Emmanuel Mamatzakis, Steven Ongena, Pankaj C Patel, Mike Tsionas, A Bayesian policy learning model of COVID-19 non-pharmaceutical interventions, Applied Economics, Vol. 56 (25), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
This article examines the impact of non-pharmaceutical interventions on the initial exponential growth of the infected population and the final exponential decay of the infected population. We employ a Bayesian dynamic model to test whether there is learning, a random walk pattern, or another type of learning with evolving epidemiological data over time across 168 countries and 41,706 country-date observations. Although we show that Bayesian learning is not taking place, most policy measures appear to assert some effect. In particular, we show that economic policy variables are of importance for the main epidemiological parameters derived from the policy learning model. In an empirical second-stage application, we further investigate the underlying dynamics between the epidemiological parameters and household debt repayments, a key economic variable, in the UK. Results show no Bayesian learning, although a higher transmission rate would increase household debt repayments, while the recovery rate would have a negative impact. Therefore, suboptimal learning is taking place. |
|
Michael Blum, Madhav Sachdeva, Yann Stricker, Rudolf Mumenthaler, Jürgen Bernard, Tag-Xplore: Interactive Exploration of Annotation Practices in Digital Editions, In: EuroVis Workshop on Visual Analytics (EuroVA), The Eurographics Association, 2024-05-27. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Digital Editions (DE) are scholarly document collections that make research artifacts accessible to both humans and machines in a structured manner, enriched with annotations. However, the interoperability and reusability of DE can be hampered by annotation inconsistencies within DE and heterogeneous annotation practices across DE. We present Tag-Xplore, an interactive and visual exploration tool for annotation practices within and across DE. Tag-Xplore offers multiple coordinated views that provide both attribute-based and document-based access to the huge search space at multiple granularities. The approach also provides rank, filter, and comparison techniques, to further support the exploration. With Tag-Xplore, data curators can validate assumptions based on existing knowledge and generate new insights about annotation practices. We demonstrate the usefulness of Tag-Xplore with two qualitative case studies on attribute ambiguity and outlier documents |
|
Ruijie Wang, Luca Rossetto, Michael Cochez, Abraham Bernstein, QAGCN: Answering Multi-relation Questions via Single-Step Implicit Reasoning over Knowledge Graphs, In: The 21st Extended Semantic Web Conference (ESWC 2024), Springer, 2024-05-26. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif)
Multi-relation question answering (QA) is a challenging task, where given questions usually require long reasoning chains in KGs that consist of multiple relations. Recently, methods with explicit multi-step reasoning over KGs have been prominently used in this task and have demonstrated promising performance. Examples include methods that perform stepwise label propagation through KG triples and methods that navigate over KG triples based on reinforcement learning. A main weakness of these methods is that their reasoning mechanisms are usually complex and difficult to implement or train. In this paper, we argue that multi-relation QA can be achieved via end-to-end single-step implicit reasoning, which is simpler, more efficient, and easier to adopt. We propose QAGCN — a Question-Aware Graph Convolutional Network (GCN)-based method that includes a novel GCN architecture with controlled question-dependent message propagation for the implicit reasoning. Extensive experiments have been conducted, where QAGCN achieved competitive and even superior performance compared to state-of-the-art explicit-reasoning methods. Our code and pre-trained models are available in the repository: https://github.com/ruijie-wang-uzh/QAGCN. |
|
Alexander Soutschek, Christopher J Burke, Pyungwon Kang, Nuri Wieland, Nick Netzer, Philippe Tobler, Neural reward representations enable utilitarian welfare maximization, Journal of Neuroscience, Vol. 44 (21), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
From deciding which meal to prepare for our guests to trading-off the pro-environmental effects of climate protection measures against their economic costs, we often must consider the consequences of our actions for the well-being of others (welfare). Vexingly, the tastes and views of others can vary widely. To maximize welfare according to the utilitarian philosophical tradition, decision makers facing conflicting preferences of others should choose the option that maximizes the sum of subjective value (utility) of the entire group. This notion requires comparing intensities of preferences across individuals. However, it remains unclear whether such comparisons are possible at all, and (if they are possible) how they might be implemented in the brain. Here, we show that female and male participants can both learn the preferences of others by observing their choices, and represent these preferences on a common scale to make utilitarian welfare decisions. On the neural level, multivariate support vector regressions revealed that a distributed activity pattern in the ventromedial prefrontal cortex (VMPFC), a brain region previously associated with reward processing, represented preference strength of others. Strikingly, also the utilitarian welfare of others was represented in the VMPFC and relied on the same neural code as the estimated preferences of others. Together, our findings reveal that humans can behave as if they maximized utilitarian welfare using a specific utility representation and that the brain enables such choices by repurposing neural machinery processing the reward others receive.Significance statementIn many situations politicians and civilians strive to maximize the welfare of social groups. If the preferences of group members are in conflict, identifying the utilitarian welfare-maximizing option requires that decision makers can compare the strengths of conflicting preferences on a common scale. Yet, there is a fundamental lack of understanding which brain mechanisms enable such comparisons of conflicting utilities. Here, we show that brain regions involved in reward processing compute welfare comparisons by representing the preferences of others with a common neural code. This provides a neurobiological mechanism to compute utilitarian welfare maximization as desired by moral philosophy in the Humean tradition. |
|
Helmut Max Dietl, Markus Lang, Johannes Orlowski, Philipp Wegelin, The effect of the initial distribution of labor-related property rights on the allocative efficiency of labor markets, Frontiers in Behavioral Economics, Vol. 3, 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Introduction
The Coase Theorem posits that frictionless markets efficiently allocate scarce resources as long as property rights are fully specified. Our empirical study investigates how the initial allocation of labor-related property rights influences the allocative efficiency in labor markets for skilled workers within a highly competitive environment—professional basketball. Specifically, we compare two regimes: one where employers can trade workers to other employers without the worker's consent, and another where workers are free agents, able to negotiate and move freely without their employer's consent.
Methods
We utilize the NBA as a “laboratory” to conduct our analysis, constructing a unique panel dataset that includes 3,132 player-season observations spanning 17 regular seasons from 2003/04 to 2019/20. To address our research question, we employ linear panel regression models to analyze the data.
Results and discussion
The findings reveal a decline in productivity among workers who transition to new employers as free agents, a phenomenon not observed among non-free agents. This observation suggests that allocative efficiency might be higher when workers are traded without their consent compared to when they exercise their autonomy as free agents. These findings highlight the significant impact that the initial distribution of labor-related property rights has on labor market efficiency, potentially challenging the assumptions of the Coase Theorem. However, the lack of a statistically significant difference in productivity changes between free agents and non-free agents moving to new employers prevents us from definitively rejecting the predictions of the Coase Theorem. |
|
Narges Ashena, Oana Inel, Badrie L Persaud, Abraham Bernstein, Casual Users and Rational Choices within Differential Privacy, In: 2024 IEEE Symposium on Security and Privacy (SP), Institute of Electrical and Electronics Engineers, Los Alamitos, CA, USA, 2024-05. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
In light of recent growth in privacy awareness and data ownership rights, differential privacy (DP) has emerged as a promising technique employed by several well-known data controller entities. This raises the question of how casual users, as the immediate recipients of privacy threats and risks, comprehend and perceive DP and its key parameter ε, as DP's provided protection depends on it. Existing studies show that ordinary users have the potential to understand the fundamental mechanism of DP and its implications for the privacy-utility trade-off when they are communicated clearly through textual and visual aids and, accordingly, make informed decisions about sharing their data under DP protection. However, these attempts either only implicitly mention a few possible values for ε, such as low, medium, and high, or altogether leave it out of the communication. In this paper, we conduct a between-subject user study (N=426) to investigate the effectiveness of nine interactive visual tools to communicate ε explicitly and on a continuous scale in a data-sharing scenario related to publishing positive COVID-19 test results. These interactive visual tools allow casual users to visualize DP's effects on data accuracy and/or privacy loss for various ε values. We found that visualizations incorporating the privacy loss component have a significant impact on assisting users in selecting values that are closer to the recommended values by experts. However, depending on the ratio between DP noise and underlying data, the accuracy loss component disparately affects users' ε decision; the bigger the relative error, the bigger the selected epsilon and vice versa. Thus, accuracy portrayals should be carried out with care. We contextualize our findings in the existing literature and conclude with insights and recommendations on effectively employing our findings to communicate DP to casual users. |
|
Elyas Meguellati, Lei Han, Abraham Bernstein, Shazia Sadiq, Gianluca Demartini, How Good are LLMs in Generating Personalized Advertisements?, In: WWW '24: The ACM Web Conference 2024, ACM Digital library, 2024-05-13. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
In this paper, we explore the potential of large language models (LLMs) in generating personalized online advertisements (ads) tailored to specific personality traits, focusing on openness and neuroticism. We conducted a user study involving two tasks to understand the performance of LLM-generated ads compared to human-written ads in different online environments. Task 1 simulates a social media environment where users encounter ads while scrolling through their feed. Task 2 mimics a shopping website environment where users are presented with multiple sponsored products side-by-side. Our results indicate that LLM-generated ads targeting the openness trait positively impact user engagement and preferences, with performance comparable to human-written ads. Furthermore, in both scenarios, the overall effectiveness of LLM-generated ads was found to be similar to that of human-written ads, highlighting the potential of LLM-generated personalised content to rival traditional advertising methods with the added advantage of scalability. This study underscores the need for cautious consideration in the deployment of LLM-generated content at scale. While our findings confirm the scalability and potential effectiveness of LLM-generated content, there is an equally pressing concern about the ease with which it can be misused. |
|
Nimra Ahmed, Xindi Liu, Ibrahim Al-Hazwani, Elaine May Huang, Cultural Dimensions and Mental Health Technology: A Systematic Review of Hofstede's Dimensions in Shaping Mental Health Experiences, In: Extended Abstracts of the 2024 CHI Conference on Human Factors in Computing Systems, Association for Computing Machinery, New York, NY, USA, 2024. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif)
This paper explores the influence of cultural factors on mental health help-seeking behaviors and the subsequent implications for the design of mental health technologies. Using Hofstede’s Cultural Dimensions as a framework, we conducted a comprehensive literature review to examine how cultural variations affect patient behaviors in seeking mental health support. This review categorically analyses literature corresponding to each of Hofstede’s five dimensions – Power Distance, Individualism vs. Collectivism, Masculinity vs. Femininity, Uncertainty Avoidance, and Long-Term Orientation. The findings reveal significant cultural influences on help-seeking behaviors, highlighting the need for culturally sensitive approaches in mental health technology design. This study underscores the importance of cultural awareness in the design and deployment of mental health technologies, offering insights for future research and development in this field. |
|
Moyi Li, Dzmitry Katsiuba, Mateusz Dolata, Gerhard Schwabe, Firefighters' Perceptions on Collaboration and Interaction with Autonomous Drones: Results of a Field Trial, In: CHI '24: CHI Conference on Human Factors in Computing Systems, ACM Digital library, 2024-05-11. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Applications of drones in emergency response, like firefighting, have been promoted in the past decade. As the autonomy of drones continues to improve, the ways in which they are integrated into firefighting teams and their impact on crews are changing. This demands more understanding of how firefighters perceive and interact with autonomous drones. This paper presents a drone-based system for emergency operations with which firefighters can interact through sound, lights, and a graphical user interface. We use interviews with stakeholders collected in two field trials to explore their perceptions of the interaction and collaboration with drones. Our result shows that firefighters perceived visual interaction as adequate. However, for audio instructions and interfaces, information overload emerges as an essential problem. The potential impact of drones on current work configurations may involve shifting the position of humans closer to supervisory decision-makers and changing the training structure and content. |
|
Liudmila Zavolokina, Kilian Sprenkamp, Zoya Katashinskaya, Daniel Gordon Jones, Gerhard Schwabe, Think Fast, Think Slow, Think Critical: Designing an Automated Propaganda Detection Tool, In: CHI '24: CHI Conference on Human Factors in Computing Systems, ACM Digital library, 2024-05-11. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
In today’s digital age, characterized by rapid news consumption and increasing vulnerability to propaganda, fostering citizens' critical thinking is crucial for stable democracies. This paper introduces the design of ClarifAI, a novel automated propaganda detection tool designed to nudge readers towards more critical news consumption by activating the analytical mode of thinking, following Kahneman's dual-system theory of cognition. Using Large Language Models, ClarifAI detects propaganda in news articles and provides context-rich explanations, enhancing users' understanding and critical thinking. Our contribution is threefold: first, we propose the design of ClarifAI; second, in an online experiment, we demonstrate that this design effectively encourages news readers to engage in more critical reading; and third, we emphasize the value of explanations for fostering critical thinking. The study thus offers both a practical tool and useful design knowledge for mitigating propaganda in digital news. |
|
Clara-Maria Barth, Jürgen Bernard, Elaine M Huang, "It's like a glimpse into the future": Exploring the Role of Blood Glucose Prediction Technologies for Type 1 Diabetes Self-Management, In: CHI '24: CHI Conference on Human Factors in Computing Systems, ACM Digital library, 2024-05-11. (Conference or Workshop Paper published in Proceedings)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Self-management of type 1 diabetes (T1D) involves multiple factors, frequent anticipation of changes in blood glucose, and complex decision-making. ML-based blood glucose predictions (BGP) may be valuable in supporting T1D management. However, it may be difficult for people with T1D to integrate BGP into their decision-making due to prediction uncertainty and interpretation. In this study, we investigate the lived experience of people with T1D focusing on their needs and expectations in using apps that provide BGP. We designed MOON-T1D, an app that shows simulated BGP and conducted a five-day study using the Experience Sampling Method coupled with semi-structured interviews with 15 individuals with T1D who used MOON-T1D. A reflexive thematic analysis of our data revealed implications for the design and use of BGP, including the complex role of emotions and trust surrounding predictions, and ways in which BGP may ease or complicate T1D management. |
|
Lauren Howe, Steven Shepherd, Nathan B Warren, Kathryn R Mercurio, Troy H Campbell, Expressing dual concern in criticism for wrongdoing: The persuasive power of criticizing with care, Journal of Business Ethics, Vol. 191 (2), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
To call attention to and motivate action on ethical issues in business or society, messengers often criticize groups for wrongdoing and ask these groups to change their behavior. When criticizing target groups, messengers frequently identify and express concern about harm caused to a victim group, and in the process address a target group by criticizing them for causing this harm and imploring them to change. However, we find that when messengers criticize a target group for causing harm to a victim group in this way—expressing singular concern for the victim group—members of the target group infer, often incorrectly, that the messenger views the target group as less moral and unworthy of concern. This inferred lack of moral concern reduces criticism acceptance and prompts backlash from the target group. To address this problem, we introduce dual concern messaging—messages that simultaneously communicate that a target group causes harm to a victim group and express concern for the target group. A series of several experiments demonstrate that dual concern messages reduce inferences that a critical messenger lacks moral concern for the criticized target group, increase the persuasiveness of the criticism among members of the target group, and reduce backlash from consumers against a corporate messenger. When pursuing justice for victims of a target group, dual concern messages that communicate concern for the victim group
as well as the target group are more effective in fostering openness toward criticism, rather than defensiveness, in a target group, thus setting the stage for change. |
|
Suzanne Tolmeijer, Vicky Arpatzoglou, Luca Rossetto, Abraham Bernstein, Trolleys, crashes, and perception - a survey on how current autonomous vehicles debates invoke problematic expectations, AI and Ethics, Vol. 4 (2), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Ongoing debates about ethical guidelines for autonomous vehicles mostly focus on variations of the ‘Trolley Problem’. Using variations of this ethical dilemma in preference surveys, possible implications for autonomous vehicles policy are discussed. In this work, we argue that the lack of realism in such scenarios leads to limited practical insights. We run an ethical preference survey for autonomous vehicles by including more realistic features, such as time pressure and a non-binary decision option. Our results indicate that such changes lead to different outcomes, calling into question how the current outcomes can be generalized. Additionally, we investigate the framing effects of the capabilities of autonomous vehicles and indicate that ongoing debates need to set realistic expectations on autonomous vehicle challenges. Based on our results, we call upon the field to re-frame the current debate towards more realistic discussions beyond the Trolley Problem and focus on which autonomous vehicle behavior is considered not to be acceptable, since a consensus on what the right solution is, is not reachable. |
|
Pablo Koch Medina, Cosimo Munari, Qualitative robustness of utility-based risk measures, Annals of Operations Research, Vol. 336 (1-2), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
We contribute to the literature on statistical robustness of risk measures by computing the index of qualitative robustness for risk measures based on utility functions. This problem is intimately related to finding the natural domain of finiteness and continuity of such risk measures. |
|
Kari A Leibowitz, Lauren Howe, Marcy Winget, Cati Brown-Johnson, Nadia Safaeinili, Jonathan Shaw, Deepa Thakor, Lawrence Kwan, Megan Mahoney, Alia J Crum, Medicine Plus Mindset: A Mixed-Methods Evaluation of a Novel Mindset-Focused Training for Primary Care Teams, Patient Education and Counseling, Vol. 122, 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Objectives
Patient mindsets influence health outcomes; yet trainings focused on care teams’ understanding, recognizing, and shaping patient mindsets do not exist. This paper aims to describe and evaluate initial reception of the “Medicine Plus Mindset” training program.
Methods
Clinicians and staff at five primary care clinics (N = 186) in the San Francisco Bay Area received the Medicine Plus Mindset Training. The Medicine Plus Mindset training consists of a two-hour training program plus a one-hour follow-up session including: (a) evidence to help care teams understand patients’ mindsets’ influence on treatment; (b) a framework to support care teams in identifying specific patient mindsets; and (c) strategies to shape patient mindsets.
Results
We used a common model (Kirkpatrick) to evaluate the training based on participants’ reaction, learnings, and behavior. Reaction: Participants rated the training as highly useful and enjoyable. Learnings: The training increased the perceived importance of mindsets in healthcare and improved self-reported efficacy of using mindsets in practice. Behavior: The training increased reported frequency of shaping patient mindsets.
Conclusions
Development of this training and the study’s results introduce a promising and feasible approach for integrating mindset into clinical practice.
Practice Implications
Mindset training can add a valuable dimension to clinical care and should be integrated into training and clinical practice. |
|
Aleksandra Urman, Mykola Makhortykh, “Foreign beauties want to meet you”: The sexualization of women in Google’s organic and sponsored text search results, New Media & Society, Vol. 26 (5), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
Search engines serve as information gatekeepers on a multitude of topics dealing with different aspects of society. However, the ways search engines filter and rank information are prone to biases related to gender, ethnicity, and race. In this article, we conduct a systematic algorithm audit to examine how one specific form of bias, namely, sexualization, is manifested in Google’s text search results about different national and gender groups. We find evidence of the sexualization of women, particularly those from the Global South and East, in search outputs in both organic and sponsored search results. Our findings contribute to research on the sexualization of people in different forms of media, bias in web search, and algorithm auditing as well as have important implications for the ongoing debates about the responsibility of transnational tech companies for preventing systems they design from amplifying discrimination. |
|
Andreas I Mueller, Damian Osterwalder, Josef Zweimüller, Andreas Kettemann, Vacancy durations and entry wages: evidence from linked vacancy-employer-employee data, Review of Economic Studies, Vol. 91 (3), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
This article explores the relationship between the duration of a vacancy and the starting wage of a new job, using linked data on vacancies, the posting establishments, and the workers eventually filling the vacancies. The unique combination of large-scale, administrative worker, establishment, and vacancy data is critical for separating establishment- and job-level determinants of vacancy duration from worker-level heterogeneity. Conditional on observables, we find that vacancy duration is negatively correlated with the starting wage and its establishment component, with precisely estimated elasticities of −0.07 and −0.21, respectively. While the negative relationship is qualitatively consistent with search-theoretic models where firms use the wage as a recruiting device, these elasticities are small, suggesting that firms’ wage policies can account only for a small fraction of the variation in vacancy filling across establishments. |
|
Anna Scolobig, Maria João Santos, Rémi Willemin, Richard Kock, Stefano Battiston, Owen Petchey, Mario Rohrer, Markus Stoffel, Learning from COVID-19: A roadmap for integrated risk assessment and management across shocks of pandemics, biodiversity loss, and climate change, Environmental Science & Policy, Vol. 155, 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
The COVID-19 pandemic demonstrated the fragility of international, national, regional, and local risk management systems. It revealed an urgent need to improve risk planning, preparedness, and communication strategies. In parallel, it created an opportunity to drastically re-think and transform societal processes and policies to prevent future shocks originating not only from health, but also combined with those related to climate change and biodiversity loss. In this perspective, we examine how to improve integrated risk assessment and management (IRAM) capacities to address interconnected shocks. We present the results from a series of workshops within the framework of the University of Zurich and University of Geneva. Initiative "Shaping Resilient Societies: A Multi-Stakeholder Approach to Create a Responsive Society". This initiative gathered experts from multiple disciplines to discuss their perspectives on resilience; here we present the key messages of the "Pandemics, Climate and Sustainability” thinking group. We identify a roadmap and selected research areas concerning the improvement of IRAM analysis capacities, practices, policies. We recommend the development of robust data systems and science-policy advice systems to address combined shocks emerging from health, biodiversity loss and climate change. We posit that further developing the IRAM framework to include these recommendations will improve societal preparedness and response capacity and will provide more empirical evidence supporting decision-making and the selection of strategies and measures for integrated risk reduction. |
|
Sebastian Ernst, Andreas I Mueller, Johannes Spinnewijn, Risk scores for long-term unemployment and the assignment to job search counseling, AEA Papers and Proceedings, Vol. 114, 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
This paper analyses how risk profiling is used to assign unemployed job seekers to job search counseling in Flanders, Belgium. We compare algorithmic selection to self-selection and selection by job search counselors. We discuss practical challenges for the implementation of risk profiling and highlight avenues for further research. We find that algorithmic assignment is used for only a small fraction of the sample and that job search counselors appear to have valuable private information on job seekers' reemployment prospects beyond what is captured by the algorithmic risk score. |
|
Rainer Winkelmann, Neglected heterogeneity, Simpson’s paradox, and the anatomy of least squares, Journal of Econometric Methods, Vol. 13 (1), 2024. (Journal Article)
![BibTex](/static/css/icons/bibtex.gif) ![PDF](/static/css/icons/pdf.png)
When a sample combines data from two or more groups, multivariate regression yields a matrix-weighted average of the group-specific coefficient vectors. However, it is possible that the weighted average of a specific coefficient falls outside the range of the group-specific coefficients, and it may even have a different sign compared to both group-level coefficients, a manifestation of Simpson’s paradox. The result of the combined regression is then prone to misinterpretation. The purpose of this paper is to raise awareness of this problem and to state conditions under which such non-convex weighting or sign reversal can arise, for a model with two regressors and two groups. Two illustrative examples, an investment equation estimated with panel data, and a cross-sectional earnings equation for men and women, highlight the relevance of these findings for applied work. |
|