Dear Experts
I collected data from managers on evaluation of their subordinates' performance. The subordinates' performance was evaluated on 4 factors i.e. knowledge, skills, abilities and personality.
I performed confirmatory factory analysis (CFA) in AMOS and found high covariance between four factors i.e. 0.93 between factor 1 and 2, 0.96 between factor 1 and 3, 0.91 between factors 1 and 4, 0.88 between factors 2 and 3, 0.88 between factors 2 and 4, 0.91 between factors 3 and 4. Factors loading are perfectly fine ranging from 0.8 to 0.91 (except two with loading 0.78 and 0.73). Moreover, model fit indices' e.g. GFI, CFI, TLI, CMIN/DF, RMSEA, and SRMR values are fully complying with standards requirements which gives me impression that model is perfectly fine.
How should I justify the high covariance between factors considering that these factors are highly similar and I cannot remove anyone from them.
Thanks for your support.
Dear Jawad, it is now common practice to test not just one model, but to test a small number of alternative or competing models. For example in your case you could test a one-factor model, a four-factor model, and a second-order model. It would probably be helpful if a second order model fitted the data as this would support the use of an overall scale score, using four separate sub-scale scores that are so highly correlated would cause collinearity problems if these are predictors and produce very little unique predictive utility for each of the sub-scales.
It's possible that the situation arises where the statistical model is indicating multi-dimensionality but in reality the use of multiple dimensions offers no real benefit over using a total scale score. This is an awkward position to be in where practical considerations have to over-ride statistical findings. This is exactly where I found myself in the following paper
www.researchgate.net/file.PostFileLoader.html?id=562dff075dbbbd4c898b45bc&assetKey=AS%3A288750878904323%401445854982971
Dear Jawad, it is now common practice to test not just one model, but to test a small number of alternative or competing models. For example in your case you could test a one-factor model, a four-factor model, and a second-order model. It would probably be helpful if a second order model fitted the data as this would support the use of an overall scale score, using four separate sub-scale scores that are so highly correlated would cause collinearity problems if these are predictors and produce very little unique predictive utility for each of the sub-scales.
It's possible that the situation arises where the statistical model is indicating multi-dimensionality but in reality the use of multiple dimensions offers no real benefit over using a total scale score. This is an awkward position to be in where practical considerations have to over-ride statistical findings. This is exactly where I found myself in the following paper
www.researchgate.net/file.PostFileLoader.html?id=562dff075dbbbd4c898b45bc&assetKey=AS%3A288750878904323%401445854982971
Hi Jawad,
You might also check the convergent and discriminant validity of the factors from your CFA (see James Gaskin's video on this: https://www.youtube.com/watch?v=yk6DVC7Wg7g ). This might further confirm Mark's excellent answer (his article was also quite interesting and germane). Good luck with your study,
Nick
Hello, i think you should calculate convergent and discriminant validities? In some cases latents are combined as 1 if they are not different enough
- Badges
- Science topic