Month: September 2012 (Page 1 of 2)

There’s been a lot of talk lately about how the presidential polls might be biased. So let’s look at how well – or poorly – some of the major survey firms are actually performing this year.

All polls contain error, mainly from the limitations of random sampling. But there are lots of other ways that error can creep into surveys. Pollsters who truly care about getting the right answer go through great pains to minimize these non-sampling errors, but sometimes systematic biases – or house effects – can remain. For whatever reason, some pollsters are consistently too favorable (or not favorable enough) to certain parties or candidates.

Since May 1, there have been over 400 state polls, conducted by more than 100 different survey organizations. However, a much smaller number of firms have been responsible for a majority of the polls: Rasmussen, PPP, YouGov, Quinnipiac, Purple Strategies, We Ask America, SurveyUSA, and Marist.

For each poll released by these firms, I’ll calculate the survey error as the difference between their reported level of support for Obama over Romney, and my model’s estimate of the “true” level of support on the day and state of the poll. Then each firm’s house effect is simply the average of these differences. (Note that my model doesn’t preemptively try to adjust for house effects in any way.) If a firm is unbiased, its average error should be zero. Positive house effects are more pro-Obama; negative house effects are more pro-Romney. Here’s what we find.

There are a number of pieces of information to take away from this table. First, none of the house effects are all that big. Average deviations are less than 1% in either direction. This is much smaller than the error we observe in the polls due to random sampling alone.

Second, even if, say, Rasmussen is getting the right numbers on average – so that PPP’s house effect is actually +1.6% – then that +1.6% bias still isn’t that big. It’s certainly not enough to explain Obama’s large – and increasing – lead in the polls. Of course, it’s possible that even Rasmussen is biased pro-Obama, and we just aren’t able to tell. But I don’t believe anyone is suggesting that.

Finally, the firms with the largest house effects in both directions – PPP and Rasmussen – are also the ones doing the most polls, so their effects cancel out. Just another reason to feel comfortable trusting the polling averages.

Here’s a plot highlighting each of the eight firms’ survey errors versus sample sizes. The horizontal lines denote the house effects. Dashed lines indicate theoretical 95% margins of error, assuming perfect random sampling. Again, nothing very extraordinary. We would expect PPP and Rasmussen to “miss” once or twice, simply because of how many polls they’re fielding.

Just out of curiosity (and no particular feelings of cruelty, I swear), which polls have been the most misleading – or let’s say, unluckiest – of the campaign so far? Rather than look at the raw survey error, which is expected to be larger in small samples, I’ll calculate the p-value for each poll, assuming my model represents the truth. This tells us the probability of getting a survey result with the observed level of error (or greater), at a given sample size, due to chance alone. Smaller p-values reveal more anomalous polls.

Here are all surveys with a p-value less than 0.01 – meaning we’d expect to see these results in fewer than 1 out of every 100 surveys, if the polling firm is conducting the survey in a proper and unbiased manner.

The single most… unusual survey was the 9/16 Suffolk poll in Massachusetts that overestimated Obama’s level of support by 7%. However, of the nine polls on the list, seven erred in the direction of Romney – not Obama. And what to say about Gravis Marketing, who appears twice – strongly favoring Romney – despite only conducting 10 polls. Hm.

It’s interesting that many of these surveys had relatively large sample sizes. The result is that errors of only 3%-4% appear more suspicious than if the sample had been smaller. It’s sort of a double whammy: firms pay to conduct more interviews, but all they accomplish by reducing their sampling error is to provide sharper focus on the magnitude of their non-sampling error. They’d be better off sticking to samples of 500, where systematic errors wouldn’t be as apparent.

There’s a general consensus among poll-watchers that Obama is currently ahead in most – if not all – of the battleground states. How likely is this lead to hold up through Election Day? And what range of outcomes are realistic to expect? Let’s set aside the forecasts being produced by my model (which combine the polls with certain assumptions about long-term election fundamentals), and instead just walk through a few different scenarios starting from where preferences stand today.

I start by accounting for uncertainty in the current level of support for Obama and Romney in each state. The idea is simply that we have better estimates of public opinion in states that have been polled more frequently. From the model results, I simulate 10,000 plausible “states of the race” for all 50 states.

Next, we have to make some guesses about how voter preferences might change between now and Election Day. So far, state-level opinion has been fairly stable; only varying within a 1%-2.5% range. Since Obama is ahead right now, the less we believe preferences are going to fluctuate over the next six weeks, the worse for Romney. So let’s generously assume that with 95% probability, voters might swing as much as 4% in either direction from their current spot, with a modal change of zero.

Here’s what the combination of these two assumptions would look like in Florida. (Recall all percentages exclude third-party candidates and undecideds.) There’s some initial variation around today’s estimate (the square); then the potential future changes are added in. The result is a 77% chance of Obama winning Florida – that is, 77% of the 10,000 simulations result in an Obama two-party vote share above 50%.

Finally, to extend the simulation to all 50 states, we have to consider that future changes in state opinion are not likely to be independent. In other words, if Romney starts doing better in Florida, he’s probably going to improve in North Carolina, Virginia, Ohio, etc. as well. So we want to build in some correlation between the state trends. Perfect correlation would be equivalent to “uniform swing” in which a constant amount is added to (or subtracted from) each state’s current estimate. The lower the correlation, the more the future state trends differ from one another.

Let’s try a moderate level of inter-state correlation: 0.8 on the range from 0 to 1. I generate 10,000 hypothetical election outcomes in all 50 states, and add up the number of electoral votes Obama receives in each. The result is a 92% chance of victory for Obama, with a median outcome of 347 electoral votes. This would be Obama winning all of his 2008 states, except for Indiana.

If we increase the correlation between states all the way to 1, Romney’s chances of winning are still just 10%. What’s going on? Obama’s lead in the polls is so large right now, that he could lose 2.5% of the vote in every single state and still have enough support to clear 270 electoral votes. The chances of him slipping more than that, if current trends continue, are slim.

One possibility is that the polls are all biased in Obama’s favor, and have been systematically overstating his level of support. Suppose we knock 1% off the model’s current estimates in each state and re-run the simulation, assuming perfect uniform swing. In that case, Romney’s chances improve to 20%.

If the polls are all biased 2% in Obama’s favor, the simulation moves Romney up to a 37% chance of winning – still not great, but at least better than 8%.

No wonder the Republicans are starting to challenge the polls. Unfortunately, there’s no serious indication that the polls are behaving strangely this year.

With the pace of polling increasing, there are going to be days when some polls seem to be especially surprising – or even contradictory. For example, a recent Washington Post survey found Obama up 8 points in Virginia, even though other polls indicate a tighter race. It’s pretty safe to say that Obama is not actually winning Virginia by 8 points. But this doesn’t mean the Post poll is biased, or wrong, or should be ignored. I imagine the Post did the best job they could. The likeliest explanation for the finding is simply random sampling error.

Even in a perfectly executed survey, there’s going to be error due to random sampling. A survey only contacts a small group of respondents, and those people won’t always be representative of the broader population. The smaller the sample, the larger the sample-to-sample variability. To see just how large sampling error can be, suppose my model is correct that Obama is currently preferred by 52% of decided, major-party voters in Virginia. Then in different surveys of 750 respondents (which is about the average size of the state polls), it wouldn’t be unusual to see results ranging anywhere from 48% to 56%, because of sampling variation alone. In fact, here’s the expected distribution of all poll results under this scenario: most should be right around 52%, but many won’t.

If we added in other possible sources of survey error (question wording, interviewer effects, mode effects, sample selection, and so forth), the distribution would become even wider. So just imagine two polls on the same day showing Romney with either 52% or 60% of the two-party vote. Astounding, right? No, not really. It happened in Missouri last week.

What is actually astounding about the polls this year is how well they are behaving, compared to theoretical expectations. For a given sample size, the margin of error tells us how many polls should fall within a certain range of the true population value. I’ll assume my model is correctly estimating the current level of preference for Obama over Romney in each state during the campaign. Then I can subtract from each observed poll result the model estimate on that day. This is the survey error. It turns out that most polls have been exactly where they should be – within two or three points of the model estimates. And that’s without any correction in my model for “house effects,” or systematic biases in the results of particular polling organizations.

Plotting each poll’s error versus its sample size (excluding undecideds) produces the following graph. The dashed lines correspond to a theoretical 95% margin of error at each sample size, assuming that error arises only from random sampling.

If the model is fitting properly, and if there are no other sources of error in the polls, then 95% of polls should fall within the dashed lines. The observed proportion is 94%. Certainly some polls are especially misleading – the worst outlier, in the lower right corner, is the large 9/9 Gravis Marketing poll that had Romney ahead in Virginia (and was singly responsible for the brief downward blip in the Virginia forecast last week). But what is most important – and what helps us trust the pollsters as well as their polls – is that the overall distribution of survey errors is very close to what we would expect if pollsters were conducting their surveys in a careful and consistent way.

More than 30 state polls have been released since the end of the DNC, and they point to an unmistakable U-turn in the trend in support for Obama over Romney. It’s the most consistently pro-Obama swing that we’ve seen at the state level all campaign, and it has been large enough to wipe out most of Obama’s losses in the polls since early August. That Obama received any bounce at all from the DNC, after Romney got nothing from the RNC, also suggests that voter preferences may not be so fixed, after all.

But while the momentum has shifted in Obama’s favor, the bounce itself is still fairly small by historical standards. My model – based only on the state polls – suggests that the overall effect of the DNC has been about 0.5% in a typical state. The size of this swing is consistent with other trends in opinion observed so far this year. If nothing else, it has been enough to keep Obama in the lead in the closest battleground states. The question now is whether the trend has peaked, and how much of the bounce will persist through Election Day.

Curiously, the post-convention bump has appeared to be larger in the national polls than it has in the individual state polls. The national polling aggregate at TPM finds Obama up 2% from early September – corresponding to a nearly 4-point current lead. The trackers at RCP and Huffington Post indicate similar gains. Compared to past years, 2% is still not a big effect – but, if accurate, it would be much more than any other swing in voter preferences since Romney secured the Republican nomination. If there really has been a shift of 2% nationally, we should be seeing it in the state polls, too. Hopefully, with more polling data, we’ll be able to resolve this discrepancy soon.

While we wait to get a sense of Obama’s post-convention bounce, it’s worth taking a look at how predictive the trial-heat polls can be at this (or any) point of the campaign. I’ll use the state-level surveys from 2008 as a guide. Applying my model to these data, I can estimate the daily level of support for Obama vs. McCain in all 50 states, from May through November. We can then compare the candidates’ standing in the polls during the campaign to each state’s election outcome, and see how those differences varied over time.

To begin, here’s the trend in voter preferences in Florida in 2008. Polls are plotted as circles, for the percent supporting Obama (out of the total for Obama or McCain). The thick line is the model’s estimate of the underlying level of support for Obama. The horizontal line indicates the result of the election, in which Obama received 51.4% of the two-party vote.

The trend is similar to what happened in most states: Obama polled behind his eventual vote share throughout the summer, then lost more ground following the RNC and selection of Sarah Palin. But once the the financial crisis began, Obama quickly gained in the polls – and by early October, the polls were about (on average) where the election ended up.

Next, I calculate the difference between each state’s daily estimate of voter opinion and the state’s election result, and average these across all 50 states. This is the mean absolute deviation, or MAD.

On average, the state polls (fed through my model) were off by 2%-3% through August, increasing to a maximum average error of 4% after the RNC. But by Election Day, the poll-based estimates only differed from the actual outcomes by an average of 1.4%. In states that were most competitive – and therefore polled more frequently – the error was even lower. If we isolate the eight closest states, where Obama finished with between 46% and 54% of the two-party vote, the average error on Election Day was a minuscule (and fairly remarkable) 0.4%.

So, in 2008, the polls were highly accurate over the last month of the campaign; and somewhat less so prior to that.

But looking at the polls isn’t the only way to predict the election outcome. In August and September of 2008, political scientists published a series of forecasts of the national-level vote, based upon long-term structural factors such as levels of economic growth, unemployment rates, whether the incumbent is running for re-election, and so forth. How well did these forecasts perform? The median prediction was that Obama would win 52% of the two-party vote. If we assume uniform swing and home-state effects relative to 2004 (as I describe here), this would have translated to an average state-level error of 3.2% – greater than that of the poll-based estimates for almost the entire campaign.

In fact, the minimum state-level MAD that any national-level election forecast could have produced in 2008 was 2.6%. Here I’ve plotted the state MAD under various assumptions about the national vote outcome (again, using uniform swing with home-state effects). The reason for this lower bound is that although uniform swing is an extremely useful model, it isn’t perfect. From election to election, state vote outcomes still tend to vary by an additional 3%-5% or more beyond the national swing.

The forecasts that I’m showing on this site are produced by combining long-term factors with estimates of current opinion based on the state polls. This stabilizes the forecasts relative to the polls alone, while also reducing the forecast MAD, as can be seen in my paper (Figure 4). In 2008, state forecasts using my model maintained an average error below 2% for the final two months of the campaign.