I voted occassionally, maybe once per year. I believe in doing it only in two situations:
1) A good player who cares a lot about their characters accidentally makes a huge mistake, like if they forgot to cast "Negative Plane Protection" even though they planned to in their discussions with the other players about how they would take on the head vampire, or if they walk into a trap that would kill their character outright and they had no warning in the same adventure that such a trap could exist.
2) If the dice seem to be the wrong approach to resolve something. In these cases, I may roll a die to see if it makes a good suggestion, but I will overrule it if I think of something better.
I believe this is great advice for people. I also believe in rolling many dice behind the screen only. There are many situations which make this more than called for in my opinion, but it's also a lot of fun to keep the players in suspense. Rolling behind the screen is actually a reliable way of doing that, which some DM's who prefer to roll all dice "out in the open" might have overlooked.
The determining factor is whether the players should know the probabilities involved, not the result; although rolling behind the screen also permits the DM to overrule, or fudge the dice when they want to.
Take the following examples:
Monster attacks. If the players see the monster needs a very low number to hit them, they will know to treat them as more of a threat. This discovery is better if it takes more time, seeing it in action a lot. Or if the monster needs a high number, they will not only know to treat them as less of a threat, they will be able to draw upon their knowledge of spells and other conditions that give penalties to the monster attacks. This is a game changer, because the closer to a natural 20 the monster needs to hit, each higher number makes an awesome change to the chance of a hit and monsters who need to roll a 17 or 18, for instance, can be neutralized almost by a simple spell whereas otherwise they would be hitting just often enough to pose a challenge.
Discovery checks. If the players search for traps, and they roll themselves, they know if they failed because they rolled poorly. The game will be more suspenseful if they are unsure if their search really ruled out a trap. And with something like the game's new spot checks, calling for a check can make them search for something they missed, which their characters aren't supposed to know anything about.
And if you let the players roll their searches for secret doors, they will know if they need to keep searching to rule out the possibility of one being there.
Damage rolls. Showing how much damage the monsters can do will let the players rapidly know this. Some of the same problems arise that do from monster attacks.
Monster saves. If you roll these out in the open, some of the same problems arise again.