Relevance and Research Question: In recent years, public opinion researchers have moved to new sources of data, especially data from the online world. For instance, researchers have analyzed the potential of replacing or supplementing survey data with information collected from Twitter, mobile devices and data from other places where people leave traces. Using data from social media instead of survey data may produce biased results, however, as individuals often promote a favorable, yet incomplete picture of their selves on social media. Other records of digital traces, such as browsing histories (i.e., domains visited) may provide a more complete picture of individuals' selves as they do not build on individuals' self-presentation in the online world. Browsing histories may reveal individual attitudes and behaviors because people tend to consume news that reinforce their existing views.

In this paper, we explore the feasibility of using individuals’ online activities to measure political attitudes and behavior. Specifically, we explore the potentials of using browsing histories and app usage to substitute traditional survey data by predicting individuals' political behavior and political attitudes from their online behavior.

Methods and Data: Members of a German commercial non-probability panel gave permission to track their browser and app usage over the four month period leading up to the 2017 German federal election. Panelists also participated in three waves of a panel survey where they were asked questions about the various ways they consume political news and information about politicians in both the offline and online world. We use these survey records to supplement the online behavioral records. Using machine learning methods, we compare predictive performance of various combinations of prediction models containing basic socio-demographic information only and/or models supplemented with records of online behavior.

Results and Added Value: Model performance does not allow the precise prediction of political behaviors and preferences, e.g., to replace survey questions (ROCs reach from .65 to .75). Yet, results may be used to target political campaigns more efficiently. Altogether, our approach allows us to learn whether political behavior and attitudes can be inferred from digital trace data, especially online browsing behavior.

Relevance & Research Question: Response latency para data stemming from online surveys are becoming more readily available and are gaining traction in the literature. We would like to contribute to their use in political science, more specifically in the realms of online survey data obtained before or in our case after a referendum vote. We aim to empirically test whether voters at the extremes of political ideologies such as the well-known economic left-right spectrum or GALTAN dimensions tend to respond faster when asked about their respective vote choice. We interpret lower response latency for vote choice as a an empirical proxy for a more solidified ideologic anchoring of voters.

Methods & Data: Response latency data was directly stored for each of the questions in the online survey automatically. Participants of this post-referendum survey were asked to recall their respective vote choice (a Yes or No question). To a certain degree we can control response latency for sensitivity due to issue salience, personality traits (OCEAN model) but also further socio-demographic and socio-economic variables. The data originate from two mixed-mode surveys (online, mail-in) held in the Canton of Aargau in September 2018 (n online survey = 453) and November 2018 (n = 463).

Results: Results from the September and November 2018 data sets show significant differences in the expected direction in a multivariate OLS regression model with vote choice response latency as the dependent variable. Not surprisingly, we see effects for control variables such as age and education. More interestingly, holding further variables constant, men, and in particular, survey respondents leaning to the political right are portraying shorter response latencies.

Added Value: a) Extending the use of para data to the case of Swiss post-referendum surveys; b) adding an overlooked element helping to further characterize voters at the extremes of the political spectrum; c) raising a new puzzle regarding gender as a variable for the response behavior in (online) surveys.

Relevance & Research Question:

Increasing media fragmentation and algorithmic personalization have led to persistent concerns about “echo chambers” and “filter bubbles” in online information consumption, which some fear could be a cause of increasing polarization in the mass public. However, well-known difficulties with self-selection bias are potentially more severe online. It is thus challenging to assess whether partisan media cause people to become more polarized, or if strong partisans select into congenial media.

Methods & Data:

To address these issues, we designed a pre-registered, randomized field experiment embedded in a nationally representative online panel survey (N = 1,500) in which we encouraged subsets to temporarily alter features of their information environment. Subjects in treatment groups were asked to (i) change their browser homepage, (ii) like a page on Facebook, and (iii) subscribe to newsletters, all corresponding to either Fox News or Huffington Post, two well-known outlets with a decidedly partisan slant. An additional control group received no such encouragement. We then followed up weeks later with post-treatment questions on attitudes, beliefs, and knowledge. Using linked data on respondents’ web visits (i.e. URL-level browsing data), which respondents agreed to provide via software that they installed on their devices, we are able to precisely measure treatment effects among compliers and gauge the extent to which their information-seeking habits were altered over time.

Results:

The field experiment was launched in October 2018. We have published our pre-analysis plan online and will be able to start analyzing the data within the next weeks. Given that our hypotheses and strategy to analyze the data are fixed and documented, we will be able to present detailed results at GOR in March.

Added Value:

We hope that our research will shed light on the power of relatively small “nudges” in online choice architecture to affect people’s media consumption behavior as well as longstanding attitudes and beliefs. We offer an experimental design that allows disentangling selection effects from actual exposure effects. Our panel design setup provides opportunities for meaningful pre-treatment post-treatment comparisons. Finally, the tracking data on respondents' browsing behavior allows measuring information-seeking behavior at previously unknown precision.