Polls, surveys and the known unknowns of the French presidential election

Un dernier billet en anglais, et je ne dirai plus rien sur l’élection présidentielle de cette année. Je ne pense même pas faire de “post-mortem” du premier tour, vu que mes prédictions n’ont rien d’exceptionnel (loin de là) et que les réponses les plus intéressantes viendront après les résultats – et l’étude des reports de voix – des législatives.

If you have read any of my two previous blog posts on the respective electoral prospects of Marine Le Pen and Emmanuel Macron in the upcoming French presidential election, you might have noticed that I did not use either polling or survey data to produce my predictions, relying entirely instead on past elections.

This is because I hold strong views about both opinion polls and electoral survey data when it comes to France. I detail these views below, as some kind of a methodological appendix to my previous posts.

Note that my electoral predictions are nothing but semi-educated guesses, and that my methodological views are just the same: I have no scientific credentials in the fields of electoral studies, opinion polling or survey methodology.

Like the previous ones, this post is written in English because it is addressed to friends who do not read French, but who keep asking about the French presidential election and the risk that it might turn into a local equivalent of the Brexit “Leave” vote or Donald J. Trump’s election to the U.S. presidency.

Update, 24-25 March: added a section on non-random polls, including a link to (and several citations from) the UK 2015 “polling miss” report, as well as a link to the French Polls Committee.

French opinion polls

As in any other wealthy representative democracy, opinion polls provide a large share of the electoral horse race spectacle. And just like anywhere else, these opinion polls have sometimes been awfully wrong, even though their average historical record is, by and large, broadly accurate (more on that later).

There are lots of reasons why one might be skeptical of opinion polls in general, and French opinion polls in particular. The classical Bourdieusian critique applies here, and French political scientist Alain Garrigou has done lots to popularise it: polls are partly performative, insofar as they frequently ask questions that respondents had not thought about before. In the context of a two-round election, this particularly applies to hypothetical “second-round” voting scenarios.1

Another problem, which is perhaps more prevalent in French opinion polling than elsewhere, has to do with sampling. In recent years, the issue of whether opt-in Internet sampling creates representative samples has attracted a lot of discussion, but an older and, in my view, more problematic issue, has to do with quota sampling, another non-probabilistic method that is widely used by French polling institutes, principally because it is much less expensive than proper random sampling.

The main problem with quota sampling is that it creates a layer of opacity around the precise re-weighting methods used to make the poll as representative as possible of the target population – an operation called “redressement” in the jargon of French polling institutes, which develop their own “in-house” methods for doing so, and then sit on it like some kind of industrial secret.2

This is a huge issue: even when quota sampling seems to be able to “emulate” simple random sampling, we have no idea how that happens. To me, this issue makes the defence of quota sampling hopeless, perhaps even more than the trivial observation that the standard error of a quota sample has no formal relationship to sample size, which makes the presentation of error margins in quota-based polling results highly misleading to the public.3

Troubles with non-random polls

Allow me to elaborate for a few paragraphs on the point made above.

The idea of “emulating” simple random sampling is, at its core, fundamentally flawed. Nature cannot be fooled: pseudo-randomness is non-randomness. Even simple random samples will contain nonrandom (missing data) components, which are hard enough to deal with. Adding non-randomness “at the root” of the sampling procedure is just madness in the present context.

And yet this is something that polling companies seem to be okay with, and not just in France: just next door, the British polling market seems to be configured exactly the same way, and even though they have less things to predict over there since the party system contains less (national) elements, they also, from time to time, get it completely wrong, precisely because their samples are non-random.

Here’s the conclusion of the committee that worked on the serious polling miss that occurred in the UK before the 2015 general election, which I mentioned earlier as the true starting point of the “Brexit-Trump” series of failed predictions:

… the primary cause of the polling miss in 2015 was unrepresentative samples… The statistical adjustment procedures applied to the raw data did not mitigate this basic problem to any notable degree. The other putative causes can have made, at most, only a small contribution to the total error. (p. 4, emphasis in original)

And a few pages down in the same document:

… all GB pollsters in 2015 took a common approach to sampling and estimation: they assembled a quota sample of eligible individuals, which was then weighted to known population totals. They asked sample members their vote intention and likelihood of voting, derived a predicted sub-sample of voters, and produced weighted estimates of vote intention for this sub-sample. (p. 22).

And then the report goes on to explain how the BES and BSA surveys are fundamentally different, because they use simple random sampling, ad because “both surveys produced good post-election estimates of the difference in vote shares between the Conservatives and Labour” (p. 27; important word here: “post” means that the surveys were run after the election, so considerations about turnout and voting uncertainty did not apply).

Issues with non-random quota samples are discussed in detail in the Francophone Wikipedia entry for opinion polls. The entry has multiple issues to do with its sources and the balance of its sections, but it otherwise makes a good read on the topic of why opinion polling is a problematic source of information, especially when it comes to voting behaviour. And if you need an example of a recent opinion poll to illustrate how black-boxed polling methodology is, here’s one.4

What can electoral polls be used for?

The flaws outlined above do not mean that opinion polls cannot produce either interesting forecasts or useful information. Measures of voting uncertainty, for instance, can be accurately measured through polls, although their generalisation to the electorate is still a very risky exercise absent of detailed information on the composition and re-weighting/redressement of poll samples. Still, compared to past elections, it seems that voting uncertainty is currently very high, especially on the left.

Another interesting use of opinion polls for electoral purposes consists in aggregating as many polls as possible, re-weight their results according to what we know of their past performance and overall reliability, and then use these results to forecast electoral results. This style of electoral forecasting has been popularised by Nate Silver‘s forecasts of U.S. elections, some of which were highly accurate, some of which were far off the mark.5

There is currently one attempt to build a Silver-style, “polls-only” forecast of the French 2017 presidential election, which came to my knowledge thanks to a tweet by François Massot. The forecast is produced by Victor Poughon and published on the “Depuis 1958” (“Since 1958”) website. The current result of the forecast, which is based on transparent polling data, unambiguously predicts that Emmanuel Macron will win the election:6

The unequivocal result of the 2017 forecast is, to me, less interesting than the “retrofit” of the model to previous elections. The Bayesian methodology of the forecast shows the model fit to the 2012 presidential election, with the difference between the actual (red) and predicted (black) results:

I also find this kind of exercise interesting because it produces a rough estimate of how much of the vote might go to the first-tier (20%+ vote shares), second-tier (10-20% vote shares) and third-tier candidates (5% or less vote shares). In 2012, these shares respectively represented around 55%, 38% and 7% of the first-round vote. For 2017, the model predicts 44%/44%/12% vote shares, which is consistent with estimates of higher voting uncertainty:

Final observations: note how wide the credible intervals reflected in the predicted vote shares shown above are. Also note that the model does not say anything about turnout, which might be lower this year. Last, note that the model cannot inform us about who is voting for whom, i.e., about voter realignments (which can also differ between presidential and legislative elections). For all these reasons, we need other forms of electoral data to supplement what polls might provide.

French electoral surveys

Especially perhaps in a multi-party system, highly detailed and sophisticated electoral surveys are much, much more important than polls to get the data generating process – i.e. what makes people cast a vote, and for whom – right. Do we have (access to) such surveys in France?

The answer is: kind of. Here are, for instance, some results from the most recent CEVIPOF/Ipsos/Steria electoral survey, the Enquête Électorale Française 2017, which also show high rates of voting uncertainty, especially when it comes to the left-wing candidates:

What credit might we lend to that survey? There is, in fact, no way to tell, except for CEVIPOF or Ipsos/Steria insiders. The survey data are not public, and the kind of sketchy technical explanations provided in the media hint that even for the official partners of the survey, French newspaper Le Monde and the Jean-Jaurès Foundation, the exact methodology that produced the data is a black box.

Lack of transparency from private polling institutes is one thing; but partly-publicly funded research is different, or at least, in my view, should be. The conclusions that insiders are drawing from the Enquête Électorale Française 2017 are highly valuable, but they do not allow outsiders to learn anything from the survey data itself, or to reuse the data in different ways. In short, French electoral data, just like many other forms of data, suffer from serious availability and cumulativity issue.

Further problems with survey data

There are several other problems with “flat microdata” (repeated cross-sections), which make up most of polling and electoral survey data.

First, to answer some questions, we need proper longitudinal (panel) data, that is, data that are produced by asking repeated questions to the same respondents at different points in time. Second, for many other questions, we need relational (network) data, of the kind produced by family-wide sampling (when the entire household is sampled). Last, in the many cases (including voting) where geography matters, we often need spatial (geolocated) data.

A few years ago, my co-blogger Joël called for the launch of a longitudinal French National Election Study, suggesting that the survey should allow multilevel analysis—that is, a form of analysis that can observe and control for the existence of hierarchical (e.g. local/regional/national) patterns:

Such a survey exists in the United States. France has no equivalent survey; the French national statistics agency does not seem to be interested in running one; and the European Social Survey, which has many of the required qualities that such a survey would call for, can only imperfectly fit the bill. Last, the French Ministry of the Interior cannot produce such a survey, out of fear of being accused to produce a partisan survey.

As a result of the many different limitations outlined above, there is no data source to turn to when one wants to analyse voting intentions in France. Sure, there are polls and electoral surveys, but neither of them are currently transparent enough to allow for independent reanalysis. Instead, what we have in France is a small club of (French) election specialists, working on disparate data sources that only very imperfectly cumulate.

Things might be getting better: France has a data repository, the Réseau Quételet, which lets researchers access some of the existing electoral studies produced under the supervision of various research teams. Note, however, that the current “French Election Studies,” a by-product of the Comparative Study of Electoral Systems (CSES), are limited to presidential-election years, and are available only for 2007 and 2012, even though they seem to have been running for longer than that.

Further reading

This blog has many posts about opinion polls, some of which focus on (older) French presidential elections: see here, here and here.

For a fuller assessment of French survey data, take a look at the appendix in Vincent Tiberj’s last book on cohort and generational effects among French citizens. The book itself is absolutely excellent, and the appendix provides a useful list of surveys – many of them cross-national – that contain data on French social and political attitudes7

  1. And outside of electoral periods, one might argue that polls co-produce the presidential approval ratings that they claim to measure. Notwithstanding, I do not agree with Garrigou that prohibiting electoral polls is the way to go. []
  2. Furthermore, there is no French equivalent of either the ICPSR or the Pew Research Center, which means that there is no repository of polling data to turn to in order to inspect raw survey data. As a consequence, the entire system of French opinion polls is effectively black-boxed. []
  3. The committee in charge of regulating polls has stated that polls should come with an empirical “margin of uncertainty” rather than a theoretical one, which actually makes things worse, since “empirical” can mean pretty much anything here. []
  4. The linked example characteristically provides none of the methodological explanations that should come with any polling result: how were the samples collected, how was nonresponse addressed, what the denominators are for population/electorate estimates, etc. []
  5. For another example of aggregated-polls-based forecasting, see Chris Hanretty’s forecast of the Italian 2013 general election. []
  6. Note that the 2017 forecast goes beyond the first round of the election, and actually attempts to predict second-round results. Also note that Macron has always been high in the approval-rating polls, which likely played a role in his candidacy. In past elections, Alain Poher and Édouard Balladur also ran for the French presidency (in 1969 and 1995 respectively) with similarly high approval ratings in the polls. Both lost. []
  7. Namely: EB, EVS, ESS, EES, CEVIPOF and TriÉlec 2012 surveys, CNCDH Racism Barometer, Insee Participation Survey, Ined TeO survey, and the 1967 French National Election Survey by Philip E. Converse and Roy Pierce. One might consider adding the ISSP and the Insee EDP (census extract) to the list. []

Guesstimating the Macron presidential candidacy

Un autre billet en anglais, en complément du précédent, au sujet de l’élection présidentielle, et en réaction aux amis qui m’ont fait remarquer que j’étais trop catégorique dans mon refus de voir Emmanuel Macron au second tour.

En bref : j’ai fait « de mon mieux » (si l’on peut dire), mais je ne vois toujours pas Macron au-delà de 7 millions de voix, ce qui n’est suffisant pour atteindre le second tour de l’élection présidentielle qu’en cas d’abstention massive.

My previous post on the 2017 French presidential election was primordially an attempt to provide a precise estimate of the (lower credible bound of the) vote for Marine Le Pen, the candidate for the extreme-right Front national party.

Several friends, however, have focused their reactions to my post on one of my side comments, about the likelihood of “Centrist” candidate Emmanuel Macron to reach the second round of the election. Here are some more thoughts on that.

tl;dr – I am not making any predictions about whether Emmanuel Macron will be on the second round of the French presidential election. Rather, I am arguing that there is no evidence to believe that he will.

In what follows, I examine Macron’s candidacy in light of (1) its “Centrist” (social-liberal) positioning, (2) its relative position within voting choices, and (3) its opportunistic potential, which I believe to be the most correct way to assess its first-round electoral potential.

Update, March 16: added a link to a research note that contains useful information to support my remarks on Macron’s position vis-à-vis the French median voter.

Macron as a “Centrist” candidate

Macron was minister of the economy under incumbent Socialist president François Hollande from 2014 to 2016, but he did not take part in the Socialist primary election. Instead, he has been running since 2016 as an independent candidate with support from various centre-left and centre-right politicians, thus making him some kind of a “Centrist” candidate.

It also seems fair to designate Macron as the “Centrist” candidate in the election because he has received the support of the former historical “Centrist” candidate François Bayrou, who ran for the (centre-right) UDF/MoDem party in the 2002, 2007 and 2012 presidential elections.

The representation of Macron as a “Centrist” candidate is reinforced by the fact that there are no other Centrist candidate in the election, and by the relative positions of left-wing candidate Benoît Hamon and right-wing candidate François Fillon, who are believed to come from the left-wing and right-wing of their respective parties. As a consequence, Macron is currently perceived as a candidate who is closer to the median voter than any of the governing party candidates.

It should be noted that the French presidential election has already been won by a “Centrist” candidate: in 1974, when Valéry Giscard d’Estaing distanced Socialist candidate François Mitterrand by 425,000 votes, in the closest presidential election since the inception of the French Fifth Republic in 1958.

Some observers are reading Macron’s candidacy as a new “Giscardian moment” in French politics, but that view is, in my opinion, highly misguided. In 1974, the French right-wing was officially divided between rival Gaullist and Centrist political parties, leading to an “endogenous switch” from the former to the latter.1 That division does not exist today.

Macron as a potential Condorcet winner

It is much more likely, in my view, that Emmanuel Macron might (just might) be a potential Condorcet winner for some part – not all – of the French electorate. There seems to be empirical support for that hypothesis in some opinion polls, but I would argue that opinion polls are, by design, cognitively biased towards these kinds of scenarios.

If we accept to consider Macron as a winner in the Condorcet paradox, then we accept that the preferences of a determinant majority of voters are of the following form:

P1 : A > B > C

P2 : B > C > A

P3 : C > A > B

… where A and B designate the left-wing and right-wing candidates of the election, and C designates Macron.

It does not matter, for our purposes, which of Hamon or Fillon are denoted by A and B, as the paradox persists regardless: under rank-ordering, there is a set of voting preferences ( P1, P2, P3 ) that eliminates candidate C from a majority vote that retains only the two largest sets of preferences, even though candidate C had initially obtained the largest set of transitive preferences.

Opinion polls are cognitively biased towards that scenario because they do not limit the depth of rank-ordering, whereas a two-round election with only two candidates on the second round does: in effect, the first round of the election does not serve to select A or B, but to eliminate all other candidates. Unless the election proceeds through cardinal voting, all preferences beyond the first choice are irrelevant.

A somewhat more realistic way to poll the electorate of a two-round election would be to ask them which pair of candidates (X, Y) they want to see reach the second round of the election, in which case the question boils down to which candidate they do not want to see get past round one:

P1 : ( A, B ) > { ( A, C ), ( B, C ) } ~ ¬ C

P2 : ( B, C ) > { ( B, A ), ( C, A ) } ~ ¬ A

P3 : ( C, A ) > { ( C, B ), ( A, B ) } ~ ¬ B

I have yet to see any polling institute frame its questions under that kind of “eliminative” voting behaviour. Note that this behaviour scales alright even if one considers more than three candidates, which would be relevant in the context of (at least this instance of) the French presidential election.

Macron as an opportunistic newcomer

Yet another way to frame Macron’s candidacy consists in comparing it to that of Édouard Balladur, a former Prime Minister who ran as an independent right-wing candidate in the 1995 French presidential election, at a time where almost all current politicians were already fighting for ministerial seats and presidential careers. Here are some pictures from Balladur’s government formation in 1993, showing, from left to right, Nicolas Sarkozy, François Fillon and Alain Juppé:

Encouraged by his high approval rates at the end of his dual-executive “cohabitation (divided government) mandate under Socialist president François Mitterrand’s second term, Balladur ran and came out out third on the first round of the election, with 5.6 million votes out of 30.4 million (18.5%). Balladur lost to Socialist Lionel Jospin, who came out first, and to the legitimate right-wing candidate Jacques Chirac, who got 690,000 more votes (roughly 2.3 percentage points) than Balladur.

Balladur and Macron are biographical twins. Both are archetypical French political elites, who studied in the same schools – Sciences Po Paris and the École Nationale d’Administration –, and then worked in the same offices, at the Secretary General of the Presidency and at the Ministry of Economics and Finance. Both have also worked for the private sector, although Balladur worked for the industry while Macron is a former banker.

Balladur and Macron might also be considered to be ideologically close. During his time as Prime Minister, Balladur (unsuccessfully) pushed for an economic reform that would have allowed employers to pay young workers less than the minimum wage. My own interpretation of Macron’s statements on the low-wage jobs offered by Uber puts him very close to that same idea, which has become a staple of social-liberal labour reforms, along with zero-hour contracts:

The caption reads: “The reality is that indeed, [Uber drivers] sometimes work 60 or 70 [weekly] hours to earn the [minimum wage].” In the rest of the interview, Macron argues that some fraction of the minimum wage is better than not working, or better than working in the criminal economy. Watch the full thing.

Last, there is some similarity between Balladur and Macron in terms of their electoral strategy. The similarity stems from the fact that, when Balladur ran for president in 1995, he was supported by many centre-right politicians, as well as by some right-wing dissidents disloyal to the Chirac candidacy. Quite similarly, Macron’s candidacy has gathered support from centre-right politicians, joined by some centre-left Socialists.

This parallel, however, also reveals some critical differences. In 1995, the centre-right was regrouped into a single political party that had not yet split into the two different parties that it has now become, one of which is currently holding to the electoral alliance that it has formed with the right-wing ahead of the forthcoming legislative elections.

Furthermore, and more importantly than the partisan structure itself, while Balladur could ground his presidential candidacy on the landslide electoral victory of the centre-right and right-wing parties in the 1993 legislative elections, Macron has no past event on which to base a tentative electoral coalition between his centre-right and centre-left supports.

That last difference is, to me, a crucial one. In my view, it shows that, beyond their common opportunism, Balladur and Macron are two different kinds of candidates, with only the latter really fitting the description of an opportunistic political newcomer trying to take advantage of low incumbent governmental performance:

Predicting Macron’s first-round score

Let’s now turn to reading Macron’s tea leaves.

While my previous post explored Marine Le Pen’s first-round electoral prospects, the methodology that I used to do so does not apply here, because Macron’s “En marche !” movement did not run in any past “intermediary” election. Instead, what we have at hand are limited similarities (and, in my view, some crucial differences) between Macron and former presidential candidates François Bayrou and Édouard Balladur.

François Bayrou scored 2 million votes (6.8% vote share) in 2002, 6.8 million votes in 2007 (18.5%), and 3.2 million votes in 2012 (9.1%). This puts his best electoral performance at the same vote share as Balladur obtained in 1993, although Balladur reached that percentage by getting 1.2 million votes less.

In the first round of the 2012 presidential election, a vote count of 6.8 million votes would have placed a hypothetical candidate in third place, just above Marine Le Pen. What I suggest is to use that count as the lower bound of a plausible interval for the Macron vote in 2017, and to use a “Bayrou 2007” vote share of 18.5% out of 47 million registered voters with 80% turnout (i.e. 6.95 million votes) as the upper bound of that same interval.

That puts my highest prediction of Macron’s first-round score at almost 7 million votes.

Given how it is computed, this prediction is highly uncertain: for instance, Macron might just as well turn to be a “Jean-Pierre Chèvenement” style of independent presidential candidate, in reference to a former Socialist minister who ran as an independent in the 2002 presidential election and ended up with just above 5% of the vote share. In other words, my lower bound does not preclude the possibility of a Macron “overhyped” candidacy that will eventually fare well below my estimate.

Furthermore, my prediction makes Macron a possible second-round candidate only if the 2017 presidential election makes 18.5% of the vote enough to reach the second round. If you read my previous post, you know that this has been the case only once since 1995, when Jean-Marie Le Pen got on the second round of the 2002 presidential election with roughly 17% of the first-round vote.

On top of that, to reach my upper bound of almost 7 million votes, Macron would need more than twice the electorate that François Bayrou managed to attract on the first round of the previous presidential election. Even if we grant Macron all of Bayrou’s former 2012 voters, that leaves Macron with 3.75 million votes to find among the rest of the electorate.2

That last number represents over a third of the votes that either right-wing candidate Nicolas Sarkozy or left-wing candidate François Hollande each got on the first round of the 2012 presidential election, and still over a third of the right-wing vote in the second round of the last regional elections (which had a much lower turnout rate than we should expect the 2017 presidential election to reach).

Assessing Macron’s candidacy

I can understand why young, brilliant, ambitious and highly connected technocratic elite Emmanuel Macron is running for president, just as I can understand why the French media are loving his independent, “above-the-rest” candidacy. But that does not affect my current prediction that Macron has very little chance to be on the second round of the 2017 presidential election.

Macron is a former banker with exactly zero electoral credentials and no preexisting electoral support. His political movement is not a mass movement, and his media dominance is nowhere near the kind of mass communication that propelled the rise of right-wing moguls like Silvio Berlusconi in Italy or Donald J. Trump.

In order for Macron to reach the second round of the upcoming presidential election, one needs to imagine a mass defection of right-wing and/or left-wing voters—not such an unlikely scenario, given the current state of affairs, but Macron is absolutely not the only candidate that disloyal voters from either side might defect to.

Taking all these parameters into account, I struggle to imagine any reason why Macron would effectively be on the second round. Perhaps his candidacy would benefit from a “mass abstention” scenario that would result in very low turnout, but that is just one more unlikely configuration that defies prediction.3

In brief, and somewhat unoriginally, Macron remains my “blind spot” in this election. He might do “well” in the first round, but to an extent that I do not deem sufficient to bring him to round two.

If you think otherwise, I would need to see some hard facts, articulated into some kind of plausible electoral scenario that does not stray too far ahead from past elections.

As way of full disclosure, I have no personal sympathy for Emmanuel Macron, and I decided not to re-register to vote when he was made Minister of the Economy, which I still consider to be the single most critical strategic mistake ever made by a left-wing French government. I have even less sympathy for Marine Le Pen and the Front national, the main topics of my previous post.

Further reading4

The reason why this post includes a short discussion of Macron as a potential Condorcet winner is because François Bayrou has been called a Condorcet winner before, in the context of the 2007 presidential election. For a detailed account of Bayrou’s 2007 high performance in that election, see this book chapter by Sylvie Strudel and this journal article by Nicolas Sauger.

After researching the issue a bit more, I also got the impression that Macron’s potential as a Condorcet winner might be even higher than Bayrou’s: many left-wing voters seem hesitant about their first-round vote,5 and the spectrum of politicians calling for a Macron vote extends to all sorts of right-wing and left-wing figures, including Greens and even (former or current) Communists, all calling for some kind of “vote utile”.

For a clearer formulation and detailed discussion of the Condorcet voting paradox, I highly recommend reading Dasgupta and Maskin’s “On the Robustness of Majority Rule” (2008):

  1. The “endogenous switch” (alternance endogène) expression comes from Jean-François Sirinelli, La Cinquième République, Presses Universitaires de France, 2009, p. 39. []
  2. Adjusting the Bayrou vote for the difference in electorate size between the two elections brings that figure down by only 200,000 votes. []
  3. See my very last footnote for a comment on why I prefer not to predict anything about turnout and voting uncertainty in this blog post. []
  4. This section was added a few hours after initial publication, based on early comments from my co-blogger Joël and a few other friends. I also added a few more links and changed a few words here and there. []
  5. Since making precise statements about (turnout and) voting uncertainty involves looking at polling data, I plan to finish this series by covering that topic in a third and last blog post. Stay tuned! []

Reading the tea leaves of the French 2017 presidential election

Un billet en anglais, pour les copains et copines qui se demandent, depuis l’étranger, si ce genre d’article, sur la possibilité pour le Front national d’emporter l’élection présidentielle à venir, est justifié. En bref : de mon point de vue, non. J’en profite, du coup, pour publier ma lecture des feuilles de thé électorales de la présidentielle 2017, à onze jours de la validation des dernières candidatures par le Conseil constitutionnel.

Recently, quite a few friends from outside France have been asking me or other friends whether Marine Le Pen, the candidate of the French Front national extreme-right party, has any chance to win the next French presidential election. Here are some thoughts on that.

tl;dr – It’s highly unlikely that the Front national will win the next presidential election, unless it manages to attract a massive amount of (mostly right-wing) voters in the second round.

Let’s start by looking at some raw numbers.

Recent first-round votes

The French electoral system features several two-round elections, and several “intermediary” (i.e. local) elections between two presidential/legislative elections, which have been synced since 2000.

Because the French presidential election has two rounds, the best estimate of how a candidate or party is doing is to look at the last two-round intermediary elections, which were the departmental elections and regional elections of 2015.1

The Front national did very well on the first rounds of both elections. In the départementales, it got 5.1 million votes (25% of the vote), and led the vote in roughly 1 canton out of 6. In the régionales, that figure rose to 6 million votes (27% of the vote), which was also the number of voters that Marine Le Pen got on the first round of the 2012 presidential election.

Lesson No. 1 is therefore that the Front national currently has a very large electoral base. In the first round of the last intermediary elections, that base was just a couple million votes below what governing coalitions got: right-wing coalitions scored 6.8/7.4 million votes, and left-wing coalitions scored 7.4/7.8 million votes (départementales/régionales, first round results only).

This is certainly good reason for concern if you support any other party–or rather, any left-wing or right-wing coalition, because despite what many people seem to think, French elections are won by (second-round) coalitions, not parties.

Which brings me to my next section.

Recent second-round votes

Second-round elections require parties to either find coalition partners, or to attract additional voters. Most of the time, the Front national has to opt for the latter strategy, because all other contending parties will coalesce against it—a strategy known as the cordon sanitaire/front républicain strategy.

There have been, however, many local ‘exceptions’ to the cordon sanitaire, whereby the Front national was able to formally ally with right-wing politicians, and several right-wing ideologues are calling for a “union des droites” (“right-wings union”) between the Front national and any other potential right-wing party. More on that towards the end of this post.

On the second round of the last two intermediary elections, the Front national lost 1 million votes in the départementales, but gained 800,000 voters (+ 13%) in the régionales.

Vote loss in the second round is a common scenario for the Front national in intermediary elections: the same thing happened in all previous similar elections, and in all pre-2015 regional elections. It is, in fact, the common scenario that all parties go through, because of the difference in turnout between intermediary and non-intermediary elections.

The important part here is therefore the result of the second round of the 2015 regional election, in which the Front national made some nontrivial progress to the point of getting more votes than in the 2012 presidential election. Despite that, it failed to win any regional council presidencies, just as it failed to win any departmental council presidencies in the previous 2015 local election.

Lesson No. 2 is therefore that the Front national is currently in capacity to win votes between the two rounds of an election, just as it did in the 2002 presidential election when Jean-Marie Le Pen gained around 700,000 voters (+ 14%) in the second round.

This is another good reason for concern if you support any other party/coalition, although right-wing coalitions enjoyed a much larger increase of (over 3.2 million) votes between the two rounds of the 2015 regional election.

Now let’s look at 2017.

Predicted 2017 candidates

Like many other political spaces, French politics are currently stricken with anti-incumbent syndrome. François Hollande, the incumbent left-wing president, is not running for reelection. Neither will Nicolas Sarkozy, the former president who did only one term right before Hollande.

Instead, the French (left-wing) Socialist Party has elected Benoît Hamon, a candidate who is further away from the French median voter than his former rivals, and who is facing serious competition from far-left candidate Jean-Luc Mélenchon.

Meanwhile, the French (right-wing) Republicans Party has elected François Fillon, a candidate who was already less appealing to right-wing Centrists than many others, and who is currently facing highly damaging legal charges.

Last, the “centrist” candidate of the election, Emmanuel Macron, who started as an outsider even though he was François Hollande’s minister of the economy, is now receiving support from both left-wing and right-wing moderates.

The problem with that setup is that governing party candidates are facing mediocre second-round prospects – even against a Front national candidate, neither Hamon or Fillon can expect much of each other’s voters to vote for them. This state of affairs is leading quite a few observers to consider Macron as the best second-round candidate against Marine Le Pen, but let’s remember that Macron, a former banker with economically liberal views, has never run for election before.

What about Marine Le Pen?

Predicted 2017 first-round vote

How much of the vote will go to the Front national in a few weeks?

Vote shares are very hard to predict under varying turnout rates, which is highly relevant in the case of the French presidential election: in the last four elections, first-round turnout has varied by -3, -7, +12 and -4 percentage points.

Furthermore, the vote shares of the candidates who made it to round 2 have also been highly variable: 21/23 in 1995, 20/17 in 2002, 31/26 in 2007, and 27/27 in 2012. Barring the 2002 vote share of Le Pen, all of those shares are within the 21-31% range, which is already wide—and if you find no good reason to discard Le Pen 2002, the range of vote shares grows to 17-31%, which definitely says something about the difficulty of predicting electoral results.

With these known unknowns in mind, my prediction is that, with – at the very least – 7 million expected votes in the first round of the presidential election, Marine Le Pen is highly likely to make it to the second round of the election, with a vote share of – at the very least – 19%. I get to that vote share by computing

V = F / ( E × T )

where F are the 7 million votes for the Front national in the second round of the regional elections, E is an electoral body of 45.2 million voters (estimated from the 2015 regional election), and T an expected turnout of 80%.

The estimate above is excessively conservative, because it is estimated from a fixed number of Front national voters that is far below the size of its potential electorate on the presidential election. To fix that, let’s have

V = ( F × βT ) / E

where F now stands for the 6 million votes for the Front national in the first round of the regional elections, βT is a multiplier of 1.6 that accounts for the expected difference in turnout (from 50% to 80%), and E is unchanged.

That formula returns a vote share of 21%. Independently, my dear occasional co-blogger Joël has made a (much better informed) first-round prediction of a Front national vote share within the 21-28% range, and other vote specialists seem okay with that estimate.

What tends to make estimates above 20% credible is the current predicament of the right-wing candidate, and the fact that the Front national has been generally considered to perform better on presidential elections than on intermediary ones. Recent intermediary elections, however, might have been highly successful for the Front national precisely because that pattern is less true today.

Predicted 2017 second-round vote

It makes little sense to produce second-round estimates before getting first-round results, but the second-round results of previous intermediary elections are, again, useful to anticipate how the 2017 presidential election will play out.

In the second round of the 2015 regional elections, left-wing and right-wing coalitions respectively scored 8 million and 10 million out of 26.5 million voters, with turnout being slightly below 60%—20 percentage points less than presidential-election second rounds.

If the second round of the presidential election includes either a left-wing or a right-wing candidate, that candidate should therefore be able to get between 9.3 and 13 million voters from his own ranks, assuming that voters have fairly stable preferences and a turnout approaching 80%. Therefore, in order to win the election on her own, Marine Le Pen would need to find at least 30% more voters than she has ever had so far.2

Something pretty unusual would be required for that to happen, and the most likely of those rare events might be a split of the right-wing Republicans party into two factions, one of which would coalesce with the Front national on round 2.3

The likelihood of that last scenario is, in my view, a good estimate of how likely it is that the Front national will be able to win the presidential election later this year. To be clear, I am talking of a really low likelihood, although one that has probably increased as of late.

The likelihood of such an event is low because the frontmen of the French right-wing are just not interested in ruling through an Austrian-style “blue-black” coalition. There would need to be a transformative event in French politics for that situation to change.

Last remarks

My highly unoriginal prediction is that the Front national will be on the second round of the French presidential election, but will then lose to whoever it faces. What is that prediction worth?

Not much, but surely more than the strange error-correction bias that is leading many people to believe that, because the establishment was betting against Brexit and Trump, then surely all bets are off in the French presidential election. Unlike the two major unforeseen outcomes of 2016, the French presidential election has two rounds and can be compared with many past data points. That, alone, should suffice to invalidate what is already a very flawed comparison of events.4

It is, however, correct to observe that were the Front national to win the French presidential election, then what would follow would be some kind of a “Trump-style” transition of power, with the party struggling to build a functional governmental apparatus, by which I mean: forming a stable and loyal legislative majority, and gathering bureaucratic support from top civil servants to run its administration. The Front national currently has none of the resources required to do any of that, and therefore to govern.5

Under the likely scenario of the Front national losing in the presidential election but making important gains in the legislative elections, I will have to update my comment with notes on what to expect next. It will have to do with the strategy of the Front national, which, under Marine Le Pen, has been trying to win enough votes to win on second rounds. That strategy has been successful so far, but there is still a gaping 30% of voters to shift towards the Front national for that strategy to reach its end-goal: the French presidency.

Since that shift might never occur, several Front national figures, such as Marion Maréchal-Le Pen, are instead suggesting to go for local alliances with right-wing Republican dissidents, under the “union des droites” scenario that is already working well in some parts of Southern France. The “Marion strategy” is local, short-term and guaranteed to work to some reasonable extent: in contrast, the global, long-term “Marine strategy” described above might never succeed.6

Further reading

If you can read French, you should be reading Joël’s book on the Front national, or read his media papers: see, e.g., this radio show, these Mediapart articles (subscription required), and two articles published by Le Monde diplomatique: one on the North/South geography of the Front national vote, and one on the front républicain strategy mentioned in this post.

Two related readings that I would like to recommend on populism and populist parties in Europe are Perry Anderson’s essay “Why the system will still win,” published this month in the English edition of Le Monde diplomatique, and Cas Mudde’s 2012 Stein Rokkan lecture, “Three decades of populist radical right parties in Western Europe: So what?

Last, I do not read much from Arthur Goldhammer‘s French Politics blog – there are simply too many daily posts for me to be able to follow –, but I know that he does a stellar job at covering the topic. Following a few of his early-March links, I ended up on the brilliantly titled Arun With A View, and found that post: “Marine Le Pen: can she win?” Make it your next stop for some earlier –and, in my opinion, much wilder– electoral predictions about the 2017 French presidential election.

Update, 7 March 2017: added a few links, removed a few words. You might also want to listen to former Prime Minister Alain Juppé explaining why he will not replace Fillon as the right-wing candidate: see this blog post, this article, and the press conference itself.

  1. There are many kinds of intermediary/local elections, but I will focus only on those that took place in 2015 at the county (canton/département) and regional levels. The Front national performed extremely well in the French EU elections of 2014, but turnout is so low in those elections that I prefer to discard information based on such small numbers of votes. Similarly, I prefer to discard the 2014 municipal elections because I find them to be more “granular” (read: personalised), and would therefore need more data than I care to look at. Last, I will not be looking at by-elections because I do not have the adequate data at hand, but they would be, of course, relevant to paint a more accurate picture of post-2012 voting patterns, and of second-round vote shifts especially. []
  2. That estimate discards the still-likely scenario that Marine Le Pen would be facing Emmanuel Macron in the second round. I do not know what to make of that scenario, since I cannot resolve myself to imagine how Macron would go from having no electoral base to, all of a sudden, having enough votes to get past the first round. If my prediction goes wrong, I expect the error to be related to that blind spot, but in all honesty, I cannot see how a banker – a banker! have you ever talked about bankers with anyone earning less than the median wage? – would magically flip French politics on its head. []
  3. The Republicans Party already hosts both elected officials and voters who are, in effect, indistinguishable from their Front national equivalents. This article about Philippe de Villiers shows how porous the frontier currently is between the French right-wing and far-right, a state of affairs for which there is also tons of non-anecdotal evidence. []
  4. Just consider the British case: Brexit, which never polled so bad as to become a completely unlikely outcome, is only the last of many electoral surprises, if you consider the results of the last general election, or even the nomination of Jeremy Corbyn as leader of the Labour Party. The fact that Britain now operates a multi-party system has likely taken its toll on electoral prediction. If you really want to draw parallels between France and the UK, then you might want to equate Hamon’s primary election to Corbyn’s—but the comparison stops right there. []
  5. This recent op-ed by the French ambassador to Japan very eloquently makes my point. []
  6. On these two strategies, see the excellent documentary “Marine Le Pen, la dernière marche ?” that aired on French TV channel France 3 in January of this year. []

Mon exemple favori d’erreur d’inférence écologique

Ce billet traînait depuis plus d’un an et demi dans les brouillons du blog : le texte date de mai 2015. Je le publie sans retouches, mais vous aurez bien noté que le « peu de temps » dont je parle au début renvoie à cette période, c’est-à-dire aux premiers mois ayant suivi l’attentat contre Charlie-Hebdo.

Il y a peu de temps, Nonna Mayer et Vincent Tiberj ont pris le temps de fournir quelques résultats sérieux sur la sociologie des participants à la « marche républicaine » qui a suivi les attentats de janvier. Leur tribune répond au dernier ouvrage d’Emmanuel Todd, qui s’est basé sur différentes corrélations géographiques pour affirmer que les participants de la « marche » étaient… autant le laisser s’expliquer :

À la suite des travaux de Durkheim sur le suicide, ou de ceux de Max Weber, mon but, c’est de faire comprendre aux gens les valeurs profondes qui les font agir et qui ne sont généralement pas celles qu’ils imaginent. Quand on observe la carte des manifestations, la première chose qui frappe, c’est ce que l’Insee appelle avec élégance la prédominance des « cadres et professions intellectuelles supérieures ». C’est elle qui permet de comprendre l’importance qu’elles ont prise à Paris, Toulouse, Grenoble, etc.

Cette « chose qui frappe » s’enseigne d’ailleurs depuis des décennies en première année de sciences sociales : la participation politique (conventionnelle, comme le vote, ou pas, comme la manifestation) est inégalement distribuée en fonction, notamment, du plus haut niveau de diplôme obtenu. Mais c’est la suite du raisonnement qui contient le ressort argumentaire principal de Todd. Accrochez-vous :

L’autre variable qui, pour moi, d’une certaine manière, est encore plus importante encore, c’est la surmobilisation des vieilles terres issues du catholicisme. Là, il faut que je rappelle ma théorie des deux France, un modèle avec lequel je fonctionne depuis longtemps déjà. D’un côté nous avons la vieille France laïque et républicaine – le Bassin parisien, la façade méditerranéenne, etc. –, la France qui a fait la Révolution en somme. De l’autre, il y a la France périphérique : l’Ouest, une partie du Massif central, la région Rhône-Alpes, la Lorraine, la Franche-Comté. Ce sont les régions qui ont résisté à la Révolution et dans lesquelles l’imprégnation catholique est restée très forte jusqu’au lendemain de la Seconde Guerre mondiale. […]

Eh bien, ce qui a inspiré ma méfiance immédiate, c’est que le 11 janvier, la mobilisation a été du simple au double entre la France de tradition athée et révolutionnaire et cette France périphérique, historiquement antirépublicaine. Ce sont les régions les moins républicaines par le passé qui ont le plus manifesté pour la laïcité, avouez qu’il y a là quelque chose d’étrange. En somme, les bastions ex-catholiques sont les endroits où on a le plus milité pour le droit au blasphème. Si on compare Marseille et Lyon, on voit même que l’intensité des manifestations est du simple au double. Qu’on ne vienne pas me dire dans ces conditions qu’il s’agit de la même laïcité que celle d’hier !1

Nonna Mayer et Vincent Tiberj rappellent ce qui doit inspirer une « méfiance immédiate » dans ce type de raisonnement, en faisant référence à William S. Robinson et à l’inférence écologique, que j’adore illustrer (pour enseigner le paradoxe de Simpson) avec cette slide, qui est d’ailleurs précédée d’un rappel de l’exemple historique de Robinson, qui concernait l’illettrisme2 :


La propension à la mobilisation au niveau régional

Bon, soyons charitable avec Todd. Supposons qu’il ait raison, et qu’une fois le niveau de diplôme pris en compte, la propension à se mobiliser « pour la laïcité » et « pour le droit au blasphème » le 11 janvier soit différente d’une région à l’autre, en rapport à l’héritage catholique régional. Après tout, la réponse que lui font Nonna Mayer et Vincent Tiberj ne dit rien de la géographie des participants.

En revanche, pour que cette différence soit autre chose qu’un artefact, il faut qu’à niveau de diplôme égal et en dehors du contexte du 11 janvier, les habitants des différentes régions françaises aient la même propension à la mobilisation, sinon le lien avec la laïcité se retrouve pris dans un ensemble d’explications possibles au différentiel de mobilisation qu’observe Todd, sans distinction possible sinon de manière subjective.

On va donc prendre les données de l’European Social Survey (ESS) et regarder ça de plus près pour les années 2010 et 2012 (les années précédentes n’utilisent pas les mêmes regroupements géographiques, ce qui rendrait l’analyse plus longue qu’une pause-café). L’une des questions posées dans l’ESS (variable pbldmn) devrait faire l’affaire :

Il existe différents moyens pour essayer d’améliorer les choses en France ou pour empêcher que les choses n’aillent mal. Au cours des 12 derniers mois, avez‐vous fait l’une des choses suivantes ? … Avez‐vous pris part à une manifestation autorisée ?

On va vérifier, en langage Todd où les régions parlent d’elles-mêmes, que les différentes régions françaises répondent « oui » dans la même proportion à cette question. Voici la carte des « réponses régionales », en pourcentage des enquêtés et en tenant compte du schéma de pondération qui fait toute la qualité de l’enquête :


Comme on pouvait s’y attendre, la propension à manifester varie dans l’espace et dans le temps, selon des régularités qui ne s’observent pas au niveau régional : les cartes ne montrent pas les erreurs standard, qui sont conséquentes (certaines régions ne sont « représentées » que par quelques dizaines d’enquêtés), et ne disent rien du profil socio-démographique des résidents, qui varie (aléatoirement) d’une vague de l’enquête à l’autre.

On pourrait s’amuser à aller plus loin, mais pour quoi faire ? Les données ESS ne révèlent qu’une seule chose de manière stable, qui correspond à ce qu’expliquent Nonna Mayer et Vincent Tiberj à partir de données différentes : la propension à la mobilisation est nettement plus faible à droite, et plus nettement forte chez les diplômés (je ne montre pas les coefficients, mais vous pouvez faire tourner le modèle chez vous).

Pour détecter un effet géographique du style de ce que décrit Todd, il n’y a que deux manières de s’y prendre. La première consiste à réunir des données aussi granulaires que possible, de les faire passer dans un modèle multi-niveaux, et d’en tirer des conclusions prudentes sur tel ou tel prédicteur. La seconde consiste à avoir une théorie et beaucoup, beaucoup d’assurance.

  1. La dernière phrase est sublime, n’est-ce pas ? C’est extrait d’un entretien au Nouvel Obs du 30 avril 2015. Le journaliste lui sert d’ailleurs complètement la soupe. Je n’ose pas citer le site Internet sur lequel j’ai récupéré l’extrait. []
  2. Les états américains où l’illettrisme était le plus fort était ceux où il y avait le moins d’immigrés, alors que le fait d’être né à l’étranger est un prédicteur individuel de l’illettrisme. Le paradoxe du vote des riches et des pauvres aux États-Unis est un autre exemple. []

Récupérer les tweets des élections départementales

Au cas où un lecteur ait à disposition une machine qui puisse tourner sans s’arrêter sur les quinze prochains jours, histoire d’avoir une bonne fenêtre de tir…

1. Il vous faut un compte Twitter

À partir de votre compte Twitter, créez une application en vous rendant à cette adresse. Le compte en question doit être lié à un numéro de téléphone. Les autres informations à fournir ne sont pas importantes : donnez n’importe quel nom et description à l’application, cochez la case en dessous des conditions d’utilisation, et vous allez vous retrouver sur une page qui permet de récupérer ces deux informations :

Screen Shot 2015-03-21 at 18.13.58

Il vous faut l’information apparaissant sous les deux barres rouges.

2. Il vous faut le package R ROAuth

Dans R, installez le package ROAuth, puis insérez vos deux codes dans les commandes suivantes :

# install.packages("ROAuth")

requestURL = "https://api.twitter.com/oauth/request_token"
accessURL = "https://api.twitter.com/oauth/access_token"
authURL = "https://api.twitter.com/oauth/authorize"
consumerKey = ""
consumerSecret = ""
my_oauth = OAuthFactory$new(consumerKey = consumerKey,
                            consumerSecret = consumerSecret,
                            requestURL = requestURL,
                            accessURL = accessURL,
                            authURL = authURL)

Une fois ces commandes exécutées, exécutez celle-ci :

my_oauth$handshake(cainfo = system.file("CurlSSL", "cacert.pem", package = "RCurl"))

Votre navigateur va s’ouvrir et vous offrir un code PIN à rentrer dans R :

Screen Shot 2015-03-21 at 18.19.02

Une fois que c’est fait, sauvegardez le résultat :

save(my_oauth, file = "oauth.rda")

Et vous y êtes presque.

3. Il vous faut le package R streamR

Le petit bout de code expliqué ci-dessus traînait sur mon disque dur. Je ne me souviens pas de la source exacte, mais c’était sans doute un tutoriel de Pablo Barbera. Le même a écrit le package streamR, qui va vous permettre de récupérer une sélection de tweets identifiés par les mots-clés et hashtags des élections départementales. Installez le package, puis lancez cette boucle :

# install.packages("streamR")

# load credentials

# scrape for 15 days
for(i in (15 * 24 * 2):1) {

  n = sprintf("%03.0f", i)
  f = paste0("tweets", n, ".json")

  cat(n, "scraping to", f, "...\n")
  filterStream(f, track = c("départementales", "departementales",
                            "#departementales2015", "#departementales",
                            "#départementales2015", "#départementales",
               timeout = 60 * 25, oauth = my_oauth)

  # leave five minutes out
  Sys.sleep(60 * 5)


La boucle produit un fichier toutes les 25 minutes et ignore dix minutes d’information par heure, histoire de laisser l’API de Twitter souffler un peu. Les résultats ne sont pas limités aux tweets en français, mais vous aurez tout le loisir de faire le tri plus tard.

Un petit test effectué ce samedi m’a renvoyé un peu plus de 4 600 résultats de bonne qualité (dont 4 100 en français) en moins de six heures. Il est très facile d’explorer ce genre de données en utilisant, par exemple, environ 700 comptes Twitter d’élus et de partis politiques français, afin de voir s’ils sont souvent retweetés :


Le codage ci-dessus est très incomplet : je n’ai codé que les comptes Twitter des médias les plus évidents (presse quotidienne, chaînes de télévision et quelques médias en ligne), et n’ai inclus dans la catégorie “Officiels” que le seul compte Twitter du Ministère de l’Intérieur. Récoltez les données demain et les jours suivants, et vous aurez un bon matériau pour faire beaucoup mieux.

Qui arrivera à décoder la bombe à retardement symbolique du patrimoine des élus ?

Les décodeurs du Monde constataient récemment que les déclarations de patrimoine des élus, issues du “scandale Cahuzac”, font l’objet d’un dispositif hyper-restrictif, empêchant notamment toute prise de note lors de leur consultation en présence d’un agent préfectoral (voir les articles 1 et 3 de l’arrêté fraîchement publié).

Screen Shot 2014-06-10 at 7.37.23 PM

Aux Pays-Bas, des petits malins ont utilisé une faille juridique pour collecter en temps réel les résultats des dernières élections européennes. Rien ne semblant interdire la lecture des déclarations à haute voix, j’imagine que le dispositif français doit aussi avoir une faille—il y en a toujours au moins une, comme en matière de sécurité informatique1.

Plus sérieusement, le Parlement français a donc décidé de s’en tenir aux effets d’annonce, ce qui fait que pour un citoyen français, il reste plus difficile de s’informer sur les ressources des élus de son propre parlement que d’obtenir la même information pour la Roumanie, qui diffuse les déclarations d’intérêts et de patrimoine de ses élus, des candidats aux élections, et plus généralement de tous les agents financés par de l’argent public, qu’ils soient fonctionnaires ou pas2 :

Screen Shot 2014-06-10 at 7.46.26 PM

La comparaison avec la Roumanie est bien évidemment faussée par la corruption endémique de son appareil d’État, qu’on se doit d’espérer moindre chez les parlementaires français. Mais dans une démocratie comparable, la Grande-Bretagne, le scandale des dépenses parlementaires de 2009 me semble avoir débouché sur quelque chose de légèrement plus pertinent : les parlementaires y bénéficient désormais d’une assistance administrative les aidant à remplir leurs déclarations de revenus et à ne pas s’acheter de nains de jardin avec leur carte bancaire de représentant du peuple3.

Bref, plutôt qu’un risque de hack généralisé du dispositif grâce à une opération massive de désobéissance civile et de crowdsourcing des données, le Parlement français me semble avoir surtout pris le risque d’affronter, un de ces quatre, le même scandale que la Grande-Bretagne—c’est-à-dire un mélange d’antiparlementarisme assez rarement étayé et de dépenses injustifiées mais souvent accidentelles et futiles.

Et dans cette dernière hypothèse, les conséquences électorales seront, à coup sûr, beaucoup plus sévères que celles occasionnées par le scandale Cahuzac, sans même parler des dégâts collatéraux que de tels scandales peuvent engendrer. C’est cette perspective, inquiétante mais à mon sens tout à fait réelle, qui me fait penser que cet arrêté est une petite bombe à retardement symbolique, placée à mi-distance des deux chambres parlementaires4.

Encore un point commun avec la Grande-Bretagne, en quelque sorte.

  1. Histoire d’être bien clair, je n’encourage personne à aller vérifier cette hypothèse. Les sans-papiers savent qu’un agent préfectoral peut obtenir l’intervention des forces de police sur place. []
  2. Le dispositif roumain est né d’une condition posée lors de l’entrée du pays dans l’Union européenne. Apparemment, “deux poids, deux mesures” s’écrit “două greutăți, două măsuri” en roumain. []
  3. Je plaisante à peine : certains députés britanniques se sont fait mettre au pilori par la presse pour avoir acheté des cabanons de jardin ou des canards en plastique pour leur baignoire. []
  4. Ce raisonnement devrait inciter les parlementaires à considérer les associations comme Regards Citoyens de manière moins négative, mais je suis biaisé par mes propres amitiés sur cet exemple. []

Visualisation interactive des cosignatures à l’Assemblée nationale

Suite et fin d’un feuilleton sur les données parlementaires et la visualisation des réseaux de cosignatures, sous-titré : « Tant qu’il y aura des données. » Le code et les données sont accessibles dans ce répertoire, et j’espère bientôt présenter ces slides.

L’an dernier, je suspectais l’arrivée d’un nouveau moteur de recherche pour les amendements de l’Assemblée nationale. Ça y est, il est en ligne, et la mauvaise nouvelle est que la nouvelle interface ne se laisse pas faire et demanderait une réécriture complète de mon code d’extraction des cosignatures.

Ce premier échantillon m’a incité à coder l’extraction des cosignatures sous la forme de fonctions génériques, auxquelles j’ai ensuite passé les mêmes listes de cosignatures issues des amendements (depuis 2002) et des propositions de loi (depuis 1986) de l’Assemblée nationale et du Sénat.

Baptiste Coulmont avait défriché le terrain il y a déjà plusieurs années, et la structure de ces données constituait un bon terrain d’entraînement pour apprendre à manipuler des réseaux pondérés et bipartites, de même que le format des données fut un bon terrain d’entraînement pour coder des fonctions d’extraction un peu complexes, utilisant des données collectées en ligne ou sorties de bases SQL.

Voici quelques résultats sous la forme d’une visualisation interactive qui passe des données et des paramètres produits avec R à la librairie de visualisation Javascript produite par deux membres du Médialab de Sciences Po, sigma.js (cliquez sur les images pour la version complète) :


La visualisation se limite aux propositions de loi de l’Assemblée nationale, qui sont disponibles sur la période la plus longue, couvrant toutes les périodes de cohabitation de la Cinquième République. Elles donnent à voir les cosignatures d’environ 2 000 députés travaillant à ces cosignatures de manière très différenciée selon qu’ils appartiennent à un groupe parlementaire ou à un autre, à droite ou à gauche, dans l’opposition ou dans la majorité1. Les voici géocodés :


L’interface de la visualisation permet de mettre l’accent sur les réseaux égocentrés des députés, ce qui ne sert pas à grand chose d’un point de vue analytique2, mais qui permet de mieux identifier les membres des bureaux chargés de centraliser les cosignatures:

Screen Shot 2014-02-20 at 4.04.34 PM

L’interface affiche quelques informations sur les députés tirées de leurs fiches dans la base Sycomore, ainsi que leurs distances moyennes, soit la « rapidité » avec laquelle on peut relier un député à tous les autres en suivant leurs co-signatures :


En ce qui concerne la construction des graphes, quelques copains assistants parlementaires (merci !) m’ont incité à structurer tous les réseaux collectés de la même manière que James Fowler, en ne retenant que les liens entre les premiers auteurs et leurs cosignataires3. J’ai découpé les données du Sénat sur les mêmes périodes temporelles que l’Assemblée nationale, par législature.

J’ai aussi calculé toute une série de mesures similaires (centralité, modularité, etc.) en tenant compte de la nature pondérée des liens entre co-signataires. Ces mesures sont expliquées en détail dans le papier qui accompagne le code du projet, et qui devrait finir par contenir des modélisations plus avancées des données4.

Avant d’effectuer les mesures, j’ai standardisé les groupes parlementaires en grandes familles sur la période observée pour pouvoir comparer d’une législature à l’autre. Voici par exemple une mesure de centralité sur la législature en cours, qui met en valeur les quelques députés intermédiaires entre les groupes minoritaires et majoritaires, à la fois dans la majorité et dans l’opposition.


Enfin, voici la contribution du groupe parlementaire à la polarisation maximale du réseau, qui mesure l’intensité du travail parlementaire intra-groupe par rapport à l’activité inter-groupe — une manière de mesurer l’importance de la segmentation partisane, déjà visuellement remarquable. Suivant cette mesure, le Parlement français est beaucoup plus discipliné que le Congrès américain, même si le Sénat apparaît systématiquement moins contraint par la discipline de parti et semble même l’être de moins en moins.


Pour produire le même style de visualisation sans trop mettre les mains dans le cambouis Javascript, voici une autre visualisation qui utilise une interface clés-en-main, tous les deux écrits par une équipe de l’Oxford Internet Institute, pour un résultat approchant de mon propre bricolage5 :


  1. Ces variables explicatives apparaissent régulièrement dans les études législatives, notamment celles utilisant les scrutins publics (roll calls). []
  2. Le récent « Que sais-je ? » d’Emmanuel Lazega sur les structures relationnelles l’explique à plusieurs reprises dans les deux premiers chapitres. []
  3. Les bureaux des groupes opèrent fréquemment par cosignature collective sous principe d’opt-in par défaut, ce qui a pour conséquence d’inclure les groupes parlementaires complets ou quasi-complets dans sur les cosignataires. Les biais induits par cette pratique sont minimisés si l’on ne retient que les liens asymmétriques et que l’on pondère les liens par le nombre de cosignataires, ce qui est la stratégie suivie par Fowler, qui peut aussi être normalisée. []
  4. Appel aux dons : si vous avez une heure à perdre au café et que vous savez lire les diagnostics d’un ERGM estimé à partir d’une distribution de Poisson, écrivez-moi maintenant, tout de suite. []
  5. Et pour cause, c’est toujours Sigma.js qui est utilisée en arrière-plan pour dessiner le réseau, avec quelques différences avantageuses comme l’emploi d’un moteur de rendu graphique compatible avec une gamme de navigateurs légèrement plus large. []

Derniers graphiques avant 2014

En faisant tourner à nouveau les scripts de réseaux législatifs dont j’ai déjà parlé plusieurs fois ici, j’ai pu vérifier les résultats précédents sur des données un peu plus propres (merci Regards Citoyens) et étendues à tous les amendements de l’année qui se termine :

Screen Shot 2013-12-30 at 7.55.26 AM

Mais j’ai aussi pu vérifier (cf. la fin de cette note) que l’Assemblée nationale doit s’apprêter à réorganiser son site Internet : entre juillet et décembre 2013, l’index des amendements déposés sous la 13e législature a diminué de 450 à 250 dossiers, et j’ai vérifié, plusieurs d’entre eux sont tout à fait valides (ce n’est pas une correction attribuable à des problèmes de datation, par exemple).

En attendant que ce mystère s’éclaircisse, voici la tête des données actualisées. Les barres grises comptent les amendements individuels, les barres rouges les amendements co-signés à moins de dix, et les barres bleues les amendements co-signés par plus de dix députés, avec des pics de co-signatures vers 200 sur les lois de finance et autres joyeusetés. Le pic gris monstrueux correspond au mariage pour tous.


Je vous épargne un graphique de réseau en plus. Non, je déconne, voici un dernier réseau législatif de l’année, basé sur presque 40 000 amendements déjà co-signés sur la législature en cours. Il me reste à calculer les mesures de centralité et de modularité en tenant compte de la pondération du réseau — c’est en cours.


Meilleurs voeux !

Quantification du travail parlementaire, pour quelques variables en plus

À l’occasion d’un coup de ménage estival, voici une mise à jour prévue depuis longtemps au dossier flegscrap, qui permet de récupérer les documents déposés sous les cinq dernières législatures de l’Assemblée nationale :

[image depots.year.png effacée, désolé ; voir les nouveaux graphes ]

Collecter des données pendant la session législative implique de récupérer des documents déposés le jour même, ce qui donne des comparatifs intéressants où l’on voit, par exemple, le déplacement des amendements sur les articles d’une loi en cours d’examen :

Screen Shot 2013-07-14 at 12.16.51 AM

Les projets et propositions de loi qui s’insèrent sur l’agenda parlementaires apparaissent d’abord dans les sources HTML que le code télécharge, puis rejoignent le jeu de données entre chaque itération, donnant à voir comment les données sont progressivement complétées :

Screen Shot 2013-07-18 at 8.27.59 PM

Les données montrent deux espaces, dominés respectivement par les projets de loi (probabilités de discussion et d’adoption élevées) et les propositions de loi (probabilités de discussion et d’adoption faibles). En rentrant dans le détail, on voit aussi que la distribution des séances de débat en séance publique s’achètent cher (auprès de la conférence des Présidents, si je ne m’abuse).

[ image depots.quadrant.png effacée, désolé ; voir les nouveaux graphes ]

En attendant que le site Internet de l’Assemblée nationale réorganise ses données, voire mieux, en rajoute. Cela ne fera pas disparaître la nécessité de bien nettoyer les données.

Un réseau de 68 203 co-signatures sur 11 569 amendements

J’ai récemment évoqué les co-signatures d’amendements, en me demandant tout du long si le réseau des co-signataires reflétait la même structuration partisane que le réseau des députés sur Twitter. Voici le résultat final sur les 11 500 amendements co-signés depuis le début de la 14e législature en cours :


J’ai rapidement annoté le graphique que m’a sorti ggnet après un calcul de douze minutes pour constituer le réseau1. Pour générer ou actualiser les données, il suffit de récupérer les dossiers parlementaires puis d’en extraire les co-signataires. Le résultat est une gigantesque liste de listes à décomposer en 68 203 paires uniques de co-signataires :


Pour tester mes versions intermédiaires, j’ai utilisé des échantillons réduits. Sur ces échantillons, le groupe GDR est complètement isolé, car il rédige moins d’amendements en valeur absolue et n’en co-signe pas souvent avec des membres d’autres groupes parlementaires. Voici le réseau avec 500 co-signatures sélectionnées au hasard :


Puis, lorsque l’on augmente la “résolution” en prenant un échantillon dix fois plus large de 5 000 amendements, les connexions commencent à apparaître, plaçant le groupe GDR dans l’opposition, en dehors du groupe de “soutien” radical-écologiste placé à côté de la majorité socialiste :


Le réseau place également les élus frontistes dans cette opposition, mais comme il manque quelques données, les points gris ne sont pas tous des élus NI. Tout est perfectible, mais je n’ai pas le temps de corriger ce problème, créé par la présence d’homonymes. J’ai rédigé cette version dans le train et vais m’en tenir à ça.

La force du marqueur partisan est aussi évident que dans le cas des députés sur Twitter : l’immense majorité des liens se tisse à l’intérieur des groupes parlementaires. Cet effet ne révèle toutefois rien d’inconnu, et surtout rien de complet : le travail parlementaire trans-groupe, déjà identifiable dans ces visualisations, est simplement plus discret que ces signaux publics que sont les liens par Twitter ou par co-signature d’amendements2.

Pour terminer, une visualisation de la centralité des points montre les inégalités d’investissement dans le travail de co-signature d’amendements (tout en sachant que ce travail exclut de facto celui, non négligeable en quantité et certainement en temps, sur les amendements déposés au nom d’un seul député) :


Le code et les données sont en ligne, avec les notes de construction et les instructions pour actualiser les données. Il doit être possible, moyennant des retouches mineures au code du scraper, de récupérer les mêmes données pour les législatures précédentes.

Mise à jour : le code permet désormais de travailler sur les deux législatures précédentes et de visualiser les données sous forme de réseaux bipartites, un type de réseau à “deux modes” où l’on représente les acteurs et les événements (ici, les co-signatures d’amendements) sur le même graphe. Ces réseaux sont à l’honneur dans le dernier numéro de Social Networks.

Voici à quoi ressemble un réseau bipartite de la législature en cours, à partir d’un échantillon aléatoire de 5 500 amendements. La stratégie d’obstruction législative dans l’opposition est parfaitement lisible, de même que la division du travail entre les groupes et leurs positions respectives vis-à-vis du nexus majorité/opposition :


Je n’avais pas prévu d’aller vers ce type de visualisation, mais Pedro Jordano m’a convaincu et a fourni le code qui sert à convertir les données entre single-mode et two-mode. En raison de la taille et de la nature du réseau, il faut également penser à le stocker dans une matrice “économe” qui compresse l’espace pris par les tonnes de “zéros” qu’on trouve dans les réseaux bipartites : le code pour cette sparse matrix est venu d’un billet de Solomon Messing sur le sujet.

  1. Ce n’est pas du “big data”, mais il faut quand même aplanir une matrice de plusieurs millions de lignes pour créer le réseau. []
  2. Il suffit de penser, par exemple, aux amendements déposés en commission, exclus des données uniquement nominatives montrées ici. []