Abstract

Abstract: We present evidence from randomized field experiments that 401(k) savings choices are significantly affected by one-to two-sentence anchoring, goal-setting, or savings threshold cues embedded in emails sent to employees about their 401(k) plan. Even though these cues contain little to no marginal information, cues that make high savings rates salient increased 401(k) contribution rates by up to 2.9% of income in a pay period, and cues that make low savings rates salient decreased 401(k) contribution rates by up to 1.4% of income in a pay period. Cue effects persist between two months and a year after the email. Keywords: nudge, cues, anchoring, goals, 401(k), retirement savings JEL codes: D03, D14, D91, G02 * Corresponding author: james.choi@yale.edu, 165 Whitney Ave., P.O. Box 208200, New Haven, CT 06520-8200. We thank Kalok Chan, Eric Johnson, David Hirshleifer, Christoph Merkle, Alessandro Previtero, Shanthi Ramnath, Joeri Sol, Victor Stango, and audiences at Columbia, Federal Reserve Board of Governors, George Washington University, Harvard, HKUST Household Finance Symposium, IZA/WZB Field Days conference, University of Mannheim, Miami Behavioral Finance Conference, Michigan State University, NBER Household Finance Meeting, NBER Aging Summer Institute, Pontifical Catholic University of Chile, Queen's University Behavioral Finance Conference, UCSB/UCLA Conference on Field Experiments, UCLA, University of Melbourne, and Yale for helpful comments, and Google and the National Institute on Aging (grant R01-AG-021650) for financial support. We are grateful for Minhua Wan's comments and assistance with database management. Part of the work on this paper was done while Emily Haisley was a post-doctoral associate at Yale and Jennifer Kurkoski was a doctoral student at UC Berkeley. 1 In this paper, we show using randomized field experiments that seeing subtle cues that make a certain savings choice salient significantly affects individuals' contributions to their 401(k) retirement savings plan, even though the cues contain little to no marginal information. The design of the three types of cues we test was inspired by psychological phenomena documented in the psychology and behavioral economics literature. Based on this literature, we predicted that savings choices would move towards the choice made salient by each of these cues. Indeed, we find that high savings cues raise 401(k) contribution rates, and low savings cues depress 401(k) contribution rates. In the terminology of Raiffa (1982) and Thaler and Benartzi (2004), our paper is a work of prescriptive economics, which aims to provide tools to improve economic outcomes. To date, the practice of choice architecture (Thaler and Sunstein, 2008)-a prominent example of prescriptive economics-in retirement savings systems has focused on defaults and the composition of the savings rate and investment option menus (Madrian and Shea, 2001; Our field experiments randomized exposure to savings cues in emails about the 401(k) that were sent to one large technology company's employees in two waves about a year apart from each other: the first in November 2009 and the second in October 2010. The only difference between the control and treatment emails was that the treatment emails included one or two additional sentences. We call the first type of cues "anchors" (Tversky and Kahneman, 1974) because they mentioned an arbitrary savings increase amount while trying to sound maximally uninformative. Psychologists have long known that the presentation of arbitrary numbers-or anchors-can shift subjects' judgments and willingness to pay for goods (Tversky and Kahneman, 1974; Johnson and Schkade, 1989; Green et al., 1998; Kahneman and Knetsch, 1993; Ariely, Loewenstein, and Prelec, 2003; Stewart, 2009). However, evidence is only beginning to emerge on the importance of anchoring for economic decisions outside the laboratory 2 Since any cue that makes a particular savings behavior salient is likely to also cause that savings behavior to become an anchor, an anchoring cue can be thought of as a constituent ingredient of all other cues and thus an interesting place to begin the study of cues. 1 The second type of cue mentioned a savings threshold that was created by the 401(k)'s employer matching contribution rules or contribution limits. The third type of cue mentioned an arbitrary savings level as an example of a goal. Locke and Latham (1990Latham ( , 2002Latham ( , 2006 summarize a large literature showing that setting concrete goals that are difficult to achieve enhances performance relative to setting unambitious or vague "do your best" goals. A number of laboratory studies have found that behavior changes even when the goals are subconsciously primed by cues in the environment rather than consciously chosen We find that relative to the control email, emails with cues that made a high savings choice salient increased 401(k) contribution rates by up to 2.9% of income in a pay period, and emails that made a low savings choice salient decreased 401(k) contribution rates by up to 1.4% of income in a pay period. Anchors affect average contribution rates with a significant delay, and high anchors have the perverse property of decreasing the probability of making a contribution rate change in the short run. In addition, very high anchors do not raise contribution rates more 1 Of course, an anchor may also be subsequently adopted as a goal, or may coincide with a certain savings threshold created by the plan's rules. 3 than moderately high anchors. On the other hand, anchor effects are the most durable out of the cue effects we test, lasting up to nearly a year after the emails were sent. In contrast, threshold and goal cues can have more immediate effects, and they are more effective the further the cue is from the recipient's status quo, but their effects dissipate within two to seven months. The cue-induced contribution rate changes we find were unlikely to be financially neutral for individuals. Although we do not observe savings outside the 401(k), the generous employer matching contributions in the particular 401(k) we study means that for most employees, even if each dollar of additional 401(k) contributions was offset by a dollar of decreased saving outside the 401(k), their lifetime wealth would still increase. Cue effects are interesting not only because they exist, but also because their magnitudes are large compared to those estimated for an expensive, commonly used economic lever to increase 401(k) savings: employer matching contributions. 2 Kusko, Poterba, and Wilcox (1998) find that, at one manufacturing firm, increasing the match rate from 25% to 150% on the first 6% of income contributed raised average 401(k) contribution rates by only 0.2% to 0.3% of income. A decrease in the match rate from 139% to 0% was accompanied by an average contribution rate fall of only 0.3% of income. Other studies that identify match effects using across-plan variation come to conflicting conclusions, finding that matches increase, decrease, or have no effect on contributions (see Our findings may provide insight into the mechanisms underlying two important savings phenomena: inertia at the default 401(k) contribution rate, and the clustering of 401(k) contributions at the maximum allowable contribution rate and the employer match threshold (i.e., the maximum level of employee contributions for which the employer will make matching contributions). Although many papers have documented strong savings inertia 3 , the causes of this inertia are still not fully understood (Bernheim, Fradkin, and Popov, 2015). 4 Choi et al. 4 contribution maximum or the match threshold, but this pattern is consistent with both rational and behavioral explanations. Our experimental results suggest that people frequently end up at the default, match threshold, and contribution maximum in part because those three choices are salient; for example, a 3% default contribution rate inevitably creates a 3% contribution rate cue. Our treatment effects are estimated on a particular sample of employees-people who are generally highly educated, very well-paid, and technology savvy-in a particular 401(k) plan. However, Goda, Manchester, and Sojourner (2014) describe a field experiment on University of Minnesota employees that shows that cues can be effective in settings that are quite different from ours. Cues are not the main focus of their experiment; they are primarily interested in the effect that providing projections of asset balances and income has on contributions to a supplemental retirement savings plan. But they did randomly vary the graphs used to deliver these projections. One set of graphs showed asset and income projections for the cases where the employees increased their savings by $0, $50, $100, or $250 per pay period. The other set of graphs showed these projections for the cases where the employees increased their savings by $0, $100, $200, or $500 per pay period. Employees receiving the graphs with the higher savings examples had a contribution rate six months after the mailing that was on average 0.2% of income higher than that of those who received the graphs with the lower savings examples. The magnitude of this treatment effect cannot be directly compared to ours because University of Minnesota employees are also covered by a primary pension that mandates high levels of saving, and changing one's savings rate in the supplemental plan was unusually onerous. 5 Our paper is distinguished from 6 The remainder of our paper proceeds as follows. Section I discusses the features of the company 401(k) plan. Section II describes our data. Section III describes the experimental design 5 Those covered by the mandatory defined contribution plan have 15.5% of their income contributed to the plan. Seventy-six percent of recipients were not enrolled in the supplemental plan before the mailing. If these recipients wanted to start contributing, they had to mail in a request for an enrollment kit, at which point they would receive the enrollment forms in a few weeks. They would then have to complete these enrollment forms and physically mail them back. Recipients who were already enrolled in the plan had to physically mail in a form to change their contribution rate. concludes. An online appendix contains details of how the joint hypothesis test in Section VII was implemented and regression tables that correspond to the figures that illustrate most of our results. I. 401(k) plan features Employees at the company we study can make before-tax, after-tax, and Roth contributions to their 401(k) plan. 7 Before March 2011, employees specified three percentages: the percent of their paycheck they wanted to contribute on a before-tax, after-tax, and Roth basis. were automatically enrolled at a 3% before-tax contribution rate unless they opted out. At the beginning of each subsequent calendar year until 2010, seasoned employees who had never actively chosen their 401(k) elections had their before-tax contribution rate automatically increased by 1 percentage point, and the default before-tax contribution rate for new hires also increased by 1 percentage point. In 2011, the default contribution rate for new hires did not change, and seasoned employees were not subject to automatic contribution rate increases. The company makes matching contributions to the 401(k) that depend upon each employee's cumulative contributions during the calendar year. The match amount during 2009 was the greater of (1) 100% of before-tax plus Roth contributions up to $2,500, or (2) 50% of before-tax plus Roth contributions up to $16,500, resulting in a maximum possible match of $8,250. 8 This match structure generated a 100% marginal subsidy on contributions up to $2,500, 7 Both principal and capital gains of before-tax contributions are taxed upon withdrawal. Only capital gains of aftertax contributions are taxed upon withdrawal. Roth contributions are made using after-tax dollars, but both principal and capital gains are not taxed upon withdrawal. 8 These matching contributions are substantially more generous than the typical 401(k) match. 6 a 0% marginal subsidy on contributions between $2,501 and $5,000, and a 50% marginal subsidy on contributions between $5,001 and $16,500. In 2010, the match structure changed to be the greater of (1) 100% of before-tax plus Roth contributions up to $3,000, or (2) 50% of before-tax plus Roth contributions up to $16,500. This new match structure shifted the 0% marginal match zone to contributions between $3,001 and $6,000. Matching contributions vest immediately. Employees receive an annual bonus that is paid each March. In 2009 and 2010, if an employee had a 5% contribution rate in effect during the pay period in which the bonus was paid, 5% of the bonus would be contributed to the 401(k) plan. As a result, many employees changed their contribution rate shortly before or during the bonus pay period in 2009 and 2010. Starting in 2011, employees could choose a separate contribution election for their bonus, and this election could specify dollar amounts to be contributed rather than percentages of the bonus. Unless actively changed by the employee, the bonus contribution election was by default set equal to the election for regular paychecks. II. Data description We use salary and employment termination date data from personnel records and 401(k) contribution rate data from Vanguard. Individuals in the data were assigned random identifiers; no personally identifying information was included. Vanguard data included cross-sectional snapshots of all 401(k) contribution rate elections recipient. This allowed us to calculate how many dollars each employee would contribute to the 401(k) during the calendar year if she left her contribution rate elections unchanged-a variable that determined which treatments the employee was eligible to be assigned to. 9 Because the changes file does not record the contribution rate elections employees had in effect when they first joined the 401(k), we cannot construct a complete contribution rate history between January Employees assigned to the low cue treatment received the additional sentences, "For example, you could increase your contribution rate by 1% of your income and get more of the match money for which you're eligible. (1% is just an example, and shouldn't be interpreted as advice on what the right contribution increase is for you.)" We call this the 1% anchor treatment because it tries to make the cued savings rate sound maximally arbitrary. Despite our efforts to make the anchor sound arbitrary, recipients may have nonetheless inappropriately inferred something about their optimal savings rate from the anchor. Our objective is not to rule out the possibility that employees performed a Bayesian update about their optimal 401(k) contribution rate based on this cue, but to see if even the most minimal, uninformative cues can have large effects on savings choices. Indeed, mistaken inference is one channel through which cues might operate. Danilowitz, Frederick, and Mochon (2011) find that about 70% of the anchoring effect magnitude identified in the laboratory using standard 10 We excluded employees hired in 2009 because we did not know how much they had contributed in 2009 to a 401(k) at a previous employer, so we could not be sure how much they were eligible to contribute to their current employer's 401(k) in 2009. 8 experimental procedures is due to mistaken inference by subjects. We could have cued the 60% maximum allowable contribution rate using the same anchoring language, replacing the 1% increase with whatever amount would cause the recipient to contribute 60% of her income to the 401(k). However, we worried that it would sound strange to mention such a large contribution increase as an arbitrary example. Therefore, employees assigned to receive a high cue had the following 60% threshold cue added to their email: "You can contribute up to 60% of your income in any pay period." Mentioning the extremely high 60% contribution rate was natural in the context of explaining the relevant 401(k) plan rule. In addition to creating an anchor at 60%, making the maximum plan contribution rate salient may activate a rule of thumb tied to the maximum, pushing people to increase their contribution rate. Although information about the 60% maximum was not present in the control email, it is such a high contribution rate that it is not a relevant constraint for most employees, so the marginal information contained in the cue is likely to be small. Since a 60% contribution rate is an implausible recommendation, the 60% threshold cue may also be less likely than the anchoring cue to be interpreted as a recommendation from the employer. and those who faced a 50% marginal match. Eligibility for assignment to experimental conditions depended on the employee's category. Employees had an equal probability of being assigned to each condition for which they were eligible. Employees on pace to contribute at least $5,000 if they left their contribution rate unchanged for the rest of the year-and thus faced a 50% marginal match-could be assigned to the control, the 1% anchor treatment, or the 60% threshold treatment. 12 The anchor cue's 11 Early drafts of this paper also reported results from a 10% anchor treatment administered in 2009. Like the 10% anchor treatment in 2010, the 2009 recipients of the 10% anchor had average contribution rates similar to the control group prior to the bonus and higher average contribution rates afterwards. We exclude the 2009 10% anchor treatment from the current paper because we discovered that by chance, randomization had created a significant difference in the average pre-email contribution rate of the 2009 10% anchor recipients relative to the control group. 12 We do not analyze employees in this projected contribution category who were not eligible for all three conditions (and they do not appear in IV. Analysis of 2009 experiments A. Control email recipient behavior The control email contains its own plethora of information and numerical cues, and also serves as a reminder about saving. 13 The cue effects that are the focus of this paper represent impacts above and beyond those of the control email. 10 rate on November 27, 2009-the first payday following the email-of control recipients employed since January 2008 was 10.7%, which is 2.3% of income higher than it was two weeks earlier. The average contribution rate increased further to 11.8% on December 11, 3.4 percentage points higher than it was on November 13. The average then fell slightly to 11.5% on December 24. By comparison, during the last three pay periods of the prior year, the sample's average total contribution rate fell by 0.5% of income. Like in the prior year, the average contribution rate of control recipients declined significantly in early 2010 as the bonus payment approached. By the first pay period after the bonus, the average contribution rate of 8.5% was only slightly above what it was in the last pay period before the email (8.4%). Some of the decline in the total contribution rate after the bonus pay period is due to employees hitting the legal calendar-year maximum of $16,500 in before-tax plus Roth contributions, at which point their before-tax and Roth contribution rates are automatically set to zero. B. Econometric methodology for estimating cue effects We identify our cue treatment effects by comparing employees in a projected contribution category who received a cue to control email recipients in the same category. In other words, we compare each treatment group in a column of Our regressions follow an event-study framework. The event date is November 17, 2009, the event is the sending of the cue, and the benchmark is the appropriate set of control email 11 recipients. As is the norm in event studies, we estimate the effect of the event at multiple individual post-event periods by running a separate regression for each post-email payday. Our main dependent variable is the difference in the total 401(k) contribution rate (before-tax plus after-tax plus Roth) between the payday being evaluated and the last payday prior to the email, and our explanatory variable is a treatment dummy. 15 We also estimate treatment effects averaged over multiple periods by using as our dependent variable the average total 401(k) contribution rate during the averaging window (equally weighting each pay period) minus the pre-email total 401(k) contribution rate. We restrict the sample to employees who were still at the company at the end of the period we are averaging over. The advantage of this approach relative to the individual payday regressions is that it concisely estimates the longer-run impact of the treatment and can have more power to detect small treatment effects that persist for many periods. The disadvantage is that when a nonzero treatment effect has a duration that is shorter than the averaging period, statistical power to detect the effect diminishes because the cumulative treatment effect size becomes small relative to the cumulative residual variance. windows at the bonus payday for two reasons. First, the bonus payday is a psychologically and economically significant date that motivates a great deal of 401(k) activity, and we see sharp changes in contribution rate trajectories at the bonus payday prior to our 2009 emails. Second, because we do not know how large each employee's bonus is, we do not know how to weight the pre-bonus, bonus, and post-bonus contribution rates to construct a cue's effect on total contributions as a percent of total compensation across all 24 paydays after the email was sent. 15 Using a first-differenced contribution rate as the dependent variable makes our cross-sectional regression equivalent to a two-period panel regression where the dependent variable is the total contribution rate and the explanatory variables are individual fixed effects, a dummy for whether the observation comes after the email date, and a treatment dummy interacted with the post-email dummy. A difference in differences regression specification, which replaces the vector of individual fixed effects with a constant and a treatment dummy, gives an identical treatment effect point estimate but has a larger standard error because it discards information from the data's panel structure. 16 12 C. Effect of the 1% anchor Our first experimental finding is that the low 1% anchor decreased subsequent contribution rates. The average contribution rate of the 1% anchor group and its control group both rose during the first two pay periods before beginning to fall, but the 1% anchor group was persistently below the control group until March 5, 2010, when the two converged as bonuses were paid. Surprisingly-given our prior expectation that anchoring effects would be strongest immediately after the email was sent-the gap between the 1% anchor group and the control group took eleven weeks to reach its peak magnitude of 1.4% on February 5. The treatment effect is not statistically significant before year-end 2009, but from January 22 to February 19, 2010, the 1% anchor decreased average total contribution rates by between 1.1% and 1.4% of income at the 5% significance level during one payday and at the 1% level during the other two. We will see in Section VI.C that when we tested anchor cues in the 2010 emails, we once more observed significant delays in their effects on average contribution rates. The average contribution rate series diverged from each other after March 5, with the 1% anchor group again consistently contributing less than the control group through October 15 by as much as 1.2% of income. Of the sixteen post-bonus paydays in The fact that the 1% anchor had no significant effect on average contribution rates in 2009 does not mean it had no effect at all that year. A linear probability regression (not shown in exhibits) reveals that 1% anchor recipients were 1.5 percentage points more likely (p = 0.035) 13 than the control group to have a contribution rate exactly 1% of income higher than their November 13, 2009 contribution rate during at least one pay period between November 27 and December 24, 2009. This increase represents a doubling of the control group's baseline probability of 1.6%. On the other hand, there is much less clustering at the cued action than there typically is at a default contribution rate. We can examine the 1% anchor effect integrated over periods of time longer than one payday. Averaged across both individually significant and insignificant paydays, the 1% anchor decreased contribution rates by 0.8% of income (p = 0.047) during the seven pre-bonus paydays between November 27 and February 19, had no effect on the March 5 bonus payday (+0.05% of income, p = 0.933), and decreased contribution rates by 0.8% of income (p = 0.076) during the sixteen post-bonus paydays from March 19 to October 15. Note that there is no contradiction between the fact that the 1% anchor mentioned a contribution increase and the fact that its treatment effect relative to the control email is negative. Recipients of the 1% anchor did raise their contribution rates on average relative to where they were before the email. They simply did not raise them as much as the control group, in accordance with the low contribution increase that was cued. The delayed reaction of the average contribution rate to the 1% anchor may be consistent with previous findings that minor psychological interventions can influence behavior after a significant delay. Research on "mere measurement" (e.g., Morwitz, Johnson, and Schmittlein, 1993) and the "self-prophecy effect" (e.g., Spangenberg, 1997) relationship with the firm, and these effects increased in magnitude for 3 to 6 months after the intervention. Alternatively, our delayed effect may be due not to a single cue exposure's effect growing over time, but to the cumulative impact of multiple exposures that occurred when employees re-read the email weeks after it had been sent in order to remind themselves of the 14 instructions on how to change their contribution rate. 17 The treatment effect on the average contribution rate is not delayed because employees who reacted to the email later are more susceptible to anchors. The average contribution rate among employees who changed their contribution rate between the email send date and year-end 2009 also exhibited a growing divergence between the 1% anchor and control groups in January, an attenuation of the anchor effect on the bonus payday, and a re-emergence of the anchor effect after the bonus. The delayed treatment effect was also not caused by fewer anchor recipients being subject to auto-escalation at the beginning of 2010 than control employees. In fact, slightly more anchor recipients (17%) than control employees (16%) increased their contribution rate by exactly 1% at that time. The 1% anchor effect's disappearance on the bonus payday might be explained by previous laboratory evidence that even tiny discrepancies between the choice domain and the anchor domain are enough to eliminate the anchoring effect. The company's employees think quite differently about bonus versus non-bonus payday contributions, a mindset reflected in the fact that in 2012, the company introduced a new set of 401(k) contribution rate elections that applies only to the bonus. Although the 1% anchor resulted in lower average contribution rate increases, did it encourage a larger fraction of recipients to make small contribution rate increases? Online Appendix anchor recipients and their control group had a different contribution rate in effect, but we find little evidence that the 1% anchor affected the probability of action at any point in time. The contrast between these null results on the probability of action and the significant results on average contribution rates indicates that the 1% anchor did not work by changing the action 17 Unfortunately, we do not have data on when emails were read. 15 bands within an Ss model (e.g., D. Effect of the 60% contribution rate threshold cue We analyze the effect of the 60% contribution rate threshold cue separately for recipients who were on pace to contribute less than $2,500, between $2,500 and $4,999, and between $5,000 and $16,499 in 2009, since each of these groups faced different marginal matching incentives. We find that this extremely high cue increased subsequent contribution rates, but only among low savers. Is the 60% threshold treatment effect on low contributors due to their learning from it that the plan's maximum contribution rate is 60%? According to this explanation, employees in the control group chose smaller contribution increases than they otherwise would have because they falsely believed they were not allowed to contribute more. 18 Online Appendix In contrast to what we saw among low contributors, the bottom two graphs in In untabulated regressions, we examine whether the 60% threshold treatment caused recipients to contribute exactly 60% of their income in any pay period between November 27, 2009 and October 15, 2010. 19 These regressions show that the 60% threshold treatment made contributing at 60% more likely only for recipients who were previously on pace to contribute less than $2,500 in 2009. The effect for these recipients is a 5.7 percentage point increase (p = 0.020) in the probability of contributing 60%, up from a baseline probability of 5.4% in the control group. Why did the 60% threshold cue have such heterogeneous effects? Four differences across the three projected contribution categories could be responsible. First, average salaries differ. If low-income individuals are generally more susceptible to "nudges" because of lower financial sophistication, and low-income individuals are more likely to be in the lowest projected contribution category, then the threshold cue would have the strongest effect in the lowest projected contribution category. Second, individuals in the lowest projected contribution category had a low average contribution rate in effect when they received the email, so there was a large gap on average between their status quo and the cued contribution rate. This larger gap may have been more motivating. Third, employees in the lowest projected contribution category had the highest marginal match rate, which may have made them more likely to act upon the motivation created by a cue. Fourth, employees in the lowest projected contribution category