In my structural equation model, I have a an observed egzogenous variable, according to modification indices, can i link it to an indicator of endogenous latent?
of course, you can do that but it does have to make sense. Google a bit about MIMIC models and differential item functioning where such scenerios are discussed. Such a situation casts some doubts on the validity of you measurement model to which the indicator belongs. Normally, an assumption would be that all indictors of a latent variable should causally function in the same way (as they are exchangable indicators). Having a differential relationship to other variables is often a sign that something is problematic.
Ceren, I guess that the modification index suggests to let the error covariance between both indicators be freely estimated? In addition to Holgers answer, one plausible reason for this covariance may be that equal wording exists or that both items are reverse coded. This would be a method effect, and it would make sense to include the covariance. There is a nice article that may be helpful for you: Podsakoff et al. (2003). Common Method Biases in Behavioral Research: A Critical Review of the Literature and Recommended Remedies. Journal of Applied Psychology 88(5):879-903.
Article Common Method Biases in Behavioral Research: A Critical Revi...
No, this is not a good idea. As Holger already pointed out, you need a justification for modifying your model. The modifications suggested by AMOS are based purely on statistics, and some of the modifications are even interchangeable. If you include one of the modifications suggested by AMOS into your model, all other modification indices may change. Some may vanish, some new modification indices may occur. It is not a good idea to just rely on suggestions of a statistical program. Don't do that!
The problem is even worse when you allow one indicator to load onto two different constructs. This raises questions about the validity of the item with a cross-loading. How would you interpret a relationship between two latent constructs, when they both share the same indicator variable?