5D Library › Bayesian Search in Unknown Landscapes, Optimal Stopping and Gittins Indices · 中文

Chapter 5 Sunk Costs: Rational Persistence, Fallacious Persistence and Stop-Loss Boundaries Set in Advance

The evidence for a sunk-cost effect

A cost that has already been paid and cannot be recovered whatever one decides now is a sunk cost. The standard conclusion of economics is that a rational decision looks only at forward value from now on, and sunk costs should not affect it. Thaler in 1980 listed “I have paid for it, so I must use it” as a typical departure from this principle in consumer choice (Thaler 1980); Staw’s 1976 experiment showed that people personally responsible for an initial decision were more inclined to commit further resources to a failing course of action, which he called being “knee-deep in the big muddy” (Staw 1976).

Evidence

Arkes and Blumer (1985). The paper reports ten experiments (Arkes and Blumer 1985). In the ski-trip problem, participants were told they had mistakenly booked two trips for the same weekend, one costing $100 and the other $50 but expected to be more enjoyable; 33 of 61 (about 54 %) chose the more expensive trip. In the field experiment with theatre season tickets, subscribers were randomly assigned to pay the normal price ($15) or to receive a discount of $2 or $7, and 54 people could finally be analysed: over the first five plays of the season the three groups attended 4.11, 3.32 and 3.29 plays on average, the full-price group the most; over the last five plays they attended 2.28, 1.54 and 2.18, with no significant difference between groups. In the “radar-blank plane” problem, with $10 million already invested and the project 90 % complete, 41 of 48 (about 85 %) were willing to invest the last $1 million; with no prior investment, only 10 of 60 (about 17 %) were willing to invest.

These results are persuasive, but they have limits: the ski and aircraft problems are hypothetical; the aircraft problem mixes “already invested” with “nearly complete”; and the theatre experiment was small, with the effect appearing only in the first half of the season.

Evidence

Later reviews and replications. The meta-analysis by Roth, Robbert and Straus, published in Business Research in 2015, included 45 studies and about 100 effect sizes, with a pooled standardised mean difference of 0.496: 0.58 for decisions about using something already paid for and 0.44 for decisions about continuing a project. In decisions about use, older samples showed smaller effects; most of the included studies used hypothetical scenarios, but in decisions about progress the effect was larger in surveys and field studies than in scenario experiments (Roth et al. 2015). In the meta-analysis of escalation of commitment by Sleesman and colleagues (2012; 166 independent samples), the average correlation between sunk costs and escalation was 0.24 (95 % confidence interval about 0.11 to 0.37), but the 80 % credibility interval included zero; in the 28 samples in which sunk costs did not vary together with “closeness to completion”, the correlation was only about 0.10 (95 % confidence interval including zero), against about 0.53 in the 5 samples in which the two were varied together (Sleesman et al. 2012). Friedman and colleagues (2007), in a laboratory task with real monetary incentives, found only a “small and erratic” sunk-cost effect: in the most favourable analysis, a high sunk cost raised the probability of staying from 0.50 to about 0.52 (Friedman et al. 2007). On the other hand, the “Many Labs” project tested 13 classic effects across 36 samples and 6,344 participants and replicated 10 of them, including a sunk-cost effect in a hypothetical scenario (Klein et al. 2014).

A more accurate statement, then, is that the sunk-cost effect is robust in hypothetical scenarios, while its size in repeated laboratory decisions with real incentives is uncertain (and the surveys and field studies in the meta-analysis of Roth and colleagues do not support the inference that it is smaller in real life), and it is often entangled with closeness to completion, personal responsibility and the wish not to appear wasteful. McAfee, Mialon and Mialon (2010) go further and point out that in many situations attending to sunk costs is rational: past investment may carry information about the project’s quality, quitting may damage one’s reputation, and constraints of money and time can make “continuing” the better deal (McAfee et al. 2010). A normative model makes these distinctions precise.

Continue or abandon: a one-armed bandit

Calculation

A project runs at most T=40T = 40 trials; each costs c=3c = 3 and succeeds with unknown probability pp, earning R=10R = 10 when it does. The prior is p∼Beta(1,2)p \sim \mathrm{Beta}(1, 2) (mean 1/3), and break-even is p=c/R=0.3p = c/R = 0.3. After any trial the project can be abandoned, with zero payoff thereafter. The state is the number of successes ss and failures ff so far: V(s,f)=max⁡{0,mR−c+mV(s+1,f)+(1−m)V(s,f+1)},m=1+s3+s+f.V(s, f) = \max\{0,\; mR – c + m\,V(s+1, f) + (1-m)\,V(s, f+1)\}, \qquad m = \frac{1+s}{3+s+f}. The 3(s+f)3(s+f) already spent appears nowhere in this recursion: it belongs to the past, and the decision depends only on the posterior (the evidence) and the number of trials remaining. The program first gives the smallest number of successes needed to continue at various points under the optimal rule, and then uses the same recursion to compute exactly the expected value of the whole project under several rules: the optimal rule; never abandoning; a short-sighted rule that abandons as soon as the posterior mean falls below 0.3 (the myopic rule); and two “sunk-cost rules” that follow the optimal rule but refuse to abandon after 10 or 20 trials have been run. All numbers are illustrative.

# Continue or abandon (illustrative numbers). Up to T = 40 trials; each costs c = 3 and
# succeeds with unknown p, earning R = 10; prior Beta(1, 2); break-even p = 0.3. After any
# trial one may abandon (0 from then on). Money already spent, 3 (s + f), appears nowhere:
#   V(s, f) = max(0, m R - c + m V(s+1, f) + (1-m) V(s, f+1)),  m = (a+s)/(a+b+s+f).
# A rule(t, s, m) -> continue? is evaluated exactly by the same recursion.
T, C, R = 40, 3.0, 10.0

def solve(rule=None, a=1, b=2):
    nxt, keep = [0.0] * (T + 1), {}
    for t in range(T - 1, -1, -1):
        cur = []
        for s in range(t + 1):
            m = (a + s) / (a + b + t)
            go = m * R - C + m * nxt[s + 1] + (1 - m) * nxt[s]
            if go > 0 and t not in keep:
                keep[t] = (s, m)                    # fewest successes to continue
            cur.append(max(0.0, go) if rule is None else go if rule(t, s, m) else 0.0)
        nxt = cur
    return nxt[0], keep

if __name__ == "__main__":
    best, keep = solve()
    print("optimal: continue before trial t + 1 iff successes >= s_min")
    print("   trials done t   s_min   posterior mean at s_min   myopic break-even")
    for t in (0, 5, 10, 20, 30, 35, 39):
        print(f"   {t:8d}        {keep[t][0]:3d}           {keep[t][1]:.3f}"
              "                  0.300")
    opt = lambda t, s: s >= keep[t][0]
    rules = (("never abandon", lambda t, s, m: True),
             ("myopic: continue iff m >= 0.3", lambda t, s, m: m >= 0.3),
             ("optimal, but never quit after 10 trials",
              lambda t, s, m: opt(t, s) or t >= 10),
             ("optimal, but never quit after 20 trials",
              lambda t, s, m: opt(t, s) or t >= 20),
             ("myopic, and never quit after 10 trials",
              lambda t, s, m: m >= 0.3 or t >= 10))
    print("expected value under the prior Beta(1, 2)")
    print(f"   {'optimal (forward-looking)':42s} {best:7.3f}")
    for name, rule in rules:
        print(f"   {name:42s} {solve(rule)[0]:7.3f}")
    print(f"prior Beta(1, 3), mean 0.25: myopic never starts (0.000);"
          f" optimal {solve(None, 1, 3)[0]:.3f}")
# optimal: continue before trial t + 1 iff successes >= s_min
#    trials done t   s_min   posterior mean at s_min   myopic break-even
#           0          0           0.333                  0.300
#           5          1           0.250                  0.300
#          10          2           0.231                  0.300
#          20          5           0.261                  0.300
#          30          9           0.303                  0.300
#          35         10           0.289                  0.300
#          39         12           0.310                  0.300
# expected value under the prior Beta(1, 2)
#    optimal (forward-looking)                   37.881
#    never abandon                               13.333
#    myopic: continue iff m >= 0.3               26.896
#    optimal, but never quit after 10 trials     35.552
#    optimal, but never quit after 20 trials     37.599
#    myopic, and never quit after 10 trials      26.822
# prior Beta(1, 3), mean 0.25: myopic never starts (0.000); optimal 15.358

The threshold of the optimal rule has a clear shape (Figure 4, left). Early on, continuing is rational even when the posterior mean is below break-even: after 5 trials, one success (posterior mean 0.250) is enough to continue; after 10 trials, two successes (0.231). The reason is that continuing brings information: if pp is in fact good, the later rewards are large; if it is poor, that soon becomes clear and the project can be dropped. The nearer the end, the fewer the trials in which the information can be used, and the closer the threshold moves to break-even: after 30 trials it requires a posterior mean of about 0.303, after 39 trials 0.310 (the threshold jumps around 0.3 because the number of successes is a whole number).

The rules differ greatly in expected value. The optimal rule achieves 37.88; never abandoning only 13.33. The myopic rule achieves 26.90, about 29 % less than optimal: its problem is not persisting too long but quitting too early, ignoring the value of information. Adding to the optimal rule the sunk-cost tendency “never quit after 10 trials” lowers the value to 35.55, a loss of about 6 %; “never quit after 20 trials” costs little (37.60), because projects that survive to the twentieth trial are mostly good anyway. The last line shows an extreme case of the value of information: with a more pessimistic prior (Beta(1,3)\mathrm{Beta}(1, 3), mean 0.25, below break-even) the myopic rule never starts, for a value of 0, whereas the optimal rule tries a few trials first and has an expected value of 15.36.

Note

Two errors. The model splits “persistence” into two kinds: persisting because of the evidence and the value of information, which is rational, and persisting because of the money already spent, which is the sunk-cost fallacy. A simple test separates them: if the same evidence and the same remaining opportunities were handed to someone who had spent nothing, would that person continue? With the parameters of this example, quitting too early costs more than sunk-cost persistence; this is a result for illustrative parameters, not a general conclusion, and with other parameters the relative size of the two errors changes.

This structure is close to that of real options: abandoning a project is like exercising an option, while continuing to experiment keeps open the chance to decide later. Volume 16, Cross-Cultural Contracts, Real Options and Game Theory of Southeast Asian Market Entry, discussed the option value of waiting and abandoning, and Dixit and Pindyck’s Investment under Uncertainty (1994) gives the systematic theory (Dixit and Pindyck 1994); this chapter revisits it only from the angle of stopping times and does not repeat its programs.

Left: in the one-armed bandit, the lowest posterior mean at which the optimal rule continues (step line) and the myopic break-even point c/R = 0.3 (dashed); the horizontal axis is the number of trials already run, whose cost is sunk. Early thresholds lie below break-even: this is the value of information. Right: five illustrative paths of a sequential probability ratio test and the two boundaries set in advance (upper: go ahead; lower: abandon).

Stop-loss boundaries written in advance: the sequential probability ratio test

In practice the hardest part is often not computing the optimal rule but keeping to it under pressure. One remedy is to write down the stopping boundaries before starting. Wald’s sequential probability ratio test (SPRT) of 1945 provides such boundaries (Wald 1945).

Calculation

Each trial converts with probability pp. The null hypothesis H0H_0 is p=0.20p = 0.20 (not worth pursuing); the alternative H1H_1 is p=0.30p = 0.30 (worth pursuing); the tolerated error rates are α=0.05\alpha = 0.05 (treating a worthless project as worthwhile) and β=0.10\beta = 0.10 (abandoning a worthwhile one). After each trial the log-likelihood ratio is updated, adding ln⁡(0.30/0.20)\ln(0.30/0.20) for a success and ln⁡(0.70/0.80)\ln(0.70/0.80) for a failure; above ln⁡A\ln A one stops and goes ahead, below ln⁡B\ln B one stops and abandons, with Wald’s approximations A=(1−β)/αA = (1-\beta)/\alpha and B=β/(1−α)B = \beta/(1-\alpha). The program turns the boundaries into a table — “after nn trials, abandon with at most so many successes, go ahead with at least so many” — and then uses 20,000 simulated runs to find, for several true values of pp, the probability of going ahead and the number of trials needed, compared with a fixed-sample test with the same α\alpha and β\beta.

# Wald's SPRT as a pre-set stop-loss rule (illustrative). H0: p = 0.20 (not worth it)
# against H1: p = 0.30; alpha 0.05, beta 0.10; A = (1 - beta)/alpha, B = beta/(1 - alpha).
# Compared with the fixed sample size giving the same alpha and beta (normal approx.).
import math, random
from statistics import NormalDist

P0, P1, ALPHA, BETA = 0.20, 0.30, 0.05, 0.10
LA, LB = math.log((1 - BETA) / ALPHA), math.log(BETA / (1 - ALPHA))
UP, DOWN = math.log(P1 / P0), math.log((1 - P1) / (1 - P0))

def sprt(p, rng, cap=2000):
    llr, n = 0.0, 0
    while LB < llr < LA and n < cap:
        llr += UP if rng.random() < p else DOWN
        n += 1
    return llr >= LA, n

def boundaries(n):        # successes needed after n trials to accept H1 / to abandon
    acc = math.ceil((LA - n * DOWN) / (UP - DOWN))
    rej = math.floor((LB - n * DOWN) / (UP - DOWN))
    return acc, rej

if __name__ == "__main__":
    z = NormalDist().inv_cdf
    nfix = math.ceil(((z(1 - ALPHA) * math.sqrt(P0 * (1 - P0))
                       + z(1 - BETA) * math.sqrt(P1 * (1 - P1))) / (P1 - P0)) ** 2)
    print(f"ln A = {LA:.3f}, ln B = {LB:.3f};"
          f" fixed-sample n with same alpha, beta: {nfix}")
    print("pre-set boundaries: after n trials, abandon if successes <= lower,"
          " go ahead if >= upper")
    for n in (10, 20, 40, 80, 120):
        acc, rej = boundaries(n)
        low = "  none      " if rej < 0 else f"{rej:6d} ({rej / n:.3f})"
        print(f"   n = {n:3d}:  lower {low}   upper {acc:3d} ({acc / n:.3f})")
    rng = random.Random(1945)
    print("   true p   P(go ahead)   mean trials   90th percentile   share of fixed n")
    for p in (0.15, 0.20, 0.25, 0.30, 0.35):
        res = [sprt(p, rng) for _ in range(20000)]
        ns = sorted(r[1] for r in res)
        mean = sum(ns) / len(ns)
        print(f"    {p:.2f}      {sum(r[0] for r in res) / 20000:.3f}"
              f"        {mean:6.1f}          {ns[18000]:4d}"
              f"             {mean / nfix:.2f}")
# ln A = 2.890, ln B = -2.251; fixed-sample n with same alpha, beta: 156
# pre-set boundaries: after n trials, abandon if successes <= lower, go ahead if >= upper
#    n =  10:  lower   none         upper   8 (0.800)
#    n =  20:  lower      0 (0.000)   upper  11 (0.550)
#    n =  40:  lower      5 (0.125)   upper  16 (0.400)
#    n =  80:  lower     15 (0.188)   upper  26 (0.325)
#    n = 120:  lower     25 (0.208)   upper  36 (0.300)
#    true p   P(go ahead)   mean trials   90th percentile   share of fixed n
#     0.15      0.001          43.9            74             0.28
#     0.20      0.043          81.4           163             0.52
#     0.25      0.465         130.8           273             0.84
#     0.30      0.906          89.8           172             0.58
#     0.35      0.988          54.5            98             0.35

The boundaries can be printed in advance: after 20 trials with no success, abandon; after 40 trials, abandon with at most 5 successes and go ahead with 16; after 80 trials, 15 and 26 respectively. When the true pp is 0.20 the probability of going ahead is 0.043, and when it is 0.30 it is 0.906, close to the chosen α=0.05\alpha = 0.05 and 1−β=0.901 – \beta = 0.90 (Wald’s approximations err slightly on the safe side). At these two hypotheses the average number of trials is 81.4 and 89.8, only 52 % and 58 % of the 156 needed by the fixed-sample test; when the true value is further away, the saving is larger (43.9 trials on average at p=0.15p = 0.15). Wald and Wolfowitz proved in 1948 that between two simple hypotheses no test with the same error rates needs a smaller expected sample size (Wald and Wolfowitz 1948).

But when the true pp lies between the two hypotheses (0.25), 130.8 trials are needed on average, and 10 % of runs need more than 273 — more than the fixed-sample test. In practice, therefore, a cap must be set in advance, with a pre-written rule for what to do if the cap is reached without crossing a boundary (for example, to treat the project as “not worth further investment”).

Note

The SPRT and optional stopping. Volume 33, Scientific Methodology, Falsifiability and the Mathematical Statistics of the Replication Crisis, discussed optional stopping: using a test designed for a fixed sample, testing every time one looks at the data and stopping as soon as the result is significant, which pushes the false-positive rate far above its nominal level. The SPRT also decides as the data come in, but its boundaries are designed in advance for sequential use, and its error rates are controlled by design. The difference lies not in “stopping as one goes” but in whether the stopping rule is written down beforehand and consistent with the test.

Hypothesis · not established

The hypothesis that “boundaries set in advance reduce escalation” (not yet established). The author conjectures that writing down an abandonment boundary before a project starts (and having someone not responsible for the project carry out the decision at the boundary) reduces the tendency to add investment out of personal responsibility. A testable prediction: in a Staw-style experiment that manipulates responsibility, participants randomly assigned to “write down a boundary in advance” add investment after negative feedback less often than controls. If the two groups do not differ, the hypothesis loses support.

Summary

The sunk-cost effect is robust in hypothetical scenarios, with a pooled meta-analytic effect size of about 0.5, but in laboratory experiments with real incentives it is small and erratic (surveys and field studies of project decisions show larger effects), and it is often entangled with closeness to completion and personal responsibility. In the normative model, money already spent does not enter the decision; the evidence accumulated and the remaining value of information do, so that continuing below break-even early on can be rational. With illustrative parameters, myopic early quitting loses about 29 %, while adding to the optimal rule the sunk-cost tendency “never quit after 10 trials” loses about 6 %. Wald’s sequential probability ratio test gives stop-loss and go-ahead boundaries that can be written down in advance and needs on average only half to six-tenths of a fixed sample at the two hypotheses, but it can run long when the true value lies between them, so a cap must be set. The next chapter applies these tools to start-up experiments and to the selection of people.

References

Arkes, Hal R., and Catherine Blumer. 1985. “The Psychology of Sunk Cost.” Organizational Behavior and Human Decision Processes 35 (1): 124–40.
Dixit, Avinash K., and Robert S. Pindyck. 1994. Investment Under Uncertainty. Princeton University Press.
Friedman, Daniel, Kai Pommerenke, Rajan Lukose, Garrett Milam, and Bernardo A. Huberman. 2007. “Searching for the Sunk Cost Fallacy.” Experimental Economics 10 (1): 79–104.
Klein, Richard A., Kate A. Ratliff, Michelangelo Vianello, et al. 2014. “Investigating Variation in Replicability: A ‘Many Labs’ Replication Project.” Social Psychology 45 (3): 142–52. https://doi.org/10.1027/1864-9335/a000178.
McAfee, R. Preston, Hugo M. Mialon, and Sue H. Mialon. 2010. “Do Sunk Costs Matter?” Economic Inquiry 48 (2): 323–36. https://doi.org/10.1111/j.1465-7295.2008.00184.x.
Roth, Stefan, Thomas Robbert, and Lennart Straus. 2015. “On the Sunk-Cost Effect in Economic Decision-Making: A Meta-Analytic Review.” Business Research 8 (1): 99–138. https://doi.org/10.1007/s40685-014-0014-8.
Sleesman, Dustin J., Donald E. Conlon, Gerry McNamara, and Jonathan E. Miles. 2012. “Cleaning up the Big Muddy: A Meta-Analytic Review of the Determinants of Escalation of Commitment.” Academy of Management Journal 55 (3): 541–62. https://doi.org/10.5465/amj.2010.0696.
Staw, Barry 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.
Thaler, Richard. 1980. “Toward a Positive Theory of Consumer Choice.” Journal of Economic Behavior and Organization 1 (1): 39–60.
Wald, Abraham. 1945. “Sequential Tests of Statistical Hypotheses.” The Annals of Mathematical Statistics 16 (2): 117–86. https://doi.org/10.1214/aoms/1177731118.
Wald, A., and J. Wolfowitz. 1948. “Optimum Character of the Sequential Probability Ratio Test.” The Annals of Mathematical Statistics 19 (3): 326–39. https://doi.org/10.1214/aoms/1177730197.

Fan, Y. (2026). Bayesian Search in Unknown Landscapes, Optimal Stopping and Gittins Indices: Exploration–Exploitation Trade-offs and Sunk-Cost Truncation in Statistical Decision Theory [未知地形中的贝叶斯搜索、秘书问题与吉廷斯指数] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23810102 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).