Say item number 1, 2, and 3 are being loaded under factor A and it seems theoretically meaningful. But another item i.e. item number 4 is also being loaded under factor A and this time it doesn't look a meaningful loading since theoretically speaking it is expected to be loaded under factor B instead of factor A. Then, what should be done in such a case?

Here are a couple of quick suggestions, Kailash Jandu:

First, I suggest that you check to ensure your data have satisfied all of the required preconditions for EFA, e.g., that you have, say, 10 x the number of participants relative to the number of items (though there are more "complex" criteria for determining the number of participants needed for EFA). I mention this because results on EFA can be quite volatile / misleading if there are too few participants.

I would also check to see whether there is something about the wording or nature of the "discrepant" item such that your participants would have given it a meaning that is different from the meaning that you as researcher(s) gave it.

In line with Robert, I just want to add that this problem often time arises when there is some chance of content validity problem, which is essentially checked qualitatively followed by pilot survey (some call it face validity). Content/face validity problem may sometimes lead the respondent to provide theoretically contradictory or inconsistent responses. I am not sure whether you've adequately checked this issue, especially if the items are developed for the first time.

I agree with the suggestion of Khan Taufique . Often most people overlook Content Validity in test development yet it is very key especially when dealing with self-reported items.

I would check simple correlations among Items 1-4. Especially, if item 4 is substantially related with other items (items 1-3), it is reasonable if you have all these items 1-4 in one factor because your sample data are constructed in that way unfortunately, different from your theoretical expectation. You may try a rotation, but I am not optimistic about it either. The last resort is to drop item 4, but dropping or deleting an item is the LAST attempt. Or since your sample data (if item 4 is highly related with other items (1-3) is made in that way, if you can, I would suggest to collect more data to see if item 4 moves to the other factor. Cheers!

However, the case is such that I have tried almost all the possible ways suggested here. The sample size is adequate (N = 450); the number of items is 15. The interesting thing (and particularly challenging for me) is that the three items of this particular dimension load onto three other different factors (one item on each factor), making the situation little complex. The wording of items looks appropriate and easy to understand. Thee three items are important for the scale so taking a decision of deleting these is difficult for me.

I have tried Principal Component Analysis with and without Varimax rotation method, Principal Axis Factoring with Promax, and Maximum Likelihood with rotation as well. Still, the results are not satisfactory.

Even I tried seeing the results after deletion of one item but still it didn't yield meaningful results.

There is a factor analysis program called Comprehensive Exploratory Factor Analysis (CEFA) and you may download it from the Ohio State Univ. website. Although CEFA is designed for exploratory factor analysis, you can allocate items to specific factors of your interest. I am enclosing the website for you: https://psychology.osu.edu/dr-browne-software. Good luck!

Kailash Jandu I am glad you have used something other than principal component analysis (PCA). Despite what many people think, PCA is not really (exploratory) factor analysis and it is probably often used because it's the default option in SPSS. I suggest you rely on principal axis factoring as your extraction method, though maximum likelihood is also a possibility if your data satisfy its requirements.

Also, I suggest you try an oblique rotation rather than varimax. Although varimax seems to be most commonly used, the methodologists I trust recommend oblique rather than orthogonal (e.g., varimax) rotations because most factors are correlated.

For what it's worth, a colleague and I recently used an oblique rotation on our data because of the advice from methodologists, but we tried a varimax rotation as a double check. Contrary to what might be expected, the outcome was much clearer with the oblique rotation. Don't ask me why!

To me it sounds more like you are using CFA (confirmatory factor analysis) rather than EFA (exploratory factor analysis). If you already have a theory about how the variables should load on certain factors then you are trying to confirm the existence of a certain structure in your data. I mostly use Principal Axis Factoring (PAF) when conducting CFA or EFA as this is recommended by Andy Field (Principal Component Analysis is not really factor analysis, strictly speaking).

I recommend parallel analysis and oblique or prominent rotation, using the generalized least squares method. You should then consider confirming your instrumental model by CFA considering the normality distribution of the data that is important for performing the factorial analysis and considering the reliability. If you present cross-loaded items as indicated, you should consider in the CFA if that indicator can theoretically also be part of another factor of the model, another way is through the error indices to perform the covariance of errors of those items that present cross-loading with other items of the proposed model that would indicate factor complexity, dependence or conceptual similarity that would be important to consider in the discussion of the study.