I myself don't have a project, but the document I provided says that decimation is done by truncation. Given that they (Actel) manufacture ADCs, I'm going to have to side with them. No offense intended. 
Maybe it's the number of samples, because to do true decimation requires more like 64 over samples. At that point, I believe the truncation is truly just noise and not meaningful information as the "division" is just a simple shift removing noise bits that never were a part of the answer. The assumption being that there is 1LSB of noise in the data that is randomly distributed white noise.