I have the following problem:
for my master thesis I am, or rather was trying to validate a scale on relationship cognitions in a different cultural context (germany/western vs. malaysia/asian) Since Malaysia is a multicultural country I want to compare the different dimensions found with the scale between the different cultures found in malaysia after this step.
As it turns out neither the cophenetic correlations between the factor matrices nor the result of the CFA indicate that I can keep the scale as it is, so so much for validating the scale. To go on with my plan to comapre the different ethnicities to each other I now need a new scale. So, I went back to EFA:
The sample size is 497, KMO = .962; Bartlett seems to be okay as well. However, as is always te problems with data in relationship studies it is really skewed. The scree-plot indicated 4-factors. I also performed velicer's MAP and parallel analysis and they recommended 5, however with 5 factors there is no loading above .5 on any item. There are 64 Items. The original scale also consisted 0f 4 dimensions. From a descriptive perspective the underlying themes of these dimensions do not seem to change.
So I built a new second-order model based on the EFA with the items achieving a factor load and it still only achieves a barely even acceptable fit. CMIN/DF: 3,032 CFI: ,846 RMSEA :,068 confidence interval:,065 to ,071 Bollen-Stine Bootstrap: .003 All of the regression weights were significant (p<.001), though Bootstrap- adjusted two tailed significence indicates that not all of them have a p-value below .01
I placed some restrictions on covariances indicated by M.I.s. but if they made sense witb the items and also deleted some items ( I use residual covariances above 2.58; number of M.I.s, bootstrap corrected p-values, communalities, the factor loeads from the EFA and theoretical considerations to chose the right items to delete).
After deleting the first 4 items however it feels rather random which ones I delete and sometimes the fit doesn't even improve afterwards. Does anyone have some tips on how to chose the right items to take out of the model in a more scientific manner? Or is this just one of those trial-and-error-things? I am using AMOS software
The fit has become better, but not good: CMIN/DF : 2,419 CFI :,906 RMSEA : ,058 confidence intervall ,055 to ,061 Bollen Stine Bootstrap = .003