Causal machine learning reveals heterogeneous effects of loss versus gain framing on retirement savings

Abstract: As Western countries face rapidly aging populations, insufficient retirement savings threaten individual financial security and the sustainability of social welfare systems. Behavioral interventions such as loss–gain message framing are often proposed as scalable solutions, yet evidence on their effectiveness is mixed, largely because little is known about when and for whom they work. We analyze heterogeneity in the effects of loss–gain framing on real-life, high-stakes retirement saving decisions.

In a large-scale randomized field experiment (\(N = 22919\)) with a European retirement savings company, individuals were randomized to loss- or gain-framed reminders designed to increase participation (extensive margin) and savings amount (intensive margin). Using causal machine learning, we estimate heterogeneous treatment effects across socioeconomic characteristics and validate them in a second, independent experiment (\(N = 18288\)). In this validation experiment, for the 25% most responsive to loss framing, conversion is estimated to rise by 9.4 percentage points relative to gain framing, whereas for the 25% most responsive to gain framing it is expected to fall by 8.4 points.

For savings amount, we estimate that loss framing raises contributions for 64% of contributors (by €737 on average) but backfires for the rest (by €615 on average). Overall, a personalized targeting policy based on causal machine learning is expected to increase participation by 51% (from 4.53% to 6.86%) and average savings by 18% (from €2,062 to €2,442 per converted customer) relative to uniform loss framing. We acknowledge support for the publication cost by the Open Access Publication Fund of Paderborn University.

S.F. acknowledges funding via the Swiss National Science Foundation (SNSF), Grant 186932. Funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under the National Research Data Infrastructure – NFDI 27/2 - 460037581 - BERD@NFDI“. Paderborn University, Paderborn, Germany Munich Center for Machine Learning, Munich, Germany The authors conducted this research in collaboration with a European FinTech company, which deployed the intervention within its production system.

This collaboration was unfunded, and the company had no role in the design of the analysis, the interpretation of the results, or the decision to publish. One author (K.B.) was employed by the FinTech company during the period in which the field experiments were conducted. The authors declare no other competing interests.

The study was conducted in accordance with applicable institutional and national regulations on research ethics and data protection. The research protocol was reviewed and approved by the Institutional Review Board of the University of Cologne (approval no. 230070MW). All participants provided informed consent permitting the FinTech company to collect, process, and use their data for analytics purposes and to contact them with savings-related information.