First, I did what I called "full stats", that, although more accurate, is revelant in less cases. Full stats means that I used the data from every user that entered in the system, including those who never spent one full minute on it, or those that came only once and did nothing before leaving. Everything's here, and the result is that users are here for an average of 3 hours and 15 minutes per week. BTW, this is the only kind of stats you can also take from Linden Labs data, getting as result an average of 4 hours and 11 minutes per month.
Then, I inserted some intelligence on the stats, and decided that those records that the system doesn't even consider as "users" wouldn't count. The result is that real users are here for an average of 6 hours and 24 minutes per week.
I wasn't happy, and decided that you should ignore those users that came a few times and never came back: after all they're probably not users at all, and are messing with the stats. I discarded all those users that stayed on the system less than a total than 24 hours. The result is that returning users are here for an average of 9 hours and 9 minutes per week.
This data still seems way lower than it feels when I'm there, so I think I might do this exercice another time, counting only with adult animals (users that spent at least 8 hours in world), and with all of them.
Now, this is real data. Polls? Mailing Lists? Nah... The problem is that you can't really trust Linden Labs figures, can you? You know they'll allways want to make them look nice...