Charles Franklin | January 18, 2010
Republican Scott Brown holds a lead in all 18 alternative models of the Massachusetts Senate race polls, now including all polls released through 6:00 p.m. Monday. Our standard trend estimate puts the race at a 6.2 point Brown lead over Democrat Martha Coakley. The less sensitive alternative linear model puts the Brown lead at 7.3 points. Across all models, Brown leads by between 1.0 and 8.9 points. Three quarters of the estimates have Brown ahead by 4 points or more.
Brown built this lead over the past week of polling with only some tentative sign of the trend flattening over the weekend. Of course the last available polls were completed Sunday evening so we do not know if any movement has occurred on Monday.
Here is a brief review of the polls and the various models estimated. First, the polls without any trend estimates:
One of the unusual features of the MA polls is the large number of leaks from Coakley's internal polling. No one leaks without a reason, and her leaks have been consistently better for her than other polling taken at the same times (with one exception). Past analysis has found that internal polls are typically about 5 points better on the margin for the leaker than are independent polls, but that the internal polls do track the trend rather well. This raises a question of how to treat Coakley's polls. Below, I estimate the models both with and without the leaks included, so we can see their potential impact. I have not discounted them for the historical five point bias with internal polls.
First, let's estimate the local regression models that are our standard here at Pollster. These are not identical to the dynamic charts because here I am estimating the Dem minus Rep margin while our charts estimate each candidate separately. I estimate the standard model, a more sensitive and a less sensitive version, and then repeat with the leaked polls included.
With the limited number of polls, the local its are not as smooth as in our usual trends with dozens or hundreds of polls. They may also be more sensitive to outliers. That is one reason to check the effect of sensitivity. For the three models without leaked polls, the sensitivity matters a bit for the trajectory but hardly at all for the endpoints. When internal polls are included, the trends end up a couple of points more Democratic, though still put Brown ahead in the end. (The more sensitive estimate touches dead even but that is due to a day with only an internal poll. The sensitive estimator chases that but then moves back down.)
Next I switch to even less sensitive linear models. The local trends show some non-linear movement, and suggest a small upturn in the last day of polling, which the linear models will miss. But with relatively few polls, much of the "bendiness" of the local trends is due to noise and overfitting the data rather than meaningful shifts. The linear fits are a hedge against the noise, at the expense of an ability to spot a reversal of trend.
With these models we can disaggregate the data by partisan affiliation of pollsters. This gives a range of estimates from most favorable to Republicans to neutral (no picking of polls) to most favorable to the Democrats. Here I include the leaked polls in the most Democratic model, using only Dem polls plus the leaked internals, and in one model with everything we have regardless of source. The result is a range of estimates. The most Democratic model shows a 1 point Brown lead. Others range from about 5 to about 9 point Republican margins.
A final variation in the models is to fit quadratic models to allow the trends to bend according to how much the data demand a bend. If there were substantial upturns (or downturns) at the end, the quadratic model could pick that up while still maintaining less sensitivity than the local regressions we started with.
As it happens, the data don't demand much bend at all. The fit is only slightly better with the bend and the end points of the lines are only modestly changed.
Finally, let's see the garbage can of all the models at once, to see if any stand out as very different.
There is a range of estimates, but all are below zero, indicating a Republican lead. The two most Democratic models (using only Dem polls plus the internal leaks) stand out as most different from the rest. Half of the models fall between -4.3 and -7.4. The two most Republican estimates put Brown's lead at about 8.8 points.
The caveats are that turnout may yet matter, for either side. Reps enjoy an enthusiasm advantage, according to the polls, but Dems might yet mobilize their voters beyond what the polling suggests. And there is the unknown of the GOTV efforts on Tuesday. But if Coakley wins, this will be a major surprise, and the pollsters will have a lot to rethink about their methods. A win for Brown will have huge implications for the Democratic policy agenda and will put the fear of God into Democrats running in November.