Alaska, this year’s new swing state?
CNalysis.com AK House predictions Alaska is one of the most interesting states this November’s election. Not only is the state competitive...
Before the conventions, a slightly narrowing race
Can events like the roll-call vote above change public opinion? We’re about to find out – and trends over the last month will make a convention bounce easy to see, since Democrats have been trending slightly downward.
Our federal poll trackers use different data streams, yet they seem to be moving together. This happened in 2016 as well. Democrats still lead for Presidency, Senate, and House. But in all cases, the race has narrowed by 2-3 percentage points.
As always, the universal yardstick for all three trackers is the Meta-Margin, defined as how much the poll margins would have to shift across the board to make partisan control a perfect tossup: Presidency 270 electoral votes, Senate 50 D+I, and House 218 seats.
Here is the Presidential tracker:
Biden reached a peak in late June, with a meta-lead of 7.2 percentage points. His lead in national polls was over 9 points. So a Republican advantage in the Electoral College of two points is still there, similar to 2016.
Today the Meta-margin is Biden +5.1%, a narrowing of 2.1 percentage points. This is also consistent with movement in Trump’s job approval rating as calculated by FiveThirtyEight, which has also moved toward Trump in the last few weeks after hitting an extreme value last month.
Compare the Presidential Meta-margin with the Senate Meta-margin, which shows a narrowing of the race by 2.4 percentage points, from a peak of D+6.0% to D+3.6% today.
Finally, the generic Congressional ballot, corrected for structural biases including gerrymandering, has narrowed from D+7.0% to D+3.3%. This measure fluctuates more, and if smoothed, the change is more like 3 percentage points.
What does this all mean? Well, there is the possibility of differential response bias – the idea that Democratic-leaning and Republican-leaning voters answer surveys with different likelihoods, and that likelihood shifts slightly over time. That’s been suggested before.
Another possibility, which I tend to favor, is the idea that these surveys measure real shifts in opinion. I would expect Biden’s meta-lead to continue dropping, at least for the next few days. After that, we can see what (if any) result comes from this week’s convention.
To be clear, this is a small movement. Identifying a consistent 2 points of change required substantial computation, and was more of a technical accomplishment rather than a shift. Probably the natural resting point of the race is at the middle of this rather narrow range.
But if this analysis got you wound up, that’s good! Now find a race in your home state or region, and go make a difference there. See the PEC 50-State Guide to focus your efforts.
The prediction and betting markets are responding to this apparent narrowing by also increasing Trump’s chance of reelection. But I wonder if betting markets are also concerned about Trump’s efforts to suppress votes and likely effort to steal close states analogous to the GOP stealing Florida in 2000, which bettors think increases Trump’s odds beyond just what the polling info indicates. Also, I wonder to what extent the meta-margin is currently a lagging indicator. There have been six major polls released post-Harris with Biden ahead by 10, 10, 10, 9, 6, 4, which yields a median of 9.5 and a mean of 8.2, which would suggest a meta-margin of 6 or 7, assuming about a 2 point advantage for Trump in the electoral college. But I have seen almost no post-Harris state polling. I anticipate the state polling, when released, will follow the national polls and the meta-margin will rise accordingly. As someone who wants to see the guy doing a spot-on impersonation of Bolsonaro lose, I hope so.
As I suspected (hoped) now that state polls are coming out post-Harris, the meta-margin is following the national polls. No knock against the meta-margin that it is a lagging indicator in this case. It really depends on the pollsters release of polls.
Looks like sampling jitter to me. If you look at the magnitude of the 1st derivative, only mid-April and mid-June (mid-July for Senate) stand out.
I don’t think so. Again, they are all moving in the same direction.
I totally left out another observation: Trump’s job approval number, per FiveThirtyEight, has crept up about 1 percentage point in the last few weeks.
Trump job approval is regressing to the mean (about 42%-ish). Fivethirtyeight do their plots with suppressed zero so tiny little fluctuations seem meaningful.
Also they do this pollster weighting thing which puts me in mind of the Very Bright Undergrads you find in physics labs who cannot use a simple voltmeter without adjusting for temperature and barometric pressure and phase of the moon.
The CNN Poll seems to have shaken a lot of folks, especially since it appears to be widely reacted to as a slightly Democrat leaning poll. It is remarkably out of line with the poll they released three or four weeks ago. How can you account for an 8 to 10 point shift in one month (without a lot of news other than Trump’s reinstating the daily briefings? Hard to figure.
Normally I delete comments regarding single polls. I’ve given you dozens of data points here aggregated into a single measure. I have nothing to offer you on the subject of what you should do with that particular data point.
Just a reminder to everyone, such questions are antithetical to what we do here. In poll aggregation, our goals are low-noise measurement, finding ways to ignore outliers and prevent cherrypicking, and allowing you to maximize your leverage as a citizen.
With just two polls, statistical noise alone is enough. Typical 95% confidence levels for polls are +-3% [and this only includes statistical effects; actual uncertainty is larger because of other effects], meaning that “Biden +4” actually means that anything between “Biden +7” and “Biden +1” is likely. Richard Wiener comment above lists 6 recent polls with Biden leads of 10, 10, 10, 9, 6, 4. All of them are perfectly compatible with Biden +7.
To put it another way: typical polls are based on ~ 1000 interviews, which are somewhat similar to 1000 coin flips. If you flip a “perfect” coin 1000 times, there is a ~10.8% chance that you get 520 heads (or more), meaning that head-tails>=40, i.e. a result where Head leads by more than 4% (even though the “lead” is a purely statistical artifact, since the coin is supposed to be “perfect”). Symmetrically, there is a ~10.8% chance to get a symmetrical result with Tail leading by 4% or more. In a vacuum, the chance of observing two consecutive polls shifting by 8% or more is ~1.2%, which is small though not entirely negligible. But here you are not looking at two random polls, since the second one is likely an outlier; in such cases, the chance of observing a shift >=8% (meaning that the “comparison” poll is an outlier in the other direction) is of the order of 10.8%, which is definitely possible.
Why not plot the average of the week-to-week differences of each pollster, rather than the difference of the average?
I guess because you would need the same pollster with same statistics and very few meet that criteria. You could try grouping similar pollsters but that is probably too much work for very little reward.
Is there any way for your meta-margin methodology to incorporate special treatment for polls that track exactly the same people over time?
Such polls will likely have some permanent bias due to pure sampling effects, but when you see a change in one you know for a fact that it is due to someone changing their mind. That makes them especially sensitive to the effects of recent events, and that sensitivity could perhaps be used to inform the general drift (or not) in the aggregate of normal polls.
If nothing else, the meta-margin could be annotated with a slight upward or downward sloping arrow indicating the effect of tracking polls.
Just a thought.
Doing that is, perhaps surprisingly, less useful than the Meta-Margin. The reason is that single polls, even if they share the same methodology, have irreducible statistical sampling error (“margin of error”) that is typically in the range of 3 points. So comparing two single polls carries lots of uncertainty.
Compare that with the Meta-margin, which is pretty stable. In past years it has done an excellent job of responding to political events within a few days. Its precision as an indicator appears to be well under a percentage point. (That is different from its accuracy – it could have a systenatic error of up to several points.)
tl;dr, what we do is probably the best way to track short-term changes over time of any calculation anywhere.