Louis Mittel writes:
Matthew Yglesias needs you to up your degenerate gambling so he can take the Economist model more seriously. He writes, “Nate’s models are the best because he is (and I say this with affection) a degenerate gambler who only cares about the statistical accuracy of his forecasts.”
That’s funny! It’s my impression that the election forecast is a bigger part of Nate’s public portfolio (however defined) than the Economist forecast is to the Economist’s portfolio, or to mine. So in that sense Nate should have a greater incentive than the Economist to get things right, or not to mess up where it counts. The overall incentives are the same—both Nate and the Economist “only care about the statistical accuracy” of their forecasts, but Nate has more incentive to stay up at night checking and debugging his method, because he is so personally associated with his forecast in a way that the Economists’ data journalists and outside collaborators are not. On the minus side, I have the impression that Nate’s approach is not fully model-based, and this can induce artifacts. On the other minus side, our forecasting approach has had problems too—indeed, last time we looked at our code, we found a bug (which we fixed, but there could be others . . .).
Also relevant in this discussion is the fivethirtyeight.com election forecast, run by Elliott Morris who arguably is even more motivated than Nate or the Economist team to get things right. Indeed, Nate, Elliott, and I discussed some of these uncertainty issues a few years ago. As Nassim would say, Elliott has skin in the game.
In practice, different forecasting methods will be giving similar results. I say this for three reasons. First, different methods are mostly using the same information—national polls, state polls, current and forecast economic and political conditions, and past voting patterns by state. As Hal Stern famously said, the most important aspect of a statistical method is not what you do with the data, it’s what data you use. Second, when developing their methods, forecasters are implicitly constrained by the consensus take on the election, as reflected by political pundits and betting markets. If you build a model and it ends up with what seems like an implausible prediction, you can adjust things to make the inference make more sense. There are so many choices in setting up a prediction model that it would in general be a mistake not to make such adjustments. We learn from the past, and our techniques improve; it’s not like we’re still using forecasting models from 2010 or whatever. Third, there’s nothing stopping anyone from altering their models midstream as problems arise. We want our models to learn from data but as a good journalist you would not want to tie yourself to the mast and commit to a forecast that you just think is wrong.
The other thing is that small changes in a probabilistic forecast can lead to big changes in the headline probabilities. We discussed this point back in 2012, where a simple calculation showed that a shift of 0.4 percentage points in Obama’s expected vote share corresponded to a change of 10 percentage points in his probability of winning.
Beyond this, there are general challenges for any forecast, regarding what uncertainties to include.
P.S. Speaking of skin in the game, here’s my post from last year on why I don’t bet on elections.
P.P.S. More on incentives in this long comment.
Andrew wrote:”the most important aspect of a statistical method is not you do with the data, it’s what data you use.”
Should be: “the most important aspect of a statistical method is not what you do with the data, it’s what data you use.”
Typo fixed; thanks.
Regarding different forecasting methods giving similar results, I completely agree with your assessment. However, right now, FiveThirtyEight’s forecast seems to be quite different from The Economist and Nate Silver’s. Of course it’s still early and this can and probably will change over the course of the campaign, but do you have any thoughts about these differences?
Raphael:
I haven’t looked in detail. The fivethirtyeight forecast is much closer to each candidate having a 50% chance of winning. Two ways to get a forecast that’s much closer to 50/50 are: (a) to have the forecast vote shares be closer to 50%, or (b) to have a higher uncertainty in the forecast. In this case, I think the difference in the forecasts is from the average more than from the uncertainty. Fivethirtyeight is currently forecasting Biden to receive 51% of the two-party vote. Oddly enough, I don’t see a national popular vote forecast on the Economist site, but it does have a current polling average that gives Biden only 49% of the two-party vote. The difference between 49% and 51% makes a difference—even though both are close enough to 50% for lots of uncertainty to remain.
Thanks! Interesting how such a relatively small difference in the polling average can impact the winning probability that much.
“Two ways to get a forecast that’s much closer to 50/50 are:”
There is a third way, push your “n” higher than everyone else’s (prestige!) by incorporating so many biased and/or crappy polls that you lose all variance and end at 50-50.
538 said that Biden has a 47% chance of winning the election, and then a couple of days later Nate Silver says Biden should drop out of the race! He even said Biden should resign the Presidency.
He refuses to say whether some other Democrat has a better than 47% chance of winning.
Roger:
About a month before the debate, Nate wrote that if Trump were to drop out, Biden would lose by 7 points. See discussion here. He recommended that Biden and Trump both drop out but did not give any forecast at the time about what might happen if that would occur. I think that his formal forecasts (as opposed to off-the-cuff comments) are based on poll averaging, and it’s pretty hard to do much with responses to survey questions about hypotheticals.
Nate Silver is no longer affiliated with 538, so this isn’t an inconsistency with one predictor’s predictions, this is a case of dueling predictions.
I had fun referencing some of the smack talk among these presidential election forecasters for my textbook. Let’s do some more this year!
D:
I don’t think of the “smack talk” as fun at all! It bothers me when people have these exchanges on twitter and don’t take the time to address the details. I’ve been frustrated with Nate in this regard; see for example the P.P.S. and P.P.P.S. of this post from 2020. I get that Nate is busy, he doesn’t owe me anything, he has no duty to respond to criticism by Elliott Morris or me or anyone else . . . still, it just seems like a lost opportunity for him not to engage. Getting thoughtful criticism from one’s peers is a great opportunity.
So where you see fun smack talk, I see a missed opportunity for scientific progress.
> is not what you do with the data, it’s what data you use
It seems like one big issue is that all of our election models are getting worse because our data is collectively getting a lot worse.
Phone survey non-response rates are through the roof.
To be fair, most of election polls are not using solely phone surveys anymore. See https://www.pewresearch.org/methods/2023/04/19/how-public-polling-has-changed-in-the-21st-century/
Interestingly, NYT/Siena, which is consider one of the highest data quality polls out there, still uses phone-only survey. However, they select from the voter file and have a carefully design sample and various weighting adjustments to try to mitigate nonresponse error.
I agree that Elliot has skin in the game. He wants to accurately predict the winner. But his social media posts belie that proposition. In 2020, he would post new updates to the (at the time) Economist model with caveats such as: don’t get mad at me. Or in the case that the update is positive for Democrats: cheers!
My point isn’t to disregard Elliot’s methodology. But he has multiple incentives, one of which is forecasting accuracy, the other is appeasing his fan’s partisan expectations. Not to mention his own partisan hackery.
Maintaining co-partisan hope—in Elliot’s case Dem’s winning—adds some insight into his decisions. I’m not convinced that he is more dedicated to the truth/predictive accuracy than seeing his arch political enemy get defeated.
Extremely polarized authors—both ideologically and affectively—should be viewed with skepticism. Especially when it comes to recalibrating a model.
Steve:
Fair enough. I guess we can embed this in a larger framework by considering all the other incentives that all of us have.
To start with, all forecasters have two somewhat contradictory goals:
1. To grab attention and stand out from the crowd. The incentive here is clear for a media-based forecast: you want to grab readers, eyeballs for your site so you can sell ads and subscriptions. But even without that direct motivation, if you’re going to the trouble of putting out a forecast, you want somebody to read it, even if those somebodies are potential private clients for off-the-record specialized forecasts.
2. To avoid making a mistake. This is the necessary flip side of goal 1: you can build a reputation with outside-the-box predictions that work out, but you can also damage your reputation by making strong predictions that fail.
The incentives here are complicated. In section 3 of our paper, 19 things we learned from the 2016 election, entitled “Overconfident pundits get attention,” we point to two amateur pundits who had received a lot of attention that year: Sam Wang, a neuroscience professor who had given Hillary Clinton a 99 percent chance of winning the election, and cartoonist Scott Adams, who had predicted a Trump landslide. Both these predictions were far off, but at the time Wang took himself off the hook by blaming the polls, while Adams was happy to have been directionally correct. Both Wang and Adams got attention in 2016 by making strong statements, but I think this ultimately hurt their reputations as pundits or forecasters. On the other hand, for the goal of getting social media engagement, any publicity can be good publicity.
The cleanest way to get the benefits of goal 1 (grab attention with a forecast that stands out from the crowd) while also satisfying goal 2 (avoid a mistake) is to get a scoop, or to get out ahead of the game in some way. That is, distinguish yourself in the time dimension rather than the forecasting dimension. Nate has done this successfully several times during his forecasting career: in 2008 he applied poll averaging to primary elections before others were doing it; in 2012 he set up an effective semi-automatic poll-averaging forecasting algorithm at a time when others were reporting public polls without converting them into a probabilistic prediction; in 2024 he was ahead of the pack of forecasters in talking about Biden’s age. Nate’s made some mistakes over the years, most notably in 2015-2016 when dismissing Trump’s chances in the primaries despite his (Trump’s) steady lead in the polls, but this was a mistake that just about all the prominent pundits were making back then, and Nate’s reputation survived the hit. By now he’s been doing forecasts for long enough that the occasional error can be seen in context.
My point here is that Nate has two incentives: one is, as Yglesias says, to get the forecast right; the other is to say ahead of the game and to not be just one more poll averager on the block.
Elliott is the new kid in town, relatively speaking, and he has an incentive to stand out too. He also has even more of an incentive than Nate to keep a wide forecast uncertainty, to not screw things up. Nate’s reputation allows him to make the occasional misstep, but if Elliott’s far from the crowd, his (Elliott’s) reputation is more at risk. Beyond this, yeah, to the extent that Elliott on social media (as compared to Elliott on fivethirtyeight) is a partisan commenter, then he has some incentive to maintain a level of partisan consistency (unless he wants to change his brand to be more contrarian) and keep his readers with him.
To get back to the forecasting for a moment, though: Suppose we accept that Elliott has a motivation, not just to get the forecast right—that’s a motivation that we all have—but also to avoid being caught out, Sam Wang or Scott Adams, with a prediction that’s strong and wrong. There are two ways to stay within the lines here:
(a) The first strategy is to hang out with the crowd: in 2016 this meant giving the nod to Clinton, in 2020, this meant leaning toward Biden but hedging your bets in case Trump again confounded the pundits, and in 2024, this means leaning Trump while trying to account for the chaos on both tickets. In short: go with the polling averages and then do some fudging to move toward the consensus.
(b) An alternative strategy is to shade your predictions toward 50/50 to cover possible unforeseen events. This is what Nate did successfully in 2016: in retrospect after that election, savvy journalists respected that Nate had threaded the needle by going with the polls and forecasting a Clinton win, but with a wide uncertainty that, unlike many other predictions at the time, gave a good chance that the polling average could be far enough off to make a difference.
Right now, you could argue that the fivethirtyeight.com model is following the second strategy here, by keeping the two candidates’ probabilities close to even. If polls continue to favor Trump, I’m guessing his forecast will move along and be close to the others, in which case I don’t see that Elliott’s forecasting algorithm should lose reputation points by being later than others to make that shift.
As in previous elections, a lot of this discussion is tricky, because much of our uncertainty comes from factors that are not in the polls, most notably what demented thing will Trump or Biden say next, while the prediction model includes this uncertainty only in abstract terms as polling bias and opinion drift.
I have the general sense that the team at the Economist has more incentive to avoid making a mistake than to stand out from the crowd. (I make this speculation about motivations as an outsider; it is not based on any conversations I’ve had with the Economist people.) They already have a successful magazine, they did solid forecasts in past elections, and probabilistic forecasting is part of their brand. They don’t gain a huge amount from being known as the absolute best forecast, but they can lose by being far off. As with fivethirtyeight, this motivates a forecast that follows the crowd while having a wide enough uncertainty to not get caught out by surprises.
As for me . . . my main incentive is to do research, and my main contribution to election forecasting has been various bits of political science research that have supplied insight and techniques for polls and voting, most notably:
1993: Why are American Presidential election campaign polls so variable when votes are so predictable?
1993: Review of “Forecasting Elections”
1997: Poststratification into many categories using hierarchical logistic regression
2008, 2009: Red State, Blue State, Rich State, Poor State: Why Americans Vote the Way They Do
2010: Bayesian combination of state polls and election forecasts
2013: Deep interactions with MRP: Election turnout and voting patterns among small electoral subgroups
2016: The mythical swing voter
2017: 19 things we learned from the 2016 election
2018: Disentangling bias and variance in election polls
2020: An updated dynamic Bayesian forecasting model for the 2020 election
2020: Information, incentives, and goals in election forecasts
2021: Failure and success in political polling and election forecasting
Forecasting is fun, and it’s a way to get some of my ideas out there, but my big ideas—Mister P, differential nonresponse, joint modeling of national and state polls, the relation between income and voting varying by state, etc.—are mostly coming from published research papers. There’s no strong incentive for me to be strongly identified with a forecast: the gain in reputation I would get by doing better, or earlier, than others in a forecast is less than the reputational damage I’d incur from being way off. So it makes sense that I’ve helped out with the Economist forecast in 2020 and 2024—I learn a lot from implementing my ideas, or some version of them, in real time—without going all in. As Nate might say, poll averaging election forecasting is a very small part of my portfolio, and he’s been doing his best to diversify too.
Can you comment on what the meaning of “mistake” is? I have a problem with these cost/benefit calculations in that all of these polls provide a probabilistic forecast and any probability not equal to 0 or 1 means that it is difficult to say a poll was “wrong.” Even going back to the 2016 election, it is difficult to say that the polls were wrong, even the more extreme ones that said Clinton had a probability of around 85% of winning. From a more technical point of view, we can say polls are wrong when we discover errors in the methodology. But I don’t understand how we can determine a poll was wrong based on its resulting probabilities. Doesn’t that interfere with this discussion of incentives of pollsters?
Dale:
Sam Wang gave Hillary Clinton a 98% chance of winning the election. She lost. You could say that this was a mistaken forecast or you could say Wang had really bad luck. Then we can go back to Wang’s model and see that it did not allow for the possibility of systematic errors in state polls, even though such errors had been present in the past. So I’d say he’d made a mistake. It’s also true that if Clinton had pulled it out, Wang’s mistake would not have been so consequential. That’s how things go: we make lots and lots of mistakes, and some of them hurt us.
Andrew
I don’t think that answers my question. If we are talking about the potential conflict between a pollster finding something different than others (getting headlines) and being accurate, is there a meaningful difference between saying the probability of candidate X winning is 55% or 60%? Or if a pollster forecasts the 2 vote share is 53% or 58%? Depending on the outcome, we could say one prediction is “more right” or “more wrong” but it is hard to see how that tempers the incentive to get a headline. Lately there have been headlines when a poll shifts from 55% to 53% for the probability of a candidate winning. I see very little that would cause a pollster to care much about the accuracy of those numbers – other than their ethical standards and their reputation among other forecasters.
Why not compare relative skill of forecasts using Bayes rule?
Precision, accuracy, and “surprise” are rewarded, as it should be.
Vague or unsurprising forecasts are not impressive and, accordingly, have low posterior probability even when accurate.
Whats the drawback from doing it this way?
Anon
I think you are right. Accuracy can be measured – in terms of the % of the popular vote, preferably state by state. I don’t think the probability of winning can be meaningfully measured in the same way. If a poll predicts 53% of the popular vote in a particular state for candidate X, and the actual vote share is 52%, then the error is 1%. If that poll predicts that the probability of candidate X winning that state is 58% and they don’t win (or if they do), it isn’t obvious how to measure the error. So, perhaps like with NHST it is the binary thinking inherent in probability of winning that is the problem. If we just stick to the accuracy of the vote share prediction, the same problem doesn’t arise.
Its just a bernoulli distribution isn’t it?
See PMF here, thats your likelihood:
https://en.wikipedia.org/wiki/Bernoulli_distribution
Forecast 1: 98% chance clinton wins.
Clinton wins, likelihood = 0.98
Clinton loses, likelihood = 0.02
Forecast 2: 50% chance clinton wins.
Clinton wins, likelihood = 0.5
Clinton loses, likelihood = 0.5
Forecast 3: 55% chance clinton wins.
Clinton wins, likelihood = 0.55
Clinton loses, likelihood = 0.45
Say Clinton does win. Then (priors are all equal so they cancel out):
p(F1|data) = 0.98/(0.98 + 0.5 + 0.55) =
0.482
p(F2|data) = 0.246
p(F3|data) = 0.271
Imagine instead the others forecasted ~90% chance Clinton wins. Then that p(F1|data) will be lower, as it should be.
Anon:
Yes, you can compare relative skills of forecasters using a proper scoring rule. The trouble is that state elections are correlated, and we only have a few national elections, so not so much can be learned from a purely data-based evaluation. Just for example, fivethirtyeight’s forecast in 2020 gave Biden a 6% chance of winning South Dakota. That seems too high! It looks to me like they fivethirtyeight team was trying to be super-careful, and in doing so they ended up with unrealistically wide forecasts for individual states. On the other hand, the election was only held once, and one can imagine a scenario in which Republican party leaders abandoned Trump . . . it’s still hard to imagine Biden winning South Dakota, but who knows? Was there really a 6% chance of that sort of landslide? The fivethirtyeight forecast that year also had notorious problems with predictive dependence between states, so a lot of that 6% was coming from the particularly implausible scenario that Biden somehow crushed it in South Dakota without winning a national landslide.
The point is that this sort of evaluations has to go beyond a simple evaluation of binary outcomes.
“Right now, you could argue that the fivethirtyeight.com model is following the second strategy here, by keeping the two candidates’ probabilities close to even. If polls continue to favor Trump, I’m guessing his forecast will move along and be close to the others, in which case I don’t see that Elliott’s forecasting algorithm should lose reputation points by being later than others to make that shift.”
I don’t agree with this. Models should be judged on all the predictions they make not just the last one before the election. The NHC doesn’t evaluate its hurricane models solely on the last prediction before landfall.
James:
Sure, and that’s consistent with what I wrote. If a model consistently plays it conservative by calling a race 50/50, then its predictions will be calibrated. The forecast can improve its accuracy by moving predictive probabilities toward 0% and 100%, but then it carries the risk of losing calibration.
If we look at forecast vote shares, it’s the same thing. If you’re legitimately concerned that there could be big swings between now and the election, with either party’s candidate having a realistic chance of receiving 55% or more of the two-party vote, then you’d want a big uncertainty about the vote—and thus a predicted probability of close to 50/50—now, with your model only tightening as you get closer to the election.
To connect to your hurricane example: I think that fivethirtyeight’s implicit model is that the hurricane is currently free to move in many different possible directions.
As noted in my comment to Anon above, a difficulty here is that we will never have a large enough sample size of elections to evaluate different predictive models on accuracy from data alone.
tldr; I’m not sure, but I think the 538 model will increasingly put weight on the fundamentals (which favor biden) as polls and fundamentals diverge.
I was trying to figure out exactly why 538’s (Elliot’s) prediction substantively differs from the Economist’s and Nate’s. The polls they incorporate clearly favor Trump: https://projects.fivethirtyeight.com/2024-election-forecast/
So I looked at the methodology page: https://abcnews.go.com/538/538s-2024-presidential-election-forecast-works/story?id=110867585
Like everyone else, 538 fits a fundamentals model, and then does some martingale-type process with the polls. I at first figured that 538’s is still just putting a lot of weight on the fundamentals since we’re so far out from the election. But, I think it makes two defensible modeling choices that are bad when put together:
Choice 1:
“Formally, what we are doing is using our fundamentals-based prediction as an informative prior for what a polls-based model should predict the Election Day vote share to be. ”
Choice 2:
“The second component of polling error is industrywide polling bias, or when surveys from different pollsters miss the outcome in a similar direction.[…] This model tells us that the average polling miss for each party’s vote share in a competitive state is a hair over 2 points, or around 3.8 points on the margin between the candidates. […] We draw potential polling errors for the future from a fat-tailed distribution — specifically a Student’s t distribution with five degrees of freedom[…]”
If I’m reading this right, they assume aggregated polls can be systematically off via a t-distribution with 5 degrees of freedom, and put an informative prior on the polls prediction based on fundamentals (which I’m guessing is normal). Distributional assumptions get you very interesting behavior when there is a conflict between the prior and data. For example, if the fundamentals prior is normal, as polls diverge from them, the model will become increasingly certain the polls are an outlier and reject them in favor of the prior. For example, if there’s a 10 point gap in polls in favor of trump, but the fundamentals still favor biden by 1, I believe the 538 model will mostly reject the polls as being an outlier and conclude it’s 50/50 who’ll win. If they model both as Ts, then you get multimodality, which is a whole other can of worms.
Obviously I don’t have their model in front of me, so I can’t verify this and could be reading things totally wrong. Anyone have thoughts?
Kj:
In theory I agree with you on the multimodality but in practice I think the uncertainty intervals are wide enough, given the possibility of systematic polling errors, that the relevant issue is the weight assigned to each data source, not the tails of the uncertainty distributions.
Something nice in 538’s presentation is the “What do the polls and fundamentals alone say?” section were they show “Polling average” (relatively narrow distribution), “Adjusted polling average”, “Forecast of polling average on Election Day” (similar point estimate but wider distribution), “Fundamentals-only forecast” and “Full forecast”.
The fundamentals pull the forecast towards Biden. In these examples for Nevada and Illinois https://imgur.com/a/laI6IlN the “Forecast of polling average on Election Day” margin of 5.9 for Trump and 8.8 for Biden respectively end at 1.6 doe Trump and 12.8 for Biden (4.3 and 4.0 points improvement for Biden).
For The Economist something similar happens but the shift seems slightly milder: 6 for Trump becomes 3 in Nevada, 4.2 for Biden becomes 8 (3.0 and 3.8 points improvement for Biden).
The Economist gives less details and it’s not even clear what the “Popular vote” shows. It talks about “who is leading the vote” but the line is labeled “Forecast vote intention”. Is it an estimate of what a perfect poll would find? An adjustment to current polls based on the propensity to actually vote? A forecast for the day of the election? The “Likely range” is a likely range of what? It seems too wide to be about current polls but it doesn’t seem consistent with the vote margin projections on top either.
I agree that 538 puts greater emphasis on the fundamentals, but something appears off.
1). Just looking at their national vote forecast, the weight appears to be 75% in favor of the fundamentals. That seems like a lot of weight given the uncertainty intervals (granted they don’t show the polls-only with the systematic bias error). And the final range seems a little tight given 75% is on the fundamentals.
2). Someone pointed out on Twitter that at least the Wisconsin forecast appears to have a bug. Current forecast is: polls only R+2.6, fundamentals D+0.2, final D+0.9. I don’t know que how they get that from their stated methodology, unless it’s all coming from state correlations. I’m not sure how it would work though.
Regarding 2), there is now a remark about that in their explanation about the different components of the forecast:
Full forecast
Our final forecast of the popular vote, based on both polls and fundamentals and accounting for the chance that polls systematically underestimate one candidate. Before Election Day, the final forecast in some states can be more Democratic or Republican than the fundamentals and polls because of patterns of overperformance in similar states.
I’ve been asking this question for years on this blog, but got no answer until now: what’s the point of all this forecasting? Just wait till you know the results. Now I finally know why people forecast: for fun and for ego boost (“I told you so”). I think that’s it, right?
More generally, in all my stats courses, I always tell my students that the best thing you can try to predict is the past, and you’ll be lucky if you can do that.
Shravan:
Please refrain from flippant insults in the comments section. As with social science in general, there are many good reasons to forecast elections besides “fun and ego boost.” For a start, I recommend our 1993 article, Why are American Presidential election campaign polls so variable when votes are so predictable?. The short answer is that, to the extent that a behavior can be modeled and forecasted, that can give us some insight into that behavior. More instrumentally, it makes sense for political practitioners to forecast because it can inform their campaigning decisions, and public forecasts serve a public purpose by somewhat equalizing the playing field.
You might as well ask, “Why study poverty? People are gonna be poor!” or “Why forecast crime? Crime’s gonna happen!”, etc.
Regarding your final comment, of course we are predicting the past. That’s what we do when we fit (or, as is said now, “train”) our models. We predict the past, and we use the performance of those predictions to get a sense of how the predictions will go in the future. Predictions are far from perfect (in your framing, “you’ll be lucky if you can do that”), which is why we have to work so hard to build these models!
This wasn’t meant to be a flippant insult. In fact, in my comment, I ended by asking what I was missing here in my understanding. It’s really odd to accuse me of flippantly insulting people here where I am just writing what I understood from the discussion.
I was just repeating what you (or maybe someone else in the comments, I forgot now) wrote as an explanation for why forecasting is so attractive.
> Forecasting is fun
> Nate should have a greater incentive than the Economist to get things righ
I am genuinely puzzled by what drives so much of the media attention on the *forecasting* of US election results. Maybe it’s good for people who are engaged in betting markets (I assume such things exist but wouldn’t know for sure).
Weather I can understand, and stock prices etc. too. I also find it useful to know whether HW grades can predict final exam performance, because if the answer is yes, then I can tell students how they can do well in exams.
I would see the study of poverty (the causes of ways of ameliorating it) and crime (ditto) useful, but forecasting poverty and crime? I’m not sure what that brings to the table. If I were working on poverty and crime with data, I would be working on trying to understand what is going on now and how to improve the situation using inferential statistics. If someone told me to use current data to forecast poverty or crime, I would tell them they are asking the wrong question.
Shravan:
Ok, just to clarify: I try to do work that is important in some way. I also like to do fun things. Forecasting is fun. Writing Bayesian Data Analysis was fun. Proving my theorem was fun. Also it is an ego boost to figure things out. When I thought about the piranha problem, that was an ego boost because I felt that I’d figured something out that had bothered me for a long time. When I worked out some of the ideas for a zero-avoiding prior for group-level variance parameters, that was an ego boost because, again, this was the solution to an existing problem. Solving problems is fun and is an ego boost.
So, at a psychological level, you could say that fun and ego boost are the reasons I do my work (as you put it in your comment above, “I think that’s it, right?”). But that would be misleading, as I don’t just choose the form of work I do; I also choose the topics I work on. The fun makes the work more pleasant, and the ego boost is a consequence of the progress I make, but this is all in the context of working on projects that I think are (somewhat) important.
I’m not saying that everything I work on is important—I do some things purely for fun and some things purely for the money—but, even there, I gain statistical insight which I think is generalizable and will be valuable to others.
To get back to election forecasting, I again point you to the 1993 paper I linked to in the above comment. I genuinely think the topic is important. I get it that you don’t think it’s important, so my only recommendation to you is to step back a bit and recognize that there are lots of topics that interest some people and not others. I am interested in voting and public opinion, I think political outcomes are important, and understanding how elections can be predicted, as well as cases where they are hard to predict (as discussed here), gives insight into politics. There’s a reason I’m a professor of political science—I’m interested in the study of politics!
Regarding your question, “what drives so much of the media attention on the forecasting of US election results”? I’ve thought a lot about this, and my conclusion goes in two parts:
1. Politics is important. Who wins the election can make a big difference in outcomes—just ask someone who wants to get an abortion. Not everybody cares about politics, but enough people do that it makes sense that the news media will cover it.
2. Two ways of covering politics are to report on candidates political activities and positions—this will give insight into how they might govern if elected—and to report on the “horse race” of who is likely to win. One reason for the predominance of horse race coverage is that the vast majority of people interested in politics have already pretty much made up their minds who to vote for, and they’re interested in who might win. In a polarized electorate, there’s less motivation to cover policies and more motivation to cover the race.
Again, I get it that you are not personally interested in this topic. It’s good that different people can work on different subjects. I hope that your work is fun to you too and also provides ego boost when you solve problems. If so, I would not say that fun and ego boost are the reasons that you do your work. Rather, I’d characterize fun and ego boost as happy side effects of you being fortunate enough to be able to work on projects that you view as interesting and important, at some level.
Andrew
My take on your point #2 is different: as a slight overstatement, I’d say most people don’t care or understand much about policies. Horse races they understand – it is sports. Politics is sports to them and predicting the winner and providing a probability of winning is simple and fits their type 1 thinking. Policies require type 2 thinking – it is hard, takes effort, and there is so much complexity that most people are turned off.
I share Shravan’s distaste for the horse race analysis. Even worse, I think the polling – while clearly important to many people involved in politics – is having adverse impacts on civil discourse and the functioning of democracy. I realize that you don’t share that opinion, but we’ll have to disagree with that. Perhaps a middle ground is that I can see much usefulness in understanding how various groups opinions of policies and candidates are formed and documenting what those opinions are. But I see the effort to produce winning probabilities as actually undermining the functioning of our political system.
https://x.com/gelliottmorris/status/1810860629051461843
Something seems pretty severely wrong with polling industry wide