Question

I am running an EFA to examine factor structure of a measure of alcohol expectancies. When interpreting the factor structure, I noticed several high, negative loadings. These were all reverse scored items (which I have double checked--I reverse scored them prior to the EFA). Is there any obvious explanation for this? Since they are negative loadings, it does not make conceptual sense why they loaded onto this particular factor. So I want to make sure I understand before I just eliminate them.

Ezgi Toplu
MEF University
Hi, Caitlin. I experienced something similar to your problem. In my case, factor 1 and factor three were highly correlated to each other (r = .81). Although the items in factor three were not reversed items (I had no reversed item), they negatively loaded onto factor three. Then, I forced a two-factor solution, all the negatively loaded items, this time loaded positively onto factor 1 and, the problem was solved. Have you checked the correlations between the factors in the correlation analysis in SPSS? Maybe this helps.

## Top contributors to discussions in this field

University of Gujrat
I recommended to use varimax rotation while processing EFA. This Will resolve the issue of negative loadings hopefully.
1 Recommendation
David Morse
Mississippi State University (Emeritus)
Hello Caitlin,
Negative polarity items are often troublesome in perception measures. First, they don't behave in arithmetically complementary ways (e.g., the mean score of the item, e.g.: "I love bananas" is almost never equal to the the complement of the mean score for the 'opposite' item: "I hate bananas." Second, it's not uncommon for negative polarity items/stimuli to "appear" to form a separate factor from that of the positive polarity items (ostensibly assessing the same target), when the more parsimonious explanation of a bipolar factor would frequently be expected. The usual justification for negative polarity items is to assure either some protection from acquiescence effect, to "force" respondents to attend to wording more carefully, or the like. The reality is, negative polarity items can impose a higher cognitive load on the respondent, as well as introduce the technical concerns mentioned above.
It is possible for a single factor to have both positive and negative affiliations with salient variables, so that outcome, by itself, isn't a reason to worry.
If your factor solution has just one factor, you may reverse the signs of all the loadings without affecting the adequacy of the solution for explaining the observed relationships among the variables. In a multifactor solution, single factors may also be reversed, but you would then have to account for revised correlations among factors (as well as signs for cross loaded variables).
If the signs of the loadings do not correspond to your expectations about the indicator/observed variables, relative to the loadings of other indicator/observed variables for the same factor, then you may have a problem. That could be due to any number of things, including respondents' not understanding the wording of the item; careless reading; double-barreled items; the score recoding was incorrect or incomplete; etc.
2 Recommendations
Marcel Grieger
Georg-August-Universität Göttingen
varimax employs orthogonal rotation and is not considered to be a "real" EFA, so do not believe that this would solve your issue if there even is one.
I know we are talking EFA here, but the idea behind any latent factor in a reflective measurement model is that your latent variable explains the manifest indicators. If you happen to have items that loads negatively onto your factor that would imply a negative correlation between the two. You stated that you recoded the items. To me, this seems to be the cause. Allow me to present an example:
Say you want to measure "optimism" and your indicators are:
- I always look on the brigt side of life. (1. "do not agree"....4. "fully agree")
- I consider myself an optimist (1. "do not agree"....4. "fully agree")
- I cannot wait to face new challenges (1. "do not agree"....4. "fully agree")
- I really see no point in going on (1. "do not agree"....4. "fully agree") --> recode for analysis
In conclusion, as far as the value of factor loadings is concerned, positive or negative does not make any difference, they are still correlations between the factor and the indicators. The interpretation, however, is up to you.
Therefore, I do not believe you need to remove any of your items.
Please, fellow researchers, do correct me if I am mistaken.
Best
Marcel
Robert Trevethan
Independent author and researcher
Caitlin Wolford-Clevenger, a colleague and I had a similar problem to yours when we used direct oblimin rotation. When we used promax rotation at the suggestion of someone on RG, the problem disappeared and everything fell into place. (Don't ask me why; it just did.)
I recommend not using varimax rotation. Many researchers use it, I suspect because they simply follow what other researchers do, but also probably because they think they will get a cleaner separation of factors.
In my experience, varimax rotations often produce muddier outcomes - and they are usually inappropriate, anyway, because many of the things/items we factor analyze are to some degree correlated with each other.
If you feel inclined, I'd be interested to hear whether using a promax rotation solves your problem. I certainly hope it does.
Ezgi Toplu
MEF University
Hi, Caitlin. I experienced something similar to your problem. In my case, factor 1 and factor three were highly correlated to each other (r = .81). Although the items in factor three were not reversed items (I had no reversed item), they negatively loaded onto factor three. Then, I forced a two-factor solution, all the negatively loaded items, this time loaded positively onto factor 1 and, the problem was solved. Have you checked the correlations between the factors in the correlation analysis in SPSS? Maybe this helps.

## Related Publications

Chapter
In diesem Kapitel wird auf die wichtigsten Aspekte bei der Durchführung einer EFA eingegangen. Es wird mit der allgemeinen Modellvorstellung in der Faktorenanalyse begonnen (Fundamentaltheorem), sodann wird die darauf basierende Varianzzerlegung in durch gemeinsame Faktoren erklärte und unerklärte Anteile dargestellt. Anschließend werden die zentra...
Article
Bilinear data matrices may be resolved into abstract factors by factor analysis. The underlying chemical processes that generated the data may be deduced from the abstract factors by hard (model fitting) or soft (model-free) analyses. We propose a novel approach that combines the advantages of both hard and soft methods, in that only a few paramete...
Article
The bleaching of indigo-carmine catalyzed by laccase was monitored on-line by UV/Vis spectroscopy. The resulted data were analyzed by evolving factor analysis (EFA) and multivariate curve resolution (MCR), and the reaction mechanism was proposed.
Got a technical question?