Study on Willpower and Study Patterns: Weekly and Daily Differences, Essays (high school) of English

Download Study on Willpower and Study Patterns: Weekly and Daily Differences and more Essays (high school) English in PDF only on Docsity! ACADEMIA Accelerating the world's research. Three field experiments on procrastination and willpower Gary Charness Related papers Download a PDF Pack of the best related papers 7 Field and Online Experiments on Procrastination and Willpower Gary Charness Field and online experiments on self-control Gary Charness Overcoming Procrastination through Planning Frank Wieber, Peter Gollwitzer Three Field Experiments on Procrastination and Willpower Nicholas Burger, Gary Charness, and John Lynham* January 8, 2009 Abstract: The issues of self-control, procrastination, and willpower have been the focus of a number of recent theoretical models. Yet there is very little clean data about behavior under financial incentives, particularly on tasks of significant duration. We therefore conducted three field experiments to investigate how people schedule and complete tasks, providing some of the first data in this area. In each of these studies, we also attempted a behavioral intervention designed to improve performance. We find clear evidence of procrastination and willpower depletion, as well as behavior suggestive of self-reputation considerations. In this respect, each of the behavioral models identifies important aspects of how people deal with tasks, yet none of the models correctly predicts behavior in all of our studies. The behavioral interventions we used led to outcomes that surprised us in all three studies, although these outcomes are largely consistent with the standard neo-classical model. We hope that our results help spark more empirical research and more theoretical work in this nascent area. Keywords: Field experiment, Incentives, Procrastination, Studying, Willpower JEL Codes: A13, A22, B49, C93, DO * Contact: Nicholas Burger, Rand Corporation, nburger@rand.org, Gary Chamness, Dept. of Economics, University of California at Santa Barbara, charness@econ.ucsb.edu, John Lynham, Dept. of Economics, University of Hawai’i at Manoa, lynham@hawaii.edu although the direction of the effect varies. The duration of the task appears to matter, as it appears that females are better at managing their time over a longer campaign, while males do better when the task is shorter and the finish line is in view. In terms of previous work, there have been a handful of studies that consider how one might overcome self-control problems. Aside from exerting willpower in the face of a disagreeable task, one approach is to bind one’s own behavior with costly restrictions. Wertenbroch (1998) presents anecdotal examples of binding behavior, including tactics such as putting savings into a Christmas-club account that does not pay interest or buying small packages of goods such as cigarettes or ice cream.’ Schelling (1992) mentions reforming drug addicts who send out self-incriminating letters, to be divulged in the case of a relapse into drug use. Another device is to set deadlines for one’s self, for example, many a researcher has agreed to present a yet-unwritten paper in the future, in the hopes that the embarrassment of being forced to either cancel or be unprepared will be a strong motivation for writing the paper prior to the presentation. In fact, many activities seem deadline-driven, particularly in our contemporary society in which people seem to be short on time. Ariely and Wertenbroch (2002 present two studies in which three tasks had to be completed over a period of time, finding that externally-imposed costly deadlines during this period are more effective than self-imposed (and binding) costly deadlines, which in turn are more effective than having no additional deadlines.* Burger and Lynham (2007) examine weight-loss bets in England, where one could bet on achieving a weight goal by a deadline; however, the vast majority of bettors lost their bets with the agency. * See Ariely and Wertenbroch (2002) for a more complete literature review. “The tasks involved tuming in term papers for grades in a class for a 14-week period (with a one percent grade penalty for each day of delay) or proofreading three papers totaling about 100 pages over a 3-week period; in the latter task, participants were paid $0.10 to find errors (there were about 100 embedded), but fined $1.00 for each day of delay. In our first study, we paid students $95 to complete 75 hours of studying at a monitored location in the campus library over a five-week period. In one treatment, participants were tequired to complete at least 12 hours during the first week, at least 24 hours by the end of the second week, etc., while there were no interim requirements in the second treatment. We expected people to procrastinate with their timing, leaving the bulk of the required studying until the end; in line with the results in Ariely and Wertenbroch (2002), we expected that extemally- imposed costly deadlines would be effective, so that the group with the weekly studying requirement would be more likely to complete the task.° However, completion rates were actually 50 percent higher with no interim requirements, the direction that would be predicted by a standard neoclassical model, since relaxing a constraint should increase or at least not affect the likelihood of success. The patterns of study time show a pronounced weekly cycle, even in the no-weekly-requirement treatment, with little difference in the aggregate from week to week. However, individual analysis reveals substantial heterogeneity, with some people logging the bulk of the hours in the first weeks and some other people doing so in the last weeks. We find evidence that, over time, students who achieve the studying goal improve their performance in the course relative to those students who did not. Having observed different behavior on weekends and weekdays, we designed a second study in which the task consisted of answering different numbers of multiple-choice questions (the order was endogenous) on each of seven consecutive days; seven groups started on different days of the week. As a behavioral intervention, we asked people to designate in advance their plans for task completion and then observed their actual behavior. We do see evidence of procrastination among even the most disciplined people in the population, as the average number 5 Fischer (2001) also presents arguments for breaking a task into smaller components. On p. 261, she states: “Therefore, the best way for a supervisor to ...reduce the risk of missing the ultimate deadline may be to break it into smaller tasks with more deadlines to better compete with the other demands on the student’s time.” of questions answered for people who succeed at the task increases steadily over the course of the seven days. Plans do not differ for those who succeed and those who do not, as people who stick with their plan are no more likely to complete the task than people who do not. For our third study, we reduced the duration of the task (answering 20 multiple-choice questions) to two consecutive days and allowed the participants to complete the entire task in one sitting, if desired. As a behavioral intervention, we also required people to complete either an easy or difficult Stroop test on the first day (Stroop 1935). The difficult version features cognitively-discordant tasks, and is considered to be willpower-depleting.® Those people assigned the difficult Stroop answered significantly fewer questions on the first day, suggesting willpower depletion. However, we were surprised that people who were assigned the difficult Stroop were actually more likely to finish during the allotted two days than those people assigned the easy Stroop. There is also evidence of lower effort on the second day, as the percentage of correct answers was significantly lower than on the first day. The remainder of this paper is organized as follows: We provide details of our experimental design in section 2, and we present some theoretical models and their predictions in section 3. We describe our experimental results in section 4, offer some discussion in section 5, and conclude in section 6. 2. THE FIELD EXPERIMENTS Study 1 Our experiment was conducted at the University of California at Santa Barbara. We obtained permission to have anonymous access to the records of students in a large introductory class and then recruited as many as possible from this class. We then advertised the session to © We thank Emre Ozdenoren, Stephen Salant and Dan Silverman for suggesting many of these design changes. Study 2 For our second study, we recruited 181 participants from a micro principles class and two intermediate micro classes at UCSB and described the task in general terms, as well as the prize. The task involved answering multiple-choice questions (drawn from test banks and previous exams in these classes) that were closely related to the material in their class. Each participant was asked to complete seven sub-tasks on seven consecutive days. One sub-task consisted of answering two multiple-choice questions online; a second sub-task consisted of answering four multiple-choice questions online, etc.’? As in Study 1, we felt that incentives (in this case, to do well in the specific class) were already in place for the participant and so did not pay the usual average per-hour rates for experiments. The prize consisted of one state-of-the-art iPod (which cost $400) for every 50 participants; we awarded four prizes, with the winners determined by drawing entries in a lottery. We gave more weight to correct answers, as otherwise people would be tempted to choose answers randomly. Each participant received one lottery entry for submitting a non-binding plan for the order in which they would complete the sub-tasks starting a week or so later, two entries for completing the seven sub-tasks on consecutive days, and four entries for answering at least 75 percent of the questions correctly. In order to investigate the day-of-the-week effects we found in Study 1, we assigned each participant randomly to one of seven groups; one of these groups started on a Monday, one group started on a Tuesday, etc. " Clearly there are advantages and disadvantages to online experiments. For a detailed discussion, see section 2.2 of Charness, Hanuvy, and Sonsino (2007). Study 3 Our third study was also conducted online with multiple-choice economics questions. We recruited participants from a micro principles class and two intermediate micro classes at UCSB; we supplemented the 134 people from these classes who signed up with 23 economics or business economics majors who were not in these classes, but who were in the campus-wide subject pool. In all cases, we described the task in general terms and the payment scheme. In Study 3, the task consisted of answering 20 multiple-choice questions similar to the questions in Study 2 (no one participated in more than one of our three studies) over a two-day period. The participant could answer all 20 of these on the first day or spread them over the two days. We paid each person $7.50 for completing the Stroop exercise and the 20 multiple-choice questions, with each correct answer earning an additional $0.75 for the student. Each participant was randomly assigned to either a Tuesday-Wednesday group or a Friday-Saturday group; based on the results of Study 1, these different two-day periods presumably reflect different stocks and/or flows of willpower. We also required people to do 250 rounds of the Stroop exercises on the first day before proceeding to answer the multiple- choice questions. Difficult (or “discordant”) Stroop exercises are used in psychology experiments to deplete willpower and consist of showing words that are the names of colors, although the actual words are printed ina color of ink different from the color name they tepresent. For example, the word “blue” might be printed in red ink. One is asked to respond by typing the color seen and ignoring the word itself. It turns out that this is much harder than it sounds. We randomly assigned each participant to process either difficult Stroop exercises or easy ones (where the word color matches the ink color). 3. THEORETICAL BACKGROUND A recent literature in economics has explored time-inconsistent behavior. Overall, one of the causes for apparent reversals in preferences over time seems to be the change in the saliency of the costs and benefits of the activity in question (Akerlof 1991). This type of systematic preference reversal is often represented by quasi-hyperbolic time discounting, under which immediately available rewards have a disproportionate effect on preferences relative to more delayed rewards, causing a time-inconsistent taste for immediate gratification. Strotz (1956), Ainslie (1992), Laibson (1997) and O’ Donoghue and Rabin (1999) discuss present-biased (quasi-hyperbolic) preferences as an explanation for persistent bad habits and addictions.'* The idea here is that the present is qualitatively different than any future date, so that the present “self” is drawn to immediate gratification. The formulation from Laibson (1997) and O’Donoghue and Rabin (1999) for preferences across a stream of utilities, where 24 is a person’s instantaneous utility in period ¢, is: T VEU" (u,,11,...-p) #5'u, +B S6'u,, where 0< Band dsl. retel Here 6 is the standard, time-consistent, exponential discount rate, whereas 6 < 1 indicates a “bias for the present”.’* The quasi-hyperbolic discounting model is consistent with considerable experimental evidence.'* 3 Frederick, Loewenstein, and O’Donoghue (2002) provide a comprehensive review of empirical research on intertemporal choice, as well as an overview of related theoretical models. We would also like to mention two very new papers. Bisin and Hyndman (2008) investigate stopp ing-time problems and characterize behavior for exponential, naive-hyperbolic and sophisticated-hyperb olic discounters. They show that an agent with standard time preferences who suffers from “temptation and self-control” would never be willing to self-imp ose a deadline. Suvorov and van de Ven (2008) develop a theory of self-regulation based on goal setting. They derive a condition under which proximal short-term goals are better than distal long-term goals. 4 We note that a precursor of this formulation appeared in Phelps and Pollak (1968). 15 However, see Rubinstein (2003) for experimental evidence not supportive of the quasi-hyperbolic-discounting models. 10 tequirements as long as the optimal amount of weekly studying is not less than the requirement. If the optimal allocation of studying is less than the requirements, then these requirements are effectively additional hurdles that must be overcome. If the cost of studying is subject to stochastic shocks, then the weekly requirements reduce the flexibility that students have to tecover from an early shock (in addition to being extra hurdles). In short, we should expect a higher success rate from participants in the treatment without weekly requirements. In terms of allocation of studying hours across weeks, in a deterministic setting, we expect students to study less early on and more later (assuming they have a daily or weekly discount rate). However, with uncertainty about the future (e.g., one’s own health later in the study), a prudent individual would log more study hours early on, thereby preserving flexibility. The intra-week pattern could either be increasing or declining depending on uncertainty. In addition, important differences in the opportunity cost of studying on different days could also be reflected in intra- week studying patterns. Without additional assumptions, this model makes no clear predictions in Study 2 about the order of the sub-tasks.”° Again, opportunity-cost considerations could affect the endogenous sub-task ordering. In Study 3, agents might avoid the additional transaction cost of logging in on the second day, thus answering all 20 questions on the first day; however, we could observe smoothing over the two days with sufficiently convex costs. In terms of performance, there should be no difference across days in the percentage of questions answered correctly. Finally, the neoclassical model predicts that the Stroop test should have no effect. The quasi-hyperbolic discounting model (with present-biased preferences) and the dual- self model generally result in similar predictions in our environments: all else equal, agents 20 Recall that participants were asked to complete sets of 2, 4, 6, 8, 10, 12, and 14 questions over seven consecutive days, with the order chosen by the participant. 13 should delay the work until the deadline approaches. In Study 1, with weekly requirements, we would expect study hours to be logged near the end of each week; without weekly requirements, we would expect a surge in hours logged late in the five-week period. However, since uncertainty could potentially overwhelm the tendency to procrastinate, the predictions for the studying profile in Study 1 will depend on assumptions about parameter values. In the absence of uncertainty, both models predict that students with weekly constraints should be more likely to complete the task. In Study 2, in principle we should observe an increasing profile over time for the number of questions in the sub-task; while this could be mitigated by uncertainty about the future, a teally bad day is fatal regardless of the chosen sub-task order. Predictions in Study 3 depend on parameter values, so once again there are no clear predictions. If § in the quasi-hyperbolic- discounting model is linked to willpower or cognitive resources then reducing willpower should induce procrastination. To date, elucidating the factors that determine an individual’s B has not been a focus of the quasi-hyperbolic-discounting literature. On the other hand, Fudenberg and Levine (2006, p. 1449) explicitly state: “increased cognitive load makes temptations harder to resist”. Thus, their model predicts that students who are assigned the difficult Stroop should procrastinate. In general, the predictions of the Bénabou and Tirole (2004) model of self-reputation depend on parameter values. However, one important aspect of the model is that (in their two- period setup) if one has been subject to external controls, these controls (weakly) reduce the likelihood that the agent puts his will to the test in period 2, as he or she doesn’t gain the needed self-confidence. In their own words: "The degree of self-control an individual can achieve is shown to ... decrease with prior external constraints” (Bénabou and Tirole, 2004, p. 848). A 14 natural interpretation of this result in our context is that a participant with weekly requirements in Study 1 will be less likely to complete the studying task. In Study 2, an increasing profile might make sense, since completing sub-tasks successfully leads to positive signals about one’s self. In Study 3, the Bénabou and Tirole model makes no clear prediction, since strength of will is fixed over time. However, if the difficult Stroop test enhances a student’s recall of their own strength of will then the model predicts students with the difficult Stroop will be less likely to procrastinate or will at least be more successful on the second day (see p. 865). The predictions of the Ozdenoren, Salant, and Silverman (2007) model depend on the stock of willpower and the relative depletion/replenishment rate. Agents with a sufficient stock of willpower will smooth their hours over time, while other agents would front-load effort provision if willpower is being depleted, or back-load effort if willpower is being replenished. Thus, we could observe differing profiles of logged studying hours over time, depending on the parameters. We could observe a declining studying-hour profile over time, with this tendency augmented by the presence of uncertainty about the future. It is interesting that if willpower is depleted during one part of the week and replenished during another, this model predicts weekly cycles in Study 1, even without opportunity costs that vary over the days of the week. In terms of which treatment group will be more successful in Study 1, the model predicts that students with weekly requirements should typically do better since the requirements substitute for using a student’s own willpower to stay on task (p. 19). In Study 2, behavior will reflect depletion/replenishment patterns, with longer sub-tasks chosen when the stock of willpower is relatively high. If willpower is depleted during the week and replenished on the weekend then this should influence when students complete tasks. In Study 3, students who have their willpower depleted by the hard Stroop should choose to answer 15 support for the hypothesis that flexibility is preferred to constraints because it allows students to tecover from shocks to the cost of studying. One interesting aspect of our data is the cyclical (weekly) patterns in the number of study hours logged. Figure 2 shows these patterns for those students who successfully completed the studying project: Figure 2 Average daily hours studied by group (winners only) Bo 3 = : iW VY zo 0 0 Sunday Sunday Sunday Sunday Sunday Day (1-35) gr2_daily_ava gr1_daily_avg target “gr1” (“‘gr2”) refers to the group without (with) weekly requirements While it may not be surprising to see a weekly, cyclical pattern when students face weekly study requirements, we had not anticipated this in the unstructured regime; if anything, the pattern is stronger for the group without weekly requirements. Cumulatively, the study hours logged for winners were very close to a target line of 15 hours per week. The average number of study hours (75.35) for the winners was close to the minimum of 75, ranging from 75.02 to 76.74. This suggests that students did not find this studying task to be innately pleasurable. 18 We turn to whether there is a difference in study hours across weeks for those who completed the 75 hours of study. Table 1 presents the average number of hours for the winners by week and by group: Table 1: Average weekly study hours (winners), by group Week Weekly requirements _| No weekly requirements 1 16.92 13.91 2 14.53 16.08 3 11.38 13.78 4 15.75 16.27 5 16.78 15.31 There is no clear trend over time for either treatment. Regressions of hours against weeks yield insignificant coefficients for the time trends (0.09 and 0.30, respectively, with corresponding ¢-statistics of 0.11 and 0.76). This does not appear to be evidence of procrastination. However, if we look at the study patterns for each individual, there is evidence that some people front-load studying hours while others delay the bulk of their studying hours.” Appendix A shows the study hours by individual, both for all participants and for only those people who completed the 75 hours. We classify people who finished as front-loaders (back- loaders) if they logged at least half of their hours in the first (last) two weeks of the five-week period. Four of the 37 people who finished were front-loaders, while 11 people were back- loaders.”° We present OLS regressions showing study patterns for both treatments in Table 2: Tn addition, one might classify those people who didn’t finish (or didn’t even start) as more serious procrastinators (we thank Jeroen van de Ven for this comment). Indeed, more extreme procrastinators meant to sign up, but didn’t get around to it (or didn’t even get around to submitting their college applications on time). 5 We see little difference across gender in this regard — one (three) of the 11 (26) male (female) winners chose to front-load, and three (eight) of the 11 (26) male winners chose to back-load; neither of these differences is close to statistical significance (Z= 0.22 and Z = 0.21 for the respective comparisons). 19 Table 2 — Regressions for weekly studying hours (winners only) Group : () (2) (3) Wessatiles Weekly requirements | No weekly requirements Pooled data Week 2 -2.39 2.17 0.32 [1.57] [1.41] [1.09] Week 3 -5.54*#* -0.13 -2.32* [1.62] [1.36] {1.11] Week 4 -1.18 2.36 0.92 [2.24] [2.82] [1.89] Week 5 -0.14 1.40 0.78 [2.82] [2.93] [2.04] Constant 16.92*** 13.91 *** 15.13*** [1.15] [1.36] [0.95] # Observations 75 110 185 R? 0.12 0.02 0.03 Week 1 is the omitted variable in these regressions. Robust standard errors clustered by subject are in brackets. * and *** indicate significance at the 10% and 1% level, respectively (two-tailed tests). The pooled regression includes a control for treatment group (not reported). The regressions confirm that there is no clear increasing or decreasing profile over the course of the experiment. Only the coefficient for the dummy for Week 3 has any statistical significance, and this would appear to reflect the effect of a closure of the library during evening peak study time due to a power outage (see the dip around day 16 in Figure 2). Table 3 shows the average hours of studying logged by winners on each day of the week: Table 3: Average study hours (winners), by day and group Day Weekly requirements | No weekly requirements Pooled data Monday 3.14 3.25 3.20 Tuesday 2.69 2.94 2.84 Wednesday 3.04 2.87 2.94 Thursday 2.57 2.21 2.36 Friday 1.01 1.05 1.03 Saturday 1.06 0.90 0.96 Sunday 1.56 1.84 1.73 20 Table 5: Logged minutes of study, by portion of day and day of week Day of week Morning Afternoon Evening Total Monday 3637 (8.9%) 19056 (46.8%) | 18062 (44.3%) 40755 Tuesday 2047 (5.4%) 20343 (54.2%) | 15171 (40.4%) 37561 Wednesday 4368 (10.8%) 20908 (51.9%) | 14989 (37.2%) 40265 Thursday 1551 (5.4%) 14535 (50.2%) | 12867 (44.4%) 28953 Friday 4395 (29.9%) 8368 (57.0%) 1928 (13.1%) 14691 Saturday 507 (4.2%) 9331 (77.3%) 2230 (18.5%) 12068 Sunday 619 (2.6%) 11752 (49.4%) | 11416 (48.0%) 23787 The percentage of each day’s study minutes is shown in parentheses. “Morning” was from 8:00-12:00 on weekdays, 9:30-12:00 on Saturday, and 10:20 on Sunday, “afternoon” was from 12:00-6:00, and “evening” was from 6:00- 12:00 except on Saturday, when the library closed at 11:00. On occasion, the library opened at 7:30 during the week. In general there is little logged study time in the mornings, with the exception of Friday, where the proportion of study minutes logged in the morning is much higher than for any other day. The amount of logged study time is about the same in the afternoon and the evening for Sunday through Thursday, but is much lower in the evening on Friday and Saturday. Since these periods were not the same length, we also present a plot of the density of study time logged by time of day in Appendix C. A final question of importance is whether completing (or even attempting) the study task was helpful in terms of performance. As mentioned earlier, 42 of our participants originated in an introductory class and we were given permission to access the (anonymous) grade records for the course, matching student ID numbers for the participants. There were quizzes, a midterm, and a final in the course. Our study commenced in the fourth week of the quarter, with two quizzes preceding our study. There is more variability in the quiz grades, with a number of people missing them, particularly after the midterm.” We therefore trust the midterm (taken in the second week of the study) and final-exam scores more, but nevertheless include an average for the first two quizzes. Table 6 shows the mean scores by group: 2° The absentee rate on the quizzes after the midterm was more than twice as high as the absentee rate on the quizzes before the midterm. 23 Table 6: Mean scores on tests, by group Group N Quiz Midterm Final Non-participants 403 3.34 8.78 7.56 Participants, non-winners 21 3.21 9.57 77 Participants, overall 42 3.43 10.14 8.48 Participants, winners 21 3.67 10.71 9.24 We see that the differences between the non-participants and the non-winners are generally small, although slightly larger for the midterm.*° In fact, Wilcoxon-Mann- Whitney (two-tailed) ranksum tests confirm that none of these differences are significant (for the midterm comparison, we find that Z = 1.37, p=0.171). On the final exam, there was no difference between non-winners and non-participants (Z = 0.10); there is also no significant difference between winners and non-winners on the midterm scores (Z = 1.58, p = 0.115); however, the difference between final scores is in fact significant (Z = 2.38, p = 0.017). Thus, the data suggest that the difference in test scores increased over the course of the quarter.*" Study 2 The task in Study 2 was rather unforgiving, in that participants were required to perform tasks on each of seven consecutive days, with no tolerance for a missed day. As a result, only 15 percent of the participants (28 of 181 people) completed the task successfully. Sixteen of 69 females (23 percent) succeeded at the task, compared to 12 of 112 males (11 percent); the difference in rates is statistically significant (Z = 2.25, p = 0.024, two-tailed test). It seems that there may have been some confusion (a serious hazard in online experiments), since some people °° We note that the midterm took place before many non-winners had stopped logging study hours. +! One issue is whether the study hours in the monitored location were simply a substitute for study hours elsewhere. The data suggest that perhaps this is not completely the case. In addition, the results from our pre- and post- experiment questionnaires reveal that only 24% of the eventual winners studied more than 15 hours per week before the experiment started and 64% of winners reported reducing their weekly study hours once the experiment ended. This provides some evidence that the experiment was an exogenous shifter of total hours studied over the five-week period. 24 performed all seven sub-tasks, but not on consecutive days.*? Overall, 60 people who signed up completed no sub-tasks, 69 people completed between one and six subtasks, and 52 people completed all sub-tasks (but only 28 people did so on time). Since so few of the people who tegistered for the experiment completed the task on time, we report data from various categories. As we asked for the planned task order a week before the sub-tasks could be performed, one natural issue is whether people were more likely to finish on time if they at least started off with their self-imposed plan.** We consider the 96 people who completed at least one sub-task on the first day. There were 46 people who stuck to their first-day plan, and 14 of these people finished on time. Of the 50 people who did not stick to their first-day plan, 14 finished on time. There is no significant difference in these proportions (Z = 0.26). Thus, having a self-imposed plan did not seem to benefit people, much as having an exogenously-imposed plan was unhelpful in Study 1. We might also be interested in whether there are differences in the plans of the people who succeed and those who do not. Figure 3 shows the average number of questions planned for each of the seven consecutive days, for people who finished on time and for people who did not: *7It is of course also possible that students viewed these questions as test preparation for the final exam, since the timing of the study was in late November and early December, but this explanation seems unlikely. A potential problem with online experiments is that there is no real chance for human feedback if a participant has a question. +3 Since everyone who completely stuck to his or her plan finished on time, the best we can do is to analyze whether people stuck to their plans for a smaller number of days, with one day being the cleanest test. 25 Table 8 — Regressions for questions answered over task days (full sample) f @) (2) (3) Wessatiles Males only Females only Pooled data Task day 0.42*** 0.02 0.02 [0.06] [0.07] [0.07] Male - - -1.78*** [0.38] Task day*Male - - -0.40*** [0.09] Constant 1.98*** 3.76*** 3.76*** [0.24] [0.30] [0.28] # Observations 271 223 494 R? 0.17 0.00 0.09 The number of questions answered is the dependent variable in these regressions. Standard errors are in parentheses. *** indicates significance at the 1% level (two-tailed test). One of the purposes behind our design in Study 2 was to examine whether day-of-the- week effects persist to some degree when the task is online, so that no physical journey is tequired. If we consider the stock of ‘studying willpower’ over the course of a week, Study 1 suggests that students replenish their supply over the weekend and deplete it once the week is over. To test whether this pattern holds, we examine the day of the week on which people ‘quit’ (first failed to complete a sub-task); this is illustrated in Figure 5: Figure 5: Quit rates by day of week 20% 15% 10% 5% Percentage of all quits O% + T T T T T T 1 Mon Wed = Thurs Fri Sat Tues Sun 28 This pattern largely supports the willpower depletion/replenishment story suggested by the weekly cycles in Study 1. Monday and Tuesday have easily the lowest quit rates and, the tate steadily increases through Saturday, with the exception of Friday. A linear regression of the quit rate against the day of the week from Monday (= 0) to Saturday gives: Quit rate = 0.110 + 0.103*Day, (0.015) (0.005) confirming that there is a significant upward trend over this period. If the opportunity cost of doing schoolwork varies substantially over the course of the week, as suggested by the weekly cycles in Study 1, we should expect substantial differences in the number of questions answered on the weekdays or on the weekend. However, there is no dramatic difference in questions answered by day of the week. This is shown for both the entire sample (labeled “All”’) and for people who finished the task on time (labeled “Success”), as shown in Figure 6: Figure 6: Questions answered by day of week Ne 10 cy —e—aAll |—a— Success Questions answered per person Mon Tues Wed Thurs Fri Sat Sun The number of questions answered is slightly higher on Sunday and Monday, while it is lowest on Saturday (and also on Wednesday for people who finished on time). We also see a Saturday effect when we compare planned tasks to actual tasks. While there is no statistical 29 difference between planned tasks and actual tasks comparing across experiment day (i.e. first day of the experiment, second day, etc.), there is one statistical difference between planned and actual tasks comparing across weekdays and this occurs on Saturday. Perhaps not surprisingly, students planned to allocate more effort on Saturday than they actually allocated. Study 3 Of the 158 people who signed up online, 100 (63 percent) completed the task successfully. In a departure from our earlier results, 60 of 85 males (71 percent) completed this less difficult task, compared to 40 of 73 females (55 percent); this difference is statistically significant (Z = 2.05, p = 0.040, two-tailed test). As might be expected, success rates were significantly higher for the Tuesday-Wednesday group, as 57 of 79 people (72 percent) in this group finished and 43 of the 79 people (54 percent) in the Friday-Saturday group finished (Z = 2.31, p = 0.021, two-tailed test). Thirty-nine people answered all 20 questions on the first day, 15 people answered all 20 questions on the second day, and the other 46 people who completed the task answered questions on both days. Ninety-two people answered some questions on the first day, while 64 people answered questions on the second day.*® Forty-nine people never answered any questions, while the remaining nine people answered between four and 15 questions. Sixty-one of the 64 people (95 percent) who answered some questions on the second day completed the task successfully. In the sample as a whole, people answered an average of 7.62 (5.55) questions on the first (second) day. With respect to the people who completed the task, the average number of questions answered on the first (second) day was 11.53 (8.47); similarly, for the nine people who 35 While the design of our second and third studies reflected our own learning process, and our intention was never to specifically compare results in one study to another (since many design features were changed simultaneously), we nevertheless note that the proportions of people who never answered any questions are very similar in Study 2 (60 of 181, 33%) and Study 3 (49 of 158, 31%). 30 females) or type of Stroop test (60 percent in both cases). In addition, the percentage of correct answers on the first day does not predict success rates (the statistic on the coefficient for number of questions is 0.21). However we do find that the percentage of correct answers on the first day is significantly higher than on the second day.** This is robust to whether we consider the whole sample of 158 people, the 100 people who finished, or the 47 people who answered questions on both days.” The percentage of correct answers is slightly higher for the Tuesday- Wednesday group (0.63 vs. 0.56, Z = 1.78, p = 0.075, two-tailed test). 5. DISCUSSION We find evidence of procrastination, although it is not as ubiquitous as some might have expected. While there was no aggregate time trend over the five weeks of Study 1, nevertheless 30 percent of the participants who completed the task deferred logging most of their studying hours to the last two weeks of the five-week period, despite the uncertainty issue. In Study 2, even the people who complete all of the sub-tasks show a clear tendency to delay completing the more difficult sub-tasks until later; however, this is only true for male participants. In Study 3, 61 of the 100 people (61 percent) who complete 20 questions don’t do so on the first day, even though there is an additional transaction cost in returning for the second day. We observe strong day-of-the-week effects. In Study 1, there are pronounced weekly cycles, with the most studying done on Monday through Wednesday and the least on Friday and Saturday. In Study 2, the lowest quit rates were on Monday and Tuesday, while completion rates are lowest for people whose starting date was a day or two before the onset of the weekend. In Study 3, people in the weekday group are significantly more likely to complete the task than the 38 The order of questions was randomized, so there should be no difference in difficulty levels across days. 5° The respective comparisons are 0.66 vs. 0.49 (Z= 4.35, p = 0.000), 0.66 vs. 0.50 (Z=4.10, p = 0.001), and 0.67 vs. 0.47 (Z = 3.43, p =0.001). All of these tests are two-tailed. 33 weekend group. However, the fact that in Study 2 there was no substantial difference across days of the week in the number of questions people chose to answer would seem to be evidence against the notion that the opportunity cost of working is higher on the weekend (at least when one can do so at home). There is strong evidence that suggests weekly cycles of willpower depletion and replenishment. In Study 1, logged study hours are highest on Monday-Wednesday, dropping later in the week and during the weekend before returning to the Monday level; in spirit this cycle seems closest to a situation where willpower is depleted over the course of the week and teplenished on the weekends; in other words, willpower is renewable. In Study 3, willpower seems considerably higher for the Tuesday-Wednesday group than for the Friday-Saturday group, as the completion rate is significantly higher. While the number of questions answered on the first day is similar with hard and easy Stroop tests for the group as a whole, people who were assigned the hard Stroop answered significantly more questions on the second day. This is consistent with the notion that succeeding at the hard Stroop serves either as a positive signal to one’s self (as in the Bénabou and Tirole model) or increases the will to complete the task. We find gender effects in each of our studies; however, at least at first glance these are inconsistent. Males are significantly less likely than females to complete the most organizationally-challenging task (Study 2), marginally less likely to complete a lengthy task with some flexibility on a daily basis (Study 1), but are significantly more likely to complete a task of relatively short duration (Study 3). While all the models are essentially agnostic on this point, our interpretation is that female students are generally more successful than males at organizing and completing tasks that require consistent attention over a period of time, males 34 have the willpower and energy to actually do better than females on tasks for which the finish line is never very far from sight, and the task can even be done in one sitting.” One feature common to each study is that behavioral mechanisms that seem attractive often lead to surprising outcomes; this illustrates how little existing data economists currently have on procrastination and willpower, and the need for further studies. Consider the effect of the weekly requirements in Study 1. The Ariely and Wertenbroch (2002) results and our own intuition suggested that this structure would help students to achieve the studying-hours goal by preventing them from falling too far behind“! However, the proportion of successful participants was 50 percent higher when we left the five-week study period unstructured, rather than imposing weekly requirements. The difference appears to be due to the lack of flexibility imposed by weekly constraints. In Study 2, people had to subnut a planned sub-task order and we thought that this might be helpful. Yet there is no difference in success rates for people who follow their plans on the first day and for people who ignore their plans. In Study 3, the difficult Stroop test was expected to deplete willpower, and yet the success rates for people who were assigned the hard Stroop were higher (with marginal statistical significance) than the success tates for people who were assigned the easy Stroop test. We now turn to how our data reflect on the standard model, as well as on the quasi- hyperbolic, self-reputation, dual-self and willpower models described in section 3. In most cases, the predictions depend on parameters whose values are unknown to us and would be “° We are clearly only speculating on this point. Nevertheless, this would be interesting to study in more detail in future work. ‘| However, there is at least one important difference between our study and Ariely and Wertenbroch (2002), in that there was no real issue of people not completing the task in their study, since a student who did not tum in a term. paper would likely receive a failing grade, and people could turn in their proofreading without having found any particular number of errors. 35 where c/(.) is the cost of studying for / hours in week ¢. Sis the stock of remaining hours: S,,, = S,-h,. The student tries to minimize the cost of studying subject to the following constraint: T s.t. Sh, =S), tel where So is the 75 hour target. In other words, the student must “deplete” the stock of studying hours (i.e. achieve the target). We can solve this problem numerically using the following dynamic programming equation (Harris and Laibson, 2001): W(S,)= min{¢(S,,/) + W(S,1)-C(Spa)(- pI}. where C(S,) =c(S,, x(S,)) and h, = x(5,) is the optimal amount of studying in week ¢ with stock S; of hours remaining given that the student knows that his future selves will choose to study h,,. = x(S,,_) inall future periods. To add in uncertainty, we assume that all future time periods are subject to a random i.i.d. shock =: 1+ A with probability 0.5 11 _ A with probability 0.5 The objective function is now: min (5), h,) +B Sorel (S,,z,h,) retel The future uncertainty will cause the student to frontload on studying, similar to a precautionary savings model in macroeconomics (Leland, 1968). To solve the two problems, we assume that the cost of studying is a simple convex function of the hours studied in each week and is 38 independent of the current stock of remaining hours (i.e. studying for one hour costs the same tegardless of whether there are 5 hours remaining or 4 hours remaining): e(h,) =h?. Inthe first case, we assume that there is no uncertainty; we set A=0, T=5, S,=75, d=1 and B=0.9. In the second case, we instead allow for uncertainty by setting A =0.5 (the student must exert 50 percent more effort during “bad” weeks and 50 percent less effort during “good” weeks), while leaving the other parameter values unchanged. The weekly predicted study hours for each case are shown in Table 10: Table 10 — Examples of quasi-hyperbolic predictions, with and without uncertainty Week Hours studied, no Hours studied, with uncertainty (A = 0) uncertainty (A = 0.5) 1 13.73 16.74 2 14.31 15.85 3 14.67 15.26 4 15.30 14.37 5 16.99 12.78 While the parameter values we have chosen are arbitrary, the directional difference in the profile over time with and without uncertainty illustrates the potential importance of taking uncertainty into account when predicting behavior on tasks requiring an extended period of time. 6. CONCLUSION Our study is one of the first to provide empirical evidence concerning procrastination, willpower, and monetary incentives in tasks involving duration. Since it is exploratory work, we did not expect to fully resolve all of the issues present in this regard. Nevertheless, we believe 39 we have identified a number of interesting patterns, offer useful feedback to the relevant theoretical models, and suggest follow-up studies to other researchers. We find clear evidence of procrastination and willpower depletion, as well as behavior suggestive of self-reputation considerations. In this respect, each of the behavioral models identifies an important aspect of how people deal with tasks, yet none of the models correctly predicts behavior in all three of our studies. It is possible that some elements of these models can be combined to produce a more descriptive model, as the picture appears to be more complex than can be accommodated by any current model. A number of puzzles remain. For example, while the difficult Stroop exercise did deplete willpower on the first day, a natural prediction would seem to be that this depletion would reduce overall success rates. Yet, overall success rates were higher for the willpower-depleted group. Is exerting willpower like exercising a muscle or does this result represent some sort of “psychological sunk cost’ for having suffered through this exercise, inducing the determination to complete the task? How much domain specificity or crossover from cognitive load is there, in the sense of Ozdenoren, Salant, and Silverman (2007) or Fudenberg and Levine (2006)? Have we identified differences in procrastination and willpower across gender, particularly in terms of differences according to whether a task is long or short in duration? The fact that the behavioral interventions in each of our three studies were ineffective or even counter-productive points out the lack of existing data on how people behave on tasks involving significant duration. It is quite clear that our results are only a beginning. We consider the area of time structuring, procrastination, and incentives to be just coming into its first full flowering. We hope that our results help spark more empirical research and more theoretical work in this nascent area. 40 APPENDIX A - INDIVIDUAL STUDY HOURS IN STUDY 1 WEEKLY STUDY HOURS IN NO-WEEKLY GROUP, BY INDIVIDUAL ID Female Week 1 Week 2 Week 3 Week 4 Week 5 Total 1 1 0 0 0 0 0 0.00 2 1 12.31 20.15 13.82 14.44 14.40 75.12 3 0 9.18 14.78 4.12 1.53 0 29.61 4 0 11.16 15.16 16.89 17.92 14.30 75.42 5 1 8.17 9.97 2.46 9.26 0 29.85 6 0 16.70 13.52 6.66 17.47 20.90 75.25 7 1 1.83 0 0 0 0 1.83 8 1 2.15 18.75 4.35 31.19 18.63 75.07 9 1 9.34 6.77 0 2.74 14.31 33.17 10 1 10.27 0.67 3.43 0 0 14.36 ll 1 17.72 17.74 20.98 10.01 8.74 75.18 12 1 8.69 18.11 17.43 10.24 20.64 75.11 13 1 12.13 12.43 5.62 5.03 39.93 75.13 14 1 2.82 0 0 0 0 2.82 15 0 18.11 11.99 12.42 18.20 14.49 75.21 16 1 12.54 13.74 10.16 19.61 19.26 75.30 17 0 14.14 22.03 18.05 14.30 7.27 75.80 18 1 0 12.87 13.76 2.74 0 29.38 19 0 16.37 6.58 27.47 8.33 16.63 75.37 44 1 13.69 9.77 13.50 13.05 25.07 75.08 45 1 14.98 12.14 16.97 22.90 8.21 75.19 46 1 24.56 28.81 16.97 5.46 0 75.80 47 1 17.68 13.94 11.94 15.77 16.56 75.89 48 1 1.33 0 0 0 0 1.33 49 1 3.81 14.67 12.08 27.67 17.48 75.72 50 1 16.32 15.49 9.53 14.44 19.38 75.16 Sl 1 15.19 17.00 11.51 14.79 16.79 75.27 52 1 6.20 11.79 13.74 23.09 20.56 75.38 53 1 5.71 15.80 6.40 29.69 17.58 TSAT 54 1 28.93 24.16 19.03 3.33 0 75.46 55 0 1.07 3.79 0 0 0 487 56 1 0 0 0 0 0 0.00 S7 0 12.93 0.99 0 0 0 13.92 58 0 10.64 0 3.53 0 0 14.17 59 0 7.60 0 0 0 0 7.60 60 1 16.93 19.94 17.73 20.99 0 75.60 43 WEEKLY STUDY HOURS IN WEEKLY GROUP, BY INDIVIDUAL ID Female Week 1 Week 2 Week 3 Week 4 Week 5 Total 20 0 0 0 0 0 0 0.00 21 0 0 0 0 0 0 0.00 22 0 12.34 11.62 0.02 0 0 23.98 23 0 19.07 15.41 15.86 13.10 13.29 76.74 24 0 14.97 17.59 8.66 12.72 21.11 75.05 26 1 19.70 14.12 12.74 7.08 21.47 7511 27 1 13.84 11.67 11.12 7.40 0 44.03 28 0 14.70 13.85 8.35 27.88 11.19 75.97 29 0 12.22 3.21 0 0 0 15.42 30 0 0 0 0 0 0 0.00 31 1 18.02 8.23 6.65 15.43 23.89 72.21 32 1 0 0 0 0 0 0.00 33 0 30.06 16.99 12.86 15.11 0 75.02 34 0 12.02 12.07 12.21 22.59 16.24 75.13 35 1 0 0 0 0 0 0.00 36 1 13.53 13.85 19.62 10.30 17.79 75.09 37 0 0 0 0 0 0 0.00 38 1 0 0 0 0 0 0.00 39 0 0 0 0 0 0 1.73 40 0 0 0 0 0 0 0.00 Al 0 11.81 4.52 0 0 0 16.33 42 1 5.31 0 0 0 0 5.31 43 1 15.55 13.10 10.23 16.69 20.50 76.06 61 0 12.43 3.82 0 0 0 16.25 62 1 13.02 15.41 10.49 12.42 23.80 75.14 63 1 15.44 28.28 4.82 9.00 17.59 75.13 64 1 14.81 13.60 9.98 13.20 23.67 75.26 65 1 0 0 0 0 0 0.00 66 0 1.41 0 0 0 0 1.41 67 1 0.95 0 0 0 0 0.95 68 1 12.53 12.32 11.93 16.05 2.34 55.18 69 0 14.40 1.61 0 0 0 16.01 70 1 15.60 12.30 7.70 16.07 6.28 57.95 71 0 17.40 7.24 20.59 24.37 5.44 75.04 72 1 19.58 15.91 6.25 21.57 12.07 75.37 7B 1 18.36 11.05 16.21 7.83 21.57 75.02 74 1 13.22 11.87 1.77 22.24 26.02 75.11 44 WEEKLY STUDY HOURS IN NO-WEEKLY GROUP, BY INDIVIDUAL (WINNERS) ID Female Week 1 Week 2 Week 3 Week 4 Week 5 Total 2 1 12.31 20.15 13.82 14.44 14.40 75.12 4 0 11.16 15.16 16.89 17.92 14.30 75.42 6 0 16.70 13.52 6.66 17.47 20.90 75.25 8 1 2.15 18.75 4.35 31.19 18.63 75.07 ll 1 17.72 17.74 20.98 10.01 8.74 75.18 12 1 8.69 18.11 17.43 10.24 20.64 7511 13 1 12.13 12.43 5.62 5.03 39.93 75.13 15 0 18.11 11.99 12.42 18.20 14.49 75.21 16 1 12.54 13.74 10.16 19.61 19.26 75.30 17 0 14.14 22.03 18.05 14.30 7.27 75.80 19 0 16.37 6.58 27.47 8.33 16.63 75.37 44 1 13.69 9.77 13.50 13.05 25.07 75.08 45 1 14.98 12.14 16.97 22.90 8.21 75.19 46 1 24.56 28.81 16.97 5.46 0 75.80 47 1 17.68 13.94 11.94 15.77 16.56 75.89 49 1 3.81 14.67 12.08 27.67 17.48 75.72 50 1 16.32 15.49 9.53 14.44 19.38 75.16 Sl 1 15.19 17.00 11.51 14.79 16.79 75.27 52 1 6.20 11.79 13.74 23.09 20.56 75.38 53 1 5.71 15.80 6.40 29.69 17.58 TSAT 54 1 28.93 24.16 19.03 3.33 0 75.46 60 1 16.93 19.94 17.73 20.99 0 75.60 WEEKLY STUDY HOURS IN WEEKLY GROUP, BY INDIVIDUAL (WINNERS) ID Female Week 1 Week 2 Week 3 Week 4 Week 5 Total 23 0 19.07 15.41 15.86 13.10 13.29 76.74 24 0 14.97 17.59 8.66 12.72 21.11 75.05 26 1 19.70 14.12 12.74 7.08 21.47 7511 28 0 14.70 13.85 8.35 27.88 11.19 75.97 33 0 30.06 16.99 12.86 15.11 0 75.02 34 0 12.02 12.07 12.21 22.59 16.24 75.13 36 1 13.53 13.85 19.62 10.30 17.79 75.09 43 1 15.55 13.10 10.23 16.69 20.50 76.06 62 1 13.02 15.41 10.49 12.42 23.80 75.14 63 1 15.44 28.28 4.82 9.00 17.59 75.13 64 1 14.81 13.60 9.98 13.20 23.67 75.26 71 0 17.40 7.24 20.59 24.37 5.44 75.04 72 1 19.58 15.91 6.25 21.57 12.07 75.37 7B 1 18.36 11.05 16.21 7.83 21.57 75.02 74 1 13.22 11.87 1.77 22.24 26.02 75.11 AS