The effects of an imposed cognitive load on economic decision-making : a higher cognitive load can increase choice inconsistency rather than change preferences

(1)

The effects of an imposed cognitive load on economic decision-making:

a higher cognitive load can increase choice inconsistency rather than change preferences

BSc L. de Koning 10334033

Master’s thesis – 15 ECTS

MSc in Economics - Faculty Economics and Business Specialization Behavioral Economics & Game Theory

Supervisor: prof. dr. T.J.S. Offerman Second examiner: dr. J.J. van der Weele

13th July 2018

Abstract

The foundation of this thesis is the recognition that cognitive load affects our economic decision-making and that current research identifies two specific interpretations of the effect of cognitive load. The first is that cognitive load affects our preferences in economic decision-making, while the second states that cognitive load increases inconsistencies in the choices we make. This thesis assumes that these theories are not mutually exclusive and can coexist. However, this study’s findings of an increase in choice inconsistency imply that an effect on our preferences is much less likely to occur than researchers currently assume. This thesis’ objective is to experimentally investigate the likeliness of each theory. The implementation of the multiple price list (MPL) procedure produces results that enable the testing of both theories. The results of this testing imply that choice inconsistency increases for choice sequences including differing levels of risk. Furthermore, the results do not support the theory that cognitive load affects preferences regarding risk taking and temporal discounting of money. Therefore, the theory that cognitive load affects preferences is not supported by any results of this experiment. Consequently, researchers should be conservative when deriving any type of conclusion from experimental data in which subjects had a high cognitive load imposed upon them.

(2)

Statement of Originality

This document is written by Student Liza de Koning who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

1. Introduction

We often make sub-optimal economic decisions in everyday situations when executing multiple tasks simultaneously. We force ourselves to divide our attention and thereby our cognitive resources, i.e. the working memory of our brains. As a result, we make decisions that differ from those we would have made if we had given a matter our undivided focus. This shift in decision-making was originally attributed to a change in preferences on the part of subjects who were experiencing a high cognitive load. Studies have primarily shown that we make more risk-adverse decisions and behave more impatiently with regard to money when a high cognitive load is imposed on the working memory (Hinson, Jameson, & Whitney, 2003) (Whitney, Rinehart, & Hinson, 2008) (Benjamin, Brown, & Shapiro, 2013) (Deck & Jahedi, 2015). Recently, a new theory has emerged stating that imposing a high cognitive load results in more inconsistent decision-making rather than a change in preferences (Franco-Watkins, Pashler, & Rickard, 2006) (Franco-Watkins, Richard, & Pashler, 2010) (Olschewski, Rieskamp, & Scheibehenne, 2018). As this thesis recognizes that these interpretations are not mutually exclusive, both possibilities were tested within one experiment. This decision led to the following research questions:

Does imposing a high cognitive load generate more inconsistencies in economic decision-making?

Does imposing a high cognitive load affect preferences in economic decision-making?

Today, it is not common to assume that all people act rationally. Individuals take too many risks and are overly impatient with regard to their short-term economic decision-making (Laibson, 1997; Tversky & Simonson, 1993). Studies in the fields of psychology and economics have found that the availability of cognitive resources has a sizable impact on this behavior. Research has often manipulated available cognitive resources by affecting cognitive load through a dual-task mechanism to test this. For example, in some studies subjects were instructed to memorize a one-digit number (corresponding to a low cognitive load) or an eight-digit number (corresponding to a high cognitive load) while they made an economic decision according to their preferences.

Results suggested that preferences changed if the imposed cognitive load increased. The dual-system framework (Kahneman, 2003) offers an explanation for this change in

(5)

preferences. According to this framework the human brain possesses two systems, a fast and a slow system, for both thinking and decision-making. The fast system acts automatically, impulsively and makes use of heuristics to minimize effort and optimize efficiency. In contrast, the slow system requires more effort to function and is more analytical, using more cognitive resources to come to a decision. A high imposed cognitive load (and its corresponding decrease in available cognitive resources) might force subjects to move from utilizing their slow system to their impulsive, fast system. Accordingly, the effects of the imposed cognitive load provide an explanation for the overly impatient and risk-taking behavior observed in subjects (Hinson & Whitney, 2006) (Whitney et al., 2008) (Benjamin et al., 2013) (Deck & Jahedi, 2015).

Burdening cognitive resources can also result in a more random response behavior, which leads to an increase in choice inconsistency. Olschewski et al. (2018) designed a mathematical model to test the effects of cognitive load and concluded that choice inconsistencies increase with higher cognitive loads. The authors found no effect on true preferences. Franco-Watkins et al. (2010) concluded that a decrease in cognitive resources complicated a full processing of a choice task and consequently forced subjects into a more random response behavior, which also resulted in higher choice inconsistency.

Earlier, Franco-Watkins et al. (2006) also provided evidence for the phenomenon that an increase in random response behavior can produce similar observable results as a change in preferences. For example, if a subject is instructed to choose between receiving money today and in the future, and initially prefers the “future” option, when the subject randomize his or her answers the results will reflect a more frequent selection of the “ ‘today” option. This occurs because of an averaging effect that pushes preferences evenly towards both options. An increase in the “today” option can now be explained by an increase in the true preference of the “today” option or a more random response behavior. Because this phenomenon might occur during experiments, researchers should take care when making statements about the effect of cognitive load in experimental results.

Although this thesis recognizes that both interpretations of the effect of a high cognitive load do not exclude one another, an increase in inconsistency implies that a change in preferences is much less likely than researchers currently assume This thesis’ objective is to provide more insight into the likelihood of either explanation for the change in observable outcomes. To achieve this, an experiment was developed using the multiple price list (MPL) method that produced results allowing the effects of both interpretations to be tested.

(6)

Evidence for one theory does not rule out the existence of the other, as both theories might be active within one sample. Across subjects, or even within a single subject, it is possible that preferences might change and random response behavior might increase. For this reason both theories must be tested within one experiment. The theory responsible for observed changes can then be distinguished through an analysis of the inconsistency in choices. A random response behavior would result in more inconsistencies in an answering sequence, whereas a true change in preferences should not result in a change of consistency.

The main contribution of this research is its introduction of a variant of the MPL method, also known as the Holt-Laury method (2002), to test both theories within one experiment. The theories were tested using a risk choice task and a temporal discounting task. This MPL method enabled the calculation of a best-fitting switching point per subject and this point’s corresponding errors. An analysis of best-fitting switching points provided insight into changes in preferences, and an analysis of the errors corresponding to the best-fitting strategy yielded insight into the degree of consistency. A between-subject analysis was implemented to impose an extrinsic cognitive load using a dual-task mechanism. The low and high cognitive load treatments consisted of a one-digit and six-digit memorization task, respectively.

This research offers a valuable contribution to the work of Franco-Watkins et al. (2010) and Olschewski et al. (2018). The robustness of their findings was investigated by introducing a new procedure to distinguish the effects of cognitive load on preferences and consistency. Differences in methodological procedures can result in different risk attitudes and different temporal discounting behavior. Researchers currently assume that the MPL procedure is reliable because of its intuitive and transparent character for subjects. However, it should be noted that the MPL procedure might be susceptible to slight framing effects. The interval of the tested choice tasks, i.e. the frame, might activate a psychological bias towards switching in the middle of the interval (Andersen, Harrison, Lau, & Rutström, 2016). In addition, the MPL procedure might be susceptible to incentive effects, since an increase in incentives increases risk aversion (Holt & Laury, 2002). The chosen interval and incentives were set equally across treatments to eliminate any differences between them. It should be noted that the results of this study cannot be extrapolated to behavior in other or wider domains than those tested in this experiment.

This research concludes that an increase in imposed cognitive load can increase choice inconsistencies. In the risk choice task, a significant increase was found in errors corresponding to the best-fitting switching points. This increase suggests that subjects with a

(7)

high cognitive load were forced into a more random response behavior in the risk choice task. The same effect was not found for the temporal discounting task. Furthermore, no effects on true preferences were found in the risk choice task or the temporal discounting task. These results do not support the theory that imposed cognitive load influences preferences in economic decision-making. As a result, it is less likely that the theory that cognitive load affects preferences is correct. Consequently, researchers should be conservative in drawing any conclusions about the effects of cognitive load on preferences.

Literature about the current discussion is analyzed in the following chapter. The third chapter elaborates the methodology and the hypotheses of this thesis. The fourth chapter describes the results of this study, which are critically discussed in the fifth chapter. Finally, the sixth chapter concludes this research with a summary.

(8)

2. Literature Review

The literature review is organized as follows. The first section discusses research on behavioral biases created by an imposed cognitive load and focuses particularly on research on risk attitude and temporal discounting behavior. The second section presents research that claims that imposed cognitive load influences risk and discounting preferences. The third section presents contrary results and focuses on how imposed cognitive load induces a more random response behavior. In the fourth section, the random response theory is refuted in a discussion of a paper by Deck and Jahedi (2015). This chapter concludes by describing the contribution of this experiment to existing literature.

2.1 Behavioral Biases

Standard economics relies on the expected utility theory. According to this theory, individuals prefer goods or lotteries with the highest expected utility over any lower utility. However, several structural deviations of behavior from the expected utility theory have been discovered. These structural deviations are hereafter referred to as behavioral biases.

According to research, several correlations exist between behavioral biases and cognitive ability. Carpenter, Graham and Wolf (2013) found a positive correlation between cognitive ability and strategic sophistication in games by testing for inductive reasoning, iterative dominance and k-level thinking. Relevant to this experiment, Graham and Wolf further found that high cognitive ability also correlates with a higher risk tolerance and more patient behavior if patience is rewarded with money. Benjamin et al. (2013) supported this finding through their discovery of a relationship between school grades and risk attitude, as well as between school grades and patience. Frederick (2005) found a relationship between SAT/ACT scores and risk attitude and between SAT/ACT scores and patience. Other scholars have found a relationship with IQ scores (Burks, Carpenter, Goette, Rustichini, & Dixit, 2009; Dohmen, Falk, Huffman, & Sunde, 2010). In sum, research has established that subjects with high cognitive ability implement more sophisticated strategies in games, make more risky decisions and are more patient if patience is rewarded by money.

A high cognitive load on working memory affects available cognitive resources and can therefore alter cognitive ability within a person. A high imposed cognitive load lowers cognitive ability and affects behavior. Bergman, Ellingsen, Johannesson and Svensson (2010) and Deck and Jahedi (2015) found an increase in susceptibility for anchoring effects when

(9)

cognitive load was increased. Evidence for the effect of cognitive load on generosity and selfishness is mixed. Benjamin et al. (2013) found that subjects acted more selfishly in a dictator game when a high cognitive load was induced. However, Van den Bos, Peters, Bobocel, Ybem and Fekke (2006) found the exact opposite effect, i.e. that subjects disliked having an advantage relative to others. Furthermore, according to the research of Carpenter et al. (2013) cooperation among subjects decreases in the prisoner’s dilemma game with higher cognitive load. In addition, the strategic sophistication in the prisoner’s dilemma games is also correlated negatively to cognitive load (Duffy & Smith, 2014).

This research focuses on the effect of cognitive load on risk attitude and temporal discounting preferences. Whitney et al. (2008), Benjamin et al. (2013) and Deck and Jahedi (2015) claim that cognitive load increases risk-averse behavior, and Hinson et al. (2003), Benjamin et al. (2013) and Deck and Jahedi (2015) claim that cognitive load increases impatience regarding temporal discounting of money. Franco-Watkins et al. (2006) (2010) and Olschewski (2018) contradict both of these interpretations and claim that cognitive load only increases inconsistencies in decision-making and does not change preferences. Table 1 summarizes the literature discussed above regarding the correlation of cognitive ability and the effects of cognitive load on behavioral biases. Later sections of this thesis consider these effects of cognitive load in more depth.

Table 1

Summary of main research regarding the effects of cognitive load and cognitive ability on behavioral biases

Paper Finding - Cognitive Load Finding - Low Cognitive Ability

Risk preferences

Whitney et al., 2008 Increased risk aversion

Burks et al., 2009 Increased risk aversion

Dohmen et al., 2010 Increased risk aversion

Benjamin et al., 2013 Increased risk aversion Increased risk aversion Deck and Jahedi, 2015 Increased risk aversion

Olschewski et al., 2018

Increased inconsistency

Temporal discounting preferences

Hinson et al., 2003 Increased impatience over money Franco-Watkins et al.,

2006

(10)

Dohmen et al., 2010 Increased impatience over money Franco-Watkins et al.,

2010

Benjamin et al., 2013 Increased impatience over money Increased impatience over money Deck and Jahedi, 2015 Increased impatience over money

Olschewski et al., 2018

Anchoring effects

Bergman et al., 2010 Increased susceptibility to anchoring effects

Deck and Jahedi, 2015 Increased susceptibility to anchoring effects

Strategic sophistication

Van den Bos et al., 2006

Decreased satisfaction with advantageous inequity

Carpenter et al., 2013 Decreased cooperation in prisoner's dilemma game

Increased strategic sophistication in games (inductive reasoning, iterative dominance and k-level thinking)

Benjamin et al., 2013 Increased selfishness in dictator game

Duffy and Smith, 2014 Decreased strategic sophistication in prisoner’s dilemma game

2.2 Cognitive Load and Individual Preferences

Kahneman and Frederick (2007) suggested that if working memory capacity is low (or if a cognitive load is imposed) the individual’s slow system will be less successful in overriding the fast system, meaning that the cognitive system as a whole becomes more susceptible to behavioral biases. Since the slow system relies on cognitive resources, an imposed cognitive load might force an individual to use the fast system. This switch in system usage can cause an increase in susceptibility to perception errors and corresponding behavioral biases. Research suggests that these biases can vary within a person depending on the availability of cognitive resources (Deck & Jahedi, 2015; Hinson et al., 2003; Benjamin et al., 2013).

(11)

To test impatience over monetary outcomes, standard intertemporal questions are often used and repeated with differing amounts and/or delays. Subjects are instructed to make multiple choices between two amounts of money that can be received at two different points in time. From these choices an individual delay discounting function can be obtained, which describes how value declines with delay. The hyperbolic discounting function is used most often and is described in the following formula:

𝑉 = 𝐴

(1 + 𝑘𝐷)1

The variable of interest is k, which is the discount rate. It is common to offer an amount of money immediately or “today” and another amount of money at a future point of time. However, receiving a certain amount of money immediately includes fewer factors of uncertainty than receiving money in the future, even if the future option is as near as the next day. This absence of uncertainty in the immediate offer is hereafter referred to as the certainty bias. Asymmetric uncertainty is created if an immediate offer is included as one of the options, whereas symmetric uncertainty is created if two future options are included for which the only differing factor is time (Harrison, Lau, Ruström, & Williams, 2005).

In psychology, Hinson et al. (2003) found that an imposed cognitive load increased impatience over temporal discounting; the variable k increased when a cognitive load was imposed, i.e. preferences for the immediate reward increased. The authors explained this increase in k as an impulse control problem induced by the cognitive load.2 It should be noted that the experimental standards in economics allow for little variation in experimental design, while studies in psychology often implement different experimental standards. Therefore, attention must be paid to designs that do not meet the experimental standards in economics (Davis & Holt, 1993; Hertwig & Ortmann, 2001). The conclusion that cognitive load induces an impulse control problem was based on two experiments. The first experiment used a large sample size but relied on hypothetical stakes. The second experiment used real-world stakes, but unfortunately its subjects included only a small sample of 20 psychology students. In addition, Hinson et al. (2003) conducted both experiments using subjective script-enactment.

1_{V refers to the value, A to the monetary amount in the choice option, k equals the discount rate and D the}

amount of the delay.

2_{Impulse control problems have been linked to working memory capacity before. Previous research found a}

link between working memory capacity and susceptibility for impulse control problems triggered by excessive alcohol consumption (Finn, Justus, Mazas, & Steinmetz, 1999).

(12)

Subjects were assured that no right or wrong answers existed. Still, the sentence “Please try to be consistent in the choices you make” (Hinson et al., 2003, p. 300) was included in the instructions, which suggested that inconsistent answers were not desirable. In addition, the experiment utilized asymmetric treatments because no memorization task was implemented in the control condition. Additionally, a certainty bias might have been present in offering the early reward “today”, which harms the ceteris paribus condition. Based on these issues researchers should be careful when interpreting the results of this experiment.

Whitney and Hinson also conducted research focusing on the effect of framing in combination with an imposed cognitive load on decision-making involving risks (Whitney et al., 2008). They suggested that if cognitive load affected working memory such that a subject is not able to fully process a choice task, subjects would rely more on their fast system of thinking, i.e. the heuristic system, than on their slow system. Accordingly, framing should have a greater influence on subjects under high cognitive loads than those under low cognitive loads. Of 60 subjects, 8 appeared to randomize their answers or selected the same choice for all trials. Unfortunately, the authors omitted the data from these subjects and included eight new subjects in their experiment instead. They justified this decision by claiming that the subjects were not “adequately engaged in the decision task” (Whitney, Rinehart, & Hinson, 2008, p. 1182). Because the data from subjects who randomized their answers are of special interest in this study, any further statistical tests and conclusions from the study of Whitney et al. (2008) are useless regarding the focus of this research.

In the field of economics, Benjamin et al. (2013) conducted research on the effect of cognitive load on risk attitude and temporal discounting preferences in Chilean high school students in their third year. Inducing a cognitive load did not influence math performance results significantly. Because mathematical skills were not impaired, there was no reliable evidence that cognitive ability was manipulated successfully. Despite the weak evidence of cognitive manipulation, Benjamin et al. still found that a high cognitive load induced higher risk aversion and impatience over money. Regarding the robustness of the experiment it must be noted that the ceteris paribus condition was harmed because the tested conditions were asymmetric as no memorization task was implemented in the control condition. In addition, the certainty bias was not eliminated. Again, care should be taken in drawing any type of conclusions from this experiment since its sample size was small and no evidence was found that cognitive load influenced cognitive ability.

(13)

2.3 Cognitive Load and Random Response Behavior

A reanalysis of the consistency of responses in Hinson et al.’s (2003) first experiment by Franco-Watkins et al. (2006) contradicted that study’s conclusions. Franco-Watkins et al. (2006) predicted the best-fitting value of k for each subject and measured the number of choices that differed from the predicted responses, i.e. the errors. They found that subjects made more errors and consequently responded less consistently in their choices under the high cognitive load condition.3 Moreover, the increase in k was positively correlated with an increase in errors. If the increase in k was purely the result of higher impatience, the number of errors should not have correlated to this increase in errors. Consequently, Franco-Watkins et al. concluded that cognitive load did not influence impatience but rather induced a more random response behavior. Also, the authors noted that on average, 25% of subjects preferred the immediate option in the control condition while 30% showed this preference in the digit memorization condition. They concluded that this increase in average preference was created by an increased random response behavior, which would push the average preference towards 50%.

As part of their reanalysis Franco-Watkins et al. (2010) conducted their own pilot study and eliminated initial preferences for one reward over another. In this study Franco-Watkins et al. did not find any changes in the rate of discounting; subjects seemed to be indifferent to both options under both conditions. This supported their view that an observed increase in impulsivity was the result of an increase in random response behavior, not a change in preferences. In a second experiment they used a hypothetical, fixed $500 future reward and a hypothetical fixed $10,000 future reward in combination with an immediate option of five possible values and eight possible delays.4 Their results, presented in Figure 1, did not show that subjects became more impulsive under higher cognitive load across time, and the value of k did not differ between the control and high cognitive load conditions. A horizontal line represents the hypothetical result if every subject implemented a random response strategy. As can be seen from the graphs, it appears that the imposed cognitive load moved subjects towards a more random response strategy since the slope of the cognitive load condition moved towards the random response strategy line. More importantly, the cognitive load condition produced more errors than the control condition. Franco-Watkins et al.

3_{Hinson and Whitney (2006) responded to this and noted that this commentary was based upon a statistical} anomaly due to the presence of outliers. They noted that if four (out of 44 subjects) most extreme values are omitted, the degree of inconsistency drops to a nonsignificant score and the increase in k is still significant.

4_{The immediate options consisted of $50, $150, $250, $350 and $450 and the possible delays of 1 month, 6}

(14)

concluded their research with the same internal consistency test of individual choices that they used in 2006 and found a negative correlation between k and an increase in errors. A decrease in k might be caused by a minor initial preference for the immediate reward option in the control condition. A random response behavior then increased the preference of the future option and decreased k. However, it must be noted that results of both experiments relied fully on hypothetical payoffs, meaning the certainty bias was again not eliminated. Also, the ceteris paribus condition was harmed again by not including a memorization task in the control condition.

Figure 1

Mean proportion of selecting the immediate choice across time for (a) $500 reward (b) $10,000 reward.

Note: Reprinted from “Taking Executive Processes Does Not Necessarily Increase Impulsive Decision Making”, by Franco-Watkins, A. M., Richard, T. C., & Pashler, H., 2010, Experimental Psychology, 198.

Using a different approach than other researchers, Olschewski et al. (2018) designed a mathematical method to investigate the effects of cognitive load on risk attitude and temporal discounting preferences. While all previously discussed research implemented a digit memorization task to impose a cognitive load, Olschewski et al. made use of an n-back task,

(15)

which had previously proven to be a strong and reliable procedure (Cohen, et al., 1994). A three-back task was implemented to impose cognitive load. The three-back task required subjects to listen to a sequence of letters and push a button when they heard the letter that corresponded to third latest letter. By using a probit and a trembling hand model, both stochastic choice models, Olschewski et al. concluded that cognitive load negatively affected choice inconsistency. No evidence was found that cognitive load affected risk attitude or temporal discounting preferences. One positive aspect of this experiment is that the ceteris paribus condition was accounted for by including a simplified version of the n-back task in the control condition. Therefore, the effects of an increase in cognitive load were tested in comparison to an introduction of cognitive load. Unfortunately, the certainty bias was again not eliminated.

2.4 Deck and Jahedi’s (2015) Original Paper

In economics, Deck and Jahedi (2015) conducted a well-designed experiment to test the effect of cognitive load on economic decision-making and were able to test for consistency across treatments. The authors’ main findings were that cognitive load led to more risk-averse behavior, higher impatience over money and a higher likelihood to anchor. They also tested for impatience over consumption but did not find any significant results. The experiment in this thesis follows a design similar to the experiment of Deck and Jahedi (2015), and for this reason this section elaborates on the details of this experiment.

During their first experiment Deck and Jahedi found that an eight-digit memorization task manipulated cognitive ability successfully. Memorizing either one digit or eight digits resulted in an equal performance in adding two numbers correctly. This is important because subjects must not be overloaded and must still be able to make meaningful decisions. Memorizing eight digits, however, resulted in a significantly worse performance in multiplying two numbers than that observed for the subjects who memorized only one digit. This indicated a successful severity of the direct induced cognitive load produced by this memorization task. Degree of risk aversion was tested by asking subjects choose between a safe option and risky option regarding gains and losses. The risky option always had an expected value that was $1 higher than the safe option, and this was repeated for differing expected values. To test impatience over monetary outcomes, standard intertemporal questions were used and repeated for differing amounts of money and differing delays in payments. The results of the first experiment revealed more risk-averse decisions, a tendency towards anchors and, contrary to expectations, more patient behavior.

(16)

A second experiment with real stakes was conducted to test whether results differed from expectations, as the first experiment relied on hypothetical payoffs only. A money and a consumption impatience task as well as an immediate and a delayed snack choice task were tested. In order to be consistent with existing research about impatience regarding consumption, Deck and Jahedi assigned zero-digit, two-digit and eight-digit memorization tasks. In both impatience tasks subjects were asked to choose between receiving an amount of money/number of snacks today or a larger amount in one week. In the snack choice tasks subjects were instructed to choose between receiving healthy and unhealthy snacks today or in one week. The money impatience task yielded results that were contradictory to Deck and Jahedi’s first experiment and demonstrated a higher impatience for real money, findings that are in line with existing literature. Unfortunately, Deck and Jahedi did not include an arithmetic test in the second experiment, so it is not clear whether these memorization tasks affected cognitive ability successfully. Moreover, they implemented similar choice tasks in the two-digit and eight-digit memorization tasks but used different choice tasks in the zero-digit task. Since the authors compared different choice tasks potential perceptional differences might have been present, resulting in behavioral biases that were not in interest of the experiment. This harmed the ceteris paribus condition and might have influenced results.5 The process of how multi-digit numbers are compared and the dependency on the numbers that are compared has been well established in literature (Schindler & Kirby, 1997; Poltrock & Schwartz, 1984; Coulter & Coulter, 2007). Also, the certainty bias was again not eliminated.

As in Franco-Watkins et al. (2006), a level of internal consistency test among individuals’ choices was carried out and subjects did not appear to be less consistent in their choices under high cognitive load. Furthermore, Deck and Jahedi argued that verification questions were often answered in a utility maximizing way, indicating that subjects were not randomizing their answers when answering. Still, it is possible that subjects possessed more than enough cognitive resources to answer the easily identifiable verification questions rationally and simultaneously possess too few cognitive resources to answer the other choice tasks rationally, forcing subjects to randomize their answers to these questions. Deck and Jahedi did mention that it is possible that a higher cognitive load generated more random choices because behavior in the high cognitive load condition is closer to 50% than that observed in the control condition.

5_{For example, the Weber-Fechner Law suggests that consumers perceive a larger difference related to the}

smaller right digits in the $23 and $22 comparison than to the larger right digits in the $19 and $18 comparison, which represents a larger relative difference. If one of the conditions measured more of these choice tasks, this behavior might have been measured in the experiment as an effect induced by the cognitive load.

(17)

2.5 Holt and Laury’s (2002) MPL Method

Risk aversion is often tested using the MPL method introduced by Holt and Laury (2002), however this method has not yet been implemented to measure the effects of cognitive load on risk aversion or any other preferences.

Holt and Laury used the MPL method to elicit risk preferences by presenting a list of 10 lotteries to subjects. Each subject was instructed to choose between (option A) a safe lottery and (option B) a risky lottery in 10 different scenarios. The payoffs of the safe lottery were earnings of $2.00 and $1.60, and in this lottery the probability of winning $2.00 increased across scenarios. The winnings of the risky lottery were earnings of $3.85 and $0.10. This lottery involved higher risks since the difference between the winnings was larger. For option B, the probability of winning $3.85 also increased across lottery scenarios. The list of lottery scenarios is presented in Table 2. Since the final choice task involves deciding between a certain $2.00 in the safe lottery and a certain $3.85 in the risky lottery, a subject should switch somewhere from the safe option to the risky option. This switching point indicates the degree of risk aversion of the subject. However, if a subject switches between lotteries more than once he or she exerts an inconsistent strategy. The number of inconsistent responses is perceived as valuable information in the economic decision-making model and analyzed in further detail in this experiment.

Table 2

The 10 paired lottery-choice decisions

Option A Option B 1/10 of $2.00, 9/10 of $1.60 1/10 of $3.85, 9/10 of $0.10 2/10 of $2.00, 8/10 of $1.60 2/10 of $3.85, 8/10 of $0.10 3/10 of $2.00, 7/10 of $1.60 3/10 of $3.85, 7/10 of $0.10 4/10 of $2.00, 6/10 of $1.60 4/10 of $3.85, 6/10 of $0.10 5/10 of $2.00, 5/10 of $1.60 5/10 of $3.85, 5/10 of $0.10 6/10 of $2.00, 4/10 of $1.60 6/10 of $3.85, 4/10 of $0.10 7/10 of $2.00, 3/10 of $1.60 7/10 of $3.85, 3/10 of $0.10 8/10 of $2.00, 2/10 of $1.60 8/10 of $3.85, 2/10 of $0.10 9/10 of $2.00, 1/10 of $1.60 9/10 of $3.85, 1/10 of $0.10 10/10 of $2.00, 0/10 of $1.60 10/10 of $3.85, 0/10 of $0.10

Note: Adapted from “Risk Aversion and Incentive Effects,” by

Holt, C. A., & Laury, S. K., 2002, American economic review, 1645.

(18)

The MPL procedure has a few effects that should be considered in the design of the experiment and when drawing conclusions from its results. Research using the MPL method presents choice tasks either simultaneously or sequentially. Presenting choice tasks sequentially does not allow for making any corrections in their answers and therefore increases choice inconsistencies (Lévy-Garboua, Maafi, & Masclet, 2012). It is common to eliminate wealth effects by informing subjects that one of their decisions will be selected at random and the chosen scenario will be run in the real world and winnings will be paid to the subject. Therefore, it is rationalized to perceive each choice task separately and not in line with the other choice tasks. The level of incentive also influences elicited preferences. Research finds that once incentives are scaled up risk aversion increases (Holt & Laury, 2002).

Susceptibility to potential framing effects is the most important consideration in the design of a MPL experiment. The frame, i.e. the interval of the tested choice tasks, might activate a psychological bias to switch in the middle of the frame (Andersen et al., 2016). The frame might serve as a kind of anchor that indicates a range in which a subject should switch their preference. Previous research has proven that subjects with a higher cognitive load are more susceptible to anchoring effects (Bergman, Ellingsen, Johannesson, & Svensson, 2010; Deck & Jahedi, 2015). If subjects are also more susceptible to framing effects, switching points will trend towards the middle of the interval when a cognitive load is imposed. Researchers currently assume that this effect is smaller in sequential presentation of choice tasks, since sequential presentation reduces the transparency of the frame of the choice tasks. Moreover, this framing effect should not create higher inconsistency in the answering sequence. No research has investigated framing effects caused by an imposed cognitive load. Therefore, it should be kept in mind that imposing a cognitive load might activate framing effects.

The main contribution of this research is its investigation of the robustness of previous findings by introducing a new procedure to determine the effects of cognitive load on preferences and consistency. The MPL method was utilized because its results allowed for testing the likeliness of two current theories on the effect of cognitive load. It was possible to test whether an imposed cognitive load led to a more random response behavior by evaluating the increase in errors in the answering sequence of the MPL method. It was also possible to assess whether cognitive load affected preferences and changed the location of the subjects’ switching points. The MPL method is assumed to be reliable because of its intuitive character

(19)

and transparent structure. However, it should be noted that its results are not generalizable to other intervals or for other incentive structures.

(20)

3. Methodology

The experimental design used in this research is comparable to the one used in Deck and Jahedi’s (2015) experiment. However, in this study the MPL method was introduced as a new component. In this thesis, Deck and Jahedi’s (2015) experiment is hereafter referred to as the DJ experiment. Deviations from the original experimental design constructed for this study were applied as necessary, and all deviations are clarified thoroughly in this chapter. In this chapter, the main design and organization of the experiment are explained. Next, the memorization task and the two main tasks are described. The experimental instructions are elaborated, followed by the payment scheme used in the experiment. Finally, this chapter concludes with its expectations regarding the ex-ante hypotheses.

3.1 Experimental Design

A total of 40 subjects were recruited for this experiment. All subjects were family, friends or acquaintances of the researcher. Subjects participated in the experiment individually online in Dutch or English and the researcher supervised all subjects during this process. Subjects required supervision to ensure that they did not cheat during the memorization task and were able to remember the numbers without the use of any tools such as a pen and paper. The experiment was conducted using Qualtrics software (Qualtrics, Provo, UT). Subjects were evenly and randomly divided between the two treatments. Treatments differed only in degree of cognitive load. A low cognitive load treatment and high cognitive load treatment were induced using a one-digit memorization task and a six-digit memorization task, respectively.

A between-subject design was chosen for two reasons. First, a between-subject design supports testing the same list of choice tasks, including the same numeric values, in both treatments without subjects recognizing a choice task. As previously mentioned, specific choice tasks could induce perception errors, which might result in behavioral biases. Since these behavioral biases were not of interest in this experiment and since they harm the ceteris paribus condition, behavioral biases had to be avoided. For this reason, the same list of choice tasks was presented to subjects in both treatments. The second reason for choosing a between-subject design was that a larger list of choice tasks could be included in each treatment since subjects were only participating in one treatment. Subjects recruited for this experiment were not rewarded with extensive monetary payments or course credits. For this reason, the duration of the experiment was minimized to increase the likelihood that subjects would

(21)

actively participate for the duration of the experiment. Nevertheless, potential data points were maximized to improve the quality of the statistical results and a total of 20 periods were conducted within the experiment.

Furthermore, some changes were made to the original design of the experiment to avoid measuring any results created by behavioral biases, which were not of interest in this experiment. It was taken into account that subjects under high cognitive loads might be more susceptible to behavioral biases created by time pressure, feedback and “rules-of-thumb.” Accordingly, any time limitations for providing answers and feedback about earnings were eliminated from the experiment. Also, the safe and risky, and early and future options were randomized between choices A and B. The randomization of option type disabled the usage of strategies such as “always choose the left option.” As explained earlier, such strategies can mask an increase in inconsistent behavior, and therefore options were randomized.

3.2 Task Design

Each of the 20 periods presented a memorization task and a main task. The main tasks were divided into two categories, risk choice tasks and temporal discounting tasks. Each period was organized as follows. A number was presented to the subject for three seconds, after which a decision was required in the main task. After this decision, subjects were asked to recall the number they had seen in the beginning of the period. Choice tasks were sequentially presented and alternated randomly between risk choice tasks and temporal discounting tasks.

The memorization element was a one-digit or six-digit memorization task. To maintain consistency across both groups a low cognitive load was selected over an absence of cognitive load in the design of the control treatment. This research utilized a six-digit memorization task for the high cognitive load treatment because this number of digits was seen as neither too easy nor too difficult to retain in one’s memory. In the trial phase conducted for this research the eight-digit memorization task resulted in a critically low number of correctly recalled numbers. The high cognitive load condition was thus reduced from an eight-digit memorization task to a six-digit memorization task to prevent subjects from giving up on memorizing the number.

The risk choice task (hereafter referred to as the risk task) elicited individual risk preferences using the MPL method with several key variations from the method used by Holt and Laury (2002). The MPL method is a suitable procedure for measuring the degree of consistency of behavior. Every subject that switches more than once from a safe to a risky option can be seen as using an inconsistent strategy. As in Holt and Laury’s (2002)

(22)

experiment, subjects were instructed to make multiple choices between a safe option and a risky option in a list of lottery choices. Unlike the experiment of Holt and Laury (2002), lotteries were sequentially presented. A sequential presentation allowed inducing a fresh cognitive load for each decision. This type of presentation also reduced the transparency of the interval of choice tasks and therefore weakened any framing effects. Another key difference between this experiment and Holt and Laury’s is that probabilities were kept constant while the values of the lotteries were varied. All risky options included a 50-50 probability of winning or losing. By raising winnings from the risky option, the expected value of the risky lottery increased, making the risky option more attractive. The exact values of the lotteries are presented in Table 3. Unlike the DJ experiment, the experiment in this study focused on behavior regarding frames of gains only, not frames of losses. Previous research has found no evidence for differences in risk-taking behavior induced by cognitive load across loss and gain frames (Whitney, Rinehart, & Hinson, 2008; Benjamin, Brown, & Shapiro, 2013; Deck & Jahedi, 2015). Therefore, this study evaluates the effects on frames of gains only.

Table 3

Choice options in time task (T) and risk task (R)

Safe Risk Tomorrow Eight

Days Lose Win R1 €11.00 €0.00 €18.00 T1 €11.00 €10.75 R2 €11.00 €0.00 €20.00 T2 €11.00 €11.00 R3 €11.00 €0.00 €22.00 T3 €11.00 €11.25 R4 €11.00 €0.00 €24.00 T4 €11.00 €11.50 R5 €11.00 €0.00 €26.00 T5 €11.00 €11.75 R6 €11.00 €0.00 €28.00 T6 €11.00 €12.00 R7 €11.00 €0.00 €30.00 T7 €11.00 €12.25 R8 €11.00 €0.00 €32.00 T8 €11.00 €12.50 R9 €11.00 €0.00 €34.00 T9 €11.00 €12.75 R10 €11.00 €0.00 €36.00 T10 €11.00 €13.00

The temporal discounting task (hereafter referred to as the time task) elicited individual time preferences through a procedure comparable to that used in the risk task. Subjects could choose between receiving an amount and money the next day or an amount of money one week later, i.e. in eight days. Increasing the amount of money that could be

(23)

received in eight days made the future option more attractive. The exact time choice tasks are also presented in Table 3.

Unlike the DJ experiment, the early option stated an amount of money that was available “tomorrow” instead of “today.” This adjustment was made for two important reasons. The first is related to the payment of the rewards and is explained in the fourth section of this thesis. The second is related to the previously explained certainty bias. According to certainty bias, an immediate payment yields no uncertainty regarding whether the payment will be received. Consequently, different factors support the choice of the immediate option versus the choice of the future option. To keep underlying motivations for choosing either of the options equal, uncertainty was introduced into the early option by moving the early option from “today” to “tomorrow.” However, it should be noted that moving the reception of the early reward from “today” to “tomorrow” might mitigate the degree of inconsistency. A high imposed cognitive load might trigger the impulse control problem if a sum of money is offered immediately; if a subject has a high imposed cognitive load, the slow system might not always be able to override the fast system that prefers the immediate reward (Kahneman & Frederick, 2007). This failure of system override might result in more errors and higher inconsistency. However, if the sum of money is not offered immediately the fast system does not push the brain to choose the early option. Therefore, moving the early option from “today” to “tomorrow” might result in fewer impulse control problems and consequently a lower degree of inconsistency.

3.3 Instructions

Subjects in both treatments read the same instructions. First, subjects were instructed regarding the risk task, including a memorization task, after which they could practice the task in a trial period. The time task was explained next, followed by another trial period. Subjects were discouraged from believing that their choices involved any computational skills. Instead, they were told that their choices revealed their true preferences. This was achieved using the following the sentence (taken from Benjamin et al. (2013)) in the instructions: “There are no right or wrong options to choose, which choice you make is a matter of personal preference.” Trial periods contained additional questions that required subjects to provide the correct answer before it was possible to continue the experiment. These additional questions were not included in the actual tasks. The questions were included to ensure that subjects understood potential payoffs. After the trial periods were complete, subjects were again informed about

(24)

how money could be earned. Subjects were instructed to begin the 20 periods if they had no further questions. The full instructions used in this experiment are included in Appendix A.

Some changes were made to the instructions to avoid any unnecessary differences between the safe and risky options and the early and future options. First, visuals that supported the choice tasks were presented in blue. This was done to avoid any perception errors that could have resulted in behavioral biases. Second, all labels used in the tasks were referred to in neutral language to prevent labels from having any biased influence. For example, the one-digit memorization task was never described as “easy” and the six-digit memorization task was never described as “hard”. Since it was possible that subjects with a high imposed cognitive load might be more susceptible to behavioral biases resulting from color differences and labels these factors were eliminated from the instructions.

In the beginning of the experiment subjects were asked to provide some standard demographic details, including their age, gender and most recently completed level of education. These data points were used as control variables in some statistical regression analyses, the results of which are explained in the following chapter. At the end of the experiment some post-experimental questions were asked using a 1-5 Likert scale to determine the degree of perceived difficulty, the effort involved in each memorization task and whether subjects trusted that they would be paid at the specified point in time.

3.4 Payment Scheme

No participation fees were paid to subjects since the experiment was not executed in a laboratory and the budget was limited. Instead, one subject and one of his or her decisions were selected at random in advance of the start of the experiment and winnings were paid to the subject after the experiment. The rewarded subject and his or her decision were chosen in advance to ensure that earnings would be received at the promised point in time. Subjects were assured that everyone had an equal probability of being rewarded once they completed the experiment and that every decision the subject made was equally important.

To maintain consistency across the early and future options of the time task the researcher decided to transfer any potential earnings by bank. At the start of the experiment,6 many Dutch banks could not guarantee that money would be received the same day it was transferred and could only ensure that money would be transferred to another bank account within one day if the two bank accounts were in the same currency, in this case euros. For this

(25)

reason, it was necessary to change the early option of receiving money from “today” to “tomorrow.” Moreover, the researcher verified that all subjects possessed a bank account in euros.

The memorization reward of correctly recalling the number was always equal to €18. This reward was always higher or equal to the maximum expected value of the options that were part of the main tasks. The instructions explained this explicitly to encourage subjects to devote their primary attention to the memorization task. This was important to ensure that the cognitive load was imposed successfully. Moreover, it was stated that the choice of paying either the earnings of the memorization task or the earnings of the main task would be selected at random. This was done to eliminate any wealth effects that might have been present. The elimination of these effects meant it was no longer possible to rationally perceive any earnings from the main task as a substitute for missed earnings from the memorization task.

3.5 Hypotheses

To test the effect of an imposed cognitive load, a best-fitting switching point (BFSP) was estimated for the answering sequence of each subject. A separate BFSP was estimated for the risk task and for the time task. A BFSP is the switching point that is the closest to the preferences of the subject and indicates when the subject should switch from the safe to the risky option or from the early to the future option. This BFSP assumes consistency in the answering sequence and every choice that differs from the strategy of the BFSP model is defined as an error. The BFSP was estimated in such a way that it minimized its corresponding errors. An error is defined as a difference between the chosen preference and its best-fitting strategy. For example, if a subject made choices for all risk tasks in Table 3, and for R2, R3 and R4 he or she preferred the safe option while for R1, R5 to R10 he or she preferred the risky option, the BFSP states that this subject should have switched from the safe to the risky option for choice task R5, and the BFSP is set to 5. The BFSP states that the subject prefers the safe option until the risky option reaches an expected value of €13 or more, at which point he or she is willing to take the risk. If this is true, his or her choice of R1 (i.e. a preference for the risky option when the expected value equals only €9) must be an error. Because of this error his or her answering sequence is inconsistent.

It was expected that inconsistencies would increase if subjects were affected by a high cognitive load. Based on this, the number of errors corresponding to the BFSPs was expected

(26)

to be higher in the high cognitive load treatment than in the low cognitive load treatment. These expectations led to the first two ex-ante hypotheses:

(1) 𝒆𝒓𝒓𝒐𝒓𝒔_{𝒕𝒊𝒎𝒆, 𝒉𝒊𝒈𝒉 𝒍𝒐𝒂𝒅} > 𝒆𝒓𝒓𝒐𝒓𝒔_{𝒕𝒊𝒎𝒆, 𝒍𝒐𝒘 𝒍𝒐𝒂𝒅}

(2) 𝒆𝒓𝒓𝒐𝒓𝒔_{𝒓𝒊𝒔𝒌, 𝒉𝒊𝒈𝒉 𝒍𝒐𝒂𝒅} > 𝒆𝒓𝒓𝒐𝒓𝒔_{𝒓𝒊𝒔𝒌, 𝒍𝒐𝒘 𝒍𝒐𝒂𝒅}

Also, the imposed high cognitive load was expected to cause more impatient and risk-averse behavior. According to the literature reviewed in this study, this change in behavior was expected to result in an increase in preference for both the immediate and the safe option in the choice tasks. Accordingly, it was expected that BFSPs would be lower in the high cognitive load treatment compared to the low cognitive load treatment. These expectations resulted in the third and fourth ex-ante hypotheses:

(3) 𝑩𝑭𝑺𝑷𝒕𝒊𝒎𝒆, 𝒉𝒊𝒈𝒉 𝒍𝒐𝒂𝒅< 𝑩𝑭𝑺𝑷𝒕𝒊𝒎𝒆, 𝒍𝒐𝒘 𝒍𝒐𝒂𝒅

(4) 𝑩𝑭𝑺𝑷𝒓𝒊𝒔𝒌, 𝒉𝒊𝒈𝒉 𝒍𝒐𝒂𝒅 < 𝑩𝑭𝑺𝑷𝒓𝒊𝒔𝒌, 𝒍𝒐𝒘 𝒍𝒐𝒂𝒅

This study tested these four ex-ante hypotheses according to the described experimental design. The corresponding results are elaborated in the next chapter.

(27)

4. Results

A total of 40 subjects were recruited and supervised during this experiment. Subjects were evenly divided between the two treatments. Each of the subjects completed 20 binary choice tasks and was instructed to memorize a number and remember it throughout their completion of these tasks. This resulted in 40 observations per subject, comprising a total of 1,600 observations for the study as a whole. In the first section of this chapter the general characteristics of the sample are described. The second section analyzes performance in the memorization task and the third section analyzes the degree of consistency by analyzing the errors corresponding to the BFSPs. The final section discusses the results of the BFSPs for the risk task and the time task.

4.1 Characteristics of the Sample

The average age of the subjects was 31.2 years, ranging between 21 and 62 years. Of the subjects, 62.5% were female and 37.5% were male. Most of the subjects were highly educated, with 50% possessing a master’s degree, 17.5% holding a bachelor’s degree and 27.5% holding an associate’s degree. All subjects were friends or acquaintances of the researcher. Subjects were supervised to ensure honest behavior and that they memorized the numbers themselves. Because subjects were supervised it is assumed that all subjects took a serious attitude towards the experiment and devoted their full attention to it.

4.2 Memorization Performance

This analysis begins with a test of whether cognitive load was imposed successfully on all subjects. Subjects were asked whether they agreed with the following statement using a 1-5 Likert scale: “During each period I have tried to remember the number.” All subjects agreed (or fully agreed) with the statement. From this it can be assumed that neither cognitive load treatment was too boring or too difficult, and that both successfully captured each subject’s attention during the experiment. Therefore, it can be assumed that the cognitive load was imposed successfully on each subject.

Two results indicate that memorizing six digits was more difficult for subjects than memorizing one digit. First, the successful memorization rate was lower for the six-digit treatment than for the one-digit treatment. The Mann-Whitney test (a non-parametric test), was applied to test the differences in results across treatments. The corresponding null

(28)

hypothesis assumes that both samples come from the same population. The Mann-Whitney test is further explained in Appendix B. According to the Mann-Whitney test average individual memorization percentages differed significantly across treatments (z = 4.280; p = 0.0000; two-tailed). From this, it can be concluded that the high cognitive load treatment impairs memorization performance significantly.

In addition, the perceived difficulty of memorizing the numbers was significantly higher for the six-digit treatment. All subjects were asked whether they perceived the memorization task as difficult based on a 1-5 Likert scale and the difference was tested with the Mann-Whitney test (z = -4.023; p = 0.0001; two-tailed).

4.3 Degree of Inconsistency

To test the degree of inconsistency two variables were calculated per task for each subject. These variables included an estimation of the BFSP and the number of its corresponding errors. As previously explained, the BFSP assumes consistency in the answering sequence and every choice that differs from this strategy is considered an error. The BFSP was estimated in such a way that it minimized the number of its corresponding errors. If two switching points resulted in the same number of errors, the errors were set equal to this amount and the BFSP was set equal to the average of the two switching points.

First, the numbers of consistent subjects were analyzed per task for each treatment. A subject was perceived as consistent in the time task if he or she preferred the early payoff and switched only once to the future payoff or did not switch at all. A subject was considered consistent in the risk task if he or she preferred the safe option and switched only once to the risky payoff or did not switch at all. An inconsistent subject switched more than once within an answering sequence. Table 4 presents the frequencies of inconsistent subjects for each type of task and for each treatment. In the time task, the number of inconsistent subjects was low, with 5% and 15% respectively in the low and high cognitive load treatments. Moreover, according to the Mann-Whitney test the difference across treatments was insignificant (z = 1.041; p = 0.2980; two-tailed). In contrast, the number of inconsistent subjects in the risk task was higher, equaling 40% in the low cognitive load treatment7 and 70% in the high cognitive

7_{In this experiment a 40% of the respondents appeared to be inconsistent in their answers regarding risk}

taking under a low cognitive load. In other research, Lévy-Garboua, Maafi, & Masclet (2012) found an inconsistency percentage of 37.55% with real payoffs, Holt and Laury (2002) found a percentage of 13% with hypothetical payoffs, Castillo et al. (2011) found 33% with real payoffs for 13-year olds and Prasad and Salmon (2013) found 30% with real payoffs. All studies also presented lotteries sequentially. The inconsistency percentage of 40% might be higher than that of previous research since a low cognitive load is imposed on the subjects.

(29)

load treatment. This difference appears large and is marginally significant according to the Mann-Whitney test (z = 1.883; p = 0.0597; two-tailed). The results of this experiment indicate that a higher imposed cognitive load results in a trend towards more inconsistent answering sequences in the risk task.

Table 4

Frequencies of inconsistent subjects Type of Task Type of Cognitive Load Number of Subjects Number of Inconsistent Subjects Percentage of Inconsistent Subjects Time Low 20 1 5% High 20 3 15% Risk Low 20 8 40% High 20 14 70% Figure 2

Distribution of the number of errors corresponding to the BFSPs in the time task

Next, the degree of consistency was investigated to gain a deeper insight into the cognitive load’s effect on consistency. The number of errors corresponding to the BFSPs was used as a measure of the degree of consistency, which tested the first and second hypotheses.

As can be seen from Figure 2 the number of errors did not differ significantly across treatments. An Ordinary Least Squares (OLS) regression was run on the risk and time tasks separately to explain the effect of the dummy variable of the six-digit treatment on the number of errors. Several control variables were included in a second OLS regression to exclude alternative explanations deriving from any effects caused by these variables. Gender and

0 2 4 6 8 10 12 14 16 18 20 0 1 2 Fre quenc y Number of errors

(30)

education were included in the regression as dummy control variables. Age was also included as control variable. Results of the OLS regressions are presented in Table 5.

According to both OLS regressions presented in Table 5 the treatment had no significant effect on the number of errors in the time task. Therefore, the null hypothesis, that the number of errors is equal across treatments, could not be rejected. From the results in Table 5 it can also be concluded that none of the control variables (gender, age and education) significantly influenced the number of errors in the time task. This finding is in line with previous research, which did not indicate demographic characteristics as influencing the number of errors. An additional analysis was executed regarding the number of true switching points within an answering sequence. This analysis was conducted to investigate the robustness of the previous result. The number of switching points is a reliable measurement tool because this number only increases if the degree of consistency increases. The number of true switching points per subject did not differ significantly across treatments using the Mann-Whitney test (z = -0.642; p = 0.5208; two-tailed). Results of both statistical tests support the following result:

Result 1 No evidence was found that a higher imposed cognitive load led to a higher degree of inconsistency in the time task.

Figure 3

Distribution of the number of errors corresponding to the BFSPs in the risk task 0 2 4 6 8 10 12 14 0 1 2 3 4 5 Fre quenc y Number of errors

(31)

For the risk task, however, the distribution of the number of errors differs. The distribution of the number of errors corresponding to the BFSPs is presented in Figure 3. According to the results of OLS regression on number of errors, the treatment did have a significant effect. Therefore, the null hypothesis for this case could be rejected. Another interesting result is that again, the control variables did not influence the number of errors in the risk task. Again, an additional test was conducted on the number of true switching points. According to the Mann-Whitney test, the number of switching points per subject differs significantly across treatments (z = -2.702; p = 0.0069; two-tailed). Both statistical tests support the following result:

Result 2 A higher imposed cognitive load led to a higher degree of inconsistency in the risk task.

Table 5

Regression output of OLS regressions on the number of errors Dependent Variable

Errors

Risk Task (1) Risk Task (2) Time Task (1) Time Task (2)

Constant 0.4* (0.112390) -0.4705943 (1.576731) 0.05 (0.05) -0.68746 (0.442626) Six-digit treatment 0.8* (0.299122) 0.9116416* (0.344381) 0.15 (0.1272172) 0.1601541 (0.120124) Female 0.1351876 (0.345416) 0.0502545 (0.158648) Age 0.0107754 (0.025742) 0.0140303 (0.009296) Master’s degree 0.4883855 (1.094233) 0.3777552 (0.300921) Bachelor’s degree -0.1316424 (1.019747) 0.1652087 (0.335012) Associate’s degree 0.6288473 (0.955085) 0.165239 (0.305463)

Master’s degree, bachelor’s degree and associate’s degree are dummies for education, data of high school graduates are included in the constant; female is a dummy for gender; six-digit treatment is a dummy for the cognitive load treatment.

Ordered logistic regression model with robust standard errors, standard errors in parentheses. * p < 0.05