Dear Robert

off the top of my head, I can only think of one reason: You have extracted only one factor. Communalities refer to single items. They measure what variance of every item is explained by all extracted factors. It stands to reason that with only one extracted factor, both initial and extracted communalities are identical which would explain their omission in the output. I am unsure, however, if that holds up for ML-extraction.

Best wishes

Marcel

Dear Marcel,

Thanks for stepping in. In this case, we have two quite strong factors (established by parallel analysis), but some of our output contains both types of communalities, whereas some output doesn't.

The mystery remains!

If you suddenly get another idea, please feel free to share it. :-)

Robt.

Dear Robert

you're welcome. It's a pity that I couldn't solve your problem. Is it possible to upload screenshots of the output-files? They may shed some light onto what went wrong. You could also try principal axis factoring instead of ML if that's applicable.

Kind regards

Marcel

Dear Marcel,

Thanks. Pls leave it with me for a little while. I'm getting someone else to "press the buttons" in SPSS for me/us (I don't have access to SPSS, but I'm familiar with what to do - so that's how we're operating), and I'm suddenly wondering whether he's not passing on all of the output to me. He knows how to press buttons under instruction, but he doesn't always realise what's important to transmit.

We've done PAF as well as ML, but the problem seems to be sporadic in both extraction methods.

Watch this space! - but it might take a while for my friend to respond . . .

Robt.

Marcel Grieger. G'day again, Marcel. I'm attaching one set of output. This is from a 40-item scale on which we had already found, from parallel analysis, that there appeared to be two factors in the data - though some cross-loadings appeared once we did this EFA.

The more important thing is that there wasn't a column of extracted communalities. That's what was mystifying me.

Incidentally, none of our items were skewed or kurtotic, and none of them had outliers, so the data were pretty "healthy".

Feel free to comment on anything!

Thanks,

Robt.

Robert, why is this two-factors? The scree suggests 1 to me (and with 40 items the eigenvalue=1 criterion isn't that useful). Also given the results factors are so highly negatively correlated it seems unless there was a particular model you were testing there is not much to argue for 2 factors.

Daniel, thanks for coming in on this. I appreciate the points you're making (in both senses of the word "appreciate").

I, too, thought that the scree test indicated only one factor in the data, but, because we were a bit uncertain, we did a parallel analysis and it came back indicating that we had two factors. Furthermore, those two factors end up being interpretable as two somewhat separate domains: one about actual teaching in schools and the other more about the professional "ethos" of being a teacher.

You might notice that the interfactor correlation was .81, which suggests a single factor. So, I'm not sure which way to bounce. Any suggestions welcome.

Incidentally, I think that we've managed to put together a set of 40 items without getting overt redundancy. It took a lot of reading and thinking about.

It sounds like you have enough ideas about how things should load that a confirmatory approach might be justified.

Hi Robert Trevethan ;

Someone has already experienced this problem. Could it be from SPSS? The link is below:

https://developer.ibm.com/answers/questions/443023/pca-communalities-table-shows-initial-values-only/

You want to re-try the analysis with Factor software. Parallel Analysis with optimal implementation can be conducted via Factor.

Bests

Daniel Wright. Thanks, again, Daniel. Sorry for not replying sooner, but here in Australia it was late when we were communicating (about 4 a.m., in fact), and I was in bed by the time you posted your most recent message.

For what it's worth, the output I posted was from inservice teachers. I've just finished writing up the CFAs for the preservice teachers in our research, who had essentially the same results but with no cross loadings. These CFAs were based on two factors and the items in the table I posted. (We had enough participants - originally more than 1,300 - in order to do EFAs with one half and CFAs with the other.) The RMSEAs were not brilliant, but the normalised chi-squares, TLIs, CFIs, and SRMRs all indicated model fit.

I am a bit mystified, as it seems, you are. I'm not sure what to do, particularly given that the factors are so highly correlated and the items seem to load pretty highly on their separate factors with only a few cross loadings on the inservice teachers' data and none in the preservice teachers'.

Back to my original question, however: I notice that someone else has stepped in with a suggestion. I'll follow that up.

If you suddenly get any additional insights, please feel free to share them.

Robt.

Abdullah Faruk Kılıç. Thanks for your suggestion above, Abdullah. I'll follow it up.

Cheers.

Sleep is important! I'd look more at the one factor model. See which don't load on the general factor. Also see where the implied and actual cor mats differ. I am thinking it may only be a few items not fitting. Also, it is not the factor model that is difficult, it is the rotation. Does an orthogonal rotation make the loadings still sensible?

Daniel Wright. Thanks for your suggestions (and I've had a couple of good sleeps since our last posts). I guess a variety of things are worth trying in order to ensure that no stones have been left unturned. First, and going back to my original problem, we have redone the analyses in which the extracted communalities had not appeared, and the second time around they miraculously did appear. (What's going on there? - so Marcel and Abdullah, thanks for offering to help, but things have self-healed . . . )

Second, although the interitem correlations across the board were high (I'd checked that out beforehand, and that's mainly why I'd gone for an oblique rotation originally), I was willing to try an orthogonal rotation as per your suggestion. We used varimax rotation, and the results were pretty horrible, primarily in terms of cross loadings.

So, it seems that an orthogonal rotation based on the two factors that parallel analysis had indicated were present in our data was, indeed, the most appropriate way to go. I'm just not sure, now, what to do given that the correlation between the two factors is hovering around the .80 mark.

Thanks, again, for your help.

Robert Trevethan , sounds like a bi-factor model is worth trying. There are exploratory versions of that if you need exploratory (e.g., Article Exploratory Bi-Factor Analysis

Daniel Wright. Thanks for this suggestion. We'll try it out.

Robt.