Élection québécoise: Ligne de départ

 Bonjour,

 Je vais couvrir les sondages de la campagne électorale qui commence comme je l'ai fait dans de multiples campagnes précédentes. Contrairement aux "agrégateurs", je ne m'intéresse pas aux projections de sièges mais plutôt à tracer l'évolution des intentions de vote de l'ensemble de la population et de divers sous-groupes. Pour estimer cette intention de vote, j'utilise ce qu'on appelle une régression locale. Il s'agit d'une procédure de lissage qui accorde plus d'importance aux intentions de vote qui sont plus proche les unes des autres et moins d'importance aux estimations qui s'écartent des autres. Le même type d'analyse est utilisé par Wikipedia (et peut même être effectué en utilisant ChatGpt). Ce type d'analyse est plus sensible aux mouvements dans les intentions de vote que les estimations utilisées par les agrégateurs.

Dans les graphiques, les points représentent des intentions de vote positionnées à la mi-période du sondage: un sondage réalisé du 24 au 26 août sera positionné le 25 août. Les lignes représentent l'évolution estimée des intentions de vote telles qu'estimées par la régression locale. 

Quelle est donc la situation en ce début de campagne? Le premier graphique montre l'évolution des intentions de vote depuis le début de janvier 2026. Il inclut le sondage Synopsis publié ce matin mais comme il est "seul", son impact sur l'analyse est minimal . Le graphique montre 

a) que le Parti québécois (PQ) aurait perdu environ trois points de pourcentage depuis janvier; 

b) que l'élection de Christine Fréchette à la tête de la CAQ (ligne bleu verticale) a fait grimper d'au moins cinq points les intentions de vote pour ce parti dans les semaines suivantes mais que cette montée s'est arrêtée deux mois plus tard, à la mi-juin; si on se fie au dernier sondage Synopsis toutefois, ces intentions de vote recommencent à progresser. Attendons au moins un deuxième sondage pour valider le tout. 

c) que la montée de la CAQ (nouvellement nommée Équipe Christine Fréchette (ECF?)) s'est surtout faite aux dépens du Parti Libéral du Québec (PLQ). Il est inexact de dire que "le PQ n'est plus seul en tête pour la première fois depuis trois ans" comme l'affirment certains medias puisque le PQ et le PLQ étaient à égalité dans six sondages entre la mi-avril et la mi-mai. Toutefois, comme on le verra par la suite, c'est très différent chez les francophones.

Qu'en est-il des Francophones?

Le même type de graphique donne un portrait instructif pour ce qui est des Francophones. Il montre que le Parti québécois a perdu près de cinq points chez les Francophones depuis janvier et que la CAQ - ECF a gagné 10 points durant la même période. Celle-ci prend des votes autant au PLQ qu'au PQ. Québec solidaire (QS) apparaît stable tandis que les appuis au parti Conservateur (PCQ) ont fluctué pour se retrouver au même niveau qu'en janvier.

 

 

Et la RMR de Montréal?

Pour ce qui est des dynamiques régionales, je me concentre sur Montréal parce que les échantillons dans les deux autres régions -- la RMR de Québec et le reste du Québec --  sont la plupart du temps trop petits pour tirer des conclusions fiables.

Le PLQ règne toujours dans la grande région de Montréal. Toutefois, la dynamique que l'on a perçu pour l'ensemble du Québec est très claire également dans cette région. Après une montée importante des intentions de vote à partir de janvier, le PLQ a perdu les appuis qu'il avait gagnés pour se retrouver au même niveau qu'en début de période. Le PQ a perdu environ deux à trois points. Encore là, la CAQ a fait des gains, surtout dans le 450 si on se fie au dernier sondage Synopsis. Le PCQ a aussi perdu un peu d'appuis et QS est stable.

 

 Un dernier mot sur les non-francophones, soit les allophones et anglophones

On parle peu des no-francophones dans cette campagne, sans doute un peu parce qu'on les considère concentrés dans certaines circonscriptions et acquis au PLQ. Les firmes de sondage, quelle que soit leur méthodologie, ont beaucoup de difficultés à les joindre et les échantillons sont vraiment petits (entre 107 et 214 répondants). Le graphique montre les intentions de vote mesurées pour ce groupe. De fait, le PLQ domine mais la courbe d'évolution est la même que pour les francophones. Toutefois, la baisse se ferait en faveur du PCQ, ce qui est peu plausible, la clientèle du PCQ se retrouvant surtout dans la région de Québec. On voit aussi une légère augmentation des appuis à la CAQ, de 6% à 12%. 

Il faut se questionner sur l'évaluation qui est faite des appuis au PLQ dans ce groupe. Ces appuis risquent d'être sous-estimés. L'analyse que j'ai faite de cette question il y a très longtemps semble encore valide. Il s'agit de ce que j'avais appelé le "paradoxe de l'anglo péquiste". En gros, l'idée est que, plus il est difficile de rejoindre un groupe de la population -- les jeunes, les allophones, etc. -- plus ceux qui sont rejoints sont susceptibles de ne pas être représentatifs de leur groupe. Toutefois, vérification faite dans les sondages publiés jusqu'à maintenant, la relation entre la proportion de non-francophones et l'intention de vote pour le PLQ chez ces derniers est quand même faible.

Un dernier point sur les non-francophones. On sait qu'ils sont très concentrés dans certaines circonscriptions. Toutefois, en 2014, j'avais montré que le vote des non-francophones , généralement fortement opposés à la Charte des valeurs proposée par le PQ à l'époque, avait fait perdre trois circonscriptions au PQ, entre autres parce qu'ils s'étaient fortement mobilisés pour aller voter. 

On trouve des circonscriptions majoritairement non francophones dans plusieurs parties de  l'île de Montréal et plus de 30% de non francophones à Laval, mais on trouve aussi 25% de non Francophones dans Camille Laurin, 27% dans Rosemont, 33% dans Mercier et 36% dans Ste-Marie-St-Jacques. Dépendamment des enjeux de la campagne, la mobilisation ou pas des non Francophones peut changer la donne dans des circonscriptions où on ne les attend pas nécessairement.

 

En conclusion

Pour les sondeurs, l'élection de 2026 peut apparaître comme "l'élection de tous les dangers", étant donné le nombre de partis compétitifs: La moins grande distance entre les partis facilite les changements de camp. Il faudra suivre les sondages attentivement pour avoir une idée de ce qui se passe, se garder de conclure à partir de petits échantillons et se rappeler qu'"un sondage ne fait pas le printemps".

Ce ne sont pas les sondages qui influencent le vote, c’est ce qu’on fait avec…

 Bonjour,

 

Je partage un article que j'ai publié dans Options politiques intitulé: Ce ne sont pas les sondages qui influencent le vote, c’est ce qu’on fait avec…

Ceci me permettra de le partager sur Facebook! 

The Hungarian Polls, a final update

 

Prepared by Tamás Bodor, University of Wisconsin, and Claire Durand. Université de Montréal

All votes have now been counted for the 2026 Hungarian national election. According to the final tally, Tisza won 53.5% of the votes, which amounts to a 14.7-point lead over Fidesz (38.9%.) This final tally, however, is not suitable for assessing the performance of pre-election polls: it includes over 300,000 mail-in ballots cast by members of the Hungarian diaspora and nearly 85,000 in-person absentee votes cast abroad, which were not included in the sampling frames of the polls.

As our Hungarian colleague Gábor Tóka pointed out, our first report on April 13 was based on a preliminary vote count that was not yet sufficiently accurate. First, by the time we obtained the preliminary count on April 13, most mail-in ballots were already included in the total. Moreover, hundreds of thousands of domestically cast absentee ballots had not yet been counted. While the former, the diaspora ballots, heavily favored Fidesz, the latter were strongly favorable to Tisza.

For our updated report, we are now using the final tally from which we excluded mail-in votes. Moreover, we also excluded in-person absentee votes cast abroad. We assumed that these 85,000 absentee votes favored Tisza 63% to Fidesz’s 31%, just as the outstanding votes as of April 13 did. Accordingly, we estimate that Tisza received 55.65% of the domestic votes, while Fidesz received 36.40%, amounting to a 19.25-point victory for Tisza domestically. We believe that these figures correspond most closely to the poll’s sampling frame. They are very similar to Gábor Tóka’s estimate of April 14 (55.7-36.3) and to the Wikipedia estimate of the constituency vote (55.3-36.7).

With these new numbers, our conclusion has changed. The new estimates are presented in the following table:

This shows that Medián and 21 Kutatóközpont who had “boosted” the difference between Tisza and Fidesz in their likely voter model were right. Their estimates are closer to the comparable election results. Overall, excluding McLaughlin, who is clearly an outlier, the polls overestimated Fidesz by an average of 3.3 points and underestimated Tisza by 1.8 points, and at least one of their estimates is outside their margin of error (except for Median and 21 Kutatóközpont). This could be partly due to pollsters’ caution, given their underestimation of Fidesz in 2022. This could also be due to voters, given the very high turnout.

Conclusion:

Our first concern was to see what would happen as the campaign went on, since two groups of pollsters published contradictory estimates. The answer: All pollsters whose estimates showed a Fidesz lead earlier, except one, -, stopped publishing weeks before the election, which speaks for itself. The one exception: McLaughlin & Associates, maintained that Fidesz led Tisza using an in-person survey, which also speaks for itself.

Our second concern was the performance of the new polling methods.  These new methods varied and accounted for most polls published in April. We did not find any substantial and systematic differences according to the mode of administration and sampling source, as the two are related.

Our last point is about the proportion of undecideds, in fact non-disclosers, that is, all the respondents who do not inform their preferences. This proportion remained quite high until the end, regardless of the mode used. This may partially explain the difference between the polls’ estimates and the votes.

Best,

Claire Durand and Tamás Bodor

The Hungarian Election Polls, the Day After

 

Hungarian Election polls, the Day After

Prepared by Claire Durand (Université de Montréal) and Tamás Bodor (University of Wisconsin)

Hi,

To follow on our post on Friday, here are our observations regarding the pollsters’ work for the Hungarian election. We examine the quality of the polls that were conducted during April.

1.      1. Let us notice that the pollsters who were considered close to the opposition or independent produced very good estimates in general. On average, there is only a 1.1 difference between the polls and the results of Tisza and 0.9 difference for Fidesz. Both parties were very slightly overestimated by these pollsters (see table below).

2.      2. Here is the table of the results, based on the official results available on Monday, April 13, 4PM. They are not likely to change much.

3.      3. However, three of these pollsters overestimated Tisza and underestimated Fidesz. Interestingly, two of these pollsters are the ones we referred to in our previous post – Medián and 21 Kutatóközpont – whose allocation of undecideds was not proportional and gave an advantage to Tisza over Fidesz. The “pure” proportional allocation would possibly have led to better estimates.

4.      4. We can hardly speak about the polls that were considered close to the government because none of them published any poll in April, besides one. The only pollster of this group who published a poll is the American pollster McLaughlin. It underestimated Tisza by 12.8 points and overestimated Fidesz by 6.2 points.  These numbers speak for themselves.

5.      5, The pollster Nézőpont, considered close to the government, issued a statement regarding its performance. It recognized that its estimates were not accurate and informed that it will analyse how it could explain the situation.

6.      6. In terms of modes of administration, only three pollsters used live phone interviews and one, in-person interviews (McLaughlin). The others used a mix of different modes (see table). Atlas Intel, a Brazilian pollster present in other electoral campaigns (U.S.A., France, at least), with usually good estimates, also appeared in this campaign. It uses a mode of administration it calls random digital recruitment, etc. It consists in recruiting respondents on the web through apps and web sites with whom the pollster has agreements. It has the advantage of randomness and being web-based.

Conclusion:

There were a few things going wrong in the polls for the Hungarian election. First, a major difference appeared in estimates almost uniquely based on the pollsters presumed affiliation. This is “not normal”. It means that there is likely some “tempering with data” based on hypotheses made on the voters’ behavior. The use of likely voter models that are not transparent is usually associated with such results, but we have never seen such a substantial difference.

We did not always have the necessary information to assess whether allocation of undecideds or likely voter models explained the differences between pollsters. In fact, the last published polls were even less transparent than the polls published before during the campaign.

Two situations are worth mentioning. First the government-leaning polls were much less transparent – three out of five did not inform on the proportion of undecideds and on the estimates before allocation of these undecideds. In addition, four out of five did not publish polls in April. These pollsters need to examine their methods and the information they provide. It is a good point that Nézőpont informs that it will do so.

Besides, among the opposition-leaning or independent pollsters, some estimates were clearly boosted in favor of Tisza. The good news is that we could notice it because the pollsters were transparent and presented their estimates before and after allocation of the undecideds. However, they did not provide this information for their polls published in April.

 

Best,

Claire Durand and Tamás Bodor.

What about the polls for the Hungarian election next Sunday?

 By Claire Durand and Tamás Bodor

We—Tamás Bodor from the University of Wisconsin - Stevens Point and Claire Durand, from the Université de Montréal—wish to draw your attention to the Hungarian elections that will be held this coming Sunday, April 12. Hungarians face a situation where pollsters are identified, according to the Wikipedia page, as being either “close to the government” (n=5) or “close to the opposition or independent” (n=7). As you can see on this Wikipedia page, https://en.wikipedia.org/wiki/Opinion_polling_for_the_2026_Hungarian_parliamentary_election, there is a significant disagreement between the two groups of pollsters.

If we consider all the polls conducted since the beginning of March, for example, the government-leaning polls forecast Fidesz, the governing coalition led by Viktor Orbán, ahead of Tisza, the opposition coalition led by Péter Magyar, by 5 to 8 points (an average of 5.8 percentage points).

Meanwhile, the polls conducted by the opposition-leaning or independent pollsters forecast a Tisza lead by 9 to 23 points, for an average of 13.4 p.p.

Who is correct? Usually, we think that pollsters, regardless of their preferences, seek to forecast the results as accurately as possible since electoral polls serve as a showcase that can help them establish their credibility. So why this situation?

We looked into a number of possible reasons. Here is what we found.

1)      We examined what happened in the 2022 election. In that election, if we take the pollsters’ last campaign poll, Fidesz was underestimated by around 3 points (2.2 to 5.1) and the opposition coalition—which was not the same as this year’s coalition—was substantially overestimated by 4.1 to 11.1 points, for an average of around 9 points. However, there were no significant or apparent differences according to presumed affiliation (most pollsters were the same as those who are active in this election).

2)      We checked whether the differences could be explained by modes of administration. There is indeed a difference between estimates according to mode of administration and there is a difference in mode use according to affiliation. All the government-leaning pollsters conduct their polls using live interviewers, while only three of the opposition-leaning or independent pollsters use this mode. The four other pollsters use mixed modes (telephone + web, sms, web + telephone), “IA assisted IVR” and “web+social media”.  However, if we keep only the live interviewer polls, there is still a clear difference remaining between the two groups of pollsters.  Using all the polls conducted since January 1, the government-leaning pollsters estimate Fidesz’s share at an average of 46.9% (45% - 51%) while the opposition/independent group estimates Fidesz’s share at 38.9% on average (35% - 41%). The pollsters using other modes—all opposition leaning or independent—estimate Fidesz’s share at 38.1 (35%-40.1%). This is not different from the polls conducted by the opposition-leaning and independent pollsters using live interviews.

3)      We examined whether there were differences in the proportion of undecided voters. Government leaning polls differ from the other polls by a much lower proportion of undecided voters, that is 14.1% on average, compared with 23.2%. However, the proportion of undecideds is not provided by three of the five government leaning pollsters. Reallocation of undecideds to one or the other coalition may contribute to explaining the differences between the two groups. Since most pollsters provided the information on the estimates before allocation of undecideds, we computed what a proportional allocation of undecideds would produce in terms of estimates. When we compared our estimates with the pollsters’ estimates, we found that three pollsters (9 polls), all in the opposition leaning group, allocated the undecideds non proportionally in a way that led to a systematic increased advantage of Tisza over Fidesz by an average of four points (between 2 and 6.4). Indeed, two of these pollsters forecast a much higher lead of Tisza than the other pollsters.

4)      We are left with not many possible other explanations for the differing estimates among pollsters. One of them may be likely voter models, if we assume that they systematically vary among pollsters according to their partisan leaning. The higher rate of “undecideds” / non-voters in opposition-leaning polls and the possibly biased allocation of these undecideds might hint at that possibility.

Moreover, the shifting political opinion climate in Hungary might have generated a profound change in patterns of respondents’ behavior. For the first time since the party’s return to power in 2010, most voters now expect Fidesz to lose the election. Accordingly, it may not be an implausible proposition that a “shy Fidesz” effect might be at play. However, theoretically, Fidesz supporters’ reluctance to disclose their party preference should affect all polls. Therefore, a hypothesized shy Fidesz effect can account for systematic differences in estimates only if it goes undetected and unaddressed by some pollsters but not by others.

It appears that Hungary is very polarized, as is usually the case when only two main groups fight in an election. This has led some people in Hungary to suggest that we should use only the polls conducted by one or the other group (usually the opposition-leaning/ independent group).  On what basis could we do that? We do not see any rationale for such a decision. The truth can probably be found in an average of all the polls. That being said, both groups agree on one point, namely, that both Fidesz and Tisza increased their share in the last few weeks. The global forecast from all the polls using local regression (Wikipedia page) has Fidesz at 42.5 and Tisza at 47.5. We need to keep in mind, however, that there are twice as many polls conducted by opposition-leaning/independent pollsters, and these polls forecast a larger difference between the two coalitions. This situation has an impact on the average forecast.

One last point: We were happy to see that, for most polls, we had access to a well-identified paragraph presenting essential information on the methodology, that is, dates of fieldwork, mode of administration, sample size, proportion of undecideds and sometimes information about the use of a likely voter model. This information was available either in the articles presenting the polls or on the pollsters’ reports. An example to follow in all election campaigns?

It will be interesting to follow the Hungarian election on Sunday. The incoming vote tally results will bring about the moment of truth and possibly a moment of reckoning: this time, not only for the losing party, but also for the polling industry in Hungary. Ultimately, we all might also be reminded that pollsters’ political preferences should not be known publicly.

 Best,

Claire and Tamas