Sunk Cost Fallacy

What it is

Money already spent is gone. It is gone whatever you do next, which means the only live question is whether the next pound, hour or afternoon is worth spending from where you now stand. Nobody disputes this. Almost nobody behaves as though it were true.

Thaler stated the hypothesis in a single sentence in 1980: paying for the right to use a good or service will increase the rate at which the good is used, all else equal. He called it the sunk cost effect, listed it alongside the endowment effect as one of the anomalies a descriptive theory of consumer choice would have to account for, and offered two examples that have been circulating ever since. A family pays $40 for basketball tickets, drives sixty miles through a snowstorm to use them, and observes in passing that had the tickets been free they would have stayed home. A man pays a $300 annual tennis club membership, develops tennis elbow after a fortnight, and keeps playing in pain because he does not want to waste the $300.

Hal Arkes and Catherine Blumer supplied the experimental programme five years later, in Organizational Behavior and Human Decision Processes. Ten experiments, run at Ohio and Oregon, plus one field study. Their contribution is not the demonstration, which Thaler had already argued for. It is the motive.

The driver they identify is a rule about waste. Not loss, not stubbornness, not attachment to the project. People have learned that being wasteful is bad, they apply that rule with more enthusiasm than accuracy, and they end up honouring an expenditure that no longer exists rather than making the decision in front of them. Two of their experiments support this from unexpected directions. In their fourth study, people in a sunk cost condition did not grimly persist against odds they understood. They inflated their estimate of the project's chance of success, 41.0 against 34.0 on a hundred-point scale. In their tenth, a college economics course made no measurable difference, including for students handed the questionnaire by their own economics professor during class, weeks after covering sunk costs in the textbook and the lectures.

Arkes and Blumer are explicit that the effect cannot be fully subsumed under the social psychological theories available to them, cognitive dissonance included.

Now the part most write-ups omit, which is how it has held up.

It replicates, and the record is unusually good for a psychology finding of this vintage. A sunk cost item was one of the thirteen effects in the original Many Labs project, Klein and colleagues, 2014: 36 samples, 6,344 participants, sixteen measures. It was among the ten effects that replicated consistently. The original effect size was d = 0.23, 95% confidence interval [.04, .50]. The median replication came in at d = 0.32, the unweighted replication at d = 0.31 with a 99% interval of [.22, .39], the weighted at d = 0.27, [.20, .34], aggregate t = 10.83 on N = 6,330. Heterogeneity across samples was not significant, Q = 35.55, df = 35, p = .206, meaning the effect showed up at roughly the same size in Poland and in Virginia.

Read that with the second column in view. The replication effect is slightly larger than the original, which is rare and to the finding's credit. It is also small, and only half of the 36 individual samples reached significance on their own.

It thins out when the money is real. Daniel Friedman and colleagues built a laboratory game in 2007 for the specific purpose of isolating the effect, a treasure-digging task in which subjects paid to move between islands, and varied eleven treatment factors including the visual display and the parameters of the sunk cost itself. Their abstract reports a surprisingly small effect that was generally insensitive to the psychological drivers they had set out to test. George Beknazar-Yuzbashev, Sota Ichiba and Mateusz Stalinski ran two pre-registered studies in 2025. The online one, N = 1,806, varied the price of entry to a fourteen-minute real-effort task by $2 and produced a moderate effect, 0.12 standard deviations, about 1.3 extra minutes of play. The field one ran across 11,328 YouTube videos, varying how long a pre-roll advertisement played before it became skippable, and had no significant effect on engagement overall, because the longer wait drove people away before the video started. Their own conclusion is that the effect exists in a controlled environment and that applying it in policy settings may prove difficult.

The construct is real. It is small, and it does not travel out of the vignette as well as its reputation suggests.