xiangyu cong posted on Wednesday, December 05, 2012 - 6:55 am
Dear Dr. Muthen:
I am doing EFA with 31 categorical indictors.
After first run , I think 8 factors is good fit. Among 8 factors extracted, only one factor has three negative loading with absolute value greater than 0.4. After I removed 7 indicators that did not have any significant loadings on any of the 8 factors, suddenly, there are three factors have negative loadings. It is very hard to interpret the results.
If the signs of all loadings switch from positive to negative, this is not a problem.
xiangyu cong posted on Wednesday, December 05, 2012 - 11:56 am
Thanks so much for your reply.
Through EFA, I now have 8 factors and 24 indicators. The factors we identified through EFA are obviously correlated.
Next, I will run CFA with my model based on my EFA results. But, two of eight factors only has two indicators with loading higher than 0.3. Should I include additional indicators for these two factors (maybe lower the cutoff from 0.3 to 0.2) to make them each has at least three indicators each?
I guess short version of my question is: "Is it OK that two out of eight factors in my model only has two indicators each in CFA?"
I would never recommend less than four factor indicators. This is a good question for SEMNET where I think there has recently been a discussion of this topic.
xiangyu cong posted on Thursday, December 06, 2012 - 8:10 am
Thank so much! Linda. Your comments are deeply appreciated.
I am going back to my EFA results and refine my model before I going into CFA.
I will look into SEMNET for the discussion you mentioned.
Daniel Lee posted on Monday, August 07, 2017 - 11:10 am
Hi Dr. Muthen,
I found several negative factor loading in a bi-factor EFA. The items in the scale are all very similar however (i.e., very positively correlated and non are in reverse). I was wondering how this might be possible. If this is potentially a product of model misspecification, do you know of any papers that can support this possibility?
I'm not familiar with that result. Try Starts = 10. Also, this may perhaps be more unusual for the general factor than for a specific factor given that the latter considers residual correlation after the general factor has been taken into account. For the general factor I would think the loading signs would reflect the sample item correlations but for the specific factors I'm not so sure.
Using EFA for categorical indicators, we settled on an 8-factor structure. Item B21 was adopted onto a factor, with a negative loading. We intended to use equally weighted subscale scores rather than factor scores, so we reverse-coded B21 when summing the scores. In order to check the correspondence between the factor score and the subscale score, we ran a CFA on the raw, unrecoded scores (same data as the EFA). Obviously, the fit was good, but the sign of B21 loading had flipped to positive.
I double checked that I was not accidentally using the reverse-coded data. The other relevant piece of information is that we only adopted 3 items onto this factor (not good practice, I know!).
Can you think of why this sign-switching might occur, or what aspects of the model I should check to find out?
ETA: I think I figured out the problem. I had omitted an item from the factor because it's too-high correlation with another variable was causing problems elsewhere in the model. The item I deleted also had a negative factor loading. When I re-ran the EFA without that deleted item, the loading of B21 was reversed (positive, as in the CFA).