With 100% of votes counted from the first round of France’s presidential election, Emmanuel Macron and Marine Le Pen are headed to a runoff. Macron, a centrist who defeated Le Pen 24% to 21.3%, is making a lot of waves as the great unifier and the future of France. Le Pen, on the other hand, is making headlines as a far-right underdog with a good shot at besting Macron. Of course, she would have to overcome her current 25 percentage point deficit in the polls to do so.
That is not likely to happen.
Unlike the misleading conventional wisdom floating around, the data simply do indicate the probability of an error large enough to elect Marine Le Pen. The model I developed to forecast the result of the France election currently gives Le Pen a 2.5% chance of victory — and that’s an optimistic estimate.
There are a few reasons why Le Pen’s 2.5% odds — while certainly prone to increase — are very, very unlikely to surpass 50%.
First, let’s discuss the accuracy of opinion polls in France.
…and round 1 of this year’s election was no exception.
After the results came, we observed that opinion polls nearly hit the nail on the head. In terms of over/underestimating each candidate…
For a more customary measurement of error, we turn to the absolute difference in the margin between the final week’s polling average and the top two finishers’ actual margin. In the first round of the past nine presidential elections (according to data from political scientists Will Jennings of the University of Southampton and Christopher Wlezien of the University of Texas at Austin), the average of those errors has been 2.8 percentage points.
|Year||Poll Margin (%)||Election Margin (%)||Error (%)|
That 2.8 percentage points of average error is bigger than error in polls of United States elections, which is in the neighborhood of 1.8 percent, but nearly half as much of the error in polls of UK parliamentary elections.
Surmise to say that polls in France are pretty good, and they performed 1.6 percent better than average (the mean absolute error of polls before Sunday’s first round vote was 3 percent).
Of course, we have moved beyond round one polling accuracy. Now, we want to know how good they will be in round two.
We turn again to the mean absolute difference in the margin between the final week’s polling average and the top two finishers’ actual margin, with data again provided by Jennings and Wlezien.
|Year||Poll Margin (%)||Election Margin (%)||Error (%)|
Generally speaking, polls in round two are just as good as those in round one! However, I’m worried about assuming that future error will match past error, given the number of data points we have for round two polls (9 elections) is very small. It is entirely possible that polls could err by 10 percent or more — of course, that does not mean it’s probable.
All of this is to say that the current polling, which shows Marine Le Pen at a ~25 percentage point deficit, is not likely to err enough to let her win. If we use a t distribution to estimate the amount of error we’d expect to see 99% of the time, the average error of past elections would need to be 6.2% to see an error big enough to elect Le Pen (at least, within that 99% interval).
The model attempts to simulate error with a similar, appropriate measurement of error, as well as taking into account other factors (like variability of undecided voters and current trends in public opinion). That is to say that I don’t expect historical error to be the only error in polls this year (and any good model ought to assume that there can be some unexpected deviation from past results).
Speaking of other indicators, what other factors should we take into account when analyzing Le Pen’s probability of victory?
Since the first two-stage election in France’s Fifth Republic, the winner of each election’s first round has gone on to win the second round 6 out of 9 times. At a macro level this gives us roughly 2/3rd confidence that Macron will go on to be President of France. However, there are some micro-level considerations to this analysis.
|Year||Round 1 Winner||Round 1 Margin (%)||Round 2 Margin (%)|
As stated, the winner of the most votes in round one has only gone on to win the election in 67% of elections — but, in modern (post-2000) politics, the round one winner has gone on to win round two in 100% of elections. This includes the 2002 election, when Marine Le Pen’s father, Jean-Marie Le Pen, lost in a particularly gruesome runoff against Jacques Chirac. Le Pen earned the support of just 18% of voters.
Still, if past trends are indicative of future patterns (as they often are) Marine Le Pen is facing an uphill battle — of course, there’s always the possibility that the past won’t predict the future. I take that uncertainty into account in my model.
Up until this point, we’ve been talking simply about what may cause Marine Le Pen to have a poor showing, as polls indicate, against Emmanuel Macron. What if she has a good showing?
In addition to Le Pen’s bad showing on Sunday (when she earned roughly 1% fewer voters than polls predicted), Nate Silver of FiveThirtyEight did a good job of breaking down reasons why Le Pen will probably not see a surge from “shy voters” — which is, of course, a theory that pollsters aren’t picking up Le Pen voters for a variety of reasons.
Nate’s research also shows that nationalist parties experience polling error of about 3.6 percentage points on average (when the party polls at 25% or more, as Le Pen’s does).
We ought to take that error into account in our modeling, although it’s not clear how to account for it when also adding historical error to our estimation. For that reason, we should simply add this uncertainty to our qualitative assessment of the election forecast.
In the end these are all reasons why we are fairly certain of Le Pen’s chances (or lack thereof) of winning the election, save for a moderate amount of historical polling error and varying evaluations of the current political environment.
Which brings me to our wrap up.
I have covered a lot of sources of potential uncertainty in polling of the 2017 France presidential election’s runoff. I have built a model that accounts for most of these sources of error, and then some more.
Even then, Le Pen only wins 2.2 percent of the time. That’s less than the odds of the Yankees winning the world series, which is six months away!
However, I’ve also discussed a lot of reasons why we should be confident in the picture being painted by public opinion.
Polls in France are relatively good (maybe even very good) estimates of what happens on election day; average polling error is less than we see in several other nations, and comparable to that of the United States. We also have no reason to believe that polls are going to be wrong, despite the recent anti-polling narrative that places too much stock in unsupported claims of massive polling misses.
Some of that conventional wisdom is evident in a forecast of the election by Ian Bremmer, a political scientist who runs The Eurasia Group. Bremmer gives Le Pen a 40 percent chance of victory, without looking at polling whatsoever, despite that those odds are 2-3 times the odds given to a Trump victory last November. Trump’s polling deficit was 3-4 percent, or 6-8 times smaller than Le Pen’s current polling deficit.
There is simply no way (according to the statistics!) that Le Pen’s odds of victory are 40%. Even the betting market PredictIt, which currently puts her chances at 12%, is probably too high. Le Pen just doesn’t have that type of support amongst the French public.
Ultimately, it’s better to trust polling than punditry on this one. Public opinion numbers have proven accurate in previous French contests, and even developing a model that is bullish on Le Pen’s chances of victory tells us that 49 times out of 50, Emmanuel Macron becomes the next President of France.
Thus, the question pundits should be asking is not “can polls err enough to elect Le Pen?” but “can Le Pen pull off one of the biggest upsets in electoral history, anywhere?”
We’ll see you then. In the meantime, keep your eyes on the forecast webpage. It is updated multiple times per day.