Who are these likely voters, anyway?
After Labor Day, most pollsters start to apply “likely voter screens,” in which they attempt to identify respondents who are not just registe...
A draft of a paper on the Meta-analysis
Dear readers, I’ve been invited to write an academic article on the Meta-analysis. I’m horribly late with it…but I do have a draft. I’d be interested in your thoughts and reactions. I’m sure I have not done justice to some important topics. The article text is here (PDF) and the figures are here (6.7 MB PDF). If you link to this, use this post. The paper is a working draft and subject to change!
Also…in the course of writing up this project, I found this benchmarking from Luke Muehlhauser and Gwern Branwen, over at the Center for Applied Rationality (the best organizational name I have seen in some time)I. They compared us to FiveThirtyEight, InTrade, Drew Linzer, and others. We came out well!
Perhaps it’s standard language but I find the clause:
“accuracy equivalent to less than ±0.N%”
to be counterintuitive. Apparently it implies great accuracy but it reads as if the accuracy were pathetic.
–bks
Drew Linzer: world-beatingly correct (way in advance, and seemingly nuts at the time!) or just lucky? It’ll be interesting to see how he does in future cycles.
I have one small and one larger remark:
1. It was unclear to me what the (first) N on page 7 was (I presume the number of polls included).
2. I liked the idea of using national polls as a more time-sensitive correction to state polls (page 19). It would seem that just changing the bias b would not suffice though, as the state data for different states are lagging by different amounts of time (depending on the number of polls). Thus the national effect has to differ from state to state, and things quickly become complicated. And this would reduce one of the features I like most about your model: its simplicity. So a nice idea, but be careful with its implementation.