I am writing my doctorate thesis and testing the developped hypotheses by calculating a SEM. Unfortunetly, even though I deleted lots of items and even some factors (based on a reliability analysis, an exploratory factory analysis, the measurement model with all items and a confirmatory factor Analysis). However the fit indeces are still poor. Now I really dont know what to do. Ad-hoc modifications (additional paths and co-variances) would be an option I just tried and they enhance the fit indices - however as far as I know this is not a good approach and also from the output I see that some paths and co-variances just dont make sence. Therefore, I am really lost. What would be a common approach? Do I really have to start again and developping a new model oder conducting the Survey again? I would really appreciate any help!
$\begingroup$
$\endgroup$
3
-
$\begingroup$ unfortunately, the entire approach seems not very defensible as described, i.e. if you were to detail this in a ms, it is unlikely to get accepted. Presumably, relationships you were certain existed were not picked up by the model? SEM has linearity assumptions, are those satisfied? What is referred to as 'items', lots of which were deleted? You are right, adding meaningless paths is not ok, so if everything else was done correctly, you may just have to report what you found. But why was SEM chosen in the first place? What is the ratio of sample size/factors? $\endgroup$– katyaCommented Nov 30, 2014 at 19:26
-
$\begingroup$ a poor model fit could be due to small sample size to large number of variables, have you tried regression and path analysis on the basic relationship and see if the direction is okay first? You may need to figure out why the model fit is bad before trying ad-hoc modification. (and ad-hoc modification is bad, as you suggested) $\endgroup$– ceoecCommented Apr 20, 2015 at 17:15
-
$\begingroup$ @katya It is true that SEM is often taught and used within the linear case, and thus believing that SEM assumes linearity is not unfounded, but non-linear cases exist in the literature. $\endgroup$– GalenCommented Apr 23, 2022 at 18:16
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
2
Did you check the normality of underlying variables? Maybe they are terribly non-normal. You can transform using Box-Cox (in Excel), ranking, or the two-step transformation (see https://www.youtube.com/watch?v=twwT6FgwlAo).
-
$\begingroup$ Box-Cox transform is available in SciPy as
scipy.stats.boxcox. $\endgroup$– GalenCommented Apr 23, 2022 at 18:07 -
$\begingroup$ I didn't learn anything about what the two-step transform is from your YouTube video. Rather, it is an SPSS how-to video explaining how to use a software implementation. A reference to Templeton 2011 would be more informative, but explaining what this transform is here would be even better. $\endgroup$– GalenCommented Apr 23, 2022 at 18:11