There are no items reverse-coded. Entirely different variables load negatively when I change the method of extraction from Principal Axis to Maximum Likelihood (both Direct Oblimin rotation). And there are no negative loadings at all when I do Principal Component with varimax rotation. The items are constant, there is no reverse coding done...why should I get negative loadings?
I am developing a new scale and my preferred method for this work is ML. How should I treat the negative loadings?
Insights and help will be deeply appreciated.
Warm regards,
Vikas.
If all of the factor loadings for a given factor are negative, that is not a problem at all, as long as that factor is negatively correlated with the other factors, (check the factor correlations matrix, it is the last table in the SPSS Output for oblique rotations). You can just ignore the differential signs of the loadings, which are mostly arbitrary (If only some of the items of a factor were negatively loaded, that would be a problem).
But you have a bigger problem. Your factor analysis results are not very stable. Your factors end up reflected by different set of items depending on the extraction method, which might indicate a small sample size or a poor item quality.
Make sure you are looking at the loadings after rotation and not the initial loadings before rotation.
Thanks Paul for your response. I have checked and have considered the loadings after rotations. For lending clarity to my question, I am attaching an excel file that has three sheets, each showing different extraction method (ML, Principal Axis, and PCA). Observations are: ML loads negatively with a set of items/factor, while Principal Axis loads negatively with a DIFFERENT set of items/factor. PCA does not load negatively with any factor/items.
I have no item that is reverse coded. Are negatively loaded items to be deleted? And if they are to be (I am not convinced they should be), different extraction methods suggest different outcomes.
I will deeply appreciate insights on how to go about.
Regards,
Vikas.
In general, Principal Components should be used with orthogonal rather than oblique rotations, due to its underlying theory about the nature of factors.
Often, experimenting with different combinations of extraction and rotation methods is more likely to lead to confusion rather than clarity. If you are doing social science research, I personally prefer ML and oblique rotation.
Thanks, David. I am attaching a file for your reference that has three sheets, showing the outputs of ML with oblique, Principal Axis wit Oblique and PCA with Orthogonal. I am doing research in social science and as suggested by you, will prefer ML with oblique rotation.
How should I treat the negative loading? Four variables are loaded negatively and range from -0.522 to -.0.794. My questions are not reverse coded. Will deeply appreciate your insights.
Regards,
Vikas.
If all of the factor loadings for a given factor are negative, that is not a problem at all, as long as that factor is negatively correlated with the other factors, (check the factor correlations matrix, it is the last table in the SPSS Output for oblique rotations). You can just ignore the differential signs of the loadings, which are mostly arbitrary (If only some of the items of a factor were negatively loaded, that would be a problem).
But you have a bigger problem. Your factor analysis results are not very stable. Your factors end up reflected by different set of items depending on the extraction method, which might indicate a small sample size or a poor item quality.
The explanation may be that principal component analysis (PCA) is not the same type of analysis than the others. Some say PCA is not factor analysis at all. What it does is to rearrange the data, but not consider measurement error which factor analysis does.
- Badges
- Science topic
- Similar topics
- Psychology
- Psychometrics