Hello, I am a new user here, and I am quite in need of help with regards to statistical problems. I really appreciate it if anyone can help me.

I am currently doing a kind of "predicting user behaviour" research using Structural Equation Model (SEM). My software of choice is LISREL 8.80.

I am currently most troubled by two things.

1. I am using local books as the main guidelines in conducting an analysis using SEM, but I also references international journals and books. In the local books, I found that they mention "variance extracted (VE)". However, in international journals and books, I read a lot of "average variance extracted (AVE)" being mentioned. I tried to find if VE and AVE are actually the same entity, or are they different. After googling I found that the formula for AVE, is the same as the VE I read in local books. But, in one of the journals I read, it wrote: "Average variance extracted (AVE) is the average VE values of two constructs". In short, is average variance extracted (AVE) the same as variance extracted (VE) or not?

2. When I did research without composite measures (one latent variable is indicated by more than one observed variables), I usually did multicollinearity diagnostics by using "Collinearity diagnostics" in SPSS, by trying out one independent variable as the dependent variable and the other independent variables as the independent variables. Then I iterate this process until all independent variables have been tested as dependent variables against other independent variables. I then check the tolerance and VIF values.

But how do I conduct this for SEM analysis, because in SEM, I only have the observed variables (for example SQ1-SQ4 as observed variables for SQ)?

I currently tried the same technique in SPSS, by putting one observed variables as the dependent, and the other observed variables (including observed variables for other latent variables) as the independent variables.

However, I have found no literature supporting what I am doing is a valid method to detect multicollinearity for SEM. For example, shouldn't I regress the latent variables against each other instead of the observed variables? (but I don't know how to do that as well!) I humbly ask, for someone to confirm if what I'm doing is correct or incorrect. And if it is incorrect, what is the correct way of doing it?

I have no idea to check for multicollinearity in LISREL.

Extra questions (confirmation only):

- Is composite reliability the same as construct reliability? Based on what I read in Hair, et. al.'s book, I am drawing conclusion that its a same.

- In SEM theory, can latent variable be said to be a kind of composite variable? I am assuming it is because in SEM, a latent variable is reflective of multiple indicators, but after reading a few papers I am now a bit unsure of it.

Thank you very much,

Iori.

I am currently doing a kind of "predicting user behaviour" research using Structural Equation Model (SEM). My software of choice is LISREL 8.80.

I am currently most troubled by two things.

1. I am using local books as the main guidelines in conducting an analysis using SEM, but I also references international journals and books. In the local books, I found that they mention "variance extracted (VE)". However, in international journals and books, I read a lot of "average variance extracted (AVE)" being mentioned. I tried to find if VE and AVE are actually the same entity, or are they different. After googling I found that the formula for AVE, is the same as the VE I read in local books. But, in one of the journals I read, it wrote: "Average variance extracted (AVE) is the average VE values of two constructs". In short, is average variance extracted (AVE) the same as variance extracted (VE) or not?

2. When I did research without composite measures (one latent variable is indicated by more than one observed variables), I usually did multicollinearity diagnostics by using "Collinearity diagnostics" in SPSS, by trying out one independent variable as the dependent variable and the other independent variables as the independent variables. Then I iterate this process until all independent variables have been tested as dependent variables against other independent variables. I then check the tolerance and VIF values.

But how do I conduct this for SEM analysis, because in SEM, I only have the observed variables (for example SQ1-SQ4 as observed variables for SQ)?

I currently tried the same technique in SPSS, by putting one observed variables as the dependent, and the other observed variables (including observed variables for other latent variables) as the independent variables.

However, I have found no literature supporting what I am doing is a valid method to detect multicollinearity for SEM. For example, shouldn't I regress the latent variables against each other instead of the observed variables? (but I don't know how to do that as well!) I humbly ask, for someone to confirm if what I'm doing is correct or incorrect. And if it is incorrect, what is the correct way of doing it?

I have no idea to check for multicollinearity in LISREL.

Extra questions (confirmation only):

- Is composite reliability the same as construct reliability? Based on what I read in Hair, et. al.'s book, I am drawing conclusion that its a same.

- In SEM theory, can latent variable be said to be a kind of composite variable? I am assuming it is because in SEM, a latent variable is reflective of multiple indicators, but after reading a few papers I am now a bit unsure of it.

Thank you very much,

Iori.

Last edited: