I collected a data of 305 people regarding their leisure time activities, to compile these, I tried to analyze using Factor analysis, wherein, I tried both PAF and PCA, PCA is giving better results, but two items, about watching news and sleeping are loading with the factors/components that are about socializing. Practically, this doesn't sound justifiable. What statistical method can help me get better results?
Hi,
Delete those items and do the SEM once again to check the factor loading. Items for a construct should be minimum 2.
Hope that helps.
Hi Aachal Taywade
This sort of problem in EFA may arise because of certain issues such as wording of items or content validity issues. I would suggest to go for Principal axis factoring or maximum likelihood instead of PCA as PCA is not considered as a method for EFA. Further, you could try for rotation as well. Do check the communality values. If below .30, you could make a decision of deleting those items.
I think you should examine your correlation matrix, which is the ultimate source of your factors. In particular, if the item for sleeping is indeed highly correlated with the items for socializing, then that is what is the data.
You should take into account the parameters for performing the AFS and the methods of extraction and rotation according to the nature and distribution of normality of the data, then also consider the CFA for the determination of the theoretical-psychometric model, if your instrument has previous studies of CFA may be more viable this analysis.
As David L Morgan says first start with scatter plots, examine these, and then correlation matrices if appropriate. THere is a lot of preliminary work to do prior to conducting a latent variable model.
The advice to simply get rid of variables or data or whatever until you get results you like is not a wise approach.
Factor analysis computes eigenvectors and eigenvalues based on the sample correlation matrix from which we calculate factor loadings. Thus, as already rightly suggested, it is a good idea to examine the sample correlation matrix first. Having said that, since you are working in an exploratory manner, one suggestion could be to estimate your factor model by imposing different number of factors to extract. You can then compare the different factor solutions.
Simple question: Both are loading on the same construct, but do they have identical signs as well. If one is negative and the other is positive, it could have an explanation. When you sleep, you can't socialize and the sign could be 'negative', but watching the news could be 'positive'.
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics