Princeton Election Consortium

A first draft of electoral history. Since 2004

Political science model matches PEC long-term forecast

October 7th, 2014, 3:55pm by Sam Wang


In TalkingPointsMemo is a rundown of political-science models of this year’s Congressional campaign. Such models are research tools that use pre-campaign fundamentals to test a hypothesis about how a campaign “ought” to turn out.

Today I point out that the Highton/Sides/McGhee Senate forecast has, in some sense, already been confirmed: it is essentially identical to PEC’s long-term forecast, which was based on polls from June to now. In that sense, the TPM article didn’t mention a pretty interesting fact: the match between our polls-only analysis and at least one non-polls-based model.

Highton et al. say their “current estimate is for the Republicans and Democrats to each control 50 seats with a 95% prediction interval of 48 to 52 seats.” Above, shown in yellow and red, are the PEC forecast as of September 9th; note specifically the prediction of 50.3 ± 0.7 Democratic+Independent seats. My “±0.7″ corresponds to the red “1-sigma interval,” and the yellow zone is approximately the 95% prediction interval: 50.3 ± 1.4 seats – or 48.5 to 52.1 seats.

A widely held view in political science is that campaigns play the role of driving public opinion toward some natural endpoint. This nearly perfect match between model and polling reality since summer suggests that in 2014, the campaign started reflecting fundamentals months ago. Of course, we’re at a point in the campaign when polls provide a much sharper view than pre-campaign fundamentals. Today, the PEC November prediction is for Democrats and Independents to have 50 or more seats with 63% probability. If outcomes match polls, the most likely outcome is looking like a 50-50 split with Republicans.

One technical note: the upheaval in Kansas in early September drove the PEC prediction upward slightly. Not counting the “Orman offset,” PEC’s long-term prediction was actually 48.0-51.6 seats. That’s still entirely consistent with the Highton et al. model.

Tags: 2014 Election · Senate

11 Comments so far ↓

  • MarkS

    OK, this post is just bizarre. Sam is now claiming that agreement with his regress-to-the-June-Sept-mean assumption (an assumption that he completely abandons 5 weeks before the election) should be considered as validation of one particular “fundamentals” model, even though he has forcefully argued aganst using fundamentals in the first place.

    • Sam Wang

      No, my point is that our starting assumption was not so off-kilter as painted in public. That’s all.

    • Richard Wiener

      If I understand correctly, Dr. Wang’s point is that it is unnecessary to include fundamentals in election forecast models. A polls only model works, because the information gained from fundamentals is already reflected in polls. A hybrid model like 538′s runs the risk of introducing systematic error by double counting the effect of fundamentals. It is an empirical question as to whether fundamentals are already baked into the polls. Agreement with a fundamentals only model suggests either both models are coincidentally wrong by the same amount or both models correctly reflect reality. If the latter, this is evidence that the information from fundamentals is already embedded in polls as Dr. Wang has hypothesized.

  • Steve Scarborough

    Pgvaidya: Also, I forgot to mention that vol. 28, issue 4, of the International Journal of Forecasting has articles on election modeling. http://forecasters.org/ijf/journal-issue/468

  • Steve Scarborough

    Hello pgvaidya. I tried to post a longer reply, twice, but it did not go through.

    In short, while I have heard of Taken’s Embedding Theorem, to be honest I am not knowledgeable. What I was thinking of was the use of forecasting models in practical business applications such as described in articles via forecasters.org. (The International Journal of Forecasting.)

  • Violet

    Sam, it looks like 538 is already trying to hedge their bets, by claiming they expect a lot of “pollster error” in the midterms, which I guess they will cite to if they have to explain why their prediction was wrong. http://fivethirtyeight.com/features/pollsters-predict-greater-polling-error-in-midterm-elections/

    • Sam Wang

      The thing is, that’s true…not just this year but any year when a contest is this close. They haven’t experienced that. The last giant suspensefest was 2004 Kerry v. Bush, which also came down to a few states.

  • Steve Scarborough

    Hi Sam. Interesting about the Highton, Sides, McGhee work.

    I and others have mentioned time series models as a possible enhancement to your meta-margin. Or for that matter, other discrete time series measures such as daily data. Example: your Senate Seat History (http://election.princeton.edu/senate-seat-history/)
    In my day, which was some 20 years ago, I was a forecaster of Medicaid caseloads. Now long in the tooth, my knowledge and methods are very dated now. However, I used methods like exponential smoothing, Box-Jenkins, and State Space models. While I have mentioned all this before in another post, to repeat, I strongly recommend evaluating such methods.

    While I do not have MATLAB (over $2k for this old guy!) I see that it has state space capabilities in it. There are numerous references for this, such as: http://www.mathworks.com/help/control/ref/genss.html; http://www.jstatsoft.org/v41/i06/paper; https://www.youtube.com/watch?v=ADPsPBxfwXE; just to mention a few.

    Idea: You and/or a student, graduate assistant, etc. take one of your data sets, like the meta-margin, and investigate using state space modeling for it. Could be an interesting project with paper opportunities and so forth. Without diving into it myself, I suspect there are powerful implications on your election modeling work using state space methods. Try it. You might like it.

    • pgvaidya

      Steve,

      As you well know, state space models are now used in Engineering, quit routinely. By State space do you mean a system embedded in higher dimension, using Taken’s embedding? Although, proposed in the early 80′s, it is not used much in data processing even today and possibly has a potential in psephology.

  • mediaglyphic

    http://campaignstops.blogs.nytimes.com/2012/10/20/googles-crystal-ball/

    Does anyone remember the above article that used google to figure out the demographic makeup of the electorate in 2012. I wonder what this work would show for 2014. Is it in line with the LV models being used?

    • Eli

      Long time reader, first time commenter. Thank you Sam for the open and thoughtful analysis over the years.

      I remember Google’s Crystal Ball very well. Very interesting experiment. I wish the had continued it.