Papers
Papers
Every paper, working paper included, exactly once.
Working Papers
- 2026
Engineering Social Networks: How Initial Group Assignment Shapes Student Social Interactions
Details
A large literature uses exogenous variation to estimate how assignment to classrooms or other groups shapes social networks. Yet most of these analyses remain dyadic, treating each link in isolation, even though ties often form through triadic closure, as a friend of a friend also becomes a friend. Using fine-grained data on phone calls, text messages, physical co-location, and social-media ties, we estimate the network formation effects of randomly assigning first-year university students to classrooms and to smaller social groups. To analyze explicitly whether group assignment interact with triadic closure, we use our random assignment to estimate a subgraph generated model of network formation. Accounting for triadic closure turns out to be crucial. For social groups in particular, group assignment affects network formation almost entirely by inducing additional triadic closure. Estimates ignoring triadic closure can thus yield misleading predictions about the network effects and benefits of group assignment policies.
- 2026
When Government Uses AI: Experimental Evidence of Selective Trust Decline and Algorithmic Nimbyism
Details
Artificial intelligence is increasingly being used in government, changing the ways in which citizens relate to public institutions. One important aspect, the impact of AI on trust in public institutions, has been the subject of substantial efforts that document an "AI penalty" on trust in public institutions when they use AI. However, little is known about why or how such a penalty arises. In two preregistered survey experiments (N=6 084) we measure the effect of AI on trust across four real-life use cases of governmental AI. First, we find large differences across the four domains—disclosing AI usage in social services and health care settings decreases trust, while trust is unaffected for policing and unemployment use cases. Second, we experimentally test a specific mechanism (algorithmic nimbyism) that considers personal exposure to the decision-making outcomes as a driver of the negative impact of AI on trust. Despite not finding any causal evidence for the notion of algorithmic nimbyism, we used rich administrative data on survey participants’ medical history and labour force participation to explore heterogeneous treatment effects that revealed descriptive and suggestive evidence of algorithmic nimbyism. Future work could consider further exploring the combination of experimental and observational data to empirically examine explanations of the impact AI has on relations between citizens and public institutions.
- 2025
Comparing AI-led to Human-led Chat-based Interviews: Motivations, Initial Results, and Challenges
Details
Chatbots powered by large language models (LLMs) have been proposed as AI-interviewers capable of collecting large-scale qualitative interview data. This paper asks whether data collected through AI-led interviews systematically differs from human-led chat interviews. In a randomized experiment (N=40), participants were assigned to synchronous text-based interviews conducted by either human interviewers or a locally hosted AI system. Human interviewers elicited longer responses per question, whereas AI interviewers conducted longer interviews overall due to faster question delivery. No significant differences in response quality—measured by specificity and relevance—were observed. These findings support the viability of AI-interviewing as a method for large-N qualitative data collection, particularly when implemented with locally hosted, open-weight LLMs that improve ethical standards, data security, and reproducibility.
- 2025
The Origins of Large-Scale Structure in Family Networks
Details
Family relations form the backbone of human social structure, yet the behavioral origins of large-scale family networks remain poorly understood. Leveraging a complete Danish longitudinal registry—including ~6 million individuals and ~8 million parent–child relations—we analyze how individual partner behaviors aggregate to shape network structure. While partner-choice homophily (e.g., similarity by education, geography, age) has been hypothesized to drive network formation, our findings show it plays a minor role. Instead, partner-change behavior—individuals leaving one partner for another—acts like 'shortcuts' in network theory, significantly reducing network path lengths and fostering meso-scale connectivity. Moreover, partner-change is self-exciting: individuals with more prior partners are more likely to change again, producing highly connected hubs. Incorporating these mechanisms into generative models captures key empirical network metrics accurately, whereas homophily-based models fail. This work provides the first quantification of the behavioral micro-foundations of family network structure.
- 2025
Removing Phones from Classrooms Improves Academic Performance
Details
Widespread phone bans are being implemented in classrooms worldwide, yet their causal effects on student outcomes remain unclear. In a randomized controlled trial involving nearly 17,000 students, we find that mandatory in-class phone collection led to higher grades — particularly among lower-performing, first-year, and non-STEM students. Importantly, students exposed to the ban were substantially more supportive of phone-use restrictions, perceiving greater benefits from these policies and displaying reduced preferences for unrestricted access. This enhanced student receptivity to restrictive digital policies may create a self-reinforcing cycle, where positive firsthand experiences strengthen support for continued implementation. Despite a mild rise in reported fear of missing out, there were no significant changes in overall student well-being, academic motivation, digital usage, or experiences of online harassment. Random classroom spot checks revealed fewer instances of student chatter and disruptive behaviors, along with reduced phone usage and increased engagement among teachers in phone-ban classrooms, suggesting a classroom environment more conducive to learning. Spot checks also revealed that students appear more distracted, possibly due to withdrawal from habitual phone checking, yet, students did not report being more distracted. These results suggest that in-class phone bans represent a low-cost, effective policy to modestly improve academic outcomes, especially for vulnerable student groups, while enhancing student receptivity to digital policy interventions.
Coverage: The Economist·The Times·RNZ
- 2024
Reducing annotator bias by belief elicitation
Details
Crowdsourced annotations of data play a substantial role in the development of Artificial Intelligence (AI). It is broadly recognised that annotations of text data can contain annotator bias, where systematic disagreement in annotations can be traced back to differences in the annotators' backgrounds. Being unaware of such annotator bias can lead to representational bias against minority group perspectives and therefore several methods have been proposed for recognising bias or preserving perspectives. These methods typically require either a substantial number of annotators or annotations per data instance. In this study, we propose a simple method for handling bias in annotations without requirements on the number of annotators or instances. Instead, we ask annotators about their beliefs of other annotators' judgements of an instance, under the hypothesis that these beliefs may provide more representative and less biased labels than judgements. The method was examined in two controlled, survey-based experiments involving Democrats and Republicans (n=1,590) asked to judge statements as arguments and then report beliefs about others' judgements. The results indicate that bias, defined as systematic differences between the two groups of annotators, is consistently reduced when asking for beliefs instead of judgements. Our proposed method therefore has the potential to reduce the risk of annotator bias, thereby improving the generalisability of AI systems and preventing harm to unrepresented socio-demographic groups, and we highlight the need for further studies of this potential in other tasks and downstream applications.
- 2024
Trading off performance and human oversight in algorithmic policy: evidence from Danish college admissions
Details
In our paper, we explore the potential of advanced AI models to improve decision-making in higher education admissions. Using a comprehensive dataset from Danish college admissions, we compare the effectiveness of sequential AI models against traditional methods based on high school GPAs or human judgment. Our results demonstrate that AI models offer more accurate and fair predictions of student dropout rates compared to simpler models, even when those models consider protected sociodemographic factors. Interestingly, the AI-driven approach also reveals that certain student profiles are better suited for specific programs, highlighting opportunities for policy-driven efficiency gains. Given the recent Danish nationwide policy to reduce admissions by 10%, we estimate that leveraging AI for admission decisions could lead to significant economic benefits. However, this efficiency comes at a cost: the potential reduction of human oversight and transparency in the admissions process. Our work underscores the delicate balance policymakers must strike between optimizing performance and maintaining accountability when deploying algorithmic systems in public policy.
- 2023
Playing the system: address manipulation and access to schools
Details
Strategic incentives may lead to inefficient and unequal provision of public services. A prominent example is school admissions. Existing research shows that applicants "play the system" by submitting school rankings strategically. We investigate whether applicants also play the system by manipulating their eligibility at schools. We analyze this applicant deception in a theoretical model and provide testable predictions for commonly-used admission procedures. We confirm these model predictions empirically by analyzing the implementation of two reforms. First, we find that the introduction of a residence-based school-admission criterion in Denmark caused address changes to increase by more than 100% before the high-school application deadline. This increase occurred only in areas where the incentive to manipulate is high-powered. Second, to assess whether this behavior reflects actual address changes, we study a second reform that required applicants to provide additional proof of place of residence to approve an address change. The second reform significantly reduced address changes around the school application deadline, suggesting that the observed increase in address changes mainly reflects manipulation. The manipulation is driven by applicants from more affluent households and their behavior affects non-manipulating applicants. Counter-factual simulations show that among students not enrolling in their first listed school, more than 25% would have been offered a place in the absence of address manipulation and their peer GPA is 0.2SD lower due to the manipulative behavior of other applicants. Our findings show that popular school choice systems give applicants the incentive to play the system with real implications for non-strategic applicants.
Coverage: Weekendavisen·Berlingske
- 2022
Statistical inference in social networks: how sampling bias and uncertainty shape decisions
Details
We investigate how individuals form expectations about population behavior using statistical inference based on observations of their social relations. Misperceptions about others' connectedness and behavior arise from sampling bias stemming from the friendship paradox and uncertainty from small samples. In a game where actions are strategic complements, we characterize the equilibrium and analyze equilibrium behavior. We allow for agent sophistication to account for the sampling bias and demonstrate how sophistication affects the equilibrium. We show how population behavior depends on both sources of misperceptions and illustrate when sampling uncertainty plays a critical role compared to sampling bias.
Publications
- 2025
Fairness Over Time: A Nationwide Study of Evolving Bias in Dropout Prediction
Details
The use of student learning data to predict educational outcomes has been widely studied, both in terms of model performance and fairness. An example of these predictive models is the use of Early Warning Systems (EWS), which identify students at risk of dropping out. They can be used continuously to make predictions, from the time of first enrollment until years into a degree program, to provide timely support. However, changes to student composition and their learning trajectories can alter the performance and group fairness of predictions over time. Using a nationwide higher education dataset, we examine changes in the fairness of a dropout prediction model at various points along the academic calendar. Our findings reveal that fairness is not static but evolves over time: the largest differences in AUC occur at 12 months after enrollment, a common evaluation point for dropout EWS. We discuss implications for the continued assessment of fairness in predictive algorithms in education.
- 2025
Unveiling the Social Fabric Through a Temporal, Nation-Scale Social Network and its Characteristics
Details
Social networks shape individuals' lives, influencing everything from career paths to health. This paper presents a registry-based, multi-layer and temporal network of the entire Danish population in the years 2008-2021. Our network maps the relationships formed through family, households, neighborhoods, colleagues and classmates for approximately 7.2 million individuals with more than 1.4 billion relations between them over the course of a decade. We outline key properties of this multiplex network, introducing both an individual-focused perspective as well as a bipartite representation. We show how to aggregate and combine the layers, and how to efficiently compute network measures such as shortest paths in large administrative networks. Our analysis reveals how past connections reappear later in other layers, that the number of relationships aggregated over time reflects the position in the income distribution, and that we can recover canonical shortest path length distributions when appropriately weighting connections. Along with the network data, we release a Python package that uses the bipartite network representation for efficient analysis.
- 2024
Attendance Boundary Policies and the Limits to Combating School Segregation
Details
What is the efficacy of redrawing school attendance boundaries as a desegregation policy? To provide causal evidence on this question we employ novel data with unprecedented detail on the universe of Danish children and exploit changes in attendance boundaries over time. Households defy reassignments to schools with lower socioeconomic status. There is a strong social gradient in defiance, as resourceful households are more sensitive to the student composition of new schools. We simulate school assignment policies and find that boundary changes that reassign areas to a highly disadvantaged school are ineffective at altering the socioeconomic composition at the disadvantaged school.
Coverage: Weekendavisen·Politiken·Politiken·Skolemonitor
- 2023
Nature Exposure is Associated With Reduced Smartphone Use
Details
Evidence links greenspace exposure with restorative benefits to cognition and well-being, yet nature contact is declining for younger demographics. Although natural settings have been shown to restore the capacity to inhibit distracting stimuli, it remains unknown whether smartphone attention capture disrupts nature contact. Here, we analyzed ~2.5 million observations of logged smartphone use, texting, calling, and environmental exposures for 701 young adults over 2 years. Participants' weekly smartphone screen-time was over double their green-time. The relationship between greenspace exposure and smartphone activity differed by exposure dose, type, and mobility state. Calling and texting increased during short recreational greenspace visits while all smartphone use declined over the first 3 hr in nature areas, suggesting that nature exposure may support digital impulse inhibition. Those with elevated baseline screen-time or green-time significantly reduced device use in nature, indicating that parts of the biosphere may provide a reprieve from the cybersphere for highly connected youth.
- 2022
Ethnographic data in the age of big data: How to compare and combine
Details
Big data enables researchers to closely follow the behavior of large groups of individuals by using high-frequency digital traces. However, these digital traces often lack context, and it is not always clear what is measured. In contrast, data from ethnographic fieldwork follows a limited number of individuals but can provide the context often lacking from big data. Yet, there is an under-explored potential in combining ethnographic data with big data and other digital data sources. This paper presents ways that quantitative research designs can combine big data and ethnographic data and account for the synergies that such combinations can provide. We highlight the differences and similarities between ethnographic data and big data, focusing on the three dimensions: individuals, depth of information, and time. We outline how ethnographic data can validate big data by providing a "ground truth" and complement it by giving a "thick description." Further, we lay out ways that analysis carried out using big data could benefit from collaboration with ethnographers, and we discuss the potential within the fields of machine learning and causal inference.
- 2022
Rising temperatures erode human sleep globally
Details
Ambient temperatures are rising worldwide, with the greatest increases recorded at night. Concurrently, the prevalence of insufficient sleep is rising in many populations. Yet it remains unclear whether warmer-than-average temperatures causally impact objective measures of sleep globally. Here, we link billions of repeated sleep measurements from sleep-tracking wristbands comprising over 7 million sleep records (n = 47,628) across 68 countries to local daily meteorological data. Controlling for individual, seasonal, and time-varying confounds, increased temperature shortens sleep primarily through delayed onset, increasing the probability of insufficient sleep. The temperature effect on sleep loss is substantially larger for residents from lower-income countries and older adults, and females are affected more than males. Those in hotter regions experience comparably more sleep loss per degree of warming, suggesting limited adaptation. By 2099, suboptimal temperatures may erode 50–58 h of sleep per person-year, with climate change producing geographic inequalities that scale with future emissions.
Coverage: The Guardian·NPR·CNN·National Geographic·The Washington Post
- 2021
Task-specific information outperforms surveillance-style big data in predictive analytics
Details
Increasingly, human behavior can be monitored through the collection of data from digital devices revealing information on behaviors and locations. In the context of higher education, a growing number of schools and universities collect data on their students with the purpose of assessing or predicting behaviors and academic performance, and the COVID-19–induced move to online education dramatically increases what can be accumulated in this way, raising concerns about students' privacy. We focus on academic performance and ask whether predictive performance for a given dataset can be achieved with less privacy-invasive, but more task-specific, data. We draw on a unique dataset on a large student population containing both highly detailed measures of behavior and personality and high-quality third-party reported individual-level administrative data. We find that models estimated using the big behavioral data are indeed able to accurately predict academic performance out of sample. However, models using only low-dimensional and arguably less privacy-invasive administrative data perform considerably better and, importantly, do not improve when we add the high-resolution, privacy-invasive behavioral data. We argue that combining big behavioral data with "ground truth" administrative registry data can ideally allow the identification of privacy-preserving task-specific features that can be employed instead of current indiscriminate troves of behavioral data, with better privacy and better prediction resulting.
Coverage: Videnskab.dk
- 2020
Assortative matching with network spillovers
Details
This paper investigates endogenous network formation by heterogeneous agents. The agents' types determine the value of linking and we incorporate spillovers as utility from indirect connections. We provide sufficient conditions for a class of networks with sorting to be stable for low to moderate spillovers; with only two types these networks are the unique pairwise stable ones. We also show that this sorting is suboptimal for moderate to high spillovers despite otherwise obeying the conditions for sorting in Becker 1973. This shows that in our sorted networks a tension between stability and efficiency is present. We analyze a policy tool to mitigate suboptimal sorting.
- 2020
The Negative Effect of Smartphone Use on Academic Performance May Be Overestimated: Evidence From a 2-Year Panel Study
Details
In this study, we monitored 470 university students' smartphone usage continuously over 2 years to assess the relationship between in-class smartphone use and academic performance. We used a novel data set in which smartphone use and grades were recorded across multiple courses, allowing us to examine this relationship at the student level and the student-in-course level. In accordance with the existing literature, our results showed that students' in-class smartphone use was negatively associated with their grades, even when we controlled for a broad range of observed student characteristics. However, the magnitude of the association decreased substantially in a fixed-effects model, which leveraged the panel structure of the data to control for all stable student and course characteristics, including those not observed by researchers. This suggests that the size of the effect of smartphone usage on academic performance has been overestimated in studies that controlled for only observed student characteristics.
Coverage: Psychology Today·Information
- 2020
Inferring transportation mode from smartphone sensors: Evaluating the potential of Wi-Fi and Bluetooth
Details
Understanding which transportation modes people use is critical for smart cities and planners to better serve their citizens. We show that using information from pervasive Wi-Fi access points and Bluetooth devices can enhance GPS and geographic information to improve transportation detection on smartphones. Wi-Fi information also improves the identification of transportation mode and helps conserve battery since it is already collected by most mobile phones. Our approach uses a machine learning approach to determine the mode from pre-prepocessed data. This approach yields an overall accuracy of 89% and average F1 score of 83% for inferring the three grouped modes of self-powered, car-based, and public transportation. When broken out by individual modes, Wi-Fi features improve detection accuracy of bus trips, train travel, and driving compared to GPS features alone and can substitute for GIS features without decreasing performance. Our results suggest that Wi-Fi and Bluetooth can be useful in urban transportation research, for example by improving mobile travel surveys and urban sensing applications.
- 2018
Academic performance and behavioral patterns
Details
Identifying the factors that influence academic performance is an essential part of educational research. Previous studies have documented the importance of personality traits, class attendance, and social network structure. Because most of these analyses were based on a single behavioral aspect and/or small sample sizes, there is currently no quantification of the interplay of these factors. Here, we study the academic performance among a cohort of 538 undergraduate students forming a single, densely connected social network. Our work is based on data collected using smartphones, which the students used as their primary phones for two years. The availability of multi-channel data from a single population allows us to directly compare the explanatory power of individual and social characteristics. We find that the most informative indicators of performance are based on social ties and that network indicators result in better model performance than individual characteristics (including both personality and class attendance). We confirm earlier findings that class attendance is the most important predictor among individual characteristics. Finally, our results suggest the presence of strong homophily and/or peer effects among university students.
- 2017
Class attendance, peer similarity, and academic performance in a large field study
Details
Identifying the factors that determine academic performance is an essential part of educational research. Existing research indicates that class attendance is a useful predictor of subsequent course achievements. The majority of the literature is, however, based on surveys and self-reports, methods which have well-known systematic biases that lead to limitations on conclusions and generalizability as well as being costly to implement. Here we propose a novel method for measuring class attendance that overcomes these limitations by using location and bluetooth data collected from smartphone sensors. Based on measured attendance data of nearly 1,000 undergraduate students, we demonstrate that early and consistent class attendance strongly correlates with academic performance. In addition, our novel dataset allows us to determine that attendance among social peers was substantially correlated (>0.5), suggesting either an important peer effect or homophily with respect to attendance.
Resting Working Papers
Not currently in active development. Still available and citable.
- 2023
Why Do Students Lie and Should We Worry? An Analysis of Non-truthful Reporting
Details
A core aspect in market design is to encourage participants to truthfully report their preferences to ensure efficiency and fairness. Our research paper analyzes the factors that contribute to and the consequences of students reporting non-truthfully in admissions applications. We survey college applicants in Denmark about their perceptions of the admission process and personality to examine recent theories of misreporting preferences. Our analysis reveals that omissions in reports are largely driven by students' pessimistic beliefs about their chances of admission. Moreover, such erroneous beliefs largely account for whether an omission led to a missed opportunity for admission. However, the low frequency of these errors suggests that most non-truthful reports are "white lies" with minimal negative impact. We find a novel role of personality and individual circumstances that co-determine the extent of omissions. We also find that estimates of students' demand are biased if it is assumed that students report truthfully, and demonstrate that this bias can be reduced by making a less restrictive assumption. Our results have implications for the modeling of preferences, information acquisition, and subjective admission beliefs in strategy-proof mechanisms.
- 2022
Voluntary Information Disclosure in Centralized Matching: Efficiency Gains and Strategic Properties
Details
Information frictions can harm the welfare of participants in two-sided matching markets. Consider a centralized admission, where colleges cannot observe students' preparedness for success in a particular major or degree program. Colleges choose between using simple, cheap admission criteria, e.g., high school grades as a proxy for preparedness, or screening all applications, which is time-consuming for both students and colleges. To address issues of fairness and welfare, we introduce two novel mechanisms that allow students to disclose private information voluntarily and thus only require partial screening. The mechanisms are based on Deferred Acceptance and preserve its core strategic properties of credible preference revelation, particularly ordinal strategy-proofness. In addition, we demonstrate conditions for which cardinal welfare improves for market participants compared to not screening. Intuitively, students and colleges benefit from voluntary information disclosure if public information about students correlates weakly with students' private information and the cost of processing disclosed information is sufficiently low. Finally, we present empirical evidence from the Danish higher education system that supports critical features of our model. Our work has policy implications for the mechanism design of large two-sided markets where information frictions are inherent.
- 2018
The price of free schools
Details
A central feature of Scandinavian welfare states is the provision of equal access to free primary education. However, if school performance is reflected in property values, economic inequality may diminish equal access. Using highly detailed geographical data for the universe of sales in Denmark in a boundary discontinuity design, we show that property values reflect the socioeconomic composition of student bodies of primary schools in Denmark. Because attendance boundaries change over time we can validate our baseline estimates under less restrictive assumptions and inspect adjustment dynamics. We document that prices begin to adjust immediately and are fully converged within three years. The estimates indicate that our baseline estimates are not inflated by omitted variable bias. Lastly, we calculate that low income households have to forego between 7 and 10 percent of consumption in order to gain access to a socioeconomically strong school. Our findings underline that even when primary school funding is centralized, there are severe obstacles to ensuring equal access to education regardless of parental background.