This was rather interesting because, although there is sampling bias, it doesn't really matter. The sample space is android devices that visit the android market. Here, the sampling bias means only devices that would use the android market are polled--which overlaps very well with the developer's target audience of people who might download their app.
75% #probability in next 4 years of magnitude 7+ quake
---------------------------------
Says based on data gathered, so presumably run through some sort of simulation and determined that way. Even though a priori, I think closest to an empirical treatment of probability.