For every set of data, there are literally an infinite number models that will fit that set. Fitting the data is literally a trivial task. For this reason, the strength of a model has zero to do with how well it fits the past data. Fitting the past data is just a minimal requirement and nothing more.
You do realize the “gender gap” for this year’s election is just about where it’s been in the past, correct? But regardless, your post is still not a serious response since-- if Obama loses a close election-- neither you nor any other Democrat will argue it’s because of some gender gap and being unable to relate to men or whatever.
First, you seem to forget the fact that Romney is actually leading among independents so I don’t know what your “enough independents are also noticing” line has to do with anything. Second, unlike Democrats when it comes to abortion, Republicans aren’t exactly single issue voters and are more apt than Democrats to vote for someone who has a different view on abortion then them (guess that cuts into the whole “playing-on-social-issues” line that gets trotted out on this board at least twice a week).
Because the OP is a stupid question which is going to elicit certain partisan type responses endemic of this board. Elections in a presidential year entail more than just voting for the president. If the GOP loses the presidential race but retains control of the House and picks up the Senate, then they’re not going to change anything anymore than the Democrats will change anything if they win the presidency, fail to retake the House and lose the Senate.
Okay, so how did they act differently after 2010? Throwing out 2010 as an aberration seems quite… convenient.
Aside from, apparently, the one you agree with.
Of course they won’t care, because if you can win an election by just appealing to your base, why would you try to broaden your appeal?
Not really. One, the people more likely to vote Democratic are generally the least likeliest to vote. Two, a state’s voting characteristics are not immutable and are influenced by things such as births, deaths or citizens moving/leaving the state, etc. It really doesn’t matter what one registers as today; it isn’t indicative of their future voting patterns.
I know it’s just one anecdote, but I’ve had an Obama sticker on my car all Summer and I actually had a guy approach me a few months ago at a gas station and ask if I ever got tired of people hassling me because of the sticker.
When I told him this was the first time it ever happened he seemed honestly flabbergasted.
Huh?
When you have a variable, and a few other variables you think cause it, you can find a model that fits better than another.
My ice cream model, for example. The best model shows ice creams sales increasing with temperature outside, using past data. It predicts the future fairly well.
You keep saying this but you’re still wrong. It doesn’t matter how well your model fits past data because it’s trivially easy to find a model that fits. There are techniques to select a good model that is more complicated than minimizing the error in your training set, which you’ve ignored. But even if apply those techniques the fact that your model fits the data you based it on is not indicative of predictive power. To establish predictive validity it has to predict new data, whether it’s actually new or you are cross validating it against historic data that you didn’t include in your set used to build the model.
Help me understand then… if national polls are irrelevant then why does the 538 site calculate an Adjusted polling average which is “adjusted for change in the national trend since the poll was conducted. Polls are also adjusted if the polling firm used a “likely voter” model, and for “house effects” (consistent patterns of partisan bias exhibited by the pollster.”
I’m not sure how much of an adjustment the 538 site is making. Anyone know?
Okay, I’m not saying that “the fact that your model fits the data you based it on is not indicative of predictive power.”
:dubious: Wanna try that again?
If you live in Berkeley, why do you get your gas in Stockton? 
(Feel free to replace these locations with the appropriate equivalents for your region…)
Yes, but most of these fittings will be overly-complex. That’s why useful models must be simple. A 4-variable model with good prediction may be better than a 12-variable model with excellent prediction.
I find few details about the Bickers-Berry model on-line. Two things I’ve gleaned about it: (1) it seems to be based on a very simple model, perhaps as few as 2 variables (unemployment and income) per state. (2) it predicts state-by-state … yet all the on-line hype is about overall result. Shouldn’t it hype the percentage of states predicted correctly?
Polls are also good predictors of course, so anything like the Bickers-Berry model could be viewed as seeking to estimate an “error” term – the difference between what a voter tells a pollster in August, and which lever he presses in November. Are polls (or past results) built into Bickers-Berry and other such models?
In most midterm elections the total vote is approximately 50-60% of a presidential vote and that hold true for both parties. As Wesley Clark pointed out the percentage of Democrats went down as expected. The Republican percentage declined by a far lesser margin. That’s great for the Republicans. It’s how you win elections. But, AFAIK, such a discrepancy is rare to nonexistent. That’s the definition of an aberration. The party line votes in 2012 will parallel those of 2008, not 2010. The GOP leadership knows this. That’s why the heavy money and the party organization went to Romney early on. That’s why Tea Party candidates have not been, for the most part, backed by the Republican establishment. The vote in 2012 will be broader and more central than in 2010. This has huge explanatory value for how the races and candidates have turned out.
Unfortunately for you I have said this consistently for the past more-than-a-year I’ve been posting in Elections. Along with all the rest of what I’m saying now.
It’s obvious, at least to me, that the Democrats are not appealing just to their base any more than the Republicans are. They both are competing for the central group of voters most likely to swing. The bases of both parties are unhappy because they are being mostly taken for granted. This is totally realistic. At worst they’ll sit it out rather than voting for the other guy.
Remember also that the swing voters are not the same as independents. They don’t care or dislike much more than they have a consistent ideology. What gets them to vote on Election Day is not the same as what get the bases to vote.
This is the counter argument that has to come true if Republicans are to go on as a viable party over the next 25 years. Predicting the future is made harder by the fact it hasn’t happened yet. But the growth in minority voters and the lessening of the voters in the Republican base requires that the GOP changes its core policies and values to appeal to groups who now actively hate and fear them. That’s hard. There are pragmatists in the party who advocate for this right now, but their voices aren’t being heard. Figuring out how to do this is the biggest what if the Republicans have to answer.
Silver believes that the trend line of a series of polls is in an indicator of the final result, and uses that as a factor. Individual polls don’t matter much. It’s the comparison of the trend line to past elections that’s important. I don’t know how much weight he puts on that, but I know it’s just one of many factors.
In this particular election, with the national totals so close - and having sat in the same spot for months - I put less weight on them than in most elections. There are only ten states that are remotely in play, and probably less than five. So I really don’t care what the national polls say. I’m much more interested in state polls.
As the OP deals with the hypothetical that Romney loses and what the Republicans will do that is an irrelevant if. And the point was that if the election is close that gap will be part of the difference. If that is normal then the point stands, Republicans need to change to make that “normal” disappear.
I did not say that this would be a decision made by just a single issue, what is clear is that on the subject I mentioned the polls are clear, independents and even Republicans do not agree with the extreme yahoos that are leading the party.
Eh, 1984, 2004. I had a brain glitch. There was still a slow but pronounced course correction following Reagan’s landslide re-election.
This is completely right. To respond to Frylock’s remark, there may be an infinite number of models, but there is not an infinite number of linearly independent ones that satisfy some given measure of goodness-of-fit. Goodness-of-fit is necessary on some level, but it is of course insufficient. There are a lot of problems with over-interpretation of fit statistics, but that’s another issue entirely.
I should have access to this journal via my university. I’m just about to take my comprehensive exams, but I’ll pull the article and post some of the salient features of the model.
Very astute. Why yes, I was driving through hicksville at the time. But this just further makes my point. The rightest of the right think on the fringes, live on the fringes, and then believe everyone else thinks exactly the same thing.
I feel like this is different than the Dems support for Obama (myself included). Obama is like Ferris Bueller and Romney is like the Snooty Waiter. Obama (like Bueller) is very beatable. But just like the Snooty Waiter CAN’T be the one to bust Bueller, Romney is the wrong man to challenge Obama.
He’s going to lose, and he’s going to lose big. And I seriously believe a bit of a course correction will be in order for the Pubbies. The smart will finally realize that they can’t win the Presidency going further to the right and they’ll recognize that eventually the same Dem support that Obama receives will descend on their districts. It’ll be survival, but it’ll also be great for a left-leaning populace.
Ok, so this issue of PS is not out yet, so I take that back for at least a few weeks.
For what it’s worth, PS is also a pretty low-rent journal as far as these things go. Even though it is published by APSA, it is not the APSR. I’m just a grad student who will soon be trying to get something published in a politics journal, and this would not exactly be my first choice.
I’d think there still would be an infinite number there. But in any case, the main point is that this is, as you said, necessary but not sufficient for rating a model as a “strong” one. For that, it must show predictive power.
You seem to be talking cross each other.
538 does several things. One prediction it makes is what the national popular vote will be. Another is state by state predictions and by way of that the electoral college prediction. Hence the national popular vote prediction of a fairly close race but given the state by state probabilities a greater than 2 to 1 (closing in on 3 to 1) odds of an Obama electoral victory. Yes a significant persistent national shift in the popular vote numbers would reflect in individual state numbers, and should bias the pot, especially in predicting states less extensively polled by a variety of known firms … and while such is not impossible it is fairly improbable to occur in this race. OTOH carpet-bombing a few key states with a huge amount of ads might shift the results in those particular markets and those states that matter most will have extensive polling of their specifics by known firms and the national numbers will mean less there.
Bolding mine. Other factors continue and additional details in the site.
And correct, fitting past data is nice to generate a hypothesis; value is generated by how well it made predictions. 538 has made predictions well.
The idea of predictive power is problematic, too. One kind of model has a dichotomous dependent variable. For instance, either Obama or Romney wins the election, so the variable itself is binary. This kind of model outputs a probability that one or the other event is going to occur. So if the model says that the probability of Candidate A winning is greater than .5, then the model says that Candidate A wins. One measure of whether this sort of model works is called “percent correctly predicted”. Observe all of the actual outcomes, compare them to the predicted probabilities, and see whether the outcome matches a model prediction of greater than .5.
You can see why this is very problematic. The two possible events might be nearly equiprobable. If the model says that one will happen with a very high probability when that really isn’t true, then even if the model makes the right prediction, it’s wrong. You are no better off by guessing.
Unless Nate Silver got an advance copy of this paper (which I doubt), we really don’t have enough to go on to say one way or another. It is not enough to predict the outcome; we really care about whether the individual state-level EV predictions are correct. It’s of no interest if the model is right for wrong reasons.
I will try to give a summary when this issue of PS comes out. I’m sure other interested parties here have JSTOR, too, so we can hash it out soon.
I wonder if he’s already done the Monte Carlo simulation (or something similar) before he posts or not…maybe I’ll go check the methodology when I get a round tuit).