Here's a fun pet project I've been working on: udreamed.com/. It is a dream analytics app. Here is the RU-vid channel where we post a new video almost three times per week: ru-vid.com/show-UCiujxblFduQz8V4xHjMzyzQ Also available on iOS: apps.apple.com/us/app/udreamed/id1054428074 And Android: play.google.com/store/apps/details?id=com.unconsciouscognitioninc.unconsciouscognition&hl=en Check it out! Thanks!
Thanks Gaskin Sir, You open up my eyes. Now, It's been able to solve the problem of endogenous variable with constraints. your stats tools package, pattern matrix builder and common latent factor connector is superb sir. Of-course this has supported my doctoral work in Tribhuvan University, Nepal. ............Once again Thank You.
Dear James and Sajeeb, thanks you for your contribution to the amateurs in this topic. But i have a few ask... For this method of fixing parameters and fixing variance of latent variables (which solved my problem), how do i justify this in my thesis or in the future paper to publish? in case it should be justified, i have a little experience in this world of research. For the other hand, my fit index are very poor. Chi-square= 1059 df= 342 probability level = ,000 CFI=0,674 NFI= 0,589 TLI= 0,640 RMSA=0,076 What do you recommend ? Thanks...
Dear Dr. James, Thank you very much for these informative videos. I have successfully published many articles with AMOS because of these videos. Thanks again and please keep on uploading your video. I will be expecting! Cheers!
Thank you so much, James,,, you have already saved me multiple times ... I seriously would love to send you a gift from Japan ... Have a wonderful amazing relaxing summer
thank you so so much for your informative video! however i wonder whether i need to report how i fix regression weight greater than 1 in my article.thanks again:)
Thanks for the wonderful videos. Can you please tell me the solution in case it still appears to be greater than 1 even after making the regression weights equal to 'aaa' on both of the lines?
Thank you Dr. Gaskin for sharing the video. It is very helpful! I have a second-order model with 5 factors, each with 3 indicators. Sample size is about 350. The coefficient between the second-order construct and one of 5 factors is 1.01. So, if I use your method, do I constrain all 5 regression weights between the second-order construct and the 5 factors? I did just that and the coefficient came down to .99. Did I do it correctly? Then I try to assess measure invariance. Do I need to keep these 5 constrained regression weights in the model? If so, what's the implication for the invariance analysis? Thank you very much!
Professor James Gaskin, thank you for these videos they are super useful and concise, and clear. I had a question about structural models. If I have a standardized regression weight on a path between an IV and DV in my structural model is this possible and justifiable? In my CFA I have none that are over 1, but in my SEM there is! I am new to this and your help would be much appreciated
Thank you for the video Dr. Gaskin. I'm facing a problem now. I conducted SEM in AMOS (including 1 second oerder factor as dependent variable, 3 mediators, and 2 outcomes, and several covariables). The standardized regression weight from the dependent variable to one outcome is greater than one. What do you think of this problem? And could do please suggest some possible solutions ? Thank you.
Make sure to confirm discriminant validity prior to assessing the causal model. If all prerequisites, such as data screening, EFA, and CFA have passed all validity criteria but there is still a problem, then make sure you have correctly specified your variables as formative or reflective.
I am getting standardized loading of (1.00) in 'Invariance constraint' test. Without constraint, I have dealt with Heywood case in 'Simple Invariance test' as described in this video. But what if value of 1 comes in Standardized Loadings during 'Invariance constraint' test. Because Regression weights are already fixed to "Parameter constraint" like W1, W2, W3, etc but still getting loading of "1" in one of observed variable. Solution?
This is a little bit trickier to fix. You would have to go to the manage models area and specify that those parameters, with their label names, should be equal, or should equal a specific value.
Dear Dr. James, Thank you very much for these informative videos. When use ML it worked but when i use ADF method it did not work. How can handle this problem when i use ADF method for estimation. Thank you.
Hi James, thanks for the videos. Regarding this video, my understanding is that retaining the revised model (where the 2 paths were constrained to be equal) would be justified only if the revised model was not a worse fit of the data than the original model. I am wondering what your thoughts were on this matter.Regards, Craig
Dear James and Sajeeb, thanks you for your contribution to the amateurs in this topic. But i have a few ask... For this method of fixing parameters and fixing variance of latent variables (which solved my problem), how do i justify this in my thesis or in the future paper to publish? in case it should be justified, i have a little experience in this world of research. For the other hand, my fit index are very poor. Chi-square= 1059 df= 342 probability level = ,000 CFI=0,674 NFI= 0,589 TLI= 0,640 RMSA=0,076 What do you recommend ? Thanks...
I would recommend fixing the model fit first, and then the heywood case will probably resolve itself. As for how to justify, you can report that you were required to constrain the paths to be equal in order to resolve a heywood case, and then you can cite the paper: Kolenikov, S., and Bollen, K. A. 2012. "Testing Negative Error Variances: Is a Heywood Case a Symptom of Misspecification?," Sociological Methods & Research (41:1), pp. 124-167.
Sajeeb Shrestha Sorry for the delay. I was hosting an SEM Boot Camp last week and then yesterday was my wife’s birthday. Today is my first day back… You are correct that an endogenous variable cannot have its variance constrained. No need for a video since it cannot be done. Instead, you should probably create composite variables in AMOS during the CFA, then use these in the causal model.
What if the Heywood case happened to your factor rather than your item? That is, in your video, what if it happened to Quality per se? And, if you show how to solve it, may you show it in Mplus? Thank you.
Heywood cases on structural connections (e.g., regressions between factors) are usually due to using nominal variables (e.g., marital status, religion) without converting them to dummy variables. It can also be due to severe non-normality or low sample size.
Dear Dr. Gaskin; thank you for these informative videos. I have a question: is there any priority in the execution steps when we use this method (constraining the variance of the construct and fixing paths to the identical term)? To make it clear, I have four indicators, and the factor loading of one of them is more than 1, while others are less than 0.5, especially two indicators with factor loadings less than 0.3). If I employ this method without removing them, the standardized factor loadings of the lowest ones (now more than 0.5) become bigger than the two other indicators (now around 0.2), which already were higher. How would it be possible? Am I doing something wrong? ? I would be so grateful if you address this issue.
In this case, it sounds like your factor is not reflectively measured. I would recommend reconsidering including this factor in the CFA. Instead, you might use this as a composite/sum/index/score etc. instead of as a latent factor.
@@Gaskination Can I first remove the indicators with coefficient factor loadings lower than 0.3 (in my case, are 2 out of 4 indicators) and then use this method to tackle the Heywood case? If I do so, I will have two indicators with factor loadings of more than 0.5. If it is not feasible, why?
@@fatemerezaei-uk1hc If the indicators are interchangeable in meaning and measurement (i.e., they truly are measuring a single dimension), then yes, you can remove these other two indicators. However, a two-item reflective factor is the weakest form. If the indicators are not truly interchangeable, then I would recommend not evaluating the factor using reflective validation metrics.
Thanks Professor James Gaskin, In my (Measurement Model) CFA model, I have 20 Observed Variables and Five Latent Variables (Four Observed Variables for each Latent Variable). When I run CFA there is one observed variable loading is negative .58. My question is, should I retain the observed variable with negative factor loading or I can delete it.
It is probably because it is reverse-coded. For example, if the latent factor is satisfaction, and you have these three, the last one is reversed. 1. I am satisfied. 2. I am content. 3. I am not pleased. In this case, you just need to re-reverse the value (subtract it from 1+scale size). So for a 7 point likert scale, subtract from 8.
Dr. Gaskin: Thank you so much for your videos. I don't know where I would be without your site and all the tools/videos you provide. I have constrained the items following your video and it did work. However, since I cannot leave the 1 on the latent variable if it is endogenous for my SEM analysis, would it be acceptable to constrain the error variance for the item above 1 to a small number (.001) like you recommend doing for negative error variances? It allows me to run the analysis when I do this, but I am not sure if this would be an acceptable workaround in this situation.
Dear James. Thank you for your video. I could fix that very same problem in my model. But does doing that have any implications on how you will later analyze your data using that model. (hope that's not a foolish question, but my knowledge of statistics is very simple)
11MCL11 Constraining the two to be equal makes the assumption that they should be relatively equal. In a reflective measure, this is fine and shouldn't have any large repercussions.
Dear James , Your video's are just awesome . The kind of stuff you have on those videos is quite interesting .I am after on-line thesis who have done analysis using EFA CFA and SEM .Can you please send me a link to thesis done in way of EFA -CFA -SEM . This would give me an idea .
+Sanjeev Ingalagi Sorry for the delay in responding. RU-vid has stopped notifying content creators when comments are made on their videos... So I just thought no one was commenting on my videos... I'm very sorry that I don't have doctoral students, so I'm not sure of a thesis that uses this approach. I think there are free online catalogs of dissertations though. Best of luck to you.
+James Gaskin I'm curious about this fix as well. I know the common conception is that these weights should not be greater than 1, but I've recently come across some articles that suggest it is permissible, though not common. It is likely an indicator of multicolinearity. So bottom line, is this fix statistically kosher, or is it simply creating a cosmetic change that will not raise red flags with reviewers?
Dear Dr James Gaskin, I have a model with 14 factors, each has 2 indicators. Many standardized regression weight are greater than 1. Is this way for my case acceptable? Thank you so much because of your videos.
Two indicators per factor is not ideal. If this is all the measures you originally had, then I would recommend not using a latent design. Instead just average the items per factor and use the averages as new variables. If you had more measures, but deleted them to achieve validity metrics, then it is likely your factors are not actually reflective (as indicated by all the deletion). In this case, I'd still recommend averaging the items (if they all have the same conceptual direction - e.g., joy and satisfaction are in the same direction, but opposite burnout).
@@Gaskination Thank you so much. I learnt a lot from your videos for my thesis.This is all the measures I originally have. In case, I generate new variables by sum or average of 2 indicators, is this different? After that, Could I run EFA and CFA for these new variables?
@@cindydo7375 Generate the new averaged variables (assuming they are on the same scale size) and then do not do an EFA or CFA, because factor analysis is only for latent factors. So, just skip directly to the causal analysis (after checking normality assumptions).
Unbelievable , i just had this problem in my cfa and randomly looked at your channel and baaaaaam. Their is the solution on the first page, uploaded 7 days ago ^^. I will try it. Is their some theoretical background to this. I mean why it is allowed to do this ?
I'm not sure about the theoretical implications. Really they are measurement implications. To be safe, you might just say that due to a low number of indicators for one of your latent factors, you had an unstable factor that needed to have its indicator regression weights constrained to be equal in order to reduce error.
Dear Dr. James, Thank you very much for your wonderful explanation. I have conducted a CFA and I have a standardized correlation value equals to 1.02!! what can I do? would you please help me with this issue thank you!
Dear Dr. James, after watching your video, I am wondering about the practical implications of fixing a Heywood case. Say I ran different factor models and one of them gives a heywood case. When you fix it and thus make the model acceptable, is it okay to compare it to the other models that didn't need any fix at all? Intuitively, I was inclined to think that models with heywood cases are worse off by default. Thanks in advance for your reply. PS Would you recommend a book or article that helps me clarify this issue further?
If you're using the same variables across different models, then it would be prudent to make the same constraints in all models to make the comparable. As for an article on heywood cases, I can't think of any specifically, but this scholar search should be helpful: scholar.google.com/scholar?hl=en&as_sdt=0%2C45&q=%22heywood+case%22&btnG=
Dear prof. Gaskin, I followed the procedure you explained in the video and it worked perfectly (I had regression weights of 2.7 and 8.5). However, all p values are *** (
dear Dr. Gaskin, I am trying to examine the actual use of ERP systems , the problem i am facing is that the actual use construct just explain .05 where other endogenous factors explain more (for example, perceive easy of use (.68) intention to use (.73). I watch the above video but as you know i cant apply the constraint (1) on the latent variables. Any Suggestions. Many Thanks
It is very common to have high r-square on perceptual variables and low r-square on observed variables. This is because perceptual variables have limited and systematically correlated variance due to various method biases (such as social desirability) or due to endogeneity (usually due to things like optimism, loyalty, etc.). Observed variables don't share this inflated covariance, so they end up with lower r-squares.
Make sure you are viewing standardized estimates. If it is still an issue, then it is probably due to something neglected earlier in the process (such as normality or validity).
Thanks a lot. I have a model of three variables (independent, mediating, and dependent) with a standardized regression weight greater than one for the effect of the independent variable on the dependent one. I cannot put a fixed value for the variance in the dependent variable, and when I name the coefficients of the path from the independent variable and the mediating variable to the dependent one with the same name, the direction of the causality from the mediating variable to the dependent one shifts from negative to positive. How can I deal with this? Sample size is 437 with 32 observed variables. No outliers or skewness exist. Thanks in advance
This is highly unusual with that sample size and normality, unless one of the variables in nominal/discrete/categorical (e.g., industry: retail, manufacturing, service). If these are all normal and validated scales, then I can't imagine what is causing the issue. If you have not yet validated the scales in a CFA, please do that first.
@@Gaskination Thanks a lot. If you mean by validating scales convergent and discriminant validity, yes there are some issues. Some AVEs are not acceptable but the composite reliability coefficients are all above 0.7 in addition, HTMT between the IV and the M is .893 .if I did not seek better cfi than .908, can I have a standardized regression weight of .94 regarding the effect of IV on the DV, or this will be regarded as too much variance explained by one variable?
@@user-gk7tb3rw3j Yikes! that sounds like there is something wrong with the data. Those variables are too highly correlated. This explains the strange estimates you're seeing. I would recommend trying to separate them more. Try doing an EFA with just the indicators for those two factors. Try separating them as much as you can by finding the items that crossload most strongly and removing them one at a time.
@@Gaskination Thanks alot. The coefficients I mentioned are already after performing EFA and CFA and eliminating many items and reaching satisfactory model fit indices. I managed to find out the standard deviation between the remaining items of IV and M and deleted 64 responses whose std deviation is less than 0.5. The HTMT is now .887 and the standardized direct effect of IV on M and DV is .718 and .483 respectively and sample size is now 377. Is there anything else I can do?
@@user-gk7tb3rw3j Sounds like you had some low quality resposnes. The more you can do to ensure the remaining responses are of high quality, the better. Check for outliers and unengaged responses.
Dear Dr. Gaskin, Please suggest if error variance (values above e1, e2, e3, etc.) can be more than 1? I see more of scale models show error variance only less than 1? what does it mean?
Thank you, Dr James for your video. How do I solve negative estimates problem for AMOS? When I used the above suggestion, my model indices failed to fit (too large/out of range values eg. RMSEA) although the estimates became positive values.
If the indicator loadings are negative, then you may want to check the direction of the scale or the wording of the questions. Also consider whether the measures are formative rather than reflective.
James Gaskin Thank you for your response. I have gone through my data once more. The items turned out to be reversed coded items so I checked the responses again to make sure I have reversed coded the items correctly, which I have done. Hence, I think my participants viewed the items in a negative fashion - opposing the factor (rather than supporting the factor). I may eliminate them from the model. If the measure is formative, should I redo/redraw the model in AMOS?
CITATION Since this is the top video I always come across for this issue, I leave a ressource here for citation purposes, that won't be a youtube-video, if your uni is as pedantic as mine :P web.pdx.edu/~newsomj/semclass/ho_improper.pdf PS: Thank you Dr. Gaskin or all your work and expertise you lend us!
Sir, I am trying to run a three factor second-order CFA. The unstandardized loadings appear to be fine but when I look into the standardized loading, two of them are equal to one. I have tried all the things mentioned in this video but have seen no improvement.
In these special cases, I would have to see the data and model to know what to do. Most likely, I would try to add or remove indicators to see if that fixes it. I would also check out the factors in a reliability test and probably run the EFA again.
In these special cases, I would have to see the data and model to know what to do. Most likely, I would try to add or remove indicators to see if that fixes it. I would also check out the factors in a reliability test and probably run the EFA again.
Hi, im doing a SEM model and I have a questions about negative variances. In my model all variances are positive, but few of them are very small (0.005 with SE=0.28 and p-value=0.8) and not significant. My question is if this is a problem, having a variance that could be negative...Thanks!
You can try moving the constraint around. You can also check modification indices to see if there is a strong systematic correlation with some other item (and then if there is a logical reason for their relationship, like very similar wording on a survey) then you could covary them.
if an indicator weight is negative, then it is probably because it is a reverse-coded question. make sure to re-reverse it by subtracting its value from the scale size +1 (so, if 5point likert scale, subtract from 6). Then save the data and re-link it.
Hi James, Thanks for this video it's really useful. I currently have this problem for a latent variable with 2 indicators. One indicator is over 1. Also the error of 1 out of the 2 indicators is minus. If I use the 'aaa' strategy- this stops the error from being negative and shares the loading but it is still over 1 (e.g. 1.02). If I just fix the regression weight to the other indicator- this fixes the problem of the regression weight being over 1- but one of the errors is still negative. Is it a problem if one of the errors is negative? Many thanks for any advice you could provide on this matter. All the best, Catrin
+catringriffiths1 Sorry for the delay in responding. RU-vid has stopped notifying content creators when comments are made on their videos... So I just thought no one was commenting on my videos... It is a bit of a problem if the error variance is negative. You can constrain it to a small positive number though (like 0.001).
Kolenikov, S., and Bollen, K. A. 2012. "Testing Negative Error Variances: Is a Heywood Case a Symptom of Misspecification?," Sociological Methods & Research (41:1), pp. 124-167.
Dear Sir, I hope you are keeping well. What to do if one or more of the standardized factor loadings of the first-order factor on a single second-order factor becomes greater than 1?
Hi James, thank you very much! I tried your methods to fix regression weights greater than 1, but when I run my model, chi-square test for the model showed that my model was rejected. In the beginning, the model is accepted. Could you please help me explain for this? Thank you in advance.
If you mean the p-value for the chi-square was less than 0.05, then this is fine. The chi-square test is a very strict measure of model fit. I would instead recommend relying on an array of measures, such as the CFI, RMSEA, and SRMR.
+batalmarah Negative covariance simply means they are inversely related. This is perfectly fine. It is like the relationship between burnout and satisfaction. When burnout increases, satisfaction decreases. They would be inversely correlated.
Is there an argument against doing this? You imposed an equality constraint on the loadings of the items, q1 and q6. Could someone (e.g., a dissertation committee) ask for a justification for assuming the loadings of those items are probably about equal?
Mister Gaskin, I"m pleased to see your video but do you know how can I solve the same problem but with Mplus instead of Amos? I'm a Phd Student in Belgium.Thank you very much
Dear Dr Gaskin, one of my indicators produced a factor loading is 0.787 which is good. But since my data is not normal, I carried out bootstrapping. And I got a lower bound of 0.478 and an upper bound of 1.362 using 95% bias-corrected CI. Is there something wrong with the indicator? By the way, this indicator belongs to a latent factor with only two indicators. Thanks a lot.
Dear Dr. Gaskin. Thank you very much for your videos. In my case, I have 10 factors but 2 of them fall into Heywood cases. I fix the problem by putting the constraints (1) on the indicators (2 Heywood cases) while maintaining the constraints (1) on the rest 8 factors. Is it fine or not? Thank you.
if im testing only the quality construct and put the constraint of 1 in the both indicator paths and in the latent quality? is this wrong? because there is some cases that the plugins of model validity runs only if a do this
On the measurement model, yes, even if there is a structural model included. But for structural paths, I would hesitate to constrain paths. Instead, check if it is due to small sample size, severe non-normality, or nominal variables (which shouldn't be included directly).
Respected Sir, I have a question regarding regression. if the data is non-parametric and there is one dependent and two independent variables then is it necessary to run regression or correlation is sufficient.
I would recommend regression when there is more than one predictor. This way you can see the effect of each predictor while controlling for the effect of the other predictors.
@@Gaskination thank you very much Sir for the reply. If there is one dependent and one independent variable then whether to go for regression or correlation is sufficient. Secondly can we run regression in non-parametric data.
@@paparinayak5039 Yes, you can still use regression. However, if the data is strongly non-normal, then this will bias your results. You might instead use something like path modeling with rubust estimation, such as MLR.
Hi Gaskin sir,after adjusting constraint on factor. CFA is solved. But problem in SEM that AMOS doesn't allow path the those endogenous variable having constraint on factor. what to do in SEM Sir?
Sajeeb Shrestha Correct, you can't constrain the variance of an endogenous variable, so you'll need to adjust the model another way. See if the model will run without the constraint.
Perhaps there is some issue with the variables. Make sure you are not using categorical indicators for that factor. Also check their Cronbach's alpha. If still no good, perhaps run the EFA to see if the two items factor together.
Sir, Can I use the factors that are obtained in CFA for Regression analysis in AMOS?. In fact, I performed CFA with 42 items through which I reduced items into 28 with good fit (GFI) value >90. Then, should I only consider these 28 for regression analysis to justify my research model? or else, should I move for 42 items instead of 28 items. Please clarify my doubt...sir.
I recommend using the results of the CFA. These will result in the strongest variables. Impute factor scores for the CFA instead of using the mean of indicators.
Dear James, Thank you for your video. May you help me with the problem in SEM. In SEM result, unstandardized estimates is below 1 and significant (0.542). But standardized regression weight resulted estimates between two variable is 1.00. How can I sort it out? Thanks in advance
+Tam Truong Duong This sometimes happens when the variable has high kurtosis (or is abnormal in other ways). Make sure to do all the data screening and EFA before moving on to CFA and causal models.
+James Gaskin Hi James, I did screening data as you mentioned. One item has high kurtosis (0.968). The scale of this variable has just 3 items, so that I cannot delete such item. May you have any solution to solve that, I mean high kurtosis or regression weight equal 1? Many thanks
+Tam Truong Duong That is not too bad for kurtosis. This can also happen when you have scales with low variance (e.g., a 3 point scale instead of a five or seven point scale). It can also happen when you have negative error variance, or when you have covaried errors.
+Tam Truong Duong I'm sorry. I'm not sure then. I would have to see the model. One other guess is that you have only two indicators for a latent factor. Sometimes this can cause the problem.
In the tool, the factors were already constructed, therefore, I did not apply EFA, but CFA. I can not go back to EFA. What should I do make it better in AMOS?
If you have the items, then you can do an EFA, even if it is already an established set of scales. If you do not have the indicators, then I'm not sure how you're doing a CFA, unless you are using a covariance matrix.
I found Heywood case when I was doing the configural invariance test on CFA based on two nationalities. Is it normal? Can I use your methods to solve this problem?
Make sure you are looking at standardized loadings, rather than unstandardized. Unstandardized loadings are not required to be less than 1. As for this method of solving the problem, I wouldn't worry about it during the invariance test.
Thank you for your reply. I found standardized loadings greater than 1.00 on two variables from ten variables in my unconstrained model based on two nationalities. Then, I constrained all paths with aa for the 1st variable and with aaa for the 2nd variable. After doing so, the Heywood case was resolved. However, I wonder whether this method will then affect the metric and scalar invariance test or not.
It will affect it because you are constraining those paths to be equal. So, you can either unconstrain them for the test, or you can claim partial invariance if you keep them constrained.
negative regression weights are due to inverse correlations. These are usually caused by reverse-coded questions. In such cases, you'll need to re-reverse the values by subtracting them from 1+ the scale size (so, if 5-point likert scale, then subtract from six).
James Gaskin thank you very much for your reply ,but what if we dont have the questionnaires data ,because my model is based on an article from the internet and not on my own work ... !
@@senam7471 If you don't have the data, are you running it with just the correlation matrix? If so, then you just have to treat it like an inverse correlation. In this case, the loading is indicating that it doesn't belong with that factor.
James Gaskin thank you ... yes its with the corelation matrix .. but im a beginner in Amos and spss ...and i didn't understand how to treat it like an inverse correlation ,is there a video about it please ?
@@senam7471 if it is negative, then you just assume that the relationship is inverted. So, if it is for an indicator of a factor, then probably remove that item. But if it is just for a relationship between two constructs, then you can just know that as one increases, the other decreases (like with burnout and job satisfaction). This is fine.