 2nd order factor. So, you have done it correctly.
Hi Dr James, in the article "Operationalizing Multidimensional Constructs in Structural Equation Modeling: Recommendations for IS Research"' the authors have suggested that for higher order reflective dimensions, a separate CFA should be run where as Hair et al (2007) recommends a "'pooled CFA"' approach even with higher order constructs having dimensions. The video above indicates you have been using pooled CFA concept. if so can you offer some references?
Hair et al 2007 I suppose is a good start. I can't think of any others off the top of my head, but a quick search of google scholar for higher order factors will probably turn up some good results.
Thanks so much James! I will email it to you.
Hello! thank you for your quick answer. No the regression weights were fine. I may have found the cause though. I am covarying one 3rd order variable with one first order variable. Am I not supposed to do that? I tried to add a variable, which I covaried with the other two and it ran without a problem. But when I took it off it didn't.
Dear Prof. Gaskin
I have some questions about second order CFA and CFA
At first CFA:
1) Is there any normality Prerequisite for CFA?
2) If yes How could we do a CFA or SEM with non-normal data with Amos?
Second order CFA:
should we apply a separate CFA for first order components or not ? if yes should we calculate validity and reliability for them ?
Yes. Your data should be normally distributed if you are using maximum likelihood estimation. Bayesian doesn't require normality. I would do 2nd order CFA with all other constructs as well. Calculate validity and reliability at the highest level.
Many thanks for you response . Can I use boot-strap instead of bayesian for non-normal data? because there is no Bayesian test in Amos and boot-strap doesn"t require normal distribution too.
If its possible for you, Please make a video about sem with Bayesian or bootstrap.
In Amos, go to Analyze, Bayesian Estimation.Bootstrapping can help. I have videos that show bootstrapping for mediation analyses.
Dear Prof.Gaskin
Thanks for your response . I have some more questions.
1)How should I do CFA with Bootstrap.
2)How we should do Moderating with Bootstrap?
Best regards
The tutorial is interesting. May I please request you to describe how to conduct a Moderation or moderated mediation in case of a second order factor(s) as the exogeneous variables.
In this case, I would recommend using a software that makes these combinations for you, such as SmartPLS. Doing latent interactions in AMOS is very difficult and unwieldy.
Hello! Thank you for the video and the Excel document. This is so helpful! I tried to test the validity of a model including a 3rd order factor but it fell. Would you have any recommendation? Thank you in advance.
Hi James,
How many constrain (I mean put “1”in regression weight) I need when I have 2nd order factors for every construct. Also the other question is I should put “1” in variance (through object properties) for every construct which have dimension or not? because if you have one construct with two dimension, it cannot run if you put just one constrain. you should put constrain for every dimension or put for variance. Is it true?
I know I am in first step to learn Amos.
Thanks so much.
Dr. regarding the last 10 secs of the video, It would be brilliant if run CFA with second-order factor solving teaching public how to solve the second order factor validity problems
I'll see if I can add that to the next series of videos.
Hello James,
Thank you for your amazing videos.
Please, I have a question regarding the reliability test (Cronbach's alpha), I have a 2nd order factor in my model which contains 2 sub-factors. How can I calculate the Cronbach's alpha for the 2nd order factor in SPSS? Should I put all the items of the 2 sub-factors together in the box to calculate Cronbach's alpha for 2nd order factor as a whole? or calculate Cronbach's alpha for each of the 2 sub-factors separately?
(p.s. I downloaded your plugin and I found CR value for the 2nd order factor, but I could not find Cronbach's alpha value)
Thank you.
If you create factor scores for the first order dimensions, you could then use those factor scores in the Cronbach's alpha in SPSS.
watch my video called: "Iteration Limit Reached in AMOS" This shows how to find problematic items (ones with really high standard errors, or ones that load over 1.0) and how to fix them. Hope this helps.
James, thanks for sharing this information.
I have one question about second-order models. When you have two or three first orders factors under a second order one, this part is not overidentified (as recommended). I saw one solution on Byrne's book, but it is not working. Do you know how to setup model to get overidentified with two or three first order under a second order factor?
Thanks,
Gustavo
Thanks very much, as always. What about a second-order formative construct? I understand AMOS doesn't handle this very well? Any thoughts would be appreciated.
Hi James
Thanks for another excellent video. If one was interested in the relationship between the lower (1st) order facets and the endogenous variable but still wanted to account for the 2nd order structure of the model (e.g. looking at how extraversion facets loading on a global extraversion factor predicted salespeoples' job performance) is it possible to have the regression arrows going from the 1st order unobserved variables to the outcome...or will the only variance left in these lower variables be the variance unexplained by the 2nd order factor? If not, could you also do a comparison between 1st order and 2nd order direct effects on the DV using this approach?
Regards
Tom
Tom Bailey If the first order factor is part of a 2nd order factor, then you should not draw arrows between that first order factor and any other factor (aside from its 2nd order factor). Instead, create two models: one with the 2nd order factor and one without.
Hello James, thank you for the video.
I want to ask you a question. I used 2nd order
Why is the result of model fit in my measurement model and structural model same?
Thank you for answering my question. However, my model doesn't show any negative error term or any standard error that is greater than 1. But I noticed that when i correlate the M.I., the regression loadings as well as the correlation loading will be exceed 1. Is there any method that can allow me to adjust the M.I. while the loadings can stay below 1?
Hi, Your videos are awesome! I'm using them to conduct a second order SEM and I havequestions/clarifications. Can you calculate the composites to be used for the structural model? If so, should you include the first order composites in the structural model since no regression lines are linked to them? Grateful for any help! Thanks again!
Sabeeha Mohamed The composites will be created for whatever latent constructs are covaried during the CFA. This means the higher order constructs. In this case, in the structural model, include the higher order constructs, but not the dimensions (lower order constructs) that were part of it.
James Gaskin Thanks alot for your feedback! Really appreciate it! :)
Hi James. Thank you for these videos. They are super helpful. I do have a question. I have a second order factor with 4 first order factors (as suggested by previous literature). When I check the loadings of the first order factors on the second order, there is one first order factor that have very low loading (0.1..) while the other 3 have 0.7-0.8. My question is what I can do. Is there any other way other than removing that first order factor? If I remove it, I feel it is hard to justify why mine is different from literature. Thanks so much.
+nganhavy It begs the question whether the 2nd order factor is actually formative instead of reflective. If formative, then the difference in loading is not a problem, but it should be modeled formatively and in PLS, not AMOS. If actually reflective, then there may be underlying issues with the items from that factor. If you plan to model it reflectively, I would perhaps recommend changing this to keep only the three strong dimensions, but then retain the other dimension outside this factor. It may work well as a predictor or covariate.
+James Gaskin Thank you. Your advice is very helpful.
Hi Dr James,
Please i need your help on this : i have a second order contruct and i find good regression weights between 1st and second order factors but before i found low correlation between 1st factors less than .5 (from .102 to .41) to go for a second order ! i'm confused since in the theory the 2nd order exist and reflect 1st order factors.
Any advice ?
thanks a lot
Dr Gaskin,
I have a quick question. In addition to constraining one of the paths between the first order factors and the second order factor do we have to also constrain the variance of the second order factor (in this case immers) to 1? Some books state that both should be applied. Thank you in advance!
AMOS will run without that additional constraint applied. So, you can use either the variance constraint or the path constraint. Either will work and the results should be about the same.
Thank you so much Dr Gaskin. I have one more question and would be grateful if you could help.
If we have a latent model, how do we compute the composite factors without doing the CMB test and imputing the factors? Let's say I have a latent model and there is no CMB, so I move on with my original data set but I still need the composite factors. Is there a way to do this?
Thank you so much for your time!
I think I found the video for it. "Imputing factor scores in AMOS". It does not say anything about CMB, so I guess that must be the right one. Thank you!
@@aozgun9 correct! :)
Dear James, how can i perform chi-square variance test of the limited and unlimited measurement patterns. Its necessary to construct discriminant validity. I need chi-square and “d. f.” results. Thanks too much your helps.
+Cemil Ceylan Try this video: ruclips.net/video/6j4_ZrkCxTc/видео.html
Thanks too much. Your videos are great.
James - Over the last several months, I've come across a few published CFAs that seem to be using higher-order factors to hide problems with discriminant validity. In one case, the correlation between two of the latent factors below the higher-order factor was .93 (!), but of course this wasn't reported, because the higher-order factor allowed this to be hidden. Interestingly, adding a higher-order factor didn't change the fit of the model at all - all it did was allow the reporting of the model to omit the correlations between latent factors. To me, this seems like statistical obfuscation to hide genuine problems with the model. Is it ever appropriate to use a higher-order factor when these discriminant validity problems exist? Do you have any recommended reading that might shed some light on this? Thanks!
This is absolutely fine if the two underlying constructs truly are subdimensions of a higher order construct. For example, I've worked with several students who want to assess job satisfaction and life satisfaction. However, these are so highly correlated, that we always run into discriminant validity issues. What we usually end up doing is placing them together into a second order factor called "overall satisfaction". Now, if we were combining something like satisfaction with productivity, that would be much different and would take some really good wordsmithing to convince readers that you really have a 2nd order factor.
James Gaskin Hi James: What are the practical implications of that move? Say I'm running a linear regression: Do I just include one measure (the second order factor) or do I include three measures (the first order factors)? In Rob's example, the two first order factors are correlated at .93. I would have concerns about multicollinearity.
Jake Jensen Multicollinearity would not be an issue if these were modeled as a reflective 2nd order factor. Then, when doing the regression, the causal path is only between the 2nd order factor and some DV.
James Gaskin That makes sense to me! Thanks James!
Hi James, I have followed your guidance and 2nd order factor analysis seems to fit to mine. Thank you. I have one more question. If I want to test the impact of fIMMERS on DV, second factor analysis is the way to do right? but what if I want to see as well the impact of fTD which is subconstruct of fIMMERS on DV? then should i eliminate fIMMERS and validate measurement model and conduct structural model analysis?
If you want to do both, then just make two models. That is perfectly acceptable.
James Gaskin Thank you so much. It has been a great help for my master thesis. I would be totally lost without your tutorials!!
Ok. Thanks .. Its same as Burnout has three dimensions i.e. Emotional Exhaustion, Depersonalization and Lack of Personal Accomplishment .. BO= EE+DP+LOPA ... Do you think AMOS is smart enough to understand this equation ... I have seen one of you tutorials on smartPLS (if I am not wrong?) ... Where you explained how a multiple dimension variable can be represented as a single dimension ...It was quite complex but it seemed some how the same as 2nd order factor?
Amos does not handle formative constructs. You would need PLS software for this.
Hi, thank you very much. Your tutorial really helps for my academic studies. However, in this video, i dont understand how you manage to run the last model with amos? Because there are both formative and reflective structure as i see, we cannot run formative structure in amos. Thank you.
I have only reflective factors in this video. The second order factor is reflective in this case.
This typically should not be done. AMOS is build on a covariance-based algorithm which doesn't handle formative constructs well. I would use SmartPLS.
Dear James,
I would like to know about the correlation between 2 second order factors (both are exogenous in the model). If it is 1.007, does it indicate discriminant validity is not achieved? Or can it be interpreted as (1) even though they are independent variables, they still have linear relationship (1 of the factors serves as independent and dependent at the same time) or (2) may have causal effect. So, discriminant validity is not affected where there will be no issues on multicollinearity.
+Nitha A A correlation greater than 1.00 indicates an estimation error. It is also highly likely that they are strongly similar and may be part of a higher order factor.
Given this situation, should I check the MI and delete/constrain the redundant items?
+Nitha A yes. that is one way to do it. Or you can do a multicollinearity test to see which ones have overlapping variance. Or you can just look at the wording of the variables and see if they are very similar.
+James Gaskin Thank you so much.
Thanks for your video!can i ask you something? I want to do CFA at a Second Order Construct with 2 factors, but as you know in this case we dont have identification. What should i do?Constrain the 2 factor loadings?constrain the variances?something else?thanks!
Dear James,Thanks for the videos! They are great. Currently I have a problem running SEM with panel data. I searched but cannot find any useful resources. May I please ask you to recommend some literature on SEM with panel/ pseudo panel? Thank you~
+Si Qi Song What do you mean by panel data? Who is responding to your survey?
Dr. James Gaskin, Thank you for the You tube video however when I try to run my data I do not seem to find / see the excel sheet where you paste the correlation and standardized regression weights you copy from the measurement model
Joanita Kataike It is on the homepage of my wiki: statwiki.kolobkreations.com
Thx James I got it.
Hi Sir,
I'm receiving this problem while running my model
"The model is probably unidentified. In order to achieve identifiability, it will probably be necessary to impose 1 additional constraint"
This means that one of the factors does not have a parameter constraint. Notice that all of the factors have one of their indicator regression weights constrained to equal 1.00. So, make sure even your second order factors have one regression weight equal 1.00. The other option is to constrain the variance of the factor to 1.00.
The model consist of 5 factors ( 4 first-order) and (1 second-order) the loading of 2nd order was -.56 and .78. when i compute the value of CR and AVE for all factors the value of 2nd order was CR lees than 0.3 and AVE lees than 0.5 because i have negative and positive value of loading. in this case what i should do.
This means that the two dimensions (lower order constructs) of the 2nd order construct are inverses of each other. You may consider reversing the negative one so that it is in the same direction as the other dimension. To do this, subtract all indicator values for that construct from m+1 where m=the maximum possible response. So, if you are on a 5-point Likert scale, then you would subtract from 6. Also, you'll need to reverse the construct label. So, if before it was "sad" now it needs to be "happy", because an increase in the scale would now represent a decrease in the original construct.
I received a notification of another reply, but I don't see it here. I will reply anyway. The first order dimensions need to be in the same direction. So, if it is confidence and anxiety. Perhaps change anxiety to calmness or serenity.
So I collected data where one of variables loaded on 2 factors -which was fine as theoretically there are 2 dimensions to it. However I don't intended to use it that way so I used it as 1 Amos. everything worked fine. but as a check, a professor asked me to do a 2nd order to make sure the 2 dimensions are equally useful/contributing enough.... I did the initial steps as per the video and 1 dimension has a value of .60 and the other .89. does that sound acceptable? Thanks!
+Val_C Sounds fine. You just want the loadings to average out to above 0.700
Hello James,
does a second order factor mean that all measurement items should be merged together to represent only one factor?
How do I determine if I should choose a second order factor model? should I look at the modification indices of the two models (the model with first order factors only and the other with second order factor) and judge on the validity of both models?
will this be enough to say that a second order factor is a plausible choice?
hishonline Usually the choice of a 2nd order factor is done based on the conceptual level. If the two factors are really just two dimensions or manifestations of a higher order construct, then they are possibly part of a 2nd order factor. Often these first order factors are highly correlated and may have difficulty achieving discriminant validity.
Hello Dr. Gaskin
While running the CFA you didn't add tthe error term or unobserved variable to the construct immerse. While in SEM error term was added. Don't we need to add it while doing the CFA also.
Kindly see if you can address my confusion. I will be more than obliged for your help.
Regards,
Dr. Namrata Chatterjee
Only endogenous variables require error terms. In the CFA, it is not endogenous.
Got it. Thanks a lot for addressing the query Sir!
Hi Prof. James, I need to prove that a construct is second order reflective construct having two first order factors but when i run it, the factor loadings doesn't show for both first order constructs (even if i constraint both of the two factors' path to 1) but the factor loading for each of first order construct's items is ok (above 0.7, items were deleted below 0.7). Squared multiple correlations for both the first order factors are both negative (-0.028 and -0.059) and also in notes for group/model "The following variances are negative" error occurs which shows that second order reflective construct also has a negative variance of -.0.036. How to fix this frustrating issue and how to prove it a 2nd order reflective construct? Plzz help.
AMOS does not like to run properly if there are no covariances between latent variables. This means that in addition to your 2nd order factor, you would need another latent variable to covary with it. This should fix the issue.
Dear Dr. Gaskin, Thanks for your wonderful videos. I am still confused about when to use reflective second order construct vs formative second order construct? Could you shed some light?
***** It all depends on the underlying factors. If the underlying factors are all the same sort of thing, then it is reflective. If the underlying dimensions are unique dimensions of the 2nd order factor, then it is formative.
James Gaskin thanks for your input.do you suggest any threshold correlation between the dimensions to determine " the underlying factors are all the same sort of thing" ?
***** Really it should be based on the construct definitions and conceptual correlations. If I had to put a statistical threshold on it, I would recommend going with a strong AVE and CR, so that all the reflective loadings averaged out above 0.700.
Dear Sir, Thank you for your very useful video! I have a problem in my thesis about second order factor and can not find an answer to it so far. There is a second-order factor including four first-order factors, and can I test its impacts on other factors directly but without any factors influencing on it? (which means this second-order factor is exogenous that no arrows at it). If so, how can I differentiate the measurement model and structure model??Thank you very much for your help!!
The way SEM works, the loadings on other factors are essentially the same in meaning as the loadings on the dimensions of that factor. So, your interpretation will be normal.
Hi..Sir..Thank you so much ...your videos are so helpful. I had a doubt..what if in case of model fit , it suggests us to covariate between error terms of two different dimensions of the same factor...in..modification indices..Can we do that?
Thanks.
Yes, for dimensions of the same factor, covarying errors can make sense.
@@Gaskination Thank you so much sir...that helps...Your videos are really blessings for beginners like us..
Hi James... Many thanks for your videos...I appreciate it
Question...I did a second order factor for my independent variables (3 independent variables under one main independent variable).....after that, I did a data imputation...got the results (composite variables)....should I include the 3 independent composite variables with the main composite independent variable...or should I add only the main composite independent variable in amos....Thanks :)
+AJ Badwan Add only the main one (the 2nd order variable), as it is comprised already by the other three.
+James Gaskin heey :) many thanks...I have a final question please....my composite variables are interval (scale) and I am trying to use the cross tab on spss to measure the strength of association....I see that on the cross tab there are options available for only nominal and ordinal...should I change to ordinal? or is there another way? thanks again :)
+AJ Badwan Interval is Ordinal. SPSS uses funny names. You can change them to ordinal.
Thanks again :)
Thank you so much for all your videos. These are great. I have a model with second order formative independent construct( composed of four variables), a mediating variable and a dependent variable. However the DV is dichotomous variable. Can smartpls handle dependent variable that is dichotomous ? If not do you know how to run regression model with formative construct.
Yes it can. But usually we do logistic regression for dichotomous DVs.
Hi James. Thank you for awesome lesson. Your videos are really helpful. Can i get the full reference of this video as I have to put it on my dissertation. Thank you.
Here you go:
How to cite Gaskination resources
:
IEEE TPC PLS article:
Paul Benjamin Lowry & James Gaskin (2014). “Partial Least Squares (PLS) Structural Equation Modeling (SEM) for Building and Testing Behavioral Causal Theory: When to Choose It and How to Use It,” IEEE TPC (57:2), pp. 123-146.
Wiki:
Gaskin, J., (2016), "Name of section", Gaskination's StatWiki. statwiki.kolobkreations.com
RUclips videos:
Gaskin, J., (Year video uploaded), "Name of video", Gaskination's Statistics. ruclips.net/user/Gaskination
Stats Tools Package:
Gaskin, J., (2016), "Name of tab", Stats Tools Package. statwiki.kolobkreations.com
Plugin or Estimand:
Gaskin, J., (2016), "Name of Plugin or Estimand", Gaskination's Statistics. statwiki.kolobkreations.com
Thank you so much.
hey how do u handle formative constructs in AMOS? how abt moderation between formative and reflective constructs, and moderation with second order constructs? any videos or links wud be really helpful, thanks for the videos, they are helpful and nice!
Dear James, please explain the direction of the arrow marks in the second order .IFI to glimmers, should it be reversed in direction
I have included it as a reflective dimension because in AMOS we do not model formative factors.
@@Gaskination Thankyou
Hello Doctor Gaskin,
For some reason, when I run the test with my second order factor (or first order), only the correlations between factors appear :( Did I omit something? (I use factors and their observed variables, everything runs great)
Thank you for your wiki and all your comments in Q&A!
Nico
I'm not sure. I'd have to see it. If you want to send it to me (the .sav and .amw files), here is my email: james.gaskin@byu.edu
that comment is from me... I was logged into my other channel.
It seems that the measurement model and the structural model looked pretty the same. Pls kindly describe the different based on your video. Thanks
Yes, the difference is that the structural model includes regressions between latent constructs, whereas the measurement model only includes covariances. The only exception to this is from the 2nd order factor to the first order factors.
Sir, I have concern about the demographic variable that you have added with covariance like age, se, yrs. Are they Moderator in your Conceptual Model?
I have a categorical moderator so confused how to include it in SEM Model or Path Analysis although I have conducted moderation analysis in SPSS and AMOS (tested significance and z-score with stats too package). I will be really grateful to get your reply.
The ones I added are just control variables. Not moderators. Categorical moderators should be included in multigroup analysis: ruclips.net/video/w5ikoIgTIc0/видео.htmlsi=6TsM_zCCNBCvyJHG
What is the rule on drawing correlations between constructs in the structural model? Please clarify. Thank you
If using a covariance-based software (such as AMOS), the exogenous variables (independent variables) should all be covaried. You cannot covary endogenous variables. However, you can covary their errors. In some software, such as Mplus, the errors of same-level (e.g., mediators) variables are covaried. This is optional in AMOS, but justified.
Hello sir, thank you for the videos. It is really helpful. Sir, i am stuck in my model. Actually i am having 4 high order constructs and i am trying to examining the relationship between them. I have taken the standardized scale for all these 4 constructs. When i run cfa, my convergent validity is perfectly fine but there is no discriminant validity as there is coming high correlation more than .80 between these constructs. Can you please suggest me how to rectify this problem? Thank you sir
This might be an indication of method bias. You can try to add a common latent factor (as shown in the CMB videos) to extract the variance common to all indicators. That should improve discriminant validity if it doesn't break your model. You can also try to reduce your model by one level. You can do this by first modeling all the lower order dimensions without the 2nd order factors. Then impute factor scores (see video called "impute factor scores in amos") and use these new factor scores as the indicators for the 2nd order factors (which will now be modeled as first order factors with factor scores as indicators). Hope this helps.
@@Gaskination thank you sir for your suggestions. I will try these.
Hi James. Do you have a video for a handling third order construct? My IV is a third order construct with three subdimensions. One of the subdimensions has further five subdimensions. I want to validate the construct and study its impact on the DV. Please refer me to a helpful video.
Another question. Should the subdimensions have equal number of items? If the items measuring one subdimension are four and for another eight, is that a problem?
Doing third order is same as doing 2nd order, but with just one more level. As for the even number of dimensions, it is recommended somewhere (not sure where) that equal/balanced is best, but I can't remember why.
Thank you.
James, when I run my CFA, the solution comes out just fine. My advisor asked me to try running with a second order factor, as well. When I do so, however, I get a negative error variance and a regression weight >1. I have no clue what could be causing this, especially given the original CFA worked fine. All I did was add in the second order latent factor exactly as you've described. Any thoughts?
This can sometimes happen when you only have two first order factors for the second order factor. If this is the case, you might need to either:
1. constrain the two regression weights (between 1st and 2nd order) to be equal and then constrain the variance of the 2nd order to be 1. Or,
2. constrain the error variance to be a small positive number (like, 0.05).
I recommend the 2nd option first.
James Gaskin I have three first order factors. Wondering if maybe something is up with the data? Or does it just mean that the model doesn't fit? Thanks for your help!
Nicole Boyko It can also happen with three. Try constraining the error variance first. That usually will fix it.
James Gaskin That worked. THANK YOU!
Dear Dr. James, i have small query regarding loading values (Path estimate values) from first order factors to second order factor, what is the ideal value of the coefficient? is there any standard acceptable limit.
If we are talking about standardized loadings, then it is the same as for a first order indicator. Ideally you want it averaging out above 0.700, but they can be as low as 0.500 and still sometimes result in good enough quality criteria.
Hi James,
Many thanks for the video on 2nd-order factor. I am having a little trouble with running the reliability/validity tests in my model as all my first-order factors are related to a 2nd order factor and I only have two 2nd order factors in the entire model. As there is only 1 covariance line that can be drawn between the two, I cannot use the excel package to extract reliability. Do you have any pointers for me as to what I can do?
Yes, I would recommend the plugin instead: ruclips.net/video/ekICmx_qcWg/видео.html
@@Gaskination Thanks!
Dear Dr.James
You showed 2nd order measurement model for reflective measures, would you please guide regarding formative measures in AMOS. Thanks
Amos was not designed to handle 2nd order formative factors. I would recommend SmartPLS for that.
Hi James,
Thanks for all your wonderful videos, which enhances our knowledge. I have a question regarding second order factor in SEM (hypothesis testing): How to take care of Heywood case in second-order factor. I can't set a parameter to the new construct I created. Whereas a first order factor shows Heywood case on one of its items
Here is a video showing how to address heywood cases in AMOS: ruclips.net/video/Vx24KFf-rAo/видео.html
Thank you for your reply. I had checked the link you have provided. For the step of aaa (same string constraint) for the 1st level construct (factor). There is no box for variance in it. If possible, can you please help me with it. I uploaded the SEM model image on google photos and the link is: photos.app.goo.gl/bb51HolmmYlaqupR2
Hi James, aside from my other issue, I have a question. can I have error term for 1st order construct. Only this model do not give Heywood case. Model fit like others is good. I haven't checked CR, and AVE for this model. Please model in the link
photos.app.goo.gl/WHA4i3nI1WZTbbR83
Click the down arrow (next to the red up arrow) before double-clicking the variable. The parameters box is disabled when the model is displaying estimates.
The way you have drawn this one is essentially correct, except that F1 does not need an error because it is not endogenous (i.e., it doesn't have arrows pointing into it).
Hi James, I run second order factors on a construct that has five latent variables. It turned out that 2 out of 5 variables have factor loading less than 0.5 (but the p values are significant). What does it mean? Does it mean that those variables which factor loading less than 0.5 should not belong to the 'parent' construct and therefore should be removed? Thx
This means that the 2nd order factor is probably formative, not reflective. In this case, it cannot be modeled accurately in AMOS. You would need to use something like SmartPLS. Here is a video about how to do such a model in SmartPLS 3: ruclips.net/video/LRND-H-hQQw/видео.html
Hi James,
I followed your steps and the path diagram seems to be working. Nevertheless, I see no p-values and SE in regression weights table for the second order constructs... They are just missing. Any ideas?
Eugene
+Eugen Bogodistov Oh, for the 2nd order factors? If you have constrained those paths, they won't receive any values. That would be the only reason that I can think of.
Hi Eugen, I am facing the same problem and would highly appreciate it if you could share whether you found a solution. Furthermore, AMOS refuses to display standardized estimates for the second order factor. Thanks a lot in advance!
Dear James, what if the independent variable is second order factor, how the treatment will be to the dependent and mediating variables?
+Adelia Wahyu If the 2nd order variable is the IV, then you don't need to go through the step of obtaining latent variable scores.