SEM Series Part 7: Building Your Structural Model

In this video I show how to build the structural model from the measurement model.
Video Rating: 5 / 5

15 thoughts on “SEM Series Part 7: Building Your Structural Model

  1. To fix the issue another way, look to see if there are any negative error variances (in notes for the model section of output) or high standard errors (in regression weight table). This will give you a clue as to where the issue might be. If those aren’t the issues, then I’m not sure what the problem might be. You could email me your model and data as a last resort.

  2. Hi James, Your videos have been very helpful and are easy to understand- thanks for that. But I keep getting an error message as I try to run my structural model. It’s telling me that the sample moment matrix is not positive definite and gives a list of reasons why this may be. I don’t think any of the reasons would apply in my case but when I click ‘allow non-positive definite sample covariances’ as they suggest as a work around there are no model fit indices (AGFI, CFI etc). Thanks.

  3. Dear James, Firstly, thank you for sharing all this useful SEM info. I have created a structural model, but it does not have good model fit. The modification indices suggest to covary an error term with a factor, which I don’t think is possible. Another suggest is to create a path from IV to DV. Is that possible? If the model still has poor model fit, does this mean that I cannot go to the next step (SEM Series Part 8: Mediation)?

  4. For the first issue, you can either covary the error term and the error term of the factor, or add a regression line between the factors. For the second issue, yes, you can create a direct path between IV and DV. If it still has poor model fit, it may be due to non-normally distributed data or outliers. You may want to review those before proceeding. Best of luck!

  5. I have never tested a non-recursive model in AMOS, although I know you can do it. I’m not sure what the caveats of such an analysis are. Here is a paper that discusses it though: “Testing Reciprocal Relations by Nonrecursive Structural equation Models Using Cross-Sectional Data”. Best of luck.

  6. Dear James, Thanks a lot for your videos! They have been incredibly helpful in discovering SEM. I just discovered that my model is a “non-recursive model” with a feedback loop (A –> B –> C –> A). I was wondering whether such a model can be modelled in AMOS as shown in your video (using cross-sectional data) or if there are special issues I need to pay attention to.

  7. Dear James! Your videos have been great help for me as I am doing my own SEM. However in this video it is unclear for me, how you have created some variables. When you do structural model second time, I dont know, how your variables like Play or Usefulness have been made. I have watched all your videos in SEM series, but I dont find answer there. I guess sum variable is not good enaugh for this kind of analysis.

  8. Dear James, Is there any way to use a binary latent variable (i.e. Adoption) in the final structural model? My final DP is the adoption and I prefer to use a binary construct rather than measuring the extent of adoption. But I am not sure if I can include it in Amos analysis. I really appreciate your help.

  9. You can. If it is binary then just include it as you would any other variable. Just make sure that you interpret it accordingly. So, if 1=yes and 0=no, then if the path coefficient leading to adoption is positive, then the indicator increases the likelihood of adoption. If negative, then it decreases likelihood.

  10. The CFA is for testing the factor structure, the validity and reliability of the latent factors. The full SEM model uses the same structure, but removes covariances among the endogenous variables and adds regression lines. This model takes into full account the measurement error of the latent factors. The path model is a simplified version of the SEM and ignores some error, but also greatly simplifies the analysis. The results will not differ greatly.

  11. continued – fit index and AIC, EICV , indicate the simpler model is best … Is that so ?? If the retained components are as good to explain the data should I say that the simpler theory predicts the data better than the other ? And if so, can I drop the other component (latent variable) and say that online attitude predicts INTENTION (also theoretically sound) ? Thanks a million !

Leave a Reply