Princeton Election Consortium

A first draft of electoral history. Since 2004

A change in the air?

September 29th, 2014, 11:58am by Sam Wang


The Princeton Election Consortium’s approach to tracking current conditions has unique advantages, which are sometimes underappreciated and misinterpreted – even by major media figures. The advantages are:

(1) We have remarkably low noise compared with a simpler approach such as at Electoral-vote.com or RealClearPolitics; and

(2) We have sharper time resolution than other sites that use smoothing, whether explicitly or implicitly, to get a more gradual curve over time. Those approaches can’t capture sharp changes as easily. Sites that appear to do this include HuffPollsterThe Upshot, and FiveThirtyEight.

Each site has its own advantages. The unique advantage of PEC’s Meta-Analysis is that it resists outliers – yet also captures real change. In other words, we publish a sensitive and accurate “electoral thermometer.”

In the last 10 days, that thermometer ticked sharply in the Republicans’ direction, past where it’s been at any point this election season. It is not time for Democrats to panic – especially since polls don’t always match final outcomes, even in the home stretch. However, it is time to choose battles carefully. At the end of this piece I will suggest some battles to choose.

CONDITIONS HAVE CHANGED SUBTLY, BUT IMPORTANTLY

You may have heard about the Des Moines Register poll for the Iowa Senate race showing Joni Ernst (R) now leading Bruce Braley (D). By itself it would be just a single data point. But if you look at the last 10 days of polling, nationally, there’s a significant change.

One way to look at it is in the average number of seats when one calculates all the millions of possible outcomes. That looks like this:

As you can see, the average seat count has dropped below the critical red-line threshold of 49.5 seats. However, also note that the gray band of uncertainty still includes outcomes in which Democrats+Independents retain control. So we are within the margin of error for saying who might win an election held today.

A more sensitive measure is the Senate Meta-Margin, defined as the amount by which all races must swing, across the board, to create a perfect tossup:

The Meta-Margin took an unprecedented plunge in the last 10 days. What caused this? In brief, it was multiple polls in Alaska, Colorado, and Iowa.

For example, the Des Moines Register poll is just one of four polls (HuffPollster) that together, show Ernst +6%, Tie, Tie, and Ernst +6%. Considering that the four polls before that were Braley +1%, Braley +4%, Braley +2%, and Tie, the change is unmistakable – a swing of 4.5% toward Ernst. I estimate that the probability that this swing occurred by chance is only 6%. It could swing back, for instance as a result of the debate the candidates had over the weekend. However, it’s undeniable that something really happened in the previous two weeks. The Des Moines Register poll is only the most recent evidence.

Likewise, I have previously written about Alaska and Colorado. Again, those moves favored the GOP candidate.

There is some chance that some of these swings were caused by a tightening of likely-voter screens by pollsters. Some commentators have focused on individual pollster biases. But I think that’s wrong. Such biases exist, but when one uses medians, as we do, those biases can only account for about 1 percentage point of change.

In summary, the last two weeks have seen a swing equivalent to 2 percentage points in across-the-board opinion – and about 1.5 Senate seats. Today, Republicans are finally in the lead in national polls.

TIME TO PANIC – OR GET THE LEAD OUT?

As prone as some Democrats are to panic – and this is a real phenomenon among some commenters here at PEC – it’s hardly time for that. For one thing, polling medians in four races are within 3 percentage points or less (AR, IA, GA, LA; see right sidebar). And that is in a range where polling error, even on Election Eve, is possible.

Let me show you something I’m working on. Here is a table of Democratic outperformance relative to polls, using mostly median-based methods. It comes out as follows.

Election Outperformance (mean±SEM)
2004 Pres. (nat’l) R by 1.4%
2006 Senate D by 2.3±0.7%*
2008 Pres. (nat’l) D by 0.2%
2008 Pres. (state) D by 2.4±0.9%
2008 Senate R by 0.5±1.1%
2010 Senate D by 2.6±1.2%
2012 Pres. (nat’l) D by 2.4%
2012 Pres. (state) D by 2.9%
2012 Senate D by 2.3±0.7%
Average D by 1.5±0.5% (SD 1.6%)

These are median polling errors in state-level races. They were calculated by comparing Election Eve state polls with actual Election Day outcomes. The Presidential findings come from comparing the final Meta-Margin with the margin of victory in the tipping-point state, i.e. the state that pushed the winner just over the electoral-vote threshold for victory. It’s a fairly impressive string of similar biases. The bottom line is that as a profession, pollsters have a small tendency to underestimate Democratic performance, by an average of 1.5%. Obviously this is off-limits to me for use in a calculation – but it is legitimate for you to think about, when interpeting the Meta-margin.

The bottom line: It’s not possible to say who would win an election held today – even with the precision snapshot that I show on this website.

WHAT DOES THIS ALL MEAN FOR THE PEC MODEL?

Since late last week, the Meta-Margin has been better than R+0.2% for the Republicans. Recall that I said our predictive model assumed that the likeliest range for the Meta-Margin was between R+0.2% and D+2.5%. That assumption is currently being violated. To state the obvious, some recalculation is necessary. That’s the bad news, from a modeling standpoint.

The good news is that we’re in the home stretch of the campaign, a time when current polling snapshot is increasingly predictive of Election Eve polls. Inevitably, the PEC calculation will move toward that. We will base our calculations on a weighted average of past conditions, plus an increasing amount of weight to current conditions. The question on my mind is what to use for past conditions. We’ve been using June-to-now, but it might be more accurate to dial that down to September-to-now. I welcome nerdy comments on this topic.

WHAT SHOULD ACTIVISTS DO?

First, Democrats should get off the ledge – and Republicans should hold back on the Champagne (or Gatorade dump). Considering the margins of error in polling, things could still go either way: further gains for Republicans, or a reversion to where the race has been for months, a smal Democratic lead.

One thing that activists should not do is waste efforts in places that are not critical. To pick an example, Kentucky is basically lost to Democrats. It would be pointless to expend effort there. That’s why outside-group ad buys for Alison Grimes have dried up. On the other hand, the most effective places to do get-out-the-vote and to spend money are the top states listed in The Power Of Your Vote in the right sidebar: Alaska, Arkansas, Colorado, Iowa, and Louisiana. Maybe Georgia and North Carolina, though that requires more watching.

*done using the RealClearPolitics averages.

Tags: 2014 Election · Senate

59 Comments so far ↓

  • Jack Kelly

    Sam, I noticed that you do not have New Mexico in the Power of the Vote list. I’d like to know how you see this race.

  • 538 Refugee

    Maybe this discrepancy has been staring us in the face the whole time and we didn’t recognize it. The Supreme Court, on partisan lines (yeah, that isn’t supposed to happen) sided with Ohio republicans to curtail early voting in the state. The move was opposed by democrats saying it would hinder voter opportunity to vote. What mastermind came up with the term “enthusiasm gap”? Maybe the poll bias is a reflection of a closing of the “opportunity to vote” gap that is being bridged by early and absentee voting? This one has been on the tip of my tongue but didn’t find articulation until I read today’s ruling and put it together with the polling discrepancies. .

  • 538 Refugee

    The midterms are getting increased scrutiny because the polls have indicated a close election in terms of senate control. What this may be telling us is we need to look a little deeper at the pollster’s numbers in these instances, especially when polling gets sparse. Did the pollster’s assumptions change? Harry Reid said that 6 weeks is the time that people get serious about their choice and voters bias towards their own personal interests. We have any poll junkies that look at crosstabs amongst us?

  • Canadian fan

    RB. Thank you again. As Sam has shown, Democrats have consistently out-performed polling turnout expectations with the exception of the 2004 race ( that includes 2010, as you point out – in a truly wave election ). That in itself tells you something – that this is either something that is not routinely picked up by the likely voter projections ( that you vehemently maintain it does ) – or that they are simply the result of highly intense Democratic outreach efforts, registering new voters, etc. – a highly likely prognostication, and one that is likely to be continually repeated. From the Republicans point of view the Democrats achieved the impossible in 2012 – improved their 2008 numbers – something Republicans strenuously and emphatically said could not be done – by using a scientific outreach effort unprecedented in scope in electoral history. Also, in U.S. presidential elections, never has a victory of such magnitude ever occurred within a five week period prior to election day after a debate precipitated a drastic and precipitous drop in the polls – only to turn it around three weeks later into a 332 sweep of the electoral college. Hurricane Sandy was undoubtedly a factor ( as was Obama’s exemplary handling of it ), but to produce such a stunning result there had to have been a ground game at the grassroots level, as well. There was. And the results were clear. In CNN’s Louisiana poll of two days ago, a run-off between Landrieu and Cassidy would result in a 50 – 47 Cassidy win among likely voters. Among registered voters it is a 51 – 45 Landrieu win. This is why the Bannock Street Project is in operation. These voters and their voting intentions already exist. Getting them to the polls is what Democrats are focusing on. 2012 ought to have been a wake-up call for Republicans. Yes, it was a presidential year, and this is a midterm election, but the reach of this grassroots project is truly remarkable, and is unprecedented for a midterm election. It is backed by 60 million dollars and over a year’s preparation. I reiterate – people are going to be surprised at the strength of the numbers. If the pattern established in the last eight years holds, the Democrats will out-perform expectations yet again.

    • Insidious Pall

      2008 and 2012 may have been aberrations in GOTV terms. While the ground strategy in those campaigns have become legend, we may want to see a couple of additional cycles. In ’08 and ’12, there was a highly motivated African American vote, based on Obama’s candidacy. In many places, A2 turnout exceeded turnout among white voters for the first time. Whether that translates into a political realignment is anyone’s guess. I remain a tad skeptical.

  • Steve Scarborough

    Hi Sam. First, I am a huge fan who is grateful that you took some time to respond to my questions. FYI, I am a long-in-the-tooth statistician who has been around the block so to speark. Here are some thoughts from an old guy:

    1. You and your colleagues are doing a great job! Stay the course.
    2. I believe that your use of the median and the SEM is on target.
    3. The use of the polynomial to aid in generating the coefficients/probabilities is nothing short of brilliant. (…And thanks for answering my dumb questions here.)
    4. I love what you are doing with the meta margin. However, it appears that some — most likely those who have failed to read your methods carefully — are often confusing the snapshot with the prediction. Hence, I suggest at some time, re-evaluating how you explain that.
    5. As one who forecasted Medicaid caseloads in a large state for many years, I can identify with the snapshot/prediction issue. I think I got in more trouble with that than anything else in my work.
    6. In my work, I often used time series methods, including the Box Jenkins ARIMA models, exponential smoothing, and state space models to forecast monthly caseloads out to as far as 36 months in the future. I suggest you take a look at the feasibility of trying out some of these approaches on your models, such as the meta margin, and see if they help. Suggest a “forecasting tournament” with your historical data to see how various methods fare.
    7. For my last comment, I offer the following 4 fundamental tenets of forecasting as a bit of brevity: 1) forecasting is difficult, especially if you are forecasting the future!; 2) the problem with forecasting is that you know you are going to be wrong, just that you do not know when and by how much!; 3) he who lives by the crystal ball soon learns to eat ground glass!; and, finally, 4)if you ever get one right, DO NOT EVER LET ‘EM FORGET IT!

    Be well,

    Steve

  • Canadian fan

    RB. Appreciate your response. You mention cash, but that is precisely what the Republicans are now in want of. They are strapped for cash, and most of the reserved TV ad spots for October have been already booked by Democrats, having purchased them months ago, thereby locking in flat rates, that in some cases have now risen 500 %. Karl Rove has recently voiced a panic about this. And the reason why he has is that controlling the airwaves means controlling the narrative. I would differ regarding the reference to an enthusiasm gap. It seems to be that if the Democrats are dominating the advance vote – as they are – by up to a two to one ratio – that speaks volumes as to who is truly motivated.

    • RB

      Canadian Fan-As far as the cash you are correct, but remember the states that are in play(AR-LA-AK-IA)…spending is cheap there. CO may be expensive, but Cory Gardner is probably the best candidate running for either party this cycle in a competitive state. The money has hurt in NC and if NH is truly up for grabs that is a bad omen for the Ds. Yes-Democrats lead big in early voting, but that is a norm(even in 2010). The LV screens would defiantly be picking up on those voters. You really have to go to pollsters delving into the enthusiasm gap(I think PEW is one of them). Remember the LV screens did pick up in a big Obama win in 08, a big GOP win in 10, and an Obama win in 12(the win may have been underestimated in the final polls due to the public digesting Hurricane Sandy)….either way we will see and still a lot of time left for events to swing this election one way or another, but for now I would put the Senate as Lean R…and the House as +5/8 seats…no math formulas here(hey I admit it) just a hunch based on the individual state polling, generic ballet question, and Obama’s JA in the low 40s.

  • A New Jersey Farmer

    The ride is getting bumpier. The meta-margin is back down from yesterday. Has the GOP trend stalled?

    • Sam Wang

      Actually, I think that’s because the fourth-oldest Iowa poll, which showed Ernst+6%, dropped out of the time window.

    • Froggy

      One new poll in Iowa showed the race there as tied, and that alone caused the meta-margin to move nearly 1%. That just goes to show you how close the battle for the Senate is. (A single poll, for example one showing Pryor leading in Arkansas, could be enough at this point to tip the meta-margin back to the Democrats.)

  • Canadian fan

    I find it amusing that some are so quick to dismiss the present and continuing discrepancy between measured registered voter preferences as opposed to what pollsters now confidently identify as likely voters. Some commentators are betting the family farm on the likely voter projections. But the fact is that both Pryor and Landrieu continue to do significantly better among registered voters. In fact, in a recent poll, Landrieu trailed with ” likely voters ” ( as has been the case in most recently released polls ) but also continues to find her leading in registered voters – by as much as six points. Just because polling agencies have now ceased to focus on registered voters doesn’t mean that registered voters have suddenly ceased to exist. The Bannock Street Project is predicated on a very firm and proven assumption, and that is that when all the voters turn up, the Democrats do better. The Republicans are banking on the midterm drop-off among Democratic voters to take them to a win. No surprise there. They may be right. But they could also be wrong. As most polls concentrate on what they call ” likely voters “, the subterranean reality of registered voters has gone off the radar of public awareness. I think consequently many are going to be surprised at the Bannock Street Project effort by election day and in the two run-off elections that follow. Already, the Democrats are holding huge advantages – in one state by nearly a two to one ratio – in mail-in ballots. This is significant. The Bannock Street Project is unprecedented in scope, and it is based on the knowledge and wherewithal of the two recent presidential elections – both of which greatly surprised Republicans, particularly the most recent one, where they truly believed that Obama couldn’t possible match – let along exceed 2008 levels – particularly five weeks after a debate, and a dismal dive in the polls – they were certain to dampen voter turnout and voter preferences. Well, a funny thing happened on the way to the forum.

    • RB

      Canadian Fan-a couple of points—Voters who have been outreached and early voted would no doubt be captured in the LV models. Similarly if the GOTV targeted at Ds was as effective as you imply it would also be picked up on LV screens. The Republicans have a big advantage in the ‘enthusiasm gap’, which again points to where LV will go. Ask yourself why is no polling firm using RV models-most likely because LV models are picking up the actual electorate which we know will be 40-45% of RV. If the Democratic GOTV we’re truly making a difference at this time it would be picked up in LV models and the enthusiasm gap would shrink. Of course we are all entitled to our opinions, but if Marist, Quin, CNN, FOX, Selzer,and other quality pollsters who use live interview with cell phone outreach thought the RV sample would be most representative of Election Day, they would go with it. And as Democrats and Republicans are want to forget the Other side is not just standing idle. I can tell you in the last month I have had every Republican asking for money, which is new-but what the Dems had been doing.

  • securecare

    “…However, it’s undeniable that something really happened in the previous two weeks….”

    Yes and Josh Marshall is, I think, close to an answer as to what has likely gone on.

    http://talkingpointsmemo.com/edblog/the-return-of-terror-politics

  • NP

    “Recall that I said our predictive model assumed that the likeliest range for the Meta-Margin was between R+0.2% and D+2.5%. That assumption is currently being violated.”

    You have really lost me here. Why isn’t the meta-margin where the race stands? Why do there have to be limits?

  • Dean

    I think that the last thing Democrats and liberals should do is unskew polls. The right wingers who did that in 2012 were the laughingstock of the polling community.

    We should use polls and be honest with them. If we lose, we lose.

    Some key Senate races are very close and could be won by Democrats. It’s entirely possible for a key race or two to be picked off by Democrats.

    Am I surprised that things are trending Republican? No, because what other first-world country fights against universal health insurance? Are there voters now in Germany or the U.K. trying to repeal universal healthcare?

    In Kentucky, a key battleground state, some voters love their new Obamacare health insurance but hate Obamacare. It’s crazy but true.

    Where else is organized labor demonized like it is in the U.S.? Our unionization rate has been dropping and is at its lowest percentage in decades.

    Where else do so many voters fight for the super-wealthy to get tax cuts? We have perhaps the worst income inequality of any advanced nation, and the CEO-to-worker pay ratio is off the charts.

    • Jena

      Which is why I admit to being very confused & disgusted right now. If I could afford to, I’d relocate to the least political, most neutral country to avoid a government dominated by Koch bought flunkies. Dems aren’t perfect by any means. But live in a place where the ideology is regressing us back to the Jim Crow era? It’s hard to accept the 2 years of obstruction hell if Republicans take the Senate.

  • JayBoy2k

    I am new to the site and tend to lean right, but I’m with the majority opinion — stay the course and follow the originally defined PEC process. I have no confidence on GOP taking the senate or whether in 5 weeks we are within a +2%D and -.02%R range. I do not believe that we could/should base change on Kansas or Alaska polls and both Iowa and Colorado might be back in the Dem fold any time in the next 6 weeks.

    • RB

      Jay-I am kind of with you but his model admittedly goes back to the summer months which include and give a fair amount of weight to RV polls. Look at CNN LA RV vsLV sample. In good Democratic Presidential years it is not a problem, but in good GOP midterm cycles it can lead to what is happening here. But perhaps he should stick with the model and if it is wrong-The RV/LV-disparity is where I might start(one more example of this-remember those 10 RV Pryor leads….pffff out came the LV models and we now have Cotten leads)

  • bks

    No changes to the topline model unless it is to correct an error in the implementation. There will be another election in two years. –bks

  • Joseph

    “The question on my mind is what to use for past conditions. We’ve been using June-to-now, but it might be more accurate to dial that down to September-to-now. I welcome nerdy comments on this topic.”

    I tend to agree that you need to “dance with the one you came with”, so if you have to pick one I’d stay with the June version.

    OTOH, I also see value in a September-on version. One can definitely argue that it wasn’t until just recently that many people actually started thinking about the election.

    So why not do both, if it’s not too much work?

  • Craig W. Barber

    Responding to Mediaglyphic: My understanding is that tCDF is the t-distribution cumulative distribution function for the data input. You’ll recall that Dr. Wang is using the t-distribution and that the choice of function, one with a broad shallow distribution versus a high spike, is part of the recent debate with Mr. Silver.

    See Dr. Wang’s discussion at: http://election.princeton.edu/2014/09/09/senate-2014-the-election-day-model-nerd-version/

    Mr. Wiener: Perhaps Mr. Koczela had already covered your comments: “Dr. Wang: Look, no matter what you do, you’re gonna catch flak.”.

  • Daniel Wiener

    The whole point of your model has to be its predictive value, otherwise nobody would care. If your model must be readjusted every time there’s a major (unexpected by your model) shift in the polls, then we might as well just watch the poll averages or else relax and wait until election day for the actual results.

    So now you have a situation in which your Meta-Margin calculation, which has consistently been predicting a Democratic-controlled Senate for most of the past few months, has suddenly and sharply plunged into Republican-controlled territory. That in itself is at least a partial fail, since it’s outside of the range of predicted variability. Maybe it’s a temporary aberration, soon to be reversed, and the Democrats will indeed hold the Senate. In that case your model will be (mostly) validated.

    But if the trend continues, and Republicans indeed take the Senate, that falsifies the model. If Republicans not only take the Senate but happen to win big in November (for example, gaining eight seats or more) that pretty much throws your model into the trash can. Sorry to be so blunt.

    Many traditional political pundits have long been forecasting a strong election or even a wave election for Republicans this year, based on a variety of fundamentals (e.g., Obama’s sinking job ratings, dissatisfaction with Obamacare, a still-struggling economy, the 6th-year effect, etc.). Perhaps, unlike in 2012, they are right this time and your model is wrong. But your underlying rationale is that you can beat the traditional experts at their own game using mathematics. We’ll know who’s right in five weeks.

    The one thing worse than having a failed model which incorrectly predicts the outcome is for you to try to adjust the model on the fly. Anyone can massage a model using hindsight, but predictions are only useful when they are prospective. So my recommendation is to leave it as is, and accept the consequences, good or bad. Now is not the time to say “Here’s alternate version 2, with this adjustment, and this different prediction.” Having more than one version, so you can later claim that “Version 1 stank, but if you’d have followed version 2 you’d have won big” is worthless CYA for any honest prognosticator.

    • axt113

      I disagree, this isn’t about whose model is better, but about building a model that is able to always predict an election precisely.

      I think it’s better for Professor Wang to try and find the model that works.

      Being willing to examine has model for mistakes shows an open mind.

    • Insidious Pall

      Dan, a couple of things before we go looking for a trash bin. One, a polls-only model will usually be more susceptible to sudden shifts in the polls. Second, the meta-margin is mostly in line with other aggregators. If Repubs pick up 8 or more seats as you say, everyone will have been wrong; we will need plenty of trash bins.

  • Lojo

    Man…. hard news to take but glad the PEC is sticking to its guns and reporting it like it (well the model) sees ‘em.

  • Randy Haugen

    In the Iowa race the only thing that really happened except for last nights debate was that the negative ads by Koch have doubled,The Debate was Good for Bruce but the gender gap looks to be large with men voting for Ernst and women at a lesser amount 15% to Braley.I am in Iowa and its just bizzare but remember we have Steve King here who always gets to keep his job no matter what he says.

  • Joe in Seattle

    The premise of this post is accurate. Democrats like me were just getting comfortable at the start of the month with the idea that we’d hold at least a 50-50 majority. Then we wake up to a nightmare: the bottom has fallen out in most of the competitive races.
    My current 2014 Senate election prediction: 52 Republicans, 48 Democrats.

    The news is better in the gubernatorial races. And as for 2016, the fallout from two years of a tea party congress will be around 11 Senate seats Democrats can target for pickup, and possibly enough of a blue tidal wave to take the House.

    • Sean

      Problem, you are missing Joe, is that the Dems will have to run in a heavy gerrymandered districts that were formed by governors and state parties back after the 2010 census. The Dems won’t have a chance to fix these districts until 2020 when the new census numbers come in.

      The key now for the Dems is to groom / elect people for the 2020 elections in state houses etc.

      Get ready for the redistricting and hope Ginsberg can last until 2016 when the Dems have the potential to take back the Senate.

  • Chris

    How about using several months back but with a different weight applied to each one, with decreasing weight for each month back. So,

    TotalScore = w_1 * MonthScore[current] + w_2 * MonthScore[current-1] + w_3 * MonthScore[current-2]

    where w_1 > w_2 > w_3 (and I’m guessing w_1+w_2+w_3 = 1, although not sure if necessary in your scoring model).

    you could also apply that on a weekly basis instead, if you think the data is fine-grained enough.

  • DB

    I was wondering if there is a pattern to this swoon and if it is possible that these polls were in the field during a time period that regularly produces results weighted towards conservative candidates.

    It would seem that data we’re getting now was in the field during early – to – mid-September. That would be right around the time that people’s lives are most disrupted by returning from summer holidays, the return of schools and – as pointed out elsewhere – LV models tighten on polls.

    Is it possible that the disruption caused by the transition from summer to fall life patterns disproportionately impacts D voters? Is there always a September polling swoon in polling data among those whose lives are impacted most by the end of August “recess” and the start of September School Bells?

    Thanks.

  • G Washington

    I think that the random walk, as implemented, is perhaps too simplistic. There are two important aspects that are not fully implemented:

    1. Correlations between races. The current model does have some covariance naively implemented, but not in the sense that the covariance is different between different races. Perhaps there is some local issue that affects CO+IA, but not the national landscape.

    2. If you ignore covariances, the random walk is somewhat simplistic since in reality there are many random walks (one for each race), which contributes to the overall picture. Even if the randomness is Gaussian for any given race, since the width of the Gaussian will be different for all races, the combined distribution will be off from Gaussian (likely with fat tails). And it may just be that a few of the races with wider distributions are out 1-2 sigma at the same time.

    The covariances is easy to understand if you want to look at the full data set, although the uncertainties will be large if you examine only the current year’s data. Similarly, the random walks from individual races will be hard to determine because of a lack of data. But could be modeled if you know the underlying function.

    As a side note, I also would not be surprised if the true variability of the race is suppressed somewhat at earlier times because of the sparser polling data. If the true variability timesecale is shorting that the polling timescale, you are effectively smoothing things.

  • Amitabh Lath

    Is the random-walk hypothesis really not viable now? What is the p-value on this fluctuation being random?

    • Art Brown

      If I’ve got Prof. Wang’s prediction model right, the fluctuation to a meta-margin of 1.42% Republican represents a 2-sigma event, which I think has a probability of 15% with the prediction’s 1 degree-of-freedom t-distribution and so is by no means impossible. The key will be whether it goes away. (I see the meta-margin is now back to 0.5% R.) However, Prog. Wang appears to be reacting to this event, so I’m probably missing something…

  • Bill

    In response to Sam’s response to my comment: Yes, I thought about this being a form of unskewing. And it would be! However, if well justified, it would arguably be a form of removing systematic bias. As I have thought about it further, maybe the way to go about this would be to weight the outperformance values with some weighting variable that represents voter turnout.

    In any event, my humble suggestion would be to leave things as they are with your projections. If you are to do anything with a scheme to remove what may be systematic error, put it on the site as a hypothesis test and see how things play out.

    Slightly off topic, if the Rs take over the Senate the next 2 years will likely be even uglier than they have been the past few.

  • Arkansas Yankee

    RB
    I am sure I will be viewed as right wing nut. I am actually a moderately coservative GOPer. I would like to point out that NH and Wisconsin won’t be as easy to be picked off in 2016 as some may think. It is not only Democrat incumbents who have staying power.

    As far as this year is concerned, I believe the tide has begun to roll for the GOP. Except for Kansas we have good candidates in the competitive states. We have an incumbent Senator in Arkansas who does not want to debate. Our candidates have done well in debates in Iowa and Colorado. I am hopefully anticipating 11-4. It is a nice thing to see Dems arguing with polls. I will sit, read, and enjoy the arguments.

    • Davey

      I’m not sure I concur with the hypothesis of GOP momentum extending into future elections. I agree that their candidates have performed better this year than 2012, but there’s not a lot of data to suggest that the party platform is resonating. And it’s a concern for the party that the gains this year will be modest when they’ll face the worst defense election of all time in 2016. My recommendation would be to spend the next two years repairing the party’s relationship with women and minorities so that they don’t lose both the House and Senate in 2016.

  • Alan Koczela

    Dr. Wang,

    Just do what is most consistent with the underlying thesis of your model — no special sauce or whatever. Let’s see how this puppy plays out. If there are multiple ways options, flip a coin. If it misses the mark, so be it. We can learn from its failure, just as we learn from its successes. There will be enough time to adjust the model, if needed, after this election cycle.

    Look, no matter what you do, you’re gonna catch flak.

    Keep up the good work.

  • Jay

    I am puzzled by a consistent pattern in many states this cycle where one poll comes out and appears to have the race settled and then another poll comes out right after that and contradicts the other poll. Do you believe variance is much higher in specific contests this cycle than it was in 2010 or 2006? And if so why?

  • Craig Barber

    So if I’m understanding this correctly (another black swan event if it ever happens!) then the expression:

    tcdf((+0.3+0.47)/0.92,1)=0.722

    is what is now being recalculated?

    I can see how the choice of time frame (June to now vs. Sept. to now) could be extremely important.

    How do we avoid “special sauce” in that choice of time frame? Mind you, special sauce tastes good too. But PEC has been all about following the polls and the math. What to do when the realistic poll time frame changes?

    A poll in September is closer to the finish line than one in June. Weight it? Invent an arbitrary rule? Go through reams of historical data and try to derive a function (people are doing that) then incorporate THAT uncertainty factor as well??

    Take fifteen aspirin and call me in the morning?! ;)

  • Stuart

    Democrats, being poorer than Republicans, are not as accessible to pollsters as Republicans. So, I think Republicans are overrepresented in polls, hence the “outperformance” for Democrats.

  • Bill

    It might be better to compute the possible bias using data only from the midterm elections. I believe that turn out is higher in presidential election years, so including those years in a bias estimate might lead to somewhat misleading results. Doing so shows, interestingly, less variability: in 2006, D +2.3, and in 2010, D +2.6. Of course, a major drawback of this method is that there are only two data points.

    • Sam Wang

      That’s probably right. However, that’s getting dangerously close to unskewing!

      At a minimum, I want to get this all out on the table so we can have a discussion. Today I buried the information deep in the piece, to attract readers like you. On a later day, we can have the evidence up top.

  • RB

    Sam-I think the starting point should be September with LV models…RV to LV is apples to oranges. Look at this weekends CNN polls as an anecdotal example.

    2016—The GOP will have 5 tough seats to defend-PA-OH-WI-IL-NH. I think Portman will be fine, but the other 4 will likely be pulled by the Presidential contest. Democrats seem fixed on the belief that Hillary will win—maybe but there is a lot of time and events that will shape that race that none of us know for now. If a Rupublican wins in 2016(I know a may be the only right of center poster here-which is why I think it is possible for them to win) then the Senate seats mentioned above will be tricky for the Ds…just my two cents.

    • Ditto

      It is not about clinton

      The electoral college and demographics are swinging hard against the odds of any Republucan winning or any Democrat losing

      If you look at the wp the GOP needs a perfect storm in the states it can win, but several of those states are moving away from the GOP demographically

    • Froggy

      Don’t forget Iowa and Florida for 2016. Grassley will be 83 in 2016, and if he retires the Democrats would face a good chance at winning an open seat in a presidential year. And Rubio didn’t get a majority in the three-way race in 2010, and he could face a primary challenge on the right from Allen West.

    • RB

      Fair points about IA and FL. A lot depends on what Grassley will do…is the Gov interested (doubt it, but ya never know). As long as Rubio is running in FL, I think he wins. Dems don’t seem to have a great bench there and Rubio is a dynamo on the stump.

    • Davey

      I also believe Republicans have a chance in 2016 for the Presidency, but the map makes it look like a tough, uphill battle. Unless conditions shift dramatically, we would put 237 electoral votes in the Dem column and 191 for the Reps. Of the remaining nine states, Obama won eight. And of those we’ve “locked away,” there’s one or two states Republicans might target, compared to some significant red states verging on becoming swing states. It seems like for now the electoral college is swinging blue in much the same way it swung red for Reagan.

  • Ditto

    I’m actually not panicked for three reasons:

    1. I expected Democrats to lose bigger. What I’m seeing is a pox on both your houses in the races rather than a wave. That still bodes badly for Democrats because of the number of races Democrats must defend. However, there are some surprising chances for Democrats to win.

    2. Even if they do win, the GOP will only hold the Senate for two years. 2016 looks far worse for them than 2014 does for the Democrats.

    3. I’m unconvinced 2014 will look like 2010. E.g.,:
    http://m.huffpost.com/us/entry/5891808?utm_hp_ref=tw

    • Matthew

      Re #3 Basically you are taking the “Rovian” approach to polls. Rove and other Republicans thought the “likely voter” models in the 2012 polls were off so they tried to “unskew” the polls and that spunds exactly like what you are doing in regards to #3.

    • Insidious Pall

      So in reading your article, HuffPo points out that there were 350,000 mail votes in IA in 2010 with 145,000 requests thus far in 2014. Two-thirds of these are Dems. Requests at 8,000 per day, even at that rate up until one day before the election (not likely), the total requests for mail ballots would be 425,000. More than 80% of those would have to be turned in to surpass 2010. Ready for Joni?

    • Ditto

      Except I’m not taking their approach either in the outcome (I assume Dems will lose, but explain why ) or in the data (I point out data that can’t be refuted by your saying it “skews” the polls). As between the polls and actual turn out , polls lose. Early ballots are turnout data. Not polling models.

    • Davey

      I can get behind the hypothesis without slamming you like other commenters, lol. In a year where voters are perturbed and Congress has a historically low approval rate, defending more seats is undesirable. Seems like solid reasoning. And defines a strategy for 2016 – if voters don’t approve of Congress, Dems might see the most success in attributing the problems with Republicans.

  • Philip Diehl

    Dems, time to give ’til it hurts to the key races Sam identifies: https://secure.actblue.com/contribute/page/pec2014

    • RB

      Arkansas Yankee-yup–you and I are probably only two conservatives here. As far as NH 2016 you may very well be right. WI will require a GOP win or close loss at POTUS (just my gut feeling) …PA as a D+1 state should be close either way. IL-tough, but the Dems have been shooting themselves in the foot there.

      As far as the model Sam uses I really think it boils down to using only LV polls from September on and applying more weight to the most recent polls. Finally —Do NOT use internals! Ask yourself why is it the Dems are only releasing internals for the most part. The answer is self evident and instructive. Does anyone really beloved the 42/42 internal Braley released immediately after the DMR released theirs.

      Orman—The race is very fluid and I do have an opinion where it will go, but that is pure speculation. Roberts went up hard on the air against Orman last week. Previously Orman was the only one on the air. Polling in the 2nd week of October will give us a better sense of where this is.