Being a nonstatistician, I wouldn't even begin to know how you would incorporate this into an aggregate or a "logistical" model, but what about the effect of massive infusions of money into this election cycle? This will be the first Presidential election - thanks to the 2010 Supreme Court "Citizens United" decision - where any one individual can throw whatever amount of money they wish to whoever they wish. A case in point is Sheldon Adelson, the Las Vegas casino executive, who has told Forbes magazine that he intends to donate up to $100 million to Mitt Romney and Republican candidates. That means that really competitive states, such as Florida, are going to be flooded with television and radio advertising - with plenty of "negative" ads. There surely must be a saturation point where people (potential voters) literally get sick of the constant unrelenting barrage and just tune out. The greater question is what effect (and how do you measure the effect) of virtually unlimited spending on these races? (I wish I owned a chain of television and radio stations down in Florida.)

In your final results, I notice the 0.5 percent advantage you have for Mitt Romney in Florida. Florida is the key state. In nearly every electoral college map I have constructed on the RCP web site, Mitt Romney has to win Florida in order to have any real chance of winning the election. In what could be a repeat of 2000 and the Bush versus Gore showdown, the Presidency could once again boil down to Florida. I have a friend living in the sunshine state who's tight with the Monroe county Elections Supervisor or the sheriff. My buddy insists that Obama is going to win Florida. (He bases this prognostication on an aggregation of several statewide polls which, he says, all show Obama out front. He also says that Florida's Republican Governor - who he calls a crook - is very unpopular and Rick Scott's unpopularity will hurt Romney.) Don't know what to make of my friend's wisdom, but if your 0.5 percent number in favor of Romney is accurate, then a couple of things suddenly make a lot of sense.

First, Republican Governor Rick Scott's controversial campaign to purge the voter rolls of "noncitizens" and, second, Obama's announcement this past Friday of a change in immigration policy. Considering the reaction of Romney (and Republicans) to the immigration announcement, it will be interesting to see how the effect of that policy change shows up in polling data - especially in states like Florida, Colorado, and Nevada. I would think the net effect of the announcement, judging by the reaction of Romney and the GOP, is a net plus for Obama. In fact, if Florida is truly as close as your model indicates, this may very well have been a major factor motivating Team Obama to announce the immigration policy change. (Judging by Mitt Romney's inability to answer a question that was put to him five times Sunday by Bob Schieffer, it does appear that Obama's executive order will have an effect on the polls. The question is how much of an effect?)

Good work with your model. This is really interesting - even for a nonmathematician like me!

Measuring the Tom Bradley Effect

Mr. goofball:

How do pollsters detect dishonest or misleading answers (which might tend to produce an incorrect or inaccurate poll) and how do they adjust for such deception? For instance, nowadays most racists try to be subtle about their racism. They never tell a pollster flat out "I'm going to vote against President Obama because he's black and I don't like black people." They might even flat out lie to the pollster (saying they will vote for Obama) knowing full well they have no intention of voting for Obama. (I think this phenomenon is called the "Tom Bradley" effect.)

Likewise there are people who dislike the Mormon religion - considering Mormonism a "cult" - and will not vote for Mitt Romney due to that belief. I suspect very few of these people will tell a pollster "I'm voting against Mitt Romney because he belongs to a cult." Nowadays people are very sensitive to being thought of as being a racist or intolerant to another person's religion - even when it's true! They may feel and believe it - but they're loath to admit it. Maybe this is due to my growing up and living in the Deep South and remembering what things were like 50-60 years ago, but I still think vestiges of those attitudes remain to this day.

Do pollsters have standard techniques for detecting when they are being lied to or misled? If not, do they simply introduce a "fudge factor" into their results which assumes that five percent of respondents (or whatever percentage) are probably lying to them so they adjust the raw numbers accordingly? (Can any poll or math model detect dishonest responses?) Also, since this election involves two candidates who each have a trait which some people find objectionable, (i.e. race and religion), what are the chances that polls could be wrong in some of these swing states?

Good question, thanks.

As a general rule polls do not attempt to correct for possibly dishonest answers. There's just no good way to do it and, except in very rare cases, dishonest answers come down on both sides and balance each other out.

In the specific cases of the president - I looked at a handful of states in the 2008 election and he did not appear to dramatically over or under perform his polling. Additionally I think racists have plenty of reasons they can cite as being anti-Olbama (socialist muslim birth control non-citizen) without feeling scared of betraying their racism.


For Romney, all there really was to look at was his election in MA (source http://en.wikipedia.org/wiki/Massach...election,_2002) and he appears to have outperformed his polling rather substantially - in kind of an anti-Bradley effect. (from around -2 to +5). So while I have no specific reason to think there is some consistent answer bias that leads Romney to overperform his polls, I certainly can't rule it out based on the MA election.

I'm have no plans to work that into the model without some serious corroborating evidence for a few reaons.
- It was a governor's election
- It was 10 years ago, Romney has changed a lot, the country has changed a lot, polling methodology has changed
- I have no idea how complete the wiki list of polls is
- I have no idea if there were others cycle specific factors that helped Romney overperform his polling.

So while I don't think the election comes close to justifying my adjusting the model, it is interesting and good to keep in the back of your mind. Plus SeattleLou can cozy up with it at night.

It's a little different. But as more polling comes in I may be able to surmise what correlations are to some degree, but since discerning from actual data relies on looking at how variables change together (and I suspect correlation studies give a lot of noise), I need to be hoping for a very volatile election to get some real results. I think I"m better off building something based on proximity and partisan index as suggested by Reno Expat.

I do think your idea to wait until the election is closer is good, for now I think my range is the best way to express an estimate.

Thanks

I think Obama's immigration move was brilliant politically (and hey good policy too) as evidence by the republicans being put between the rock of opposing overwhelmingly popular policy and the hard place of agreeing with the president.

I agree with you that Florida will end up very close and dear god there hasn't been a poll out of there in two weeks. WTF Florida.

pump?

Thanks!

Wouldn't looking at Republican primary results give a more thorough picture?

It definitely would. Go for it I'd be interested in what you find

Yeah, sorry I'm taking so long. It's simple and clean in an excel spreadsheet but more difficult to translate into words. I'm working on it.

*** 6/24 Update ***

Electoral Map




State Level Details



PA, WI, NH and VA all sit in the same order atop the list. Michigan drops off thanks to two new bad polls for the president on 6/18 in that state showing up -2 +1. It's very surprising but for the moment it looks like while Romney's name may not win him Michigan he could force the president to use some resources in that state.

Outside Michigan it was a very good polling week for the president with polls showing him +7 in CO, +6 in WI, +4 in FL and +5 in NH. It is important to remember however that this early in the race one week's polling can't move the status of the race dramatically. The president is showing a 1 tick upgrade in winning chances across the Monte Carlo simulation.

Monte Carlo

When I built the model I calibrated the logistics using a sample size of 10,000, this is why I use square root ratio to estimate the error introduced by samples which fall short (or exceed) that standard.

Let's try again with an abbreviated version.

Adjusting Error for Effective N

A poll ha 3 main sources of error:
1) The sample doesn't adequately represent the population
2) The population doesn't adequately represent the voting population (response bias, cell phone bias, question bias)
3) No one know what will happen between now and election day.

Since (1) is the only source of error that is calculable a priori I imagine I live in a world where it is the only source of error (perfect poll done on election day). If that were the case I could calculate exactly the error for any given sample size. I can also calculate a candidates true chance of winning for any given poll outcome comparing the two normal distributions (polling mean and error) For example a candidate down by 1 in a perfect poll of 500 voters on election day has a 38% chance of winning.

I can then modify the logistic formula by the ratio of the sample size and 10,000 then calculate that same candidates winning percentage using this modified logistic. For the candidate above the logistic gives them a 36.6% chance of winning.

I've posted 3 graphs below. They are for a candidate who is down 1, 2 and 3 points respectively in the election day poll described. The blue line is the a priori (two awesome uses of a priori holla) calculated winning percentage for that candidate as described above, and red line is the same value calculated by the modified logistic.

The x-axes are sample size and the y-axes are candidates change to win. Using direct ratio and ln ration these graphs matched up very poorly. I'm still amazed by how well they match up when using the sqrt ratio.

Bunch of polls out today. OR isn't a state I was officially modeling not a big enough sample but the +8 is super good news for Obama.

I dunno WTF is going on with that Romney +4 in VA - there hadn't been a poll there for a while so it had a healthy effect moving obama down to about 60% in VA.

The rest of today's polls (CO O+4, OH O+3, NC R+3) were all good for the president.

Ohio +8 for Obama today!!!

Update! Based on actual math.