Colleagues,
I have a co-author who is struggling with the Heywood case of onon-posiitve definite, essentially you have a variance of "0" in the SEM analyiss. There is a large sample size, over three hundred, and there are many measurement indicators.
Have any of you successively dealt with this? If so, how? Also what has been done so far is to only test the latent variables (essentially treat all constructs as reflective) and not include the measurement indicators in the model.
Your thoughts and suggestions on how to deal with this anomaly in SEM?
thanks,
Louise Kelly
This article by Chen et al. does a pretty good job of covering improper solutions that may result in a Heywood case.
From your description, the model you're fitting is incorrect (a Heywood case may be the result of a small sample size -sampling variability- or model mispecification). I would suggest simply dropping that indicator. If this cannot be done, you will have to find an alternative model. Consider a bifactor model, they always provide a good fit to data.
I am also struggling with the same situation. The Chen et al (2001) paper gave me some ideas how to deal with that.
I am currently working on a paper titled, "HEYWOOD CASES: POSSIBLE CAUSES AND SOLUTIONS".
It is a real problem for me as well! Looking forward to hear your proposed solutions.
Regards
Ghaith
Great to hear that, Rayees Farooq . I am looking forward to read you paper.
if you are doing in AMOS this vdo is useful https://www.youtube.com/watch?v=Vx24KFf-rAo
I too face this problem. Can we fix the negative error component to some positive value and get standard path coefficients less than 1?
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling