Apparent sunk cost effect in rational agents
Ott and colleagues challenge the conclusion of the cross-species sunk cost study on statistical grounds. Their argument is a form of attrition bias. In a task where subjects can quit at any moment, the ones still waiting at a late point are not a random sample of those who started; they are disproportionately the subjects who happened to value that particular trial highly, so the apparent unwillingness to quit after long waiting can be produced by selection rather than by any sensitivity to time already spent. They demonstrate this with a computational model in which a purely rational, reward-maximising agent with no sunk cost mechanism reproduces the published pattern. They also propose a better task design that separates the two explanations. The wider point they draw is the useful one for this vault: analyses of behaviour that condition on subjects who have not yet stopped are vulnerable to this error, which is easy to miss.
Ott and colleagues challenge the conclusion of the cross-species sunk cost study on statistical grounds. Their argument is a form of attrition bias. In a task where subjects can quit at any moment, the ones still waiting at a late point are not a random sample of those who started; they are disproportionately the subjects who happened to value that particular trial highly, so the apparent unwillingness to quit after long waiting can be produced by selection rather than by any sensitivity to time already spent. They demonstrate this with a computational model in which a purely rational, reward-maximising agent with no sunk cost mechanism reproduces the published pattern. They also propose a better task design that separates the two explanations. The wider point they draw is the useful one for this vault: analyses of behaviour that condition on subjects who have not yet stopped are vulnerable to this error, which is easy to miss.
Written from the abstract, OpenAlex.
What our sources record
Reproduced as it was written when this source was checked, figures and all.
Science Advances (2021). doi:10.1126/sciadv.abi7004. Argues attrition bias reproduces the pattern in a reward-maximising agent, and that the Sweis task cannot dissociate sunk costs from ordinary variation in valuation. Abstract only.
Where we use it
- Why You Finish the Series You Stopped Enjoying, Why We Act
Reliability of this reference
The link below was matched against the publisher's record on title, first author and year, and all three agree.
The source
https://doi.org/10.1126/sciadv.abi7004
DOI: 10.1126/sciadv.abi7004