Confirmatory factors analysis CFA in panel data

[Maria MartinezBravo]

Dear readers,

I am carrying out a study with panel data from 2011 to 2016 for several cities. I created some factors based on the principal component analysis PCA results (command "pca"), on the exploratory factor analysis EFA results (command "factor"), and on the literature review.

However, I want to analyze the goodness of fit indices for the factors (e.g., CFI, RMSEA, factor loadings significance...). To do so, I run the "confa" command and I get always an error. I paste it below. Please note that I called the factors that I previously created f20 and f17 and the name of the items that form the factors start with "ie...".

. confa (f20:ieq11d ieq16d ieq17d ieq110d ieq111d ieq112d ieq29d) (f17:ieq23d ieq24d ieq27d ieq28d)

initial: log likelihood = -4904.2775
rescale: log likelihood = -4904.2775
rescale eq: log likelihood = -303.87926
could not calculate numerical derivatives
flat or discontinuous region encountered
convergence not achieved
r(430);


I was planning to run "estat fitindices" afterwards but the program does not allow me to do it.

. estat fitindices
last estimates not found
r(301);


I have also tried with the options "missing", "from", "level"...and I always got an error.

Am I addressing the issue correctly? Is there any specification for the "confa" command that I am not taking into account?

Thank you very much for your help.
Best,
Maria

[Joseph Coveney]

There's no guarantee that it will converge any better, but try

sem (ieq11d ieq16d ieq17d ieq110d ieq111d ieq112d ieq29d <- F20) (ieq23d ieq24d ieq27d ieq28d <- F17)
estat gof, stats(all)

You could also look into adding a term for covariance of the two factors.

[Maria MartinezBravo]

Dear Joseph,

Thanks for your answer. Effectively, I got the results ! Even if I find a CFI of 0.23. It seems quite low... Any other ideas ?

On the other hand, where should I take the factor loadings from ? From the results of the "factor" command I guess, am I right?

Best,
Maria

[Phil Bromiley]

There is always an issue about modifying models after you estimate them. That said, did you try Joseph's suggestion of allowing the factors to covary?

I think you will find that there are some modification indicators available in the SEM documentation. These suggest changes in your assumptions that may improve the fit. Obviously, this raises the question of data mining, but since you want to get a better fit index it may be what you need to look at. Sometimes you find that you have constrained some things to be uncorrelated that quite reasonably could be correlated. I guess another thing to check would be outliers in the original data. And, there is always the possibility that simply this is not a good model for these data.