Publications

With a dataset of testing and case counts from over 1,400 institutions of higher education (IHEs) in the United States, we analyze the number of infections and deaths from SARS-CoV-2 in the counties surrounding these IHEs during the Fall 2020 semester (August to December, 2020). We find that counties with IHEs that remained primarily online experienced fewer cases and deaths during the Fall 2020 semester; whereas before and after the semester, these two groups had almost identical COVID-19 incidence. Additionally, we see fewer cases and deaths in counties with IHEs that reported conducting any on-campus testing compared to those that reported none. To perform these two comparisons, we used a matching procedure designed to create well-balanced groups of counties that are aligned as much as possible along age, race, income, population, and urban/rural categories—demographic variables that have been shown to be correlated with COVID-19 outcomes. We conclude with a case study of IHEs in Massachusetts—a state with especially high detail in our dataset—which further highlights the importance of IHE-affiliated testing for the broader community. The results in this work suggest that campus testing can itself be thought of as a mitigation policy and that allocating additional resources to IHEs to support efforts to regularly test students and staff would be beneficial to mitigating the spread of COVID-19 in a pre-vaccine environment.

The public often turns to science for accurate health information, which, in an ideal world, would be error free. However, limitations of scientific institutions and scientific processes can sometimes amplify misinformation and disinformation. The current review examines four mechanisms through which this occurs: (1) predatory journals that accept publications for monetary gain but do not engage in rigorous peer review; (2) pseudoscientists who provide scientific-sounding information but whose advice is inaccurate, unfalsifiable, or inconsistent with the scientific method; (3) occasions when legitimate scientists spread misinformation or disinformation; and (4) miscommunication of science by the media and other communicators. We characterize this article as a “call to arms,” given the urgent need for the scientific information ecosystem to improve. Improvements are necessary to maintain the public’s trust in science, foster robust discourse, and encourage a well-educated citizenry.

Scholars have long documented unequal access to the benefits of science among different groups in the United States. Particular populations, such as low-income, non–white people, and Indigenous people, fare worse when it comes to health care, infectious diseases, climate change, and access to technology. These types of inequities can be partially addressed with targeted interventions aimed at facilitating access to scientific information. Doing so requires knowledge about what different groups think when it comes to relevant scientific topics. Yet data collection efforts for the study of most science-based issues do not include enough respondents from these populations. We discuss this gap and offer an overview of pertinent sampling and administrative considerations in studying underserved populations. A sustained effort to study diverse populations, including through community partnerships, can help to address extant inequities.

Research on complex contagions suggests that individuals need social reinforcement from multiple sources before they are convinced to adopt costly behaviors. We tested the causal foundation of complex contagions in a country-scale viral marketing field experiment. The experiment used a peer encouragement design in which a randomly sampled set of customers were encouraged to share a coupon for a mobile data product with their friends. This experimental design allowed us to test the causal effects of neighboring adopters on the product adoption of their own neighbors. We find causal evidence of complex contagions in viral marketing: contact with one neighboring adopter increases product adoption 3.5-fold, while contact with a second neighbor increases it 4-fold. We also find that social reinforcement crucially depends on the local network structure that supports or constrains peer influences: the more friends two individuals have in common—the more embedded their relationship is—the stronger the effect. While the effect of social reinforcement is quite large, we show that the ability to generate such reinforcement in a realistic setting may be limited when the marketer cannot directly control the messages that customers send to their friends.

Popular online platforms such as Google Search have the capacity to expose billions of users to partisan and unreliable news. Yet, the content they show real users is understudied due to the technical challenges of independently obtaining such data, and the lack of data sharing agreements that include it. Here we advance on existing digital trace methods using a two-wave study in which we captured not only the URLs participants clicked on while browsing the web (engagement), but also the URLs they saw while using Google Search (exposure). Using surveys paired with engagement and exposure data collected around the 2018 and 2020 US elections, we found that strong Republicans engaged with more partisan and unreliable news than strong Democrats did, despite the two groups being exposed to similar amounts of partisan and unreliable news in their Google search results. Our results suggest the search engine is not pushing strong partisans into filter bubbles, but strong Republicans are asymmetrically selecting into echo chambers. These findings hold across both study waves, align with work on social media and web browsing, and provide a rare look at the relationship between exposure and engagement. Our research highlights the importance of users' choices, and our approach moves the field closer to the independent, longitudinal, and cross-platform studies it needs to evaluate the impact of online search and social media platforms.

Given that being misinformed can have negative ramifications, finding optimal corrective techniques has become a key focus of research. In recent years, several divergent correction formats have been proposed as superior based on distinct theoretical frameworks. However, these correction formats have not been compared in controlled settings, so the suggested superiority of each format remains speculative. Across four experiments, the current paper investigated how altering the format of corrections influences people’s subsequent reliance on misinformation. We examined whether myth-first, fact-first, fact-only, or myth-only correction formats were most effective, using a range of different materials and participant pools. Experiments 1 and 2 focused on climate change misconceptions; participants were Qualtrics online panel members and students taking part in a massive open online course, respectively. Experiments 3 and 4 used misconceptions from a diverse set of topics, with Amazon Mechanical Turk crowdworkers and university student participants. We found that the impact of a correction on beliefs and inferential reasoning was largely independent of the specific format used. The clearest evidence for any potential relative superiority emerged in Experiment 4, which found that the myth-first format was more effective at myth correction than the fact-first format after a delayed retention interval. However, in general it appeared that as long as the key ingredients of a correction were presented, format did not make a considerable difference. This suggests that simply providing corrective information, regardless of format, is far more important than how the correction is presented.

Social media data can provide new insights into political phenomena, but users do not always represent people, posts and accounts are not typically linked to demographic variables for use as statistical controls or in subgroup comparisons, and activities on social media can be difficult to interpret. For data scientists, adding demographic variables and comparisons to closed-ended survey responses have the potential to improve interpretations of inferences drawn from social media—for example, through comparisons of online expressions and survey responses, and by assessing associations with offline outcomes like voting. For survey methodologists, adding social media data to surveys allows for rich behavioral measurements, including comparisons of public expressions with attitudes elicited in a structured survey. Here, we evaluate two popular forms of linkages—administrative and survey—focusing on two questions: How does the method of creating a sample of Twitter users affect its behavioral and demographic profile? What are the relative advantages of each of these methods? Our analyses illustrate where and to what extent the sample based on administrative data diverges in demographic and partisan composition from surveyed Twitter users who report being registered to vote. Despite demographic differences, each linkage method results in behaviorally similar samples, especially in activity levels; however, conventionally sized surveys are likely to lack the statistical power to study subgroups and heterogeneity (e.g., comparing conversations of Democrats and Republicans) within even highly salient political topics. We conclude by developing general recommendations for researchers looking to study social media by linking accounts with external benchmark data sources

The backfire effect is when a correction increases belief in the very misconception it is attempting to correct, and it is often used as a reason not to correct misinformation. The current study aimed to test whether correcting misinformation increases belief more than a no-correction control. Furthermore, we aimed to examine whether item-level differences in backfire rates were associated with test-retest reliability or theoretically meaningful factors. These factors included worldview-related attributes, namely perceived importance and strength of pre-correction belief, and familiarity-related attributes, namely perceived novelty and the illusory truth effect. In two nearly identical experiments, we conducted a longitudinal pre/post design with N = 388 and 532 participants. Participants rated 21 misinformation items and were assigned to a correction condition or test-retest control. We found that no items backfired more in the correction condition compared to test-retest control or initial belief ratings. Item backfire rates were strongly negatively correlated with item reliability (⍴ = -.61 / -.73) and did not correlate with worldview-related attributes. Familiarity-related attributes were significantly correlated with backfire rate, though they did not consistently account for unique variance beyond reliability. While there have been previous papers highlighting the non-replicable nature of backfire effects, the current findings provide a potential mechanism for this poor replicability. It is crucial for future research into backfire effects to use reliable measures, report the reliability of their measures, and take reliability into account in analyses. Furthermore, fact-checkers and communicators should not avoid giving corrective information due to backfire concerns.

Topic model evaluation, like evaluation of other unsupervised methods, can be contentious. However, the field has coalesced around automated estimates of topic coherence, which rely on the frequency of word co-occurrences in a reference corpus. Contemporary neural topic models surpass classical ones according to these metrics. At the same time, topic model evaluation suffers from a validation gap: automated coherence, developed for classical models, has not been validated using human experimentation for neural models. In addition, a meta-analysis of topic modeling literature reveals a substantial standardization gap in automated topic modeling benchmarks. To address the validation gap, we compare automated coherence with the two most widely accepted human judgment tasks: topic rating and word intrusion. To address the standardization gap, we systematically evaluate a dominant classical model and two state-of-the-art neural models on two commonly used datasets. Automated evaluations declare a winning model when corresponding human evaluations do not, calling into question the validity of fully automatic evaluations independent of human judgments.

Science rarely proceeds beyond what scientists can observe and measure, and sometimes what can be observed proceeds far ahead of scientific understanding. The twenty-first century offers such a moment in the study of human societies. A vastly larger share of behaviours is observed today than would have been imaginable at the close of the twentieth century. Our interpersonal communication, our movements and many of our everyday actions, are all potentially accessible for scientific research; sometimes through purposive instrumentation for scientific objectives (for example, satellite imagery), but far more often these objectives are, literally, an afterthought (for example, Twitter data streams). Here we evaluate the potential of this massive instrumentation—the creation of techniques for the structured representation and quantification—of human behaviour through the lens of scientific measurement and its principles. In particular, we focus on the question of how we extract scientific meaning from data that often were not created for such purposes. These data present conceptual, computational and ethical challenges that require a rejuvenation of our scientific theories to keep up with the rapidly changing social realities and our capacities to capture them. We require, in other words, new approaches to manage, use and analyse data.