That's exactly what I was missing, thanks! Added: https://bitminter.com/stats/rewards
I removed the first shifts (proofs of work are grouped in shifts, each of size difficulty/10). I started recording these before I switched to PPLNS, so the first ones naturally had no rewards.
This is just 46x difficulty worth of proofs of work. We are a small pool, and it wasn't always PPLNS, so we don't have that much data to go on yet.
You can see some bad luck in the beginning put us below average. Then a nice streak of good luck put us at about twice the expected earnings. From there up until now it's been pretty average. Right now we are having a really bad block, which is the reason for no earnings towards the end. Luckily we had some short blocks just before that.
Obviously the big lucky streak at the beginning takes a while to even out. Ending at 29 rather than 23 BTC is a big difference.
Since switching to PPLNS we made 58 blocks rather than the expected 46. Too small sample size maybe? Any other pool ops have some data?