Although highly dependent on the game itself, I did some investigation back in 2008 into testing various Elo modifications on a pretty large dataset of individual players when playing in a 5v5 team game, and geometric mean of individual ratings was the best simple predictor for which team would win (the update method was essentially 10 individual matchups of each player vs the aggregate rating of the enemy team).