More often than not (depending upon context and teams) I prefer to fully qualify references, particularly in cases where there are any possibilites of collisions now or in the future (e.g. a lot of vaguely named data and business entities fall in this category, with nondescript names that could easily apply elsewhere). This just seems like good defensive programming, and yields code that isn't fragile. It won't shatter if someone adds a class to your namespace (or any other namespace that you're blindly using) that collides with a class in another namespace. For precisely the same reason I often preface instance references with this.

In this preference I am more of an exception than the rule, however - most prefer to add a using atop their namespace, and simply rely upon the compiler to sort things out. The premise being that the fewer characters exist in the code, the cleaner and more maintainable it is: If you can add a using System.Data.SqlClient to your file, then that namespace should never appear again. There is definitely a lot of validity to that position as well.

(Of course with something like the "CodeML" I described earlier, the XML would store the fully-qualified names, and developers can choose whether to display the qualified or unqualified names based upon their own preferences. It's just a display issue)