Are Sunk Costs Fallacies?
Human and animal sunk costs often aren’t, and sunk cost bias may be useful on an individual level to encourage learning. Convincing examples of sunk cost bias typically operate on organizational levels and are probably driven by non-psychological causes like competition.
It is time to let bygones be bygones.
Khieu Samphan1, Khmer Rouge head of state
The sunk cost fallacy (“Concorde fallacy”, “escalation bias”, “commitment effect” etc.) could be defined as when an agent ignores that option X has the highest marginal return, and instead chooses option Y because he chose option Y many times before, or simply as “throwing good money after bad”. It can be seen as an attempt to derive some gain from mistaken past choices. (A slogan for avoiding sunk costs: “give up your hopes for a better yesterday!”) The single most famous example, and the reason for it also being called the “Concorde fallacy”, would be the British and French government investing hundreds of millions of dollars into the development of a supersonic passenger jet despite knowing that it would never succeed commercially2. Since 1985’s3 forceful investigation & denunciation, it has become received wisdom4 that sunk costs are a bane of humanity.
But to what extent is the “sunk cost fallacy” a real fallacy?
Below, I argue the following:
sunk costs are probably issues in big organizations
but maybe not ones that can be helped
sunk costs are not issues in animals
sunk costs appear to exist in children & adults
but many apparent instances of the fallacy are better explained as part of a learning strategy
and there’s little evidence sunk cost-like behavior leads to actual problems in individuals
much of what we call “sunk cost” looks like simple carelessness & thoughtlessness
Subtleties
One cannot proceed from the informal to the formal by formal means.
A “sunk cost fallacy” is clearly a fallacy in a simple model: ‘imagine an agent A who chooses between option X which will return $10 and option Y which will return $6, and agent A in previous rounds chose Y’. If A chooses X, it will be better off by $4 than if it chooses Y. This is correct and as hard to dispute as ‘A implies B; A; therefore B’. We can call both examples valid. But in philosophy, when we discuss modus ponens, we agree that it is always valid, but we do not always agree that it is sound: that A does in fact imply B, or that A really is the case, and so B is the case. ‘The moon being made of cheese implies the astronauts walked on cheese; the moon is made of cheese; therefore the astronauts walked on cheese’ is logically valid, but not sound, since we don’t think that the moon is made of cheese. Or we differ with the first line as well, pointing out that only some of the Apollo astronauts walked on the moon. We reject the soundness.
We can and must do the same thing in economics—but ceteris is never paribus. In simple models, sunk cost is clearly a valid fallacy to be avoided. But is the real world compliant enough to make the fallacy sound? Notice the assumptions we had to make: we wish away issues of risk (and risk aversion), long-delayed consequences, changes in options as a result of past investment, and so on.
We can illustrate this by looking at an even more sacred aspect of normative economics: exponential discounting. One of the key justifications of exponential discounting is that any other discounting can be money-pumped by an exponential agent investing at each time period at whatever the prevailing return is or loaning at appropriate times. (George Ainslie in The Breakdown of Will gives the example of a hyperbolic agent improvidently selling its winter coat every spring and buying it just before the snowstorms every winter, being money-pumped by the consistent exponential agent.) One of the assumptions is that certain rates of investment return will be available; but in the real world, rates can stagger around for long periods. “Hyperbolic discounting is rational: valuing the far future with uncertain discount rates” (2009)5 argues that if returns follow a more geometric random walk, hyperbolic discounting is superior6. Are they correct? They are not much-cited or criticized. But even if they are wrong about hyperbolic discounting, it needs proving that exponential discounting does in fact deal correctly with changing returns. (The market over the past few years has not turned in the proverbial 8–9% annual returns, and one wonders if there will ever be a big bull market that makes up for the great stagnation.)
If we look at sunk cost literature, we must keep many things in mind. For example:
organization versus individuals
Sunk costs seem especially common in groups, as has been noticed since the beginning of sunk cost research7; et al 2000 found that culture influenced how much managers were willing to engage in hypothetical sunk costs (South & East Asian more so than North American), and a 200521ya meta-analysis that sunk cost was an issue, especially in software-related projects8, agreeing with a 200917ya meta-analysis, Desai & Chulkov. 2011 interviewed principals at Californian schools, finding evidence of sunk cost bias. Wikipedia characterizes the Concorde incident as “regarded privately by the British government as a ‘commercial disaster’ which should never have been started, and was almost canceled, but political and legal issues had ultimately made it impossible for either government to pull out.” So at every point, coalitions of politicians and bureaucrats found it in their self-interest to keep the ball rolling.
A sunk cost for the government or nation as a whole is far from the same thing as a sunk cost for those coalitions—responsibility is diffused, which encourages sunk cost9 (If Kennedy or other US presidents could not withdraw from Vietnam or Iraq10 or Afghanistan11 due to perceived sunk costs12, perhaps the real problem was why Americans thought Vietnam was so important and why they feared looking weak or provoking another “who lost China” debate.) People commit sunk cost much more easily if someone else is paying, possibly in part because they are trying to still prove themselves right—an understandable and rational choice13! Other anecdotes from 1992 suggest sunk cost can be expensive corporate problems, but of course are only anecdotes; Robert Campeau killed his company by escalating to an impossibly expensive acquisition of Bloomingdale’s but would Campeau ever have been a good corporate raider without his aggressiveness; or can we say the Philip Morris-Proctor & Gamble coffee price war was a mistake without a great deal more information; and was Bobby Fischer’s vendetta against the Soviet Union sunk cost or a rational strategy to Soviet collusion or simply an early symptom of the apparent mental issues that saw him converting to and impoverished by a peculiar church and ultimately an internally persecuted convict in Iceland?
And why were those coalitions in power in the first place? France and Britain have not found any better systems of government—systems which operate efficiently and are also Nash equilibriums, which successfully avoid any sunk costs in their myriads of projects and initiatives. In Joseph Tainter’s 198838ya Collapse of Complex Societies, he argues that societies that overreach do so because it is impossible for the organizations and members to back down on complexity as long as there is still wealth to extract, even when margins are diminishing; when we accuse Pueblo Indians of sunk cost and causing their civilization to collapse14, we should keep in mind there may be no governance alternatives. Debacles like the Concorde may be necessary because the alternatives are even worse—decision paralysis or institutional paranoia15. Aggressive policing of projects for sunk-costs may wind up violating Chesterton’s fence if managers in later time periods are not very clear on why the projects were started in the first place and what their benefits will be. If we successfully ‘avoid’ sunk cost-style reasoning, does that mean we will avoid future Vietnams, at the expense of World War IIs?16 Goodhart’s Law comes to mind here, particularly because one study recorded how a bank’s attempt to eliminate sunk cost bias in its loan officers resulted in backfiring and evasion17; the overall results seem to still have been an improvement, but it remains a cautionary lesson.
Whatever pressures and feedback loops cause sunk cost fallacy in organizations may be completely different from the causes in individuals.
Non-monetary rewards and penalties
“Individual organisms are best thought of as adaptation-executers rather than as fitness-maximizers.” What does this mean in a sunk cost context? That we should be aware that humans may not treat the model at its literal face value (without careful thought or strong encouragement to do so, anyway), treat the situation as simply as ‘$10 versus $6 (and sunk cost)’. It may be more like ‘$10 (and your—non-existent—tribe’s condemnation of you as greedy, insincere, small-minded, and disloyal) versus $6 (and sunk cost)’18. If humans really are forced to think like this, then the modeling of payoffs simply doesn’t correspond with reality and of course our judgements will be wrong. Some assumptions spit out sunk costs as rational strategies19. This is not a trivial issue here (see the self-justification literature, eg. 1981) or in other areas; for example, providing the correct amount of rewards caused many differences in levels of animal intelligence to simply vanish—the rewards had been unequal (see my excerpts of the essay “If a Lion Could Talk: Animal Intelligence and the Evolution of Consciousness”).
Sunk costs versus investments and switching costs
Many choices for lower immediate marginal return are investments for greater future return. A single-stage model cannot capture this. Likewise, switching to new projects is not free, and the more expensive switches are, the fewer switches is optimal (eg. et al 2017).
Demonstrated harm
It’s not enough to suggest that a behavior may be harmful; it needs to be demonstrated. One might argue that an all-you-can-eat buffet will cause overeating and then long-term harm to health, but do experiments bear out that theory?
Indeed, meta-analysis of escalation effect studies suggests that sunk cost behavior is not one thing but reflects a variety of theorized behaviors & effects of varying rationality, ranging from protecting one’s image & principal-agent conflict to lack of information/options (2012), not all of which can be regarded as a simple cognitive bias to be fixed by greater awareness.
Animals
“It really is the hardest thing in life for people to decide when to cut their losses.”
“No, it’s not. All you have to do is to periodically pretend that you were magically teleported into your current situation. Anything else is the sunk cost fallacy.”
John, Overcoming Bias
Point 3 leads us to an interesting point about sunk cost: it has only been identified in humans, or primates at the widest20.
1999 (“The Sunk Cost and Concorde Effects: Are Humans Less Rational Than Lower Animals?”) claims (see also the very similar 1987):
The sunk cost effect is a maladaptive economic behavior that is manifested in a greater tendency to continue an endeavor once an investment in money, effort, or time has been made. The Concorde fallacy is another name for the sunk cost effect, except that the former term has been applied strictly to lower animals, whereas the latter has been applied solely to humans. The authors contend that there are no unambiguous instances of the Concorde fallacy in lower animals and also present evidence that young children, when placed in an economic situation akin to a sunk cost one, exhibit more normatively correct behavior than do adults. These findings pose an enigma: Why do adult humans commit an error contrary to the normative cost-benefit rules of choice, whereas children and phylogenetically humble organisms do not? The authors attempt to show that this paradoxical state of affairs is due to humans’ overgeneralization of the “Don’t waste” rule.
Specifically, in 197254ya, Trivers proposed that fathers are more likely to abandon children, and mothers less likely, because fathers invest less resources into children—mothers are, in effect, committing sunk cost fallacy in taking care of them. 1976 pointed out that this is a misapplication of sunk cost, a version of point #3; Arkes & Ayton’s summary:
If parental resources become depleted, to which of the two offspring should nurturance be given? According to Trivers’s analysis, the older of the two offspring has received more parental investment by dint of its greater age, so the parent or parents will favor it. This would be an example of a past investment governing a current choice, which is a manifestation of the Concorde fallacy and the sunk cost effect. Dawkins and Carlisle suggested that the reason the older offspring is preferred is not because of the magnitude of the prior investment, as Trivers had suggested, but because of the older offspring’s need for less investment in the future. Consideration of the incremental benefits and costs, not of the sunk costs, compels the conclusion that the older offspring represents a far better investment for the parent to make.
Direct testing fails:
A number of experimenters who have tested lower animals have confirmed that they simply do not succumb to the fallacy (see, eg. Armstrong & Robertson, 1988; et al 1989; Maestripieri & Alleva, 1991; Wiklund, 199021).
A direct example of the Trivers vs Dawkins & Carlisle argument:
A prototypical study is that of Maestripieri and Alleva [199135ya], who tested the litter defense behavior of female albino mice. On the 8th day of a mother’s lactation period, a male intruder was introduced to four different groups of mother mice and their litters. Each litter of the first group had been culled at birth to four pups. Each litter of the second group had been culled at birth to eight pups. In the third group, the litters had been culled at birth to eight pups, but four additional pups had been removed 3 to 4 hr before the intruder was introduced. The fourth group was identical to the third except that the removed pups had been returned to the litter after only a 10-min absence.
The logic of the 1991 study is straightforward. If each mother attended to past investment, then those litters that had eight pups during the prior 8 days should be defended most vigorously, as opposed to those litters that had only four pups. After all, having cared for eight pups represents a larger past investment than having cared for only four. On the other hand, if each mother attended to future costs and benefits, then those litters that had eight pups at the time of testing should be defended most vigorously, as opposed to those litters that had only four pups. The results were that the mothers with eight pups at the time of testing defended their litters more vigorously than did the mothers with four pups at the time of testing. The two groups of mothers with four pups did not differ in their level of aggression toward the intruder, even though one group of mothers had invested twice the energy in raising the young because they initially had to care for litters of eight pups.
Arkes & Ayton rebut 3 studies by arguing:
1980: digger wasps fight harder in proportion to how much food they contributed, rather than the total—because they are too stupid to count the total and only know how much they personally collected & stand to lose
1995: cichlid fish successful in breeding also fight harder against predators; because this may reflect an intrinsic greater healthiness and greater future opportunities, rather than sunk cost fallacy, an argument similar to 1984’s criticism of apparent sunk costs in economics22
1979: savannah sparrows defend their nests fiercer as the nest approaches hatching; because as already pointed out, the closer to hatching, the less future investment is required for X chicks compared to starting all over
To which 3 we may add tundra swan feeding habits, which are predicted to be optimal by 201123, who remark “we show how optimization of Eq. 3 predicts the sunk-cost effect for certain scenarios; a common element of every case is a large initial cost.”
(2004, “The Sunk Cost Effect In Pigeons And Humans”, claim sunk cost effect in pigeons, but it’s hard to compare its strength to sunk cost in humans, and the setup is complex enough I’m not sure it is sunk cost.)
Humans
Children
Arkes & Ayton cite 2 studies finding that committing sunk cost bias increases with age—as in, children do not commit it. They also cite 2 studies saying that
1997 tested children at three different age groups (5–6, 8–9, and 11–12) with the following modification of the 1981 experiment …the older children provided data analogous to those found by : When the money was lost, the majority of the respondents decided to buy a ticket. On the other hand, when the ticket was lost, the majority decided not to buy another ticket. This difference was absent in the youngest children. Note that it is not the case that the youngest children were responding randomly. They showed a definite preference for purchasing a new ticket whether the money or the ticket had been lost. Like the animals that appear to be immune to the Concorde fallacy, young children seemed to be less susceptible than older children to this variant of the sunk cost effect. The results of the study by Krouse (198640ya) corroborate this finding: Compared with adult humans, young children, like animals, seem to be less susceptible to the Concorde fallacy/sunk cost effect.
… Perhaps the impulsiveness of young children (Mischel, Shoda, & Rodriguez, 198937ya) fostered their desire to buy a ticket for the merry-go-round right away, regardless of whether a ticket or money had been lost. However, this alternative interpretation does not explain why the younger children said that they would buy the ticket less often than the older children in the lost-money condition. Nor does this explanation explain the greater adherence to normative rules of decision making by younger children compared with adults in cases where impulsiveness is not an issue (see, eg. Jacobs & Potenza, 1991; Reyna & Ellis, 1994).
I think Arkes & Ayton are probably wrong about children. Those 2 early studies can be criticized easily24, and other studies point the opposite way. et al 1993 asked poor and rich kids (age 5–12) questions including an 1985 question, and found, in their first study no difference by age, ~30% of the 101 kids committing sunk cost and another ~30% unsure; in their second, they asked 2 questions, with ~50% committing sunk cost—and responses on the 2 questions minimally correlated (r = 0.17). 2004 found that correct (non-sunk cost) responses went up with age (age 5–12, 16%; 5–16, 27%; and adults 37%). Bruine de et al 2007 found older adults more susceptible than young adults to some tested fallacies, but that sunk cost resistance increased somewhat with age. et al 2008 studied 75 college age students, finding small or non-statistically-significant results for IQ (as did other studies, see later), education, and age; still older adults (60+) beat their college-age peers at avoiding sunk cost in both et al 2008 & et al 2011.
(Children also violate transitivity of choices & are more hyperbolic than adults, which is hardly normative.25)
Uses
Learning & Memory
18. If the fool would persist in his folly he would become wise.
46. You never know what is enough unless you know what is more than enough.
Felix Hoeffler in his 200818ya paper “Why humans care about sunk costs while (lower) animals don’t: An evolutionary explanation” takes the previous points at face values and asks how sunk cost might be useful for humans; his answer is that sunk cost forfeits some total gains/utility—just as our simple model indicated—but in exchange for faster learning, an exchange motivated by humans’ well-known risk aversion and dislike of uncertainty26. It is harder to learn the value of choices if one is constantly breaking off before completion to make other choices, or realize any value at all (the classic exploration vs exploitation problem, amusingly illustrated in Clarke’s story “Superiority”).
One could imagine a not too intelligent program which is, like humans, over-optimistic about the value of new projects; it always chooses the highest value option, of course, to avoid committing sunk cost bias, but oddly enough, it never seems to finish projects because better opportunities seem to keep coming along… In the real world, learning is valuable and one has many reasons to persevere even past the point one regards a decision as a mistake; et al 2007 (remember the exponential vs hyperbolic discounting example):
Consider a project that may take an unknown expenditure to complete. The failure to complete the project with a given amount of investment is informative about the expected amount needed to complete it. Therefore, the expected additional investment required for fruition will be correlated with the sunk investment. Moreover, in a world of random returns, the realization of a return is informative about the expected value of continuing a project. A large loss, which leads to a rational inference of a high variance, will often lead to a higher option value because option values tend to rise with variance. Consequently, the informativeness of sunk investments is amplified by consideration of the option value…Moreover, given limited time to invest in projects, as the time remaining shrinks, individuals have less time over which to amortize their costs of experimenting with new projects, and therefore may be rationally less likely to abandon current projects…Past investments in a given course of action often provide evidence about whether the course of action is likely to succeed or fail in the future. Other things equal, a greater investment usually implies that success is closer at hand. Consider the following simple model…The only case in which the size of the sunk investment cannot affect the firm’s rational decision about whether to continue investing is the rather special case in which the hazard is exactly constant.
If this model is applicable to humans, we would expect to see a cluster of results related to age, learning, teaching, difficulty of avoiding even with training or education, minimal avoidance with greater intelligence, completion of tasks/projects, largeness of sums (the risks most worth avoiding), and competitiveness of environment. (As well as occasional null results like 2006.) And we do! Many otherwise anomalous results snap into focus with this suggestion:
information is worth most to those who have the least: as we previously saw, the young commit sunk cost more than the old
in situations where participants can learn and update, we should expect sunk cost to be attenuated or disappear; we do see this (eg. et al 200727, et al 199028, et al 199929, 198130, 198631, et al 199132, 201233)
the noisier (higher variance) feedback on profitability was, the more data it took before people give up ( et al 1998, et al 2003)
sunk costs were supported more when subjects were given justifications about learning to make better decisions or whether teachers/students were involved (199534)
extensive economic training does not stop economics professors from committing sunk cost, and students can be quickly educated to answer sunk cost questions correctly, but with little carry-through to their lives35, and researchers in the area argue about whether particular setups even represent sunk costs at all, on their own merits36 (but don’t feel smug, you probably wouldn’t do much better if you took quizzes on it either)
when measured, avoiding sunk cost has little correlation with intelligence37—and one wonders how much of the correlation comes from intelligent people being more likely to try to conform to what they have learned is economics orthodoxy
a ‘nearly completed’ effect dominates ‘sunk cost’ (1993, 1998, 2002, et al 2007)
for example, the larger the proportion, the more costs were sunk (1991)
it is surprisingly hard to find clear-cut real-world non-government examples of serious sunk costs; the commonly cited non-historical examples do not stack up:
199538 studied the NBA to see whether high-ranked but underperforming players were over-used by coaches, a sunk cost.
Unfortunately, they do not track the over-use down to actual effects on win-loss or other measures of team performance, effects which are unlikely to be very large since the overuse amounts to ~10–20 minutes a game. Further, “The econometrics and behavioral economics of escalation of commitment: a re-examination of Staw and Hoang’s NBA data” (1999), claims to do a better analysis of the NBA data and find the effect is actually weaker. As usual, there are multiple alternatives39.
et al 199340 is a much-cited correlational study finding that small entrepreneurs invest further in companies they founded (rather than bought) when the company apparently does poorly; but they acknowledge that there are financial strategies clouding the data, and like Staw & Hoang, do not tie the small effect—which appears only for a year or two, as the entrepreneurs apparently learn—to actual negative outcomes or decrease in expected value.
similar to et al 1993, et al 200941 tracked ‘exit routes’ for young companies such as being bought, merged, or bankrupt—but again, they did not tie apparent sunk cost to actual poor performance.
in 2 studies42, Africans did not engage in sunk cost with insecticide-treated bed nets—whether they paid a subsidized price or free did not affect use levels, and in one study, this null effect happened despite the same household engaging in sunk cost for hypothetical questions
Internet users may commit sunk cost in browsing news websites43 (but is that serious?)
an unpublished 200125ya paper (Barron et al “The Escalation Phenomenon and Executive Turnover: Theory and Evidence”) reportedly finds that projects are ‘significantly more likely’ to be canceled when their top managers leave, suggesting a sunk cost effect of substantial size; but it is unclear how much money is at stake or whether this is—remember point #1—power politics44
sunk cost only weakly correlates with suboptimal behavior (much less demonstrates causation):
2005 and Bruine de et al 2007 compiled a number of questions for several cognitive biases—including sunk cost—and then asked questions about impulsiveness, number of sexual partners, etc., while the latter developed a 34-item index of bad decisions/outcomes (the DOI): ever rent a movie you didn’t watch, get expelled, file for bankruptcy, forfeit your driver’s license, miss an airplane, bounce a check, etc. Then they ran correlations. They replicated the minimal correlation of sunk cost avoidance with IQ, but sunk cost (and ‘path independence’) exhibited fascinating behaviors compared to the other biases/fallacies measured: sunk cost & path independence correlated minimally with the other tested biases/fallacies, Cronbach’s alphas were almost uselessly low, education did not help much, age helped some, and sunk cost had low correlations with the risky behavior or the DOI (eg. after controlling for decision-making styles, 0.13).
et al 1993 found tests of normative economic reasoning, including sunk cost questions, correlated with increased academic salaries, even for non-economic professors like biologists & humanists (but the effect size & causality are unclear)
Dissociation in hypotheticals—being told a prior manager made decisions—does not always counteract effects (2006)
Sunk costs may also reflect imperfect memory about what information one had in the past; one may reason that one’s past self had better information about all the forgotten details that went into a decision to make some investments, and respect their decision, thus appearing to honor sunk costs (2011).
Countering Hyperbolic Discounting?
Use barbarians against barbarians.45
Henry Kissinger, On China 2011
The classic kicker of hyperbolic discounting is that it induces temporal discounting—your far-sighted self is able to calculate what is best for you, but then your near-sighted self screws it all up by changing tacks. Knowing this, it may be a good idea to not work on your ‘bad’ habit of being overconfident about your projects46 or engaging in planning fallacy, since at least they will counteract a little the hyperbolic discounting; in particular, you should distrust near-term estimates of the fun or value of activities when you have not learned anything very important47. We could run the same argument but instead point to the psychology research on the connection between blood sugar levels and ‘willpower’; if it takes willpower to start a project but little willpower to cease working on or quit a project, then we would expect our decisions to quit be correlated with low willpower and blood sugar levels, and hence to be ignored!
It’s hard to oppose these issues: humans are biased hardware. If one doesn’t know exactly why a bias is bad, countering a bias may simply let other biases hurt you. Anecdotally, a number of people have problems with quite the opposite of sunk cost fallacy—overestimating the marginal value of the alternatives and discounting how little further investment is necessary48, and people try to commit themselves by deliberately buying things they don’t value.49 (This seems doubly plausible given the high value of Conscientiousness/Grit50—with marginal return high enough that it suggests most people do not commit long-term nearly enough, and if sunk cost is the price of reaping those gains…)
Thoughtlessness: the Real Bias
One of the known ways to eliminate sunk cost bias is to be explicit and emphasize the costs of continuing (Northcraft and Neale, 1986, Tan and Yates, 1995, et al 198251 & conversely 198152, 1986), as well as setting explicit budgets (1992, 199553, et al 1997). Fancy tools don’t add much effectiveness54
This, combined with the previous learning-based theory of sunk cost, suggests something to me: sunk cost is a case of the ur-cognitive bias, failure to active System II. One doesn’t intrinsically over-value something due to past investment, one fails to think about the value at all.