Odoxa’s last survey for Public Sénat, released on July 4, 2024, gave the National Rally 210-250 seats. The New Popular Front was put second with 140-180, and the presidential bloc third with 115-155. After the second round, the order was reversed: the Front won 182 seats, Ensemble 163, and the National Rally finished third with 143.
That gap says a good deal about how French election polls are built, and where they stop working.
The quota method
Most French pollsters rely on quota sampling. Larousse describes it as building “a priori a reduced model of the parent population, such that one finds in it the statistical proportions of that population.” Interviewers fill set quotas on variables such as gender, age, place of residence and income. The method, Larousse notes, “gives good results and has the advantage of being more economical and faster than the random method.”
The alternative is a probability sample, which means drawing people “under conditions close to chance” from a complete list of the population. That is slower and costs more.
Rules and oversight
A 1977 law banned publication of polls in the week before a vote. The Court of Cassation struck that down in 2001, and a 2002 law kept the ban only for the day before the vote and polling day itself. The Commission des sondages checks that polls are “conducted and sold in accordance with the law.”
Seat projections sit outside that safety net. On June 18, 2024 the commission warned that such projections “must be interpreted with caution” and said it does not review their methodology the way it does for ordinary polls.
Why 577 races break the model
A national poll can sample the country. A legislative election cannot be sampled that way. “Since legislative elections involve 577 separate races and we cannot survey 577 representative samples, we must make projections based on assumptions,” said Christelle Craplet, director of BVA Opinion. Her institute chose not to publish seat numbers at all.
Those assumptions differ between firms. Sociologist Hugo Touzet said institutes build matrices with different parameters, for instance by giving outgoing deputies a different weight. Two projections from the same raw data can therefore point in different directions.
What the polls missed
Observers pointed first to the Republican front, in which voters backed whichever candidate was best placed to beat the National Rally. An Ipsos Talan estimate found that 72 percent of Front voters supported Ensemble candidates in runoffs against the National Rally. Ensemble voters, in contests that pitted La France Insoumise against the National Rally, picked the former 43 percent of the time and the latter 19 percent. Craplet said the scale of it “was difficult to foresee.”
Alberto Bitonti of the Università della Svizzera italiana makes a broader point: polls capture “voting sentiment at the time of the survey and not definitive predictions about elections.” Shy voters, turnout shifts and strategic second-round choices all work against precision. Good polling, he says, needs large samples, stratification by gender, social class and education, and open disclosure of methods and margins of error.
It was not a new problem. In 2022, surveys put the left-wing NUPES alliance at 145-210 seats, and it won 131. They projected at most 50 seats for the National Rally, which took 89.
Sources
- Sondage (Larousse encyclopedia), translated
- Did polling institutes miscalculate the 2024 legislative elections? (Public Senat), translated
- French elections and the limits of polls (USI, Università della Svizzera italiana)
