Even if the sampling was rigorously random and no projects were dropped after starting, there would still likely be a bias towards small, "long tail", projects. It seems plausible that these types of projects would have more errors (or vice versa).