These are two related, but not the same criteria. The Fornell-Larcker criteria is about how much variance two factors share with each other. The VIF criteria is about how much variance in the dependent variable is explained by the same factors. So, in this case, you have no multicollinearity problems, but you have a discriminant validity problem. To fix discriminant validity, you might try running an EFA with just the indicators of the two factors that are too highly correlated. Then try to separate them in the EFA before returning to the CFA.
Thank you so much, sir for helping me finish my thesis. I have a question regarding multicollinearity test. In the video that you did in 2011, you did collinearity test by putting the independent variables one by one as the dependent variable. However, in the SEM bootcamp series you just put the endogenous variable as the dependent variable. Can you help me clarify the difference between the two methods? Once again thank you so very much!
I've learned since 2011 that it is sufficient (and more accurate) to use the DV in the DV spot. The reason I did it the other way in 2011 is because that is the way I was taught it during my PhD program. However, having the DV in the DV spot is important because we need to assess the redundant explained variance in the DV, which we can't do if we don't include the DV.
@@Gaskination Thank you sir, that explains a lot. I was self-doubting all the way till my deadline since you didn't really explain it and I also received contradictory advice from different people.
Hi James! I am testing a structural model with 3 IVs, 1 mediating variable and 3 DVs. In addition, I want to investigate the effect of 1 moderator on the relationship between the mediating variable and the 3 DVs. 3 control variables are also included into the model. All 11 constructs are measured by multiple items on a 7-point Likert scale. When testing for multicollinearity by means of the correlation matrix and variance inflation factors (VIF), do I only have to examine the 3 IVs or also the correlation between additional variables, such as the mediating, moderating and control variables? Thank you very much in advance!
@@Gaskination Thank you for your answer! So you would compute the VIFs for all 3 IVs, 1 mediator, 1 moderator, 1 interaction term and the 3 control variables if they have a direct link to the DVs in the structural model?
I appreciate the many videos of yours that I have watched. So, first I want to say "thank you so much" - Second, I have a question: I have a specific variable with 95 values. The variable is "Counties" - I have searched, rephrased my wording - yet, I have not found how how to analyze my dataset that describes details per county, such as income, population, and hoping to see if certain counties have more cases of Obesity deaths than others and are the lower income or is there no significance to the income etc. This is a dataset obtain via CDC specific to obesity and the state of Tennessee. Can you give me some direction please? I have been using SPSS for quite a number of years, but this is a first for me.
This is a simpe one. Just use ANOVA. The dependent variables will be the ones you mentioned: Obesity, income, and anything that you want to compare across counties. The factoring variable will be counties. This should work pretty well, as long as you have multiple entries for each county. If each row in your dataset represents a single county, and each county is represented only by a single row (i.e., you have 95 rows of data), then you'll need to instead just visualize the data using Excel or SPSS chartbuilder. Statistics won't be possible on data with only one datapoint per group (county)
@@Gaskination Thank you so much. This was much simpler than another researcher suggested. He suggested I create a multiple regression model with these data. I could use income as the predictor and obesity as the outcome while controlling for key variables (e.g., population). Then recommended that I use dummy code for the County variable and enter it into the regression model to see what the model suggests about the county. #1 - I don't know how to "create" a regression model, #2 - not sure what is meant here by "dummy code" -- so I went with your advice and got what I needed..
Quick question - is it ok to use this procedure to inform which variables I should remove from a generalized linear model? For example, I have temperature, temperature gradient as dependents among others predicting species abundances. Can i use the VIF's produced here to inform my generalized linear mode. TY
Hello! Mr. James! If my dependent variable has multiple items(e.g. Three questions with 7 scales asking similar things that all measures satisfaction), should I compute the average and create a new DV to run this linear regression? Thanks in advance~
@@Gaskination Thank you so much again! The other videos of yours teaching how to impute the factor score from CFA in AMOS to SPSS is fantastic! If the VIF for my data is lower than 10 with the independent variables, I think it would not be a problem for my master degree dissertation?
Hello James. I would like to know if this applies to logistic regression? I need to do the multicollinearity test in SPSS but my dependent variable is binary (yes and no)
@@Gaskination So can I use this method anyway and then continue with my analysis by the logistic regression option? The result would be the same, right? Thank you for returning
Hi James, May i know why there would be a pop up chat said an error occurred while attempting to fit the model. Is there any mistake that i made or how can i solve the problem?
If you mean this happened in SPSS, then this could be for several reasons, including insufficient sample size, predicting something with itself, erroneous inclusion of categorical variables in a regression, inclusion of a variable that does not vary (e.g., if gender variable is all male), etc. In AMOS, it can be for many similar reasons, but AMOS will usually give more explanation.
Kindly , I would like to know that answer about VIF ? My model includes one independent variable only , two mediation variables and two dependent variables. Can I use VIF test in that situation or no? Thanks a lot prof .
You can conduct multicollinearity with a random value outcome. To do this, create a new variable with completely random values (you can do this in Excel with the Rand function). Then make this variable the dependent variable and use all other variables in your model as predictors. Then the VIF will be produced for all of them. I should make a video for this...
@@Gaskination Ok, thank you. So to test the second-order factor I would compute the mean score of all the items which represent the second order factor? Apologies if this is a silly question.
Also, can I use this method to assess common method bias? Or is this approach shown in this video different to the common method bias video for smartpls. Sorry, I am very confused.
No. Multicollinearity is an assessment of the redundance of predictors (i.e., do two predictors explain the same variance in the outcome). Multivariate normality is an assessment of the "normalness" of specific cases (rows) in your data with regards to specific multivariate relationships based on the general relationship (i.e., if most people are more satisfied when they get paid more, but respondent 124 reported being less satisfied, then they are multivariate non-normal).
@@atikhomthienthong9310 Pretty much any SEM is a network of interdependent relationships. As such, any time you add or take away any parameter or variable, there is the chance that some or all other parameters will change textbook should suffice. I like Hair et al 2010 "Multivariate Data Analysis"
Very useful, thanks a lot!
Thank you
Thank you for all the videos! I have a question: If SQRT (AVE) < Correlation but all VIFs are below the 3. What would you do?
These are two related, but not the same criteria. The Fornell-Larcker criteria is about how much variance two factors share with each other. The VIF criteria is about how much variance in the dependent variable is explained by the same factors. So, in this case, you have no multicollinearity problems, but you have a discriminant validity problem. To fix discriminant validity, you might try running an EFA with just the indicators of the two factors that are too highly correlated. Then try to separate them in the EFA before returning to the CFA.
Thank you so much, sir for helping me finish my thesis. I have a question regarding multicollinearity test. In the video that you did in 2011, you did collinearity test by putting the independent variables one by one as the dependent variable. However, in the SEM bootcamp series you just put the endogenous variable as the dependent variable. Can you help me clarify the difference between the two methods? Once again thank you so very much!
I've learned since 2011 that it is sufficient (and more accurate) to use the DV in the DV spot. The reason I did it the other way in 2011 is because that is the way I was taught it during my PhD program. However, having the DV in the DV spot is important because we need to assess the redundant explained variance in the DV, which we can't do if we don't include the DV.
@@Gaskination Thank you sir, that explains a lot. I was self-doubting all the way till my deadline since you didn't really explain it and I also received contradictory advice from different people.
Hi James
Shoud we include control variables along with independent variables for multicollinearity diagnostics ?
Yes. That makes sense. All predictors.
Hi James! I am testing a structural model with 3 IVs, 1 mediating variable and 3 DVs. In addition, I want to investigate the effect of 1 moderator on the relationship between the mediating variable and the 3 DVs. 3 control variables are also included into the model. All 11 constructs are measured by multiple items on a 7-point Likert scale. When testing for multicollinearity by means of the correlation matrix and variance inflation factors (VIF), do I only have to examine the 3 IVs or also the correlation between additional variables, such as the mediating, moderating and control variables? Thank you very much in advance!
You should test it between all variables predicting the DVs.
@@Gaskination Thank you for your answer! So you would compute the VIFs for all 3 IVs, 1 mediator, 1 moderator, 1 interaction term and the 3 control variables if they have a direct link to the DVs in the structural model?
@@martinb.8231 you would calculate the VIF for all the predictors, and you could use the DV or a random variable as the DV in the test.
I appreciate the many videos of yours that I have watched. So, first I want to say "thank you so much" - Second, I have a question: I have a specific variable with 95 values. The variable is "Counties" - I have searched, rephrased my wording - yet, I have not found how how to analyze my dataset that describes details per county, such as income, population, and hoping to see if certain counties have more cases of Obesity deaths than others and are the lower income or is there no significance to the income etc. This is a dataset obtain via CDC specific to obesity and the state of Tennessee.
Can you give me some direction please? I have been using SPSS for quite a number of years, but this is a first for me.
This is a simpe one. Just use ANOVA. The dependent variables will be the ones you mentioned: Obesity, income, and anything that you want to compare across counties. The factoring variable will be counties. This should work pretty well, as long as you have multiple entries for each county. If each row in your dataset represents a single county, and each county is represented only by a single row (i.e., you have 95 rows of data), then you'll need to instead just visualize the data using Excel or SPSS chartbuilder. Statistics won't be possible on data with only one datapoint per group (county)
@@Gaskination Thank you so much. This was much simpler than another researcher suggested. He suggested I create a multiple regression model with these data. I could use income as the predictor and obesity as the outcome while controlling for key variables (e.g., population). Then recommended that I use dummy code for the County variable and enter it into the regression model to see what the model suggests about the county.
#1 - I don't know how to "create" a regression model, #2 - not sure what is meant here by "dummy code" -- so I went with your advice and got what I needed..
Quick question - is it ok to use this procedure to inform which variables I should remove from a generalized linear model? For example, I have temperature, temperature gradient as dependents among others predicting species abundances. Can i use the VIF's produced here to inform my generalized linear mode. TY
Multicollinearity just tells you if two predictors are redundant in the variance they explain in an outcome variable.
Hi, Mr. James. Is it possible to conduct multicollinearity test if I have 2 dependent variables?
Yes. Just do it for each DV separately.
Hello! Mr. James! If my dependent variable has multiple items(e.g. Three questions with 7 scales asking similar things that all measures satisfaction), should I compute the average and create a new DV to run this linear regression? Thanks in advance~
Yes, that is one way to do it. The other way is to impute a factor score during the EFA or CFA, and then use that factor score.
@@Gaskination Thank you so much again! The other videos of yours teaching how to impute the factor score from CFA in AMOS to SPSS is fantastic! If the VIF for my data is lower than 10 with the independent variables, I think it would not be a problem for my master degree dissertation?
@@andrewdogdog1 yes, that is fine. Lower than five is better, but lower than ten is adequate.
@@Gaskination Thank you very much for the reply! Best wishes for you~
Is a multicollinearity test necessary for as linear regression with one mediator ?
Multicollinearity should be tested if an outcome variable has more than one predictor.
Hello James. I would like to know if this applies to logistic regression? I need to do the multicollinearity test in SPSS but my dependent variable is binary (yes and no)
As far as I'm aware, there are no multicollinearity options in logistic regression.
@@Gaskination So can I use this method anyway and then continue with my analysis by the logistic regression option? The result would be the same, right? Thank you for returning
@@sschiefelbein Since the outcome is binary, then yes, the result should be the same.
Hi James, May i know why there would be a pop up chat said an error occurred while attempting to fit the model. Is there any mistake that i made or how can i solve the problem?
If you mean this happened in SPSS, then this could be for several reasons, including insufficient sample size, predicting something with itself, erroneous inclusion of categorical variables in a regression, inclusion of a variable that does not vary (e.g., if gender variable is all male), etc. In AMOS, it can be for many similar reasons, but AMOS will usually give more explanation.
Kindly , I would like to know that answer about VIF ?
My model includes one independent variable only , two mediation variables and two dependent variables.
Can I use VIF test in that situation or no?
Thanks a lot prof .
You can conduct multicollinearity with a random value outcome. To do this, create a new variable with completely random values (you can do this in Excel with the Rand function). Then make this variable the dependent variable and use all other variables in your model as predictors. Then the VIF will be produced for all of them. I should make a video for this...
Hi James! Can I please ask if I can use this method to test multicollinearity between a second-order construct to a dependent variable?
Yes, but only if the dependent variable has multiple predictors.
@@Gaskination Ok, thank you. So to test the second-order factor I would compute the mean score of all the items which represent the second order factor? Apologies if this is a silly question.
Also, can I use this method to assess common method bias? Or is this approach shown in this video different to the common method bias video for smartpls. Sorry, I am very confused.
If I may add, I tried to do the CLF approach but it broke my model and now I am unsure how to test for common method bias..
@@ievamasevic7360 Correct. Or impute a factor score for all latent factors (rather than an average): ruclips.net/video/dsOS9tQjxW8/видео.html
Is this Multicollinearity the same as Multivariate Normality. i.e., Does Multicollinearity explain the multivariate normality?
No. Multicollinearity is an assessment of the redundance of predictors (i.e., do two predictors explain the same variance in the outcome). Multivariate normality is an assessment of the "normalness" of specific cases (rows) in your data with regards to specific multivariate relationships based on the general relationship (i.e., if most people are more satisfied when they get paid more, but respondent 124 reported being less satisfied, then they are multivariate non-normal).
@@Gaskination Thanks for the clarification
What should we do if we do not want to remove any independent variable?
Don't remove IVs unless they are strongly violating multicollinearity (e.g., VIF > 10)
@@Gaskination Do you have any publication to cite?
@@atikhomthienthong9310 Pretty much any SEM is a network of interdependent relationships. As such, any time you add or take away any parameter or variable, there is the chance that some or all other parameters will change textbook should suffice. I like Hair et al 2010 "Multivariate Data Analysis"
@@Gaskination Thank you for your help.