Events

Causal machine learning (Causal ML) is an emerging branch in ML/AI research aimed data-driven decision-making by integrating robust causal inference with advanced predictive algorithms. A key advantage of Causal ML is the ability to prediction under intervention, that is, to predict the outcomes of a treatment at the individualized level while adjusting for various confounders. Causal ML can explicitly model how the treatment impact varies across subpopulations, thus uncovering rich, individual-level heterogeneity that can be leveraged for personalized targeting and more effective decisions. In this talk, we explore the methodological foundations of Causal ML, discuss critical guardrails necessary for its rigorous and responsible deployment, and explore applications in behavioral science and policy. In particular, we introduce the “AI Heterogeneity Explorer”, which allows to uncover the differential effectiveness of behavioral interventions and thus identify for whom interventions are effective. The “AI Heterogeneity Explorer” provides a systematic recipe for understanding the heterogeneity of behavioral interventions, optimizing the personalized delivery of interventions, validating the targeting strategy—which offers a powerful alternative to one-size-fits-all approaches often used in data-driven decision-making. Finally, we illustrate how this explorer can be leveraged in the context of climate interventions to advance behavioral and climate science.

Contact person: Malte Toetzke

Subscription to the invitation mailing list and more information on the seminar page.

Venture capitalists, business angels, funding agencies, and incubators evaluate ventures, a difficult task where decision uncertainty is high. We examine how the degree of cognitive uncertainty affects judges’ admission recommendations at an incubator. Judges read an application, use preset criteria to score it, and form an intuitive overall judgment to accept or reject the application. We model how cognitive uncertainty affects this judgment through a Bayesian classification model. We test how judge expertise and venture complexity affect classification accuracy and cognitive uncertainty, the key mechanism of the model that produces different judgments. Judges demonstrate moderate accuracy in evaluating venture quality, performing better than random but with substantial room for improvement. Bayesian models of judgment capture much of the decision process but struggle to fully explain judgments that are less clear-cut. Complexity raises uncertainty and lowers classification accuracy, while expertise reduces uncertainty and improves accuracy; the expertise premium is largest at intermediate complexity levels.

Contact person: Daehyun Kim

Subscription to the invitation mailing list and more information on the seminar page.

PR network meeting of communicators from the Max Planck Institutes in the Munich region exchange of information on the A-HA Science Communication Hub.

Cancer is a leading cause of death in developed countries, and cancer treatments are the top category of pharmaceutical spending in the United States and Europe. This paper assesses (1) whether the use of novel cancer therapies are associated with a reduction in mortality, and (2) the cost per statistical life year saved. Using panel data from 28 countries, we study the relationship between mortality attributed to a specific cancer site and the use of pharmaceutical treatments approved to treat that site. The cross-country and cross-site variation over time allows us to isolate the decline in mortality attributable to new drugs from that due to changes in lifestyle and environmental factors, and we distinguish between the effects of treatments based on their evaluated therapeutic benefits by an important health technology assessor. We correct for the endogeneity of mortality and the availability of new treatments using instrumental variables. On average, our results show a decline in mortality associated with the use of innovative treatments for a cancer site. The gains vary across countries and cancer sites. (Joint work with Pierre Dubois)

Contact person: Elisabeth Hofmeister

Subscription to the invitation mailing list and more information on the seminar page.

Like physical products, new technologies are developed using globally sourced inputs. But while the supply chains behind physical goods are well understood, we know far less about the international ”supply chain” of scientifi c knowledge that powers U.S. innovation—and how vulnerable it may be to disruption. Here, I uncover the structure of the U.S. knowledge supply chain by tracing multi-generational citation paths that connect NSF-funded research to downstream patents, and assess its fragility by simulating barriers that impede the fl ow of scientifi c knowledge across the U.S. border. The results reveal that U.S. innovation is deeply reliant on foreign science: 56% of the intermediary papers linking NSF research to patents are produced outside the United States. Cross-border restrictions reduce the connectivity of these paths, increase their length, and lower innovation productivity, as measured by the U.S. patent-to-publication ratio. Most consequentially, such restrictions strand promising knowledge trajectories outside the U.S.: I estimate there are 104,149 NSF-stimulated paths currently under development outside the U.S. Under the status quo, 67,965 are projected to return to the U.S. for patenting. However, under scientifi c autarky, virtually none would, representing a loss of approximately $10.7 billion in invested capital. These impacts also affect U.S. fi rms that are critical to national priorities, including innovation, energy, and security. For example, autarky reduces outstanding path capture at Microsoft, ExxonMobil, and Lockheed Martin by between 48% and 57%.

Contact person: Daehyun Kim

Subscription to the invitation mailing list and more information on the seminar page.

Open Source AI: Strategic Motives for the Selective Revealing of AI System Components
Presenter: Leonard Hanschur (TUM)
Discussant: Ulrike Morgalla (MPI)

The recent open revealing of numerous AI systems challenges the notion that the exclusivity of an AI system’s data and model constitutes a source of competitive advantage. We explore the mechanisms behind revealing AI and the characteristics of AI systems associated with it. Specifically, we examine two dimensions of the selective revealing of AI systems: Its completeness, describing which components are revealed (none, the model, or model and data), and its degree determined by the license type (proprietary, restrictive, permissive). Employing a mixed-methods approach, we draw on 24 interviews with decision-makers at AI-focused organizations and prior theory to construct hypotheses that we empirically test on a sample of 659 AI systems. We hypothesize, and find supported in the data, that organizations tend to reveal larger and more innovative models to a lesser degree and less completely. Further, we find that data modality shapes revealing completeness, and that model size moderates this association. These findings suggest, in line with our qualitative findings, that revealing AI system components serves to promote their adoption and to establish a lock-in across AI system versions rather than collaborative development. Our study contributes to the academic discourse on open innovation and competitive advantage. For strategists and policymakers, we provide guidance in navigating their pathways toward opening AI.

Strategic Reserves: Shelved Innovation as a Real Option
Presenter: Elisabeth Hofmeister (MPI)
Discussant: Denzel Glandel (ISTO)

I investigate the role of shelved innovation - R&D projects suspended despite promising results - in firms’ strategy, drawing on evidence from the pharmaceutical industry. Initially, I construct a novel dataset linking clinical trials to their published results, enabling the systematic identification of shelved drug development projects. Using the exogenous nature of trial failures, I evaluate whether firms restart shelved projects following project failures in the same market. The results show that firms restart shelved projects in response to late-stage failures in Phase III clinical trials. Further, I find that the decision to restart is moderated by the thickness of the market for technology and the firm’s level of co-specialized complementary assets. Overall, these findings demonstrate that shelved innovation is not merely an incidental byproduct of the R&D process but a strategically managed asset.

Selective Promotion of Complements on Online Auction Platforms: Evidence from the Automotive Industry
Presenter: Alexey Rusakov (ISTO)
Discussant: Tim Hahne (TUM)

Platforms can steer demand by selectively promoting complements in platform markets. But how does selective promotion affect the overall demand when the products are idiosyncratic, such as on auction platforms, and does this effect differ for similar products and a competing platform? By studying a car auction platform with unique cars, we find that promotional car reviews on YouTube positively affect the prices and bid volumes of reviewed cars in the same category. However, the latter effect is largely due to short-term attention spillovers, while the sentiment of the reviews can have unexpected consequences for the bid prices. Moreover, the effect on the competing platform is rather limited and probably only occurs when users first visit the focal platform and then switch to the competing platform.

Contact person: Elisabeth Hofmeister

Program as pdf

How novel ideas are adopted and recognized is crucial to scientific progress, but not all ideas from all groups are equally recognized. This paper studies whether and how increasing gender diversity at universities may lead to greater inclusion and recognition of research traditionally associated with women. Leveraging the transitions to coeducation of 76 all-male universities and novel text-based measures of research content, we show that coeducation led to overall modest shifts toward female-associated research. This aggregate effect masks substantial heterogeneity across fields: in disciplines with higher early female representation, we observe a pronounced increase in female-associated research driven by both existing faculty and new entrants. Male-dominated fields, by contrast, exhibit little change or even declines in female-associated research, primarily due to changes in hiring practices. These findings highlight that while diversity can foster innovation, its effects may only be concentrated in areas already receptive to the new perspectives. (joint work with Drew Hendrickson and Ashley Wong)

Contact person:  Marina Chugunova

Subscription to the invitation mailing list and more information on theseminar page.

On 15/16 December 2025, the Max Planck Institute for Innovation and Competition will host the 8^th Research on Innovation, Science and Entrepreneurship Workshop (RISE8), an annual workshop for Ph.D. students and Junior Postdocs in Economics and Management. 

The goal of the RISE8 Workshop is to stimulate an in-depth discussion of a select number of empirical research papers. It offers Ph.D. students and Junior Postdocs an opportunity to present their work and to receive feedback.

Keynote speaker of the RISE8 Workshop is Matt Marx (Cornell University)

Program RISE8 2025

For more information see RISE Workshop.

How does big data change the search for innovation? Data-driven predictions can identify promising opportunities even when their underlying mechanisms are not understood. This has raised concerns that decoupling innovation from theoretical understanding weakens incentives to develop new theory and yields findings whose consequences are poorly understood. In this paper, I argue that big data can instead catalyze the generation of new theory. Data-driven search broadens the space of combinations explored and increases the variability in outcomes relative to the filtering provided by existing theory. As a result, this search strategy uncovers more empirical anomalies that stimulate, rather than substitute, new theorizing. I test these ideas in the domain of human genetics, where genome-wide association studies (GWAS) operate as a data-driven search for the genetic roots of disease. Compared with traditional theory-based approaches, GWAS introduce gene–disease combinations that span a wider portion of the genetic landscape, more frequently fall at both extremes of scientific quality, and often defy expectations drawn from prior knowledge. Rather than crowding out theory generation, GWAS findings are followed by a surge of research aimed at clarifying their causal mechanisms. Together, the results reveal a complementarity between theory and data in search, suggesting that big data can fuel virtuous cycles of theorizing by accelerating the identification of anomalies.

Contact person: Daehyun Kim

Subscription to the invitation mailing list and more information on the seminar page.