Hi everyone. I'm running an exploratory factor analysis on data from a Paranormal Belief Scale of 26 items, with a sample size of 1373 subjects. Using the principal axis extraction method and applying varimax and oblimin rotations, I found that the same subset of items load to the same factor, but with the oblimin rotation, the coefficients are negative.
This set of items corresponds to a subscale of the questionnaire in the original version. My question is: how do I interpret the change of coefficient values when applying varimax and oblimin rotations?
P.S.: I'm using SPSS.
Hi Angel,
Note that Varimax is orthogonal rotation (which is used for uncorrelated factors) and Oblimin is oblique rotation (used for correlated factors).
That your data coefficients turned negative when you applied Oblimin rotation showed that your data are uncorrelated, hence you should stick to Varimax or Quartimax.
However, if you are interested in seeing different pattern and loadings, you could change Principal Axis Extraction to Principal Component Extraction, but maintain the orthogonal rotation.
Good luck.
Dear Angel,
I'm a little bit perplexed by your question. I would have expected the items with the largest absolute values for their loadings on the first factor to have been positive. However, it is the relative sign that counts anyway so it is not really an issue. Are you looking at the pattern matrix rather than the structure matrix?
Also, did you consider trying to extract more than one factor?
Finally, I suggest you should provide an argument as to which rotation is preferable. We recommend starting with an orthogonal one, such as promax, then comparing the withing factor correlations to the between factor correlations. If the former is not significantly higher than the latter then we recommend considering an oblique rotation such as oblimin. However, you are currently only extracting a single factor so this is not relevant. To be honest, if you are only extracting one factor, I cannot see the need to perform any rotation.
Thank you both for your prompt answers, I think I should clarify a couple points:
1) I'm adapting an already validated scale (Paranormal Belief Scale) to a different population, and I ran the exploratory analysis to see if the scale factorial structure turned out similar. The scale contains seven subscales originally.
2) I tried both rotations because I wasn't sure if my factors would be correlated. Since I'm measuring a psychological construct ("paranormal belief") and all subscale items correspond to different aspects of the same trait (traditional religious belief, witchcraft, superstition, etc.) I expected to find a high degree of correlation among factors. However, I wanted to see if the factorial structure was similar with both methods.
3) I didn't set up an a priori number of expected factors, I kept the criteria of higher than 1 eigenvalues, checked the scree plot and: 1) the rotated factors matrix in varimax rotation, 2) the pattern matrix in the oblimin rotation. In both cases I obtained a 6 factor structure that explained 58% of the total variance (but factor 6 only has 2 items and it's not suitable for a subscale). Now, applying the varimax rotation, I find that items 24, 10, 3 and 17 all load to the same factor (3) POSITIVELY, while in the oblimin rotation, THE SAME ITEMS load to factor 5 NEGATIVELY. I need to determine with method is more reliable and how do I interpret the negative factor loadings in the oblimin rotation.
In addition, I choose PRINCIPAL AXIS FACTORING over MAXIMUM LIKELIHOOD as extraction method because both the response distribution of every item and the global scale score were non-normal.
- Badges
- Science topic
- Similar topics
- Analytical Chemistry
- Extraction