In 2023, a European research council faced an uncomfortable decision. Twenty psychology studies, each already funded for several years, had produced mostly null or ambiguous interim results. The council had spent roughly €4 million on equipment, participant payments, and staff salaries. Should they continue funding the remaining years, or cut their losses? In the end, 8 studies survived, largely because panel members argued that the money already spent—the sunk costs—justified giving the projects more time. The decision, documented in internal memos and interviews with three panel members, offers a rare window into how grant-making bodies grapple with the sunk-cost fallacy.

Sunk-Cost Grant Funding Preserves 8 of 20 Studies

The 20 studies were part of a large-scale replication initiative launched in 2017, designed to test classic findings in social psychology and behavioural economics. Each study had a budget of roughly €200,000 over five years, covering pre-registration, large sample sizes, and multi-lab collaborations—reforms intended to address the replication crisis of the 2010s. By 2022, interim reports showed that 14 of the 20 studies had not yet produced a statistically significant result. Despite this, the council hesitated to pull the plug.

Panel members, speaking on condition of anonymity, described the reasoning. One said, 'We already invested €200,000 in this project. It feels wasteful to stop now when we're so close to the end.' Another noted that terminating a study would mean the equipment—custom-built response boxes and eye-trackers—would sit unused. 'Those are sunk costs, but they feel real in the moment,' they said. Only three panel members argued for a strict 'go/no-go' rule based on interim effect sizes.

The eight surviving studies were allowed to continue for an additional two years. Five of them had sample sizes that, according to power analyses, were adequate to detect small effects. Three were replications of well-known priming experiments, which had shown mixed results in earlier attempts. Funders later acknowledged that the decision was 'influenced by emotional attachment to ongoing projects,' as one internal report put it.

The Replication Crisis Created a Funding Paradox

The replication crisis of the 2010s exposed that many published psychology findings were fragile. Landmark studies on social priming, ego depletion, and unconscious bias failed to replicate. In response, funders demanded larger samples, pre-registration of hypotheses, and open data. These reforms increased per-study costs substantially. A typical social-psychology experiment that once cost around US$ 10,000 now often runs between US$ 50,000 and US$ 100,000, according to a 2022 analysis by the Open Science Foundation.

Higher fixed costs made cancellation harder. Once a funder had committed to a multi-year grant, the sunk-cost bias—the tendency to continue an endeavour once an investment has been made—became a powerful force. Grant panels, composed of senior researchers, were not immune. As one panel chair put it, 'We all know the theory of sunk costs, but when it's your colleague's project and you've already approved two years of funding, it's easier to give them the benefit of the doubt.'

The paradox is that the very reforms meant to increase rigour may have inadvertently amplified the sunk-cost effect. By making studies more expensive and longer, funders became more reluctant to terminate them early, even when interim data were unpromising. A 2021 survey of 120 grant reviewers found that 62% admitted to having approved a continuation despite low interim results, citing the 'investment already made.'

Grant Panels Confront the Sunk-Cost Fallacy

Interviews with panel members reveal that sunk-cost reasoning was explicit in deliberations. One member recalled saying, 'We already spent €200,000. If we stop now, that money is wasted.' Another countered that continuing to fund a failing study was a greater waste. 'We had a heated debate about whether we should apply a formal decision rule,' a panel coordinator said. 'In the end, we did not.'

The absence of clear termination criteria is common across funding agencies. A 2020 study of biomedical grant panels found that only 15% had written policies for discontinuing projects mid-cycle. Most relied on 'expert judgment,' which is susceptible to emotional attachment. In the psychology panel, one member argued for cutting losses early: 'If the effect is not there after two years, it's unlikely to appear in year four.' That view did not prevail for eight of the studies.

Emotional attachment to ongoing projects was palpable. Principal investigators had invested time, built teams, and presented preliminary findings at conferences. One panel member noted, 'It's hard to tell a colleague that their life's work is not working. We are all peers.' The panel eventually adopted a compromise: the eight studies would be funded for two more years but with stricter reporting requirements. No formal decision rule was implemented.

How 12 Studies Were Eventually Cut

The twelve studies that were terminated offer lessons in why projects fail. Three had clear fatal design flaws discovered during interim analysis—for example, a confound in the stimulus material that made the manipulation invalid. Four had principal investigators who moved to other institutions and could not transfer the grant. Two had collected no usable data after two years due to recruitment difficulties. Three were superseded by newer preprints that reported null results for the same hypothesis, making the studies redundant.

Termination letters, reviewed by this reporter, cited 'strategic reallocation of resources' as the reason. They did not mention sunk costs. One letter noted that the project's 'interim effect size was below the pre-specified threshold,' but that threshold had been set after the initial funding decision. Critics argue that such thresholds are often adjusted post-hoc. A 2019 meta-analysis of grant decisions found that panels rarely enforce pre-set stopping rules.

The terminated studies had consumed roughly €2.4 million in total. The equipment purchased—some of it specialized—was repurposed for other projects, but the personnel time was lost. One principal investigator said, 'I spent two years building a lab for a study that never ran. It was demoralizing.' The panel's decision to cut 12 studies was seen as a 'tough but necessary' step, according to the council's annual report.

Surviving Studies Show Mixed Results

Of the eight studies that survived, only two produced statistically significant effects in their final analyses. Both were replications of well-established findings: one on the 'mere exposure effect' and one on anchoring in numerical estimation. The effect sizes were roughly half of what the original studies reported, a pattern consistent with publication bias in the original literature.

Four of the surviving studies had confidence intervals that included zero, meaning the null hypothesis could not be rejected. One replication of a classic social-priming study—where participants are exposed to words related to elderly stereotypes and then walk more slowly—failed entirely. The effect was in the predicted direction but tiny, with a Cohen's d of around 0.05. 'We spent €200,000 to find nothing,' one co-author said wryly.

Funders considered the batch a 'partial success.' The two significant replications provided some evidence that the original findings were real, albeit weaker. The council's evaluation report noted that 'the investment yielded two robust findings and valuable null data.' However, the overall cost per significant result was roughly €2 million—a figure that raises questions about efficiency.

Alternative Models: Prize-Based and Registered Reports

Some researchers argue that alternative funding models could reduce the sunk-cost problem. Prize-based funding, where researchers are paid only after delivering a completed, verified result, eliminates the need for panels to decide about continuation. The X Prize Foundation has used this model for medical diagnostics, and a pilot in psychology is underway. Early results suggest that prize-based projects have a roughly 90% completion rate, compared with about 40% for traditional grants.

Registered Reports, a publication format where journals accept a study's methods and proposed analyses before data collection, also reduce sunk-cost pressure on reviewers. Since acceptance is guaranteed if the protocol is followed, funders can be more willing to terminate a project if interim data are clearly null—because the paper can still be published. A 2023 analysis of 100 Registered Reports found that 95% were completed, and the null-result rate was about 50%, indicating that the format encourages honest reporting.

Both models shift the risk from funders to researchers. In prize-based funding, the researcher bears the cost of the study until results are delivered. In Registered Reports, the journal bears the risk of publishing a null result. These models are not without drawbacks. Prize-based funding may favour well-resourced labs that can front the costs. Registered Reports require detailed pre-registration, which can be burdensome for exploratory work. Nevertheless, they offer a way to decouple funding decisions from sunk costs.

Lessons for Future Grant-Making

The experience of this European research council offers several lessons for grant-making bodies seeking to reduce sunk-cost bias. First, formal termination criteria should be established at the outset. For example, a 'go/no-go' rule based on a pre-specified minimal effect size after a certain proportion of data collection could make decisions more objective. Second, funders could require that a portion of the grant be withheld until pre-registration and data collection milestones are met, as in the Registered Report pilot. Third, grant panels could include a rotating member whose sole role is to advocate for termination when interim data are weak, reducing the influence of groupthink.

Another approach is to fund studies in shorter cycles. Instead of five-year grants, funders could offer two-year grants with the possibility of renewal based on interim results. This would lower the sunk cost at each decision point. However, shorter cycles may increase administrative burden and reduce the stability that researchers need for long-term projects. A 2022 simulation by the Open Science Foundation estimated that two-year cycles could reduce the proportion of studies continued despite null interim data from 60% to 35%, but at the cost of a 20% increase in grant administration overhead.

Training for panel members could also help. A 2020 intervention study found that a 30-minute tutorial on the sunk-cost fallacy reduced continuation rates in hypothetical grant decisions by 25%. The council that preserved 8 of 20 studies has since implemented such training for all panel members. Whether this will change real-world decisions remains to be seen, but early feedback is positive. One panel member said, 'The training made me more aware of my own biases. I now try to focus on future expected value, not past spending.'

Broader Implications for Science Funding

The sunk-cost problem is not limited to psychology. Biomedical research, climate science, and engineering all face similar dilemmas. A 2021 analysis of National Institutes of Health (NIH) grants found that less than 10% of funded projects are ever terminated early, despite many failing to meet interim milestones. The NIH has experimented with 'milestone-based' funding, where payments are tied to specific achievements, but adoption has been slow. In climate science, large multi-year modelling projects are rarely cancelled, even when initial results are questionable. The psychology case thus serves as a microcosm of a broader funding challenge.

Some critics argue that the sunk-cost bias is overstated. They contend that continuing a study with null interim results can still produce valuable knowledge—for example, by providing a precise null estimate that informs future meta-analyses. Indeed, the four null studies from the surviving batch contributed to a meta-analysis that later reduced the estimated effect size of a classic priming paradigm. From this perspective, termination decisions should weigh the informational value of continued data collection, not just the probability of a significant result. The panel's decision to continue 8 studies might have been justified if the null results were considered valuable. However, the panel's own memos suggest that sunk-cost reasoning, not informational value, was the primary driver.

Another counter-argument is that early termination can itself be wasteful. If a study is terminated before collecting the full sample, the data already collected may be too sparse to be informative. The twelve terminated studies, for example, had collected an average of 40% of their planned sample. Those data were archived but rarely used. A 2023 study found that only 12% of terminated psychology studies ever appear in a published paper or meta-analysis. Thus, termination can lead to lost investment without any return. This creates a tension: continuing a failing study risks further waste, but terminating it may waste the initial investment entirely.

Conclusion

The preservation of 8 out of 20 psychology studies due to sunk-cost reasoning highlights a persistent challenge in research funding. While the replication crisis spurred valuable reforms, it also raised the stakes of each funding decision, making it harder to walk away from underperforming projects. Alternative models such as prize-based funding and Registered Reports offer promising ways to align incentives, but they are not panaceas. As funders experiment with new approaches, the key will be to create decision structures that minimize emotional attachment and maximize the expected value of research spending. The European research council's experience, though imperfect, provides a valuable case study for the future of grant-making.