Hi Amita,
I guess you won't want to hear this but actually, you should not use MI at all. A MI value tells you how much the chi-square value will drop if you freely estimate a respective parmater. It's a purely technical perspective in which the software estimates this drop for every fixed parameter. The software is dumb and thus cannot elaborate whehter a change makes sense. This is your job as a researcher. You could now inspect the list of MIs and evaluate the theoretical appropriateness but this has two problems
1) This is a post-hoc approach and you come up with a hypothesis after seeing the data (i.e., the MIs). If your solution is so viable, the question is why you did not incorporate it in the first place...
2) More importantly, the use of MI assumes that the set of already-estimated parameter is correct and that the only problem is one or several wrongly-fixed parameters. This is often not correct. See for instance, factor models (CFA) in which the fundamental set up (the time-to-factor assignment) is the problem and not only some left-out parameters.
I can recommend two articles by David Kaplan who also discusses the value of expected parameter change in addition or instead of MI
Kaplan, D. (1989). Model modification in covariance structure analysis: Application of the expected parameter change statistics. Multivariate Behavioral Research, 24(3), 285-305.
Kaplan, D. (1990). Evaluating and modifying covariance structure models: A review and recommendation. Multivariate Behavioral Research, 25(2), 137-155.
Beyond that, I would spend more effort in printing and interpreting the "standardized residuals" which are the standardized cell-wise differences between the emprical covariances and the model-implied covariances, and then try to locate the problem of the model (which is difficult but makes more heavy use of your theory-skills than simply adhering to proposals of the software).
I have demonstrated that in this paper (for which you can also get the R-code).
Rosman, T., Kerwer, M., Steinmetz, H., Chasiotis, A., Wedderhoff, O., Betsch, C., & Bosnjak, M. (2021). Will COVID‐19‐related economic worries superimpose health worries, reducing nonpharmaceutical intervention acceptance in Germany? A prospective pre‐registered study. International Journal of Psychology. doi:doi.org/10.1002/ijop.12753
Hope that helps
Holger
Rather than using any kind of arbitrary threshold value for the modification indices, I would suggest simply looking at the largest ones and trying to figure out why they are large for the given parameters. This can sometimes help you identify areas of your model that underfit the data.