The money has gone, but it remains at the table. It speaks when the project is reconsidered, when the expensive ticket meets an unappealing evening, when another year is offered to an undertaking that has already taken several. We have put too much into it to stop. The marginal principle gives its sharpest reply here: the decision can change what happens next, and cannot retrieve a payment that every available alternative has already lost. Yet the payment feels like a reason, sometimes like an obligation. The original question asks why almost no one can bear to ignore the past. The difficulty is real; its reach, and the circumstances in which persistence is mistaken, need to be examined rather than assumed.
A sunk cost is an expenditure that cannot now be recovered and does not vary across the alternatives under consideration. It can have been grave, avoidable and badly chosen. None of that gives it a further power to distinguish alternatives that bear it equally. Suppose an invented project has consumed ninety. Finishing it will require twenty more and produce a certain receipt of thirty; stopping yields no receipt and no salvage. Continuing improves the position by ten from today, although the whole project will still have lost eighty. Stopping would leave the full ninety lost. The fact that the undertaking was a bad idea at the beginning does not make abandoning it the best idea now.
Change only the remaining cost to forty and the decision reverses. Completion now adds a loss of ten, turning the historical loss of ninety into a hundred. No size assigned to the vanished payment can improve that future bargain. It is tempting to treat the loss already incurred as an argument for a more strenuous rescue, but the rescue must be judged by what it can still rescue. The proper comparison is neither a rule to finish nor a rule to abandon. It can require finishing a project whose lifetime account will remain disastrous, and abandoning another that once looked excellent. The difference lies in the consequences still open to choice.
Something bought in the past may nevertheless retain a present alternative use. Restore the completion cost of twenty, but suppose the unfinished asset could instead be sold for fifteen. Completion leaves ten from today; sale leaves fifteen. Selling is better by five. Calling the asset already paid for would conceal the opportunity being sacrificed by using it. The original payment is sunk, but the asset is not valueless merely because its invoice is old. Recoverable deposits, resale proceeds and alternative uses belong in the current comparison. So do costs of cancellation that have not yet become unavoidable. The discipline is to remove the common irrecoverable amount while retaining every consequence that still differs.
The difficulty is not confined to managers facing complicated forecasts. In Arkes and Blumer's 1985 questionnaire, students were asked to imagine having bought two nonrefundable, nonresalable ski trips for the same weekend. One cost a hundred dollars, the other fifty, and the cheaper trip was explicitly expected to be more enjoyable. Thirty-three of sixty-one chose the more expensive trip. The larger payment could not be recovered through that choice, and the scenario had already specified which experience was preferred. This is a compact demonstration of the past entering where the prospective comparison gives it no place. It is also a bounded observation: a hypothetical choice by this group of respondents, not a finding that everyone makes the error or that every costly commitment should be broken.
The same paper moved beyond the questionnaire by varying what theater subscribers paid. The first sixty intending purchasers were assigned tickets at full price or one of two discounts; six purchasers buying as couples were excluded to avoid treating their joint decisions as independent. In the planned comparisons for the first half of the season, full-price buyers used more tickets than either discounted group. The differences did not recur in the second half. The experiment thus gives the earlier expenditure a possible afterlife in actual attendance, while also showing that the afterlife need not last. Its small sample and the distinction between the planned comparisons and the overall analysis matter. The result does not support turning a demonstrated tendency into an immutable feature of every chooser.
Richard Thaler's account of mental accounting offers a way to understand the hold of the payment. An advance purchase can open an account that remains unfinished until the experience is consumed. Attending lets the purchaser place the expenditure in the ordinary category of things bought and used; staying home can force recognition of a loss. The distinction is psychologically consequential even when the money has gone in both cases. In this account, making use of the purchase can protect the earlier decision from a painful evaluation. A new burden — the unpleasant journey, the lost evening — may be recorded separately or receive less attention than the old payment it is supposed to redeem. The accounts have become part of the decision, rather than a neutral record of it.
The felt obligation to avoid waste has an intelligible source. Before a purchase, thrift compares what the money could obtain elsewhere. After an irrecoverable purchase, repeating the same concern can misidentify the waste: an unused ticket becomes more visible than the good evening sacrificed to use it. The rule that protected resources before the transaction can spend additional resources after it. This is the peculiar loyalty in the sunk-cost error. The chooser tries to vindicate an earlier self, as though consuming something unwanted could turn the original expenditure into a better decision. A candid account of the loss can feel like causing it, even though the next decision is only where its recognition occurs.
Yet knowing the past can be indispensable to judging the future. McAfee, Mialon and Mialon examine projects whose total cost of completion is uncertain. Expenditure without success changes the information available about how much more will be needed. In their model, whether greater past expenditure raises or lowers the chance of success from a further investment depends on the distribution of the unknown completion cost. Both directions are possible. The sunk payment still cancels from the direct comparison of future gains and losses, while the history it records changes those future gains and losses. Ignoring the payment as an additional debit is consistent with using the experience as evidence.
For a small illustration, suppose the total cost needed to complete a project is either sixty or a hundred, each initially equally likely. Completion pays ninety. The decision maker is risk neutral, knows this distribution and can make only one further payment of ten. After fifty has been spent without success, another ten has a one-half chance of completing the project. Its expected receipt is forty-five and its new cost ten, so it is worth making. After sixty has been spent without success, however, the low-cost possibility has been ruled out. Another ten cannot complete the project, and no later payment is available. The same proposed expenditure is now unattractive. More has been sunk, but the sensible response is less persistence because failure has taught the chooser something adverse.
The past can also change what others will do next. A manager who abandons a project may reveal information about an earlier judgment and suffer a future career cost. Continuing can then serve the manager while harming the organization whose resources pay for it. Alternatively, a credible commitment to remain in a joint undertaking can encourage a partner to invest, changing the gains available to both. McAfee and his coauthors distinguish these reputational mechanisms. They explain why an observed association between prior investment and persistence cannot, by itself, diagnose a mistake. A private incentive, a cooperative commitment and an irrecoverable cost mistakenly counted again may produce similar outward behavior while answering different questions about benefit and loss.
This leaves room for loyalty without making every lost expense sacred. A relationship can contain present affection, future obligations and another person's reliance, all of which a narrow cash account could omit. A damaged undertaking can still be worth repairing because the repair creates something valuable. The sunk-cost principle does not decide what should count as value or release anyone from an obligation merely because it began earlier. It asks whether the stated reason distinguishes the futures now available. The pain of acknowledging an error may itself be real, but allowing that pain to purchase indefinite further sacrifice gives the original error a continuing claim on the life that remains. Recognition and repetition are different ways of carrying a past.
The past can teach the next decision without charging it the same cost again.
The last unit thus draws a difficult boundary through memory. An irrecoverable payment supplies no additional gain from continuing; what it has taught, what it has built and what stopping will now change can matter greatly. The arithmetic is exact once the alternatives are faithfully described. Faithfully describing them is the human work. The chooser must separate a reason to persist from a wish that persistence would vindicate what has already happened. This is why the rule can be understood long before it becomes easy to obey. The next plate reaches the boundary of the marginal country itself, where choices arrive in lumps and the available measure can place a well-funded whim above an urgent need. The argument remains open.