For the answer based on an unspecified two million game database, have a look here.
Frankly, there's very little that's surprising there, and of course the results aren't normative: one isn't obliged to make sure that his d-pawn and knights leave the board first and his h-pawn is the last thing to go. The only thing that did surprise me a little was that the h-pawn's survival rate was as high as it is: I assumed it would be high, but thought that h4-h5 attacking ideas against fianchettos (as in the Yugoslav Attack vs. the Dragon) would lower its average lifespan by a larger degree.
One slightly interesting finding was came when I played around with the numbers. If my math was correct, on average White finishes .103 pawns ahead, assuming the traditional material scale according to which a pawn = 1, bishops and knights = 3, rooks = 5 and queens = 9. If you turn a typical engine on at the beginning of the game it will award White an advantage greater than .1 pawns, so either the engines get it wrong, players with White consistently underperform, or part of White's advantage is not based on the quantity of material but on what can be done with that material. (Of course, this must be part of the story, as at the beginning of the game the computer prefers White, even though the material is completely even.)
I'm sure clever readers can find more interesting applications of the data than I did, so have at it.
HT: James Turnbull & Phil Salathé