Hi Rabia

as I understand that these results for two different samples with different time, and also sample sizes is problematic here (150 vs 760)

to conduct results use EFA & CFA you must do it in a suitable sample size, split it, then analyze first half use EFA, and validate the results by analyzing other half use CFA. 

Rana

If the CFA dont support the original factor structure, then yes you may need to conduct EFA then to be followed by CFA to establish its validity and supporting indeces. Good luck

Thank you all for suggestions

Hi Rabia, 

If the studies are conducted at a comparable time frame and not very far apart (temporally), and there is no reason to expect that there should be difference between them, then it might be a good idea to combine the 2 data sets and then randomly split it into two (like earlier suggested), and conduct EFA and CFA. Hope this helps, good luck.

Thanks Pat for your kind suggestions. A better fit has provided higher TLI and CFI values (.94) than other models (.91) tested for the sample. RMSEA was .04. I have used Cronbach's alpha to compute reliabilities. I dropped the items for conflict scale that showed poor reliability (.18 only) and merged the two sub scales which had high multicollinearity (.92). resulting in 2 factor solution having good fit and reliabilities higher than .70. I wanted to confirm if if i did the right thing. 

Dear Rabia,

My responses may be redundant but here are my suggestions. I assume both data sets measure the same thing, then you may combine them to make one large sample whose sample size becomes N=820 (=150+670). Then you may randomly split them into two samples (e.g., N1=410 and N2=410). You may use N1 (the first half randomly split from the original sample) as a calibration sample and N2 as (the other half randomly split from the original sample) as a validation sample.

With N1, you systematically check with 2-factor, 3-factor, 4-factor, and 5-factor solution with fit indexes. However, you have to note that if your input variables' scores are categorical (e.g., Likert scale scores or any types of ordered categorical values, such as "1", "2", "3", "4"), then you have to use polychoric  correlations to identify your factor solution (either LISREL or Mplus computes correlations for categorical variables). Otherwise, your solution could be underestimated or biased. Once you choose the best solution, you confirm the factor solution (from N1) with N2 (your validation sample). When you do validation with N2, you may choose the "bootstrapping" option, if available, to check stability of your estimates (e.g., factor loadings).

Good luck!

Hello to all participants.

I find it very interesting discussion. I want to ask to Patrick about some reference on methodological literature to which you are refered about the limitations of the split method.

And, secondly, I want to recommend to Rabia the FACTOR 9.2. program, which can be found and downloaded for free (at the link below) and would allow developing many of the methods that have been suggested. You could make an EFA from the matrix of polychoric or tetrachoric correlations, would have several methods to estimate the adequate number of factors to retain, to obtain measures of internal consistency estimates for the factorial solution, you would have several methods of rotation and could even perform a semifactorial analysis. In addition, the page has good methodological information on the program itself (easy to use) and on appropriate statistical in each case.

I hope you find it useful.

Best regards.

http://psico.fcep.urv.es/utilitats/factor/Documentation.html 

If you mean that the measure was originally published with five factor model with good fit, then you are getting unexpected results that may be explained by particularities of your sample.  I would suggest screening the sample for multivariate outliers within the unreliable factor. If you say that the fit is beter in the 5-factor model and that the loadings patterns correspond to theory, you are fine, but make sure you have a sample 10 times the number of items in your measure. 

Although you did not need to run an EFA if the measure has been published before, an EFA on the big sample can help you explain why the CFA results looked bad, and suggest alternative hypotheses regarding the number of factors to extract. Maybe the loading patterns in a 3 or 4 factor solution will make sense and thus be more parcimonious than the 5 factor solution. 

Maybe EFA is more adapted at the current stage of your research? Take advantage of the big sample to assess the factor extraction and loading patterns properly with sufficient sample size, and keep CFA for later. Besides the use of CFA to confirm EFA results is debated (consult the articles below):

http://pareonline.net/getvn.asp?v=17&n=15

https://www.researchgate.net/publication/241866804_Confirmatory_analysis_of_exploratively_obtained_factor_structures

http://pareonline.net/getvn.asp?v=17&n=15

Article Confirmatory Analysis of Exploratively Obtained Factor Structures

look at this paper. It answer your questions.

https://www.researchgate.net/publication/269039467_Studying_the_Factorial_Structure_of_Ghanaian_Twelfth-Grade_Students_Views_on_Mathematics

Chapter Studying the Factorial Structure of Ghanaian Twelfth-Grade S...

The following research papers may answer your questions.

https://www.researchgate.net/publication/269039467_Studying_the_Factorial_Structure_of_Ghanaian_Twelfth-Grade_Students_Views_on_Mathematics

Chapter Studying the Factorial Structure of Ghanaian Twelfth-Grade S...