Modeling non-IID data

I'm working on a problem where the data we're getting is not very IID, and was wondering what can be done to account for this?

Long story short, I can't be overly specific, but I'd like some papers/discussion on what people have done in the past.