Dougal666 wrote:philip1201 wrote:In other words, learn to use statistics before you start complaining about others using statistics wrong. Linear approximation is a valid method of estimation as long as you know how to handle it, which Randall does.

Oh my, the negativity.

A first order approximation of one third (one year out of three) of a clearly non-linear graph, and extrapolated to half that distance (six months)?

I'm really not clear what statistical approach is appropriate here. Ideally you could get the actual numbers rather than the normalized data. Google claims that if you download the data then they will give you relative standard errors. You could use that, if you trust them.

Looking at the fluctuation (it looks to me like they plot every two weeks but it could be weekly) it looks so regular that probably the numbers are pretty big and you don't need much statistical power to show that yes, something is happening.

So something changed in the last couple of years, and people are increasingly asking Google about tumblr, while they are asking Google about blogs a bit less than they used to. If that increase doesn't slow down, the tumblr searches will overtake the blog searches pretty soon.

Would it be better to model the tumblr data as exponential growth, or as two linear pieces? Do you care that much? A logistic model is probably better on theoretical grounds -- things which look like exponential growth often turn out logistic -- but how well could you tease that extra parameter from the noise?

I think the basic conclusion here is that something has changed. And if nothing else changes then it will keep going on the way it is now.