Interesting. I would probably introduce a new boolean variable GaveSomething / GaveZero and look at the subsets separately. All the zeros are screwing up the distributions and so, model fits.
For the GaveZero subset you can't do much more than count, but the GaveSomething subset should me more well-behaved and easier to analyse.
It's mostly the non-EAers who give $0, it seems. Removing those datapoints would diminish the difference between groups and make EAers look worse than they are - we care about how much total was given, not a random dichotomization like $0 vs non-$0.
http://www.gwern.net/EA%20donations