I’m not unskewing anything!
Just to put that right up front. I’m also going to assume everyone is good with how polls are conducted (call tens of thousands of people, get a few hundred responses, weigh those responses to match the likely electorate, publish).
However! There are multiple ways to model a predicted electorate. And Siena, who are polling 100 house races this cycle and have completed 24, have been kind enough to provide us with a variety of those models in addition to their topline numbers. These models all reflect different assumptions about who is going to turn out for the election and Siena could have easily chosen any of them to be their toplines.
However, Siena have gone with a model that weighs their respondents against voter file information (age, party registration, gender, race, region, education) and to balance their likelyhood to vote between their self-reported intention to vote and their historical voting practice. Which I actually think is probably the best model to use. But plenty of very good pollsters with long track records of accurate results use other models, similar to the ones I’ll discuss below.
In the interests of not flooding you with spreadsheets of data I’ve put results for each seat into five bins; Likely Dem (D+10 or more), Lean Dem (D+3-D+9), Toss-Up (D+2-R+2), Lean R (R+3-R+9), Likely R (R+10 or more). I’ve also attached a picture of a summary spreadsheet at the end of the diary in case anyone want to look in a bit more detail.
Siena (Published toplines; uses age, party registration (or predicted party/primary vote), gender, race, region, education from L2 voter files and balances self-reported intention to vote and historical voting practice)
Likely D; CA-49, CO-6, IA-1 (3)
Lean D; MN-3, CA-45, KS-3, PA-7 (4)
Tossup; CA-48, KS-2, NM-2, MN-8, KY-6, IL-12, IL-6, NJ-7, TX-32, CA-25 (10)
Lean R; TX-7, FL-26, VA-7, ME-2, WI-1, TX-23, WV-3 (7)
Likely R; (0)
Ochre (Siena, but doesn’t weigh by party registration, predicted party, or primary vote)
Likely D; CA-48, MN-3, CO-6, IA-1, CA-49 (5)
Lean D; TX-32, TX-7, CA-45, IL-6, KS-2, KS-3, PA-7 (7)
Tossup; NJ-7, KY-6, MN-8, IL-12, CA-25, VA-7 (6)
Lean R; ME-2, NW-2, WI-1, TX-23 (4)
Likely R; FL-26, WV-3 (2)
Umber (Siena, but weighs using Census data instead of voter files)
Likely D; TX-32, MN-3, TX-7, CO-6, IA-1, CA-49 (6)
Lean D; KY-6, IL-6, KS-2, PA-7, CA-48, KS-3 (6)
Tossup; CA-45, MN-8, CA-45, NJ-7, IL-12, ME-2 (6)
Lean R; VA-7, NM-2, WI-1, FL-26 (4)
Likely R; WV-3, TX-23 (2)
Vermilion (Siena, but doesn’t weigh by education)
Likely D; CA-49, MN-3, CO-6, IA-1 (4)
Lean D; MN-8, NM-2, KS-2, CA-45, KS-3, PA-7 (6)
Tossup; CA-48, IL-6, KY-6, CA-25, IL-12, NJ-7, TX-32, FL-26 (8)
Lean R; TX-7, VA-7, ME-2, WV-3, TX-23, WI-1 (6)
Likely R;
Ultramarine (Siena, but only includes voters who self report as certain to vote)
Likely D; KS-3, MN-3, CA-45, CO-6, PA-7, IA-1 (6)
Lean D; CA-45, KY-6, IL-6, MN-8, NJ-7, CA-48, CA-49, TX-7 (8)
Tossup; KS-2, IL-12, TX-23 (3)
Lean R; WV-3, NM-2, WI-1, TX-32, FL-26 (5)
Likely R; VA-7, ME-2 (2)
As you’ve noticed the alternative models are all more bullish on Democratic chances than the Siena toplines. Which isn’t to say the are more likely to be correct. But it is to say say that they aren’t particularly more likely to be incorrect. If a equally respected polling firm other than Siena was running the exact same polls and got the exact same responses the published toplines could easily look exactly like the Ochre results, or the Umber, or the Ultramarine.
Take-home message 1; Polls aren’t particularly accurate in tight elections because weighing can easily shift a topline 5-10 points. So don’t sweat 5 point differences in a single poll.
Take-home message 2; The Siena house polls look really good for Democrats and the toplines are about as pessimistic as it is possible to be without doing something unreasonable like assuming 2014 turnout.
Read More