Sunk cost fallacy

Arkes and Blumer (1985) named the mechanism the desire not to appear wasteful. Stopping converts a prior spend into an admitted waste, which is aversive; continuing keeps the earlier decision looking sensible. The management literature calls the same thing self-justification, after Staw, and the lab version is not merely a tolerance for a worse bet: in the same paper's Experiment 4, respondents who had already sunk $10 million into a project rated its chance of financial success at 41.0 out of 100 (n = 76), against 34.0 among respondents shown the identical project with nothing spent (n = 82), t(156) = 2.02, p < .05. People who have paid revise the forecast, not just the decision.

Two things sit next to the mechanism and are routinely mistaken for it. The first is the endowment effect — once something is yours you value it more. Ronayne, Sgroi and Tuckwell (2021) separated the two inside one 528-subject design: 23% of the 268 subjects who earned a lottery through a real-effort task (counting letters in blocks of Latin text) declined to switch to a strictly dominant one paying $10 on a 20% rather than a 10% chance, against 7% of the 197 who were simply handed the inferior lottery, and none of the 25 offered the two as a straight choice. The difference in proportions was 0.16, p < .001, which leaves the endowment effect accounting for approximately only a third of the overall effect. The second confusable is proximity to completion, which the escalation literature calls goal substitution: as a project nears its end, finishing becomes a goal in its own right, an argument Sleesman and colleagues trace to Conlon and Garland (1993). This one matters enormously for product work, because the meta-analytic record suggests completion proximity is doing much of the work usually credited to sunk cost.

For an interface this distinction is the whole job. There is a defensible pattern — reflecting value a user genuinely accumulated, so they can see what they have and decide well — and there is the fallacy, which is a reasoning error you would be exploiting rather than serving. The first is closer to endowed progress and the goal gradient effect; where the user built a specific artifact the IKEA effect governs, and where the question is liking rather than continuing, effort justification does. The second is the thing cancellation flows do when they recite how much you have spent. Nothing in the literature licenses inflating switching costs; the evidence is that people already overweight their own past spend without your help.

The baseline field demonstration is old, direct and small. Arkes and Blumer sold season tickets to the Ohio University Theater's 1982–1983 season at the normal $15, a $2 discount or a $7 discount, the ticket types randomly ordered beforehand and the discount revealed only after the buyer had announced an intention to buy; six people buying as couples were dropped, leaving 18, 19 and 17 subjects. Across the first five plays the no-discount group used significantly more tickets (4.11) than the $2 group (3.32) or the $7 group (3.29), one-tailed t = 1.79 and 1.83, p < .05. Read the second half of the season before you quote this as a retention lever: use was 2.28, 1.54 and 2.18 and the groups did not differ. The only significant source of variance in the full analysis was season half, F(1, 51) = 32.32, p < .001 — 3.57 tickets per subject over the first five plays against 2.09 over the last five. The authors' own summary is that the effect lasted for six months following the purchase and was not manifested six to nine months after it.

The scenario evidence is far larger, which should make you suspicious of scenario evidence rather than confident about the effect. In Experiment 3, 41 of 48 respondents (85%) endorsed spending the last 10% of a research budget to finish a radar-blank plane that a competitor had already beaten to market, while only 10 of 60 (17%) endorsed starting the same project from scratch for the same remaining million: chi-square(1, N = 108) = 50.6, p < .001. Experiment 5 removed the obvious confound by pricing the from-scratch plane at the same $10 million and got identical answers, 10 yes against 50 no. The weaker items in the same paper deserve equal billing. In the much-quoted ski-trip question only 46% chose the trip they expected to enjoy more (99% CI 30 to 62%), and the result reproduced when both trips were won in a radio competition rather than bought: 44 of 86 still took the $100 trip, a comparison that did not approach significance. An item that behaves the same way when nothing was sunk is not measuring only what its reputation says. Experiment 10 is the null worth keeping: across 181 students, 61 of them handed the questionnaire by their own economics professor in a class that had covered sunk costs that semester, the chi-square did not approach significance. Knowing about the bias does not switch it off.

The meta-analysis is where the honest story is. Sleesman, Conlon, McNamara and Miles (2012) screened an initial list of 917 articles for determinants of escalation of commitment. Sunk costs had a significant main effect: k = 34, n = 5,524, mean observed r = .195, corrected rho = .243, 95% CI .114 to .373. Two neighbours beat it. Time investment reached rho = .432 (k = 7, n = 1,664, CI .247 to .617), and proximity to project completion rho = .393 on a wider base with a tighter interval (k = 14, n = 3,073, CI .308 to .478). Then the moderator analysis took most of the sunk cost effect back. Where studies explicitly covaried sunk cost with project completion information — the classic nine-of-ten-million-spent, ninety-percent-complete vignette — rho was .528 (k = 5, n = 945, CI .332 to .725); where sunk cost was varied without that covariation, rho was .100 (k = 28, n = 3,821) and the 95% confidence interval ran from −.036 to .236. The authors' own conclusion is that the sunk cost effect may not be as robust as the literature would suggest, and that studies kept attributing to sunk cost what was partly or wholly a completion effect even though the completion argument had been in the literature for over fifteen years. Personal responsibility for the initial decision, from the same table, has the largest evidence base of any determinant they coded: k = 54, n = 8,625, rho = .258, CI .216 to .301.

Field evidence outside the lab exists for consumption, not for software retention. Ho, Png and Reza (2018) used Singapore's certificate-of-entitlement system, which moves the sunk cost of a new car independently of its running costs, across 8,264 cars and 45,195 service visits, and estimated the elasticity of driving with respect to sunk costs at 0.048 (s.e. 0.016). Between January 2009 and January 2013 the average policy-related sunk cost rose 116%, from S$11,278 to S$24,316; the elasticity implies that this was associated with 5.6% more driving, 86 kilometres a month against a sample average of 1,545. They describe it as the first field evidence of the sunk cost fallacy in usage of a major durable good, and their model has the excess usage attenuating over time. Note the magnitude: more than doubling an unrecoverable cost moved usage by about a twentieth. Nothing comparable exists for subscription software, where the closest thing to field data is enforcement data — Gibson Dunn's 2026 survey of negative-option actions records a $7.5 million settlement by the California Automatic Renewal Task Force with a meal-delivery service and, in November 2025, a $4.8 million settlement by 34 states with an online clothing retailer over its paid membership programme. Grade the principle mixed on that record: the direction survives controlled tests, the magnitudes are modest, and the headline correlation is confounded with completion information across much of the literature it rests on.