My MOE has always been less than 0.5% using a somewhat modified formula of the 18.4% rule. It is a simple modification to account for regional biases-- candidates from this part of the state always run ahead of those from elsewhere regardless of party label. It correctly forecast that Toomey's U.S. senate election in 2010 would be a squeaker and that he would win by a considerably smaller margin than Corbett for governor.
It showed other amazingly accurate tracking results dating back to 2004, when we first volunteered to be poll watchers and I first begin checking data in our precinct vs. statewide.
In 2012, everything still tracked quite closely for the statewide races except one: the top of the ticket and it was way off . . . so far off that the chances of massive unprecedented fraud were very nearly 100%.
Our precinct is very well run and the Judge of Elections whom I am replacing ran a very good show. So the problem aren't the stats we generated, they are what went on statewide.
1. Depressed rural turnout. Part of that might be population loss. Some areas in the midwest had counties flipping from Obama 08 to Romney, but had less actual votes for Romney than McCain.
2. Severe bluing in inner cities. Michigan's map was about a normal map for a 3% loss, not a 8% loss. Was it fraud? Probably some in the cities like normal. $20 says it's in the absentees.