The model in the picture indeed looks like it should be over-identified (and far from being under-identified unless many of the indicator correlations are close to zero).

Negative correlations of indicators within the same factor should simply result in some of the loadings being negative. However, some SEM programs may run into convergence problems when some variables have positive and others negative loadings. This can be addressed either by recoding items or by providing user-defined starting values for the loadings. In case this doesn't solve the problem, could it be a data error? How strong are the correlations of the indicators within the same factor?

Thank you Christian for your prompt feedback. Recoding would not be an option, as the data is already collected and this is for a thesis.

How do I know, which starting values to set? Could you provide a source, so I can read on that procedure?

For the correlations between items for one factor see files of the SPSS output.

Hi Jonas Schmitt

why would recoding not be an option? You can easily do this after the data has been collected by re-arranging the numbers on the response scale so that all are in ascending or descending order (i. e. 1 = low, 4 = high). Whether that would solve the problem at hand - I concur with Christian Geiser , under-identification is not the issue here - I cannot say for certain. It would not hurt to give it a go. On a different note, your inter-factor-correlations appear to be rather low. For example "When I am at work, I have to do what I am told" only shows one significant correlation with all the other items within the same factor.

Best

Marcel

Hi Marcel,

maybe I did understand that option wrong. I thought recoding meant to recode the whole item wordings and collect again.

Okay, so I will try that, but I am not sure what to do here. I already have a 7 point likert scale. Could you advise?

Thanks a lot!

Hi there,

after taking a look into the question wordings, I have problems to understand, how these could be conceptualized as indicators of the same underlying factor. The low correlations among the indicators show that, too. I guess, this is the reason for the error: The model is misspecified.

There are two options. Select a few indicators that may indeed measure the same thing (for instance, the incompetence factor, measured by #1, #3 and #7 could work

But still, some correlations among items that could share some common cause, a remarkably low. For instance, in your table 2, item #4 ("when I am at work...") shows a correlation of r = .03 with #2 ("I feel I can make lot of inputs.."). Both seem to reflect some kind job autonomy. Why the correlation is so low, is really strange. Did you pretest the item (that is, perosns could have severe problems to understand the items..)

The second option would be: Go the "index-road" and create a composite. As any index however, you loose some part of relevance (as you are investigating a thing that you have just created). And if my presumption about the large measurement error is true, the index won't solve that entirely.

All the best,

Holger

Hi Jonas

glad we cleared that up. Well, it is relatively straightforward. Let us assume you want 1 to be the lowest option and 7 to be the highest then all responses already starting with 1 being the lowest option (e. g. "do not agreee at all") need not be changed.

If you assigned "do not agree at all" to 7, then you need to recode the items, mirroring them, so to speak. After recoding, 7 = 1, 6 = 2, 5 = 3, 4 = 4, 3 = 4, 2 = 6, 1= 7, missing value = missing value.

Having done that, I would recommend running the CFA again.

Best

Marcel

EDIT: Please create a new variable! Do not overwrite the old values. We mark the recoded variables with "[item designation]_re" for recoded (e. g. f18_re).

Hi Marcel Grieger

unfortunately mirroring the items did not change anything in the amos output. The model is stil unidentified and I should add 5 additional constraints.

Best Jonas

Hi Holger Steinmetz

the model is an adaptation from the Basic Psychological Need Satisfaction at Work Scale from Ilardi, Leone, Kasser, & Ryan, 1993. So, they are not my creation. I did modify the scale by removing the relatedness dimension. In hindsight that might have been a mistake and I should have kept it to just not consider it later on, as it has no relevance to my research. I should maybe add that I have a Pre-test/Post-test, Control/Treatment group design. For the CFA I only considered to use the Post-test, does that make sense?

A composite was also considered, but this does not really help to confirm the factors, I guess. I did run the analysis, that did work and I got results from AMOS. Altough fit indices were not that good: around 0.5 for NFI, IFI, TLI, CFI; Chi square/df at 2.053. RMSEA at 0.099. Std. residuals do not exceed 2.58, but are generally very high. The analyse also showed that establishing covariance among error terms will result in Change of around 1 for several parameters (not sure if that is much).

I did pilot the items and I got no complaints about the understandibility. I know my data is a bit weird, there should not be even negative correlations. I am not sure why this happend, also it makes no sense that correlation between similar items is that low.

I am considering to remove items from the scale and use fewer items as you suggested. Prior analysis considering reliability, showed that alpha values considerably improved, when removing item 7 and 13 across pre- and post-test scales. The composite factor analysis now for the post-test confirmed, that Factor loadings(-0.24 &-0.15) for these items were not significant (all others were significant). Would I be right to conclude that these items should be removed?

What does it mean for my model that I have negative factor loadings and correlations (they should be positive in theory)?

Thanks a lot for all your help!

As this is my first time doing CFA, I am not that sure about the correct procedure or options I could use and any suggestion really helps.

Best,

Jonas

Hello Jonas,

did the authors of the scale really say/write that it should be a *factor*? and if yes, did they sucessfully (optimally with a non-sign. chisquare) test the factor structure? For decades psychologists did not understand that the term "factor" means something (actually the whole big 5 research rests on the wrong assumption that PCA is the same as factor analysis.

Hence, my hunch is that the did the usual thing: PCA with variamax, alpha is ok, done. Or in case they applied CFA they showed the same level of explain-away-mentality as it is applied in the majority of scales.

Unfortunately you have now the meass to deal with. As I said, if the cause of the low correlation is NOT measurement error, you may nonetheless use it as an index for your experiment. As you have a pre-post design, estimating the retest reliability could answer this (at least of the duration between pre and post is not too narrow).

All the best,

Holger

Hi Jonas, you already got some very good advice. I just wonder about the following: How did you deal with the missing values? How did you code them, and is it verified that AMOS actually interprets these values as missing? This could also be a cause for the problem. And did AMOS use the correct number of subjects, did you check this?

Otherwise, I would recode the negatively formulated items and perform an EFA to get an idea about the structure of the data.

HTH, Karin

Karin Schermelleh-Engel Holger Steinmetz Marcel Grieger

Thanks for all your suggestions. It helped a lot!

Karin Schermelleh-Engel the missing values turned out to be acutally a part of the problem, as they were not recognized by AMOS. Results are much cleaner now. Thanks for the clue!

On another note for anyone with the same problem. It did help to set the constraint on another variable (that was not reverse coded) to achieve AMOS to recognize the identification of the model. This saves some trouble reflecting the variables.