6th August 2012 at 11:06 am #2201SFernMember
My sample size is 188 and I want to use SEM to test my hypotheses. I need to run EFA as I have substantially modified mesurement scale of one of my constructs (as a result of a qualitative study conducted first).
So, my question is about possibility (and viability) to run an EFA to make sure the dimensions I have idetified for that particular construct will be supported by the exploratory analysis as well (and perform the necessary modifications to improve the loading etc.) and then move to CFA?
My concern is about the sample factor as it might not be a good practice to use the same sample to do the both EFA and CFA, however I can’t do any other round of data collection equally can’t think of any other options.
Any comments/advice?6th August 2012 at 1:56 pm #2204Iain NobleMember
My answer to queries like the first tends to be: run the analysis and post the results. It’s much easier to comment on specific cases with that kind of detail than debating it in principle. In most cases, the stats provided by the analysis package are what enable people to say whether or not it’s a worthwhile thing to do. The stats police are not looking over your shoulder.
For the second query: again why not? It’s a fairly common professional and academic practice for exactly the reason you mention: people can’t afford to do two waves of data collection and, unless they’re using a validated pre-existing set of factor items developed on other surveys, they have no choice. And even if you’re using a pre-existing set you should run EFA again to see if they work the same way on your sample, you might be surprised to learn how many times they don’t – even with samples drawn by the same method from the same population and the same data collection method.
My main concern here would be sample size, when I did (or commissioned) this kind of thing I was usually dealing in samples of thousands, sometimes tens of thousands and things could be uncertain even with these but, again, you have to do the analysis first to find out.6th August 2012 at 3:28 pm #2203Sunny BoseMember
First you got to tell us what are the total number of items on which you want to run the EFA? Assuming the number is ‘n’ (say) in that case you can run the EFA with a sample of atleast 5n where 5n should not be less than 70. In that case you can split your sample likewise and use the rest of the data for subsequent CFA. However, CFA based on SEM is very data sensitive and therefore, smaller sample size (<200) might throw up some not so acceptable output.
Therefore, my advice to you would be to atleast get another 20-30 respondents. Else you can always replicate data, however, how you deal with this ethical issue in rsearch is totally up to you.6th August 2012 at 8:59 pm #2202Rafael GarciaParticipant
What I usually suggest is that you randomly split the sample, do the EFA on half and then CFA on the other half.
Another approach would be to just look at the correlation matrices. First, check that the modified measure has high reliability (drop items if you need to). Second, look at the correlations between that average modified measure and the construct related measures. If the modified measure is in the wrong direction or near zero, it will not load onto the factor properly, so think carefully about what that means. Looking at the correlation matrices allows you to look at the relationships BEFORE running the CFA.
Alternatively, you can run nested CFA models to test for parsimony (nested model comparisons). The purpose of CFA is to test theoretically-specified models, if you think that they will load onto that factor, test it with the CFA. If the fit is poor, drop it, and see if there is a significant increase in the model fit (Chi-squared).
I don’t see why you need to do the EFA. The CFA will test your theorized relations. If they’re wrong, THEN I would do an EFA.
The general rules I do my analysis by:
Confirmation > Exploration
Multiplism > Singularism
- The forum ‘Default Forum’ is closed to new topics and replies.