Each weekend I've been running a Monte Carlo simulation of the general election using data from Andrew Tanenbaum's www.electoral-vote.com. The Votemaster has stopped updating data for a possible  McCain/Clinton matching, so I'm no longer running simulations of that, either.

My previous simulations have used just the 4% margin of error common in state polls as the only source of variability, which implies that the only source of error in state polls is sampling error. At best this method gives a snapshot of the current state of the election, using the most recent polls. But we're still five months away from the election, so voter preferences will shift, and the model I've been using doesn't allow for any other changes.

As  a first cut, to adjust for this, I've estimated a wider margin of error for the polls, using data from www.electoral-vote.com from the 2004 election. Taking the polling available on a given date (the site has polling data and dates from September 1, 2004 until the election), I've computed the standard deviation of the polling errors from that date, and then I've done a linear regression of that data using the number of days until the election as the independent variable. From this I can extrapolate a margin of error to use for data today. This model gives me a margin of error of about 13%, over 3 times the 4% typically reported from the polls.

So I've run two version of the simulation, one using 4%, and the other using 13%, doing 10,000 trials of each. Results:
4% Margin of Error
Obama wins 88.0%, averages 289.4 EV
McCain wins 10.3%, averages 248.6 EV
Electoral tie 1.7%

13% Margin of Error
Obama wins72.2%, averages 284.0 EV
McCain wins 26.4%, averages 254.0 EV
Electoral tie 1.5%

This is Obama's best showing since I've been doing this, but the current results haven't changed much since last week. Obama now has a slim edge Missouri, where he previously had trailed by varying degrees, and while there have been other new polls, they just confirmed fairly large leads for the candidate leading in a given state, and thus had little impact on the simulations.

The 13% margin of error shows greater variability in the outcome, as it makes it more likely that states will swing to the other column. Intuitively, I find it easier to believe McCain has about a 25% chance of winning than that he has just a 10% chance. But that may still understate the current closeness of the race. In both these models, I'm assuming treating state's result as an independent random variable. While this is reasonable when you're only trying to reflect sampling bias, it's not such a good assumption when modeling opinion shifts. What causes a candidate to improve in one state is more likely to cause improvement in other states also. A different model which accounts for that would likely provide better projections.

In any event, current polling shows Obama ahead, and he may stretch that lead further if the end of the nomination race provides an extra bounce. The results using a 4% margin of error are, at best, a snapshot of the most recent polling data, which is often quite old. The 13% margin of error allows for changing opinions, but in a way that each state's shifts are completely independent of each other state's, which would understate the chances of the trailing candidate to make a comeback (or the leading candidate to win in a landslide).