Working papers & ongoing projects

Who finishes an AI-administered interview, and how does that group differ from who finishes a standard survey? We report a pre-registered four-arm randomized experiment (N = 1,591) assigning Danish online-panel adults to a standardized survey, a standardized chat, an adaptive AI chat, and an extended adaptive AI chat, all covering same substantive content. Completion falls monotonically from 91.8% to 66.7% across arms. The chat interface drives most of the gap (a 14.6 percentage-point drop), limited adaptive probing adds almost nothing (2.3 percentage-point), and extended probing leads to exhaustion and attrition (8.2 percentage-point). Dropout is socially selective: in the most demanding arm, respondents aged 18 to 34 are more likely to complete than those aged 56 and older, and completers are more educated and more likely to use a computer rather than a smartphone. AI-interview samples are therefore younger, more educated, and more desktop-oriented, with direct consequences for inference of results.

Private hosting of refugees is a high-cost prosocial behaviour: households share their homes for months. Whether it can scale during displacement crises depends on who steps forward, but population-level evidence is scarce because hosts are rarely identifiable in the full population. Using Denmark's linked address, population and housing registers, we identify about 3,900 households that privately hosted Ukrainian refugees after Russia's 2022 invasion and compare them to all other households. Hosting is highly selective along two margins. Ukrainian-linked households are far more likely to host, and workplace exposure to Ukrainians is associated with hosting. Hosts also come disproportionately from larger households in spacious homes. Selection on socioeconomic resources, including owner-occupation, is concentrated among non-Ukrainian hosts, whereas Ukrainian-linked households host across the income distribution. The findings clarify when private hosting can provide flexible civic reception capacity during displacement shocks.

Sudden displacement crises require rapid expansion of refugee accommodation, yet little is known about how different types of initial housing affect integration outcomes. Exploiting linked administrative registers covering the full population of Ukrainian refugees arriving in Denmark in Spring 2022 and a representative survey, we classify each refugee’s initial accommodation from address and co-residence records and track outcomes for 18 months. The majority of arrivals during the peak were absorbed in pop-up shelters (36%) and private hosting (43%). These scalable alternatives to the inelastic supply of conventional refugee housing proved durable, with mean stays of about seven months. Moreover, by leveraging quasi-random assignment generated by within-municipality capacity and time constraints, we estimate effects of accommodation type while conditioning on locality, arrival timing, and sociodemographics. Relative to conventional public housing, private hosting led to higher early employment, higher earnings, persistently lower public-transfer receipt, and improved psychological well-being. Pop-up housing performed at least as well on labor-market outcomes and showed modest gains in social integration. By holding locality constant, we show that how refugees are housed within municipalities has an independent, first-order effect on integration—distinct from the well-studied importance of where they are placed. These findings highlight the potential for civic-led accommodation to complement public systems during displacement shocks and shape long-term refugee trajectories.

This study examines how interface- and adaptiveness changes the quality of data generated from semi-structured interviews. In a randomized controlled trial with approximately 800 to 1200 adults, respondents are asked to reflect on the deservingness of strangers in need using five core items based on the CARIN framework. Participants are assigned to one of four treatment arms, or interview modes: a standard survey, a standardized chat interface, an AI interviewer with one adaptive follow-up, or an AI interviewer with multiple adaptive follow-ups. This design isolates the effects of interface and adaptiveness of the interviewer. Comparing arm 1 and 2, we can estimate the effect of interface, and comparing and 2, 3 and 4 we can estimate the effect of adaptiveness of the interviewer. We create several outcome measures to capture the various factors that impact the overall assessment of the quality of the interview data. We present pre-specified hypotheses we intend to examine and test statistically.

This project examines whether deservingness considerations become more morally salient when individuals reflect on a completed act of helping (Experiential condition) rather than evaluate helping prospectively (Evaluative framing). Drawing on the CARIN framework, which identifies Control, Attitude, Reciprocity, Identity, and Need as core criteria structuring deservingness judgments, the study tests whether moral reasoning shifts when helping becomes self-relevant. We implement a randomized controlled design using AI-administered semi-structured interviews with a Danish sample of 400 to 600 respondents. Participants are randomly assigned to either an evaluative condition, where they consider whether to help a stranger who violates a CARIN principle, or an experiential condition, where they reflect on having helped such a person. Each respondent evaluates all five CARIN dimensions, generating within-person variation across dimensions. Outcomes include a coded measure of the moral importance assigned to each criterion, indicators of injustice and exploitation language, and measures of within-person salience dispersion and decisiveness. The study estimates causal effects of framing on moral reasoning and contributes to research on deservingness, moral emotions, and the situational structure of helping norms.