Hi all,
Need your suggestion on my research.
I am working in CFA which is derived after EFA with 2 factors having 7 and 8 items, but issue is in CFA, AVE is coming less than 0.5 hence I am forced to delete the items with less than 0.4 standardized beta. But now only 4 items remain in the each factor {and all model fit indices , AVE and CR is in acceptable range }against 7 and 8 which where derived from EFA.
Is this acceptable to delete so many item after EFA?
Note : EFA and CFA is done on different data.
Kindly suggest.
Uday Arun Bhale Below are some of the links to previously asked similar questions on research gate.
1. https://www.researchgate.net/post/Am-I-allowed-to-delete-factors-in-CFA-in-order-to-improve-model-fit
2. https://www.researchgate.net/post/How_to_purify_constructs_drop_items_through_confirmatory_factor_analysis
3. https://www.researchgate.net/post/Can_we_remove_items_displaying_poorFactor_Loadings_in_SEM
Hello Uday,
First, do recognize that the "stability" of a given structure across samples (even if the samples are drawn from the same population) is less than 100%; sometimes considerably so. So, the idea that you get one result from an EFA (with one sample), and seemingly different results (see below) with another should not come as a complete surprise.
Second, I'm assuming that you decided the EFA structure you adopted made sense theoretically and rationally as well as meeting suitable guidelines (such as how well the resultant solution could reproduce the observed inter-item correlations). As well, you don't indicate what the item/variable-factors were after rotation, so we have no way to judge whether the AVE values you obtained from the CFA were consistent or wildly inconsistent with the EFA results.
You didn't indicate why it is you believe that a "good" model must have variable-factor correlations of about .71 or better, on the average. There are all kinds of suggestions / guidelines / recommendations for what comprises an appropriate threshold for variable-factor correlations to be considered salient. Cattell offered one example in which loadings as low as .15 would be considered constructive for factor identification and affiliation.
There are a number of other goodness of model-data fit indices in the CFA/SEM world that inform about the adequacy of a given model for accounting for the observed relationships among the measured variables (e.g., RSMEA, RSMR, and a host of others...see Kenny's brief overview of fit indices: http://davidakenny.net/cm/fit.htm).
If you jettison 50% (or so) of your items, there is the clear possibility that you are materially changing the meaning of your obtained factors. As these decisions are changing your structure, you'd really be best served by obtaining yet more data to see whether the second model holds up. So, I'd first have a look at other model-data fit indices first before tossing out indicators left and right.
Good luck with your work.
You're good to go at least 3 items for individual construct. You might have 2 unless you have big samples (eg. More than 500).
Hi David,
Thanks for your contribution.
In my case ,Issue is with AVE which is not qualifying the 0.5 criteria. Hence I need to delete the items. Is AVE less than 0.5 is acceptable in CFA? Any supporting journal on this ? In my case it's coming as 0.30 without deleting items.
In practical conditions, the AVE value of above 0.45 is still tolerable for a development of a new measure (Netemeyer, R. G., Bearden, W. O., and Sharma, S. (2003). Scaling procedures: Issues and application. Thousand Oaks: Sage Publication.)
- Badges
- Science topic
- Similar topics
- Correction