I'm actually very glad that you brought this up - gave my a nudge to test this myself!
I remember Ashley making mention of something like this sometime ago with regards to someone's question about performance with individual rain drops as sprites. He suggested then that it'd be more efficient to move the rain drops to the top of the screen when they hit the ground, instead of destroying and recreating them.
I made a quick test to see this sort of problem in action, and, provided my testing methodology is accurate (I think it is), it's definitely more efficient to move objects than to destroy and recreate them. With this test, on my machine, I started to see frame-rate wobbles (or "pulses", as it appears in this test) around the mark of 3,500 instances with the destroy/create method, whereas with the move method, I could take the instance count up to 15,000 before the frame-rate dropped below 60 fps (once it got up to speed, that is). And even then the frame rate, when lower than 60, would be very stable up to about 35,000 instances.
Try out this demo yourself. You might need to adjust the repeat or sprite count checks, depending on your machine's hardware, to find the point at which the difference between the two techniques is the most noticeable.