Tom had our preview of the PA-12 special election yesterday, but just a quick statistical accompaniment to that.
Here is a chart showing the Democatic share of the two-way vote (that is, excluding third parties) in all open seat (non-incumbent) House races in the 2006 and 2008 cycles. As you can see, the trend is generally fairly linear with respect to PVI:
In a district like PA-12, which has Partisan Voting Index of R+1, we'd have expected the Democrat to win by a margin of about 54-46 if conditions were the same as in 2006 and 2008. These, of course, were very good cycles for Democrats.
Here is the same information for 2004, which was a good cycle for Republicans.
Under 2004 conditions, we'd have expected the seat to go to Republicans by a margin of 52.5-47.5.
Now, this doesn't really tell us very much. As you can see from the graphs, there's quite a lot of variance around the regression lines, even if the overall trend is fairy linear. That is to say, the contingencies of individual candidates and districts matter a lot -- and this is especially so in special elections.
But if the Republican Tim Burns were to win by about 5 points, that would give us one indication that the cycle was shaping up to be more like 2004, from which the Republicans emerged with 232 House seats. And if Democrat Mark Critz were to win by, say, 8 points, that would give us one indication that the cycle was shaping up to be more like 2006 or 2008.
Of course, most people expect this November's election to wind up somewhere in between 2004 and 2006/08. And most people expect tonight's race in PA-12 to feature a very close outcome. The only point I'd emphasize is that, from a forecasting standpoint, a Crtiz win by 50.1-49.9 is essentially no different than a Burns win by the same margin. Even under 2006/08 conditions, a Democrat would have about a 30 percent chance of losing this seat, and even under 2004 conditions, he'd still have about a 40 percent chance of winning it. It's really only if one of the candidates wins by middle-to-high single digits (or more) that it might tell us something, and then not necessarily very much.