How the probabilities are estimated
The details are expained in my AER paper. I forecast each state election
outcome 19482004, using the same information each year (opinion
polls, past votes, etc.). I then measure the prediction error in each
state: the difference between the forecasted election outcomes and the
actual. Starting from the predicted outcomes for each state in 2008, I
add errors randomly drawn from the distribution of past prediction
errors. This gives me a a distribution of hypothetical election
outcomes. From this, I can compute different statistics, such as the
probability that either candidate wins.
It is crucial to allow for national swings, otherwise the uncertainty in
the election is underestimated. I allow the states to swing together by
estimating the national average prediction error in each election, and
the indpendent state errors.
The varibles included. At the national level, the model uses national
opinion polls, lagged national vote share, incumbent president and
incumbent party, second quarter economic growth. For recent years, at the state level, I
use statelevel opinion polls, lagged votes and the ADAscores of the
state's congressmen and senators.
One could draw errors from the actual error distribution. I instead assume
that the distribution of forecast errors is normally distributed and
draw from a distribution with the same variance (and mean) as that of the
prediction errors. For state outcomes, the actual error distribution is
quite close to normal. For the national errors, we only have 15
observations so it is hard to judge whether the errors are normal.
