I am doing a CFA in SPSS Amos, and I face the following problem. 2 of the latent constructs are highly correlated (Amos provides the estimate of greater than 1).
Could please someone advise how to deal with this issue?
Thank you in advance!
Kind regards,
Anastasia
What estimate you are talking about?
If correlation is greater than 1, definitely you have a problem. Check for items redundancy and delete one of the two redundant items.
If the factor loading is greater than 1, also you have a problem. You need to check for negative error variance and correct it.
Hi Anastasiia,
Could you post a path diagram of your model? It sounds like it might be a specification issue.
Hello Robert,
My path diagram looks really messy now, most of the observed variables overlap in the picture, so I attach a file with the path diagram, standardized factor loadings and correlations.
I am quite new to SEM, so I would really appreciate any kind of help!
Anastasia
That is quite the CFA. May I ask how much data you have? That may be causing estimation problems. If you have sufficient data then you may want to combine the factors that have the high correlation into a single factor.
I am targeting at least 250 responses, however, this CFA is run with 140 responses.
The over-1 correlated factors describe two types of strategic alliance stability, so actually they view one issue from different perspectives. However, it is critical for the research that I keep them separate to identify different kinds of effects other factors are supposed to have on them.
Is there probably a way to identify what exactly causes the problem?
Anastasia
So one problem could be the number of observations you are using. You have quite a few parameters and it is my guess that you do not have enough observations to properly estimate the path coefficients and are getting an error. The minimum for reliable estimates is about 5 observations per parameter. Are all of these factors of different surveys or are they all part of the same survey? If they are from different surveys then you can run a separate CFA per survey structure. This will help with your observation to parameter ratio for estimation. If that does not work then you may not be able to distinguish between those two factors given this data set.
Robert, thank you a lot! Probably, I really need to merge them, I will try to do that and see what happens.
I have collected responses by running one survey (I have to admit that it was pretty long), so I have to stick to one CFA, I guess.
Or else, should I run PLS instead of the SEM in case of a small number of observations?
What estimate you are talking about?
If correlation is greater than 1, definitely you have a problem. Check for items redundancy and delete one of the two redundant items.
If the factor loading is greater than 1, also you have a problem. You need to check for negative error variance and correct it.
Multicollinearity is a real pest at times. You might start by checking residuals, and modification indices to see if particular items are problematic. You may not want to change a validated scale, but sometimes similarly worded items, or even a common stem (i.e., introduction to items) can be reason to correlate certain item residuals. This can help at times.
You mentioned that the two constructs are similar (strategic alliance stability), so it might be helpful to know if they have been used together in prior studies; have the constructs shown discriminant validity before?
Correlations above 1 are impossible, so I would suggest starting with just the two factors, and then build from there to better understand the issue - although a more nuanced review is through standardised residuals.
If they are still highly correlated - even after fixing the problem of +1 correlation - then you are left with either combining them (in essence, they mean the same thing to the respondents), or choosing which of the two is most useful to your study. A bifactor analysis might be useful in combining them, while still holding on to some semblance of each construct.
Food for thought ... all the best, Anastasiia.
Thank you a lot for the help!
I have ended up with combining the two factors into one, so the problem has been solved.
Kindly,
Anastasiia
Hi
I need to obtain the level of significance for the correlations between the latent variables in AMOS. How can I do that?
Olá Graça Miranda Silva,
First you need to check the option Standardized Estimates, within the Analysis Properties menu / Output tab. Then run your analysis as usual.
Then in the output (View Text buttom) you need to look for the Estimates sub-menu. You will find several tables, look for two in particular: Covariances and Correlations. In the former you are going to see (for each pair of variables defined to be covariant within the model) the estimate (the covariance value itself), S.E. (the standard error of that estimate), C.R. (critical ratio) and then the p-value of that estimate (that´s the one you´re looking).
In the next table (Correlations), you´ll find the correlations for the same pairs of variables you found in the previous table.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling