1
$\begingroup$

I am facing problem in getting the good fit indices. I have only 163 respondents with a total of 4 variables and 70 items (A: 5 components, 3-4 items each; B: 3 components, 8 items each; C: 4 components, 4-5 items each, D: 11 items). The initial values for fitness indices were:

ChiSq P-value = .000; RMSEA = .071; GFI = .586; SGFI = .558; CFI = .675; TLI = .663; NFI = .488; ChiSq/df = 1.815

Then, my supervisors suggested me to run EFA for the model but it ended up the items in the same group were being distributed to different groups, where the grouping of those items doesn't make sense at all as I adopted these constructs from past literature. The number of factors was not the same as the one proposed in past literature. Q1: Should I modify the model according to the EFA result?

Without modifying the model, I tried to do CFA for each construct separately and deleted items by referring to Modification Indices (MI). And I ran CFA for entire model again by deleting what I've deleted in the last step. However, the fitness indices are still poor. If I further delete items with highest MI, the result is even worse.

Then, I started over again by using Cronbach's Alpha to identified the items with lower reliability and deleted it. And I repeated the same steps by running CFA for individual constructs, deleting items according to MI, running CFA again for whole model, and deleting items according to MI. I've deleted 17 items from the model. The final result is as below:

ChiSq P-value = .000; RMSEA = .060; GFI = .690; SGFI = .660; CFI = .831; TLI = .822; NFI = .649; ChiSq/df = 1.587

Q2: May I know is there any other way that I could do to improve the fitness indices? If the problem is with my data, would I have to give up using SEM for analysis?

Thank you!

$\endgroup$
1
$\begingroup$

Don't use GFI, it's pretty frowned upon.

You have two problems here. First, your sample size isn't very large, and you have a very large model.

Second: Your RMSEA is not bad. But your CFI is dreadful. This tells me that your null model (i.e. the model that has no relationships between the variables) is not very bad. Try fitting it - a model with no parameters (except variances). This should be a completely appalling model - I'm going to guess that yours is not. This means you have low quality data - the variables are not sufficiently highly related. Perhaps they're not reliable, perhaps they're just not related. In addition, with this small sample size, your chi-square should be better. (Another tip: don't tell us chi-square/df, tell us chi-square and df - with that information I'd know more about your model).

In summary: Stop trying to fit this with SEM, it's never going to work. Sorry to be blunt, but you have (a) not enough data, (b) the data you have is low quality and (c) your model is very poor.

$\endgroup$
0
$\begingroup$

You can try pool CFA, or using original dimension (Please check orginal article of your instruments esp result on EFA)or formative construct

$\endgroup$
  • $\begingroup$ Welcome to the site, @Fatin. This is rather sparse by our standards. Can you expand on your points here? $\endgroup$ – gung - Reinstate Monica Nov 7 '16 at 20:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.