It seems clear to me that on average, guilded players have a slightly longer average playing time than non-guilded. The difference becomes more profound at higher levels. I also suspect that the data for non-guilded players is less accurate at the higher levels because, as your other data found, there are far fewer non-guilded players at the higher level. So it only take a few outliers to make it appear playing time for non-guilders exceeds that of guilders at the high levels.

Bruce

"In other words, controlling for level, being in a guild does not increase playing time."

Just a caveat about causality - these are correlations and it's not clear what's causing what.

If the data show that, controlling for Z, X the independent variable is positively associated with Y the dependent variable, you could conclude either that X causes Y or that Y causes X, or that there's a third variable, not in Z, that increases both X and Y. The likelihood of the latter goes up as the number of control variables goes down.

Also, when you start having N values in the tens of thousands, the difference between statistical and substantive significance becomes important. With a million observations, you could probably find that, I don't know, left-handed people are statistically significantly taller than right-handed people. That's because any measured difference, tiny though it would be - 0.0001 mm - would still render as 'statistically significant' because the standard error would be even smaller - just because of the huge N. But substantively, nobody would call 0.0001 millimeters significant in any way.

Causality: important
Statistical significance, t-tests, etc.: not evidence of importance

Excellent post -- I expected a clear-cut increase in playing time for guilded players, but that's apparently not the case.

Just to follow up on Bruce and Edward's point, it would be really helpful to see more stat results -- for the ANOVA test, it would be great to see the R-square and coefficient of variation. Similarly, it would be great to see sample sizes for each stratum of player level, just to see whether assumptions at higher player levels can be valid.

Also, it might be neat to run post-hoc tests after running the ANOVA -- I'm not sure if you're using SAS, but some post-hoc tests that come to mind include Duncan's multiple-range test, Student-Newman-Keuls', Tukey's, or Scheffe's.

Anyway, I'm *very* glad to see a hard statistical look at MMORPG data -- I'm really looking forward to seeing what you have to show in the future! *hint* *hint* cluster analysis *hint*

PS: By the way, what program did you use to create the plot? It looks awfully nice!

Nick is using the latest version of SPSS (13, I believe).

Ted - Thanks for pointing out the caution on the causality. On the statistical significance part, I think that's why the *non-significance* with this particular analysis is so striking. Even with the large N, there is no significant difference.

And re: David - SPSS 13 does amazing things :)

I'm away from my work computer for DiGRA, but will add the additional stats when I get back.

Nick

Another problem here, is that you would really need to track by accounts versus characters. In any given play time, I may be on with 3 differnt characters at various levels, depending on the needs of myself or guild/group activities. And across a few days, with mule characters, etc. it gets even more compounded. Good start though.

Michael,

This is an excellent point. Ahh, if we could get our hands on server-side data...

While I behave just a you describe, switching back and forth between characters, my impression is that the majority of my guildmates tend to play one character per session.

And while it would be wonderful as a researcher to know which characters were played by the same player, there have been a few cases where I've had an enjoyable RP session with someone I knew, but I was incognito. (And the converse problem exists: characters played by more than one player.)