University of Amsterdam Faculty of Economics and Business Master thesis Business Economics: Organisation Economics Number of ECTS: 20 Date: 12-06-2015
“Too much invested to quit”
1
or
“too much invested to gamble”?
2
An analysis of the relationship between sunk costs, risk
behaviour and decision making
Author: Kelly Groot
10049916
Supervisor: Prof. T. Buser
1
Teger (1980) 2
Foreword
This master thesis examines the influence of sunk costs on risk behaviour and decision making and is written to complete the Master in Organisation Economics at the University of Amsterdam. The Master specialization focuses on understanding and improving organizational designs and issues. In this thesis, insights from standard economy, behavioural economics and psychology are integrated to analyse the influence of sunk costs, aspiration levels, goals and the irrational behaviour involved. Writing my thesis, I have had some difficulties in pinpointing what I exactly wanted to study and how to optimally perform my experiment. After thorough research and adjustments of the design, I found interesting insights and I can gladly say that I am proud of the result.
At this point, I would like to thank some people for their help. First and foremost, I want to express my gratitude to my supervisor prof. T Buser for his guidance, suggestions, feedback and further involvement. Secondly, I would like to thank all respondents and particularly friends, family and colleagues who helped to spread the word for me. Thirdly, I would like to give special thanks to my sister Jolien Groot who was willing to help me out with some statistical questions all the way from Australia. Gratitude is also expressed towards my 93-year-old grandma who was very enthusiastic about my subject and filled out the survey -using her Ipad- and motivated other family members to do so as well.
Kelly Groot
ABSTRACT
This thesis reports on the relationship between behavioural sunk costs, risk behaviour and financial decision making under uncertainty. Existing literature is inconsistent about whether sunk costs cause risk-seeking behaviour or risk aversion. Those relationships are clarified with an extensive experimental design, manipulating the amount of sunk costs and expected payoffs. The empirical findings are in accordance with the predictions based on prospect theory and suggest that sunk costs do not directly lead to risk seeking or risk averse behaviour. More specifically, sunk costs influence people’s aspiration level, which in turn influences risk behaviour and decisions. The priority seems to be to reach the aspiration level, after which the safest option that enables reaching the aspiration level is chosen, even when this option has a lower expected payoff than other options. This research offers new insights to find solutions for (irrational) decision making when sunk costs are involved.
Table of Contents
1. INTRODUCTION ...5
2. LITERATURE REVIEW...6
2.1 ESCALATION OF COMMITMENT... 7
2.2 DESIGN AND RESULTS OF PAST ESCALATION EXPERIMENTS ... 9
2.3 SUNK COST LITERATURE... 13
2.4 PROSPECT THEORY ... 14
2.4.1 Important features of prospect theory ... 14
2.4.1 Prospect’s value function ... 15
2.5 CONCLUSION... 17
3. METHODOLOGY ... 18
3.1 RESEARCH DESIGN ... 18
3.2 METHODOLOGICAL CHOICES... 21
3.2.1 Recruitment of participants... 21
3.2.1 Behavioural sunk costs... 22
3.2.2 Answer options ... 22 3.2.3 Payment ... 23 3.3.4 Control questions ... 23 3.4 HYPOTHESES ... 24 4. RESULTS ... 26 4.1 DESCRIPTIVE STATISTICS ... 26 4.1.1 Sample ... 26 4.1.2 Summary statistics ... 27 4.2 MAIN RESULTS... 29 4.3 POSSIBLE LIMITATIONS ... 33 5. DISCUSSION ... 35 5.1 FINDINGS ... 35 5.2 THEORETICAL CONTRIBUTIONS ... 37 5.3 MANAGERIAL IMPLICATIONS... 38
5.4 RECOMMENDATIONS FOR FUTURE RESEARCH ... 39
6. REFERENCES ... 41
1. Introduction
Imagine that, as a product developer, you invested a large amount of time and money on developing a new kind of sports car. One year later, the project is completed for 90% but unfortunately, given the current market circumstances; the chances for success are low. What would you do: keep investing (last 10%) to launch the new sports car or quit the project?
Similar information is often provided to participants in escalation experiments (Teger, 1980). Many choose to continue investing because it feels like a waste of resources otherwise. A commonly used argument is: “If I stop right now, I will never be able to make up for the investments” (McCarthy et al., 1993; Moon, 2001). However, this kind of reasoning is misleading since “prior investment in money, effort, or time” named sunk costs cannot actually be recovered once spent (Arkes & Blumer, 1985, p. 124). When decisions are based on those fallacies, it can have bad consequences. For example, when people irrationally base their decision to continue on the amount of sunk costs and thus ignore future negative expected utility ((McCarthy et al., 1993; Moon, 2001), it is likely that the project fails eventually, after spending a lot of money. The right thing to do would have been to quit investing earlier. This rational conclusion would have been drawn when only future benefits and costs were measured and sunk costs were ignored. In short, there may be both good and bad reasons to continue a project but the decision should not be based on prior investments.
Unless the mentioned consequences, it is an established fact that prior investments are taken into account when making decisions (Arkes & Blumer, 1985; Thaler & Johnson, 1990). Much research has been done to identify causes, discover influencing variables and to find solutions for this irrational behaviour. Results are not consistent; there are conflicting ideas in the literature about the underlying framework of this behaviour. On the one hand, there is quite some literature that illustrates that taking sunk costs into account will lead to risk seeking behaviour (Kahneman & Tversky, 1979; Thaler, 1980; Teger, 1980; Arkes & Blumer, 1985). For example, in the situation outlined above, the chances for success are low but many product developers still keep investing so take a high risk (McCarthy et al., 1993). It seems like they have too much invested to quit (Teger, 1980). The phenomenon of taking more risk after incurring sunk cost is called the ‘sunk cost effect’. In contrast, some researchers state that sunk costs will actually lead to risk aversion (Schaubroeck & Davis, 1994; Zeelenberg & van Dijk, 1997). In experiments where multiple options
are provided, participants choose the safest option that can make up for the incurred sunk costs and thus enables reaching the aspiration level3. They apparently have too
much invested to gamble and this leads to risk averse behaviour (Zeelenberg and van Dijk, 1997).
The ambiguity in literature of risk behaviour after incurring sunk costs in combination with the consequences of taking sunk costs into account, make it interesting to investigate the exact relationship between sunk costs, risk behaviour and decisions. Based on the above, the research question is formulated as follows:
How are risk-taking behaviour and financial gain decisions (under uncertainty) affected by behavioural sunk costs? .
This study contributes to this area of research by exploring and clarifying the exact relationships between sunk costs and decisions with the help of an improved research design. The obtained insights have a high practical relevance since they provide guidelines to make more rational and effective decisions. For this research, prospect theory’s model is used to clarify and predict behaviour. Answers are found with the help of an extensive experiment in which the amount of sunk costs and answer options are manipulated. This way, the effect of investing time and effort on decisions concerning, sometimes uncertain, positive payments is analysed. An adjusted framework is proposed based on the results of the experiment, which show that decisions are indeed influenced by sunk costs, especially by the desire to reach the aspiration levels instead of by an actual change in risk behaviour.
The thesis is structured as follows: the next section provides a review of related literature, starting with a general review of escalation and sunk cost literature, after which prospect theory’s features and predictions are outlined. Section 3 proposes the experimental method and discusses adjustments and improvements compared to preceding experiments. Results are given in section 4 and the subsequent sections discuss results and draw conclusions. References and instructions for the experiment can be found in the last two chapters.
3 Aspiration level: “the level of future performance in a task that an individual explicitly undertakes to
2. Literature review
In this section, the existing literature concerning sunk costs and its influence on decisions and risk taking will be addressed in order to give an overview and to clearly define the research topic. First, escalation and sunk cost literature are discussed and existing reasons and solutions for the problem are outlined. Next, strong points and major shortcomings of the typical escalation or withdrawal experiments are highlighted and alternative methods and experiments are reflected on. Subsequently, information on prospect theory is provided, to finish with a short conclusion that emphasizes the need for this research.
2.1 Escalation of commitment
The example in the introduction illustrates that (in experiments) decision makers often decide to stay with their initial choice and keep investing, even when negative feedback is received and the expected utility of alternative options seems higher. This behaviour is referred to as escalation of commitment or the sunk cost effect. This phenomenon is observed for monetary as well as nonmonetary sunk costs (Arkes & Blumer, 1985). An example with nonmonetary sunk costs is that people sometimes stay with their job even though they find it disastrous, simply because they did so much effort and spent years on training to get the position and do not want to ‘waste’ that investment. Another illustration from real life is that when tickets for a theatre show are already bought (a sunk cost is incurred), people are more likely to drive through a snow storm (take more risk) than when they had not incurred the sunk cost. They would have been risk averse when no sunk costs were incurred: namely stayed home (Thaler, 1980). This shows again that the incurred sunk costs influence the decision: paying for the right to use a service seems to increase the user rate. Rationally, the decision to go or not should have been based on future costs and benefits instead of sunk costs.
Nonetheless, there is extensive escalation literature that has shown that prior events and investments have an influence on future actions and decisions (Arkes & Blumer, 1985; Cheng, Schulz, Luckett, & Booth, 2003; Chow, Harrison, Lindquist, & Wu, 1997; Garland, 1990; Kanodia, Bushman, & Dickhaut, 1989; Staw, 1976; Thaler, 1980). On the one hand, this behaviour makes sense since past events and investments often contain information and can therefore be useful to take into account. However,
the fact in itself that investments have been done should not influence any decision. Sunk costs are already spent (in time or money) and whatever your future decisions are, there is no way to get the investment back. Sunk costs are therefore irrelevant and it would be misleading to consider them. Instead, factors that should influence decisions are expectations about future performance and future incremental costs and benefits (Keys & Schwartz, 2007; McCarthy, Schoorman, & Cooper, 1993). Those variables should thus be considered regardless of prior investments. Still, there are many examples where sunk costs are taken into account when making decisions. Thus far, research has achieved to define reasons, moderating factors and solutions for this irrational behaviour. The fundamental cause seems to be that people dislike the prospect that their investments have been in vain (Arkes & Blumer, 1985; Frisch, 1993; Garland & Newport, 1991; Tan & Yates, 1995; Zeelenberg & van Dijk, 1997). People seem to reason that investments are lost when they do not continue to invest and therefore want to avoid such waste of money. This first cause converges with another dominant reason for escalation behaviour, namely justification of behaviour that appears to have been in error (Schulz & Cheng, 2002; Staw, 1976). The decision maker feels the need to justify the initial investment choice either to him or herself or to the outside world and does this by reinvesting in the same project. The need for justification may thus be strong enough to make people ignore feedback indicating negative perspectives for the chosen project (Arkes & Blumer, 1985). Apparently, the small chance of ‘making up’ for their investments weighs stronger than the relatively high risk of further losses through continuation. It could also be that decision makers focus more on positive outcomes and underestimate the probability of a loss (Brockhaus, 1980; Vlek & Stallen, 1980).
Some researchers argue that people become more risk seeking once they have invested time or money and that this results in continued investments. (Thaler, 1980; Whyte, 1986; Garland, 1990). The underlying framework of this behaviour will be discussed more extensively in paragraph 2.4. Other psychological explanations include mental accounting (Thaler, 1980) and framing (Laughhunn & Payne, 1980). Note for example that investments are perceived as smaller when they take place in the context of larger absolute investments (Garland, 1990; Garland & Newport, 1991). Also, participants assess the likelihood of success higher when they have invested time or money into something. It is not yet clear whether this inflated estimate stimulates continued investment, is the consequence of the decision to
continue or a combination of both (Arkes & Blumer, 1985). Finally, some theories state that the sunk cost effect is a result of hiding private information and reputational concerns (Kanodia et al., 1989).
Positively moderating variables are level of responsibility for the project, competition, publicity of the choices and a culture of consistency (Staw & Ross, 1987). Manipulating those factors downwards, that is less responsibility, competition and publicity and less cultural norms for persistence, seem to decrease escalation. The presence of alternatives and clear information regarding the probabilities of future outcomes may also mitigate the problem (McCain, 1986).
2.2 Design and results of past escalation experiments
Most experiments on this topic used a certain design. This particular design and the most important results will be outlined followed by a discussion of the strong points as well as shortcomings. More specific designs concerning risk-taking behaviour will be discussed subsequently.
Most escalation experiments have a similar design as the sports car experiment in the introduction. Participants get the role of president of a company or for example research and development manager (Arkes & Blumer, 1985; Garlands & Conlon, 1998; Tan & Yates, 1995; van Dijk & Zeelenberg, 2003). Then, they are told that they invested a certain amount in a research project. In most research, this investment is about 10 million dollars and the project is to develop for example a radar blank plane (as in Arkes’ and Blumer’s (1985) experiment). They are informed that the product is completed for 90% and also that another enterprise launches a similar product with a certain probability of being superior to their product. The probability is not always explicitly stated; in some cases subjects are asked to give their perceived chance of financial success (on a 0-100 scale). Subsequently, participants need to decide whether to continue investments for the last 10% or to terminate the project. In other words: participants escalate or withdraw.
In those scenarios, sunk costs and responsibility for the prior investment are manipulated. The major finding in all of those experiments is that participants who invested some resources were more willing to continue the project than participants who had not incurred sunk costs. Moreover, the higher the incurred costs, the higher the willingness to invest (Arkes & Blumer, 1985; Garland, 1990; Keil, Truex, & Mixon, 1995). Therefore, risky investments (given the small chance to succeed) with
the possibility to make up for the loss seem to be more attractive than choosing for a sure loss (terminating the project). From those outcomes is inferred that people become risk seeking when sunk costs are involved. Referring to the title of this thesis, the phrase “too much invested to quit” (Teger, 1980) thus applies to the escalate and withdrawal literature. Figure 1 illustrates the cause and effect relationship found in those experiments.
This research forms the foundation of the sunk cost literature and helps to determine some major behavioural tendencies. The main conclusion is that most research shows that people are risk seeking after investing time or money into a project, or when they suffered another kind of loss and are in the loss domain (Hershey & Schoemaker, 1980; Kahneman & Tversky, 1979; Payne, Laughhunn, & Crum, 1980; Slovic, Fischhoff, & Lichtenstein, 1982).
Figure 1
Relationship between Sunk Costs, Risk Behaviour and Decisions in Escalate or Withdraw Experiments
Apart from the relevant contribution those experiments deliver, there are some major points of criticism for this method. First of all, the experiment offers participants very limited options: they have to make a decision between continuing investment, which is risky and offers a small chance of a satisfying result or terminating the project, which is a safe option but never satisfying. In real life, there are mostly more choice opportunities. For example, one could invest the money in an alternative project and
Sunk costs increase •Options: continue or quit People become more risk seeking
•Small possibility to make up for the loss (continuing) is more attractive than choosing for a sure loss (terminating)
Decision is influenced
•People prefer the riskier option (continue)
so opportunity costs should be taken into account (Schaubroek & Davis, 1994). This does not happen in the previous experiments since there is no other choice than to reinvest or withdraw. It could be that this design moves participants in the direction of continuing since they are certain of a loss otherwise. Some researchers therefore argue that this design leads to biased results (Schaubroeck & Davis, 1994; Zeelenberg & Van Dijk, 1997).
The second point of criticism is that the studies do not give clear and relevant information about marginal costs and expected benefits (Heath, 1995). When the costs and benefits of continuing are not explicitly clear, it becomes particularly difficult to conclude whether participants acted rationally. For example, in the scenarios in which responsibility for the project is manipulated, it shows that high responsibility subjects show escalation behaviour. Heath (1995) emphasizes the expectation of extra costs responsible subjects could have compared to non-responsible subjects: in real life, they are likely to get punished for the poor performance and could experience reputational and social penalties. When people take this into account, it would indicate that the future costs of discontinuing a project are larger for a subject with full responsibility. Even though additional costs for responsible decision makers are not explicitly mentioned, it could be that people take them into account (unconsciously). The height of those costs are difficult to track but this example does show that it can be a rational choice for the individual to continue investment, even though it is a poor decision for the organization.
Thus far, the literature showed risk seeking behaviour as a result of sunk costs. Remember that the discussed experiments offer limited options, namely risky reinvestment or riskless withdrawal. Incorporating alternatives in the answer options would eliminate the certain loss that withdrawal entails. Schaubroek and Davis (1994) use the same design but extend it and offer participants three options instead of two: withdrawal, reinvestment (risky) and investment in an alternative project (risky). The riskiness of the last two projects is manipulated; higher risk is compensated with higher possible return. Note that success with both of the last two options clearly makes up for the sunk cost. The results show that participants do not structurally escalate anymore. Instead, they prefer the option that is more conservative. The authors conclude that “the presence of alternatives may change the risk orientation of the decision maker from risk prone to risk averse.” (Schaubroeck & Davis, 1994, p. 78).
Zeelenberg and van Dijk (1997) go one step further by including a riskless option that enables ‘recovering’4 the sunk costs. Furthermore, the design is different
from the experiments described above. Participants have to make a choice between options that differ in payoff and risk. Participants are informed that they did a dirty job for someone and had to work hard. Subsequently, they have to make a decision between two options: a safe option to receive 50 sure guilders and a gamble resulting in a gain of 100 guilders with 50% chance and 0 guilders with 50% chance. The influence of behavioural sunk costs, namely effort and time, is analysed. They find that participants in the behavioural sunk cost conditions are more likely to choose the safe option than the other participants and thus that sunk costs make people more risk averse. Their results demonstrate that this is the case when people have the opportunity to choose for an option that enables reaching the aspiration level. The relationships are depicted in figure 2.
Figure 2
Relationship between Sunk Costs, Risk Behaviour and Decisions found in Zeelenberg and van Dijk’s (1997) and Schaubroeck and Davis’ (1994) Research
4 Sunk costs can not actually be recovered but people do perceive it that way. The word recover is still
used in this thesis but always in quotation marks.
Sunk costs increase
•Options: safe 50 gain and risky 100 gain Aspiration level increases People become more risk averse
•Reaching the aspiration level is satisfactory and larger gains will only result in small increases in value Decision is influenced •People prefer the safest option
A point of concern with the research of Schaubroeck and Davis (1994) as well as with Zeelenberg and van Dijk (1997) is that the researchers do not test whether individuals are more risk seeking when the more conservative or safe alternative is not offered. Also, options that never reach the aspiration level are not included in any of the conditions. Theory would predict that options that are not able to make up for sunk costs are foregone but they do not experimentally test this. To clarify the exact relationship and to make the research as complete as possible, both the effect of offering only a risky alternative that reaches the break-even point as the effect of offering safe alternatives are tested in this thesis.
Furthermore, the investment and expected return are explicitly and clearly stated in Schaubroeck and Davis’s (1994) experiment while this is not the case in the research of Zeelenberg and van Dijk (1997). Participants are told that they did a dirty job for someone but no explicit duration, effort level or opportunity costs are mentioned. This is critical since it could lead to different aspiration levels for every participant. Zeelenberg and van Dijk (1997) seem to assume that 50 guilders is a plausible aspiration level but they are unable to check this. The unclear information about sunk costs could, as discussed, lead to a bias and make it difficult to draw robust conclusions (Heath, 1995).
Summarizing, there are contrasting findings concerning risk behaviour after incurring sunk costs. It seems that high risks are taken when withdrawal is the only alternative to reinvestment but risk aversion is preferred when there are safer possibilities to recover. The framework that puts the findings in perspective and can offer an explanation for decisions under uncertainty is prospect theory. This theory will be described carefully in paragraph 2.4.
2.3 Sunk cost literature
The terms ‘sunk cost effect’, ‘sunk cost fallacy’, ‘escalation behaviour’, ‘escalation error’ and ‘escalation of commitment’ are used interchangeably in the literature (Kanodia, et al., 1989). The last three terms describe the same thing but a difference exists between those escalation terms and the sunk cost effect/fallacy. It is necessary to consider this distinction since it will clarify the direction of this research. Escalation refers to the continuation of investments, especially when prospects are negative. Escalation literature tries to identify reasons and solutions for this phenomenon but do not manipulate sunk costs. The sunk cost effect specifically
involves the effect of sunk costs on decision making and is thus a less broad term than escalation (Duxbury, 2012). The current paper specifically focuses on the influence of sunk costs on choice (under uncertainty) and manipulates sunk costs as an independent variable and therefore belongs to the sunk cost literature.
2.4 Prospect theory
The first thing one can conclude from the previous findings is that the behaviour observed in experiments as well as in real life is inconsistent with standard economic expected utility theory. Kahneman and Tversky (1979) propose an alternative descriptive model for choice under uncertainty, namely prospect theory. Features of prospect theory are also used to describe and explain the behaviour that is observed after incurring sunk costs. Prospect theory can help to clarify the contradicting results in risk behaviour found in the literature. The most important characteristics of prospect theory that play a role in sunk cost literature will now be outlined.
2.4.1 Important features of prospect theory
First of all, prospect theory involves that gains are treated differently than losses (Kahneman and Tversky, 1979). For example, when people are given a choice between 4000 with 80% chance and 0 with 20% or 3000 for sure, the majority will choose the second option. In contrast, when a similar choice option is given with losses (-4000, 0.80; -3000, 1.0) most prefer the first, and thus riskier option. The preference for negative prospects are thus opposite of the preferences for positive prospects. This mirror effect is called the reflection effect.
Another feature of prospect theory deals with the way people perceive probabilities. Outcomes seem not to be exactly weighted by their probabilities, as expected utility theory describes but people sometimes over- or undervalue probabilities (Kahneman & Tversky, 1979). First of all, people overvalue certain gains while they undervalue certain losses. For gains this means that people value a certain gain too high relative to what would be expected for a probable gain. In contrast, certain losses are valued too low compared to losses with a probability below 1.0. This is referred to as the certainty effect (Kahneman & Tversky, 1979). This could explain why many people favour investing which would lead to a very small possibility of success over withdrawing which would result in a certain loss, as observed in escalation situations.
A third aspect of prospect theory concerns that the shape of the value function is convex for losses and concave for gains. This is the case because of the psychological phenomenon that the difference between 0 and 100 appears to be greater than the difference between 1000 and 1100. This psychological bias explains the risk-seeking tendency for losses and preference for risk aversion for gains. A last characteristic of the value function is that it exhibits loss aversion. The value function is steeper for losses than for gains; meaning that losing e.g. money hurts more than that gaining the same amount brings joy (Kahneman and Tversky, 1979).
2.4.1 Prospect’s value function
The value function, which is an adjusted version of the utility function, forms the foundation of prospect theory. The function describes the relationship between gains and losses and the subjective value a person attaches to each of those (Arkes & Blumer, 1985). An illustration of the value function is shown in figure 3 (Zeelenberg & van Dijk, 1997). One essential feature of prospect theory is that decisions are evaluated in relation to a reference point. Figure 3 shows three different reference points, which will all be considered now.
Figure 3
Reference point A: This is usually the status quo (point A in figure 3). From this point every negative outcome (everything below the reference point) is seen as a loss and every outcome above zero (higher than the reference point) will be considered as a gain.
Reference point B: After an investment has been made, prospects are not been seen from the status quo (point A in figure 3) but from the loss side, which would be point B in figure 3. This is the case because sunk costs are not neglected and seem to create a loss frame (Arkes & Blumer, 1985). Imagine that someone has invested €3000 in a project and now has to choose between a sure €1500 or a 50% chance to receive €3000 and 50% chance to get nothing. This should be approached as a choice between (€1500) and (€3000, .5) but when the decision maker takes the sunk cost into account, he/she will perceive the problem as a choice between (- €1500) and (- €3000, .5). This means that even positive numbers (for example the sure €1500) are not perceived as gains when they do not make up for the sunk cost (of in this case €3000). Only outcomes that make up for the sunk cost are seen as a gain.
Notice that from point B gains cause large increases in value while losses cause relatively small decreases. This leads to the situation that people are more willing to take the risk of a small loss in order to obtain possible large gains instead of taking the sure loss (Zeelenberg & van Dijk, 1997). On the basis of prospect theory, this is the explanation of the observed risk-seeking behaviour after incurring sunk costs.
Reference point C: According to Zeelenberg and van Dijk (1997), it is of great importance to add a third reference point when applying prospect theory to sunk cost situations. When people expect a certain outcome, such as a salary or a return on their investment, the decision is not evaluated in relation to their current assets but to the outcome they expect. This alternated reference point is called the aspiration level, shown as point C in figure 3 (Zeelenberg & van Dijk, 1997). Kahneman and Tversky (1979) also describe this special reference point but do not use it in their experiments. In their experiments, only the status quo is used as reference point.
When sunk costs have been made, research shows that outcomes are also evaluated in relation to this aspiration level, meaning that everything above this point is seen as satisfactory and everything below is not. The aspiration level can be on the gain as well as on the loss side of the value function. The aspiration level is on the
loss side of the value function when one expects a withdrawal of for example €1000. When only €300 is withdrawn, it is seen as a gain since it is higher than the aspiration level. The aspiration level is on the gain side when a positive outcome is expected: this situation is illustrated in point C. In this research, the aspiration level will be on the gain side so the situation is similar to the one in point C.
Recall the example of someone who invested €3000 in a project and had to choose between a sure €1500 or a 50% chance to receive €3000 and 50% chance to get nothing. Note that in this example, only the risky option enables to make up for the prior loss. When one implements a safe option that makes it possible to reach the aspiration level, the decision frame will be changed. Imagine the choice options are a sure €3000 and a 50% chance to receive €6000 and 50% chance to get nothing. This choice will probably be perceived as a choice between (€0) and (- €3000, .5; €3000, .5). In this case, the aspiration level of €3000 is used as the reference point and so point C in figure 3 would be at the origin of the graph, implying that point A is on the loss side. The sure €3000 offers participants a safe option to reach the aspiration level. Reaching the aspiration level is satisfactory and larger gains will only result in small increases in value. The status quo (point A) is regarded as a loss and since losses cause larger decreases than gains cause increases, the safe option will be more attractive than the risky one. Even though the risky one can result in higher gains, the safe option is the only one that always ensures a satisfying result. Zeelenberg and van Dijk (1997) argue therefore that it is not the full recovery of sunk costs that make people risk averse but the increased aspiration level induced by sunk costs.
Altogether, prospect theory offers important explanations for decision making under risk. Three essential characteristics of prospect theory are that the value is defined by changes from the status quo rather than from the final asset position, the function is concave for gains and convex for losses and finally, the loss function is steeper than the gain function. Those characteristics can all be inferred from the value function.
2.5 Conclusion
Previous experiments in the area of sunk costs show contrasting results. Especially, the literature is inconsistent about whether sunk costs cause seeking or risk-averse behaviour. Also, it is clear that there are difficulties with proposing a robust research method. With the help of prospect theory and an extended research design,
this research will investigate the exact relationships between sunk costs, risk-seeking behaviour and decision making.
3. Methodology
After thoroughly discussing the literature on sunk costs and decision making under uncertainty, the research design and method will be explicated in this section. First the research design is clarified. Subsequently, information on the sample is given and methodological choices are justified. Finally, the hypotheses are outlined.
3.1 Research design
As mentioned, the literature is inconsistent about whether sunk costs cause risk-seeking behaviour or risk aversion. In order to clarify the relationships, a research design is developed based on the experiment by Zeelenberg and van Dijk (1997), which has been discussed in paragraph 2.2. Some major changes and extensions are made in order to improve the design.
Two major shortcomings in the experiment of Zeelenberg and van Dijk (1997) are that the design is incomplete and furthermore unclear regarding the amount of sunk costs. To solve the first problem, conditions with different sunk costs are included. This enables testing both the effect of offering a safe alternative that always reaches the aspiration level as well as the effect of the absence of a safe alternative to reach the aspiration level (in other words: offering only a risky alternative that has a chance to reach the aspiration level). Note that Zeelenberg and van Dijk (1997) do not have a condition in which the outcome of the safe option does not reach the aspiration level and so were only able to test the first situation.
Secondly, the amount of sunk costs and expected returns are made clear in the adjusted design. The experiment involves behavioural sunk costs; the value of effort and time is likely to differ between people and is therefore difficult to translate to one specific value. Still, this is done by informing participants about the duration of the task and about the costs they could otherwise have earned in their regular job (opportunity costs). This way, participants are expected to use the intended aspiration level as the reference point.
After informing the participants about the work they did, they can choose between three options: a safe, a moderately risky and a risky one. Since the amount of
sunk costs will be manipulated, it differs for each condition which of the options can reach the aspiration level. In table 1, the chance to reach the aspiration level for each group is outlined in percentages. Table 2 denotes a simplified version of the first table. Zeelenberg and van Dijk (1997) did not manipulate the sunk costs and only offered a safe and risky option. With the additions in the current extensive method, it becomes possible to test behavioural tendencies in more situations.
Table 1
Chance (in %) that Aspiration Level can be Achieved for each Answer Option and per Condition
Answer option Control group (sc = 0) Treatment 1 (sc = 60) Treatment 2 (sc = 80) Treatment 3 + 5 (sc = 100) Treatment 4 (sc = 125) Option A 100% 100% 0% 0% 0% Option B 100% 75% 75% 0% 0% Option C 100% 50% 50% 50% 0%
The experiment has seven different scenarios consisting of one control group, five variations in the level of sunk costs and one additional scenario testing the effect of nonequivalent expected payoffs. The different conditions will be explained in more detail below.
Sunk Cost Absent vs. Sunk Cost Present
The amount of sunk costs participants have incurred varies among the different conditions. Zeelenberg and van Dijk (1997) have a Sunk Cost Absent (€0) and a Sunk Table 2
Simplified Chance that Aspiration Level can be Achieved for each Answer Option and per Condition
Answer option Control group (sc = 0) Treatment 1 (sc = 50) Treatment 2 (sc = 70) Treatment 3+5 (sc = 100) Treatment 4 (sc = 120)
Option A Always Always Never Never Never
Option B Always Sometimes Sometimes Never Never
Cost Present (€50) condition. In this thesis, there is a Sunk Cost Absent (€0), Sunk Cost present (€60), Sunk Cost Present (€80), Sunk Cost Present (€100) and Sunk Cost Present (€125) condition. The first two groups are similar as in Zeelenberg and van Dijk’s (1997) experiment but the last three form an extension. The addition of the Sunk Cost Present (€80) group is important since it tests which option will be favoured when the safe option does not reach the aspiration level but the moderately risky (75% chance) and risky option (50% chance) sometimes do. The Sunk Cost Present (€100) condition enables investigating behaviour when the aspiration level can only be reached by choosing the risky option.
The additional Sunk Cost Present (€125) group makes it possible to research what happens when none of the options offer an opportunity to reach the aspiration level. This has not been investigated before and is therefore added to this research.
As an example, one of the scenarios is given below. This is the Sunk Cost Present (€60) scenario. An overview of all conditions and the corresponding scenario instructions can be found in the appendix. Note that the answer options are exactly the same in the different conditions. Of course, control questions (age, gender and study background) are added to enable controlling for potential biases afterwards. The control questions are similar for every group so they are only depicted in the control group instructions in the appendix.
Imagine the following situation:
You just finished a job for someone, it took you six hours and it was hard work. You have not agreed on a specific payment on forehand so you are not sure what sum you will get. To do this job, you had to take time off from work, where you would otherwise earn €10 each hour so in total you will receive €60 euros less from your regular boss.
The person you did the job for, comes to you and offers you the following three options to receive your money. Which option do you choose?
a) Receive €60 for sure (chance = 100%).
b) Have a 75% chance to receive €80 and 25% chance to receive €0. Clarification: you can grab one marble from a basket with 4
marbles: 3 red marbles and 1 white. When you grab a red marble, you get €80. When you grab a white one, you get nothing.
c) Have a 50% chance to receive €120 and 50% chance to receive €0. Clarification: you can grab one marble from a basket with 4 marbles: 2 red marbles and 2 white. When you grab a red marble, you get €120. When you grab a white one, you get nothing.
Equivalent expected payoff vs. nonequivalent expected payoff
The aforementioned answer options all have an equal expected payoff of €60. This is the case for the control group, treatment 1, 2, 3 and 5. Those conditions are referred to as the equivalent expected payoff conditions. This is similar as in Zeelenberg and van Dijk’s experiment although their expected payoff is €50 in both choices.
This approach is taken one step further in treatment 4, in which the answer options are adjusted and the expected payoff is not the same for every option. This is the nonequivalent expected payoff condition. Participants in this condition can choose between a safe option with an expected payoff of €60, a moderately risky option with an expected payoff of €60 and a risky option with an expected payoff of €50. In this case, the safe and moderately risky options clearly have a higher expected payoff and rational decision makers should therefore pick one of those options. However, the risky option is the only one that offers an opportunity to ‘recover’ sunk costs. This condition tests if the desire to reach the aspiration level is strong enough to result in the irrational behaviour of choosing the risky option. This will only be tested for the Sunk Cost present (€100) condition since it is this group for which the risky option is the only one that makes it possible to reach the aspiration level.
3.2 Methodological choices
In order to make the research as reliable as possible, certain methodological choices are made. These are discussed and justified in this section.
3.2.1 Recruitment of participants
This ‘experiment’ is carried out in the form of a survey since that enables achieving a high response rate in a relatively short time frame. A high response rate is very important for robustness reasons, especially given the amount of different conditions. There is no focus on a specific target group since the sunk cost effect is a
psychological phenomenon which is expected to be uniformly present. Therefore, every response is useful. The survey is available online so participants will be recruited via social media (especially fellow students, friends and family) and through online forums. A link to the Qualtrics survey is distributed and participants are randomly redirected to one of the six conditions. Paper versions are also available in order to recruit participants in local municipalities. Municipalities are chosen since everyone needs to go there from time to time. Therefore, a variety of people from different ethnic groups will be recruited and there is no significant bias in the type of visitors compared to the total population of the Netherlands. On top of this, multiple local municipalities are chosen (Amsterdam Center, Amsterdam West and De Goorn) to reduce the bias from residence. Twenty paper surveys of each condition are printed and randomly jumbled up. The surveys are distributed, starting at the top of the stack, to all people waiting in the municipality. Everyone who was able to read Dutch, accepted to fill out the entire survey. The final sample consists of 313 participants, around 50 for each condition.
3.2.1 Behavioural sunk costs
The experiment involves behavioural sunk costs, meaning effort and time instead of monetary sunk costs. Most previous research focused on the influence of financial investments. In daily life, many of our decisions are influenced by behavioural sunk costs as well. Therefore, they are investigated in this research. A problem with behavioural costs could be that they are not really accurate since it is more difficult to measure people’s value for time and effort than for money. This potential problem is solved by mentioning monetary opportunity costs and in this way inducing an aspiration level.
3.2.2 Answer options
This design offers participants three answer options, which is one more than in Zeelenberg and van Dijk’s (1997) experiment. Of course, this is still a simplification of reality since it is likely that there are more than three options for a decision in real life. However, since a safe, moderately risky and risky option are offered, it should be enough to recognize certain tendencies and patterns of behaviour without pushing participants in a certain direction, a problem observed in standard escalation literature.
3.2.3 Payment
There are many reasons to argue that paying people according to the decisions they make is important. Just telling people that they invested time and effort and subsequently asking what decision they would make, measures behavioural intentions instead of real behaviour. Unfortunately, problems arise when payment is introduced as well. When one wants to base the payment on the situation and decisions, a choice should be made whether to incur the sunk costs/opportunity costs or not. For example one could state: 𝑃𝑎𝑦𝑚𝑒𝑛𝑡 =100+ 𝑂𝑢𝑡𝑐𝑜𝑚𝑒−𝑜𝑝𝑝𝑜𝑟𝑡𝑢𝑛𝑖𝑡𝑦 𝑐𝑜𝑠𝑡𝑠
50 .
If a participant in treatment 1 (Sunk Cost Present (€60), equivalent expected payoff) chooses option B, there is 75% chance that the payoff will be 100 + 80−60
30 = €4 and
25% chance that the payoff will be 100 + 0−60
30 = €1,33. In this case, the sunk costs are
reflected in the payment function so they will probably be explicitly taken into account. This would push participants in a direction and hinders measuring actual behaviour. Instead, one could also use 100 + 𝑂𝑢𝑡𝑐𝑜𝑚𝑒
50 as payoff function but in this
case, sunk costs will probably be ignored since they are not reflected in the payoff. Both of the structures will push participants in a certain direction and possibly make them ignore some of the other information. This means that for this design, paying participants according to their decisions could do more harm than good and therefore participants will not be paid. Also, note that previous studies, including Kahneman and Tversky (1979) use hypothetical decision scenarios as well. They find significant results consistent with prospect theory so there is not yet a reason to question whether behaviour could be manipulated by hypothetical sunk costs.
3.3.4 Control questions
Control questions are added to the experiment to investigate certain relationships and correlations and to control for potential biases. For example, it seems important to control for gender effects since multiple studies have found that women are more risk averse than males (Byrnes, Miller and Schafer, 1999; Jianakoplos and Bernasek, 1998). Also, age and education background could influence the results.
One could argue that it is also important to control for risk preferences: some people are more risk averse, risk neutral or risk loving than others. However, there is no need to add a control question that investigates risk preferences since participants
are randomly distributed over the conditions and the control group will already illustrate which choice people tend to choose most often. This group does not involve sunk costs so the choice will be based on the general risk preference. This result will be used as a baseline and will be compared to the other conditions. When preferences for answer options significantly differ for the other groups, an effect of the sunk costs is already demonstrated.
3.4 Hypotheses
This research investigates prospect theory’s predictions about individuals’ risk preferences so the hypotheses will be formed based on prospect theory.
Equivalent expected payoff
For the equivalent expected payoff conditions, economic theory would predict no difference between the groups’ choice preference. Prior outcomes and investments are irrelevant for future decisions and therefore the decision will not be affected by the presence of sunk costs (Keys, & Schwartz, 2007). If prospect theory is at the base of the sunk cost effect, differences are expected.
Hypothesis 1: The presence of sunk costs has an influence on risk-taking and subsequently on decisions.
According to prospect theory, sure losses are undervalued and particularly aversive so it is not likely that participants choose an option that will never reach the aspiration level (Kahneman & Tversky, 1979). They would rather choose options that offer at least a chance to reach the aspiration level.
Hypothesis 2: When sunk costs are involved, people will be more risk seeking when there is no safe way to ‘recover’ sunk costs.
o Individuals in the Sunk Cost Present (€80) condition are more likely to choose the moderately risky option than individuals in the Sunk Cost Absent and Sunk Cost Present (€60) condition.
o Individuals in the Sunk Cost Present (€100/€125) condition are more likely to choose the risky option than individuals in the Sunk Cost Absent and Sunk Cost Present (€60/€80) condition
When an option satisfies the aspiration level, it is satisfactory. Larger gains will only result in small increases in value; losses will cause respectively large decreases. Even though the riskier choice can result in higher gains, the safe(r) options have a higher chance (100% / 75%) of a satisfying result. For this reason, the safest option will be most attractive.
Hypothesis 3: When sunk costs are involved, people will be more risk averse when there is a safe(r) way to ‘recover’ sunk costs.
o Individuals in the Sunk Cost Present (€60) condition are more likely to choose the safe option than individuals in the Sunk Cost Present (€80/€100/€125) condition.
o Individuals in the Sunk Cost Present (€80) condition are more likely to choose the moderately risky option than individuals in the Sunk Cost Present (/€100/€125) condition.
When none of the options satisfy the aspiration level, all of the answers will be unsatisfying. This is expected to change people’s mind-set: they are not motivated to take higher risks since even risk taking will not be rewarded as the outcome will never be satisfying. It could be that people are in this case willing to ignore the sunk costs and choose the answer they would have chosen when there were no sunk costs at all. Prospect theory states that for gains, certain options are overvalued and people are more likely to choose a lower but certain amount than a higher but uncertain option. Therefore, when there are no sunk costs or when sunk costs are ignored, people are expected to be risk averse.
Hypothesis 4: Individuals will ignore sunk costs when there is no way to ‘recover’ them.
o The behaviour in the Sunk Cost Present (€125) condition is similar to the behaviour in the control group: the safe option will be chosen.
Nonequivalent expected payoff
In the nonequivalent expected payoff condition, the safe and moderately risky option have a higher expected payoff than the risky one. However, the risky option is the only one that offers an opportunity to ‘recover’ sunk costs. This would mean that the risky option is the only one that could lead to a satisfactory outcome. Since this is the
case, participants are expected to choose the risky option, even though the expected payoff is lower.
Hypothesis 5: When sunk costs are involved, people will be more risk seeking when there is no safe way to ‘recover’ sunk costs, even when the expected payoff is lower than for safe options.
o Individuals in the Unequal expected payoff, Sunk Cost Present (€100) condition are most likely to choose the risky option.
4. Results
The previous section illustrates the methodology used for this research. In this section, the results are provided. First of all, some descriptive statistics will be given. Subsequently, chi-square tests of independence and Fischer’s exact tests are performed to explore relationships between the conditions and the other variables. Furthermore, an analysis is done for specific subgroups and possible limitations are pointed out.
4.1 Descriptive statistics 4.1.1 Sample
The total number of participants is 332. From those, 19 participants did not complete all questions so are deleted from the sample, which results in a final sample of 313 participants. 192 Of the participants were recruited online, the remaining 121 participants were recruited in local government offices. The participants for this study consist of 161 men (54.44%) and 152 women (48.56%). The majority of the participants is aged between 20 and 40 years (60.38%) and only 0.64% are aged between 80 and 100. Participants between 0 and 20 make up 14.38% of the sample, 20.77% are aged between 40 and 60 and few between 60 and 80 (3.83%). As for education, most participants attended MBO5 (28.43%), HBO6 (34.19%) or WO7
(24.60%). Some participants’ level is primary school (1.28%), VMBO8 (3.19%) or
5 Intermediate Vocational Education 6 University of Applied Science 7
University of Science 8
HAVO/VWO9 (8.31%).
The participants do not significantly differ per condition on the sociodemographic characteristics age [F(5, 313) = 1.23, p = 0.295], gender [χ2 (5, N =
313) = 6.205, p = 0.287] and education [χ2 (15, N = 219) = 7.567, p = 0.940]10. This
indicates that the participants distributed over different conditions are comparable.
4.1.2 Summary statistics
Figure 4 gives an overview of the chosen answer options for the six different conditions. As can be read from the graph, there seem to be differences in the chosen answers per condition. More accurately, table 3 shows the frequencies, percentages and corresponding chi-square values for each answer option in every condition. The results in the control group indicate which answer people tend to choose when there are no sunk costs involved, this will be used as a baseline. In 84.21% of the cases, answer option 1 is chosen, against 7.02% for option 2 and 8.77% for option 3. This is consistent with prospect theory’s prediction that the safest option is preferred when no sunk costs are involved.
Note that table 3 shows a high value of the chi-square test on the total sample [χ2 (10, N=313) = 129.319, p < .001]. This points out that there is a strong relationship
between the amount of sunk costs and the answer options. One can derive the direction of the relationships from figure 4 and table 3 already. For example, treatment 1 does not differ much from the control group while treatment 2 differs from the control group in that the frequency of chosen answer 2 is much higher and the frequency of answer 1 is lower. Treatment 3, 4 and 5 show a different pattern, namely that option 3 is chosen more often than in the control group. The percentage of people who chose answer 1 is lower in those cases as well. More specific analysis needs to be done to investigate the strength of the relationships. An alpha level of .05 is used for all statistical tests.
9 Higher General Secondary Education 10
The 4 participants with primary school and 10 with VM BO as educat ional level have been deleted for this analysis to enable performing the chi-square test since there are too many cells to perform a Fischer’s exact test.
0 10 20 30 40 50 1 2 3 Treatment 4 0 10 20 30 40 50 1 2 3 Treatment 2 0 10 20 30 40 50 1 2 3 Treatment 5 0 10 20 30 40 50 1 2 3 Treatment 3 0 10 20 30 40 50 1 2 3 F re q u e n c y Control group 0 10 20 30 40 50 1 2 3 Treatment 1 Table 3
Statistics (Frequency, Percentage and Chi-squared) of Chosen Answer Option for Each Condition
Anwer Control Group SCA (€0) Treatment 1 SCP (€60) Treatment 2 SCP (€80) Treatment 3 SCP (€100) Treatment 4 SCP (€100) Treatment 5 SCP (€125) Option N % χ2 N % χ2 N % χ2 N % χ2 N % χ2 N % χ2 1 48 84.21 9.1 39 79.59 5.6 17 32.69 4.6 29 55.77 0.0 16 30.77 5.4 22 43.14 1.2 2 4 7.02 4.5 6 12.24 1.3 32 61.54 47.2 5 9.62 2.6 7 13.46 1.0 7 13.73 0.9 3 5 8.77 6.4 4 8.16 5.9 3 5.77 8.1 18 34.62 1.5 29 55.77 18.0 22 43.14 5.9 Note. Total 2 = 129.3194*, df = 10. Figure 4
Frequency of chosen Answer Option per Condition
4.2 Main results
In order to compare the behaviour of participants in the control group (no sunk costs) with the behaviour in the treatment groups (sunk costs), the data of the five sunk cost present groups have been merged and a chi-square test of independence is calculated. A significant interaction is found [χ2 (2, N=313) = 24.603, p < .001]. The chosen
answer options differ significantly between the groups without sunk costs and with sunk costs. Given that participants are randomly assigned to each group, this indicates that participants’ decisions are influenced by sunk costs. Hypothesis 1 is confirmed:
Hypothesis 1: the presence of sunk costs has an influence on risk taking and subsequently on decisions.
In the above analysis, all treatment groups have been merged together. Now, each treatment group is individually compared to the control group to analyse the relationship between the separate conditions and the control group. Chi-square tests of independence are performed in most of the cases and Fischer’s exact test is used when the Chi-square test is not appropriate. This is the case when not all requirements are fulfilled because there are not enough observations and more than 20% of the cells have a value of less than 5. The results of the tests are reported in table 4. All treatment groups differ significantly from the control group, except for treatment 1, which is the treatment with a sunk cost of €60.
Besides the comparisons between the control groups and treatment groups, all treatment groups are compared to each other as well. This is needed to test the remaining hypotheses; the results are reported in table 4 as well. First of all, note that there is a non-random association between the answers in the control group, treatment 1 and the Sunk Cost Present (€80) indicating that individuals in the Sunk Cost Present (€80) condition are more likely to choose the moderately risky option than individuals in the Sunk Cost Absent and Sunk Cost Present (€60) condition. This already partly confirms hypothesis 2:
Hypothesis 2: When sunk costs are involved, people will be more risk seeking when there is no safe way to ‘recover’ sunk costs.
To be able to totally confirm hypothesis 2, it should be the case that individuals in treatment 3, 4 and 5 more often chose answer option 3 than individuals in the control
group and in treatment 1 and 2. As shown in table 4, there are significant relationships between the given answers for those conditions except for treatment 1 and treatment 3, which have a marginally significant difference in the chosen answers [χ2 (2, N =
101) = 10.391, p = 0.006]. The relationship is in the expected direction, namely that the risky option is chosen more often in treatment 3, 4 and 5 than in the control group and treatment 1 and 2.
Altogether, there are many significant and one marginally significant result that show that people are more risk seeking when there are sunk costs involved and there is no safe way to ‘recover’ sunk costs; hypothesis 2 is accepted.
Subsequently, the next hypothesis is tested:
Hypothesis 3: When sunk costs are involved, people will be more risk averse when there is a safe(r) way to ‘recover’ sunk costs.
It has already been found that individuals in treatment 2 prefer the moderately risky option and individuals in treatment 3, 4 and 5 choose the riskiest option more often than participants in the other conditions. Participants in treatment 1 clearly prefer the safe option, which means that they are significantly more likely to choose the safe option than individuals in treatment 2, 3, 4 and 5. This also entails that individuals in the Sunk Cost Present (€80) condition chose the moderately risky option significantly more often (61,54%) than individuals in treatment 3 (9,62%), 4 (13.46%) and 5 (13.73%). Altogether, people tend to choose the safest option available to ‘recover’ the sunk costs so they are more risk averse when there is a safe(r) way to ‘recover’ sunk costs. Hypothesis 3 is confirmed.
Table 4
Chi-Square Tests and Fischer’s Exact Tests Present Relationship Between Every Condition and the Chosen Answer Options
Condition 0 1 2 3 4 5 0) Control Group SCA (€0) 1) Treatment 1 SCP (€60) (2, N = 106), p = .653 2) Treatment 2 SCP (€80) (2, N = 109), p < .001* (2, N = 101), p < .001* 3) Treatment 3 SCP (€100) χ2 (2, N = 109) = 11.943, p = .003* χ2 (2, N = 109) = 11.943, p = .003 χ2 (2, N = 101) = 10.391, p = .006 χ2 (2, N = 104) = 33.547, p <.001* 0.001* 4) Treatment 4 SCP (€100) χ2 (2, N = 109) = 33.601, p < .001* χ2 (2, N = 101) = 28.571, p < .001* χ2 (2, N = 104) = 37.181, p < .001* 0.001* χ2 (2, N = 104) = 6.663, p = 0.036* 5) Treatment 5 SCP (€125) χ2 (2, N = 108) = 20.910, p < .001* χ2 (2, N = 100) = 17.243, p < .001* χ2 (2, N = 103) = 31.100, p < .001* 0.001* χ2 (2, N = 103) = 1.685, p = 0.431 χ2 (2, N = 103) = 1.899, p = 0.387
