Princeton Election Consortium

A first draft of electoral history. Since 2004

Meta-Margins for control: House D+1.0% Senate R+4.2% Find key elections near you!

Meanwhile, in real news…

September 17th, 2014, 8:33pm by Sam Wang

1) Currently, the critical race for Senate procedural control (i.e. whether Dems+Inds keep 50) is in Iowa. Braley’s up by a median of 1.0% over Ernst. That alone is driving the daily snapshot most strongly. Iowa. Is. Important.

2) Scotland is voting on independence tomorrow. “No” is ahead by 4+/-1%. For such a consequential question, that’s very close!

That is all.

Tags: 2014 Election · Politics · Senate

12 Comments so far ↓

  • bks

    For Scottish referendum, Ladbrokes (UK betting) is paying 1 to 5 if you want to bet on No, 7 to 2 if you want to vote on Yes:

    • shma

      4±1% puts into odds is about 31,500:1 against a yes vote, for comparison.

    • Sam Wang

      It also depends on the systematic error. (That seems like a big topic today around here.)

    • Gregory Primosch

      >> “No” is ahead by 4+/-1%. For such a consequential question, that’s very close!

      Actually, this speaks directly to the confidence questions Silver raises. If the model really says that that No is 4±1% then that it is NOT very close (virtual certainty of No success). How do you reconcile your intuitive nervousness of the outcome with the near certainty of confidence intervals you are citing? What about the model would have to be changed to account for this?

    • Sam Wang

      This is a good point, and I do have an answer.

      Here is the problem: when a margin is 4%, the outcome is nearly always decided. The only exception I can recall is the 2010 NV Reid vs. Angle race. In retrospect, there were good reasons why the polls were off: a mobile population, lots of Hispanics who are hard to reach, that kind of thing. But those reasons only became apparent afterward.

      In my view, the answer is to take two steps: 1) add a small uncertainty to account for actual median poll-vs-election error (sigma_systematic, which I defined on the other thread). 2) Feed the margin into a distribution that is not bell-shaped, but has long tails. This, by the way, is why I like to go off about t-distributions. They are much “tail-ier” than a bell-shaped curve, and capture that “hey, freaky things can happen once in a blue moon” vibe of live politics. I have come to love t-distributions.

      In this case, Gaussian statistics using the 1% uncertainty would give a “No” win probability of 99.997%, which isn’t right, as per Silver’s criticism of my 2010 calculation. However, using the approach I outline here, if we call the true uncertainty 2.0%, and used a t-distribution with 3 degrees of freedom, the probability becomes 93%. This seems more plausible.

    • shma

      Even assuming a 2% overall error and higher chance of a black swan event, 1:5 odds are good enough to take.

      But not as good as the even money bet that YES will get between 45 and 50% of the vote. Unless they’re counting spoiled ballots in that total, that’s just ridiculous odds.

    • Elithrion

      Similar (and slightly better) odds on Betfair, e.g. 1 : 2.12 for “Yes” to be 45-50%. I feel like these betting sites usually have more last-day uncertainty than is warranted.

    • Sam Wang

      They usually do. I demonstrated this back in the day.

  • Froggy

    With a new Fox poll today, we now have three post-Taylor-withdrawal polls for Kansas:, with the median being Orman +1%.

    Is it time to use these three-way-race polls, or hold off for the time being, given the uncertainty as to whether Taylor will be on the ballot, and the effect that has on the poll results? (The Fox poll has Roberts +2% for the three-way race, but Orman +6% for a straight Roberts vs. Orman contest.)

    • Sam Wang

      We’ve been using two-way race data, but at this point it might not hurt to wait until the Kansas Supreme Court has ruled on the case.

  • FJ

    Final result in Scotland is NO wins by 10.6%. That is pretty far from the polls. Even the student t-distribution that you propose (centered around 4%, variance 2%, 3 degrees of freedom) only gives a result at least this good for the NO side a probability of 2%. (The Gaussian estimate of this chance is 2* 10^(-11), so that approximation of where the race was at was almost certainly wrong.)

    Just a fluke of the polls, or something more fundamental?

    • Steve Schran

      The discrepancy between the polling and result of the Scottish referendum is that the final polls can be attributed to two factors. First, less importantly, there was movement toward “No” in the final polls that wasn’t incorporated in the median. But second, I think it is very reasonable to believe that there was a significant “Silent No” effect, coincidentally perhaps of similar size exhibited in the Quebec ’95 Referendum. /wiki/Quebec_referendum,_1995.

Leave a Comment