I've seen some similar questions to this, but nothing spot on so I thought I'd ask.

I'm running an EFA with 54 items (and going with the 3 factor solution as it has the least cross-loadings and simplest structure) and it is largely turning out as I would hope with a few exceptions. Most notably, I have a good bit of cross-loading items. However, often they cross-load with the opposite sign!

So for example, item A should theoretically load on Factor1, which is "positive things" and it does @ .52. It also loads on the "negative things" Factor2, however, @ -.35.

Technically, this is a GREATER than .2 difference and I can ignore it, however, I remember someone once telling me that the signs matter less than the magnitude. So, if I look at just the absolute values, it is less than a .2 difference and I should discard the item. Which is correct?

I am using promax rotation (oblique) with an N of about 400.

• Hi Bonnie, and welcome to the CV community! Could you share a bit more information about why you retained 3 factors--have you considered conducting parallel analysis, nested model comparisons, or evaluating model fit indexes to see if they support your decision? Also, could you describe all of your factors (you say Factor 1 = "positive things" and Factor 2 = "negative things", so what is Factor 3)? Finally, given the oblique rotation, what is the correlation between your factors, especially between Factor 1 and 2? Commented Jan 5, 2016 at 21:23
• Thank you for your response! I retained three factors because the scree plot indicates that this makes sense. In addition, the three factor solution provides the simplest structure with the least cross-loading items. I examined 2, 3, 4, 5, and 6 factor solutions as the measure I am looking at theoretically has 6 subscales (3 positive and 3 negative). Commented Jan 6, 2016 at 15:38
• I have not examined model comparisons / model fit as I have been told by another colleague that it will almost always demonstrate that more factors = better fit, so it doesn't tell you much. Please let me know if this is in error (and if you have a citation that would be great!). If you have syntax for running a parallel analysis in MPlus I would love to try it! I have been unable to find any online. I am using Mplus as it lets you specify your variable type (SPSS treats all as continuous even when specifying otherwise) and I am treating all my variables as ordinal. Commented Jan 6, 2016 at 15:39
• The three factors are (this will not map on exactly to what I said above as that was just an example): Factor1 = Dysregulation (a negative way to respond to threat); Factor 2 = Disassociation (a negative way to respond to threat); Factor 3 = all of the positive ways to respond to threat. As I mentioned, there are 6 proposed subscales. The way they map on here is that all 3 positives load on the 3rd factor, 2 of the negative subscales load beautifully onto Factors 1 and 2, and the 3rd negative subscale is spread across the two. Commented Jan 6, 2016 at 15:39
• Most of the cross loadings I am discussing here are from ONE of the positive subscales (in Factor3) cross loading onto both Factors1 and 2. The correlations between factors are: bw 1 & 2 (the two neg) = .241; bw 1 & 3 = .49; bw 2 & 3 = -.119. Commented Jan 6, 2016 at 15:39

So for example, item A should theoretically load on Factor1, which is "positive things" and it does @ .52. It also loads on the "negative things" Factor2, however, @ -.35.

This is what you would expect.

Imagine that you had two different factors one that represents positive emotion and one that represents negative emotion.

Now imagine we have the following item.

I feel that life is very rewarding. (1-5 likert scale)

We would predict that this item would load positively on our positive emotion factor and negatively on our negative emotion factor.

Technically, this is a GREATER than .2 difference and I can ignore it, however, ... So, if I look at just the absolute values, it is less than a .2 difference and I should discard the item

I don't know whether you should ignore values <.2. I have not heard this particular advice. Whereas, the sign change we can explain the decision on what you should ignore is more subjective.