Résumé
In line with academic research asserting the existence of correlations between the analysis of tweets and of other social phenomena offline several studies have suggested that observation of Twitter might help predict election results (O-Connor et al., 2010; Tumasjan et al., 2010; Jungherr, Jürgens, Schoen, 2011). While these new research technologies challenge the predictive claim of surveys and the assumption that opinions are quantifiable, they also introduce two significant epistemological breaks. The first concerns the collect of data and the substitution of a passive position of observation for active questioning. Online opinions are not artificially elicited by a researcher’s questions but rather present themselves as raw material to analyse. This set-up makes it possible to overcome one of the main biases of declarative surveys: the imposition of a problem. The second break concerns the level of analysis. A transition from sampling to a comprehensive analysis of the study population is enabled by the digital nature of the field and the development of software able to process very large amounts of data. However, by giving up population sampling and questioning this type of research has also lost the power to socially place the opinions studied. Researchers are limited to studying the network’s opinion as a whole, without separating the various social subgroups involved in the debate (students and professionals, left-wing and right-wing activists, etc.). In order to overcome this limitation, we propose a hybrid method for measuring opinions on Twitter that combines the advantages of web tracking with those of survey approaches. We created a Twitter panel. This method allow us to analyse the opinions of twitterers without ever interacting with them while being able to place them socially and politically. Our panel include 1,228 political twitterers whose socio-demographic characteristics we identified through a questionnaire. After having established a comprehensive list of French political twitterers and formed a representative sample of accounts from it with a keyword search using the AMI Opinion Tracker software, we collected more than 2,800,000 tweets from 248,628 accounts. We then took a random sample of 20,000 accounts that we tested for eligibility. The account had to meet two criteria: be held by only one individual and be drafted in French. We then administered a short questionnaire including socio-demographic questions to the 10,299 eligible accounts. 608 individuals agreed to answer the questionnaire and be part of the panel. Thereafter we had no more interactions with these respondents. We simply gathered and analysed all of their tweets using our software. We analysed the activity of both the respondent and non-respondent panels for 11 months, from 1 March 2013 to 31 January 2014. During this period the panel posted 840,251 tweets, including 81,606 political tweets (9.7%). Our first key findings are that the political tweet output of our panels is highly concentrated and socially determined on the one hand, and very uneven over time and correlated with the level of politicization of the news on the other hand.