Self-Control at Work: Evidence from a Field Experiment Supreet Kaur, Michael Kremer, and Sendhil Mullainathan *

Transcription

1 Self-Control at Work: Evidence from a Field Experiment Supreet Kaur, Michael Kremer, and Sendhil Mullainathan * Abstract When workers have self control problems a key feature of agency theory changes. In addition to workers not working as hard as employers would like, they now do not work as hard as they themselves would like. As a result, the employer s value changes. Instead of only providing insurance, he also increases productivity (and utility of the patient self) by helping the patient self solve his self-control problem. We designed and implemented an experiment in the field with data entry workers that tests key predictions of a self-control agency model. First, consistent with worker sophistication, we find that workers choose dominated contracts as a way to motivate themselves to work harder: they pick a contract which pays less for every output level but has steeper slope. Second, consistent with very high short-term discounting, workers efforts increase significantly as the (randomly assigned) payday gets closer. Third, these two effects are linked: the demand for dominated contracts (and their benefits) is concentrated amongst those with the highest payday effects. We fail, however, to find a horizon effect: dominated contracts are chosen at the same rate for the next morning as for the same day. Ex post analysis suggest the temporal resolution of uncertainty may moderate this effect: when workers face uncertainty that will be resolved in the morning, they wait to choose the dominated contracts; when they do not, they choose them the night before. Finally, workers appear to learn about the need for these contracts. As workers gain experience, their self-control problems do not change. But they do appear to learn about them: the correlation between the payday effect and the demand for the dominated contract grows with worker experience. We argue that these results are hard to reconcile with other explanations such as worker confusion, social influences or signaling to employers. I. Introduction In agency theory, there is a tradeoff of insurance against incentives. To a farmer, a fixed wage contract provides greater insurance but less incentives than a sharecropping contract. Workers must be compensated for the risk embodied in higher powered incentives. This model presumes a simple view of work: workers will not work as hard as the employer would like. Introspection suggests a different problem: the worker may not work as hard as the worker himself would like. Choosing for the future, people would like to act patiently: tomorrow, I would like to work hard. Choosing for the present, though, they act impatiently: today, I would rather watch TV. * Supreet Kaur: Department of Economics, Harvard University, Cambridge, MA (kaur@fas.harvard.edu); Michael Kremer: Department of Economics, Harvard University, Cambridge, MA (mkremer@fas.harvard.edu); Sendhil Mullainathan: Department of Economics, Harvard University, Cambridge, MA (mullain@fas.harvard.edu). For helpful comments, we thank Lawrence Katz, David Laibson, and seminar and conference participants at Advances with Field Experiments (Wharton), CEPR Development Economics Workshop (Barcelona), Cornell, Dartmouth, Harvard, LSE, NEUDC Conference (MIT), Summer Institute in Behavioral Economics (Trento), UCLA, University of Chicago, USC Marshall, and Washington University. Raluca Dragusanu, Matthew Darling, and Nicolas Papamichael provided excellent research assistance. We thank Dominik Bulla, Ramya Krishna, Karthik R, and the Centre for Innovative Financial Design (IFMR) for operational support. This project received financial support from Harvard University, the International Finance Corporation, and the National Science Foundation. All errors are our own. 1

2 Work effort involves a self-control problem. 1 How does the addition of a self-control problem change the logic of agency theory? 2 We argue that it significantly alters the incentive-insurance tradeoff. A simple example illustrates how. Suppose an employer receives a contract to enter data. If it does not finish by a certain date it will be penalized an amount C. It hires a worker to enter the data at some wage w. To incentivize the worker to meet the deadline beyond her wage she also receives a penalty for not meeting the deadline. But there is risk here the data could turn out to be complex to enter so that even at his best effort, the worker may not be able to meet the deadline so the firm imposes a penalty c < C. Suppose the firm wanted to increased the penalty c. The worker would certainly work harder and the deadline would more likely be met. This increases the firm s revenues. It also increases worker earnings. But if workers were optimizing, this would only decrease their utility. They were already working as hard as they would like; now they are exerting more effort with no greater reward (only the penalty has gone up). Worker wages would have to be increased to offset the decrease in utility. This is not necessarily the case when workers have self-control problems and recognize that they have them. From the standpoint of period t, a worker contemplating an increase in the penalty in period t+1 faces a different calculus. She wants the future self to exert more effort. An increase in the penalty forces the t+1 self to work harder harder than the t+1 self would like, but closer to what period t would like. 3 As a result, a time inconsistent worker may voluntarily take on the higher penalty without any extra compensation. The traditional tension between worker and employer is replaced by a joint interest between the employer and the worker. The worker favors steeper incentives because they bind future selves to work harder; the employer favors them because they produce greater output. This insight means that self control at work if important could alter our understanding of work arrangements. 4 Absent self-control, incentive contracts increase welfare by providing insurance but do so at the cost of productivity workers are most productive when they are the 1 See Frederick et al 2002 and DellaVigna 2009 for reviews of time inconsistency. Theoretically, models from hyperbolic discounting to dual-self games model time inconsistency. Prominent models of time inconsistent behavior include Laibson (1997), O Donoghue and Rabin (1999; 2001), Gul and Pesendorfer (2001; 2004), Fudenberg and Levine (2006), and Banerjee and Mullainathan (2009). Empirically, self-control has been explored in a diverse set of areas ranging from retirement savings (Angeletos et al 2001) to gym memberships (DellaVigna and Malmendier 2006). Frederick et al (2002) and DellaVigna (2009) provide reviews of the growing body of lab and field evidence on time inconsistency. 2 O Donoghue and Rabin (2006) are amongst the first to ask this question. They show how a principal uses inter- temporal incentives (e.g. deadlines) to motivate a procrastinator. They focus on naïve agents and on one- time tasks (completing a project). In contrast, our focus is on (at least partly) sophisticated agents and on continuous tasks (data entry). Because, however, they assume that the naive agents use the IR constraint that sophisticated agents would, their agents also demand incentive contracts with commitment benefits (e.g. deadlines). They also produce interesting implications for screening which we do not address. DellaVigna and Malmendier (2004), on the other hand, derive contract design implications in a long-term purchasing context, e.g. gym memberships. 3 We are assuming here as in the paper that workers are sophisticated about their self- control problem (as discussed in O Donoghue and Rabin 2004). They must realize they will not work as hard as they would like in order to generate an additional value to steeper incentives. 4 See Kaur et al (2010) for a discussion of the ways in which workplace arrangements can mitigate self control problems. 2

3 full residual claimants. 5 In the self-control view, employers can actually increase productivity beyond what workers would do as the full residual claimant. 6 The employer is an interested party with both the motive and the means to mitigate self-control problems. 7 All this depends on two key assumptions: self-control at work is quantitatively important and workers are sufficiently sophisticated as to value the implicit self-control benefits of stronger incentives. Are these assumptions accurate? We performed an experiment with data entry workers to find out. 8 We recruited and hired individuals to perform full-time work in an office setting for one year. We varied the workers incentives and workplace conditions to test basic predictions of a behavioral agency model. In the experiment, workers are paid weekly according to a piece rate that depends on the number of accurate fields entered each day. We randomize them into two conditions. Workers in the control condition receive a simple linear scheme: getting a piece rate w for each accurate field entered that day. Workers in the choice condition choose a target T. If they enter T or more fields, they receive w per accurate field. If they fail to achieve the target they receive w/2 per accurate field entered. Choosing T >0 is dominated by choosing T=0, which replicates what they get on control days; like in the hypothetical example, choosing T > 0 only increases the penalty without increasing the reward. In these treatments we can compare whether workers actually choose these dominated contracts and what impacts they have on productivity. Moreover, to test how horizon affects these choices, we randomly assigned workers to a Morning and Evening choice condition. In Morning choice, workers choose the target for that day when they arrive at work. In Evening choice, they choose the target the night before. 5 Complications to the simple agency model could generate a productivity gain from the employment relationship. The most important case of this is the case of team production. Cheung (1969) and Alchian and Demsetz (1972), for example, emphasize how workers would pay for a boss when they must produce output together. The boss reduces the public good problem of effort and increases productivity. We will argue a similar effect even in cases of individual production. 6 Clark (1994) makes this argument for changes in the organization of production during the industrial revolution. He argues that the shift away from the pre-industrial putting out system (in which workers were paid piece rates and were full residual claimants) to the more rigid workplace system that is the norm today (employment in factories with features like assembly lines, production minimums, rigid work hours, and hefty punishments for momentary lapses in behavior) was in part an attempt to solve self-control problems. 7 Ashraf et al (2006) and Gine et. al. (2010) show demand for commitment in savings and smoking cessation. The employment context is unique for two reasons. First, the commitment device here is implicit in the structure of incentives rather than existing independently it is merely a feature of an incentive scheme. Second, employers play a unique role in providing commitment to work because they are privately incentivized to do so- they share in the resulting productivity gains. It is less clear that banks benefit as much from providing commitments against towards savings or which firms stand to gain from providing commitments to quit smoking. A third part with independent incentives is particularly important in the presence of partial naivete (O Donoghue and Rabin..) where workers will demand commitment devices but under-demand them. 8 Two papers illustrate self-control problems at work. Ariely and Wertenbroch (2002) hired students to complete a set of proofreading tasks over the course of a semester and examined the power of deadlines as a commitment device to work harder. Burger et al (2008) examined students studying patterns, once again using interim targets as commitment devices. Both papers entailed short durations, student populations and smaller stakes. We expand on this work by testing predictions of a behavioral agency model among full time workers over the course of a year. A limited group of field studies investigate other behavioral effects in labor markets but focus on concepts other than time-inconsistency, such as fairness (Gneezy and List 2006) and framing (Hossain and List 2009). A larger body of work uses field experiments to test other types of labor market hypotheses (Shearer 2004; Bandiera et al 2007; Fehr and Goette 2008). 3

4 Our experimental design tested another auxiliary implication of the behavioral agency model. As paydays get closer, the benefits of work are discounted less and effort should rise. 9 In other words, if the time gap between effort and reward causes self-control, reducing that gap should increase effort. To test this, we randomly assign workers to paydays: some workers are consistently paid on Tuesday, some on Thursday and some at the end of the week. This random assignment allows us to control for calendar effects on work effort such as the rest provided by the weekend or the desire to leave work early to enjoy a Friday night. Our experiment takes place over a 13-month period in a small Indian city. Workers were typically high school graduates in their twenties (see Table 2), most of whom lived in the rural areas surrounding the city. Our job was the primary source of earnings for most of our sample. The typical worker in our sample earned Rs. 170 ($3.60) per day. This is consistent with average daily wages paid by other data entry firms in the city, which typically range from Rs. 125 to Rs The payday assignments were randomized across people, so that some were allocated to Tuesday, some to Thursday, and some to Saturday paydays. Once workers were assigned to a payday, they were always paid on that day. Contract assignments, however, were randomized daily. 10 We find a significant payday effect. Workers earn 8% more on paydays than in the beginning of the weekly pay cycle. 11 We find these effects are not concentrated on paydays alone but instead production rises smoothly as the payday gets closer. We calibrate the magnitude of these effects in two ways. First we estimated the OLS coefficient for education on our sample. Second, we use exogenous variation to estimate how a change in the piece rate w (for the control condition) affects output. Using these estimates, the payday effect seems comparable a little more than a 1- year increase in education and an 18% increase in the piece rate wage. Workers also robustly demand the dominated contract. Positive targets are chosen 35% of the time (again, absent workers are considered as not choosing the dominated contract). Workers are also more productive in the choice condition, showing a Local Average Treatment Effect of 6%. This suggests the ability to choose the dominated contract affects output at the same level roughly of a year of education or a 14% increase in piece rate. When we examine hourly production, we find the target works as expected workers work harder in the hours leading up to the target and then slack off once it is reached. We also find an intimate relation between the payday effect and the choice of dominated contracts. The payday effect shows significant inter-personal heterogeneity: some workers show 9 More precisely, by itself, the payday effects are only a test of high (immediate) discount rates not of time inconsistency per se. Combined with the demand for dominated contracts, however, these treatments can paint a fuller picture of self-control at work. 10 In addition to the control and two choice conditions above, we also randomize workers into one of three commitment contracts without choice. While not useful for understanding self-control, these provide useful calibration benefits. This is reported in greater detail in Section IV. We also randomized workers seats so that every 1 to 3 weeks they moved to different seats. This allows us to estimate peer effects as discussed in Kaur, Kremer and Mullainathan (2010, 2011). 11 To avoid selection bias due to selective attendance, in all our results we report the impact on total production with non-attending workers being given zero production. All workers know their contract assignment the day before. 4

5 much stronger payday effects than others. These large payday workers are also the ones accounting for the bulk of demand for dominated contracts. They are 49% more likely to choose the dominated contracts. They also benefit more from them. They show local average treatment effects of choice of 20% (equivalent to roughly 2.5 years of education or a 50% increase in the piece rate). In other words, it appears as if some workers have large self-control problems (as proxied by their payday behavior) and those are the ones who value the commitment benefits of the dominated contracts the most. While behavior around paydays predicts demand for dominated contracts, more traditional measures of self-control fare less well. We gave workers monetary tradeoffs (for real stakes) to measure both discount rates and preference reversals. We also conducted surveys to elicit subjective measures of self-control at work 12 and in other domains (such as smoking). As a whole, these measures show at best modest predictive power. Though they point in the correct directions, they are rarely significant. We cannot rule out that we simply lack power to detect modest predictive power. Given the striking predictive power of the payday outcomes, it does suggest that behaviors in related contexts may be more powerful than subjective questions or abstract behaviors in predicting behavior. Some of our evidence is inconsistent with the simplest self-control models. Specifically, there is no difference in targets chosen between the evening and morning choice conditions. Ex post analysis suggests one hypothesis for this failure to find an effect. As we discuss in Section VII, workers may have greater demand for commitment the night before but may also face greater uncertainty that makes them less willing to commit. Both uncertainty due to computer speed and in being able to get to work on time appear to mediate this effect: as these sources of uncertainty decrease, the demand in the evening increases. In fact, demand in the evening is larger than in the morning when our measures of uncertainty are low, and this reverses when uncertainty is high. Self-control models that allowed for such uncertainty that is revealed over time could easily explain these results but we ourselves did not make such a prediction prior to running the experiment. Finally, we find workers seem to learn about their self-control problems over time. Early on, many workers experiment with dominated contracts. As they gain experience, workers diverge: some workers choosing it much more regularly and others decreasing take-up. As before, the workers who increase demand are also the ones who show the largest payday effect. Consistent with this logic, we find evidence that the productivity impact of being offered dominated contracts increases with worker experience: as workers sort better, the average treatment effect increases. In addition, we find that the effect of paydays neither declines nor increases with experience. These results suggest that workers learn about the extent of their self-control problems but do not necessarily learn away these problems (suggesting the availability of external commitment is low). Could our results be driven by factors other than self-control? In section VIII we discuss several possibilities. While alternative explanations may explain individual findings, we argue that none can explain the full pattern of results: the payday effect; the demand and impact of dominated contracts; and the correlation of contract effects with the size of the payday effect. For example, 12 For example, agreement with statements like, Some days I don t work as hard as I would like. 5

6 demand for the dominated contract might be driven by confusion: workers may choose this contract not out of a desire to bind themselves but out of a failure to understand that it is dominated by the control contract. It is less clear how confusion could explain the evening vs morning effect or the other findings. As another example, the demand for commitment may be explained by external distractions (family or friends) rather than self-control: the dominated contract may allow one to beg off these distractions without losing face. It is unclear why this would explain the correlation between take-up of the dominated contract and the payday effect. We argue that the pattern of results together suggests self-control as the likely explanation. The rest of the paper proceeds as follows. In section II, we present a model of time inconsistency in work effort and derive a set of testable predictions. We lay out our experiment design in Section III and describe the experiment context and protocols in Section IV. In Section V, we describe the data that will be used for our empirical tests. We present results in Section VI. In Section VII, we evaluate the plausibility that explanations other than time inconsistency could drive our results. In Section VIII, we provide discussion of the insights generated from our findings and conclude. II. Model We present a simple principal-agent model. Since our empirical work focuses on the demand for contracts, we do not derive the optimal contract. Instead we will use this framework to derive worker behavior under and demand for different contracts. A risk-averse agent makes an effort decision in one period and receives compensation for resultant output in a later period. We allow the worker to have time inconsistent preferences, and examine the impact of payment horizon and commitment on effort and utility. In doing so, we do not attempt to derive the optimal incentive scheme. Rather, we look at how effort responds to changes in a fixed contract, and how agents would rank different contracts. This enables us to derive a set of testable predictions on behavior under time consistent versus time inconsistent preferences. III.A. Set-up Suppose there is an agent employed in a one-period production activity. She exerts unobservable continuous effort e to produce stochastic binary output, y. Output equals 1 with probability p(e) and equals 0 otherwise. Output is perfectly observable, and the agent is paid L and H for low and high output. We will write Δ = H-L so that the agent s expected pay can be written as L + p(e)δ. She exerts effort in period 1 and receives compensation in period T+1. We model discounting with a time-horizon discount function. The agent discounts a choice at horizon τ (i.e. τ periods in the future) with discount function d(τ) 1. We assume that d(0)=1 and d(τ) decreasing in τ. If the agent is time consistent, we assume she has an exponential discount function d(τ ) = δ τ where δ equals the daily discount factor and τ is measured in days. This agent is consistent because a payoff moving from τ to τ+k is discounted by δ k which does not depend on τ-when that payoff occurs. The impatience in receiving payment a day later does not depend on whether the payment is to be received today or in a week. 6

7 Following the hyperbolic discounting work, we assume that for a time inconsistent agent the rate of time preference does depend on the horizon. We assume people are particularly impatient for receiving a payment today versus tomorrow, less impatient for tomorrow versus day after tomorrow and even less impatient for 39 days versus 40 days from now. More generally, we denote the hyperbolic discount function as d(τ), and assume that the impatience d(" + s) for a delay of s periods is decreasing in τ for any fixed s. In the empirical work we will d(") further assume that d(1) is noticeably smaller than 1, that is agent s discount even a given day at a significant amount. 13 We assume that the agent is risk averse, with concave utility, u( ), over consumption (and therefore income). The agent s cost of effort, c(e), is convex in e. We also assume separability in production, utility, and contracts. In period 1 effort happens today and payoffs occur T periods from now. Thus in period 1, a time consistent agent has utility: U 1 C = δ T p(e)u(h ) + (1 p(e)u(l) c(e) and since the agent is time consistent, her period 0 utility is simply U 0 C = δu 1 C. 14 For time inconsistent individuals, their utility is written as: U 1 I U 0 I = d(t )p(e)u(h ) + (1 p(e))u(l) c(e) = d(t + 1)p(e)u(H ) + (1 p(e)u(l) d(1)c(e) Time inconsistency means that period 1 and period 0 utility can no longer be written as U 0 I d(1)u 1 I. For a given incentive contract, a time consistent s effort is determined by the first order condition: δ T p'(e C * )[u '(H ) u'(l)] = c'(e C * ) The agent trades off the marginal return to effort (higher probability of higher pay) against the marginal cost of effort. For a time inconsistent individual, the optimal effort differs between period 1 and period 0. Consider period 1, who actually makes the effort choice. His first order condition is: 13 A hyperbolic discount factor d(τ) = (1+ ατ) -γ/α, where α captures deviations from exponential discounting (Lowenstein and Prelec 1992), will satisfy this property. Quasi-hyperbolic discounting (see Laibson 1997) or more broadly, present bias (see Benhabib et al 2007) satisfies these assumptions with one possible caveat. A quasihyperbolic function has d(τ) = βδ τ so that d(1) d(τ+1) = βδ and = δ for τ > 0. So the rate of time preference is d(0) d(τ) strictly decreasing between τ= 0 and τ= 1 and the same between other periods. The conflict is purely about now versus later with no conflict in future periods. For our purposes we are interested in conflict even in future horizons: for example we will look not just at a payday effect but also affects as the payday gets closer. 14 In writing this utility function we are assuming that the agent only consumes pay when it is given. If agents have access to perfect credit markets and suffer from no other psychological biases (e.g. mental accounting) then consumption utility would not depend on the actually date of pay. 7

8 d(t )p'(e 1 * )[u'(h ) u'(l)] = c'(e 1 * ) Comparing these formulas already produces a simple but useful prediction. Suppose the payday comes closer (T goes down). For a time consistent agent, since we are interested in paydays moving closer by at most a week, and δ 1 there should be no noticeable changes in e * C. 15 In contrast, for a time inconsistent agent, since we are assuming that d(1) < 1 by a noticeable amount, effort will be decreasing in T. This leads to the simplest prediction: e 1 * Prediction 1: As the lag between effort and compensation decreases, a time inconsistent agent will supply greater effort. In contrast, there will be no noticeable changes in effort provision if the agent is time consistent. This of course refers to the effort level that ends up being chosen by period 1. Period 0 desires a different effort level. Period 0 has the first order condition: d(t + 1)p'(e * 0 )[u '(H ) u'(l)] = d(1)c'(e * 0 ) The key difference between period 1 and 0 is that in period 1 the benefits and costs of effort have relative weight d(t ) d(t + 1) whereas in period 0 they have relative weight > d(t ) 1 d(t ) 1. As a result period 1 chooses less effort than 0 would like: e * * 0 > e 1. This is the heart of the time inconsistency problem. Now consider how changes to the contract affect the agent s utility. Suppose we change the payment in the low state L. For the time consistent worker: U 1 C L ( ) = δ T [1 p(e * C )](u '(L)) + e * C L δ T p'(e * C )[u '(H ) u '(L)] c'(e * C ) = δ T [1 p(e C * )](u '(L)) > 0 This is intuitive. For a time consistent agent, a marginal change in pay in the low state has only an income effect. The incentive effect can be disregarded because of the envelope theorem the agent was already equating marginal cost and benefit of effort. As a result, a decrease in L lowers utility and an increase raises utility. Notice a very similar result follows for the impact on time 1 utility for the time inconsistent worker. This is because as far as 1 is concerned marginal costs and benefits of effort have been equated. So for an inconsistent agent at time 1, a change in L also only has an income effect: U 1 I L = d(t )[1 p(e * 1 )]u'(l) > When presenting results in Section VII, we return to this assumption and explicitly calibrate the exponential discount rate that is implied by our empirical results. 8

9 Not so for period 0. As far as she is concerned, the marginal costs and benefits of effort have not been equalized. So the impact of a change in L on incentives must also be considered: U 0 I ( ) L = d(t + 1)[1 p(e * C )](u'(l)) + e * 0 L d(t + 1)p'(e * 0 )[u'(h ) u'(l)] d(1)c'(e * 0 ) As before the first term is positive: change L and income changes. The second term, however, the incentive effect is negative. With 0 s utility weights, the net benefit of reducing incentives is negative on utility. This is easiest to see for a reduction in L. For a time consistent and a time inconsistent in period 1, the effect of this marginal reduction is just the effect of a loss of income. A time inconsistent in period 0 suffers this income effect but he also values positively the increased effort that period 0 will now exert. Thus, if the agent s self-control problem is sufficiently severe, in period 0 utilty may rise when L falls. In essence, she may prefer a dominated contract one that holds payment for high output fixed and pays less for low output. Although this decreases expected earnings for any given effort level, it can raise overall utility by bringing effort provision closer to desired levels. This is the essence of commitment value. Prediction 2a: A (sophisticated) time inconsistent agent may prefer a dominated contract that increases the marginal returns to effort. In contrast, a time consistent agent would never prefer such a contract. Since a dominated contract is potentially appealing solely because of its ability to reduce distortions on effort: Prediction 2b: Providing a (sophisticated) time inconsistent agent the option to select a dominated contract will increase effort, output, and earnings. Of course, lowering L would incentivize both time consistent and time inconsistent agents to work harder. However, since a time consistent agent would never select a dominated contract, providing the option of such a contract should not affect effort if there is no self-control problem. Note that a direct implication of our model is that there will be a correlation between our first two sets of predictions. Time consistent workers will not be affected by the timing of compensation, nor will they select dominated contracts. In contrast, all time inconsistent workers will increase effort as the lag between effort and compensation narrows; in addition, a subset of these workers (those that are sophisticated) are the only ones that would select a dominated contract 16 : 16 Note that paydays should lead to production increases among all time inconsistent workers, whereas only time inconsistent sophisticates should demand commitment. In a model that incorporates naivete as well as sophistication, this prediction presumes a sufficiently strong correlation between the level of the self-control problem and the degree of sophistication. In addition, the continuity of time means that workers may face selfcontrol problems even on paydays (e.g. in the morning) or on days when they have chosen a positive target (e.g. after reaching the target). We do not, therefore, have priors on which set of treatments will produce stronger effects or how they will interact. 9

10 Prediction 3: An agent that is affected by the timing of compensation will be more likely to select and benefit from a dominated contract. When will time inconsistent sophisticates be most likely to prefer the dominated contract? Trivially in our model period 0 could choose a dominated contract, whereas period 1 would not. If we expand our model to allow for s+1 periods prior to effort the term in front of the incentive * "e effect 1 d(t + s +1) would become " d(1) and would therefore be increasing in s. In other "L d(t + s) words, demand for the dominated contract would increase with the horizon between the choice and effort period. Inconsistent individuals are more likely to choose a dominated contract for next week than for tomorrow Prediction 4: A time inconsistent agent will be more likely to prefer a dominated contract farther in advance of the effort period. An easy implication of this observation is that the principal can use the incentive scheme to do better. Consider the case of a risk netural agent where u(w) is linear. In this case the optimal incentive scheme for a time consistent agent is L=0 and H =1, that is the agent owns the output. From before, we know that for a time inconsistent agent, less effort will be realized under this * scheme than 0 would like. Following the previous notation let e 1 be the effort when the agent owns the output. Suppose the firm perturbs this scheme by x so that: L = p(e * 1 )x and H = 1 + (1 p(e * 1 ))x. It is easy to see that the cost to the firm for a small perturbation x is zero. The income effect is also zero. But the incentive effect on period 1 is positive. As a result, we can see that this scheme generates positive net benefits to both firm and worker. The worker would prefer and the firm would weakly prefer it (this can be made a strong preference by letting the firm share in some of the revenues. This leads to the following implication: Prediction 5: A time consistent agent has maximum productivity when he owns the output. For a time inconsistent agent, however, will produce more with an employer providing a different incentive scheme than him simply owning the output. In traditional agency theory, the firm provides insurance. Here we can see the firm can increase productivity as well. 17 In a quasi-hyperbolic model, this prediction would be more trivial: there will be demand for commitment for tomorrow but no demand for commitment today. The horizon of commitment would not otherwise since the model is exponential except for today versus tomorrow. 18 Our stylized model predicts that workers will only demand targets in advance of the effort period. However, the continuity of time means that even in the morning of the workday, workers may value targets for example, because the morning self wants the afternoon self to work hard. Thus, our prediction is simply that there will be less demand for the dominated contract in the morning than the evening, not that there will be none. 10

11 We do not directly test this implication but we see it indirectly when we examine nonlinear (dominated) incentive contracts and examine impacts on productivity and workers willingness to choose them. The derivation of this prediction implicitly shows another prediction. Time inconsistency generates incentive schemes that give super-normal incentives: workers will be incentivized at a rate that is larger than the actual impact of effort on output. 19 III. Experiment Design III.A. Ideal Workplace Features The experiment took place with full-time data entry workers, who were recruited and hired in an area that is one of India s major data entry hubs. Workers used data entry software to type information from scanned images into fields on their computer screen. To control for quality, we measured accuracy using dual entry of data, with manual checks by separate quality control staff when there were discrepancies. These are standard practices in the data entry industry. Workers were paid piece rates for production. The specific piece rate schedule depended on the contract assignment (see Section IV.C.) but all contracts were functions of accurate fields entered. We imposed no additional requirements on behavior. Employees were free to select their work hours they could arrive, take breaks, and leave when they wanted. There was also no penalty for absences. To mitigate potential incentives from long-term contracting, workers were explicitly told they were being hired on a temporary basis for a one-time job. Together, these features ensured that worker incentives were governed solely by the piece rate contracts. The project ran for approximately one year. Wages from the job constituted workers primary source of earnings to the best of our knowledge. Several features of this context make it an appealing place to test the model above. This constitutes an ideal task for individual piece rate payment. First, as in an agency model, effort speed of typing and care to prevent errors is unmeasurable here. But we can measure output, which is imperfectly correlated with effort for example, some groups of forms are harder, Second, there are no unobservable aspects of output which might suffer under incentives. Quality here is completely captured by accuracy, which we measure carefully as described above. As a result, when we find changes in production, we can be sure that it also reflects changes in profitability to the firm (Holmstrom Milgrom). Third, production is largely or completely individualistic. This means that incentive schemes do not need to be concerned about production externalities. 20 Fourth, the incentives to work come primarily from the piece rate contracts. There are, for example, less reputational concerns or potential for promotion to drive effort since workers are hired for a one-time job. Of course there may be some residual career concerns and we discuss how this would affect our findings in Section VIII. The incentives in the contract also summarize the pay consequences of effort: there are no penalties for being late or leaving early. Such penalties would have complicated the work effort calculus. Finally, work effort is largely independent across days. High effort on a particular day likely does not change the cost of effort 19 Both these predictions are seen in O Donoghue and Rabin (1996). The first is implicit in their Propositions 2 and 3 where the employer is able to increase probability of task completion. The second is a direct consequence of employers in their model willing to give sharper penalties for missing a deadline than project value. 20 This does not rule out social externalities in production. See Kaur et. al. (2010; 2011) for an analysis of those in this context. 11

12 on a subsequent day. Workers may be a bit more tired but these effects seem minor. As discussed in Section VIII below, this assumption allows us to randomize contracts daily and compare the difference in outcomes each day. Of course, this is an empirical question and we statistically examine inter-temporal substitution of labor supply in Section VIII. Finally, the workers here use this job as their primary source of earnings. As discussed above, existing experiments on inconsistency in labor supply involve small stakes. In our context, where workers are reliant on earnings for their livelihood, self-control problems (and their magnitude) are more informative about the extent of such problems in other labor supply contexts. The length of the experiment also helps boost external validity since workers have numerous chances to see the different contract treatments. With these considerations in mind, we turn to describe the experiment design. III.B. Payday treatments Self control arises in our model because of the costs of effort are borne now while its benefits occur in the future. A direct implication of this is that the self control problem diminishes as the benefits of effort become more immediate. In the work context, the date of the payday can generate this kind of variance. If paydays are on Friday, effort on Friday produces more immediate rewards than effort on Tuesday. To test this prediction, we randomized employees into three payday groups Tuesday, Thursday, and Saturday. One-third of workers were assigned to each group at the beginning of the study, and these assignments determined which day of the week each worker was paid. For example, on Tuesday evening of each week, employees in the Tuesday payday group were paid for work completed since the previous Wednesday. Randomly assigning paydays removes other reasons that specific days might impact effort. In this design, the same day is a payday for some (randomly chosen) workers and not for others. As a result, we can identify the effect of aligning compensation with effort by comparing production on paydays with production on non-paydays. This delivers our test of the model s first prediction: 21 Test 1: Production will increase as the payday gets closer. Note that even time consistent preferences would generate the prediction that workers will work harder as the payday approaches. Since this is due to a daily discount rate, however, workers would need to have incredibly high discount rates to exhibit statistically measurable payday spikes. We return to this point below when presenting the payday treatment results. III.D. Contract treatments We also use demand for certain types of contracts to test for time inconsistency. Specifically, we focus on two types of contracts. The first is a linear control contract that paid a piece rate wage 21 In a quasi-hyperbolic model, where time horizon is defined by a day, one can make a sharper prediction: production will only be higher on the payday itself. Of course if the time horizon is longer than a day or if discounting is hyperbolic (and not quasi-hyperbolic) this sharp prediction will fail. 12

13 of w for each unit of production. The second is a nonlinear dominated contract that imposed a production target. Under this latter contract, workers received the piece rate of w if they met the target, but only received w/2 for each entered field if they fell short of the target. As shown in Figure 1, the control contract dominates the commitment contract in earnings for any given production level, earnings are always weakly higher under the control contract. These contracts form two of our four treatment cells: i) Assignment to Control. Employees were assigned to the linear control contract. ii) Assignment to a Target. Employees were assigned to the dominated contract, with an exogenously chosen production target. The imposed target was selected from three target levels low, medium, or high. 22 The control treatment will serve as a baseline for comparison. The target assignment does not tell us about self-control but does ensure that all workers would learn about the target contracts and how they functioned. A time consistent individual should always prefer the control contract when given the choice this is the heart of our test. In contrast, a time inconsistent worker might prefer the dominated target contract. She knows she will work less than she intends to and the dominated contract may help prevent this shortfall. By increasing the cost of failing to meet the production target, she will be far less likely to stop before reaching it. As a result, a worker who is sophisticated about her time inconsistency problem may prefer a positive target because it enables her to achieve higher realized earnings. In contrast, a time consistent worker derives no benefits from the dominated contract and should always prefer a target of zero. Thus, if workers are time inconsistent (and sophisticated about their time inconsistency problem), we should see a demand for targets. To test these ideas, we imposed two choice conditions in which workers chose their target level. They had the option of choosing a target of 0, which is equivalent to selecting the control contract. The difference between the two treatments is the timing of when the choice had to be made: iii) Evening Choice. Workers selected their preferred target the evening before the workday. iv) Morning Choice. Workers selected their preferred target the morning of the workday. A sophisticated person with a sufficiently severe self-control problem should prefer the dominated contract and benefit from its provision: Test 2a: At least some workers will demand the dominated contract by selecting positive targets when assigned to the choice treatments (treatments (iii) and (iv)). Test 2b: Being assigned to choice treatments (treatments (iii) and (iv)) will increase production and earnings relative to being assigned to the control contract. 22 For about half the randomization period (mostly in the first half of the study), the Target Assignment treatment consisted of assignment to low or medium targets only. Assignment to the high target was added later, as worker production levels increased. These target levels are explained in greater detail below. 13

14 The payday and contract treatments provide two separate tests for time inconsistency in the same population of workers. The payday and choice treatments should each affect workers only if they are time inconsistent. This has a direct implication: outcomes under our first two tests should be correlated: 23 Test 3: Workers that show the biggest production increases on paydays should be more likely to demand and benefit from the dominated contract. Our model demonstrates that demand for the dominated contract depends on the time horizon workers will be more likely to bind future selves to hard work. We use variation in the timing of when the target choice must be made to test this prediction: 24 Test 4: Demand for targets will be higher under Evening Choice than Morning Choice. As discussed in Section III, Prediction 4 may be mediated by uncertainty. As we describe below, workers faced two types of uncertainty that affected production and which were at least partially resolved in the morning the network speed in the office and time of arrival to the office. In Section VII, we examine the mediating role of these shocks in influencing how demand for targets changes with the time horizon. We randomized contract treatments daily at the individual level. We used a balanced design, where every worker received each of the four contract treatments in random order exactly 25 percent of the time over every 8-day or 12-day work period. 25 This ensured that each of the four treatment cells had an equal number of observations, both within each worker and across the entire sample. The vector of assignments was independent across workers. As an example, Table 1 displays the contract assignments for 5 workers in the sample over a 24-day period. Daily variation enables us to generate a large number of data points within the study period. This will give us sufficient power for a richer set of analyses, such as looking at trends in behavior over time. 26,27 23 Note that paydays should lead to production increases among all time inconsistent workers, whereas only time inconsistent sophisticates should demand commitment. Thus, Test 3 presumes a sufficiently strong correlation between the level of the self-control problem and the degree of sophistication. In addition, the continuity of time means that workers may face self-control problems even on paydays (e.g. in the morning) or on days when they have chosen a positive target (e.g. after reaching the target). We do not, therefore, have priors on which set of treatments will produce stronger effects or how they will interact. 24 Our stylized model predicts that workers will only demand targets in advance of the effort period. However, the continuity of time means that even in the morning of the workday, workers may value targets for example, because the morning self wants the afternoon self to work hard. Thus, our prediction is simply that there will be less demand for the dominated contract in the morning than the evening, not that there will be none. 25 During the period when the Target Assignment treatment consisted of assignment to only low or medium targets, randomizations were on an 8-day cycle: 2 Control; 2 Target Assignment (1 low and 1 medium target); 2 Evening Choice; and 2 Morning Choice. During the period when the Target Assignment treatment consisted of assignment to low, medium, and high targets, randomizations were on a 12-day cycle: 3 Control; 3 Target Assignment (1 low, 1 medium, and 1 high target); 3 Evening Choice; and 3 Morning Choice. The proportional weight on each of the four treatments therefore remained 25% at all times. 26 A drawback of daily randomization is the potential for bias from intertemporal substitution in effort. In section XII below, we empirically test for this concern. 27 Note that daily randomization reduces the likelihood of potential Hawthorne effects. It is unlikely that workers would persistently alter their behavior each day for a year in response to knowledge of their treatment status. 14

15 In addition to the above predictions, our design provides the opportunity to test whether heterogeneity in treatment effects is predicted by correlates of self-control that have been posed elsewhere in the literature. These include preference reversals in estimated discount rates; workers own assessments of their self-control problems; failed attempts at quitting addictive behaviors; and education and IQ. In Section XII, we will discuss how potential alternate explanations for our predictions for example, workers choose targets in order to avoid external commitments can be ruled out using the experiment design and results. IV. Context and Protocols IV.A. Recruitment and Training Workers were data entry operators in an Indian data entry firm. The office was located in a small city, in a region that is one of the country s major data entry hubs. Employees were recruited through the standard procedures used by the firm with which we worked from the pool of resumes submitted by walk-ins to the firm and solicitations via posters and announcements in surrounding villages. Applicants were required to have completed tenth grade education and be at least eighteen years of age. Employees were hired in order of application. Upon joining the firm, workers received about 2 weeks of training. This included technical instruction on the data entry software, production task, and other aspects of computer usage. They were also trained on the two types of incentive contracts and the four contract treatments. During the initial part of training, workers were paid a flat stipend of Rs. 100/day while they learned the task. Trainees then worked under assignment to the control contract. At the end of the training period, they were assigned to the dominated contract for two days under the low and medium targets, respectively. This gave them the opportunity to observe their production under both types of incentive schemes before beginning contract randomizations. 28 IV.B. Production Task Workers entered information from scanned images into fields on their screen (see Figure 3). Once a worker finished entering data from an image, the software automatically sent the data to a central server and fetched the next image. This meant workers could not select the images on which they worked. Output was measured as the number of accurate fields entered. The data entry software displayed both the total and accurate number of fields entered so far that day (with about a 15 minute delay), so employees always had real time information on their own output. Workers faced some uncertainty in production due to shocks. Two types of shocks are particularly relevant in our context. First, the office experienced network speed fluctuations that impacted productivity. Some computers were more sensitive to these fluctuations than others. As a result, workers were randomly assigned to seats in the office and these assignments changed every 1-3 weeks. Second, many employees commuted from surrounding villages using buses and 28 The training period for some workers (particularly those that were the first to joined the project) lasted longer than 2 weeks. However, the structure of the training remained the same, regardless of duration. 15

16 trains, with some traveling up to two hours in each direction. Those from more remote locations faced increased uncertainty in morning arrival times and therefore production. IV.C. Payday Treatment Workers received their wages in cash on their assigned payday. Once they finished work for the day, they reported to the office manager, who computed and paid out their earnings for the previous week (including that day). If employees were absent on their payday, they could collect their owed earnings when they returned to work at no penalty. 29 IV.D. Contract Treatments As described above, workers were paid piece rates based on output. The control contract paid Rs for each accurate field entered, regardless of production amount. The dominated contract paid Rs per accurate field if the worker met the day s production target, and Rs per accurate field otherwise. 30 Under the Assignment to a Target treatment, workers were assigned to low, medium, and high targets. These were set at 3,000, 4,000, and 5,000 accurate fields, respectively. In the first month of randomizations, these corresponded to the 30 th, 50 th, and 70 th percentiles, respectively, of worker production under the control contract. 31 During the last month of contract randomizations, we changed these levels to 4,000, 5,000, and 6,000 accurate fields to correspond to increases in worker production over time. Before leaving work each day, employees were required to report to an office staff member in a separate area of the office. At that time, they were told their contract assignment for the next day. For example, employees were informed of Wednesday s assignment on Tuesday evening. If the assignment was Evening Choice, they also selected their target for Wednesday at that time. If the assignment was Morning Choice, then they selected their target upon arriving in the office Wednesday morning. This exchange was confidential and took place away from other workers. IV.E. Attrition The office held 64 data entry operators at a time. Due to employee turnover, 111 workers participated in the experiment. 32 When an employee quit, the management staff hired a replacement from a database of persons that had submitted applications for the job. As in the initial recruitment, workers were hired in order of application date. The gender composition of employees was kept fixed if a female quit, the worker hired to replace her was a female. Each 29 The office was closed every alternate Saturday; in those weeks, the Saturday pay group was paid on Friday. More generally, if the office was closed on a scheduled payday due to a holiday, the payment day for members of that group was moved to another day (almost always the day immediately before or after). Such adjustments to the payday schedule were announced in advance and also marked on the posted monthly office calendar. 30 Under both contracts, workers were also paid a small flat daily wage of Rs. 15 conditional on attendance. Since the base wage constituted about 8 percent of mean daily earnings, the overwhelming majority of worker compensation was tied explicitly to output. Data entry firms in the region commonly use an incentive structure that combines a base wage with payment tied to production. Earnings of workers in the experiment were at par with or slightly higher than those of workers in other data entry firms in the region. 31 As noted in footnote XX, we did not assign employees to the high target for most of the first half of the study given its level of difficulty workers were hired over the course project. However, 27 of these workers quit before completing training or were only employed outside the 11-month randomization period. As a result, the sample consists of 111 workers. 16

17 new worker inherited all the assignments of his or her predecessor payday group, vector of contract assignments, and seating assignment. The payday and seat assignments took effect immediately. New hires began their scheduled contract assignments after completing training. IV.F. Office Structure and Timeline The office was open each day from 8:45 am to 6:30 pm, five to six days per week except holidays. Employees could choose when they worked, except for two 15-minute periods each day when work activity was halted for server maintenance. In accordance with the norms of the firm with which we worked, employees were given tea in an outside area at 11 am and 3:30 pm each day. Workers could select the length of their tea breaks and lunch breaks. They were also free to check , play computer games, or leave the office at any time. The project ran for 15 months. During the first 2 months, the management staff established protocols, recruited subjects, and trained the new hires. After this, the contract and payday randomizations ran for 4 months. There was then a 2-month break while the office underwent changes to the data entry software and task. During this time, workers were generally not paid the standard piece rates and there were no contract randomizations. The contract and payday randomizations then resumed for another 4 months. In the final 3 months of the project, we ran endline activities and surveys. We did not randomize workers into the four contract treatments during this time, but we continued to adhere to the payday assignments. Thus, the contract treatments ran for an approximate total of 8 months and the payday treatments for 11 months. 33 V. Data V.A. Summary of Worker Characteristics and Survey Responses Panel A of Table 2 displays participants characteristics. Most workers were males (74%). 63 workers reported their age on resumes or elsewhere in their application. These workers ranged in age from 19 to 38 years, with a mean of 24. We collected information on educational attainment and experience during baseline surveys administered to 101 of the 111 workers. 34 Employees had 13 years of education on average. The majority had taken a computer course and had an address prior to joining the firm. Panel B of Table 2 reports the results of tests administered to new hires. We tested comprehension of the contract treatments using a written quiz, which required workers to calculate earnings for various production and target levels under the different contracts. 35 The mean score was 93%, indicating that workers had a strong grasp of the incentive contracts and treatments. In addition, workers were given the Raven s Matrix IQ test and the Digit Span IQ test forwards and backwards in both English and the local language. 33 There were small stretches of time when operations were interrupted due to problems like breakdown of the electricity generator a reality of business operations in developing countries. These periods never lasted longer than 2 weeks, and often only lasted 1-3 days. Randomizations were suspended during these periods. There were also 3 instances where we paused randomizations while workers were retrained before a change in the data entry task. 34 In this and other information presented in Table 2, some of the employees that were hired in later stages of the project were not surveyed because of clerical oversight workers who had difficulty with the quiz were retrained and given a new quiz before beginning randomizations. Their final quiz score is reported. 17

18 At the end of the project period, we administered a discount rate exercise to the 58 workers employed at that time. We asked workers to trade off amounts of cash between different time horizons. There were three sets of cash tradeoffs: Rs. 20 vs. Rs. 24; Rs. 50 vs. Rs. 57; and Rs. 100 vs. Rs Workers selected among the options in each set of tradeoffs under two time horizons: the smaller reward today vs. the larger reward three days from now; and the smaller reward in 14 days vs. the larger reward in 17 days. This 3-by-2 design generated 6 sets of questions that comprised the discount rate questionnaire. To ensure workers took the exercise seriously, at the end of the questionnaire, we randomly selected one of the six questions and paid workers the amount selected on the appropriate date (e.g. Rs. 57 in 17 days). 36 Panel C of Table 2 reports the results from this exercise. The typical worker chose the smaller reward 31% of the time (i.e. in about 2 out of the 6 questions). In addition, on average workers displayed a preference reversal where they chose the smaller immediate reward when deciding between today and 3 days from now, but chose the larger delayed reward when deciding between the same amounts 14 vs. 17 days from now 17% of the time. We also administered an endline survey that asked workers to assess their self-control problems and constraints. 37 We use the responses to compute several proxies for self-control, which are reported in Panel D. Nine questions asked workers to agree or disagree with statements about self-control such as, Some days I don t work as hard as I would like to. For each of these questions, responses ranged between values of 1 ( Disagree strongly) and 5 ( Agree strongly ). 76% of workers agreed or agreed strongly with the above statement, while 40% agreed or agreed strongly with: At the end of the day, I get tempted to leave work earlier than I would like. Our first proxy for self-control is an index that equals the mean of each worker s responses to all 9 self-control questions. The typical worker has a self-control index score of Our second selfcontrol proxy follows previous studies in asking respondents about failed attempts to quit addictive habits (e.g., Chabris et al 2008a). We asked male workers whether they had tried and failed to quit chewing tobacco, smoking, or drinking; 12% of male workers answered yes for at least one of these behaviors. 38 Finally, we conducted a Factor Analysis using responses from all the survey questions. The factor with the largest eigenvalue, which explains the greatest amount of variation in responses, is defined primarily by self-control questions. We compute each worker s self-control factor score to form our third proxy for self-control. 36 Notice that because the discount rate questionnaire was administered in a firm, it did not suffer from two major concerns that can bias results towards finding impatience and preference reversals in other contexts. The first potential concern is trust issues. Respondents may be more likely to select an immediate reward today rather than deferring to a future period (when the payment may not materialize), but perceive no difference in risk in the long horizon (14 days vs 17 days). A second potential concern is fixed costs. If respondents have to show up to a location to collect their payment, they may prefer to select the immediate reward today to avoid the fixed cost of showing up at a future period, but show patience when choosing for the longer horizon since the fixed cost cannot be avoided (Chabris et al 2008b). Neither of these is a concern in our context. Workers are full-time employees who have been paid regularly for months. In addition, since they have to show up for work anyway, there is almost no difference in fixed costs between the choices. 37 This portion of the survey was administered to 70 workers those that were currently employed plus 12 previous workers that we were able to contact. The survey also included questions about peer effects, which are discussed in Kaur et. al. (2011). 38 We did not pose the questions about addictive behaviors to female workers due to Indian cultural norms. 18

19 The endline survey also asked workers to agree or disagree with 4 statements about their external constraints. Summary statistics for these questions are reported in Panel E. For example, 61% of workers agreed or agreed strongly with If I miss one bus or train, the next one I can take is much later. As before, we compute a constraints index for each worker by taking the mean of her responses to the 4 constraints questions. The typical worker has a constraints value of We will employ these various proxies to explore heterogeneity in the analysis that follows. Using a large number of survey questions to create statistics raises concerns about the potential for data mining. We deal with this in a few ways. First, we compute index measures by taking means of all survey questions that pertain to each topic self-control or constraints. Of course, there may remain concerns about judgment in deciding which questions pertain to a topic and the decision to take means rather than some other statistic. Second, we employ Factor Analysis to create a self-control proxy. This relies completely on a defined mathematical approach and is therefore immune to researcher judgment. Finally, as will become clear below, we supplement these survey-based measures with ones that rely on demonstrated behavior such cash rewards chosen in discount rate exercises to create additional proxies. By using measures from multiple sources and comparing them in the analysis side-by-side, we mitigate concerns that the explanatory power of any one proxy is driven by data construction rather than underlying fundamentals. V.B. Outcome Variables and Observations Used in Analysis The key outcomes of interest are worker output and commitment demand. As defined above, output is measured as the number of accurate fields entered in a day. We have 2 measures of commitment demand under the Choice treatments an indicator for whether a positive target was chosen and the target level chosen. Table 3 reports summary statistics for measures of these outcomes. Column (1) provides means and standard deviations for 8-month period during which the contract randomizations were run. This constitutes the main Analysis Sample when both contract and payday treatments occurred simultaneously and is comprised of 102 workers and 8,423 observations. Attendance was 88 percent and mean production conditional on attendance was 6,094 accurate fields per day in the Analysis sample. When assigned to the Choice Treatment, workers chose a positive target 35 percent of the time. The mean target level chosen was 974 this reflects that the chosen target was zero 65 percent of the time. Column (2) reports statistics for the 11-month period during which the payday randomizations were run. This constitutes the Full Payday Sample and is comprised of 111 workers and 11,744 observations. While mean attendance was the same in the full payday period, mean production was somewhat higher (6,433 fields). This difference stems from the fact that the payday period ran for 3 additional months, and therefore reflects production increases over time by workers. For consistency of analysis, we will use the Analysis Sample throughout the empirical analysis that follows. To demonstrate that restricting analysis to this sample does not impact the payday results, all tables pertaining to the payday treatments also contain a column showing regression estimates for the Full Payday Sample. 19

20 Table 4 shows how the 8,673 observations in the Analysis Sample are spread across the treatment cells. Due to the balanced randomization design, observations are split approximately evenly across the three payday group assignments. Similarly, each of the four contract treatments constitutes 25% of the sample. The minor differences in observations in each cell are caused by worker turnover vacancy time until worker replacement, variation in start day from first payday, and the random order of contract assignments. VI. Results VI.A. Payday Effects on Production (Test 1) Do employees work harder on paydays or closer to paydays? To answer this, we estimate an OLS model of the form: Y i,t = α 0 + α 1 Payday i,t + α 2 W i + α 3 D t + α 4 S i,t + µ i,t. (1) Payday i,t is an indicator for whether worker i had a payday on date t. W i, D t, and S i,t are a vectors of worker, date, and seating assignment dummies, respectively, and are included to increase the precision of the payday coefficient estimate. µ i,t is the residual error. Y i,t is output for worker i on date t. As discussed above, we measure output as the number of accurate fields entered in a day. However, recall that workers are absent from work on some days. If we simply dropped absentee observations from the sample, it would constitute a form of attrition that could bias the payday coefficient. As a result, on days when workers are absent, we define their output as 0 and include these observations in the sample. This is appropriate because worker earnings which are tied to production through the piece rate wage are 0 when absent. We follow this convention throughout the paper in all regressions where the dependent variable is output. Note that this estimation strategy (and the others we use below) makes an implicit assumption: separability of labor supply across periods. Of course nonseparability is possible. For example, the cost of effort could depend on effort in recent periods workers may be tired if they worked hard recently. Alternatively, nonseparability might arise through income effects: high earnings in one period could reduce the willingness to work in later periods. This of course does not create a bias since treatments are randomly assigned from day to day. But it does create an interpretational issue: if harder work on paydays is offset by less work on later days, then the interpretation of the coefficient is very different. Similarly, the magnitude of the impact of dominated contracts would be interpreted differently. The net effect on worker welfare could be much smaller (arbitrarily smaller) than the main coefficient on today s output. When we present the contract treatment results below, we describe a test for nonseparability. Our results there suggest that in fact separability is a reasonable assumption: workers appear to show little interday substitution of labor. Since workers are randomly assigned to the three payday groups, and the paydays are spread evenly across the workweek, Payday i,t will be uncorrelated with any other factors that could 20

21 create weekly cyclicality in work effort: E[Payday i,t µ i,t ]=0. α 1 is therefore interpretable as the causal impact of paydays on production. 39 Columns (1)-(4) of Table 5 provide estimates of the payday effect on output in the Analysis Sample. Column (1) estimates the specification in regression model (1). It shows that workers produce 215 fields more on average on paydays than non-paydays. This effect is significant at the 1% level and is equivalent to 4% of mean daily production. Column (2) adds controls for lagged production. The coefficients indicate that there is substantial serial correlation in production, and the inclusion of the lagged controls therefore reduces the magnitude of the payday coefficient to 140 fields. Columns (3) and (4) add dummies for the days leading up to paydays to check for patterns in the payday cycle. They indicate that workers increase production on the days leading up to their paydays. Production is particularly high on paydays and the 2 days before paydays. For example, column (4) shows that workers produce 428 fields (8% of mean production) more on their payday and 539 fields (10% of mean production) more on the day before their payday than they do 6 or more days before their payday. Figure 4 graphs the regression coefficients from column (4) of Table 5. The days that immediately follow a payday, and are therefore furthest from the next payday, are when employees are least productive. Production then rises steadily over the pay cycle. 40 Employees are likely to face self-control problems in not only how hard they work while in the office, but also in the decision of whether to show up to work. Columns (5)-(8) of Table 5 directly test for the effect of paydays on attendance in the Analysis Sample. Column (5) estimates a Linear Probability Model in which the dependent variable is a binary indicator for attendance and the covariates are as shown in regression (1) above. The results show that workers are 4.7 percentage points more likely to come to work on their payday. Columns (6) and (7) add controls for the days leading up to the payday. These show substantive spikes in attendance the day before the payday and on the payday itself. As a check on our estimates, in column (8), we estimate a probit model where the dependent variable is attendance and the independent variables are the same as those in column (7). The probit marginal effects estimates are similar to the OLS estimates in column (7). In columns (9) and (10) we rerun the specifications in columns (4) and (7), respectively, on the full Payday Sample of 11,744 observations. The results are essentially the same as those above. 39 Note that α 1 is not interpretable as the policy impact of paying workers everyday. In the experiment, we do not vary the frequency of paydays they always occur once a week. α 1 is estimated off the difference in output on weekly paydays versus non-paydays. 40 Figure 4 shows that production steadily increases as the payday approaches. It does not, however, pin down the shape of this increase one could fit a linear, convex, or concave curve through the confidence intervals. Some time inconsistency models predict convexity. At the extreme, β-δ and dual-self models where time periods are defined as days would predict that all the effect should be concentrated on the payday itself; our results would seem to refute this special case. However, this is not necessarily true under alternate ways of conceptualizing time periods or the horizon of β and δ. More generally, predictions in any model will be sensitive to how the discount function is defined and how time periods are specified. As discussed above, we are interested in testing core predictions that are common across self-control models. We therefore focus here on the qualitative prediction that production should be higher closer to paydays a result that bears out strongly in the data. 21

22 This demonstrates that the payday effect also held strongly in the entire 11-month experiment sample. How should we assess the magnitude of the payday effects? One useful benchmark is the production impact associated with an increase in the piece rate wage in the same population of workers. Toward the end of the experiment (after contract randomizations were finished), we randomized workers into two piece rate wages: Rs (their usual wage) and Rs per accurate field. Each worker received each wage 5 times over a 10-day period in random order, with approximately half the workers in the office assigned to each wage within any given day. During this period, workers produced 11% more on average under the 25% wage increase an elasticity of What does this imply for the payday effects estimates? In Table 5, the production difference between paydays and the beginning of the pay cycle is 428 fields, or 8%. Thus, the production increase on paydays is roughly comparable to the impact of raising the wage by 18% between the beginning and end of the pay week. 41 The long horizon of the study provides an opportunity to test for heterogeneity in individual treatment effects of paydays. Since we observe each worker s production on paydays and nonpaydays repeatedly over the course of many weeks, we can estimate each worker s mean payday impact by interacting her fixed effect with the payday dummy: Y i,t = φ 0 + φ 1 Payday i,t + φ 2 W i + φ 3 Payday i,t *W i + φ 4 D t + φ 5 S i,t + φ 6 Y i,t-1 + φ 7 Y i,t-2 + µ i,t. (2) Payday i,t *W i is a vector of interactions between the vector of worker dummies and the payday dummy, Y i,t-1 and Y i,t-2 are lag production controls, and all other variables are as defined above. If the treatment effect of paydays is homogeneous across workers, then the coefficients in the φ 3 vector should be 0. The p-value of the F-test of joint significance of the coefficients in φ 3 is We therefore conclude that there is heterogeneity in the extent to which workers are impacted by paydays. Could the payday effects in Table 5 be explained by reasons other than time inconsistency? One potential concern is that the payday effect reflects impatience, not time inconsistency. However, since payments are made weekly, the maximum gap between paydays and non-paydays is about 6 days. Could exponential discounting explain a payday spike of 8%? Suppose the worker has an extremely myopic yearly discount rate of 100%, which implies a discount factor of 0.5. Denoting the worker s incentive scheme by f(e), she chooses effort to equate marginal effort costs today with discounted marginal benefits 6 days from now: c (e) = 0.5 6/365 u ()f (e). Moving payment 6 days forward (so that it occurs on the workday) is equivalent to increasing marginal incentives by 1.002, or 0.2%. A 7% production increase from a 0.2% increase in the piece rate implies an output elasticity of 35. Such an elasticity is hard to justify. It implies, for example, that one could double worker output by increasing incentives by 6%. Thus, even if we allowed for extremely myopic agents, exponential time discounting could not explain the payday effect. 41 The above calibration assumes a constant elasticity of effort with respect to the wage. Of course, this may not be exactly true in practice. The calibration also pertains to the effects of temporary, not permanent, wage changes. We abstract from these issues since the above exercise is not intended as a policy evaluation of competing ways in which firms can impact worker production. Our purpose here is only to provide a simple benchmark with which to compare the payday effects. 22

23 A variant of impatience is credit constraints. Credit constraints combined with utility shocks (e.g. need for cash for health care for a relative) or investment shocks (e.g. a high return investment breaks down and must be repaired) could generate a high return on capital. This would then generate a sharp need for money which might induce individuals to work harder to meet these shocks. To fit our facts this explanation would need to be further refined. First these shocks would need to be extremely concentrated so that if cash is not delivered on a particular day there is little return. Otherwise the shocks would generate high work effort throughout not just on the payday. For example if there is a shock that occurs three days before payday we would see high effort all the way up to and after the payday as long as there is a high demand for capital that week and the following weeks. It is only if shocks unmet that day have no rewards. It is hard to imagine such date specific shocks happening at such a high frequency as to generate payday effects every week. Second, even if they existed, such shocks would need to be of sufficiently high return as to generate the implicit discount rates we have calculated above. While returns to capital are thought to be high among poor and middle-income populations in India (Dasgupta 1989; Banerjee and Duflo 2010), there are no estimates we know of which generate returns equivalent to these. 42 A more practical concern would be that there are transaction costs to showing up at work and individuals may simply be coming to collect their paychecks. This would generate an attendance effect on paydays that would look in our regressions as a production effect since our production variable includes a zero for attendance. We cannot address this problem by simply focusing on production of those who attend since that is a selected sample whose mean production could be different. 43 To resolve it we note a prediction of a (continuous) hyperbolic model: individuals should show increasing production as the payday gets closer. The payday effect need not be concentrated on the payday itself. In fact, the results in Table 5 show this to be true. There is an effect on production in the day before the payday and a smaller one the day before that. These pre-payday effects could not be explained by a transaction cost of attendance. We therefore interpret the results in Table 5 as strong evidence that time inconsistency leads workers to supply greater effort on paydays than non-paydays Note that under time inconsistency, the payday results do require the existence of credit constraints. If time inconsistent workers had access to perfect credit markets, they d borrow against future income and increase production just before loan payment was due. If loans were long-term, we may not see a weekly payday effect. However, it s reasonable to assume the population from which our workers are drawn doesn t have easy access to long-term credit (particularly non-collateralized credit against future wages). In addition, we can assume that a decent portion of what is earned in one paycheck is consumed before the next paycheck. 43 In Appendix Table A1, we show regression results where the dependent variable is production conditional on attendance. However, for the reasons stated above, these results are difficult to interpret. Since attendance is higher closer to paydays due to the self-control benefits of paydays, the composition of workers in attendance is correlated with distance from payday. For example, when workers face negative productivity shocks like sickness, they may be more likely to come into work on paydays than non-paydays (i.e. the higher cost of effort is more likely to be justified when benefits of effort are more immediate). Alternately, if low ability workers also have greater selfcontrol problems, the mean ability of the worker pool will be lower on paydays. Such selection problems could lead average production conditional on attendance to be lower on paydays than non-paydays because different groups of workers are being compared across days. 44 We find another piece of evidence consistent with our model. In India, festivals involve large expenditures by households (Banerjee and Duflo 2006). Under convex effort costs, time consistent workers should not show large 23

24 VI.B. Demand for and Treatment Effects of Dominated Contracts (Test 2) We now turn to explore worker behavior under the contract treatments. Recall that under the two Choice treatments Evening Choice and Morning Choice workers had the option to bind their future selves to positive production targets. We interpret selection of a positive target as revealing demand for a dominated contract with self-control benefits. Table 6 presents summary statistics on take-up of the dominated contract a binary indicator for whether the worker selected a target above zero when assigned to a Choice treatment. The first row of the table summarizes choice behavior when the worker was present both the day before and the day of Choice the assignment. In these observations, workers were notified of contract assignment the evening before as per the experiment protocols, selected their target at the appropriate time, and were actually bound by their chosen target because they were present at work. Column (1) shows that take-up was 35% in the Analysis Sample as a whole. In Column (2), we compute worker-level means for the 101 employees that were assigned to Choice at least once in the Analysis Sample the mean of the workers choice rates is 36% in this sample. The observations when workers were present constitute a selected sample. In the second row of the table, we include all 4,193 observations. We define target choice to be 0 if a worker was absent the day of Choice assignment. This is sensible if we think that not showing up to work indicates a preference for a 0 production level. For consistency, we also define target choice to be 0 if a worker was absent the day before Choice assignment, and therefore did not receive notice of assignment as per protocols and could not select a target if assigned to Evening Choice. These conventions provide a lower bound on the level of demand for dominated contracts by workers. Under this definition, the take-up rate across observations is 28%, as is the mean of the workers take-up rates. In Figure 5, we plot a histogram and kernel density estimate of worker take-up rates. As in the first row of Table 6, we consider only those observations where workers were present the day before and day of Choice assignment. The figure reveals substantial variation in demand for targets. Some workers (16% of the sample) always chose a target of 0. The bottom quarter of the distribution chose positive targets less than 10% of the time. The top quarter of the distribution chose positive targets at least 60% of the time. How should we assess the magnitude of the targets chosen by workers? The penalty for missing one s target is substantial: half of one s piece rate earnings for the day. If shocks generate uncertainty in output (see Section XX), then choosing overly aggressive targets can be extremely costly either due to the financial penalty, or from having to achieve the target even on days when the cost of effort turns out to be high. When deciding on a target level, time inconsistent workers would weigh the motivational benefits to their future selves against the probability of incurring these costs. Given this, it would be surprising, for example, for workers to select targets at their median production level this would amount to signing up for a 50% chance of a production spikes in the days leading up to festivals (which are perfectly foreseeable); time inconsistent workers, however, would be expected to show such spikes. Indeed, we find that average production increases by 15% in the week prior to major festivals (significant at 1%). 24

25 very costly penalty. Indeed, we would expect targets to be set at levels that are achievable in most states of the world under the effort level preferred by the period 0 self, but have motivational impact because there is a positive chance of incurring the penalty if the period 1 self does not exert sufficient effort. To explore this intuition, we estimate the probability that workers production would fall below their chosen targets if they had been assigned to the control contract. We use the counterfactual of the control contract since it represents what the period 1 self would produce if not tied to a positive target by the period 0 self. This statistic is computed as follows. For observations where workers were in attendance, we estimate a regression of production on worker, date, and computer fixed effects; lag production controls; payday distance dummies; contract assignment dummies; and log experience. For each of the 1,168 observations in which a worker was assigned to Choice, selected a positive target, and was present, we predict the worker's production under the control contract on that day using the estimates from the above regression. To this predicted value, we add the worker's vector of residuals from the above regression to arrive at a vector of potential production values, which we fit to a lognormal distribution. Evaluating the CDF of this distribution at the chosen target level provides an estimate of the probability that the worker would have missed her chosen target under the control contract. Row 1 of Table 7 presents the summary statistics from this exercise. The mean probability workers would have missed their chosen targets (if they had been assigned to the control contract) is 8.3% across the 1,168 observations. The mean of the worker averages for this statistic is 8.8%. Figure 6 presents the distribution of worker averages; it shows considerable variation among workers in the aggressiveness of their selected targets. For about 60% of workers, the mean target miss probability was 5% or less. The top quarter of the distribution selected targets associated with 15% or higher probability of incurring the penalty. In Row 2 of Table 7, we show the proportion of times workers actually failed to reach their targets under Choice (conditional on choosing a positive target). The mean worker missed her chosen targets 2.5% of the time. These findings suggest that targets were selected at non-trivial levels. Using hourly data, we present additional descriptive evidence that workers considered the targets a meaningful goal. If workers use the targets to push their future selves to exert a minimum level of effort, we would expect a decrease in production immediately after the target is reached. We check for this by examining how workers behaved in the hours around when they reached the target threshold. Of course, such an analysis cannot have a causal interpretation. The decision to select a target, the level at which it is set, and the hour in which it is reached are all endogenous outcomes. However, such an analysis is empirically interesting as it helps us understand how workers production behavior moves in relation to their (self-chosen) targets. In Table 8, we examine production behavior around when targets are reached. The dependent variable in each regression is hourly production (the output of worker i on date t in hour h) where hours are defined as calendar hours (8-9 am, 9-10 am, etc). In columns (1)-(3), we examine behavior in observations in which workers selected a positive target under Choice and were present. In column (1), we begin by regressing hourly production on a dummy that equals 1 for the calendar hour in which the worker hit her target that day and 0 otherwise, along with worker, 25

26 date, and computer fixed effects. 45 Workers produce 597 fields (70% of mean hourly production) more in the hour when they reach their target than in other hours of the day. In column (2), we add a series of dummies for the hours immediately before and after the target was hit; the omitted category is 4 or more hours before the target was reached. The estimates show that on average, production increases steadily, spikes in the calendar hour in which the target is reached, and then declines after the target is reached. This pattern can be clearly seen in Panel A of Figure 7, where we graph these coefficients along with confidence intervals. In column (3), we include a covariate that measures worker i s mean production in hour t on days the worker was assigned to the Control contract. Conditional on workers usual production in a given hour, production increases slightly in the hour before the target is reached, spikes in the hour when it is reached, and then declines. In both specifications (2) and (3), we reject that production in the hour when the target was reached equals production in the hour after it was reached at the 1% significance level. 46 As a benchmark, we repeat this analysis for the Assignment to Target observations in columns (4)-(6) of Table 8 and Panel B of Figure 7. We find similar patterns in how workers behave around the point at which they hit the targets we exogenously impose on them. Thus, on average, workers do appear to reduce production once they achieve their targets both when the targets are self-chosen and when externally imposed. We now turn to quantify the benefits of the dominated contracts by testing whether their provision actually translates into greater daily output and earnings. We begin with an Intent to Treat analysis of the impact on production of the contract treatments: Y i,t = λ 0 + λ 1 Choice i,t + λ 2 Target i,t + λ 3 W i + λ 4 D t + λ 5 S i,t + λ 6 Y i,t-1 + λ 7 Y i,t-2 + µ i,t (3) Choice i,t is an indicator for whether worker i was assigned to one of the Choice treatments on day t; Target i,t is an indicator for the Target Assignment treatment. As before, Y i,t measures production, W i, D t, and S i,t are a vectors of worker, date, and seating assignment dummies, respectively, and Y i,t-1 and Y i,t-2 are controls for lagged production. Due to random assignment, E[Choice i,t µ i,t ] = E[Target i,t µ i,t ] = 0. The key coefficient of interest is λ 1. It represents the reduced form average treatment effect of offering workers the option to take-up the dominated contract relative to assigning them to the control contract. 45 Production is defined as 0 in hours when workers did not work. In cases where a worker did not manage to reach her target during the day, the hour when she reached it is coded at 7 pm (which is when the office closed for the day). 46 Note that there are compositional issues in the sample off which the distance from the target dummies are estimated. For example, workers who reach their targets at 5 pm are never observed 3 hours after the target is reached the value of the 3 hours after target dummy will always be 0 in these cases. This raises selection problems that make it difficult to interpret the coefficients on the distance dummies. The comparison in which we are most interested is between the hour when the target is reached and the hour right after. We observe workers in the hour after the target is reached in 98.7% of cases; so while the selection problem undermines the validity of this comparison, it is unlikely to drive the difference in the coefficients. In Appendix Figure A1, we graph the proportion of worker-days observed in each distance category. 26

27 In columns (1)-(4) of Table 8, we estimate variants of the above regression model for the Analysis Sample. In Column (1) we regress production on a dummy for assignment to Choice; dummies for assignment to low, medium, and high targets under Target Assignment; worker, date, and seating assignment fixed effects; and lag production controls. The results show that assignment to Choice increased production by 111 fields on average (2% of mean production). This effect is significant at the 10% level. Being assigned to the low target did not significantly increase production in the sample overall. In contrast, assignment to the medium and high targets led to average production increases of 213 fields (4% of mean production) and 335 fields (6% of mean production) respectively. These estimates are significant at the 5% level. In column (2), we test separately for the treatment effect of Evening Choice and Morning Choice treatments. We find workers produce 150 fields (3% of mean production) more under Evening Choice than under assignment to Control. This effect is significant at the 5% level. The estimated effect of Morning Choice is 73 fields, but is not significant. In columns (3) and (4), we limit analysis to those observations in which workers were assigned to the Control or Choice treatments; the Target Assignment observations are excluded. The results are similar to those in the first two columns. In columns (5)-(8), we check whether the contract assignments impact attendance. We estimate a Linear Probability Model in which the dependent variable is the binary attendance indicator and the independent variables are the same as in columns (1)-(4), respectively. Unlike the payday treatment, which led to substantial attendance increases, the contract treatments do not appear to impact whether employees show up to work on average. The above estimates show that the Choice treatments increase production by about 2%, implying a local average treatment effect of approximately 6%. In Section VII.A, we used random increases in the piece rate wage to obtain a production elasticity estimate of Together, these estimates imply that for time inconsistent sophisticates, the production impact of providing the dominated contract is comparable to a 14% increase in the piece rate wage. Finally, as we did for the payday treatment effects, we check for heterogeneity in treatment effects of Choice. Using only Control and Choice observations, we regress production on: a dummy for assignment to Choice; worker fixed effects; interactions of each worker fixed effect with the Choice dummy; and date fixed effects, computer fixed effects, and lag production controls. The p-value of the F-test of joint significance of the interaction coefficients is We interpret this as evidence for heterogeneity in the individual treatment effects of Choice. As discussed in the payday results above, our interpretation of treatment effects estimates relies on the assumption that labor supply is separable across periods. In the case of the contract treatments, if hard work under targets increases the cost of effort on future days, this would change our interpretation of the estimates in Table 8. To test for this concern, we exploit the random ordering of contract treatment assignment. If effort costs are not independent across periods, then today s production should be lower if the worker was assigned to a high effort treatment last period (such as choice or assignment to the high target). In Appendix Table A1, we explore various specifications in line with this approach and find no evidence that there is dependence in effort between periods. For example, in column (1), we regress production on 27

28 dummies for yesterday s contract assignment. We cannot reject that being assigned to choice or a target (relative to being assigned to control) has no impact on the next day s production. In fact, the coefficient estimates in all specifications are usually positive (though insignificant). In contrast, if there were intertemporal substitution, these coefficients should be negative. Another potential concern arises from the block randomization design of the contract treatments. Since workers are assigned to each treatment a fixed number of times within each 8- or 12-day period, treatment assignment on a given day is correlated with the probability of future treatments within each block. For example, conditional on receiving Choice today, a worker is less likely to receive Choice (and more likely to receive the Control contract) tomorrow. This could induce a mechanical correlation that affects what is being captured by the coefficients on the treatment assignment dummies a concern that would not arise under iid randomization. We test for this concern in Appendix Table A3. For each observation in the Analysis sample, we compute the probabilities of receiving each contract assignment in that worker-day; these probabilities are determined by the worker s previous assignments in that randomization block. We then directly control for these probabilities in a regression of production on the contract dummies. The results indicate that the assignment probabilities have little predictive power and their inclusion has little impact on the estimated treatment effects. This is confirmed by an F-test of joint significance of the probability controls the test p-value is The above results confirm the predictions of our model. Workers showed substantive demand for the dominated contract (at least 28% by our most conservative estimate). They selected targets at non-trivial levels and showed responsiveness in production around the point at which they reached targets indicating that the targets were meaningful. Most importantly, offering workers the option to select the dominated contracts led to increased output and earnings. VI.C. Correlation Between Payday and Contract Effects (Test 3) The payday and contract results each support the predictions of time inconsistency models. As noted above, we also see substantial heterogeneity among workers in the payday and contract effects. We now explore this heterogeneity by checking whether the two sets of effects are correlated whether those that are most affected by the payday treatments are also those that select and derive the greatest benefits from the dominated contracts. We first construct a measure of the payday effect for each worker. We define the payday differential as: Payday Difference = (Mean Production on Paydays) " (Mean Pr oduction on Nonpaydays) Mean Production in Sample We compute this differential for each worker in the Analysis Sample, using only those observations in which the worker was assigned to the Control contract. We then define a worker as having a high payday difference if her differential is above the mean differential in the sample. 47,48 47 In the results presented in Section VII.A., we see substantial production increases in the days leading up to paydays. However, we construct our self-control proxy by differencing the worker s mean production on paydays 28

29 In Table 9, we test whether workers with high payday differentials are more likely to demand the dominated contract. We use two outcome variables to measure take-up: the target level chosen (which includes targets choices of 0) and a binary indicator that equals 1 if the worker chose a positive target. Following the conventions described above, in this and future regressions, we define both take-up dependent variables as 0 on days that the worker was absent the day before or day of assignment to Choice (see Section VII.B). In each column, we regress the dependent variable on the high payday differential dummy and controls. On average, workers that are more affected by paydays select a target that is 353 fields higher and are 13.8 percentage points more likely to select a positive target. These coefficients correspond to a striking 47% and 49% of the mean target level and take-up rate, respectively, and are both significant at the 1% level. In Table 10, we explore whether workers with high payday differentials derive more benefit from the contracts. In column (1), we provide the estimates of the average treatment effects of Assignment to Choice and Assignment to a Target for reference. In column (2), we estimate: Y i,t = θ 0 + θ 1 Choice i,t + θ 2 Choice i,t *HighDiff i, + θ 3 Target i,t + θ 4 Target i,t *HighDiff i + θ 5 W i + θ 6 D t + θ 7 S i,t + θ 8 Y i,t-1 + θ 9 Y i,t-2 + µ i,t (4) HighDiff i is the indicator for whether worker i has an above average payday differential. We are interested in the coefficients on the interactions θ 2 and θ 4. If the workers that are most affected by paydays are also those that benefit the most from the dominated contracts, then these coefficients will be positive. The results indicate that this is indeed the case. The average treatment effects of Choice and Target Assignment for workers with low payday differentials are statistically indistinguishable from 0. In contrast, compared to the production of low differential workers under each of the treatments, high differential workers produce about 480 fields more under Choice and Target Assignment on average. These coefficients are significant at the 1% level and their magnitudes correspond to 9% of mean production. Since in our sample as a whole, an additional year of education is associated with a 466 field increase in production on average, providing high payday difference workers with simply the option to select commitment leads to production increases comparable to a one year increase in education. In addition, using our benchmark production elasticity of 0.44 (see section VII.A), the local average treatment effect for high difference workers is comparable to a 48% increase in the piece rate wage. In column (3), we explore how these treatment effects vary on paydays versus non-paydays. The payday difference statistic measures the extent to which a worker is affected by aligning the compensation period with the effort period. It constitutes an imperfect proxy for the level of a and non-paydays for two reasons. First, our statistic relies on the one prediction regarding paydays that is common across various models of time inconsistency (see footnote 36XX). Second, there is no objective way to incorporate information from earlier parts of the pay cycle into our proxy since the theory does not provide clear guidance on earlier days, any attempt to do so would be motivated by our ex-post findings in Table 5XX. 48 Note that since we can only compute this statistic for workers that were assigned to the Control contract on both paydays and non-paydays during their employment, it cannot be computed for some workers that were in the sample for shorter periods of time. This reduces our sample size for this analysis from 8,423 to 8,240 observations. 29

30 worker s self-control problem. On non-paydays, when the effort and compensation period are not aligned, time inconsistency is likely to create greater distortions on effort for workers with greater self-control problems. We therefore see that it is these workers that benefit the most from the provision of targets on non-paydays. Specifically, on non-paydays, high payday difference workers produce 735 fields (14% of mean production) more under Choice than Control. At the same time allowing for heterogeneity in the extent to which both types of treatments help a worker overcome her time inconsistency problem the high difference workers are the ones that are most helped by paydays. They therefore have less need for dominated contracts to solve the self-control problem on paydays than those workers for whom paydays don t produce large benefits. As a result, we see that on paydays the Choice treatment is relatively more beneficial for workers with a low payday difference. The estimated coefficients on Target assignment in column (3) tell a similar story. In columns (4)-(6), we repeat this analysis for attendance as the dependent variable. While the average treatment effects on attendance are indistinguishable from zero, we see in column (5) that this masks substantial heterogeneity. The Choice treatment does not affect attendance of low payday difference workers, and increases the attendance of high difference workers by 5.8 percentage points. This effect is significant at 1%. If workers face a self-control problem in not just how hard they work in the office, but also in the decision to show up to work (as implied by the payday results), then this effect is consistent with a model of time inconsistency with sophisticated agents. There is a sizeable fixed cost of attendance for example, up to a 2-hour commute. Workers that are sophisticated enough to pick targets are also sophisticated enough to know that in the absence of a target, they will be tempted to exert low effort. Consequently, they're more likely to go in when they can select targets, because they know their earnings on those days will justify paying the attendance fixed cost. This is consistent with the results in column (6), which match the trends in column (3). Choice boosts attendance for high difference workers especially on non-paydays, whereas on paydays the impact of Choice is relatively lower for high difference workers than low difference ones. 49 In Appendix Table 5, we perform the analogous exercise for payday treatment effects we examine whether those most impacted by the contract treatments have higher increases in effort on paydays. We find that treatment effects under Choice and Target Assignment are highly predictive of payday spikes. However, take-up of dominated contracts is not predictive of payday effects. VI.D. Demand for Dominated Contracts: Timing of Contract Offer and Uncertainty (Test 4) Our model predicts that since workers should be more willing to bind future selves to hard work, they should be more likely to select targets under Evening Choice. We begin by testing whether there are differences in demand for the dominated contract in Evening versus Morning Choice: 49 Is the large production effect of Choice on high difference workers driven completely by the impact on attendance? In Appendix Table A4, we estimate the average treatment effects of Choice on production and attendance as 395 fields and 4.4 percentage points, respectively, for high payday difference workers. For these workers, mean production conditional on attendance is As a simple calibration, 5581*0.044 = 245 < 395. This implies that the entire effect of Choice on Production for these workers is not driven by attendance increases. In column (3), we regress production conditional on attendance on the contract treatment dummies. While the coefficients are positive and significant, they are difficult to interpret (see footnote 38XX). 30

31 Take-up i,t = δ 0 + δ 1 Eve i,t + δ 2 W i + δ 3 D t + δ 4 S i,t + µ i,t. (5) Eve i,t is an indicator that equals 1 if worker i was assigned to Evening Choice on date t and equals 0 if the worker was assigned to Morning Choice. Take-up i,t measures take-up of the dominated contract by worker i on date t. As before, we use two measures of take-up the target level chosen and a binary indicator for whether a positive target was chosen. Since contract treatments are randomly assigned, E[Eve i,t µ i,t ]=0 and δ 1 is interpretable as the change in take-up caused by assignment to Evening Choice instead of Morning Choice. Table 11 provides estimates of δ 1. It indicates that in the sample as a whole, there is no discernible difference in demand for the dominated contract between the evening before and morning of the workday. However, as noted in Section III, this prediction assumes worker expectations are the same the evening before and morning of the workday. If employees face uncertainty in production that is resolved the morning of the workday, this prediction may not necessarily hold. We now turn to explore whether two sources of uncertainty faced by workers network speed shocks and commuting constraints mediate ex-ante demand for targets. We first document the extent of the volatility in production across days in the Analysis Sample. We regress the production of worker i on day t on a vector of worker dummies and controls for worker i's production on dates t-1 and t-2. For each date of the experiment, we compute the mean of the residuals from this regression. This gives us a measure of daily office production from which the variation explained by individual worker productivity and prior shocks has been netted out. Figure 5 plots the daily mean of the production residuals over the course of the experiment. The figure reveals substantive amounts of day-to-day volatility in production. An important source of this volatility is network speed fluctuations. These fluctuations affected the rate at which workers could send data entered from an image to the central server and retrieve the next image for entry. The wait time between images could range from one second to over five minutes. Since workers typically entered over 150 images per day, network speed was an important determinant of output. Some computers in the office were more sensitive to network fluctuations than others. Since we re ultimately interested in how uncertainty impacts commitment demand, we want to look at uncertainty as it s perceived by workers themselves. For this reason, we asked the office management staff to consult workers in identifying the set of computers that were perceived as more sensitive to network slowdowns. Management did not know the list would be used for this purpose. We first explore whether the computers identified as more uncertain are indeed more sensitive to network fluctuations. In Table 11, we estimate: e i,t = β 0 + β 1 BadComputer i,t + β 2 ( 1 N 1 e j,t ) + β 3 BadComputer i,t * ( 1 N 1 j i j ie j,t ) + µ i,t (6) 31

32 e i,t is the production residual, as defined above. ( 1 N 1 j i e j,t ) is the mean of the production residuals across workers on day t, excluding worker i s own residual. BadComputer i,t is a binary indicator for whether worker i was assigned on day t to a computer that was sensitive to network fluctuations. Recall that workers are randomly assigned to seats in the office and these assignments change periodically there is therefore no correlation between computer quality and worker characteristics. The coefficient of interest is β 3. If bad computers are more sensitive to network fluctuations, then the production on these computers should vary more with the daily fluctuations captured in ( 1 N 1 j i e j,t ). We therefore expect β 3 to be positive. Column (1) of Table 12 shows estimates for this regression model. Workers on bad computers have lower levels of production on average (139 fields). They are also more sensitive to network fluctuations the coefficient on the interaction is positive and significant at the 5% level. In column (2), we compute the mean of production residuals within each day separately for good and bad computers (again excluding the worker s own residual), and interact these with the bad computer indicator. The mean shock experienced by bad computers (which experience the brunt of the network shocks) is predictive of the production residual for good computers as well, since they also are connected to the office network. However, as expected, it is especially predictive of the production residual for bad computers (the interaction is positive and significant at 5%). In addition, once the mean of bad computers is controlled for, the mean residual of the good computers has little predictive power. In Figure 9, we show that the mean residuals of bad computers tracks the mean office-level shocks closely, much more so than the mean of good computers. Note this does not mean that workers on bad computers face less ex-ante variance in production; rather, the figure shows that ex-post, the shocks experienced by bad computers are more closely correlated with the mean shock in the office on any given day. What are the implications of this for the Evening versus Morning Choice decision? When workers arrive to the office in the morning, they receive new information on the network speed and can use this to inform their target choice. This information is especially valuable for workers on bad computers, since network shocks will greatly impact their productivity. Thus, not only do workers on bad computers face more volatility from network shocks, they are those workers for whom a substantive portion of production uncertainty is likely to be resolved by the morning. To test this, we estimate: Take-up i,t = γ 0 + γ 1 Eve i,t + γ 2 BadComputer i,t + γ 3 Eve i,t *BadComputer i,t + γ 4 W i + γ 5 D t + µ i,t (7) All variables are as defined above. We predict γ 1 >0 and γ 3 <0. We test these predictions in Table 13. Columns (1)-(2) provide some evidence that workers assigned to bad computers are less likely to demand targets in the sample as a whole. In columns (3)-(4), we estimate the above OLS regression model. Our predictions hold strongly in these results. When assigned to a good computer, selected targets are 168 fields higher on average in the evening than the morning. However, when assigned to bad computers, selected targets are 82 fields lower on average in the evening than the morning. These are sizeable magnitudes (equivalent to 22% and 11%, respectively, of the mean target levels chosen by workers in the sample overall). The results are similar if we use our binary measure of commitment demand. Workers on good computers are 6.6 percentage points more likely to pick a positive target in the evening than the morning. In 32

33 contrast, those on bad computers are more likely to demand commitment in the morning than the evening. A second source of uncertainty faced by workers stemmed from external constraints on time. For example, as discussed in Section V, workers that lived in more remote areas faced long and uncertain commute times. These impacted morning arrival time and therefore how much the worker could produce in a day. In addition, some workers had duties or binding constraints on time outside the office. This made it more difficult to absorb production shocks. For example, if the network unexpectedly slowed down, it would have been harder for workers with external constraints to stay late in the office to ensure their targets were met. Much of the uncertainty from these sources was resolved by the morning of the workday by then employees knew their arrival time at the office and would have had a better sense of duties at home for that day. Thus, we expect workers with greater external constraints to be relatively more likely to demand the dominated contract in the morning. To test this prediction, we estimate: Take-up i,t = ϕ 0 + ϕ 1 Eve i,t + ϕ 2 Constraint i + ϕ 3 Eve i,t *Constraint i + ϕ 4 D t + ϕ 5 S i,t + µ i,t (8) Constraint i is a measure of the external constraints faced by worker i. In Table 14, we present results for two binary measures of the constraint variable. In columns (1)-(2), the variable measures workers response to a question in the endline survey that asked them to agree or disagree with the statement: The bus/train schedules really impact whether I can get to work on time because if I miss one bus or train, the next one I can take is much later. The high constraint indicator takes a value of 1 if the worker s response was Agree Strongly and 0 otherwise. The results indicate that workers with more uncertain commute times select targets more often under Morning Choice than Evening Choice, and the opposite is true for workers with less uncertain commute times. Recall that the endline survey asked four questions related to external constraints. For the analysis in Columns (3)-(4), we compute a Constraint Index for each worker by averaging his or her answers to the four questions. The high constraint indicator equals 1 if the worker s constraint index score was above the sample mean score and equals 0 otherwise. The results in columns (3) and (4) are similar to those in Columns (1) and (2), respectively. Collectively, our results tell the following story: In the sample as a whole, there is little difference in demand for the dominated contract between the evening before the workday and the morning of the workday. However, this masks substantial heterogeneity that arises from uncertainty. Uncertainty in production interacts strongly with the timing of contract offer to influence take-up levels. Specifically, we see evidence that when uncertainty is low, take-up is higher in the evening than the morning. In addition, we find strong evidence that when uncertainty is high, this pattern reverses and take-up is higher in the morning after a substantive portion of the uncertainty has been resolved. VI.D. Other Results In this section, we present auxiliary results of interest: the aspects of worker behavior that changed in response to the treatments; an analysis of whether other correlates of self-control 33

34 posed in the behavioral economics literature predict treatment effects; and evidence of learning and trends in effects over time. VI.D.1 Channels of Impact What aspect of worker behavior changes in response to paydays and contract assignment? In Table 15, we investigate whether workers work longer hours or work more intensively. Panel A focuses on behavior around paydays. The dependent variable in columns (1) and (2) is workday length the number of minutes between a worker s first login and last logout in the data entry software within a day. When workers are absent, workday length is defined as 0 minutes worked. In column (1), we regress workday length on a payday dummy and controls. On average, workday length is 20 minutes longer on paydays than non-paydays. In column (2), we add dummies for distance from payday. Workers spend about 40 more minutes in the office on their payday and the day before their payday than they do 6 or more days before their payday. This constitutes a 10% increase in mean workday length and is significant at the 1% level. In columns (3) and (4), we look for evidence on whether employees work more intensively on paydays. In column (3), we regress production on a payday dummy, standard controls, and a quadratic function for the number of minutes spent logged into the data entry software during the day. The payday coefficient measures the mean change in production on paydays once the variation in output explained by minutes worked has been partialled out. Since minutes worked is an endogenous variable and is correlated with the payday dummy, the payday coefficient cannot have a causal interpretation. Rather, this approach is an accounting exercise if the payday dummy is positive, this would mean that output increases on paydays are not fully accounted for by changes in minutes worked on paydays. In regression (3), the payday coefficient is statistically indistinguishable from 0. In column (4), we add dummies for distance from payday. As before, we see no evidence of increases in production intensity on paydays. However, the coefficients on the three days immediately before the payday are positive and significant. In Panel B of Table 15, we repeat this analysis for the contract treatments. In column (1), we regress workday length on dummies for Choice assignment, Target assignment, and controls. On average, workday length is 9 minutes (2% of the mean) longer when assigned to Choice than when assigned to the Control contract. In column (2), we add interactions with the High payday difference indicator. High difference workers spend almost a half hour more in the office when assigned to Choice. This corresponds to 6% of mean workday length and is significant at the 1% level. In columns (3) and (4), we regress production on the contract treatment indicators, controls, and a quadratic function of minutes logged into the data entry software. The results in column (4) indicate that high payday difference workers produce more under Choice and Target assignment, even after the variation in production explained by minutes worked has been partialled out In Appendix Table A6, we repeat this analysis for only those days when workers are present. See the discussion in footnote 38XX on the selection issues caused by conditioning on attendance. 34

35 Overall, the results in Table 15 indicate that employees work longer days in response to paydays and the contract treatments. In addition, they provide some suggestive evidence that all the production increases we see in response to paydays and the contract treatments are not accounted for by increases in minutes worked. Thus, it seems that workers respond on both margins duration as well as intensity of work. VI.D.2 Other Sources of Inter-personal Heterogeneity The literature on psychology and economics has proposed a range of correlates of self-control problems. Through the endline surveys, we collected information on some of these correlates. In Table 16, we examine the extent to which these correlates predict demand for the dominated contract and explain heterogeneity in treatment effects for the contract and payday treatments. Each column conducts this analysis for a different potential correlate of self-control problems; the correlate analyzed in each column is specified at the top of that column. In Panel A, we show coefficient estimates from a regression of target level chosen on the correlate and controls. In Panel B, we regress the binary take-up indicator on the correlate and controls. In Panel C, we report estimates from a regression of production on: the correlate; dummies for Choice and Payday; interactions of each of these dummies with the correlate; and controls. In columns (1)-(3), we look at measures of self-control problems based on self-reports by workers in the endline surveys. The correlate in column (1) is the Self-control Factor, obtained from a factor analysis on the endline survey data. In column (2), we construct a Self-Control Index from the endline survey responses by averaging each worker s responses to the 9 selfcontrol questions in the endline survey. Both the Self-control Factor and Self-control index values have been de-meaned in the analysis. In column (3), we use self-reports of addictive behaviors by male workers. In this column, the correlate equals 1 if the worker said he had tried to quit drinking, smoking, or chewing tobacco and failed, and equals 0 otherwise. Each of these three columns shows similar results. These three correlates from the endline surveys positively predict demand for the dominated contract, and also positively predict treatment effects of the contracts. However, among these, only the coefficients on the Self-control Factor are generally significant. None of the correlates predicts the payday effect. In columns (4)-(5), we look at outcomes from the discount rate exercise, in which we asked workers to trade off cash rewards between different time horizons a standard way of testing for self-control problems in the literature. In column (4), we look at impatience. Our measure of impatience is defined as the proportion of times in the 6 questions the worker chose the smaller immediate reward rather than the larger delayed reward. In column (5), we look at preference reversals. Here, our self-control correlate is defined as the proportion of times a worker chose the larger immediate reward in the short horizon, but then displayed patience when choosing between the same amounts in the long horizon. As in the case of the Self-control Index and Addictive behaviors, these correlates do not appear to predict demand for the dominated contracts the coefficients in Panels A and B are positive but significant. Also as before, we see in Panel C that workers with greater self-control problems (as measured by these correlates) are less productive on average. We also see evidence that these workers benefit more from the contract treatments. For example, the coefficient on the interaction between proportion of impatient responses and Choice assignment is 706 and is significant at the 5% level. It indicates 35

36 that workers that show impatience in all 6 questions produce 706 more fields (13% of mean production) under Choice than Control. As before, the correlates do not seem to predict payday behavior. Finally, in columns (6)-(7), we look at two other correlates of interest education and IQ (the sum of the worker s scores on the Raven s Matrix and Digit Span tests). While these do not stem directly from predictions of time inconsistency models like the other correlates, we check whether they have predictive power. We largely find that they do not. Workers with higher education levels are more likely to demand the dominated contracts. For example, an additional year of education corresponds to a 2.9 percentage point average increase in take-up; this estimate is significant at the 10% level. In contrast, IQ does not predict take-up. In addition, neither education nor IQ predicts treatment effects of the contracts or paydays. Overall, we find some support that proxies of self-control posed in the literature correlate with behavior under the contract treatments. In contrast, none of these proxies is correlated with the payday effects. Thus, while these survey-based proxies have some predictive power, we find that the strongest predictor of effects under each set of treatments is workers behavior under the other set of treatments (see Section VII.C). VI.D.3 Time Trends in Treatment Effects Did treatment effects change over time? Could take-up of dominated contracts reflect mistakes by confused workers for example? Or alternatively, do workers learn either about the value of the dominated contracts or perhaps find other ways around their self-control problems? Examining trends requires dealing with attrition. Due to worker turnover, on any date of the project, the employees at the firm have participated in the experiment for varying lengths of time. As a result, looking at time trends over calendar days could provide a misleading picture of trends in worker behavior. For example, if workers select positive targets only during their first month at the firm, we could see persistent demand for the dominated contract over the entire eight-month period because new workers join each month. Therefore, to test for persistence in behavior, we first construct a measure of worker experience. We define experience i,t as the number of workdays worker i has been in the Analysis Sample on date t. 51 In Figure 10, we explore how demand for the dominated contracts evolved with worker experience for high and low payday difference workers. For each value of the experience variable, we compute the proportion of high difference workers that choose a positive target under assignment to Choice (and were present the day of and day before Choice assignment); these values are plotted in closed circles. The open circles plot the value of this statistic for the low difference workers. The figure shows little initial difference in mean take-up rates between high and low difference workers. However, as workers gain experience, we see a divergence. Over time, those that have the largest self-control problems (as measured by our payday 51 Recall that days during which workers are in training are not included in the Analysis Sample. As a result, experience i,t =1 on worker i's first day of contract randomizations. Note also that the experience variable suffers from selective attrition. 36

37 difference proxy) end up demanding the dominated contract at substantially higher rates than the workers that do not have large payday impacts. 52 We explore these trends more formally in Table 17. In columns (1) and (2), we regress each of our measures of take-up on the log of the experience variable and controls. In both columns, we cannot reject that mean demand for the dominated contract does not change over time in the sample as a whole. In columns (3) and (4), we add the high payday difference indicator and an interaction of high payday difference with log experience. The results are consistent with the trends in Figure 10. As low difference workers gain experience, they decrease their take-up of the targets; a 1% increase in experience is associated with about a percentage point decrease in take-up (significant at 5%). In contrast, we cannot reject that the demand among high difference workers stays constant over time. The F-tests of whether the log experience coefficient and interaction coefficient sum to zero are insignificant (with p-values of and in columns (3) and (4), respectively). As a result, at higher values of experience, the high payday difference workers exhibit substantially higher demand on average for the dominated contracts than the low difference workers. These results are consistent with a story in which workers initially try the dominated contracts. Over time, they continually receive opportunities to observe their production under targets both through Target assignment and potentially also when on Choice assignment. Those workers for whom the targets do not yield utility benefits stop selecting the dominated contract. In contrast, the workers with large self-control problems see that the targets are helpful, and continue to select them. Consistent with Test 3 of our model, this latter group of workers correlates with the group that is most affected by paydays. Next, in Table 18, we test whether the treatment effects on production persist over time. For reference purposes, in column (1), we first regress production on: log experience; dummies for Choice assignment, Target assignment, and Payday; and our vector of standard controls. As before, we define experience as the number of workdays the employee has been in the Analysis Sample. Not surprisingly, we see that production increases strongly with experience. The remaining results in column (1) are consistent with those presented in earlier tables. In column (2), we add interactions of log experience with each of the treatment variables of interest: Choice, Target assignment, and Payday. We are interested in the coefficients on the interactions. If treatment effects diminish over time for example, once the novelty of the treatments wears off then these coefficients will be negative. Instead, the results in column (2) reveal positive interaction coefficients. The interaction of log experience with Choice assignment is positive and significant at the 5% level. This is consistent with the findings in Table 17, which indicate that the workers that derive the largest benefits from the dominated contracts are the ones that are most likely to select them over time. In addition, the interaction with Target assignment is also positive and significant at the 10% level. The coefficient on the payday interaction is essentially 0, indicating that the payday effect is constant over time on average. 52 The figure also shows that there is variation in the level of take-up among both groups of workers over time. This is not surprising since the composition of workers assigned to Choice changes each day with the contract randomizations. In addition, day-to-day shocks (such as network speed fluctuations) impact take-up. 37

38 In column (3), we repeat this exercise using a different measure of experience: a binary indicator for whether the worker has been in the Analysis Sample for more than two calendar months. The coefficients Choice and Target interaction coefficients are positive (and insignificant). The interaction on the payday coefficient is now negative, but insignificant. In column (4), we check for persistence in the payday effect in the full 11-month payday sample and again find no change in effects over time. Together, columns (2) to (4) provide compelling evidence that the treatment effects of the contract treatments and paydays persist over time, and some evidence that the effects of Choice increase over time. Overall, we see that workers select and derive steady benefits from the dominated contracts throughout the experiment. Similarly, the production increases on paydays persist week after week. Given the long horizon of the study, our results imply that time inconsistency is a perpetual problem in the workplace. They lend credence to our view that many workplace features can plausibly be interpreted as arrangements that seek to solve self-control problems. VII. Alternate Explanations The results described in the previous section are consistent with the predictions of our model. We now turn to evaluate whether these results could plausibly be explained by reasons other than time inconsistency. While various alternate explanations may explain individual effects, we argue none can explain the full pattern of results: production increases on paydays; demand for dominated contracts and treatment effects of Choice; the correlation between the payday and contract effects; differences between evening and morning choice mediated by uncertainty; correlation of the contract effects with other measures of self-control; and time trends, including the differences in take-up over time among high and low payday difference workers. For concreteness, we apply this argument to several potential alternate explanations below. One alternate explanation for the contract effects is that workers use dominated contracts to avoid external commitments or pressure: by selecting high targets, employees can justify having to stay in the office. This enables them to get out of duties at home or avoid peer pressure from coworkers. For families, this would somehow require that workers cannot simply tell their family they have to work irrespective of actual contract: it is unclear how the family could ever know the details of the contract. This explanation is more plausible if it is a commitment against co-workers, who could plausibly learn about contract choice. It is less clear however how this could explain the payday effect. Why would external pressures lead one to work harder when the pay is coming today than when it is coming tomorrow? The financial returns are the same, only the timing matters. Moreover, it is hard to see why the demand for targets should be highest for those with the highest payday differentials, or with some of the other heterogeneity we find. Finally, it is unclear why the desire to get out of external commitments would be higher in the evening than the morning and then reverse under uncertainty. Another potential explanation is that workers select targets to signal ability to the employer. To mitigate this in the initial design, we advertised the job as a one-time employment opportunity. Of course that might not have worked fully. We feel this is still not a viable explanation for our results for a variety of reasons. First, it is not clear that demand for the dominated contract actually should serve as a positive signal. Since the employer observes production directly, there 38

39 is no reason to believe a worker that can achieve high production under the control contract should not appear more impressive than one that needs a dominated contract to increase output. Second, this explanation too would struggle to explain a payday effect. Finally, it would struggle to explain why the desire to signal ability should be different in the evening versus the morning, and be higher for workers with larger payday differentials. We therefore reject a signaling interpretation for demand for the dominated contracts. In addition, one may worry that demand for targets might be driven by confusion: workers may choose dominated contracts not out of a desire to bind themselves to hard work but out of a failure to understand that they are dominated by the piece rate contract. As described above, after training workers on the contract treatments, we administered a quiz that tested their comprehension of the contracts. The mean score was 93%, indicating that workers understood the contracts. 53 In addition, demand for these contracts does not decline with worker experience for the high payday difference workers; it seems unlikely that confusion could persist over such long horizons. It is also less clear how confusion could explain the evening versus morning differences and their relationship to uncertainty or the other findings. Turning to the payday results, an alternate behavioral explanation that could produce payday effects is income targeting. If workers target a fixed weekly income level, then small amounts of impatience or the realization of shocks could lead workers to backload effort closer to the payday. We can test for such behavior directly in the data. Income targeting implies a sharp decrease in marginal utility for income levels above the weekly target (see Camerer et. al. 1997). If a worker experiences an unexpected production increase today, putting her closer to her weekly goal, she will lower production in all subsequent days of the pay cycle to compensate. However, as we saw above in our test for intertemporal substitution, exogenous production increases caused by target assignment do not lead to production decreases on subsequent days (see Appendix Table A2). In addition, a targeting model delivers an even finer testable prediction: an unexpected production increase today will lead to a larger reduction in tomorrow s effort if the worker is closer to her payday, because there are fewer subsequent days over which the adjustment needs to be made. We test for this in Appendix Table A8. Under income targeting, the interactions in columns (2) and (3) should be negative. However, the interactions are positive (though largely insignificant). 54 Finally, an income targeting story that did not 53 In Appendix Table A7, we test whether quiz score is correlated with demand for the dominated contracts. If workers mistakenly chose dominated contracts because they did not understand the contract treatments, then we would expect quiz score to be negatively correlated with take-up. Instead, quiz performance positively predicts takeup, although the coefficients in columns (1)-(4) are insignificant. In addition, as noted above, education strongly predicts take-up of the dominated contract. 54 We also test an additional prediction of income targeting. Since the impact of day-to-day shocks is adjusted within the payweek to arrive at the weekly target, the variance in production among payweeks should be less than the variance in production among weeks defined according to some other arbitrarily cycle, such as calendar weeks. To check this, we compare production across workers payweek cycles with production across 4 artificial weekly cycles, created by shifting forward days from the worker s actual pay cycle. For example, for a worker assigned to the Saturday pay group, her true pay week is from each Monday to Saturday. The 4 artificial cycles for this worker would be from Tuesdays to Mondays, Wednesdays to Tuesdays, Thursdays to Wednesdays, and Fridays to Thursdays. For each worker, we then compute the standard deviation of weekly production across her actual payweeks and across each of the 4 associated artificial weekly cycles. On average, the standard deviation of weekly production for actual payweeks is The mean standard deviations for each of the 4 artificial cycles are lower 39

40 involve self-control problems would also be difficult to reconcile with demand for the dominated contracts and the other contract treatment results. We therefore conclude that the cyclicality in effort around paydays is not driven by weekly income targeting. VIII. Conclusion We use a field experiment to test for time inconsistency in labor supply among data entry workers. We use two treatments payday assignments and provision of a dominated contract with self-control benefits to test core predictions of time inconsistency models. Our tests provide strong evidence that self-control problems distort effort provision at economically meaningful magnitudes. What are the implications of our findings for explaining workplace arrangements? Take the basic prediction that the self-control problem increases as the returns to work are further in the future a prediction validated by the payday findings. This mechanism suggests a new look at a variety of naturally occurring production function differences. In agriculture in developing countries, should we view productivity in long-horizon crops differently from productivity in short-horizon crops? Might farmers choose shorter horizon crops as a commitment device? Might the move from farm work to formal sector work with regular pay have self-control productivity benefits? Self-control considerations also suggest a re-examination of various workplace practices. Might certain contract features (such as nonlinear contracts) be thought of as partly reflecting self-control benefits? Might discipline at the workplace or workplace rules be thought of as demand for features to help workers avoid the temptation to shirk? Might the organization of production itself, such as the presence of a boss and task division with deadlines serve to mitigate self-control problems? Incorporating time inconsistency into agency models may therefore deliver new insights into optimal workplace design, and augment our understanding of observed arrangements. For example, we often see the coexistence of many different arrangements for the same production tasks. Within developing countries, any given crop is grown by small individual landholders working their own farms, through tenancy arrangements, and also on large plantation farms with hired workers. During a harvest season, agricultural laborers in the same village may be paid piece rates for production, hired on fixed daily wage contracts, or enter into long-term seasonal contracts. Similarly, we see the same good, such as textiles, produced both at home by micro-entrepreneurs and also by workers employed in large factories. Such diversity exists everywhere, but is especially pronounced in poor countries today. It was also prevalent in the west in earlier stages of development, such as in early industrial Europe. Differences in these arrangements are typically understood through features of the production function, agency considerations, insurance demand by workers, and labor market institutions. While these considerations are undoubtedly important for explaining workplace contracts, our results indicate the role for an additional consideration self-control problems. Notice the arrangements above provide substantial variation in self-control benefits. They alter the lag between effort and compensation: this lag is small for a factory worker, whereas a microthan this, ranging from 1731 to Overall, all 5 standard deviation estimates are close to each other (within 5% or less from the payweek mean). This provides further support against weekly income targeting. 40

41 entrepreneur may have to wait weeks or months before receiving compensation through sale of her good. Similarly, the arrangements provide variation in discipline from an external boss: a factory manager ensures workers exert constant effort daily, while the micro-entrepreneur must solve the motivation problem on her own. Indeed, the process of development which entails movements from agriculture to manufacturing, or from cottage industry to factory work may increase labor productivity not just through technological innovation, but also because the work arrangements associated with these advances mitigate workers self-control problems. (See Clark 1994 and Kaur et. al for further exposition of this view). 55 These arguments are, of course, speculative. However, given that we find strong evidence that self-control problems distort worker effort at economically meaningful magnitudes, a closer exploration of these possibilities is warranted in future research. 55 For example, one of the major changes in the organization of production in economic history has been the transition from the putting out system (under which workers were paid piece rates according to work performed and could choose production levels and work hours themselves) to the more rigid workplace system that is the norm today (with features like assembly lines, production minimums, rigid work hours, and hefty punishments for even momentary lapses in behavior). One interpretation is that increases in capital since the industrial revolution place a premium on increasing labor productivity (Clark 1994). Finding ways of reducing worker self-control problems could be one response to this problem. Under this view, the rise of the factory and its associated disciplinary infrastructure was in part an attempt to solve self-control problems. 41

42 References Angeletos, George-Marios, David Laibson, Andrea Repetto, Jeremy Tobacman, and Stephen Weinberg The Hyperbolic Consumption Model: Calibration, Simulation, and Empirical Evaluation. Journal of Economic Perspectives, 15(3): Ariely, Dan and Klaus Wertenbroch Procrastination, Deadlines, and Performance: Selfcontrol by Precommitment. Psychological Science, 13(3): Ashraf, Nava, Dean Karlan, and Wesley Yin Tying Odysseus to the Mast: Evidence from a Commitment Savings Product in the Philippines. Quarterly Journal of Economics, 121(2): Bandiera, Oriana, Iwan Barankay and Imran Rasul Incentives for Managers and Inequality Among Workers: Evidence From a Firm Level Experiment. Quarterly Journal of Economics, 122: Banerjee, Abhijit and Esther Duflo The Economic Lives of the Poor. MIT Working Paper No Banerjee, Abhijit and Esther Duflo Giving Credit Where it is Due. Mimeo, MIT. Banerjee, Abhijit, and Sendhil Mullainathan The Shape of Temptation: Implications for the Economic Lives of the Poor. Mimeo, Harvard University. Baumeister, R.F., E. Bratslavsky, M. Muraven, and D.M. Dice Ego Depletion: Is the Active Self a Limited Resource? Journal of Personality and Social Psychology, 74: Benhabib, Jess, Alberto Bisin, and Andrew Schotter Present Bias, Quasi-Hyperbolic Discounting, and Fixed Costs. Available: Burger, Nicholas, Gary Charness, and John Lynham Three Field Experiments on Procrastination and Willpower. Available: Chabris, Christopher F. et al. 2008a. Individual Laboratory-Measured Discount Rates Predict Field Behavior, Journal of Risk and Uncertainty. Chabris, Christopher, David Laibson, and Jonathan Schuldt. 2008b. Intertemporal Choice. Palgrave Dictionary of Economics. Clark, Gregory Factory Discipline. Journal of Economic History, 54: Dasgupta, A Reports on Informal Credit Markets in India: Summary. New Delhi: National Institute of Public Finance and Policy. 42

43 DellaVigna, Stefano Psychology and Economics: Evidence from the Field. Journal of Economic Literature, 47: DellaVigna, Stefano and Ulrike Malmendier Contract Design and Self-Control: Theory and Evidence. Quarterly Journal of Economics, 119: DellaVigna, Stefano and Ulrike Malmendier Paying Not to Go to the Gym. American Economic Review, 96(3): Fehr, Ernst and Lorenz Goette "Do Workers Work More if Wages Are High? Evidence from a Randomized Field Experiment." American Economic Review, 97(1): Frederick, Shane, George Loewenstein, and Ted O Donoghue Time Discounting and Time Preference: A Critical Review. Journal of Economic Literature, 40(2): Fudenberg, Drew and David Levine A Dual Self Model of Impulse Control. American Economic Review, 96: Giné, Xavier, Dean Karlan, and Jonathan Zinman Put Your Money Where Your Butt Is: A Commitment Contract for Smoking Cessation. Available: Gneezy, Uri, and John List Putting Behavioral Economics To Work: Testing For Gift Exchange In Labor Markets Using Field Experiments, Econometrica, 74(5): Gul, Faruk and Wolfgang Pesendorfer Temptation and Self-Control, Econometrica, 69(6): Gul, Faruk and Wolfgang Pesendorfer Self-Control and the Theory of Consumption. Econometrica, 72(1): Hossain, Tanjim and John List The Behavioralist Visits the Factory: Increasing Productivity Using Simple Framing Manipulations. NBER Working Paper No Kaur, Supreet, Michael Kremer, and Sendhil Mullainthan Self-Control and the Development of Work Arrangements. American Economic Review Papers and Proceedings, 100(2): Kaur, Supreet, Michael Kremer, and Sendhil Mullainthan Works Well with Others: Peer Effects and Social Motivation. Mimeo, Harvard University. Laibson, David Golden Eggs and Hyperbolic Discounting. Quarterly Journal of Economics, 112(2): O Donoghue, Ted and Matthew Rabin Doing It Now or Later, American Economic Review, 89(1):

44 O Donoghue, Ted and Matthew Rabin Choice and Procrastination, Quarterly Journal of Economics, 116(1), O Donoghue, Ted and Matthew Rabin Incentives and Self Control. Society Monographs, 42: Econometric Schelling, T Self-command: A new discipline. In J. Elster and G.F. Loewenstein (Eds.), Choice Over Time (pp ). New York: Russell Sage Foundation. Shearer, Bruce "Piece Rates, Fixed Wages and Incentives: Evidence from a Field Experiment," Review of Economic Studies, 71(2): Thaler, Richard and Shlomo Benartzi Save More Tomorrow: Using Behavioral Economics to Increase Employee Saving. Journal of Political Economy, 112(1): Thaler, R.H., and H.M. Shefrin An Economic Theory of Self-Control, Journal of Political Economy, 89, Wertenbroch, K Cosumption Self-control by Rationing Purchase Quantities of Virtue and Vice, Marketing Science, 17,

45 Figure 1: Incentive Contracts Notes: This figure displays the two types of incentive contracts offered to workers. The linear control contract paid a piece rate wage of w for each accurate field entered. The nonlinear dominated contract imposed a production target, X; workers were paid w for each accurate field if they met the target, but only received w/2 for each field if they fell short of the target. Thus, earnings are equivalent under both contracts for output levels above X. However, if a worker fails to achieve X, earnings are substantially less under the dominated contract. 45

46 Figure 2: Treatment Design Control Contract (0.25) Contract Assignment (Assignment changes daily) Assigned to Target (0.25) Evening Choice (0.25) Morning Choice (0.25) Payday Assignment (Assigned once in beginning of study) Tuesday Payday (0.33) Thursday Payday (0.33) Saturday Payday (0.33) Notes: This chart provides an overview of the treatment design. One-third of workers were assigned to each of the three payday groups. This assignment was done once for each worker, when the worker joined the firm, and remained fixed for the duration of the project. Workers were orthogonally assigned to each of the four contract treatments exactly 25% of the time. The assignments changed daily. 46

47 Figure 3: Workers Data Entry Screen Notes: The figure displays a screen shot of a typical data entry screen. Workers viewed scanned images of records in the top half of their screen and entered information from these images into the appropriate fields at the bottom half of the screen. Identifying information from the records has been covered for confidentiality. 47