Tuesday, January 15, 2008

Diebold effect sticks around, need a proper statistician

I have incorporated Brian's work in my latest analysis. I tried different linear and generalized linear models combining Clinton's score, the total number of votes, the usual demographic data, employment rates, education and latitude and longitude.

I took the latest available data sets, got the list of towns using electronic voting from the official site, completed and corrected town names by hand.

The significant factors I found are, in order of decreasing signifiance:
  • Percent of people holding bachelor's degrees,
  • Voting method.
It is interesting to note that when with the new, improved data, the percent of people holding a bachelor's degree becomes extremely significant (about p = 3e-9 vs. about p = 0.001 for voting method.)

I'd like someone who know his statistics well to check the data and tell us if the voting method is indeed significant. The fitted models are linear and for all I know, it could be acting as a non-linear proxy for population size or some other funny explanation...

Anyway the new data is available here, feel free to check, improve and re-publish it. Note that you need the maptools R package. You can install it by typing install.packages(c("maptools", "maps"),dependencies=T)) in R.


Chris Chatham said...

Thanks for the analysis - I am finding the same results. please contribute at black box voting where we're actively running these analyses as new demographic info comes in (hopefully we'll have population density soon).


Chris Chatham said...

Sorry, that URL was supposed to go here

John said...

I suggested this to Chris as well, but given this study would it be interesting to run the regression with Clinton and Obama Diebold votes swapped and see how it affects the significant factors?