See
http://en.wikipedia.org/wiki/Birthday_attackA basic estimate is that the probability of collision is (n^2 / 2) / 2^160, where n is the number of hashes that have been generated. For example, if you've generated 2^64 hashes, you have a 1/2^33 chance of collision. This works pretty well until you get too close to 2^80, and larger.
Edit: That formula actually is a shorthand for (n * (n-1)/2) / 2^160, you get it because on your first hash, you have 0/2^160 chance of collision, and on your second hash, you have 1/2^160 chance of collision, and on your third, you have 2/2^160 chance of collision, and on your nth, you have (n-1)/2^160 chance of collision (assuming you haven't had any collisions already). So you can add up those probabilities to get (n * (n-1)/2)/2^160 -- you can just add the probabilities because the chance of having _two_ collisions is super unlikely (until it is no longer super unlikely, around 2^80 or so). (Really it's 1 - (1 - 1 / 2^160)^k and we're approximating that with k * (1 / 2^160.)
