Based on my experience of this issue, I suggest you check the "Modification indices" results. Most of the time higher correlation among the variables leads to this issue.

By correlating the error terms containing high values of "Modification indices" can fix this issue.

Meanwhile, here is a useful video containing the proper solution to this issue.

https://www.youtube.com/watch?v=JkZGWUUjdLg&t=10s

Hope it will work for you.

Thanks.

I think (this is my current, before coffee, opinion, but I I can be convinced otherwise) is that it is okay because the EFA is not influenced by the CFA, and the EFA would be described as Exploratory. The CFA should also be described, since that failure in the original hypothesis is still likely an important component of the paper.

Thinking about it pragmatically, if you have no length restrictions you would usually print the correlation matrix as part of the descriptives, and the EFA can be viewed as just two steps beyond this (just reporting the eigenvalues and vectors [PCA] might be considered one step as this involves no choices that might be influenced by the CFA).

I'm interested what other people think. I know this suggests if you are first planning an EFA, you get a free model test if you throw in a CFA, but if you can justify that this is probably good for science. Second, if you run a CFA and it doesn't fit, it does not mean that the data cannot inform science.

Fakhar Shahzad, thank you for trying to help out, but, sorry, I think you've not really addressed the issue I raised. Furthermore, the YouTube presentation you've suggested is really about conducting CFA (unless I'm mistaken), and I wasn't concerned about that.

But, again, thank you for trying to help.

It is a good question, I would also apply what Daniel refers to, although I have not seen any reference or article to support it, hopefully someone can report some evidence.

Respected Robert Trevethan I am sure anybody would love to apply the solution proposed by Mr. Daniel Wright and endorsed by Mr. Cristian Ramos-Vera but i do not think we can find support for this in literature. CFA means trying to validate some existing scale. So, poor fit indicates data set is not good either because of the poor items' quality or because of the poor understanding (lack of interest etc.) of the target population. Recently a researcher from a very reputed research group informed me that they could not proceed further on a project because of the poor quality of data they received. So, in this case I think new data set should be collected. Thank you.

Robert Trevethan , could you tell us more about the data?

Poor fit might occur with small sample sizes (

Muhammad Zia Aslam, thank you for your post. In our case, we did not have a prior scale, but, rather, a prior theory put forward as a conceptual proposition in an extended journal article.

We created a set of items that was carefully created to conform with the theory, with six items on each of four proposed factors/constructs within the theory.

The resulting scale was administered to two sets of participants (each approximately 200 people) and the two sets of data were submitted to CFA. For neither data set did most of the goodness-of-fit indices (GoFs) meet satisfactory criteria. (We used the normed chi-square, CFI, TLI, RMSEA, and SRMR).

To be honest, I think that maybe 3 of the 24 items were not "good", but possibly not so defective as to destroy the CFA. In order to check why the CFAs were so poor, we went back and subjected each data set to EFA and discovered that, for both samples, there was only one factor in the data.

I should add that we checked the data carefully and we removed a very small number of records in which there seemed to be invalid responses or missing data.

To pick up on the possibilities you suggest, I think that neither the items nor our participants' "deficiencies" (lack of interest, etc.) accounted for the problem. Instead, we concluded that the original theory did not apply - at least to the people we sampled.

In my view, what we did was permissible (and provided valuable insights), but I thought I'd like to check out what others (on RG) thought. Essentially, did we do something that was methodologically misguided or wrong - or did we engage in a desirable strategy?

Peter Moorer, thanks for taking an interest in my question. Some of what I wrote in the post immediately above here addresses your request for information about our data.

I should point out that we didn't "conduct EFA to avoid using fit criteria". Instead, we conducted EFAs to get some leverage on why the CFAs had so obviously disconfirmed theoretical expectations.

I hope that makes sense. In essence, I think we conducted EFAs to do part of what you're suggesting: We tried to identify the reason(s) for misfit. In our case, I think misfit was because our participants' responses did not conform to the theoretical expectations. Whether other participants' responses would conform to those theoretical expectations or not is another matter.

Actually, from the start I was suspicious about the theory being put forward. We tried to put that theory to the test in as fair a way as we could - and it didn't come up to scratch!

If several previous studies that carry out CFA determine that the structure of such an instrument is two-dimensional with re-specifications of error, this may be an indicator of being able to perform EFA?

Robert Trevethan , did you try a parallel analysis? It is implemented in R and JASP (an alternative to SPSS, but not as advanced, although in some points it is).

Sometimes items are correlated stronger then expected. This shows up in minor factors between specific items. Sometimes this is due to shared viewpoints between items (some items referring to I, You, We, Them etc). I have found that this happens regularly.

Cristian Ramos-Vera, you could well be correct, but in our case there had been no prior studies in which CFA had been conducted. We based our CFAs on prior theory.

Peter Moorer, yes, certainly, we did perform parallel analysis. Actually, I omitted one part of our analyses for the sake of simplicity. The parallel analysis for both samples indicated the presence of only two factors (the scree plots were ambiguous, and I don't trust the Kaiser criterion in EFA).

When we sought a two-factor solution for the first sample, 20 of the 24 items loaded on the first factor, one item loaded uniquely on the second factor, one item cross loaded on the two factors with a difference of only .03, and two items loaded on neither factor.

For the second sample, 10 items loaded on the first factor, 11 items loaded on the second factor, one item cross loaded with a difference of only .02, and one item loaded on neither factor.

When we removed the two unsatisfactory items in each sample, the remaining items loaded on only a single factor.

I hope I've satisfied your curiosity - at this stage, at least. And thank you for that curiosity.

Dear Robert Trevethan , Well while developing measurement instruments based on theoretically valid concepts, finding a different than proposed factor structure is a novelty according to my limited knowledge, and certainly is a contribution towards the body of knowledge, which should be tested in future for further validation.

Actually in your case EFA should have been a starting point before going for CFA. So, if the necessary scale development process is justified (as sounds in your case), I do not think there should be any issues to proceed and recommend new findings for improvement based on the inferences drawn. Thank you.

Muhammad Zia Aslam, we do, indeed, hope to have contributed to the relevant body of knowledge, and we have suggested ways in which future research might be conducted.

I am not sure why you say that we should have started with CFA rather than EFA, however. We created a scale carefully, based on a particular theory, and we used that as a basis for assessing the validity of the theory. Under the circumstances, I think that CFA was the most appropriate kind of factor analysis to employ in the first instance. When that (CFA) didn't support the theory, our hunch was that EFA would then be appropriate as a means of seeing how the proposed dimensions / factors within the theory DID relate to each other. I wasn't sure whether we'd done the right thing, thus my question at the start of this thread.

I have just found out that what we did has been written about, in the following article, as being justified:

Schmitt, T. A. (2011). Current methodological considerations in exploratory and confirmatory factor analysis. Journal of Psychoeducational Assessment, 29(4), 304-321.

Thank you, everyone, who has contributed helpful ideas to this thread. For me, the issue has been convincingly resolved near the end of the article that I provided the citation for in my post immediately above here.

There (final paragraph on p. 315), Schmitt wrote: "On a final note, researchers should be aware that it is reasonable to follow up a poor-fitting CFA model with an EFA."

Thanks, again, for your contributions.