When to Stop, and What Not to Count

What a sunk cost is

Stopping a project is a decision, and decisions run on arithmetic. The first step is to know which numbers belong in it. The OpenStax managerial accounting textbook defines a sunk cost as one that cannot be avoided because it has already occurred. It adds that sunk costs have no bearing on future events and are not relevant in decision-making. What matters is the difference between the alternatives, which OpenStax calls differential analysis, and the cost of not choosing the other option, which it calls opportunity cost.

OpenStax gives a plain example. A car owner has already spent 1,600 dollars on repairs and tires. That money is sunk. The choice is a further 1,200 dollars of repairs now against a newer car with 24 payments of 150 dollars, 3,600 dollars in all, less a 200 dollar trade-in, a relevant cost of 3,400 dollars. The 1,600 dollars appears on neither side.

Why we keep going anyway

The textbook also concedes the difficulty. It says managers may be reluctant to ignore sunk costs, especially after an unwise prior decision, because letting go of a past choice means admitting a mistake.

The classic experiments are decades old. Arkes and Blumer, in a 1985 paper on the psychology of sunk cost, describe the sunk cost effect as a greater tendency to continue an endeavor once money, effort or time has been spent. In one questionnaire, 61 people chose between a 100 dollar trip and a 50 dollar trip they preferred, for the same weekend. Thirty-three chose the 100 dollar trip. In a field study of theater subscribers, buyers who paid full price used 4.11 tickets in the first five plays, against 3.32 and 3.29 for those given discounts. The effect was gone in the second half of the season.

A later survey is closer to a founder's position. Meyer, in a 2013 paper presented to the Project Management Institute, described a project planned at 25 million dollars over 15 months. At an early review it was forecast to take 19 months and cost 33.5 million dollars, with 2.5 million dollars salvageable if stopped. Of 482 respondents, 308 (64 percent) chose to continue and 174 (36 percent) chose to stop. The paper notes the project would lose money.

The arithmetic of stopping

The numbers in this section are illustrative, not real data. A small studio has spent 120,000 dollars on a product. Finishing needs 40,000 dollars more. We estimate a 40 percent chance that the finished product earns 130,000 dollars, so expected receipts are 0.4 x 130,000 = 52,000 dollars. Continuing is worth 52,000 - 40,000 = 12,000 dollars.

Stopping brings 6,000 dollars from selling what exists. The 40,000 dollars and the same months of work could instead fund a second product with expected receipts of 58,000 dollars, worth 58,000 - 40,000 = 18,000 dollars. Stopping is therefore worth 6,000 + 18,000 = 24,000 dollars. The difference is 12,000 dollars in favour of stopping. The 120,000 dollars already spent is in neither figure. Framed as "we would waste 120,000 dollars", the same project feels like an easy yes.

The rule is: stop when Salvage + Best alternative > Expected receipts - Remaining cost. The inputs are estimates, so we would test them. If the chance of success were 60 percent, expected receipts would be 78,000 dollars and continuing would be worth 38,000 dollars, which beats 24,000. The decision turns on one probability, and that is the number to argue about.

What the evidence does not show

The experiments above do not show that stopping earlier improves outcomes in real ventures. Meyer's survey was hypothetical, about half the participants were part-time project management students, and only financial benefits were tested. His literature review also notes that some researchers argue sunk cost and project completion are confounded. The summary at Behavioral Economics lists further studies, including work finding that cognitive reflection can reduce the behaviour. The evidence supports a tendency, not a law.

Base rates give context, with limits. Bureau of Labor Statistics data show that of private-sector establishments born in March 2013, 79.6 percent were operating a year later, 50.6 percent after five years and 34.7 percent after ten. The BLS survival table extends this to 30.4 percent by March 2025. The same table shows that survivors keep surviving: 91.5 percent of establishments alive in 2017 were alive in 2018. These are establishments, not firms, and BLS cautions that a closing is not necessarily a permanent end. An establishment with no employment in March may reopen. The data cannot tell us whether any closure was a wise stop.

Going deeper

Three ideas sit behind the example.

• Relevant cost: only avoidable costs that differ between alternatives matter, per OpenStax.
• Expected receipts = Probability of success x Payoff if successful, summed over outcomes.
• Stop if Salvage + Best alternative > Expected receipts - Remaining cost.

The harder problem is behavioural, so the fix is procedural. Vermeulen and Sivanathan, in a Harvard Business Review article from 2017, describe escalation of commitment as driven by reinforcing biases, including the sunk cost fallacy and loss aversion. Their first remedy is to set decision rules in advance. Meyer's review reports a related technique: minimum goals that trigger review or withdrawal, which Simonson and Staw found produced the largest reduction in escalation among the methods they tested. We would write the rule before the spend, for example: if the chance of success falls below 25 percent at the next review, we stop. We note that we have only read the abstract of the HBR article.

To read the survival data yourself, open Table 7 in the BLS series. It gives survivors since birth and survivors of the previous year for each opening-year cohort. As a check, 293,400 establishments alive in 2019 out of 318,339 in 2018 is 92.2 percent.

MaxxForge is designed to serve outside founders as well as our own portfolio, which we treat as a worked example, not the limit. General information, not financial advice. Every figure above links to the page we read, and if one fails to match, tell us and we will correct it.