Neural mechanisms and personality correlates of the sunk cost effect

The sunk cost effect, an interesting and well-known maladaptive behavior, is pervasive in real life, and thus has been studied in various disciplines, including economics, psychology, organizational behavior, politics, and biology. However, the neural mechanisms underlying the sunk cost effect have not been clearly established, nor have their association with differences in individual susceptibility to the effect. Using functional magnetic resonance imaging, we investigated neural responses induced by sunk costs along with measures of core human personality. We found that individuals who tend to adhere to social rules and regulations (who are high in measured agreeableness and conscientiousness) are more susceptible to the sunk cost effect. Furthermore, this behavioral observation was strongly mediated by insula activity during sunk cost decision-making. Tight coupling between the insula and lateral prefrontal cortex was also observed during decision-making under sunk costs. Our findings reveal how individual differences can affect decision-making under sunk costs, thereby contributing to a better understanding of the psychological and neural mechanisms of the sunk cost effect.


All 23 pairs of destinations in the fMRI task
The defined costs of the tickets in the control condition are provided in parentheses.

Prior research concerning the sunk cost effect
This section briefly describes previous empirical evidence regarding the sunk cost effect.
The first discussion of sunk cost effects came from organizational behavior scholars.
They demonstrated in hypothetical lab experiments that subjects were more likely to continue investing in a project if its previous (sunk) costs were high 1 . This type of decision came to be called a "progress" decision. An influential 1980 paper described In establishing the sunk cost effect, it is extremely important to control for perceptions of any marginal costs (e.g. from investing further in progress decisions) and marginal benefits 5,6 . For example, suppose a company builds a building and expects it to cost $10 million, and values it at $15 million. After finishing half of the building, spending $5 million, they realize it will cost $12 million more to complete it. Suppose the half-finished building is worthless unless it is finished. Then the company must decide whether to spend an additional $12 million to have a building worth $15 million.

13
The company should continue to spend money because the marginal benefit minus the marginal cost is positive (=15-12=+3). When they are done, they will have paid $17 million for a building worth $15 million. But the mistake that was made was underestimating the chance of a cost overrun. Paying the marginal $12 million was not a sunk cost mistake. This age effect is especially important because 72% of the effects came from young college student subjects.
An interesting variable with mixed effects is the length of time between when a cost was first sunk, and when a later decision must be made 10 . Time passage reduces sunk costs effects for utilization decisions but increases it (more weakly) for progress decisions. One theory rationalizes the sunk cost effect as an optimal heuristic if estimated benefits are forgotten over time. If sunk costs are proxies for those forgotten benefits, then continuing projects based on their sunk costs could be rational 11 . However, the main prediction of this theory is that sunk cost effects should get stronger with longer time passage (presuming forgetting increases with time passage too), which is certainly not the case for utilization decisions.
Some of the most conclusive studies have come from field experiments. A pioneering study gave unexpected discounts to people buying tickets to a series of six plays 2 . The people who got discounts went to fewer of the first three plays (the difference then disappeared for the last three plays). A similar study offered unexpected discounts to diners at an all-you-can-eat pizza restaurant 12 . The diners who paid full price ate one more slice of pizza than those who paid less. In contrast, varying prices charged for a water purification product in Zambia did not show a sunk cost effect (in the form of higher usage by those who paid more) 13 .
Analyses of field data without experimental control have also been consistent with sunk cost effects. Consider NBA basketball players who are chosen higher up in the draft, because they are expected to play better (and are usually paid more). Controlling for actual performance, the higher-drafted players play more minutes even three seasons later 5,14 . Driving a car in Singapore could create sunk cost effects because licenses to drive are extremely expensive 15 . Indeed, one particularly careful analysis showed that drivers who paid more for a license drove more (an effect which grew smaller over time as their license fees were mentally 'amortized'). In penny auctions, players pay a nonrefundable fee to top the bid of a previous "leader". If enough time passes after a topping bid the auction ends. The selling company collects the nonrefundable fees and sells the good at the last price. Analysis of bids indicates that players who have sunk a lot of costs (in the form of bid fees) then bid more aggressively, as if the object's value is increasing in the amount of sunk cost 16 .
There are interesting data on whether animals and children exhibit sunk cost fallacies.
An early study proposed that female digger wasps are sensitive to sunk costs, because when two females fight over a nest to which they have both brought dead katykids (for food), the one who brought more katydids wins more often 17 .
However, the number of katykids are both a sunk cost and have future marginal benefit, so the digger wasp example does not clearly imply a suboptimal tendency to look back at costs that are thoroughly sunk and have no association with marginal benefit, as in ideal experimental designs 18 . However, there is other evidence that animals act as if lever-pressing conditioned stimuli that require more effort to yield reward are more valuable, which is consistent with an effect of sunk cost on valuation 19 .
Finally, there is some intriguing evidence, albeit from tiny samples, that 5-6 year-olds do not exhibit a sunk cost effect in the "lost ticket" (snowstorm) paradigm described above 20,21 . Eight-to-nine and 11-12 year-olds exhibit close to adult behavior, preferring to buy another ticket if they lost their ticket but not if they lost money.
Younger children typically do not make decisions about how to spend scarce resources (such as money, or taking time to cook food). The development of a "no waste" heuristic, as a device to promote careful planning of resource spending, is therefore of little use to younger children. On this presumption, the unusual lack of a sunk cost effect among the 5-6 year olds is consistent with the general "no waste" explanation that drives adult behavior.

Supplementary Results
Additional analyses using the revised sunk cost effect measure Because some participants showed preference reversals in some trials of the control condition, we also analyzed the data using the sunk cost measure taking this into  S2).

Additional analyses concerning activation associated with decision-making under sunk costs
In the current task, the ticket prices were systematically higher in the sunk cost condition compared with the control condition because the ticket price for the non-preferred destination was presented as 50% higher than the ticket price from the preferred one in the sunk cost condition. Therefore, we performed additional analyses Control trials (C): control trials with costs of the options of ¥40,000, ¥50,000, ¥80,000, or ¥100,000 (both options were identical in the control trials).
Control trials (D): control trials with costs of the options of ¥30,000.
The mean costs of the non-preferred options and the sum of both (preferred and non-preferred) tickets in the "sunk cost trials (A)" were ¥70,000 and ¥116,667, respectively. The mean costs of the non-preferred options and the sum of both tickets in the "control trials (C)" were ¥75,000 and ¥150,000, respectively. Therefore, (1) the mean costs of the non-preferred options and (2) the mean costs of the sum of both tickets in the "sunk cost trials (A)" were nearly equal to those of the "control trials (C)" [those in "the sunk cost trials (A)" were even slightly lower than those in the "control trials (C)"].
Then, we analyzed the data using the following GLM, which included

preference phase
The GLM also included motion parameters as regressors of no interest. We compared the brain activation during decision-making in the "sunk cost trials (A)" and the "control trials (C)". We also found that our identified regions during sunk cost decision-making (e.g., insula, IFG and ACC) were robustly activated [the statistical threshold was defined at a cluster-level q < 0.05 after correcting for multiple comparisons using FDR (at voxel-level uncorrected p < 0.001), Supplementary Table   7]. Functional images were corrected for differences in slice-acquisition timing and were then spatially realigned to correct for head motion. The realigned images were spatially normalized to fit the EPI template supplied in SPM8. The functional images were resampled into 2 mm × 2 mm × 2 mm voxels during the normalization process. Finally, all EPI images were smoothed using a Gaussian kernel with a full width at half maximum of 8 mm in the x, y, and z-axes.