With the latest batch of Senate polling results, Neil Stevens of Likely voter has updated his Senate Projection. According to his latest model, he projects R+7, with a 3% chance of Republicans reclaiming the majority (R+10 or greater).
The problem I have with the Likely Voter projection model is two-fold: one, it factors in races that are, for all intents and purposes, already decided; and two, it seems to assume a normal distribution. If a projection were performed in which seemingly non-competitive races were removed, and the ANOVA based solely on the actually competitive races, intuition tells me that such a projection (especially with a right-leaning distribution) would have to center around R+8, if not +9. To wit:
Introducing Uncertainty Based on Non-competitive Races
Range of Possible Outcomes
First, let's set the table:
- Seats not up for election: 63 (23R, 40D)
- Seats up for election: 37 (18R, 19D)
- Bondary of Possible Results: R-18 to R+19
Now, let's add in some realistic boundaries to those results.
Range of Realistic Outcomes
Lower Boundary of Realistic Outcomes
- All 18 R-held seats are 100% likely R. (R+0)
- The following D-held seats are 100% likely R: AR, IN, ND, PA (R+4)
It is logical to conclude that, at this point, any projection that shows anything less than R+4 is just not consistent with reality.
Upper Boundary of Realistic Outcomes
- The following D-held seats are 100% likely D: HI, MD, NY, OR, VT (R+14)
It is logical to conclude that, at this point, any projection that shows anything more than R+14 is just not consistent with reality.
Range of Realistic Outcomes
So, at this point, the range of realistic outcomes is R+4 to R+14. Anything outside of these numbers should be considered 0% likely.
Range of Likely Outcomes
- The following D-held seats are 90% likely R: CO, WI (R+6, lower)
- The following D-held seats are 90% likely D: DE, NY (s) (R+12, upper)
So, at this point, the range of likely outcomes is R+6 to R+12. Anything outside of this range should be considered unlikely.
Likely Voter's current projection distribution curve has a mean of R+7, and R+5 - R+8 accounts for 77% of all outcomes (R+7 22.6%, R+6 22%, R+5/R+8 35.4%). If I assume that Likely voter's probability curve is normally distributed, then, IMHO, the mean simply must be shifted too far left. There is just no possible way that R+5 has 18% probability. I'd say, at the absolute upper end, it has 5-10% probability. Balancing the 90% Likely R pickups against the 90% Likely D holds lowers the probability even further.
So, just using back-of-mental-napkin calculations, I would say:
- <R+5: 0% likely
- R+5: 5% likely
- R+6 - R+12: 90% likely
- R+13: 5% likely
- >R+13: 0% likely
Analysis of Actually Competitive Races
The eventual outcome will be determined entirely by the results of six races: CA, CT, IL, NV, WA, and WV.
Two or three weeks ago, I would have rated those races as follows:
- Lean-R: The following D-held seats are 55% likely: IL, WV
- Toss-Up: The following D-held seats are 50% likely: NV, WA
- Lean-D: The following D-held seats are 45% likely: CA, CT
However, things have shifted a bit; I would now rate these races as follows:
- Lean-Likely-R: The following D-held seats are 60% likely: WV
- Lean-R: The following D-held seats are 55% likely: IL, NV
- Toss-Up: The following D-held seats are 50% likely: CA, WA
- Lean-Likely-D: The following D-held seats are 40% likely: CT
As you can see, aside from CT (which, to be honest, I am close to writing off as a potential Republican pick-up), all of the competitive races have shifted in the Republicans' favor. I put together a quick Monte Carlo simulation of my own, and here are the results:
So, my model projects a mean +9 seat gain for Republicans, and a 40.1% chance that Republicans will regain control of the Senate (a gain of +10 or more seats). Results:
- n = 10,000
- μ = 5.2
- σ = 1.4
- max = +13
- min = +5
- +8 - +10 = 73.1%
- 10+ = 40.1%
At first blush, these numbers appear to me to be more realistic, given the current state of the races in play (and not in play).
Evaluating the Normal Distribution Model
It seems that the Likely Voter projection model is based upon the assumption that the outcomes of competitive races will be normally distributed. I wouldn't expect a normal distribution for these outcomes, even in a "normal" election year - but especially not in a "wave" year.
Just as the outcome distribution of competitive races was biased toward the Democrats in 2006 and 2008, I fully expect the distribution to be biased toward the Republicans in 2010. This bias is due primarily to two factors that are not easily accounted for through pre-election polling: the enthusiasm gap and shifts in party affiliation.
In short, pollsters simply don't have a reliable means of estimating the breakdown of voter turnout, and it is entirely likely that they will tend to err on the side of a conservative estimation of the shift from 2006/2008 to 2010.
In a later post, I will examine some of these factors in each of the six competitive races.