Model fit and factor loadings are not directly related (although with higher factor loadings, you do have more power to reject a model). That is, even a well-fitting model can have parameter estimates (including loadings) that are not significant and/or not meaningful in terms of their size.

This may indicate that your items aren't very strongly related. Perhaps the items for a given factor are rather heterogeneous?

To fix this, one way is to re-examine the items that are loading poorly on the latent construct and see if they are appropriate for measuring that construct. If not, you may need to rephrase or remove those items. You may also want to consider adding more items or re-conceptualizing the construct to ensure that it is well-measured.

Another way is to consider alternative estimation methods such as Maximum Likelihood (ML) or Bayesian estimation, as these methods may produce more accurate factor loadings.

Additionally, it's important to keep in mind that the goodness of fit statistics (CFI, TLI, RMSEA) are not necessarily indicative of the factor loadings, they only provide an overall measure of how well the model fits the data. Therefore, it's important to consider both the factor loadings and the goodness of fit indices together when evaluating the model.

Both Christian and Dinesh have rightly pointed out that the goodness of fit statistics is not necessarily an indication of the establishment of the structural model (reliability and validity of constructs). Lower item loadings are an indication of heterogeneity of the construct, that is, the items may not necessarily measure the given construct. If there are 3+ items in the construct, check if dropping one or more items improves the situation. If you're running a pilot test, then consider adding new items.

Also, check if the construct you are measuring has two dimensions - 3 items providing excellent loadings and the other 4 together make a good construct (just an example). If so, your construct may be a higher-order construct with more than one dimension.

two of my construct consist of Likert-type questions (to measure behavior) and the other two consist of 8 binary items each (to measure knowledge, each 1 represents a correct answer). The problem is when i test for validity and reliability the factor loadings of Likert-type items is relatively high (0.7+) while the binary is extremely low (0.1-0.3). Is this a good way to measure the “Knowledge” construct? If so, why is the loadings of binary items are so so low? Ali Farooq Dinesh Kumar Christian Geiser

Thank you for sharing details Do Huy Phuc. Likely, the knowledge construct is not a formative construct. In this situation, item loadings may not be high, but their weights significantly load on the given construct. I had this experience in one of my publications where I was measuring knowledge. Have a look at this paper:

Farooq, A., Jeske, D., & Isoaho, J. (2019). Predicting students’ security behavior using information-motivation-behavioral skills model. In ICT Systems Security and Privacy Protection: 34th IFIP TC 11 International Conference, SEC 2019, Lisbon, Portugal, June 25-27, 2019, Proceedings 34 (pp. 238-252). Springer International Publishing.

The paper has supplementary material as well. Let me know if you cannot access the paper online, and I will send you the text in DM/PM.

My understanding is that you are using reflective measurement models where the items "depend" on the latent variables (factors).

Are the knowledge items measuring knowledge regarding the same construct/dimension/issue? If not (i.e., if they measure knowledge about different areas), then they may not be highly correlated. A person may know about Issue A but not Issue B. In that case, it would not make sense to use the items as indicators of a common latent variable (common factor) in a reflective measurement model. I would check the tetrachoric correlations for those binary items. They are probably close to zero.

Correction in my above comment: read "... knowledge is a formative construct..."