[byrneseyeview]

Linear fits? Why? Why?

Karma is a function of time since joining, participation, and quality of contributions. And it starts at 1. 'Participation' can be determined by looking at contributions per length of time. Quality is average score of each submission -- separating it from participation is a useful way to extend this to a more complicated model taking into account the fact that people stop using HN. So the line of best fit should be something closer to 1 + t * q * p, or the sum of 1 + (t0 * q * p0) + .... (tn * q * pn) to describe folks who are off-and-on contributors.

[timr]

If nothing else, it's legitimate to filter non-participants before doing a regression analysis on the rest. It's technically true that you can't predict karma over time, but that's not really an interesting statement until you eliminate the large number of people who sign up, then never post or comment.

My instinct is that once you filter these people, you'll see a much stronger linear relationship between time and karma, since karma isn't normalized by the number of contributions, and number of contributions is probably a poisson process.

[breck]

Removing all 1's and 2's improves the relationship somewhat (moreso with log(k)) but still not a whole lot.

[timr]

Sounds like there's a vast gulf of people with little (but not zero) contribution, then. Can you plot # of contributions versus membership time?

[breck]

don't have the contributions data, just the karma score.

[DougBTX]

> And it starts at 1.

That's just a constant offset, it matters more whether the next karma level is 2 or 10. When fitting trend lines a "linear fit" would normally satisfy y=mx+c, without limiting yourself to c=0. Note that the posted linear fit has c = -1... apparently everyone starts with -1 karma. Lies, damned lies and statistics I say!

[breck]

Agreed. If I had data for the quality of contributions I could have done that.