The Number That Is Irrelevant Both Ways

This publication has an article on sunk cost already. It is a practitioner's piece about governance, warning signs and Canadian mechanics, and it does its job. This one is about the evidence underneath it, which is more contested, more interesting, and in one respect considerably worse than the practitioner literature admits.

The Verdict, Stated First

Four claims, in descending order of how confident we are, because this article is long and a reader deserves the conclusion before the argument.

One. The correct decision rule is not in dispute and is almost never applied properly. Whether to continue depends on the remaining cost against the remaining value. The amount already spent does not enter, in either direction. We demonstrate below that the two most common business heuristics, we have spent too much to stop and we have spent too much to keep going, are the same error, because both consult a number that is irrelevant.

Two. A large share of what gets diagnosed as sunk cost reasoning is ordinary forecasting failure. Our own arithmetic shows that a thirty percent optimism bias in estimating remaining cost flips the correct decision in three of four illustrative cases, with no sunk cost reasoning involved anywhere. Those two problems have different remedies, and calling both of them sunk cost guarantees the wrong one gets applied.

Three. The phenomenon is real but it is not a single thing. The meta-analytic literature treats escalation as the product of numerous determinants across psychological, social, project and structural categories, and a reading of that synthesis concludes the behaviours involved vary in how rational they are.

Four. The founding study did not replicate. A 2021 attempt with 365 participants found no escalation effect and reported that participants simply invested more in more promising projects. We treat this seriously and we do not treat it as decisive, for reasons set out at length below.

Our Grades For These Claims

Applying the scheme from the first article in this series.

Grade A that escalation behaviour occurs. Thirty-five years of research, a meta-analytic synthesis in a top management journal, and documented cases across finance, marketing, accounting, information systems and project management[4].

Grade B for the meta-analysis's specific determinants. We obtained fragments of the paper including several hypothesis statements and one numerical result, and we did not obtain its effect table.

Grade C for the responsibility effect specifically, which is the founding finding, because a well-powered replication attempt failed and because the meta-analysis itself notes recent evidence challenging that mechanism.

Grade C for the one social determinant reported significant, which rests on three studies.

Grade A for the decision arithmetic, which is not empirical and can be checked by anyone with a calculator.

Our position, stated plainly: the practitioner literature on sunk cost is more confident than the research literature supports, and the research literature is more confident about mechanism than about magnitude.

A Note On Method

Everything here is verified to August 2026.

We obtained fragments of the 2012 meta-analysis from a hosted copy, including its abstract, several hypothesis statements, one numerical result and one summary sentence[1][5]. We did not obtain its effect size tables, its moderator results in full, or its methods section, and we report no effect size from it beyond the single correlation quoted.

We obtained the abstract of the 2021 replication attempt in full[2] and not the paper.

We obtained the abstract of a 1992 review of self-justification explanations in full[6].

We did not obtain the 1976 founding study, the 1985 sunk cost paper, or any of the moderator studies discussed in the second half of this article, and report them from reference lists and citing descriptions.

One source used here is a personal website, cited once for a summary judgment about the meta-analysis and flagged at that use[3].

All arithmetic is ours and uses invented project figures.

This article discusses research on decision making. It is not investment, capital allocation, project management or accounting advice.

Knee-Deep In The Big Muddy

The founding paper, and the phrase that named a literature.

Reference lists identify Staw, B. M. (1976), Knee-deep in the Big Muddy: A study of escalating commitment to a chosen course of action, Organizational Behavior and Human Performance, 16(1), 27–44[7].

A description of what it showed: it "demonstrated that individuals responsible for an investment experiencing negative outcomes tend to persist in continuing that course of action, effectively ignoring warning signs."[4]

Three observations, ours.

The word carrying the weight is responsible. The founding claim is not merely that people continue failing projects. It is that the people who chose the project continue it more than people who inherited it, which is a claim about personal responsibility rather than about arithmetic.

That distinction is what makes the finding psychological rather than economic. An error anyone would make is a reasoning failure. An error that depends on who made the original decision is about the self.

And we did not obtain this paper. Everything in this article about it comes from descriptions in later work, which for a paper this influential is an uncomfortable position to be in and one we will return to.

Where The Name Came From

A small point of provenance, because the literature's vocabulary shapes how it is read.

The 1976 title's phrase, the Big Muddy, gave the field its running metaphor, and the 2012 meta-analysis titled itself Cleaning Up the Big Muddy in direct reference[1][7].

A survey of the area records the popular idioms attached to it: "Throwing good money after bad," and "In for a penny, in for a pound."[4]

Two observations, ours.

A field whose central metaphor is wading further into a swamp has embedded a conclusion in its vocabulary. The image presupposes that continuing is wrong, which is precisely the question at issue in any real case.

And the proverbs cut both ways in ordinary usage, which is worth noticing. In for a penny, in for a pound is not always foolish advice. Sometimes the pound is worth spending, and the whole difficulty is telling those cases apart.

The Psychology Of Sunk Cost

The economics-side treatment, published nine years after the founding management paper.

Reference lists identify Arkes, H. R., and Blumer, C. (1985), The psychology of sunk cost, Organizational Behavior and Human Decision Processes, 35(1), 124–140[8][7].

A survey records that economists use "related terms like the 'sunk-cost fallacy' (Arkes and Blumer 1985) and 'lock-in'" to describe phenomena related to escalation of commitment[9].

We did not obtain this paper and report no findings from it.

Two observations, ours.

Two literatures grew up around the same behaviour with different names, different journals and different explanations. Management called it escalation of commitment and explained it through self-justification; economics called it the sunk cost fallacy and explained it through mental accounting and prospect theory.

This is the twenty-eighth article's problem in a new setting, and we flag it early: when a single behaviour carries two names from two disciplines, the advice built on each transfers badly to the other. The management version is about who decided. The economics version is about how the numbers are represented. Those imply different remedies.

The Three Conditions

The cleanest definition we found, and it is more restrictive than common usage.

A 2018 paper in Psychological Science states: "Escalation of commitment to a failing course of action occurs in the presence of (a) sunk costs, (b) negative feedback that things are deviating from expectations, and (c) a decision between escala"ting or not, with our source truncating[10].

A later definition from the same research group describes the bias as "the act of 'carrying on' with questionable or failing courses of action."[11]

Three observations, ours.

All three conditions are required, and the second is the one commonly missing in business usage. Escalation requires negative feedback. Continuing a project that is going well is not escalation, however much has been spent.

The first condition, sunk costs, is necessary but not sufficient, which is the distinction that collapses in ordinary conversation. Every ongoing project has sunk costs. That fact alone diagnoses nothing.

And the third condition specifies a decision point. Escalation is a choice made at a juncture, not a general disposition, which means the remedy has to operate at junctures.

Thirty-Five Years, One Synthesis

The paper that tried to organise the field.

Sleesman, D. J., Conlon, D. E., McNamara, G., and Miles, J. E. (2012), Cleaning Up the Big Muddy: A Meta-Analytic Review of the Determinants of Escalation of Commitment, Academy of Management Journal, 55(3), 541–562[1][12].

Its abstract states the problem: "The topic of escalation of commitment has intrigued the organizational sciences for over 35 years. A variety of theoretical explanations have been offered for why escalation occurs, and numerous constructs have been examined as antecedents of escalation behavior. However, little effort has been made to systematically investigate these various accounts. Using meta-analysis, we present a comprehensive overview of the many determinants found in the literature, an analysis of the power of different theoretical perspectives, and an examination of the relative efficacy of the various theories."[12]

Three observations, ours.

That middle sentence is a serious indictment of thirty-five years of work, written by people sympathetic to it. Numerous constructs examined, little effort to investigate the accounts systematically. A field can accumulate a great deal of evidence without ever asking which of its explanations is doing the work.

The paper's stated ambition is comparative: not just whether escalation happens but which theoretical perspective explains it best. That is the right question and it is asked far too rarely.

And the structure it imposes is worth carrying. The paper organises determinants into categories including psychological, social, project and structural groupings, of which we obtained references to the social and structural ones directly[5].

Fourteen Of Sixteen

The headline result of the synthesis, in the authors' own words.

"In total, we found support for 14 of our 16 main effect predictions. Only two antecedents, both in the social determinants category (public evaluation of the deci"sion, and a second our source cuts off, failed to reach support[5].

Two of the individual hypotheses we obtained show how the predictions were framed. On decision risk: "Subjective expected utility theory: Risk increases the likelihood of loss and the salience of loss potential to decision makers, lessening the likelihood of escalation even in the face of information on previous performance."[13] On opportunity cost: "Opportunity cost information provides a clear" signal, with our source truncating[13].

Three observations, ours.

Both hypotheses we obtained predict negative relationships with escalation. That is, both risk and opportunity cost information are predicted to reduce escalation, which means a substantial part of what the meta-analysis established is about what suppresses the effect rather than what causes it.

That is more useful to a practitioner than a catalogue of causes, and it is the part that never reaches business writing.

And we obtained two of sixteen hypothesis statements. We are reporting the headline count from the paper's own summary sentence, and we cannot tell you what most of the sixteen were.

We Tested That Claim

Because a count of supported predictions is a claim that can be examined, and because fourteen out of sixteen sounds more impressive than it may be. Our own arithmetic, a crude directional check, not the authors' test.

Each of the sixteen predictions specifies a direction. If the underlying theory carried no information at all, and directions were guessed at random, each has a fifty percent chance of being right.

Under that null, the probability of getting 14 or more of 16 correct by chance is 0.00209, or about one in 478.

Three observations.

So the pattern is not chance. The theoretical framework predicts directions considerably better than a coin, and that is a real result.

But this test is weak in a specific way, and we want to be the ones to say so. It asks only whether directions were guessed correctly. It says nothing about whether the effects are large, whether the underlying studies were sound, or whether the predictions were formulated before or after the authors saw the literature they were summarising.

And that last point matters. A meta-analysis synthesising a mature literature is not in the same position as a preregistered study. The authors knew the field before framing the hypotheses, which is unavoidable in a review and which makes a hit rate of 14 out of 16 less surprising than it first appears. We report our own test and immediately discount it.

The One Social Determinant

The single numerical result we obtained from the meta-analysis, and it repays attention.

"The only social determinant having a significant relationship with escalation was group identity or cohesiveness strength (H15; ρ = .307, n = 138, k = 3)."[5]

Our own conversions, standard formulas: a correlation of .307 is roughly d = 0.65, which by the benchmarks used throughout this series is a moderate to large effect and would sit near the top of the distribution the fiftieth article assembled.

Three observations, ours.

If it holds, it is commercially significant and rarely discussed. It says the strength of group identity predicts escalation, which locates part of the problem in team cohesion rather than in individual reasoning.

That is an uncomfortable finding for the usual management prescription, because cohesion is normally treated as an unambiguous good. On this result, the same property that makes a team effective may make it worse at abandoning its own project.

And note what it implies about the rest of the category: the only social determinant reaching significance. Public evaluation of the decision, which is the intuitive candidate, is named among the two that failed.

And It Rests On Three Studies

The qualification, which the reported statistic makes unavoidable.

The result is k = 3. Three studies, 138 participants in total.

Our own arithmetic: at n = 138, a rough within-study confidence interval around ρ = .307 runs from about .16 to .46.

Three observations, and the third is the important one.

That interval is already wide. At its lower bound the effect is less than half what the point estimate suggests.

But the interval we just calculated understates the uncertainty, and we want to be explicit that our own number is the optimistic version. It treats 138 observations as though they came from one study. They came from three, and a proper meta-analytic interval accounts for variation between studies as well as within them. We have no heterogeneity estimate, so we cannot compute the correct interval and are showing the scale rather than the answer.

And three studies is thin for a headline finding. We grade this C, and we would grade it C even though it appears in a top journal, because the number of underlying studies is a fact about the evidence that no journal's prestige alters.

It Is Not One Thing

The reading of the meta-analysis that we think is correct, and which comes from a source we flag carefully.

A personal website's essay on the topic concludes: "meta-analysis of escalation effect studies suggests that sunk cost behavior is not one thing but reflects a variety of theorized behaviors and effects of varying rationality, ranging from protecting one's image and principal-agent conflict to lack of information/options, not all of which can be regarded as a simple cognitive bias to be fixed by greater awareness."[3]

This is a personal website, not a peer-reviewed source, and we cite it once, here, because it states a reading of the meta-analysis more directly than any academic source we located. A reader should weigh it accordingly.

Three observations, ours, and we think this framing is right for reasons independent of the source.

Principal-agent conflict is not a cognitive bias. A manager who continues a failing project because abandoning it ends their career is not making a reasoning error; they are correctly optimising a different objective function. The behaviour looks identical from outside.

Lack of options is not a cognitive bias either. A firm with one project and no alternative use for the capacity faces a genuinely different calculation from one choosing between projects, which is why the meta-analysis predicts opportunity cost information reduces escalation.

And this is the practical heart of the matter. If the same observable behaviour has causes ranging from self-deception to rational career management to genuine constraint, then a single diagnosis is worthless and a single remedy is worse than worthless.

The Self-Justification Account

The dominant explanation, stated by a review devoted to assessing it.

A 1992 review in Academy of Management Review states: "The present article first reviews evidence suggesting that escalation is determined, at least in part, by decision makers' unwillingness to admit that their prior allocation of resources to the chosen course of action was in vain (the self-justification explanation)."[6]

It then distinguishes rival accounts: "A distinction is drawn in the second part of the article between alternative (to self-justification) explanations of escalating commitment: Some are designed to replace self-justification, whereas others are intended to supplement self-justification, that is, to add explanatory power beyond that which can be accounted for by self-justification."[6]

And its verdict: "There is little evidence that the replacement theories provide a better explanation than does self-justification; however, theories designed to supplement self-justification are likely to lead to a more complete explanation."[6]

Three observations, ours.

The distinction between replacement and supplement theories is unusually clear thinking and we have not seen it made this crisply elsewhere in the behavioural literature.

The verdict is conservative and defensible: nothing beats self-justification, several things add to it. That is what a mature explanatory position looks like.

And a later paper describes the mechanism as operating indirectly, arguing that "the need for self-justification affects escalation of commitment indirectly via other cognitive processes", specifically "selective perception, sunk cost effect, and overoptimism"[14]. We did not obtain that paper and report its abstract's claim.

The Challenge To It

A qualification the meta-analysis itself records, which is the sort of thing that rarely survives into secondary accounts.

The 2012 paper writes: "Drawing on recent evidence (Schulz-Hardt et al., 2009) challenging the role of self-justification and the responsibility effect, we predicted that personal responsibility for the initial decision to begin a later failing course of action would lead to higher levels of escalation."[5]

We did not obtain the challenging study and report only that the meta-analysis names it as challenging both self-justification and the responsibility effect.

Two observations, ours.

The responsibility effect is the founding finding. A 2009 paper challenging it, cited inside the field's own meta-analytic synthesis three years later, is a significant fact about the state of the evidence.

And notice what the meta-analysis did with it: predicted the traditional direction anyway, and by their summary count found support. That is a legitimate way to handle a contested question and it means the meta-analysis is, on this point, evidence against the challenge rather than merely a report of it.

The Founding Study Failed To Replicate

The most consequential item in this article, and we obtained its abstract in full.

A 2021 paper in Economics and Business Review is titled Escalation of commitment is independent of numeracy and cognitive reflection. Failed replication and extension of Staw (1976)[2].

Its abstract: "First demonstrated by Staw (1976), escalation of commitment is the tendency for an individual to increase their commitment to a failing course of action when they are personally responsible for the negative consequences. An attempt was made to replicate this finding and test whether individual differences in numeracy and cognitive reflection could help explain such an effect. No evidence for escalation of commitment amongst the participants was found (N = 365). Participants simply invested more in more promising projects. Also, no evidence was found that numeracy or cognitive reflection predict escalation behaviour. The validity of escalation of commitment behaviour is discussed which suggests that future work should look to explore the boundary conditions of such an effect."[2]

Four observations, ours.

N = 365 is a substantial sample by the standards of this literature and of this series.

The sentence that matters most is the one that is easy to skim: "Participants simply invested more in more promising projects." That is not a null result. It is a description of participants behaving correctly.

The secondary finding is also worth noting. Neither numeracy nor cognitive reflection predicted escalation, which if the effect existed would suggest it is not a failure of computational ability, and given that it did not appear, tells us little.

And the authors' own framing is measured: they call for work on boundary conditions, which is a call to find where the effect operates rather than a declaration that it does not.

What That Does And Does Not Mean

Our reading, and we want to be careful here because a single failed replication is easy to over-read and this series has criticised others for doing so.

Four points.

It does not mean escalation does not occur. A meta-analytic synthesis of thirty-five years of work sits on the other side, and the fiftieth article's own audit found that failed replications in this series more often qualify findings than destroy them.

It does mean the responsibility effect specifically is less secure than its textbook status implies. This is the founding claim, it has a 2009 paper challenging it cited inside the field's own meta-analysis, and it has a 2021 attempt with 365 participants failing to find it. That is a pattern, not an isolated result.

It should change how confidently the phenomenon is invoked. A manager told that they are escalating because they made the original decision is being told something the evidence supports less strongly than they will assume.

And it does not settle the question, and we did not obtain the paper. We have an abstract. We cannot assess the design, the incentives, the sample, or whether the operationalisation matched the original closely enough to count as a test of it.

The Only Test That Matters

Setting the literature aside for a section, because the decision rule is not empirical and does not depend on any of the above. Our own arithmetic, invented figures, and the reasoning can be checked by anyone.

A project will be worth 900 on completion. It has consumed S so far and needs R more to finish. The correct test is R < V, and S appears nowhere.

Five cases, with the two popular heuristics scored against the correct answer.

Spent 100, still to come 1,200: correct answer STOP. Spent 400, to come 950: STOP. Spent 600, to come 880: GO. Spent 800, to come 400: GO. Spent 950, to come 150: GO.

Three observations.

Read the second row. Only 400 spent, and the project should be killed, because finishing costs 950 to produce 900. Early-stage projects can be worth abandoning immediately, and the small amount spent is not a reason to continue.

Read the fourth row. 800 spent, budget comprehensively blown, and continuing is correct, because 400 more produces 900. A manager who says we have already spent too much, cut our losses destroys 500 of value.

And notice the shape across the table. As spending accumulates, R falls, so the correct answer moves toward continuing. The sunk cost fallacy is real, but the naive correction for it points the wrong way on exactly the projects where most money is at stake.

Both Popular Heuristics Are Wrong

The point this article is named for. Ours.

Score the two common business rules against those five cases.

"We have spent too much to stop" gives the wrong answer on the first two rows, where the project should be killed.

"We have spent too much to keep going" gives the wrong answer on rows three, four and five, where continuing creates value.

Three observations.

Both heuristics consult S. Neither consults R against V. They are the same error, and the fact that they recommend opposite actions disguises this.

The second heuristic is the one taught as the cure for the first, which is the reason this matters. A generation of managers has been trained that large sunk costs are a warning sign, and on the arithmetic, large sunk costs are weakly associated with the correct answer being continue, because they imply less remains to be spent.

And we would state the rule this way. The amount already spent is not merely unhelpful. It is a number that, consulted in either direction, produces error, and its only legitimate use is as an input to estimating R.

The Case That Is Not Sunk Cost At All

The section we think is the most practically important in this article, and it concerns a failure mode with no psychology in it whatsoever. Our own arithmetic, invented figures.

The rule R < V is correct and requires knowing R. Nobody knows R. R is an estimate, produced by people who are attached to the project, and this publication's article on the planning fallacy documents that such estimates run optimistic.

So we asked: what happens to the decision if R is underestimated by thirty percent, with no sunk cost reasoning anywhere?

True remaining cost 1,200, reported as 840: the reported figure says GO, the truth says STOP. Wrong call.

True 1,200, optimism fifty percent, reported 600: says GO, truth says STOP. Wrong call.

True 1,000, reported 700: says GO, truth says STOP. Wrong call.

True 700, reported 490: says GO, truth says GO. Correct, by luck.

Three Of Four Wrong Calls

Reading that table honestly, including a correction to our own first description of it.

Our initial summary of these figures said the optimism bias flipped the decision in two of four cases. It is three of four. We caught it by reading the table rather than our own sentence about the table, which is the third time in this series that method has caught an error and the only method that reliably has.

Four observations, ours.

A thirty percent optimism bias on remaining cost is enough to reverse most of these decisions, and thirty percent is modest against what the project overrun literature reports.

Not one of those wrong calls involves sunk cost reasoning. Every decision-maker in that table is applying the correct rule, correctly, to a wrong number.

Which means a large share of continued-too-long projects may be forecasting failures wearing a psychological label, and the two have entirely different remedies. Sunk cost calls for changing who decides. Optimistic R calls for changing how the estimate is produced, which is a reference class problem.

And this is testable inside a firm. Ask whether the remaining-cost estimate has moved. A project genuinely worth continuing has an R that falls as work proceeds. A project in trouble has an R that stays roughly constant no matter how much is spent, and that pattern is visible in the numbers without any judgment about anyone's psychology.

Time Is Not Like Money

A boundary condition with a title that states it.

Reference lists identify Soman, D. (2001), The Mental Accounting of Sunk Time Costs: Why Time is not Like Money, Journal of Behavioral Decision Making, 14(3), 169–185[15].

We did not obtain this paper and report its title, which asserts a difference, without its findings.

Two observations, ours.

If sunk time behaves differently from sunk money, then a substantial part of professional practice sits in the less-studied case. A firm's largest sunk investment in most engagements is hours, not cash.

And we would not guess the direction. We report that a paper exists asserting the difference and that we could not obtain what the difference is, which is unsatisfying and accurate.

Windfall Money Behaves Differently

A second boundary condition, same limitation.

Reference lists identify Soman, D., and Cheema, A. (2001), The Effect of Windfall Gains on the Sunk-Cost Effect, Marketing Letters, 12(1), 51–62[15].

We did not obtain it.

Two observations, ours.

The existence of the question is itself informative. If the source of the money moderates the effect, then the effect is not a pure property of the amount, which is what a mental accounting account would predict and what a strict economic account would not.

And the commercial analogue is direct and worth naming even without the finding: grant-funded, insurance-funded and windfall-funded projects may face different escalation dynamics from ones funded out of operating cash. We flag that as our own conjecture, unsupported by anything we obtained.

Groups Are Not Individuals

A third boundary, connecting to the one meta-analytic number we have.

Reference lists identify Smith, C. M., Tindale, R. S., and Steiner, L. (1998), Investment Decisions by Individuals and Groups in Sunk Cost Situations: The Potential Impact of Shared Representations, Group Processes and Intergroup Relations, 1(2), 175–189[14].

We did not obtain it and report the title only.

Two observations, ours.

This sits beside the meta-analysis's finding that group identity or cohesiveness strength was the only significant social determinant, at ρ = .307 on three studies.

Together they point at a location most escalation advice ignores. The unit of analysis may be the committee rather than the individual, and if so, replacing the decision-maker does nothing while replacing the decision procedure might.

Action, Inaction, And The Framing Result

A more recent line, published in a leading journal.

A 2018 paper is titled When Action-Inaction Framing Leads to Higher Escalation of Commitment: A New Inaction-Effect Perspective on the Sunk-Cost Fallacy[10].

A related paper is identified as van Putten, M., Zeelenberg, M., and van Dijk, E. (2010), Who throws good money after bad? Action vs. state orientation moderates the sunk cost fallacy, Judgment and Decision Making, 5, 33–36[10].

We obtained neither paper and report titles and citations.

Three observations, ours.

Both titles assert that how the choice is framed as action or inaction moderates the effect, which is the twenty-eighth article's framing literature arriving in this one.

The practical implication, if the titles describe what the papers found, is that whether continuing is presented as the default or as an active choice changes behaviour. Most project reviews present continuing as the default, and stopping as the action requiring justification.

And that is a design choice a firm controls completely. A review in which continuation must be actively re-approved is a different instrument from one in which continuation happens unless someone objects, and this series' article on defaults documents how much that distinction can matter elsewhere.

The Monuments

The cases the literature returns to, reported as the literature reports them.

A survey names studies demonstrating escalation in project planning and management including Expo 86 in Vancouver, the Sydney Opera House, the Shoreham nuclear power plant, and the Denver International Airport[4].

We did not obtain any of those case studies and report only that the survey names them.

Three observations, ours.

These are selected on outcome, which is the thirtieth article's problem in this literature. Famous overruns are studied because they overran; projects that were continued through difficulty and succeeded are not in the sample.

The Sydney Opera House is the sharpest illustration of the difficulty. It overran catastrophically and is now among the most recognised buildings in the world. Whether it is an escalation case depends entirely on the V in R < V, and V includes six decades of tourism and civic identity that nobody estimated in 1963.

And that is not a quibble. Case studies of escalation are retrospective judgments about whether V was correctly estimated, made by people who know how it turned out, which is precisely the forty-third and forty-fourth articles' problem.

Why Awareness Is Not A Remedy

The prescription that dominates business writing on this topic, and our objection to it. Ours.

Four points.

The standard advice is to teach people to ignore sunk costs. On the arithmetic above, that advice is incomplete in a dangerous way: it tells people which number to ignore without telling them which numbers to estimate, and a manager who ignores S while accepting an optimistic R still gets the wrong answer.

Where the driver is principal-agent conflict, awareness is beside the point. A manager whose career ends with the project is not confused, and explaining the fallacy to them addresses nothing.

Where the driver is group cohesion, on the one meta-analytic result we obtained, the individual's awareness is not the operative variable at all.

And the forty-fourth article established the general form of this objection with unusual clarity: subjects who believed outcomes should not affect their judgments still let outcomes affect their judgments. Stating a principle is not a control.

What Actually Works

Our own prescription, derived from the arithmetic and the moderators rather than from any intervention study. Untested, and offered as reasoning.

Track R, not S. Require every project review to produce a fresh estimate of remaining cost and remaining value, and record both. The amount spent should not appear in the decision document except as an input to estimating what remains.

Watch whether R falls. This is the diagnostic we would put first. On a healthy project, remaining cost declines as work is completed. On a failing one, R stays flat while S climbs, and that signature is visible in the numbers without any judgment about anyone's motives.

Make continuation the action. If the framing literature's titles describe what those papers found, then requiring active re-approval to continue is a different instrument from requiring an objection to stop.

Separate the estimator from the advocate. The optimism analysis above is a forecasting problem, and the person producing R should not be the person whose project it is.

Supply the opportunity cost. The meta-analysis predicts opportunity cost information reduces escalation, and a project review that does not name what else the capacity could do has withheld the most decision-relevant fact available.

And attend to the committee, not just the manager. The only significant social determinant reported was group cohesion, and if that holds the remedy is procedural rather than personal.

Our Honest Position

Where we would leave a reader, stated without hedging.

Five statements.

Escalation is real and the responsibility effect is shakier than you have been told. A 2009 paper challenges it, cited inside the field's own meta-analysis; a 2021 attempt with 365 participants failed to find it.

The arithmetic is not in dispute and is routinely misapplied in both directions. This is the part we would teach first, because it requires no research literature at all and it corrects an error that the standard cure actively causes.

Most of what a firm will encounter is not the fallacy. It is optimistic estimates of remaining cost, career risk attached to abandonment, and the absence of a named alternative use for the capacity. All three produce the same visible behaviour and none is fixed by explaining a bias.

The evidence is better organised than most literatures in this series and thinner than its prominence suggests. A comprehensive meta-analysis exists, which is more than most topics here can claim, and its single reported social determinant rests on three studies.

And the famous cases prove less than they appear to. They are selected on outcome and judged in hindsight, and at least one of them is now a landmark.

What To Do

Delete the amount spent from the decision. The correct test is remaining cost against remaining value, and the amount already spent produces error whichever direction you consult it in.

Distrust the cure as much as the disease. "We have spent too much to keep going" is wrong on three of our five illustrative cases, and it is the advice usually given as the remedy for sunk cost thinking.

Interrogate the remaining-cost estimate first. On our own arithmetic, a thirty percent optimism bias flips three of four decisions with no sunk cost reasoning involved anywhere.

Ask whether R has fallen since the last review. A healthy project's remaining cost declines with work done. A failing project's does not, and that is a number rather than a judgment.

Name the alternative use of the capacity. The meta-analysis predicts opportunity cost information reduces escalation, and most project reviews omit it entirely.

Make continuing require a decision. Two papers assert that action versus inaction framing moderates the effect, and continuation is the default in almost every review process.

Look at the committee. The only significant social determinant reported was group identity or cohesiveness strength, which locates part of the problem in team dynamics rather than individual reasoning.

Do not rely on awareness training. Where the cause is career risk or genuine constraint, awareness addresses nothing, and this series has documented elsewhere that believing a principle does not stop people violating it.

Be sceptical of famous escalation cases, including the ones you like. They are selected on outcome and judged with hindsight, and the completed value of a long project is frequently unknowable at the time the decision is criticised.

The Limits Of This Analysis

Several caveats matter, and this article's sourcing is uneven enough that they deserve care. This article discusses research on decision making and is not investment, capital allocation, project management or accounting advice. Everything is verified to August 2026. We did not obtain the 1976 founding study, and everything said about it here comes from later descriptions, which for a paper of this influence is a serious limitation. We did not obtain the 1985 sunk cost paper and report no findings from it. We obtained only fragments of the 2012 meta-analysis: its abstract, one summary sentence, two of sixteen hypothesis statements, and one numerical result. We did not obtain its effect tables or moderator results, and we report no effect size from it other than the single correlation of .307. Our confidence interval around that correlation understates the true uncertainty, because it treats 138 observations as a single study when they came from three, and we lack the heterogeneity estimate needed to do it properly; we report it to show scale and not as an answer. Our binomial test of the 14-of-16 claim is our own crude directional check, not the authors' analysis, and it is weakened by the fact that hypotheses in a review of a mature literature are not formulated blind. We obtained the 2021 replication's abstract and not the paper, so we cannot assess its design, incentives, or whether its operationalisation was faithful to the original. We did not obtain the 2009 study challenging self-justification, any of the four moderator papers, any of the named case studies, or the paper on indirect self-justification effects, and report all of these from citation records or descriptions. One source is a personal website, cited once for a summary judgment and flagged at that use. All arithmetic is ours and uses invented project figures throughout; our first description of the optimism table was wrong and is corrected in the body. The prescriptions in the final sections are our own reasoning from the arithmetic and the reported moderators, are supported by no intervention study we obtained, and are untested.

Related Reading

No. 53 The Mythical Number Two The framework behind every corporate behavioural training deck was called a convenient and seductive myth in a leading journal. Its defenders replied... Behavioural Finance The Planning Fallacy Why the remaining-cost estimate is the number that actually decides these cases. Behavioural Finance They Knew They Should Not, And Did Why stating a principle is not a control, and why famous cases are judged in hindsight.

Frequently Asked Questions

What is the correct rule?

Compare the remaining cost of finishing against the value of the finished thing. If remaining cost is lower, continue. The amount already spent does not enter the calculation, and its only legitimate use is as an input to estimating what remains.

Isn't "we've spent too much to stop" the fallacy?

Yes, and so is its opposite. On our five illustrative cases, "we have spent too much to stop" is wrong twice and "we have spent too much to keep going" is wrong three times. Both consult the amount spent, which is irrelevant. The second is usually taught as the cure for the first.

Why does more spending push toward continuing?

Because remaining cost falls as work is completed. A project with 800 spent and 400 remaining against a value of 900 should be finished. A manager who abandons it because the budget is blown destroys 500 of value, and that is the naive correction for sunk cost doing the damage.

Did the founding study replicate?

A 2021 attempt with 365 participants reports no evidence of escalation, and that participants simply invested more in more promising projects. The meta-analysis also cites a 2009 paper challenging the responsibility effect. We obtained the replication's abstract and not the paper, and we would not call the question settled.

So how much of this is really sunk cost?

Less than the label suggests. On our own arithmetic, a thirty percent optimism bias in estimating remaining cost flips the correct decision in three of four cases with no sunk cost reasoning anywhere. A reading of the meta-analysis concludes the behaviours involved range from image protection to principal-agent conflict to simple lack of options.

What is the single best diagnostic?

Whether the remaining-cost estimate has fallen since the last review. On a healthy project it declines as work is done. On a failing one it stays flat while spending climbs. That pattern is arithmetic rather than a judgment about anyone's psychology, and it is visible in documents a firm already produces.

Does training people about the bias help?

We doubt it, for three reasons. Where the driver is career risk the manager is not confused. Where it is group cohesion the individual is not the operative unit. And this series has documented that subjects who explicitly believed outcomes should not affect their judgments were affected anyway.

References

1. Sleesman, D. J., Conlon, D. E., McNamara, G., & Miles, J. E. (2012). Cleaning Up the Big Muddy: A Meta-Analytic Review of the Determinants of Escalation of Commitment. Academy of Management Journal, 55(3), 541–562. Citation confirmed independently across multiple scholarly reference lists, including a Springer journal article's bibliography, a Springer book chapter's bibliography, and an academic preprint's reference list. Note: citation confirmation across independent reference lists. The paper itself was obtained only in fragments, from the hosted copy at reference 5. link.springer.com
2. Journal record for Escalation of commitment is independent of numeracy and cognitive reflection. Failed replication and extension of Staw (1976), Economics and Business Review, June 2021, reproducing the abstract in full, on escalation of commitment having been first demonstrated by Staw (1976) as the tendency for an individual to increase their commitment to a failing course of action when personally responsible for the negative consequences; on an attempt having been made to replicate this finding and to test whether individual differences in numeracy and cognitive reflection could help explain such an effect; on no evidence for escalation of commitment having been found amongst the participants, with N = 365; on participants simply having invested more in more promising projects; on no evidence having been found that numeracy or cognitive reflection predict escalation behaviour; and on the validity of escalation of commitment behaviour being discussed, suggesting that future work should explore the boundary conditions of such an effect. Note: an open journal directory record reproducing the abstract in full. We did not obtain the paper, and cannot assess its design, incentives, or whether its operationalisation was faithful to the 1976 original. doaj.org
3. Personal website essay on whether sunk costs are fallacies, stating that meta-analysis of escalation effect studies suggests sunk cost behavior is not one thing but reflects a variety of theorized behaviors and effects of varying rationality, ranging from protecting one's image and principal-agent conflict to lack of information or options, not all of which can be regarded as a simple cognitive bias to be fixed by greater awareness, and citing the 2012 meta-analysis in support. Note: a personal website, not a peer-reviewed source. Cited once, for this summary judgment about the meta-analysis, and flagged as such in the body. A reader should weigh it accordingly. gwern.net
4. Academic preprint on escalation of commitment in large language models, literature section, on organizational scholars having been intrigued since the late 1970s by decision makers' tendency to persist in failing courses of action; on escalation of commitment having been first described by Staw (1976), demonstrating that individuals responsible for an investment experiencing negative outcomes tend to persist in continuing that course of action, effectively ignoring warning signs; on the tendency having been covered at length in finance, marketing, accounting and information systems; on it having been demonstrated in project planning and management studies including Expo 86 in Vancouver, the Sydney Opera House, the Shoreham nuclear power plant and the Denver International Airport; on economists having described similar tendencies as the sunk-cost fallacy and lock-in; and on the idioms throwing good money after bad and in for a penny, in for a pound. Note: an academic preprint's literature review, not a primary source. We obtained none of the case studies it names. arxiv.org
5. Hosted copy of the 2012 meta-analysis, from which we obtained fragments including the statement that the only social determinant having a significant relationship with escalation was group identity or cohesiveness strength, reported as H15 with a correlation of .307, n = 138 and k = 3; the summary statement that in total the authors found support for 14 of their 16 main effect predictions, with only two antecedents failing, both in the social determinants category and including public evaluation of the decision; the statement that, drawing on recent evidence challenging the role of self-justification and the responsibility effect, the authors predicted that personal responsibility for the initial decision would lead to higher levels of escalation; and hypothesis headings covering volition in the initial decision, covariation of sunk cost and project completion effects, and influence of opportunity cost information. Note: a hosted copy obtained in fragments only. We did not obtain the paper's effect tables, moderator results or methods section, and report no effect size from it beyond the single correlation quoted. iot.ntnu.no
6. Publisher record for a 1992 review article in Academy of Management Review on the escalation of commitment to a failing course of action, reproducing the abstract in full, on escalating commitment referring to the tendency for decision makers to persist with failing courses of action; on the article first reviewing evidence suggesting that escalation is determined at least in part by decision makers' unwillingness to admit that their prior allocation of resources was in vain, being the self-justification explanation; on a distinction being drawn between alternative explanations, some designed to replace self-justification and others intended to supplement it by adding explanatory power beyond it; and on there being little evidence that the replacement theories provide a better explanation than self-justification, while theories designed to supplement it are likely to lead to a more complete explanation. Note: the publisher's record, reproducing the abstract in full. We did not obtain the review itself. journals.aom.org
7. Springer book chapter reference list confirming Staw, B. M. (1976), Knee-deep in the Big Muddy: A study of escalating commitment to a chosen course of action, Organizational Behavior and Human Performance, 16(1), 27–44; Staw, B. M., and Ross, J. (1987), Behavior in escalation situations: Antecedents, prototypes, and solutions, Research in Organizational Behavior; Arkes, H. R., and Blumer, C. (1985), The psychology of sunk cost, Organizational Behavior and Human Decision Processes, 35(1), 124–140; and Sleesman and colleagues (2012), Academy of Management Journal, 55(3), 541–562. Note: a reference list; citations only. We obtained none of the works listed. link.springer.com
8. Springer book chapter reference list independently confirming Arkes, H. R., and Blumer, C. (1985), The psychology of sunk cost, Organizational Behavior and Human Decision Processes, 35(1), 124–140, together with Soman, D. (2001), Journal of Behavioral Decision Making, 14(3), 169–185, and Soman, D., and Cheema, A. (2001), Marketing Letters, 12(1), 51–62. Note: a second independent reference list, used to confirm the 1985 citation. We did not obtain the paper. link.springer.com
9. Academic preprint surveying behavioural biases in project management, on escalation of commitment being the tendency to justify increased investment based on cumulative prior investment despite new evidence that the decision may be wrong; on it applying to individuals, groups and whole organizations; on it having been first described by Staw (1976) with later work by Brockner (1992), Staw (1997), Sleesman and colleagues (2012) and others; on economists using related terms such as the sunk-cost fallacy, citing Arkes and Blumer (1985), and lock-in; and on the rational approach when deciding whether to invest further being to disregard what has already been invested. Note: an academic preprint's survey section, not a primary source. arxiv.org
10. Publisher record for a 2018 article in Psychological Science, When Action-Inaction Framing Leads to Higher Escalation of Commitment: A New Inaction-Effect Perspective on the Sunk-Cost Fallacy, stating that escalation of commitment to a failing course of action occurs in the presence of sunk costs, negative feedback that things are deviating from expectations, and a decision between escalating or not, our source truncating; together with its reference list identifying van Putten, M., Zeelenberg, M., and van Dijk, E. (2010), Who throws good money after bad? Action vs. state orientation moderates the sunk cost fallacy, Judgment and Decision Making, 5, 33–36; Whyte, G. (1986), Escalating commitment to a course of action: A reinterpretation, Academy of Management Review, 11, 311–321; and Tykocinski, O. E., and Ortmann, A. (2011), The lingering effects of our past experiences: The sunk-cost fallacy and the inaction-inertia effect, Social and Personality Psychology Compass, 5, 653–664. Note: a publisher record; we obtained a truncated definition and the reference list, and none of the papers named. journals.sagepub.com
11. Academic preprint evaluating cognitive biases in language models, describing escalation of commitment, also known as commitment bias, as the act of carrying on with questionable or failing courses of action, citing Sleesman and colleagues (2018); noting it was first examined in Staw (1976) and has been studied in finance, governance, and research and development; and describing an experimental paradigm based on Staw (1981) in which a past decision is attributed either to the subject or to another independent actor. Note: an academic preprint, not a primary source. The quoted definition is attributed there to a 2018 work by the same lead author as the 2012 meta-analysis, which we did not obtain. arxiv.org
12. Bibliographic service record for the 2012 meta-analysis, reproducing its abstract, on the topic of escalation of commitment having intrigued the organizational sciences for over 35 years; on a variety of theoretical explanations having been offered and numerous constructs examined as antecedents; on little effort having been made to systematically investigate these various accounts; and on the authors using meta-analysis to present a comprehensive overview of the many determinants found in the literature, an analysis of the power of different theoretical perspectives, and an examination of the relative efficacy of the various theories. Note: a bibliographic service record. Our source for the abstract; the paper itself was obtained only in fragments. semanticscholar.org
13. Repository copy of the 2012 meta-analysis carrying its hypothesis table, on hypothesis 1 predicting a negative relationship for decision risk on the reasoning that risk increases the likelihood of loss and the salience of loss potential to decision makers, lessening the likelihood of escalation even in the face of information on previous performance; and on hypothesis 2 predicting a negative relationship for opportunity cost information on the reasoning that such information provides a clear signal, our source truncating. Note: a repository copy obtained in fragments. We obtained two of sixteen hypothesis statements and cannot report the remainder. researchgate.net
14. Publisher record for an article on how self-justification indirectly drives escalation of commitment, Schmalenbach Business Review, reproducing the abstract, on self-justification being the most examined and empirically supported explanation of escalation of commitment; on the authors arguing, from motivated reasoning theory, that the need for self-justification affects escalation indirectly via other cognitive processes, influencing the decision maker's selective perception, sunk cost effect and overoptimism; and on the authors investigating escalation in the venture capital industry to strengthen the external validity of previous laboratory studies; together with its reference list identifying Smith, C. M., Tindale, R. S., and Steiner, L. (1998), Investment Decisions by Individuals and Groups in Sunk Cost Situations: The Potential Impact of Shared Representations, Group Processes and Intergroup Relations, 1(2), 175–189. Note: a publisher record; we obtained the abstract and reference list, and not the paper or the 1998 study it cites. link.springer.com
15. Springer book chapter reference list identifying Soman, D. (2001), The Mental Accounting of Sunk Time Costs: Why Time is not Like Money, Journal of Behavioral Decision Making, 14(3), 169–185; Soman, D., and Cheema, A. (2001), The Effect of Windfall Gains on the Sunk-Cost Effect, Marketing Letters, 12(1), 51–62; and Soman, D., and Gourville, J. T. (2001), Transaction Decoupling: How Price Bundling Affects the Decision to Consume, Journal of Marketing Research, 38(1), 30–44. Note: a reference list; citations only. We obtained none of these papers and report their titles, which assert boundary conditions, without their findings. researchgate.net

This article discusses research on decision making and is not investment, capital allocation, project management or accounting advice. The 1976 founding study and the 1985 sunk cost paper were not obtained. The 2012 meta-analysis was obtained only in fragments, and no effect size from it is reported beyond a single correlation whose confidence interval, as calculated here, understates the true uncertainty because it treats three studies as one. The 2021 replication was obtained as an abstract only. One source is a personal website, cited once and flagged. All arithmetic is the authors' own and uses invented project figures; one description of it was wrong on first pass and is corrected in the body.