Election Simulations on 538
By: Stephen CollierOne of the blog phenomena of this electoral cycle has been 538, a blog written by a statistician who uses simulation techniques to create predictive models for electoral outcomes. I haven’t looked into the details, but in their broad structure these models are similar to the simulations that we have been looking at from the 1960s on in the context of defense and emergency management. They incorporate a bunch of electoral and demographic data, and then run simulations using a randomizer (like a Monte Carlo simulation). Effectively, this randomizer produces a large number of different “worlds” — which are just outcomes of the simulator. Back in the day it took weeks to run one such simulation. But now, with massive computing power, every time new data comes in — in the form of new polls — they plug them in and run the simulation again. It is then possible to run standard statistical analyses on the outcomes of these simulations, essentially treating them like an archive of past events. If you check out the charts on the right side of the home page, you see an “electoral vote distribution” graph. This essentially shows the number of simulations that produced a given outcome in terms of electoral votes. From this you get some probabilities that a given candidate will win or lose, but also win or lose with different combinations of state-level outcomes.
In fact, this sort of thing is becoming increasingly routine. I have seen similar techniques applied, for example, to baseball statistics. (One particularly interesting example was an attempt to use simulations to figure out how likely it was that the record for consecutive games with a base hit would be tied or broken — the answer is fairly likely). And this is definitely the technique used in many formal catastrophe models.
July 26th, 2008 at 1:28 pm
As Stephen notes, my impression is also that these simulation techniques are becoming more wide spread. Two anectodal points:
1) My housemate is a super-quant ph-d student in political science. As I had in the past pushed him to tell me how he feels about simulation techniques as opposed to tradition statistical analysis methods, his response consistently has been that they are not fundementally different. For him, and I suspect in general, simulation seems to be a more convinient way to come up with data; and in this respect from a practical point of view it is not fundementally different from traditional statistical analysis. I thought the difference may be comparable to another shift that took place in 19th century (i think) in terms of data collection. As Hacking notes, at some point statisticians realized that partial data collection, i.e. sampling, was more efficient and had less error than collection of data from the entirity of a population (i.e. the sample size is as large as the population size). Can we see simulation as an other step towards further abstraction/seperation of signifiers from the signified?
2) One might dispute the significance of simulation in cases such as baseball games. Even in the case of electoral voting, it seems quite essoteric (though I am sure it has profound consequences in terms of the results of elections). However, if we go back to the issue of simulating the economy, it seems like Goldman Sachs is doing this in a routine basis for taking positions in the market place. They simulting the US and the World Economy to generate worst case economic crissis scenarios. I suspect this is used on a wide scale for pricing of assests in general.
July 26th, 2008 at 2:09 pm
Simulating electoral outcomes seems esoteric? Not so sure about that. But in any case I agree with you that other applications like financial markets modeling are very interesting (I had a long conversation with an economist about real estate models used by hedge funds yesterday, but forgot to explore this angle of it).
One question: Is “step toward abstraction/separation of signifiers from signified” the right way to see this? It is interesting to think about this viz your observation about the 19th century move away from “polling” an entire population because weighting samples proved *more accurate*. I am reminded of an argument that Henry Brady made after the 2000 election debacle, when he said that good exit polls might actually be more accurate than the actual ballot returns because the error rate was so high in the counting, and with the registering of votes (due to hanging chads, butterfly ballots, and all those other things we learned about in November/December 2000). So isn’t this a case in which sampling procedures or even simulations actually bring us *closer* to the signified, if we still want to speak in such metaphysical terms?
July 26th, 2008 at 2:39 pm
Just on a side note: the blogger behind 538 is Nate Silver. He actually won fame in the baseball statistics world for developing a system similar to the one he uses for political polling that forecasts the career of baseball players (its called PECOTA).
July 26th, 2008 at 2:50 pm
Maybe it was actually he who did this other analysis I happened to read.
This makes me think, rather tangentially, of another question in all of this: how widespread or specific are these kinds of techniques. What kinds of experts engage in this sort of modeling? When I was working on disaster models, I was struck by the extent to which there seemed to be a pretty small universe of experts and organizations that comprised the modeling community. They were in various places — insurance companies, some doing contract work for the government, etc. — but they seemed to have a sense that they had a distinctive approach. I am not sure whether that is because there was a common contrast with a more dominant form of risk analysis in the insurance world, or whether it was really that this was an important locus for this kind of modeling more generally. But in any case, it would be good to know who exactly the “practitioners” are in these worlds — particularly as this moves into being another taken-for-granted approach to risk assessment (maybe it has already become so long ago).
July 28th, 2008 at 8:50 am
Stephen,
As regards how widespread this is and has been, you might want to check out the history of Operations Research. As the writers of this blog know, operations research (OR) is the “discipline of applying advanced analytical methods to help make better decisions” (from http://www.scienceofbetter.org/). In the early 2000s, The Institute of Operations Research and Management Science, the professional society of OR researchers, had its 50th anniversary, and they released an online number on the history of OR. Below is a link to an article that traces the organizational backdrop of the OR professionals:
http://www.lionhrtpub.com/orms/orms-10-02/frhistory.html
It seems the field as a professional practice is initiated during World War II and then turned after that to industrial and civil time problems. The early meetings in the 1950s were attended by a couple of hundred people, almost all of them employed by industry, government or the military. Seems to have been a flourishing professional practice, and this social integration might have created what you term “a sense that they had a distinctive approach”.
Inside the article are also documented interesting disagreements on what constitutes the object of OR research (military or business), and thoughts on why the membership of these organizations has declined since the 1970s despite the development of the computers and the increased access to data. All very contemporary problems as well judging by someone who studies OR.