MSc in Economics
Track: Behavioral Economics and Game Theory
Master Thesis
(15 ECTS)Strengthening the case for nudging:
The influence of a good default on subsequent self-control
by
Bernardo Buarque
(11088540)
Supervisor:
Dr. Ailko van der Veen
Statement of Originality
This document is written by Student Bernardo Sousa Buarque who declares to take full re-sponsibility for the contents of this document. I declare that the text and the work presented in this document is 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.
ABSTRACT
This dissertation attempts to evaluate empirically the influence of soft-paternalism on subsequent self-control. Grounded on recent insights from neuroscience, I hypothesize that nudging might reduce the mental costs of making choices, such that it can re-allocate cognitive resources to a future self-control task and indirectly enhance performance on said exercise. I constructed, therefore, a two-stage experiment to investigate the influence of having a good default option during an initial choice task on participants’ executive functions in the later stage. In line my hypothesis, I find that the good default significantly reduces subjects’ posterior Stroop effect; moreover, a two-stage regression establishes the channel through which the intervention decreases reaction time during the first-stage and subsequently enhances performance on the Stroop task. At last, I believe the outputs from this experiment may strengthen the case for nudges by exposing potential benefits from soft-paternalism neglected by most policy-makers; namely, that nudges can improve decision-making on independent tasks.
Contents
1 Introduction 4 2 Literature Review 6 3 Methodology 9 3.1 The First-Stage . . . 9 3.2 The Second-Stage . . . 113.3 Questionnaire and Experimental Specifics . . . 14
4 Results 15 4.1 Regression Model for the Stroop Effect . . . 18
4.2 Cognitive Capacity Re-allocation . . . 21
4.3 Delayed Gratification Exercise . . . 23
4.4 Alternative Experiment . . . 26
4.5 Methodological Concerns . . . 29
5 Discussion and Conclusion 31
Appendices 37
Appendix A Tables and Figures 37
1
Introduction
In 2010, the United Kingdom devised the first Behavioural Insights Team (BIT), a social purpose company jointly owned by the British Government and whose mission is the application of behavioral sciences to “enable people to make better choices for themselves” (BIT, 2014). This event symbols a growing international tendency for policy-makers to use insights on human behavior in an attempt to increase welfare without restricting choice or significantly changing agents’ economic incentives1,2 (Thaler & Sunstein, 2008). However, as is the case with any new approach to government, the popularity of nudges as an alternative to direct mandates has become a focus of debate in the literature, with many scholars equally demonstrating enthusiasm and skepticism over the effectiveness of these policies.
Along these lines, the present research attempts to strengthen the case for soft-paternalism by empirically examining the unacknowledged influence that a good default option (most com-mon type of nudge) may have on subsequent self-control - the capacity of altering one’s responses to align them with one’s standards and goals (Baumeister et al, 2007). In light of recent devel-opments in behavioral and neural sciences, I here speculate that a nudge can reduce the mental costs of comparing the alternatives and making a decision; hence it can re-allocate cognitive ca-pacity to a future self-regulation task, and indirectly increase performance during this exercise. That is to say, by facilitating choice, soft-paternalism could positively influence subjects’ poste-rior self-control, and so this research could reveal a potential benefit from soft-paternalism that has not yet been addressed by the literature. Therefore, this investigation would contribute to the debate by empirically demonstrating an advantageous consequence of choice architecture that has not been previously considered by most policy-makers.
To test the hypothesis described above, I conducted a web-base two-stage experiment. Par-ticipants were recruited via Amazon’s Mechanical Turk and were randomly divided into treat-ment and control groups. In the first stage, they performed a series of 40 multi-attributes choice tasks with an unambiguous solution (which was hard to find). For the treatment group, the optimal alternative was always pre-selected as the default, whereas in the control group, the default option was generated at random. That is, the only difference between the two groups was the nature of the default option, while subjects in the treatment group received a good default, the control group faced a random default. The second-stage of the experiment, in turn, assessed subjects’ self-regulation through two methods - participants performed both an executive function task (Stroop test) and a delayed gratification examination. The hypothe-sis elaborated before, thus, dictates that individuals in the treatment group, who experienced a good nudge, should have more resources to spend at the second-stage, and so they should
1 Libertarian paternalism, soft-paternalism, choice architecture and nudge are all terms used in the literature to refer to how government can use lessons from behavioral science to influence choice and increase welfare, without restricting agents’ freedom to choose.
2 Figure A and B, in the appendix, exhibit trends in popularity of ‘nudge’ synonyms both on English publications and entries on Google search engine.
exhibit greater levels of self-control on this later stage.
I believe that the outputs from this experiment will boost the debate on nudges by highlight-ing two positive implications of soft-paternalism that have not yet been rigorously documented by previous literature. First, past research on choice architecture focuses almost exclusively on how nudges improve decision in one particular task and ignore how nudges could positively influence choice on an outside question, which at first glance should not be affected by the policy. More importantly, although Haan and Linde (2014) demonstrate that subjects who encountered a good default during the initial rounds of their experiment became “less critical” and were more likely to underperform when later facing the same question but with a randomly generated default, the present research complements the work done by the authors by estimat-ing if the good default intervention can nevertheless improve performance when subjects’ face an uncorrelated subsequent task. Therefore, I understand that the results of this paper could increase our understanding of the costs and benefits of choice architecture by recording the unanticipated impact of nudges on an independent exercise.
By the same token, this research could contribute to the literature by documenting to what extent do nudges mitigate the cognitive costs of making decisions, thus possibly alleviating the negative consequences of having too many options. Vohs et al (2008) remarks that having too many choices could have a detrimental effect on human well-being; indeed, the author provides a robust analysis of the mechanisms through which making many sequential choices depletes agents’ mental resources and ultimately impairs their ability to exert self-control3. Taking the
work by Vohs et al (2008) as a benchmark, the present research examines if nudges can reduce the negative influence of choice on well-being, by estimating if the good default option can hinder the ego depletion observed by the authors. In other words, based on the example set by Vohs et al (2008), I believe that estimating the influence of nudges on posterior self-control, not only allows me to measure the external effects of nudges on a future choice but also to assess if nudging can reduce the costs of making many decisions.
In summation, I believe that the results from this experiment could strengthen how we evaluate choice architecture by demonstrating that: a) a nudge can have a positive effect on a future uncorrelated task; b) nudges reduce the mental costs of making choices, and thus alleviates subjects’ cognitive constraints. Furthermore, I believe that this research could im-prove how policy-makers appraise and design nudges by stressing the importance of accounting for external effects previously unconsidered by the literature - in light of the results from this paper, for example, policy-makers could attempt to improve decision on a fundamental matter by nudging individuals on a less important task, thus reallocating cognitive capacity between the exercises and increasing welfare.
Comparing the two groups of interest, I found that participants in the control group exhib-ited a statistically larger Stroop effect - a 21 microseconds difference and a p-value calculated from a one-sided t-test equal to 0.10. In addition, a linear regression reveals that the interven-tion had a significant impact on performance during the Stroop task; namely, after controlling
3 The literature often refers to the exhaustion of this limited pool of mental resources used in the exertion of self-control as ‘Ego Depletion’.
for both participants’ enjoyment and individual features, I found that the treatment statisti-cally enhances subjects’ executive functions. Finally, a two-stage linear regression illustrates the channel through which the treatment affects posterior behavior, that is, by reducing the reaction time during the initial stage, the default intervention improved performance on the Stroop task - corroborating the hypothesis that nudging can re-allocate cognitive resources amongst tasks and thus increase self-control on a subsequent exercise.
I acknowledge, on the other hand, that the treatment had no significant influence on perfor-mance during the delayed gratification exercise. Although individuals in the treatment group demanded a much lower premium in order to accept a one month delayed payoff (they demanded approximately 100 dollars less than their counterparts), the difference between the two groups of interest were not statistically significant. In this dissertation, I shall propose alternative ex-planations for this outcome; most remarkably, I understand that time preferences offer a more stable measure of self-regulation, less prone to variations from experimental interventions, and so better suited to assess individual differences in aggregate self-control than the differences in ego depletion caused by the treatment. Likewise, I remark that, given the hypothetical nature of the examination, there was no real temptation from an immediate reward, and so the delayed gratification does not measure the exertion of self-control, such that the Stroop effect is more reliable to study the hypothesis in hand.
After this introduction, the rest of the paper is organized as follows. Section 2 presents a concise literature review regarding both soft-paternalism, and self-control. Section 3 describes the methodology adopted in the paper, while Section 4 outline the main results from the experiment. Finally, Section 4 concludes.
2
Literature Review
Thaler and Sunstein (2008) defend and illustrate how one can use lessons from behavioral science to help (nudge) individuals to make better decisions, thus promoting welfare without restricting their freedom to choose. Indeed, a growing literature on choice architecture documents, for example, how social norms can be used to foster environmental conservation (Schultz et al, 2007; Goldstein et al, 2008), how product positioning can stimulate a healthy diet (Kroese et al, 2015), and most remarkably how a good default option can promote organ donations (Johnson & Goldstein, 2004) and increase the number of workers enrolled in pension funds (Thaler & Benartzi, 2004). Nevertheless, the authors’ influential book has also been subject to divergence. Hausman and Welch (2010), for instance, postulate that while soft-paternalism does not restrict agents’ freedom of action, it does exert control over individuals’ deliberation process, and hence it violates their autonomy. Other authors, like Mitchell (2008), suggest more practical concerns regarding the ample use of nudges - the researcher notes that given the heterogeneity of preferences/behaviors, nudges that are designed to help some individuals may hurt others; that is, choice architecture could have a non-intended redistributive effect, where
‘rational’ agents will bear the costs of increasing the benefits of ‘irrational’ persons4.
In summation, despite the growing support for behavioral oriented policies, before we can promote nudges as a viable alternative to increase social welfare without restricting the ability to choose, much empirical research is required in order for us to properly understand what are the costs and benefits, what are the overall implications of these new policies. In other words, we need to understand what are the advantages/disadvantages of choice architecture in comparison with more traditional, choice limiting, policies; otherwise, as Bubb and Pildes (2014) eloquently suggest, if the benefits attained by nudges are minimal in contrast with a direct mandate, we could end up defending a policy which sole function is “to preserve an illusion of choice that has little consequences”. Hence, in this dissertation, I will attempt to foster the debate on choice architecture by exposing alternative benefits of nudges not mentioned by most scholars.
I find, in particular, that very few past studies have attempted to investigate the indirect influence that nudges may have on subsequent decisions. That is to say, whilst previous research focused on how to improve a single decision, few authors remarked how nudges could affect the process of choice in general, such that a present nudge could inhibit/enhance decision biases in a future task. Cass Sunstein (2013) acknowledges for instance that by preventing agents from making a ‘bad’ decision, choice architecture could equally prevent a learning-by-doing experience and thus it could hinder future decision-making - if nudging prevents learning, then when the subject faces a similar choice, but without the nudge, he/she is more likely to make a mistake. Likewise, Caplin and Martin (2011) empirically document that individuals being nudged become ‘spoiled’, as they tend to put less effort in subsequent tasks. Finally, Haan and Linde (2014) speculate that decision heuristics (e.g. status quo bias) might have an ‘endogenous’ component, and so a positive nudge (a good default) could reinforce the original bias and, ultimately lead to worse decisions in the future. Indeed, they performed a controlled lab experiment and observed that participants who experienced a good default were more likely to stick with the recommended option even when they faced a bad default.
We must note, however, that even the few papers that attend for the indirect influence of nudges on subsequent choices focus exclusively on how choice architecture may affect decisions in the same or very similar context. I know of no previous literature that investigates the influence of nudges on a subsequent and uncorrelated task - most importantly, I know of no previous literature that evaluates if soft-paternalism can indirectly enhance future performance by reducing the time and effort required to make a decision in the present. Along these lines, I hypothesize that while choice architecture may negatively impact future similar tasks (due to the reinforced heuristic bias), it could also have a positive influence on uncorrelated exercises, due to the redistribution of cognitive resources. In summation, the main contribution of this paper is to investigate the external influence of nudges on an independent task, and so to show how nudges may indirectly promote better decisions.
4 As an example, Mitchell (2008) argues that if a cafeteria director decides to use a nudge to help diners with low self-control to eat less, then she may need to raise her prices in order to maintain the original profit, meaning that diners with high self-control, those who already ate only a sufficient amount, may now bear some of the costs of the soft-paternalism policy.
The empirical support for the hypothesis detailed above comes from a recent literature on the neuroeconomics of attention. Krajbich et al (2010) describe that visual fixation is posi-tively correlated with decision accuracy and negaposi-tively correlated with reaction time; namely, the model proposed by the authors predicts that by directing agents’ attention towards an op-timal alternative, one can positively influence behavior and reduce the time/energy required for making the decision. Furthermore, Gottlieb et al (2014) summarize recent findings on visual selective attention and indicate that individuals learn to account for exogenous stimuli based on the rewards provided by the cue.
In light of this new field of neuroeconomics, therefore, I propose that a good nudge increases subjects’ attention towards the optimal solution and, simultaneously, reduces the computa-tional energy required to make an accurate decision. For this reason, when Haan and Linde (2004) experimentally remove their nudge, subjects continue to fixate in the default option and consequently underperformed on their experiment. However, if the nudge alleviates subjects’ cognitive constraints, then it is possible that individuals who experienced a good nudge will perform better when presented with an uncorrelated task - they will have more resources to spend on these exercises and hence will also have a greater accuracy. In sum, I imagine that, while the self-enforcing attention bias can negatively affect a subsequent/repeated decision (as observed by Haan and Linde), the use of exogenous cues and attention manipulation can reallocate time and effort among tasks, and thus externally improve decision-making during uncorrelated exercises.
For this analysis, I am particularly interested in the influence that choice architecture may have on a subsequent self-control task. An ample body of empirical evidence suggests that the exertion of self-control depends on a limited and common resource, such that self-control appears to suffer from deterioration from repeated exertions (Baumeister et al, 2007). The influence of cognitive resources on self-regulation was well described by Shiv and Federikhin (1999). The researchers observed that making subjects remember a seven digit number (in comparison with a two-digit number) takes much cognitive resource and ultimately results in less self-control. Moreover, the work by Gailliot and Baumeister (2007) evince how self-control exertion depends on energy sources, as the authors demonstrate that blood-glucose levels have a decisive role on agents’ ability to exert self-control. More importantly, at least for the scope of this analysis, Vohs et al (2008) found that subjects that made many choices among consumers goods (treatment) exhibited lower levels of self-regulation in subsequent tasks, when compared with those who were presented with a different initial task - that is to say, the authors found that making choices leads to ego depletion and reduced self-regulation. The focus of this paper, along these lines, is to observe if nudging subjects towards the optimal solution could mitigate the depletion of this common cognitive resource, and ultimately mitigate the results observed by Vohs et al (2008). In other words, I am interested in observing if subjects exposed to a positive nudge will exhibit greater levels of self-regulation in a late exercise.
Although the influence of choice architecture on self-control is interesting on itself, previous research exposes how self-regulation is positively correlated with both educational attainment and interpersonal success (Tagney et al, 2004), in this paper, I am using a well-documented
correlation between choices and ego depletion to investigate if nudges can indeed re-allocate cognitive capacity amongst tasks, and thus if nudges foster posterior decision-making on an independent task - i.e. the self-regulation task allows me to establish a connection between nudges, cognitive capacity and future decision. Furthermore, I believe that by observing the effects of the intervention on self-control, this research can complement the work done by Vohs et al (2008) and estimate if nudges can partially alleviate the perverse effects of having too many choices - as documented by the authors.
At last, to empirically study if choice architecture can improve decisions in an external exercise, I conducted an experiment to observe the impact of having a good default option in an initial choice task on a subsequent self-control task. Given past literature from neuroeconomics, I understand that a good default reduces the amount of time/energy required to make a decision in the present, thus, allowing individuals to reallocate cognitive resources to a future task. In addition, the Strength Model of Self-Control proposes that if choice architecture can mitigate the mental costs of making decisions, if it can re-allocate cognitive capacity to a future exercise, then individuals who experienced a good default should have more resources to spend on the later task, and so they should also exhibit greater levels of self-regulation. Following this logic, I here propose that by observing the effects of choice architecture on subsequent self-control, I will be able to determine if nudges can mitigate the cognitive costs of making sequential choices, if it can indeed redistribute energy amongst tasks, and finally if it can indirectly increase accuracy on uncorrelated exercises.
3
Methodology
To empirically test if choice architecture can mitigate ego depletion and enhance subsequent self-control, I conducted a two-stage experiment. In the first-stage, participants were asked to perform a set of 40 choice tasks. Here, subjects in the treatment group were presented with an optimal default option, while participants in the control group faced a random default. The only difference between treatment and control consisted on the nature of the default option, the remainder of this initial task (e.g. the number of alternatives in each round) was identical for both groups. Furthermore, the second-stage, which assessed participants’ self-control through more than one method, was also identical for both treatment and control. My hypothesis dictates, therefore, that individuals being nudged in the first part of the experiment should exhibit greater levels of self-regulation in the later stage.
3.1
The First-Stage
The design adopted for the first-stage was inspired by the work of Haan and Linde (2014). Like said paper, in each round participants had 40 seconds to choose from a table consisting of six options. Each option, in turn, was composed of six categories with different weight values. Participants were informed that his/hers selected option would generate credits equal to the sum of points in each category multiplied by the weight of that category.
Figure I illustrates the choice task faced by subjects on this first-stage. As we may observe, each column represents a different category. Moreover, these six columns/categories and their relative weights were the same for all 40 rounds. Subjects were instructed, thus, that an option would always render credits equal to: 6×points in column one + 5×points in column two + 4×points in column three + 3×points in column four + 2×points in column five - points in column six (labelled as the price).
Figure I: Example of a Choice Task
Note: Each round, participants had to select an option from the list and confirm their decision by pressing the ‘make choice’ button. Participants had 40 seconds to make a decision and received a bonus for completing the task before the limit. The credits generated by an option are equal to the sum of points in each column multiplied by the weight of that column. The default option highlighted in green (Option 4) would generate, for example: 6×38 + 5×22 + 4×28 + 3×3 + 2×30 - 291 = 216 points.
The number of points attributed to each category, in turn, was randomly generated under the condition that each option should produce between 70 and 250 points. In addition, the best option always generated at least 10 points more than the second best alternative.
On top of the credits generated by their preferred option, subjects also received a bonus for completing the choice task before the 40 seconds limit. This bonus started at 20 credits and decreased by 1 credit every two seconds a decision was not made. The bonus was relatively small in comparison with the potential gains of making an accurate decision and represented the opportunity cost of spending too much time on a single task. Following this logic, the bonus was implemented to encourage participants to select an option as soon as they reached a decision; otherwise, participants who had already decided could have waited the full 40 seconds with no additional costs.
This multi-attribute task was previously employed by Kalayci and Potter5 (2011), and it
is supposed to simulate a situation where one must choose between a set of similar products with different qualities and price. In this exercise, the category weight represents the relative importance of different characteristics, and the points depict the quality of the product in that
5 The authors experimentally analyzed the role of choice attributes on buyer confusion and market prices. The experiment adopted by them did not use a default option, as they were not interested in estimating the influence of soft-paternalism.
component. Furthermore, to make sure that subjects understood how the choice task should work, this information was given to them before the first round started.
Finally, at the start of each round, one option was preselected and highlighted at a different color. Participants were instructed that this was their default option and that, if a choice was not made, their payoff in that round would be equal to the amount of points generated by said option. Participants were also told to interpret this default as a recommendation to buy a certain product, however, no further detail specified why this default option was recommended to them. Moreover, the instructions reinforced that subjects were free to opt-out from this default at any round. As previously explained, the only difference between treatment and control was that, in the former, the best alternative (the one that produced most credits) was always preselected as the default6; whereas, in the control group the default option was
determined at random. Given the previous literature on the matter, I predict that individuals in the treatment group should observe, on average, a greater reward per round and a smaller reaction time during this first-stage.
After completing 40 rounds of this choice task, participants had to answer a few questions regarding their decision in the initial stage. In a small questionnaire, individuals were asked, among other things, if they used a specific strategy during the task, if they followed the de-fault/recommended option, and how they evaluate the task in general (e.g. was it difficult, or was it fun). These questions were later used in order to better understand participants motivations, and most importantly to control for potential unwanted biases.
3.2
The Second-Stage
The second-stage of the experiment was designed to measure participants’ levels of self-control. Subjects were informed at the beginning of the experiment that it consisted of two stages however they were not informed about the nature of this second-stage. Previous research on self-regulation reveals that individuals change their behavior to conserve mental energy when they are informed that they will need to exert self-control in the future. (Muraven et al, 2006). Hence, to prevent any undesired confounding error, to make sure that individuals would not modulate their behavior in the first-stage, I decided to not inform participants about the nature of the second-stage. Yet, as I shall propose in the discussion, it could be interesting to observe, in the future, how this information affects agents’ behavior - perhaps, if choice architecture indeed mitigates ego depletion, subjects will attempt to conserve self-control strength in the first-stage by following the recommended option.
Duckworth and Kern (2011) conducted a meta-analysis on the various methods used to as-sess self-control and concluded that: “self-control is a coherent but multi-dimensional construct best assessed using multiple methods”. Along these lines, the second-stage of this experiment
6 This is where I depart from Haan and Linde (2014). The authors were interested in observing the influence of a good default on participants’ behavior during the same exercise but with a randomly generated default. Hence, in their paper, the treatment group only had an optimal default for the first 25 round. In the remainder of the experiment, treatment and control faced the same random default. In this examination, the treatment group had a good default during all rounds, and I measured the influence of the treatment on a subsequent self-control exercise - the second stage of the experiment.
used more than one method to measure participants’ self-regulation. In particular, subjects were asked to perform an executive function task, followed by a hypothetical delayed gratifica-tion examinagratifica-tion. The initial executive funcgratifica-tion task required subjects to repress an impulsive reaction, such as not pressing a button when a stop signal appears; whereas the delayed grat-ification examination presented participants with a choice between an immediate reward or receiving their payoff with a premium somewhere in the future. Here, my hypothesis is that individuals in the control group will experience greater ego depletion in the first round, such that they will have a harder time repressing the impulsive reaction (they will underperform in the executive function task) and will more likely exhibit myopic time preferences.
There are several other alternatives that I could use to measure self-regulation. In order to test if making choices impair subsequent self-control, for example, Vohs et al (2008) used a cold pressor test7 and observed subjects’ persistence while taking an unsolvable task. However, most alternatives were impractical, relayed on methods not endorsed by experimental economics (e.g. deception), or simply were not suitable to test my hypothesis - does the good default option affects self-control. Moreover, given a moderate correlation and convergence validity across most methods (Duckworth & Kern, 2011), I understand that introducing too many tests would not necessarily provide extra information (the two methods adopted already capture different facets of self-control), while it could prolong the experiment, cause excessive fatigue, and ultimately disturb my results.
The literature provides numerous variations of the executive function task. For this analysis, I adopted the most frequent and best documented of them: the Stroop task. In this exercise, subjects are presented with color names in different print colors. Participants are then required to ignore the word meaning and respond only to the print color. For example, in Figure II.b, the participants would be required to ignore the name ‘green’ and respond to the print color ‘red’ by pressing an associated button.
Figure II: Example of a Stroop Task
(a) Congruent (b) Incongruent
Note: Subjects are required to ignore the name of the color and respond only to the print color. Panel a exhibits the name ‘green’ printed in green, it exhibits thus an example of a congruent trial, when the meaning of the word and its color match. Panel b shows the name ‘green’ printed in red, thus it illustrates a incongruent trail, the case where meaning and print are different. The conflict between word-recognition and color-recognition determines that participants will require extra processing time to respond to an incongruent trial. Difference in reaction time during congruent and incongruent trials can be used as a measure of self-control (Duckworth and Kern, 2011).
Stroop (1935) first documented that word-recognition is faster and stronger than color-recognition, such that individuals have a hard time inhibiting the more natural response that is reacting to the color name. Therefore, the concept behind the Stroop task is that participants must repress an immediate reaction, pressing the button associated with the color name, in favor of a less familiar action - respond to the print color. In other words, the executive function task measures participants’ ability to control their impulses in order to achieve a predetermined goal, that is, it measures their levels of self-regulation.
Participants were required to complete 100 repetitions of the Stroop task. During each round, the combination of color name and print was generated at random. At the end of the experiment, rounds were divided into two groups: a) Congruent, when the name and print colors matched; b) Incongruent, when the name and print colors were not compatible. The Stroop effect was then calculated as the difference between the average reaction time (in microseconds) during these two groups of repetitions (MacLoad, 1991). Using the congruent trials as a control for individual differences in reaction time, the Stroop effect measure the extra processing time required for individuals to regulate their behavior when word-recognition and color-recognition conflict; hence, the Stroop effect measures how long it takes for subjects to repress his/hers initial impulse and react to the print color. Along these lines, my hypothesis dictates that participants in the control group should have fewer resources to modulate their responses thus they should take more time to react to the print color, and so they should exhibit a greater Stroop effect - that is, the treatment should have a negative impact on the Stroop effect.
After completing the Stroop task, subjects were given two questions concerning a delayed gratification problem. In the first question, subjects were asked to choose between an immediate hypothetical payment of 100 dollars, and a hypothetical payment of $100 with a 10% premium in a month time. The second question, in turn, asked directly what would be the minimum amount (in dollars) to be received in one month time that would make them indifferent to an immediate hypothetical payment of $100. Both assessments were designed to measure partici-pants’ time preferences, and so should capture a similar effect of the good default intervention on participants’ ability to resist temptation from an immediate reward, that is, their ability to regulate their action in favor of a future-oriented goal. Therefore, if my hypothesis is correct, if individuals in the treatment group indeed are capable of exerting greater levels of self-control, then they would be more likely to wait one month for the first question, and also would require a smaller premium in the second question.
The only substantial difference between the two delayed gratification questions was the specification of the problem. Hence, as previously explained, the two inquiries should provide similar results. Nevertheless, given the relatively small premium offered in the first question, I understand that even though subjects in the treatment group were perhaps more likely to resist temptation, the small premium offered could be not sufficient to make them wait one month time, and so the first question would fail to capture participants’ differences in time preferences. Following this logic, I decided to include the second question in order to increase the accuracy of the assessment, in order to assure I would capture any differences on participants’ levels of self-regulation. Moreover, since the experiment was conducted online without the presence of an
experimenter (a problem that I shall better address in the discussion), the second inquiry allows me to test for participants’ attention and their understanding of the problem. In particular, by observing subjects consistency during both assessments, I was able to control for their comprehension of the problem and the quality of the data generated by them - for instance, three subjects selected the immediate payment on the first round, but demanded a premium smaller than $10 on the second enquiry; likewise, nine participants did the opposite, selected the delayed gratification on the first round, but demanded a premium greater than $10 on the second question.
A methodological concern regarding this delayed gratification examination is that it relies on hypothetical payments. In other words, the results from this analysis could suffer from a hypothetical bias - the tendency for hypothetical willingness to pay to overestimate real willingness to pay (Whitehead & Cherry, 2004). Along these lines, it is possible that our results misrepresent subjects’ real time preferences. Nevertheless, given that both treatment and control groups faced the same hypothetical question, and that our primary focus is not to accurately estimate subjects time preferences but to compare the ability to resist temptation between agents facing a good or random default, the hypothetical bias should not undermine the results of this experiment. Furthermore, I remark that Duckworth and Kern (2011) found a strong homogeneity amongst hypothetical and real gratification experiments, with a p-value calculated from a Cochran Q-test8 equal to 0.90 - which implies that the effect sizes estimated
by these two methods are not statistically different.
3.3
Questionnaire and Experimental Specifics
Once subjects completed both sections of the experiment, they were required to fill in a brief and final questionnaire. Among other questions, participants performed a 10-Item Self-Scoring Self-Control Scale, where they had to indicate how much they identify themselves with state-ments like: “I have a hard time breaking bad habits” (Tangney et al, 2004). Self-reported questionnaires are seldom used for assessing ego depletion since they do not require the ‘per-formance’ of self-control, that is, they do not require participants to inhibit their impulses in favor of a goal-oriented action. Nevertheless, the 10-Item Scale questionnaire provides a reliable control for participants’ differences in personality. Hagger et al (2010) summarize that there are two theories of self-control: the strength model prospects that self-control exertion depends on a limited resource; whereas, the capacity-based theory “conceptualizes self-control as a dis-positional, trait-like construct that differs across individuals”. Furthermore, they argument that the two theories are not mutually exclusively; in fact, one could suggest that “people high in dispositional self-control will have more resources at their disposal”, and so they will be less sensible to ego depletion. Along these lines, I understand that individual traits/personality could partially determine subjects’ performance during both the Stroop task and the delayed gratification examination; for this reason, I use participants’ responses to the 10-Item Scale as a
8 A classical measure of heterogeneity, which is calculated as the weighted sum of squared differences between individual study effects and the pooled effect across studies.
control for differences in ‘dispositional’ self-control - to assure that both groups were comparable and that any observed results were due to ego depletion and not differences in personality9.
By the same token, the questionnaire assessed subjects’ socioeconomic characteristics, which were later used for control proposes. Among others, participants were asked about their age, gender, profession, and nationality. In addition, given the growing body of evidence that correlates cognitive reasoning and decision bias (Oeachssler et al, 2009), I also included a cognitive reflection task10 (Fredericks, 2005) in the questionnaire. Finally, for similar reasons,
the questionnaire also contained a risk aversion assessment - where subjects had to choose between a hypothetical and certain payment of $10.00 and a hypothetical lottery that payed $20.00 with probability of 75% and zero otherwise. These two variables should allow me to control for individual differences in skills, and cognitive capacity.
The experiment was computerized with PHP/MySQL and conducted online. Subjects were recruited for this web-based experiment using Amazon’s Mechanical Turk (MTurk) webpage. Originally created in 2005, the crowdsourcing labor market has recently gained popularity among social scientists as a source for inexpensive experimental data. Despite the concern regarding the quality of the data generated by this new research instrument, recent studies indicate that the population on MTurk is as representative of the American population as traditional subject pools, and most importantly that the data obtained from this market are at least as reliable as those gathered via traditional methods (Paolacci et al, 2010; Buhrmester et al, 2011). In sum, I understand that MTurk provides this experiment with a larger sample, a more diverse sample, and a reliable dataset.
The online assessment was carried between the 07th and 08th of July 2016. Subjects took on average 40 minutes to complete the experiment and were paid approximately 1.50 dollars per hour worked - value above the median reservation wage of $1.38/hour (Horton & Chilton, 2011). In total, 78 participants completed the experiment, of which 38 were part of the treatment group and 40 were part of the control. Approximately 54% of the subjects were male, and the average age was 33 years. Nearly 73% of subjects were Americans, while 16% was from India. More than 3/4 of all subjects had completed at least an undergraduate degree, of which 20 participants reported finishing some level of post-graduate studies.
4
Results
In accordance with the hypothesis detailed in the literature review, subjects in the treatment group, that experienced a positive nudge, made better and faster choices during the first-stage
9 Ideally, the questionnaire should take place before the experiment. Nevertheless, given the results from Muraven (2006) I feared that by having the questionnaire before the first-stage, I would frame participants to conserve mental energy in the first-stage - subjects could anticipate they would be required to exert self-control during the experiment, and so they would modulate their behavior. Finally, since self-reported questionnaires are not used to measure ego depletion (Hagger et al, 2010) and given the higher convergence validity of questionnaires as a measure of self-control (Duckworth & Kern, 2011), I believe that the good default intervention should not affect participants’ answers in the survey, and so it provides me with a proxy for participants’ personalities.
of the experiment. Indeed, as we may conclude from Figure III, individuals in the treatment group took on average 12.8 seconds to make a decision, whilst subjects in the control only reached a decision after approximately 16 seconds. Finally, a Wilcoxon rank-sum test reveals that the perceived differences between treatment and control are statistically significant.
Figure III: Average Reward and Reaction Time on First-Stage
p−value = 0.03 0 50 100 150 200 CONTROL TREATMENT
(a) Average reward per round
p−value = 0.00 0 5 10 15 CONTROL TREATMENT
(b) Average reaction time per round
.
Note: The vertical axis measures the reward (Panel a) and reaction time (Panel b) per round. Reward is measured in credits, while time is measured in seconds. The bars indicate the average values among control and treatment. P-values refer to a Wilcoxon-Mann-Whitney test. Null hypothesis states that subjects on control and treatment groups exhibited the same reward and reaction time per round. With these p-values we may reject the null hypothesis.
Data from the first-stage demonstrate that a positive nudge can reduce the time required to compare the alternatives and make an accurate decision; hence, it indicates that soft-paternalism could alleviate the cognitive costs associated with making choices. However, the primary focus of this research is to observe if choice architecture, by reducing the reaction time during the first-stage, can redistribute resources amongst tasks, such that individuals who experienced a good nudge can exert greater levels of self-control in a subsequent exercise. In other words, although the results exposed above help to corroborate my original hypothesis, to estimate if soft-paternalism can indeed mitigate the mental costs of making many choices and thus enhance performance on an independent task, I am mostly interested in observing differences between treatment and control on the second-stage of the experiment.
Before we can attribute any potential outcomes to the treatment, however, we must con-firm if the randomization process was efficient, that is, we must concon-firm that treatment and
control groups are indeed comparable. Along these lines, Table I presents a summary statistics for all key control variables. As we may observe, there are no significant differences between treatment and control in terms of gender, nationality, or education. Additionally, both groups reported similar levels of self-control in the 10-Item scale and performed just as well on the cognitive reflection test. Finally, I understand that there are no significant differences between the two groups in terms of risk aversion. This result is important because, although risk and time preferences are apparently correlated (Anderhub et al, 2003), risk aversion does not seem to be affected by ego depletion (Gerhardt et al, 2015) and thus should not depend on the treatment. In sum, Table I suggests that participants in treatment and control groups exhibit similar socioeconomic characteristics, and so we are led to conclude that any observed results can be attributed to the treatment (Angriest & Pishke, 2009).
Table I: Summary Statistics of Control Variables
Variables Control Treatment Difference p-value
Mean s.d. Mean s.d. Male 0.50 0.08 0.58 0.08 -0.08 0.49 Age 32.05 1.44 31.10 1.33 0.94 0.63 Years of Education 16.57 0.54 16.68 0.67 -0.11 0.89 American 0.70 0.07 0.76 0.07 -0.06 0.53 Cognitive Reflection 1.70 0.18 1.81 0.18 -0.11 0.65 Self-Control 3.43 0.12 3.31 0.13 0.13 0.47 Risk Aversion 0.80 0.06 0.84 0.06 -0.04 0.63 N.OBS 40 40 38 38 78 78
Note: Male is a binary variable that is equal to one if the participant is male and zero other-wise. American is also a dummy variable which corresponds to one if the subject is from the USA. Cognitive reflection measures how many correct answers the subject got on the Fred-ericks (2005) test. Self-Control is an scale constructed with subjects’ answer to the 10-Item questionnaire, which goes from 1 to 5, and where 5 corresponds to not at all self-controlled and 1 is equal to very self-controlled. Finally, risk aversion measures the probability that an individual will choose a $10 guaranteed payment over a risky lottery with a greater ex-pected return ($15). At last, p-values refers to a two-sided t-test, where the null hypothesis dictates that the observed difference between the two groups are identical to zero. We may reject the null hypothesis if the p-value is smaller or equal to 0.10.
Once we assured that the randomization was efficient, we may start our analysis on the second-stage of the experiment. Comparing the two groups of interest, we observe that in-dividuals in the treatment group exhibited, in line with my hypothesis, a smaller reaction time during both congruent and incongruent trials. Furthermore, we note that subjects under the treatment also exhibited a lower Stroop effect - 110.82 microseconds in comparison with the 122.04 microseconds observed by the control. At last, for a more precise description of the results, I conducted a robust statistical analysis of the data to investigate if the observed differences between treatment and control are indeed significant.
Table II displays the outputs from this statistical investigation. It suggests that individuals who experienced a good default could be outperforming their peers in a subsequent self-control
task. In particular, I remark that the estimations from a one-sided t-test inform that partici-pants in the control group exhibited a statistically greater Stroop effect (p-value=0.10), that is, they exhibited a statistically significant lower self-regulation. However, given the results from a Wilcoxon-Mann-Whitney test, one could suggest that the observed differences in performance between the two groups were actually insignificant; in other words, one could affirm that the results from Table II are not robust enough for us to completely reject the null hypothesis that a positive nudge had no impact on the subsequent exercise.
Table II: Statistical Analysis of the Stroop Task
Variables Control Treatment T-test Rank-sum
Mean s.d. Mean s.d. (p-value) (p-value)
Congruent 969.52 31.12 868.28 19.87 0.00 0.01
Incongruent 847.48 27.91 767.45 19.82 0.01 0.03
Stroop Effect 122.04 12.36 100.83 10.66 0.10 0.35
N.OBS 40 40 38 38 78 78
Note: Variables are measured in microseconds. T-test refers to a one-sided test that mea-sures if the reaction time observed by the control group was greater than in the treatment group. Rank-sum refers to a Wilcoxon-Mann-Whitney test, which the null hypothesis was that both groups exhibited the same reaction time. We may reject the null hypothesis if the p-value is smaller or equal to 0.10.
The main distinction between the two tests exposed before is that the Wilcoxon rank-sum examination does not assume the dependent variable (Stroop effect) is normally distributed, which could imply that said test produces more robust/reliable statistics. However, before we dismiss the results from the one-sided t-test in favor of the Wilcoxon-Mann-Whitney, we must observe that both Stroop (1934) original experiment, MacLoad (1991) review of the many replications of the Stroop task, and the data collected from my experiment (Figure D in the Appendix) suggest that the Stroop effect follows a bell-shaped distribution. In addition, I understand that the onesided ttest more accurately capture the hypothesis of this paper -individuals in the treatment group should exhibit a smaller Stroop effect.
In summation, I understand that, even though the results from Table II are not particularly robust, the one-sided t-test is sufficient to indicate that nudges could have a positive effect on participants’ posterior self-control, and so we should not dismiss the hypothesis elaborated in the literature yet. Moreover, to better understand the effects of nudges on subsequent self-control, to gather more evidences favoring/refuting my hypothesis, I decided to perform a parametrized regression, and also to evaluate the effects of the intervention on the hypothetical delayed gratification exercise.
4.1
Regression Model for the Stroop Effect
To further evaluate the effects of a good default option on subjects’ posterior self-control, I adopted a parameterized regression model. Here, I am interested in estimating the impact of
the treatment, a binary variable that corresponds to one if the subject faced a good default in the initial stage and zero otherwise, on the Stroop effect. In addition, given past evidence from the literature, to increase both the robustness and the predictive power of this model, I decided to include controls for participants’ enjoyment and their socioeconomic attributes.
Vohs et al (2008) propose and demonstrate empirically that, depending on the amount of choices made in sequence, subjects’ enjoyment of the task could have an influence on ego depletion. By the same token, one could propose that subjects facing a random default (control group) may find the initial task more demanding, hence resulting in a lower enjoyment11. If
that is to be the case, the observed differences between treatment and control could be partially attributed to different levels of enjoyment, and not to a redistribution of cognitive capacity -that is, the results could be biased. Along these lines, to add robustness to the regression, I constructed a parametric control that captures individuals’ enjoyment during the first-stage. Participants were asked to grade, on a scale from 1 to 7, how interesting, difficult, and fun was the initial choice task. Using this information I constructed an index variable called Enjoyment, which was later added to the regression model - this index was calculated as the average score on the ‘fun’ and ‘interesting’ scales, and the inverse score on the ‘difficult’ scale.
MacLoad (1991), in turn, provides and ample literature review on the Stroop task, and remarks that socioeconomic characteristics such as gender, age, literacy and language could explain individual differences in the Stroop effect. Therefore, even though the randomization was successful, I opted to include additional socioeconomic controls in order to enhance the precision of the model.
In sum, we may write the complete parametrized model as:
Stroopi = β0+ β1T reatmenti+ β2Enjoymenti+ ΘXi+ i (1)
where Stroopitrepresents our dependent variable, which measures in microseconds the difference
between the average reaction time during congruent and incongruent trials; T reatmenti is a
binary variable that is equal to 1 if the subject ‘i’ experienced a good default and zero otherwise; Enjoymenti is an index variable ranging from 1 to 7, and which is equal to 7 if participant ‘i’
found the task very enjoyable; Xi corresponds to a vector of socioeconomic control variables;
β0 is a constant intercept parameter; and i is considered to be a disturbance term.
Table III presents the estimations for an OLS regression of the model. Column one exhibits the results for the simple linear regression with no control; column two controls for participants’ enjoyment, and column three controls for both enjoyment and socioeconomic characteristics. All three columns include robust standard errors.
In line with the results from the initial statistical analysis, Table III reveals that the good default intervention decreased participants’ Stroop effect - that is, the positive nudge enhanced participants’ performance on the self-regulation task. More importantly, although I initially found no statistically significant effect, once I controlled for both enjoyment and socioeconomic
11 Indeed, treatment and enjoyment are positively correlated (corr = 0.1842), thus suggesting that partici-pants’ in the treatment group found the choice task more enjoyable.
Table III: Regression Analysis of the Stroop Task Variable (1) (2) (3) Treatment -21.20 -25.14 -33.46* (16.32) (17.19) (17.54) Enjoyment 7.50 8.72 (6.02) (6.06) Male 46.08*** (15.79) Age 1.10 (0.82) Years of Ed. -1.35 (1.96) American 68.88** (29.31) English -23.14 (26.70) Constant 122.03*** 90.32*** 18.99 (12.37) (26.39) (49.59) R-sq 0.0215 0.0427 0.2362 N.OBS 78 78 78
Note: The table presents coefficients from OLS regressions, where the dependent variable is the Stroop effect (measured in microseconds). All re-gressions use robust standard errors. Treatment is a binary variable that is equal to one if the partic-ipant receive a positive nudge, and zero otherwise. Enjoyment is an index variable ranging from 1 to 7, and which is equal to 7 if the participant found the choice task very enjoyable. Constant refers to a intercept parameter. ∗∗∗ p<0.10,∗∗ p<0.05 and ∗ p<0.01.
features, the results from the linear regression model are statistically significant. In other words, when I control for individuals’ characteristics, I find evidence that a good default significantly reduces the size of the Stroop effect among subjects; thus, indicating that a positive nudge can alleviate the ego depletion associated with making choices.
The output from column three (Table III) can be explained by the growth in the precision of the model. Past literature documents how the Stroop effect covaries with many socioeconomic characteristics; hence, when I parametrically controlled for individuals’ traits, I reduced the residual variance of the estimations and increased the adjustment of the model - indeed, as we can note by contrasting the results in columns one and three, the R-square of the model goes from 0.0215 to 0.236212. Therefore, the outputs from the complete model depicted in
column three provide a more robust estimate of the impact of the treatment on participants’
12 The Adjusted R-square of the model goes from 0.0086 to 0.1598.The Adjusted R-square is a more robust statistic since it compares the explanatory power of regression with different numbers of parameters.In particular, after the introduction of a new parameter, the Adjusted R-Squared only increases if the new term improves the model more than what would be expected by chance.
posterior self-control, and so it provides empirical evidence favoring the hypothesis that choice architecture can foster decision on an independent task.
In summation, a one-sided t-test reveals that participants in the control group exhibited a statistically greater Stroop effect; furthermore, following the indication from previous liter-ature, when I controlled for enjoyment and socioeconomic characteristics, I reduced the noise in my estimations and found a statistically significant impact of the treatment on participants’ executive functions. Last but not least, the results from Table I suggests that individuals in the two groups of interest possess similar characteristics, in particular, they exhibit similar levels of ‘dispositional’ self-control and cognitive ability; hence, it indicates that the observed outputs were indeed caused by differences in ego depletion. In other words, the results from the regression analysis lead me to propose that the good default intervention had an impact on subjects’ executive functions, and so that a good nudge can improve decisions on a subsequent uncorrelated task.
4.2
Cognitive Capacity Re-allocation
Thus far, we observed that a positive nudge reduces the time required for participants’ to make an accurate decision, and also estimated that choice architecture may enhance subsequent self-control; yet, we still don’t have enough evidence favoring the mechanism through which, by reducing the cognitive costs of making choices, a positive nudge can promote better decisions in a posterior uncorrelated task. In other words, the primary hypothesis of this paper dictates that our treatment should affect self-control by reallocating resources amongst tasks; thus, to understand the process through which a good default option affects subsequent self-control, I constructed a two-stage regression model to measure the impact of a reduced reaction time in the choice task (used here as a proxy for cognitive capacity consumption) on the Stroop effect.
Mathematically, we may represent this new parametrized model as:
T imei = α0+ α1T reatmenti+ i
Stroopi = γ0+ γ1T imei+ γ2Enjoymenti+ ΘXi + µi
(2)
where the first-stage measures the impact of our treatment on subjects reaction time during the initial choice task (measured in seconds), and the second stage calculates the effect of the estimated reduction in reaction time on the variable of interest (Stroop task). All the control variables are the same as our original model, and we can calculate the parameter of interest via a two-stage least squares (2SLS), such that:
β2SLS = γ1 α1
(3)
Table IV exhibits the result of the two-stage least squares. Analogous to Table III, column one presents the estimations with no parametric control; while column two controls for enjoy-ment and column three controls for both enjoyenjoy-ment and socioeconomic characteristics. Once
more, all columns use robust standard errors.
Table IV: Two-Stage (2SLS) Analysis of the Stroop Task
Variable (1) (2) (3) Time 6.64 6.08 8.71* (5.85) (4.65) (5.29) Enjoyment -3.36 -4.27 (8.35) (7.633) Male 40.19** (19.26) Age -0.24 (1.47) Years of Ed. -2.30 (2.12) American 83.24 (56.64) English -12.33 (50.49) Constant 15.92 38.95 -22.24 (84.93) (49.35) (53.41) N.OBS 78 78 78
Note: The table presents coefficients from 2SLS re-gressions. The dependent variable is the Stroop ef-fect (in microseconds). Time measure individuals reaction time during the first-stage of the experi-ment, and it was instrumented by the good default intervention. Therefore, time measures the impact of a reduced reaction time (in seconds) on the de-pendent variable. Constant refers to an intercept parameter. All regressions include robust standard errors. ∗∗∗ p<0.10,∗∗ p<0.05 and∗ p<0.01.
The two-stage model provides a powerful interpretation of the mechanisms behind my main hypothesis. The statistically significant estimator from column three suggests that, when I control for both enjoyment and subjects’ characteristics, the good default option enhances participants’ self-control (it has a negative impact on the Stroop effect) by reducing the average reaction time during the first part of the experiment. I find, therefore, that a positive nudge can foster self-control by reallocating resources amongst tasks. That is to say, the two-stage model corroborates with the hypothesis developed during the literature review - i.e. choice architecture mitigate the mental costs of making choices, it re-allocates resources to an independent task, and ultimately it enhances performance during this separate question. Furthermore, the outputs from this two-stage model strengthen the hypothesis that choice architecture can partially prevent ego depletion, and so it indicates that choice architecture can alleviate the negative consequences of making many choices observed by Vohs et al (2008).
At last, given the outputs from the one-sided t-test, the linear regression model, and the two-stage least squares, I understand that this paper provides robust evidence that choice
architecture can both mitigate the cots assosiated with making choices, and improve decision making in a future exercise.
4.3
Delayed Gratification Exercise
We may continue our investigation on the effects of nudges on posterior self-control by re-producing the analysis conducted so far on the hypothetical delayed gratification assessment. Comparing treatment and control, we observe that although the proportion of individuals in the control group that accepted the delayed payment with a 10% premium in the first question was surprisingly larger than in the treatment group (a difference of eight percentage points), subjects who experienced a positive nudge demanded a much lower premium in order to accept a one month delayed reward in the second question - they demanded on average a premium of 64.50 dollars, while their counterparts asked for a $167.70 premium. It is important to acknowledge, however, that the observed differences between our two groups of interest were not statistically significant; that is, the p-values calculated from both a one-sided t-test and a Wilcoxon-Mann-Whitney test were not sufficient for us to refute the null hypothesis that participants in the two groups exhibited similar time preferences13.
For a more robust analysis of the problem in hand, I adjusted the regression model described by Equation 1, to estimate the effect of a good default intervention on two new variables of interest: Delayed and Premium. In other words, I first replaced the dependent variable in Equation 1 (Stroopi) with a binary variable that is equal to one if the participant selected the
delayed gratification in the first round of the assessment (Delayedi), and later I replaced it again
for a variable that measures the premium required to make the subject indifferent between an immediate payment of $100 and a one month delayed reward (Premiumi).
Table V displays the outputs from this parametric investigation. Panel A describes the results for the complete sample, whereas Panel B presents the results containing only consistent answers. As before, column one exhibits the results calculated with no control, while column two controls for enjoyment and column three controls for both enjoyment and socioeconomic characteristics. At last, once more all estimations used robust standard errors14.
The calculations from this parametrized test corroborate with the hypothesis that a good nudge had no significant impact on participants’ decisions during the delayed gratification assessment. Indeed, the outputs from Table V reveal that our treatment had an insignificant impact on both the probability of choosing the delayed gratification in the first question, and the size of the premium demanded on the second inquiry.
In contrast with the results obtained from our analysis on the Stroop effect, therefore, I found no statistical evidence that the intervention had an impact on participants’ time preferences. There are several alternative explanations for this outcome. First, the hypothetical nature of the assessment and the lack of an experimenter allows for reduced attention and misunderstanding of the problem at hand, which could undermine the quality of the data generated by the later
13 Table A, in the Appendix, shows a descriptive statistics for the results presented above.
Table V: Regression Analysis of the Delayed Preferences Variables (1) (2) (3) A. Complete Sample Delayed -0.08 -0.10 -0.8 (0.11) (0.11) (0.10) Premium -103.06 -79.43 -96.61 (124.22) (113.69) (136.69) N.OBS 78 78 78 B. Reduced Sample Delayed -0.07 -0.09 -0.08 (0.12) (0.12) (0.11) Premium -116.58 -92.48 -97.76 (150.31) (152.06) (156.55)
Constant YES YES YES
Enjoyment NO YES YES
Individual traits NO NO YES
N. OBS 66 66 66
Note: The table presents coefficient from OLS regressions on the influence of a positive nudge on participants’ time preferences. Panel A shows the results for the complete sample set, while Panel B exhibits the results of the re-duced sample, containing only consistent answers. ‘Vari-able’ depicts the dependent variable, where Delayed refers to a binary variable that is equal to one if the participant selected the delayed gratification in the first round of the assessment, and Premium measures the value required by participants’ in order to make them indifferent between an immediate payment of $100 and a one month delayed re-ward. Column 1 presents the estimations with no paramet-ric control. Column 2 controls for enjoyment, and column 3 controls for both enjoyment and socioeconomic charac-teristics. All estimations include an intercept parameter, and all estimations use robust standard errors. ∗ p<0.10, ∗∗ p<0.05 and∗ p<0.01.
method. This concern is evident once we consider the number of inconsistent answers and the somewhat counterintuitive outputs from this section - participants in the control group were more likely to accept the 10% premium, while simultaneously demanding a larger premium in the second inquiry. However, I remark that even when I control for participants’ attention (by dropping from the sample all inconsistent answers), the estimations remain statistically insignificant (Table V.B) - thus, suggesting that misunderstanding and lack of motivation are not a substantial explanation for the observed outputs.
Moreover, as previously stated in the methodology, self-regulation is a multidimensional concept and the two assessments used in this paper were designed to capture different spheres of this complex notion. Hence, given the results in hand, one could propose that while choice architecture can enhance agents’ ability to repress an automatic/habitual response, it is not sufficient to increase their capacity to resist temptation from an immediate reward. Growing evidence from neuroscience suggests that different mental processes are responsible for
execu-tive function and time preferences - execuexecu-tive function usually aggregates different processes responsible for attention, response inhibition, task switching and information updating (Miyake et al, 2000), whereas time preferences are more similar to simple choices, in which the brain computes and compares noisy evidence favoring rewards at different points in time (Fehr & Rangel, 2007; Kable & Glimcher, 2007). Along these lines, continuing our analogy on cognitive resources, one could suggest that the amount of ‘energy’ required to collect evidence favoring a future reward and thus make a self-controlled decision is far greater than the quantity of re-sources used for inhibition, such that the extra cognitive resource provided by the good nudge is sufficient to increase executive functions but not enough to influence time preferences. In other words, one could suggest that choice architecture efficiently reduces the cognitive costs of mak-ing choices, however, the gains from a positive nudge are limited, and so choice architecture does not influence all tasks that demand self-control alike. Nevertheless, previous literature suggests that ego depletion effect should be consistent across all spheres of self-control (Hagger et al, 2010); hence, once we consider past investigations on the matter, this explanation seems less credible.
It is possible, on the other hand, that choice architecture influences aggregate self-control but the delayed gratification task was not suitable to measure the differences in ego depletion caused by the intervention. Duckworth and Kern (2011) estimate that 53% of the overall variance in the self-control effect size can be explained by the method adopted; likewise, the authors propose that delayed gratification assessments have a stronger convergence validity than executive function tests, and they also exhibit a stronger correlation with informant-report questionnaires. Hence, one could posit that delayed gratification is a more robust/stable measure of self-control, less prove to fluctuations in response to experimental manipulations, and so an ideal method to measure individual differences in self-regulation but not the effect of ego depletion. In other words, it is possible that the delayed gratification was not suitable for the experiment in hand, as it does not accurately capture the effects of the treatment.
To measure ego depletion, the second exercise of the two-stage experiment must require participants to exert self-control; that is to say, the accuracy of the second-stage relies on the “requirement of the effortful suppression of an impulse” (Hagger et al, 2010). Following this logic, the delayed gratification assessment can only capture differences in ego depletion if subjects’ are required to effortfully resist the temptation from an immediate reward in favor of a future-oriented goal. However, given the hypothetical nature of the assessment and the distance between participants and this tempting immediate reward, one could affirm that there was no real temptation, and so there was no need for impulse suppression. In essence, like the 10-Item questionnaire, perhaps the delayed gratification did not require participants’ to exert self-control, and so it cannot measure ego depletion but only individual differences in ‘dispositional’ self-control. Therefore, although the hypothetical delayed gratification is a robust measure of agents’ self-control (Duckworth & Kern, 2011), it may not be an accurate measure of ego depletion, and thus it is not adequate for the objective of this experiment.
In summation, to strengthen our understanding of the influence of a good default on sub-sequent self-control, I decided to include a delayed gratification assessment as an alternative
