Working papers

(with Andreas Bjerre-Nielsen and Sune Lehmann)

Using a unique dataset covering the entire Danish population (~6 million individuals, 8 million parent–child relations), this study examines how individual behaviors shape the emergent large-scale structure of family networks. Contrary to expectations, partner-choice homophily has only a minor effect. Instead, two key dynamics drive network structure: (1) partner-change behavior acts like ‘shortcuts’ (akin to the small-world model), shortening path lengths and accelerating connectivity, and (2) this behavior is self-exciting—individuals with more prior partners are more likely to change partners further—thus creating highly connected hubs. Including these behaviors in network models accurately reproduces real-world family network properties; models relying on homophily alone do not.

(with Alp Sungu and Pradeep Kumar Choudhury)

This large-scale randomized controlled trial involving nearly 17,000 students tests the impact of removing smartphones from classrooms. Mandatory in-class phone collection improved grades—especially for lower-performing, first-year, and non-STEM students—while reducing disruptions and increasing support for phone-use restrictions. The intervention produced no detectable negative effects on well-being or motivation, making it a promising, low-cost strategy for schools.

(with Semra Yuksel Guven, Tobias Gård­hus, and Hjalmar Bang Carlsen)

This preprint reports results from a randomized experiment (N=40) comparing synchronous text-based interviews conducted by human interviewers versus a locally hosted AI system. Human interviewers elicited longer responses per question, while AI interviewers conducted longer interviews overall by delivering questions more quickly. Crucially, no significant differences were found in response quality—measured by specificity and relevance—suggesting that AI-interviews can produce data quality comparable to human-led interviews. The study also discusses experimental design challenges and advocates for using open-weight, locally hosted language models to enhance data security and research reproducibility.

(Magnus Lindgaard Nielsen, Jonas Skjold Raaschou-Pedersen, Emil Chrisander, David Dreyer Lassen, Julien Grenet, Anna Rogers)

Our study shows that AI models can improve college admissions by predicting dropout rates more accurately than traditional methods. While this boosts efficiency, it raises concerns about reduced human oversight and transparency in policy decisions.

(with Terne Sasha Thorn Jakobsen and Robert Böhm)

This study proposes a simple method to address annotator bias in crowdsourced data without requiring a large number of annotators or instances, by asking annotators to report their beliefs about others’ judgements rather than providing direct judgements. In two experiments involving Democrats and Republicans (n=1,590), this approach consistently reduced bias, offering potential to improve AI generalizability and reduce harm to underrepresented groups.

(with Lykke S. Christensen, Mikkel H. Gandil & Hans H. Sievertsen)

We study the role of manipulation beyond preference to gain admissions to schools and show how the design of admission affects such manipulation and the consequences for admission.

(with Martin B. Busch)
We examine how individuals form expectations about societal behavior through statistical inference and social connections, focusing on the effects of sampling bias and uncertainty. Our study characterizes equilibrium in strategic settings, exploring how sophistication in understanding these biases influences behavior and population dynamics.

(with Emil Chrisander)

We investigate why students in Danish college admissions misreport preferences, finding that omissions are mostly due to pessimistic admission beliefs and influenced by personality, with minimal negative impact. Our findings suggest that assuming truthful reporting biases demand estimates, highlighting implications for modeling in strategy-proof admission mechanisms.

(with Emil Chrisander)

We propose two mechanisms for college admissions that allow voluntary disclosure of private information to address fairness and welfare, demonstrating improved outcomes when public and private student information weakly correlate and processing costs are low, supported by empirical evidence from Denmark.

(with Mikkel H. Gandil)

Scandinavian welfare states provide free primary education, but in this study from Denmark we show property values reflect school socioeconomic status, impacting equal access. Property prices adjust within three years, and low-income families must sacrifice 7-10% of consumption to access better schools, indicating centralized funding alone doesn’t ensure equal access.