Behavioral Economics and Public Policy: a Pragmatic Perspective†

American Economic Review: Papers & Proceedings 2015, 105(5): 1–33 http://dx.doi.org/10.1257/aer.p20151108

RICHARD T. ELY LECTURE

Behavioral Economics and Public Policy: A Pragmatic Perspective†

Starting with Simon 1955 , Kahneman and of each viewpoint in different settings e.g., List Tversky 1979 , and Thaler( 1980) , a large body 2004; Levitt and List 2007; DellaVigna( 2009 . of research( has) incorporated( insights) from psy- In this paper, I approach the debate )on chology—such as loss aversion, present bias, and behavioral economics from a more pragmatic, inattention—into economic models.1 Although policy-oriented perspective. Instead of pos- this subfield of behavioral economics has grown ing the central research question as “Are the very rapidly, the neoclassical model remains the assumptions of the neoclassical economic benchmark for most economic applications, and model valid?,” the pragmatic approach starts the validity of behavioral economics as an alter- from a policy question for example, “How native paradigm continues to be debated. can we increase savings rates?”( and incorpo- The debate about behavioral economics is rates behavioral factors to the extent) that they often framed as a question about the founda- improve empirical predictions and policy deci- tional assumptions of neoclassical economics. sions.2 This approach follows the widely applied Are individuals rational? Do they optimize in methodology of positive economics advocated market settings? This debate has proved to be by Friedman 1953 , who argued that it is more contentious, with compelling arguments in favor useful to evaluate( economic) models on the accu- racy of their empirical predictions than on their assumptions.3 While Friedman used this reason- * Harvard University, 226 Littauer Center, Cambridge, ing to argue in favor of neoclassical models, I MA 02138, and NBER e-mail: [email protected] . Prepared for the Richard T.( Ely Lecture, American Economic) argue that modern evidence calls for incorpo- Association, January 3, 2015. A video of the lecture is rating behavioral economics into the analysis of available. I thank Saurabh Bhargava, Stefano DellaVigna, important economic questions. Nathaniel Hendren, Emir Kamenica, Lawrence Katz, David I classify the implications of behavioral eco- Laibson, Benjamin Lockwood, Sendhil Mullainathan, Ariel nomics for public policy into three domains. Pakes, James Poterba, Matthew Rabin, Josh Schwartzstein, Andrei Shleifer, and Dmitry Taubinsky for helpful com- Each of these domains has a long intellectual ments and discussions. I am very grateful to my collabo- tradition in economics, showing that from a rators John Friedman, Nathaniel Hendren, Lawrence Katz, pragmatic perspective, behavioral economics Patrick Kline, Kory Kroft, Soren Leth-Petersen, Adam represents a natural progression of rather than Looney, Torben Nielsen, Tore Olsen, and Emmanuel Saez a challenge to neoclassical economic( methods. for their contributions to the studies discussed in this paper. ) Augustin Bergeron, Jamie Fogel, Michael George, Nikolaus First, behavioral economics offers new pol- Hildebrand, and Benjamin Scuderi provided outstand- icy tools that can be used to influence behavior. ing research assistance. This research was funded by the National Science Foundation. † Go to http://dx.doi.org/10.1257/aer.p20151108 to visit 2 I focus on factors that can be changed through policy, the article page for additional materials and author disclo- but much of the analysis in this paper also applies to predict- sure statement. ing the effects of changes in other exogenous factors, such 1 Although the implications of psychology for economics as technology. have been formalized using mathematical models only in 3 In a widely cited example, Friedman points out that recent decades, some of these ideas were discussed quali- the behavior of an expert billiards player may be accurately tatively by the founders of classical economics themselves, modeled using complex mathematical formulas even though including Adam Smith Ashraf, Camerer, and Loewenstein the assumption that the player himself knows and applies 2005 . ( these formulas is likely to be incorrect. ) 1 2 AEA PAPERS AND PROCEEDINGS MAY 2015

Insights from psychology offer new tools—such year to subsidize retirement saving in 401 k as changing default options or framing incen- and IRA accounts Joint Committee on Taxation( ) tives as losses instead of gains—that expand the 2012 . I summarize( recent evidence showing set of outcomes that can be achieved through that such) subsidies have much smaller effects on policy. This expansion of the policy set parallels savings rates than “nudges” Thaler and Sunstein the transition in the public finance literature 2008 such as defaults and automatic( enrollment from studying linear commodity taxes Ramsey plans) that are motivated by behavioral models of 1927 to a much richer set of nonlinear (tax poli- passive choice. These new policy tools allow us cies )Mirrlees 1971 . to achieve savings rates that may have been unat- Second,( behavioral) economics can yield tainable with the tools suggested by neoclassical better predictions about the effects of existing model. These empirical findings are very valu- policies. Incorporating behavioral features such able irrespective of the underlying behavioral as inertia into neoclassical models can yield model, although theory remains essential for better predictions about the effects of economic extrapolation e.g., predicting behavior in other incentives such as retirement savings subsidies settings and (for welfare analysis e.g., deter- or income tax policies. Moreover, these behav- mining )whether policymakers should( be trying ioral features can help econometricians develop to increase savings rates to begin with . new counterfactuals control groups to identify The second application illustrates that) behav- policy impacts. ( ) ioral models can be useful in predicting the Third, behavioral economics generates new impacts of existing policies even if they do not welfare implications. Behavioral biases such produce new policy tools. Here, I focus on the as inattention or myopia often generate differ( - effects of the Earned Income Tax Credit EITC — ences between welfare )from a policymaker’s the largest means-tested cash transfer program( ) in perspective, which depends on an agent’s expe- the United States—on households’ labor supply rienced utility his actual well-being , and the decisions. The EITC provides subsidies that are agent’s decision( utility the objective) the agent intended to encourage low-wage individuals to maximizes when making( choices . Accounting work more. I discuss recent evidence showing for these differences between )decision and that individuals living in areas with a high den- experienced utilities improves predictions about sity of EITC claimants have greater knowledge the welfare consequences of policies. The dif- about the parameters of the EITC schedule and, ference between the policymaker’s and agent’s accordingly, are more responsive to the program. objectives in behavioral models parallels non- These differences in knowledge across areas pro- welfarist approaches to optimal policy Sen vide new counterfactuals to identify the impacts ( 1985; Kanbur, Pirttilä, and Tuomala 2006 and of the EITC on labor supply decisions and reveal the techniques used to identify agents’ experi) - that the program has been quite successful in enced utilities resemble those used in the long increasing earnings among low-wage individu- literature on externalities Pigou 1920 . als. These results demonstrate that even if one I illustrate these implications( of behavioral) cannot directly manipulate perceptions of the economics for public policy using a set of appli- EITC, accounting for the differences in knowl- cations drawn from recent research. The appli- edge across areas is useful in understanding the cations focus on three major decisions people effects of the existing incentives. make over the course of their lives: how much to The first two applications focus on the posi- save, how much to work, and where to live. Each tive implications of behavioral economics, i.e., application is motivated by a policy question predicting the effects of policies on behavior. that has been studied extensively using neoclas- The third application shows how behavioral sical models. My objective here is to illustrate models also provide new insights into the wel- how incorporating insights from behavioral eco- fare consequences and optimal design of poli- nomics can yield better answers to these long- cies. I illustrate these normative implications by standing policy questions. considering policies such as housing voucher In the first application, I show how behavioral subsidies whose goal is to change low-income economics offers new policy tools to increase families’ choice of neighborhoods. Recent retirement saving. The US federal government empirical studies have shown that some neigh- currently spends approximately $100 billion per borhoods generate significantly better outcomes VOL. 105 NO. 5 RICHARD T. ELY LECTURE 3 for children yet do not have higher housing costs. change behavior and increase welfare if agents Both neoclassical models and models featuring suffer from behavioral biases without distorting behavioral biases e.g., present-bias or imperfect behavior if agents optimize. Model uncertainty information can (explain why families do not can thus provide a new argument for the use of move to such) neighborhoods, but these models behavioral nudges that is distinct from the com- generate very different policy prescriptions. The mon rationale of libertarian paternalism Thaler neoclassical model says that there is no reason and Sunstein 2003 . ( to intervene except for externalities. Behavioral Together, the three) applications illustrate that models call for policies that encourage families incorporating behavioral features into economic to move to areas that will improve their chil- models can have substantial practical value in dren’s outcomes, such as housing voucher sub- answering certain policy questions. Of course, sidies or assistance in finding a new apartment. behavioral factors may not be important in all The optimal policy in this setting depends on applications. The decision about whether to agents’ experienced utilities—their willingness incorporate behavioral features into a model to pay for a better neighborhood in the absence should be treated like other standard mod- of behavioral biases. Many economists hesitate eling decisions, such as assumptions about to follow the policy recommendations of behav- time-separable utility or price-taking behavior ioral models because of concerns about pater- by firms. In some applications, a simpler model nalism, i.e., giving policymakers’ perceptions of may yield sufficiently accurate predictions; in an individual’s experienced utility precedence others, it may be useful to incorporate behav- over the individual’s own choices. I discuss ioral factors, just like it may be useful to allow three nonpaternalistic methods of identifying for time nonseparable utility functions. This experienced utilities that have been developed pragmatic, application-specific approach to in recent research: i directly measuring expe- behavioral economics may ultimately be more rienced utility based( on) self-reported happiness, productive than attempting to resolve whether ii using revealed preference in an environ- the assumptions of neoclassical or behavioral ment( ) where agents are known to make choices models are correct at a general level.4 that maximize their experienced utilities, and This paper builds on several related literatures. iii building a structural model of the differ- The applications discussed here are a small subset ence( ) between decision and experienced utili- of a much broader literature that takes a pragmatic ties. These methods can provide more accurate approach to behavioral economics and public and robust prescriptions for optimal policy than policy. Thaler and Sunstein 2008 ; Congdon, those obtained from neoclassical models, ulti- Kling, and Mullainathan 2011( ; )Keller-Allen mately increasing social welfare if individuals and Li 2013 ; and Madrian( )2014 provide suffer from behavioral biases. examples( of the) new policy tools (and predictions) In some situations—including the neighbor- generated by behavioral economics. Bernheim hood choice application—one may have to make 2009 and Mullainathan, Schwartzstein, and judgments about optimal policy without being Congdon( ) 2012 provide further discussion of certain about whether the currently available ( ) data are generated by a neoclassical or behavioral model. Economists are inclined to use the 4 The relevance of behavioral economics is application-specific because deviations from rationality vary widely neoclassical model as the default option when across settings. In some markets, behavioral phenomena can faced with such model uncertainty. A more prin- be diminished by experience effects, arbitrage, or aggrega- cipled approach is to explicitly account for model tion that cancels out idiosyncratic mistakes see, e.g., List 2004; Farber 2014 . But the rarity of important( decisions uncertainty when solving for the optimal policy, ) as in the literature on robust control Hansen and e.g., buying a house or choosing where to go to college , limits( to arbitrage Shleifer and Vishny 1997 , and the lack) Sargent 2007 . Using some simple (examples, I ( ) ) of returns to debiasing consumers Gabaix and Laibson show that model uncertainty does not neces- 2006 may lead behavioral anomalies (to persist in other set- sarily justify using the neoclassical model for tings.) This context-dependence makes it difficult to answer welfare analysis. On the contrary, the optimal the question of whether individuals are “rational” or not at a general level. The pragmatic approach discussed here deals policy in the presence of model uncertainty may with these issues of external validity and generalizability by be to use behavioral nudges such as changes in ( directly focusing on the relevance of behavioral economics defaults or framing , because such nudges can for the question of interest. ) 4 AEA PAPERS AND PROCEEDINGS MAY 2015 normative issues in behavioral models. All of such as labor supply or neighborhood character- these applications of behavioral economics build istics. Let p denote the pretax price vector for the directly on prior research translating lessons c goods and Z the individual’s wealth. from psychology to economics and documenting Following Kahneman, Wakker, and Sarin empirical evidence of deviations from neoclassi- 1997 , let u c denote the agent’s experienced cal models. Conlisk 1996 , Rabin 1998 ; and utility—his( ) actual( ) well-being as a function of DellaVigna 2009 provide( ) an excellent( )over- choices—and v c his decision utility—the view of this earlier( ) body of work. objective he seeks( to) maximize when choosing c . Finally, the empirical applications discussed As discussed by DellaVigna 2009 , in a setting in this article are all examples of recent studies in without uncertainty, the agent’s( decision) utility applied microeconomics that use administrative can differ from neoclassical specifications either datasets with millions of observations. This big because he has nonstandard preferences—e.g., data approach often leads researchers to identify a utility function that exhibits reference depen- empirical regularities that are unrelated to their dence—or because he is influenced by ancillary initial hypotheses and sometimes do not match conditions Bernheim and Rangel 2009 , such as neoclassical predictions, making it useful to the way in (which choices are framed. The) ancil- draw on insights from behavioral economics. As lary conditions do not enter the agent’s expe- economics becomes an increasingly empirical rienced utility and budget set, and hence have science, economic theories will be shaped more no effect on behavior in a neoclassical model. directly by evidence, and the pragmatic approach It is useful to divide the ancillary conditions to behavioral economics described here may into two groups: those that can be manipulated become even more prevalent and useful.5 by policymakers such as defaults , which I The paper is organized as follows. Section I label “nudges” n following( Thaler and) Sunstein formalizes the three pragmatic implications of 2008 , and a set of other ancillary conditions behavioral economics for public policy using d( that )may affect agent behavior but cannot be a stylized model. Section II discusses the new manipulated through policy, such as perceptions policy tools offered by behavioral economics, or overconfidence. focusing on retirement savings. Section III illus- The planner’s objective is to choose a set trates how behavioral models can help us bet- of tax rates t and nudges n that maximize the ter predict the effects of income taxes and labor agent’s experienced utility u c subject to a ( ) supply. Section IV discusses the welfare impli- revenue requirement R̅ and a standard incen- cations of behavioral economics in the con- tive-compatibility condition for the agent:6 text of neighborhood choice. Each section also briefly reviews other applications that illustrate 1 max t,n u c s.t. the implications of behavioral models for other ( ) ( ) questions. I conclude in Section V by discussing 2 t ⋅ c R̅ some lessons for future research. ( ) = 3 c arg max c v c n, d s.t. p t ⋅ c Z . I. Conceptual Framework ( ) = { ( | ) ( + ) = } Neoclassical economics solves a special case This section formalizes the implications of of this general optimal policy problem, which behavioral economics for public policy using a typically imposes the following additional con- simple representative-agent model. Let c denote a straints on 1 .7 vector of choices made by the agent. In canonical ( ) examples, c represents a set of different consump- 6 tion goods or consumption at different times, but The assumption that policymakers should maximize individuals’ experienced utilities has been a standard bench- one can also interpret c as including other choices mark in normative economics since Bentham’s formulation of utilitarianism, but many other objectives have also been proposed e.g., Sen 1985 . See Kahneman and Sugden 5 Hamermesh 2013 documents the increasing influence 2005 for (a discussion of whether) maximizing experienced of empirical evidence( in) economics by studying publication utility( ) is a reasonable criterion in behavioral models. patterns. He reports that the fraction of empirical articles 7 The definition of a “neoclassical” model varies across published in general interest economics journals increased papers. A minimal requirement is that choices satisfy consis- from 38 percent to 72 percent between 1980 and 2010. tency and transitivity, but most applied economists impose VOL. 105 NO. 5 RICHARD T. ELY LECTURE 5

ASSUMPTION 1 Neoclassical Restrictions : or changes in information provision is as ad hoc The planner does not( have any policy nudges n) , an assumption as restricting attention to linear the agent’s decision utility is a smooth, increas- taxes or limiting attention to taxes on a subset ing, and concave function of consumption of goods in the economy. Although any of these choices, and experienced utility equals decision assumptions may be useful simplifications to utility: make progress in a given application, there is no deep justification for giving priority to models 4 n that restrict the policy set. For example, consider ( ) = ϕ the well-known result that linear consumption 5 d , v c smooth, increasing, taxes become superfluous once one permits non- ( ) = andϕ (concave) linear income taxation under fairly general conditions Atkinson and Stiglitz 1976 . This result 6 u v. prompted( researchers to reevaluate )the rationale ( ) = for taxes on capital income and commodities in Behavioral economics can be interpreted as Mirrlees’s framework rather than continue to relaxing the constraints in 4 , 5 , and 6 . work in Ramsey’s framework. Similarly, if one There is a long methodological( ) ( tradition) ( of) were to find that changes in default provisions in relaxing such constraints in economics, and in retirement savings plans obviate the need for tax this sense behavioral economics represents a subsidies, it would be difficult to justify retain- natural progression of widely accepted methods ing the assumption in 4 when studying optimal in the economics literature. I consider the impli- savings policies. ( ) cations of relaxing each of the three constraints Relaxing 5 yields better predictions about in turn.8 the effects of( existing) policies. A model of deci- Relaxing 4 yields new policy tools. For sion utility that incorporates nonstandard prefer- example, policymakers( ) may be able to influence ences and ancillary conditions can be helpful in the agent’s choice of c by making certain features predicting the effects of taxes dc j dt i regard- of the choice set more salient or changing default less of whether it offers new tools( to/ manipulate) options. Expanding the policy set broadens the behavior n . Building models of behavior set of feasible allocations that can be achieved, whose predictions( = ϕ) more accurately match data which could ultimately increase welfare u c . is a core focus of positive economics. As one This expansion of the policy set parallels (the) example, consider recent evidence that the drop shift from studying linear taxes on commod- in expenditure around retirement may be better ities Ramsey 1927 to a mechanism design explained by a model that features complemen- approach( that permits) general, nonlinear taxes tarities between consumption and labor in the Mirrlees 1971 .9 Ruling out the use of defaults utility function Aguiar and Hurst 2005 . Few ( ) would insist on retaining( the assumption of) separable utility when studying consumption pat- stronger additional assumptions, such as smoothness of utility and concavity which rule out phenomena such as ref- terns around retirement in light of such evidence. erence points or exponential( discounting to rule out time Similarly, if one can better explain the data in a inconsistent choices) . The precise delineation( between “neo- ) relevant application by incorporating features classical” and “behavioral” models is a matter of terminol- such as inattention or reference dependence into ogy and is not central for the main arguments in this paper, the model of individual decisions v c n, d , there which focus on the implications of relaxing the restrictions ( | ) made in existing models. would be little justification for excluding these 8 Assumption 6 subsumes assumption 4 ; hence, drop- factors. Importantly, these modeling decisions ping 4 requires( dropping) 6 as well. If( decision) utility ( ) ( ) are application-specific: for someapplications coincides with experienced utility, policy nudges which e.g., understanding the effects of income taxes by definition do not enter experienced utility cannot( affect ) on( labor supply , a model featuring separable behavior. I write 4 as a separate assumption to distinguish ) violations of 6 (that) yield new policy tools n from those that do not. ( ) 9 Ramsey 1927 solved 1 subject to 2 – 6 as well as requirement, since this leaves behavior undistorted. Mirrlees the additional( condition) that( not) all goods( can) ( be) taxed. If 1971 expanded the set of policy tools under consideration all goods can be taxed, the problem is trivial: the optimal by( allowing) for nonlinear taxes on income, and subsequent policy is to impose what is effectively a lump-sum tax work in the mechanism design literature allows for a general by taxing all goods at the same rate to meet the revenue set of taxes on consumption and income. 6 AEA PAPERS AND PROCEEDINGS MAY 2015 utility might yield perfectly reasonable predic- this reason e.g., Mullainathan, Schwartzstein, tions, and most economists would not insist on and Congdon( 2012 . Measuring the internal- allowing for complementarity between con- ity e c requires identifying) the impact of an sumption and labor in such cases. Applying the agent’s( ) choice on his own experienced utility, same approach to behavioral economics would much as measuring a traditional externality call for incorporating only the behavioral ele- requires identifying the impact of an agent’s ments that are essential for obtaining accurate choices on other agents’ experienced utilities.10 predictions for the application at hand. Correspondingly, recent research has developed Thus far, I have focused on the positive various methods of estimating internalities e c implications of behavioral economics, as that resemble those used in the literature (on) in Friedman 1953 . Relaxing 4 – 6 also externalities, which are discussed in Section IV yields new welfare( ) implications(. )If( )agents below. have nonstandard experienced utilities, such as The pragmatic value of behavioral econom- reference-dependent preferences, then the wel- ics—new policy tools, better predictions of the fare consequences of policies naturally differ effects of existing policies, and new welfare from the predictions one would obtain from a implications—can ultimately be evaluated only neoclassical model. However, as long as the in the context of real-world applications. The decision and experienced utility are identical next three sections of the paper illustrate these i.e., 6 holds , one can still conduct welfare ideas more concretely in the context of such (analysis( ) using) revealed preference methods applications. analogous to those in the neoclassical model because an agent’s observed choices reveal his II. New Policy Tools: Increasing Retirement experienced utility u c . Savings Welfare analysis ( )in behavioral models becomes more challenging when experienced In this section, I illustrate the ways in which and decision utilities differ, as is the case when behavioral economics can expand the set of pol- agents suffer from behavioral biases such as icy tools available to policymakers. The central inattention or present bias. Since the plan- application that I focus on is increasing retire- ner’s objective is no longer directly related to ment savings, an area where behavioral eco- the agent’s decision utility, one cannot use the nomics has already had a significant impact on agent’s observed choices to recover the welfare policy Madrian 2014 . I begin by summarizing ( ) function u c . As discussed by Kanbur, Pirttilä, recent evidence on the impacts of neoclassical and Tuomala( ) 2006 , this problem is formally tools t, namely, tax subsidies for retirement analogous to nonwelfarist( ) approaches to optimal saving,( ) and then discuss new policy tools policy, in which the planner’s objective differs n—defaults and automatic enrollment plans— from maximizing the agent’s private utility. For that( ) emerge from behavioral models. In the example, Sen 1985 discusses social welfare final subsection, I briefly review other examples functions that (incorporate) notions of individ- of policy tools that have emerged from behav- uals’ capabilities and freedoms in addition to ioral models, such as information provision to their hedonic utilities, while Besley and Coate increase college enrollment rates and loss fram- 1992 model the planner’s objective as a func- ing to increase the impacts of incentive pay for tion( of) income levels rather than utility. teachers. The problem of measuring social welfare when experienced and decision utilities differ bears many similarities to the classic problem 10 of measuring social welfare in the presence of One conceptual difference between externalities and internalities is that other agents’ utilities are exogenously externalities Pigou 1920 . This can be easily ( ) affected by the actions of a given agent in the case of exter- seen by writing the planner’s objective in 1 nalities. With internalities, the agent makes the choice in as v c e c where e c u c v c mea( )- question herself, and hence the planner arguably needs sures( )the+ “externality”( ) (that) = the( agent) − (imposes) a stronger rationale to intervene and overrule the agent’s endogenous decision. That is, the very fact that an agent her- on himself by making suboptimal choices. The self made a choice c increases the probability that the choice term e c is frequently labeled an “internality” ( ) might have been optimal and thus raises the bar for policies in the behavioral public economics literature for that seek to change c. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 7

A. Neoclassical Tools: Subsidies shown in panel A of Figure 1, which plots mean for Retirement Saving capital pension contributions versus taxable income. The figure is constructed by grouping There is growing concern that many people individuals into DKr 5,000 income bins based may not be saving adequately for retirement on their current taxable income relative to the e.g., Poterba 2014 , and policymakers have top tax cutoff, demarcated by the dashed vertical expressed( interest in) increasing household sav- line. It then plots the mean capital pension con- ings rates. What is the best way to achieve this tribution in each bin in each year from 1996 to goal? 2001 versus income. The relationship between The traditional approach to increasing retire- income and capital pension contributions is ment savings is to subsidize saving in retirement stable from 1996 to 1998, the years before the accounts changing t in the model in Section I . reform. In 1999, the marginal propensity to save The United( States federal government spends) in capital pension accounts falls sharply for more than $100 billion per year on subsidies for those in the top bracket: each DKr of additional retirement savings accounts such as 401 k s and income leads to a smaller increase in capital IRAs by granting saving in these accounts( )favor- pension contributions. The changes are substan- able tax treatment Joint Committee on Taxation tial: mean capital pension contributions fell by 2012 . A large empirical( literature has evaluated nearly 50 percent for individuals with income the effects) of these subsidies on savings rates 25,000–75,000 DKr above the top income tax by testing predictions derived from neoclassical cutoff. life-cycle models. This work has obtained mixed The aggregate patterns in panel A of Figure 1 results Poterba, Venti, and Wise 1996; Engen, appear to support the predictions of neoclassical Gale, and( Scholz 1996 because of limitations life-cycle models of savings behavior: reducing in data availability and )because the neoclassical the subsidy for saving in a particular account model does not predict observed savings pat- reduces contributions to that account. However, terns well, as I discuss below. the individual-level responses underlying these In a recent study, Chetty et al. 2014a use aggregate patterns point in a different direc- 41 million observations on the savings( of) all tion. Panel B of Figure 1 plots the distribution Danish citizens from 1995–2009 to present new of changes to individual capital pension contri- evidence on the effects of subsidies on savings butions as a fraction of lagged contributions behavior. I focus on this study here because it for individuals( who were contributing to capital) illustrates the value of incorporating behavioral pensions in the prior year. The sample in this fig- economics into the analysis of canonical policy ure consists of individuals whose incomes place questions. them 25,000–75,000 above the top tax cutoff, The Danish pension system is similar to that the “treatment” group affected by the subsidy in the United States, except that Denmark has reduction.11 The figure plots the distribution two types of tax-deferred savings accounts: cap- of changes in contributions for this group from ital pensions that are paid out as a lump sum 1998 to 1999 the year of the treatment and upon retirement and annuity pensions that are from 1997 to 1998( as a counterfactual. ) paid out as annuities. In 1999, the Danish gov- Panel B of Figure 1 shows that many of the ernment reduced the tax deduction for contribut- individuals in the top tax bracket leave their cap- ing to capital pension accounts from 59 cents per ital pension contributions literally unchanged Danish Kroner DKr to 45 cents per DKr for in 1999 despite the fact that the capital pension individuals in the( top) income tax bracket. The subsidy was reduced. Since any optimizing agent cutoff for the top tax bracket was DKr 251,200 US $38,600 in 1998, roughly the eightieth per- ( ) centile of the income distribution. The deduction 11 The treatment group is defined starting with individuals was unchanged for those in lower tax brackets, DKr 25,000 above the top tax cutoff rather than exactly at the top tax cutoff itself because individuals( face uncertainty and the tax treatment of annuity pension contri- ) butions was also unchanged. in their taxable income when making retirement account Chetty et al. 2014a begin by analyzing the contributions during the year. Since individuals close to the ( ) cutoff might not have expected to be in the top bracket at impacts of this reform on mean capital pension the end of the year, including them could understate the true contributions. The results of this analysis are response to the subsidy change. 8 AEA PAPERS AND PROCEEDINGS MAY 2015

Panel A. Mean retirement account contributions not describe the behavior of all the individuals 12 in the economy. Moreover, a large fraction of 15,000 individuals stop contributing to capital pensions entirely, as shown by the spike in the distribution

10,000 at 100 percent in 1999. Chetty et al. 2014a show− that the entire aggregate reduction( in cap)-

ital pension contributions shown in panel A of 5,000 Figure 1 is driven by the additional 19.3 percent of individuals who stopped contributing to cap-

0 ital pensions when the subsidy was reduced in Pension contribution ( DKr ) 1999. The remaining 80.7 percent of the popula- 0 tion appear to have made no change in their sav- 25,000 50,000 75,000 75,000 50,000 25,000 − − − ings plans in response to the change in subsidies, Income relative to top tax cutoff DKr again contradicting the predictions of the neo- ( ) 1996 1997 1998 classical model. Hence, 80.7 percent of individ- 1999 2000 2001 uals are “passive savers” who are unresponsive to changes in marginal incentives, while 19.3 Panel B. Distribution of changes in percent are “active savers” who behave as the retirement account contributions neoclassical model would predict. 60 Next, Chetty et al. 2014a assess whether the 19.3 percent of individuals( who) stopped contrib-

uting to capital pension accounts reduced their 40 total amount of saving or shifted this money to 19% who stop contributing account for entire aggregate other accounts. They find that roughly one-half reduction

of the reduction in capital pension contributions 20 was offset by increased contributions to annuity pension accounts and the rest was almost Percent of Individuals

0 entirely offset by increased saving in taxable accounts e.g., bank and brokerage accounts . 0 50 ( ) 50 100 Based on this analysis, they conclude that each 100 − − $1 of tax expenditure on retirement savings sub- Percent change in pension contributions 1997 to 1998 1998 to 1999 sidies increases retirement saving by approximately $0.01, with an upper bound on the 95 Figure 1. Effects of Reduction in Subsidy for percent confidence interval of $0.10. Retirement Savings Accounts in Denmark There are two lessons of this analysis from the perspective of behavioral public economics. Notes: This figure reproduces Figures Va and VIb in Chetty First, responses that appear to be consistent with et al. 2014a . Panel A plots mean total individual plus employer( capital) pension contributions for (individuals with optimization in the aggregate may mask signifi- income in) each DKr 5,000 income bin within DKr 75,000 cant deviations from optimization at the individ- of the top income tax threshold, in each year from 1996 to ual level. Second, the standard tools suggested 2001. Panel B plots the distribution of changes to individual by neoclassical models are not very successful capital pension contributions, as a percentage of lagged indi- at least in some settings in increasing savings vidual pension contributions, for individuals who are DKr ( ) $25,000–$75,000 above the top tax cutoff in 1998 and 1999 rates because they appear to induce only a small (the treatment group affected by the subsidy reduction). group of financially sophisticated individuals Panel B restricts the sample to individuals who were con- to respond, and these individuals simply shift tributing to capital pension accounts in the previous year. assets between accounts. These results naturally lead to the question of whether other policy tools—perhaps those that directly target passive at an interior optimum should change capital 12 A neoclassical life-cycle model can generate zero pension contributions by some nonzero amount response if wealth and price effects happen to offset each when the subsidy is reduced, this fact immedi- other exactly. However, this is a knife-edge measure zero ately implies that the neoclassical model does case. ( ) VOL. 105 NO. 5 RICHARD T. ELY LECTURE 9 savers—can be more effective in increasing 12 Employer Pensions 5.64 = saving. Individual Pensions 0.56 = − Taxable Savings 0.02 = B. New Policy Tools: Defaults and Automatic 8 Enrollment Employer pensions Taxable saving A large body of research over the past 4 Individual pensions decade has found that employer defaults have

a large impact on contributions to retirement 0 accounts despite leaving individuals’ incentives Contribution or t axable saving rat e ( percent of income ) 4 2 0 2 4 unchanged. In an influential paper, Madrian and − − Shea 2001 show that an opt-out system—in Year relative to firm switch ( ) which employees are automatically enrolled into Figure 2. Effects of Employer Contributions to their company’s 401 k plan but are given the Retirement Accounts option to stop contributing—increases( ) participa- Notes: This figure reproduces Figure Ib in Chetty et al. tion rates in 401 k plans from 20 percent to 80 ( ) 2014a . The figure analyzes a set of workers who switched percent at the point of hire. This result has since to( a new) firm whose employer pension contribution rate was been replicated in numerous other settings e.g., at least 3 percentage points of labor income higher than their Choi et al. 2002 . Similarly, Thaler and Benartzi( previous firm. The x-axis of the figure is the calendar year 2004 show that) individuals who enroll in plans relative to the year the worker switched firms, so that year 0 ( ) represents the first year the worker is employed at the new to escalate retirement contributions over time firm. The figure plots mean employer and individual retire- rarely opt out of these arrangements in subse- ment account contributions as well as individuals’ taxable quent years. savings in each year, all measured as a percentage of cur- While defaults clearly have substantial effects rent labor income. The sample consists of workers for whom data is available for event years 4, 4 , so that the number on contributions to retirement accounts, it is crit- of observations is constant through[− +the] figure. The sample ical to determine whether these larger retirement also includes only workers with positive individual pension contributions come at the expense of less saving contributions prior to the switch, i.e., those who are able to in non-retirement accounts or actually induce reduce individual pension contributions when employer con- individuals to consume less as required to raise tributions rise. The figure lists the change in each form of ( savings (as a percentage of labor income) from year 1 to total savings rates . Most studies to date have year 0. − not been able to estimate) such crowd-out effects because they do not have data on individuals’ full portfolios. Chetty et al. 2014a are able to fully offset changes in employer contributions in resolve this problem because( the )Danish data this manner. they use contain information on savings in all Chetty et al. 2014a test this prediction and accounts. They study the impacts of defaults on estimate the causal( effect) of employer pension total savings by exploiting variation in employ- contributions on savings rates using an event- ers’ contributions to retirement accounts across study research design, tracking individuals firms. In Denmark, employers and individuals who switch firms. This design is illustrated in contribute to the same accounts, so changes Figure 2, which plots the savings rates of indi- in employer contributions are analogous to viduals who move to a firm that contributes at changes in defaults. Consider an individual least 3 percentage points more of labor income who is contributing DKr 2,000 to his retirement to their retirement accounts than their previous account. Suppose his employer decides to take firm. Let year 0 denote the year in which an indi- DKr 1,000 out of his pay check and contribute vidual switches firms and define event time rel- it to his retirement account, so the individual’s ative to that year e.g., if the individual switches total compensation stays fixed. Since the indi- firms in 2001, year( 1998 is 3 and year 2003 vidual could fully offset this change by reduc- is 2 . The sample consists of− individuals who ing his personal contribution to DKr 1,000, the are+ observed) for at least four years both before employer contribution effectively changes the and after the year of the firm switch to obtain a “default” contribution rate without changing the balanced panel and who make positive( individ- individual’s budget set. Indeed, the neoclassical ual pension contributions) in the year before they life-cycle model predicts that individuals should switch firms to limit the sample to individuals ( 10 AEA PAPERS AND PROCEEDINGS MAY 2015 who are able to offset the increase in employer levels of wealth, and are more likely to have contributions . taken finance courses in college. Hence, defaults The series )in squares in Figure 2 plots total not only have a larger impact on aggregate sav- employer contributions to capital and annuity ing, but also target those who are saving the least accounts . By construction,( employer contri- for retirement more effectively than existing butions jump) in year 0, by an average of 5.64 price subsidies. percent of labor income for individuals in this The broader lesson of this work is that defaults sample. The series in triangles plots the individ- make it feasible to achieve outcomes that cannot ual’s own pension contributions. Individual pen- be achieved with subsidies. Given an exogenous sion contributions fall by 0.56 percent of income policy objective of increasing saving, this empir- from year 1 to year 0, far less than the increase ical finding has practical value even if the under- in employer− contributions. Finally, the series in lying behavioral assumptions remain debated.14 circles in Figure 2 plots savings in all other tax- Indeed, in light of the work by Madrian and Shea able accounts. Savings in taxable accounts are 2001 and Thaler and Benartzi 2004 , defaults essentially unchanged around the point of the have( already) started to be systematically( ) applied firm switch. These findings show that increases to increase retirement savings by both private in employer pension contributions are not offset companies and governments. significantly by less saving in other accounts— Although the empirical findings on defaults that is, employer defaults effectively increase have great value, understanding the theory that total saving. Building on event studies of this explains savings behavior remains useful for form, Chetty et al. 2014a estimate that a $1 two reasons. The first is extrapolation: predict- increase in employer( retirement) account contri- ing the impacts of defaults in other contexts butions coupled with a $1 reduction in salary so e.g., larger changes in default rates requires that total compensation is unchanged increases( a( theory of saving that explains why) defaults individuals’ net savings rate by approximately) matter, such as the model of procrastination pro- $0.85. These savings increases persist for more posed by Carroll et al. 2009 . Second, welfare than a decade and lead to greater wealth balances analysis requires a model( of )savings behavior. at retirement, showing that employer defaults Should we be trying to increase the amount peo- have long-lasting effects on savings behavior. ple save for retirement? If so, what is the optimal Since the neoclassical model predicts full default savings rate? These optimal policy ques- offset of changes in employer defaults, the fact tions cannot be answered without specifying the that a $1 increase in defaults raises total savings underlying behavioral model. I return to these by $0.85 implies that 85 percent of individuals normative questions in Section IV. are “passive savers” who are inattentive to their From a methodological perspective, the retirement plans and simply follow the default research on retirement savings over the past option.13 This estimate is consistent with the decade captures the essence of the pragmatic finding discussed above that roughly 80 per- approach to behavioral economics. Much of the cent of agents respond passively to changes research in this literature has been motivated by in subsidies. The 15–20 percent of individuals finding the most effective way to increase sav- who respond actively to price incentives are ings rates rather than testing the assumptions also much more likely to offset employer pen- of neoclassical models. For example, Chetty sion contributions by reducing saving in other et al. 2014a did not set out to test whether accounts. These active savers tend to be more agents( optimize) in making savings decisions; financially sophisticated e.g., they rebalance their portfolios more frequently( , have higher ) 14 For example, the evidence on defaults could be explained by a model with inattentive agents or a signaling model in which individuals who are uncertain about how 13 Crowd-out could be less than 100 percent even in the much they should save treat the default as an informative neoclassical model if individuals hit the corner of 0 individ- signal about the correct savings rate. Distinguishing between ual pension contributions. Chetty et al. 2014a show that these “behavioral” and “rational” models is only useful if the this effect accounts for very little of the( imperfect) crowd- two models generate different predictions in some domains; out that is observed because even individuals who are well from a pragmatic perspective, there is no inherent advantage within the interior do not offset most of the changes in to the “rational” signaling model if it does not provide better employer contributions. predictions. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 11 instead, the goal of the study was to evaluate cantly increases the probability that their chil- the effectiveness of alternative policies to raise dren attend college. Similarly, Hoxby and Turner retirement saving, with an initial focus on tax 2014 show that providing high-achieving stu- subsidies t. In the process of studying the dents( )from low-income families with simple data, it became( ) evident that individuals’ behav- information about the college application pro- ior was better explained by behavioral models cess and colleges’ net costs given their families’ that generate passive choice. This naturally led particular financial situation increases the prob- to the exploration of new policy tools n such ability that children apply to and attend more as employer defaults. Although one could( ) have selective colleges. The interventions imple- approached this policy question from a strictly mented in both of these studies are inexpensive; neoclassical perspective—focusing exclusively for instance, the intervention implemented by on the impacts of price subsidies—the analy- Hoxby and Turner 2014 cost $6 per student. sis of new policy tools motivated by behavioral Information and application( ) assistance thus pro- models yields richer insights and ultimately bet- vide new tools to raise college attendance rates ter methods of increasing retirement saving. that may be much more cost-effective at the margin than existing policy tools, such as grants C. Other Applications or loans. Loss-Framing and Teacher Performance.— In this section, I briefly summarize four other Fryer et al. 2012 show that framing teacher applications in which insights from behavioral incentives as( losses) relative to a higher salary economics have been used to develop new pol- rather than bonuses given for good performance icy tools.15 increases the impact of these incentives on student performance. In particular, teachers who Simplification and Choice of Health Plans.— are given bonuses in advance and told that the Bhargava, Loewenstein, and Sydnor 2014 money will be taken back if their students do study a large US firm where employees (choose) not improve sufficiently generate significantly from a menu of health insurance plans that vary higher student test scores than those paid a con- in several dimensions e.g., deductibles, copay ventional performance bonus. Such loss-framing rates, out-of-pocket maximums,( etc. . They has no additional fiscal cost to the government show that many individuals choose strictly) dom- and thus provides an attractive new policy tool inated health insurance plans, i.e., plans that to improve students’ outcomes. reduce their payoffs in all states of the world. Their findings imply that simplifying the set Social Comparisons and Energy of options given to individuals can potentially Conservation.—Allcott 2011 shows that improve their decisions. Interestingly, Bhargava, sending households a letter( informing) them Loewenstein, and Sydnor 2014 find that sub- about their energy usage relative to that of their optimal choices are particularly( common) among neighbors reduces mean energy consumption. low-income households, suggesting that com- This finding is consistent with models of social plexity may have negative distributional conse- comparisons in which individuals are concerned quences in addition to reducing average welfare. about how their behavior compares with others’ behavior. Such social comparisons are now com- Application Assistance and College monly used by utility companies alongside con- Attendance.—Bettinger et al. 2012 show that ventional policy tools such as price increases. offering information and assistance( )in complet- All of these studies exemplify the pragmatic ing the Free Application for Federal Student Aid approach to behavioral economics: their goal FAFSA form to low-income families signifi- is to evaluate the efficacy of new policy tools ( ) suggested by behavioral models rather than test-specific assumptions of neoclassical or 15 Perhaps the most concrete evidence that behavioral behavioral models. In some cases, it is not even economics has expanded the set of policy tools available to fully clear exactly what the underlying behav- policymakers is the creation of “nudge units” in the US and UK governments that are tasked with formulating and testing ioral model is. For instance, application assis- new policies that do not involve direct changes in financial tance could matter because individuals exhibit incentives, such as defaults, framing, and social persuasion. inertia, lack information, or procrastinate in 12 AEA PAPERS AND PROCEEDINGS MAY 2015 filling out forms. Similarly, there are various EITC amounts depend upon a tax filer’s tax- potential theories—“rational” models based able income, marital status, and number of chil- on signaling effects and “behavioral” models dren. Panel A of Figure 3 plots EITC amounts as based on relative comparison utilities—that a function of taxable income for single tax filers could explain tastes for conformity in electricity with 1 versus 2 or more children, expressed in consumption. Despite this uncertainty about the real 2010 dollars. EITC refund amounts first underlying assumptions, the new policy tools increase linearly with earnings the “phase-in” identified as a result of incorporating behavioral region , then are constant over (a short income considerations have pragmatic value in expand- range )the “plateau” , and are then reduced lin- ing the set of outcomes that policymakers can early (the “phase-out”) region . The phase-in achieve.16 subsidy( rate is 34 percent for taxpayers) with 1 child and 40 percent for those with 2 or more III. Better Predictions: Effects of Income Taxes children; the corresponding phase-out tax rates on Labor Supply are 16 percent and 21 percent. Because individuals face payroll and other taxes, they obtain the Even if they do not generate new policy tools, largest tax refund when their taxable income behavioral models can still be useful in predict- exactly equals the first kink of the EITC sched- ing the impacts of existing policies. This sec- ule, which is $8,970 for filers with 1 child and tion demonstrates this point by showing that the $12,590 for those with 2 or more children.17 effects of the Earned Income Tax Credit EITC One of the primary goals of the EITC is to on labor supply decisions are better predicted( by) increase the labor supply of low-wage work- a model that allows for imperfect knowledge of ers by increasing their effective wage rates. A the tax code, an ancillary condition d that plays large literature has evaluated whether the EITC no role in neoclassical models of labor( ) supply. is effective in achieving this goal by estimating I begin by discussing recent evidence which labor supply elasticities in neoclassical mod- shows how differences in knowledge about the els of labor supply. This work has found clear EITC across areas lead to spatial variation in evidence that the EITC increases labor force its impacts on reported taxable income. I then participation, but mixed evidence on the effects turn to the program’s impacts on real labor sup- of the EITC on hours of work and earnings con- ply decisions. In the final subsection, I discuss ditional on working Eissa and Hoynes 2006; experiments that evaluate whether information Meyer 2010 . ( provision can be used as a new policy tool to Chetty, Friedman,) and Saez 2013 —hence- increase the impacts of the EITC. forth, CFS—study the impacts( of )the EITC using new data from de-identified federal A. Effects of the Earned Income Tax Credit on income tax returns covering the US population Reported Income from 1996–2009. These administrative data permit a much more precise analysis of the EITC’s The EITC is the largest means-tested cash impacts because they are several orders of transfer program in the United States. In 2012, magnitude larger than the survey datasets used 27.8 million tax filers received over $63 billion in prior research. CFS’s core analysis sample in federal EITC payments Internal Revenue includes 78 million taxpayers and 1.1 billion Service 2012, Table 2.5 . The( federal EITC observations on income. was expanded to its current) form in 1996, and CFS’s initial research plan—which had remained essentially unchanged over the next 15 no connection to behavioral models—was to years aside from inflation indexation. exploit state-level differences in EITC “top up” policies to identify the effects of the EITC. For example, Kansas has a state EITC program that 16 As noted above, understanding the underlying theory is provides a 17 percent match on top of the fed- still valuable for making extrapolations and for welfare anal- eral EITC amount, whereas Texas has no state ysis. For example, Allcott 2014 shows that the efficacy of the social comparison intervention( ) is highly heterogeneous across cities. If one had a precise theory of why social comparisons matter, one might be able to better predict which 17 Tax filers with no dependents are eligible for a very places would benefit most from this new policy tool. small EITC, with a maximum refund of $457 in 2010. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 13

Panel A. Federal EITC schedule for single filers minus the income threshold for first kink of with children in 2008 the EITC schedule shown in panel A the ( $5K refund-maximizing kink . The figure plots the percentage of tax filers )in $1,000 bins centered $4K around the refund-maximizing kink. In Texas, EITC claimants have a sub- $3K stantial excess propensity to “bunch” at the

amount refund-maximizing kink, a result first docu- $2K mented at the national level by Saez 2010 .

EITC ( ) $1K More than 5 percent of EITC claimants report income within $500 of this kink in Texas, much $0K higher than the density at surrounding income $0 $10K $20K $30K $40K levels. This is precisely the behavioral response Total earnings real 2010 $ that one would expect in a neoclassical model ( ) with a nonlinear budget set: since the effec-

One child Two or more children tive wage rate falls by 40 percent once one crosses the kink, many optimizing individuals should choose to report income exactly at the Panel B. Taxable income distributions for EITC refund-maximizing kink. claimants in Kansas and Texas The degree of sharp bunching is much lower 5% in Kansas than in Texas. In Kansas, the fraction of individuals at the refund-maximizing kink is 4% only slightly higher than at other nearby income levels. This lower degree of responsiveness to 3% the EITC is not what one would have predicted from a neoclassical model, as Kansas offers its 2% residents a larger EITC than Texas. To understand what drives this heterogeneity 1% Texas in EITC response across areas, CFS estimate the Kansas

Percent of EITC claimants degree of sharp bunching at the refund-maxi- 0% mizing kink across all the three-digit ZIP codes $10K 0 $10K $20K − ZIP-3 in the United States. CFS define the Taxable income degree( )of sharp bunching in a ZIP-3 c in year relative to first kink of EITC schedule t bc t as the percentage of EITC claimants with Figure 3. Effects of the EITC on Reported Taxable children who report total earnings within $500 Income of the first EITC kink and have nonzero self-employment income. CFS focus on self-employ- Notes: Panel A shows the EITC schedule for single filers ment income to define sharp bunching because with children in 2008 as a function of their reported taxable the excess mass at the refund-maximizing kink income. The vertical lines denote the refund-maximizing kink for each group. Panel B reproduces Figure 1a in Chetty, is driven entirely by self-employed individu- Friedman, and Saez 2013 using data for EITC claimants als. The distribution for wage earners exhibits with children in Texas( and) Kansas. These distributions are no spike in its density at the kink see panel A histograms with $1,000 bins centered around the first kink of Figure 6 below . Sharp bunching( is driven of the EITC schedule. Taxable income is the total amount of ) earnings used to calculate the EITC, and is equivalent to the purely by the self-employed because self-em- sum of wage earnings and self-employment income reported ployed individuals directly report their income on Form 1040 for most tax filers. to the IRS, making it easier for them to manipulate their reported income to exactly match the amount needed to obtain the largest refund.18

EITC program. Panel B of Figure 3 plots the 18 Wage earners have much less scope to manipulate distribution of taxable income for EITC claim- their reported income, as it is reported directly to the IRS by ants with children in Kansas and Texas. The employers. I discuss the effects of the EITC on wage earners x-axis of panel B of Figure 3 is taxable income in the next subsection. 14 AEA PAPERS AND PROCEEDINGS MAY 2015

Panel A. Sharp bunching in 1996 Panel B. Sharp bunching in 1999

Panel C. Sharp bunching in 2002 Panel D. Sharp bunching in 2005

Panel E. Sharp bunching in 2008 Amount of sharp bunching

4.1% > 2.8–4.1% 2.1–2.8% 1. 8 –2.1% 1. 5 –1.8% 1. 2 –1.5% 1. 1 –1.2% 0.9–1.1% 0.7–0.9% 0.7% <

Figure 4. Geographical Variation in EITC Sharp Bunching by Year

Notes: This figure reproduces Appendix Figure 3 in Chetty, Friedman, and Saez 2013 . It plots self-employed sharp bunching rates by 3-digit ZIP code ZIP-3 in 1996, 1999, 2002, 2005, 2008. Self-employed( sharp) bunching is defined as the fraction of EITC-eligible households( with) children whose total income falls within $500 of the first kink point of the EITC schedule and who have nonzero self-employment income. The observations are divided into deciles pooling all years so that the decile cut points are fixed across years. Each decile is assigned a different color on the maps, with darker shades representing higher levels of sharp bunching.

Figure 4 presents heat maps of the amount of bc t into ten deciles, pooling all of the years of the self-employed sharp bunching across ZIP-3s in sample so that the decile cut points remain fixed the US in 1996, 1999, 2002, 2005, and 2008. This across years. Deciles with higher levels of sharp figure is constructed by dividing theestimates of bunching bc t are represented with darker shades VOL. 105 NO. 5 RICHARD T. ELY LECTURE 15 on the map. In 1996, shortly after the EITC 120 expanded to its current form, sharp bunching was prevalent in very few areas southern Texas, 100 ( 59.7 New York City, and Miami . Bunching then β = 5.7 ) ( ) spread gradually from these areas to other parts 80 of the country over time. Much of the variation 60 in these maps is within states, again suggesting 6.0 p-value for diff. = β 6.2 that differences in state EITC policies are not ( ) in slopes: p 0.0001 40 < the key determinant of variation in behavioral Change in EITC refund ( $ ) 1% 0.5% 0% 0.5% 1% responses to the program. − − In light of this evidence, CFS set out to deter- Change in ZIP-3 sharp bunching mine why behavioral responses to the EITC Figure 5. Effects of Moving to Areas with Higher vs. vary so much across areas of the United States. Lower Sharp Bunching on EITC Refunds Given the spatial diffusion pattern in Figure 4, one plausible model is that the variation stems Notes: This figure reproduces Figure 3b in Chetty, Friedman, and Saez 2013 . The sample in this figure consists of tax- from differences in knowledge about the EITC’s ( ) incentive structure and learning over time. While payers who move across 3-digit ZIP codes. The figure plots changes in EITC refund amounts from the year before the the neoclassical model typically assumes that move to the year after the move vs. differences in prior res- all individuals are fully informed about the tax idents’ sharp bunching rates in the new and old ZIP-3s. The code, in practice many families seem to have lit- figure groups individuals into bins of width 0.05 based on tle understanding of the marginal incentives cre- their change in sharp bunching and plots the mean change in EITC refund amounts within each bin. The solid lines ated by the EITC e.g., Smeeding, Ross Phillips, show linear regressions estimated on the individual-level and O’Connor 2002( . ) data separately for the observations above and below 0. The To test whether differences in knowl- estimated slopes are reported next to each line along with edge explain the spatial variation, CFS con- standard errors clustered by bin. sider individuals who move across ZIP-3s. The knowledge model predicts that moving to a higher-bunching area e.g., from Kansas CFS go on to show that areas with a larger to Texas should increase (responsiveness to density of EITC claimants tend to have much ) the EITC. But moving to a lower-bunching higher levels of sharp bunching bc t , consistent area e.g., from Texas to Kansas should not with a model in which knowledge diffuses affect( responsiveness to the EITC,) as individ- through local networks. In sum, a model that uals should not forget what they have already accounts for differences in knowledge and learn- learned. Figure 5 shows that this is precisely ing—i.e., a model where decision utility v c d what one finds in the data. This figure is a depends upon information d —makes much( bet| )- binned scatter plot of changes in EITC refund ter predictions about the effects of the EITC than amounts from the year after the move relative a model which assumes that all agents are fully to the year before the move versus the change informed about the tax code.19 in sharp bunching rates among prior residents in the destination and origin ZIP-3s. The EITC B. Earnings Responses: Using Behavioral refund amount is a simple summary measure Models to Generate Counterfactuals of the concentration of the income distribution around the refund-maximizing kink. The figure As discussed above, the sharp bunching is constructed by binning the x-axis variable response to the EITC is driven entirely by Δ bc t into intervals of width 0.05 percent and plot- ting the means of the change in EITC refund within each bin. Individuals to the right of the 19 One may argue that models of imperfect knowledge dashed line are moving to higher-bunching and learning are not “behavioral” because they can poten- areas, while those to the left are moving to low- tially be explained by a neoclassical model with search costs er-bunching areas. There is a sharp break in the for acquiring information. The key point here is that incorporating such features into the analysis of taxes and labor slope at 0: increases in bc t raise EITC refunds, supply is useful. Whether a model is labeled as “neoclas- but reductions in bc t leave EITC refunds sical” or “behavioral” is inconsequential; what matters is unaffected. whether that model accurately predicts behavior. 16 AEA PAPERS AND PROCEEDINGS MAY 2015 self-employed individuals. Audit data reveal Panel A. US population that most of this sharp bunching is driven by 3.5 4k misreporting of self-employment income rather than real changes in work patterns Chetty, 3 EITC 3k Friedman, and Saez 2013 . While understand( - 2.5

) amount ( $ ) ing the effects of the EITC on reported income 2 age-earners 2k is useful, the objective of the EITC is to change 1.5 the amount that people actually work and con- 1 tribute to the economy, not just the income they 1k report to the IRS. To study the impacts of the 0.5 EITC on labor supply decisions, CFS charac- Percent of w 0 0k terize the program’s effects on the distribution $0 $10K $20K $30K of wage earnings, excluding self-employment W-2 wage earnings income. Because wage earnings are directly reported by employers to the IRS on W-2 forms, Panel B. Highest sharp bunching decile versus individuals have little scope to misreport wage lowest sharp bunching decile earnings. Misreporting rates for wage earnings 3.5 4k are below 2 percent Internal Revenue Service ( 3 EITC 1996, Table 3 . Hence, changes in wage earn- 3k ings can be )interpreted as changes in real 2.5 amount ( $ ) labor supply behavior rather than just reported 2 age-earners 2k income. 1.5

Panel A of Figure 6 plots the distribution of 1 1k wage earnings using data from W-2 forms in 0.5 Lowest sharp bunching decile the United States( as a whole for EITC claim) - Highest sharp bunching decile Percent of w 0 ants with one child. Unlike with the self-em- 0k ployed, there is no sharp spike in the density $0 $5K $10K $15K $20K $25K $30K $35K around the refund-maximizing kink. This is W-2 wage earnings because wage earners face frictions in choos- Figure 6. Effects of the EITC on Wage Earnings ing their labor supply. For example, workers typically cannot Altonji and Paxson 1992 , Notes: This figure reproduces Figure 5a in Chetty, Friedman ( ) and Saez 2013 . Panel A plots W-2 wage earnings distribu- making it difficult for them to target a specific ( ) level of earnings precisely. Because of these tions for single wage earners with one child, pooling data from 1999–2009 in the United States as a whole. Panel B frictions, any effects of the EITC on real wage focuses on the subset of individuals living in the highest and earnings are too diffuse to detect without a lowest self-employment sharp bunching deciles, i.e. areas counterfactual—i.e., an understanding of what with the highest and lowest levels of information about the the earnings distribution in panel A would look EITC. Self-employed sharp bunching is defined as the percentage of EITC claimants with children in the ZIP-3-by- like in the absence of the EITC. This problem year cell who report total earnings within $500 of the first lies at the root of why estimating the effects EITC kink and have nonzero self-employment income. The of the EITC has been challenging, as there are series in triangles includes individuals in ZIP-3-by-year cells few good counterfactuals for programs that are in the highest self-employed sharp bunching decile, while implemented primarily at the national level and the series in circles includes individuals in the lowest sharp bunching decile. Each distribution is a histogram showing are changed relatively infrequently. the percentage of observations in $1,000 bins. The EITC The spatial variation in knowledge about the schedule is shown on the right y-axis, and the dashed ver- EITC proves to be very useful in obtaining such tical lines depict the plateau region of the EITC schedule a counterfactual and identifying the impacts of where individuals obtain the largest EITC refund amounts. the EITC on wage earnings. The idea is straight- forward: areas with no information about the EITC can be used as a counterfactual for behav- To implement this strategy, CFS proxy for ior in the absence of the marginal incentives cre- the level of information about the EITC in each ated by the program. Intuitively, individuals who ZIP-3 using the level of sharp bunching among do not know about a program cannot respond to the self-employed, bc t. Panel B of Figure 6 plots its marginal incentives. the distribution of wage earnings for individu- VOL. 105 NO. 5 RICHARD T. ELY LECTURE 17 als with one child living in ZIP-3s in the high- Panel A. Wage earnings distributions in est decile of sharp bunching such as southern the year before rst child born ( Texas versus those living in the lowest decile of 6% ) sharp bunching such as Kansas . There is sig- Lowest sharp bunching decile nificantly more ( mass in the plateau) region of the Highest sharp bunching decile EITC—between the income levels of approxi- 4% mately $9,000–$16,000—in high-information age-earners high self-employed sharp bunching areas 2% than( low-information areas. This suggests) that the EITC induces individuals to take jobs that generate earnings that are roughly in the range Percent of w 0% that yields the largest EITC refunds, even if they $0 $10K $20K $30K $40K cannot perfectly target the refund-maximizing W-2 wage earnings kink itself. The comparisons across areas in panel B of Panel B. Wage earnings distributions in Figure 6 could be biased by omitted variables; the year of rst child born for instance, the industrial structure in southern 6% Texas is different from that in Kansas, which Lowest sharp bunching decile could lead to differences in the distribution Highest sharp bunching decile of wage earnings for reasons unrelated to the 4% incentive structure of the EITC. To address this age-earners concern, CFS study changes in wage earnings 2% around childbirth. Individuals without children are essentially ineligible for the EITC, and hence the birth of a first child generates sharp variation Percent of w 0% in marginal incentives. Panel A of Figure 7 plots $0 $10K $20K $30K $40K the distribution of wage earnings for individuals W-2 wage earnings in the highest- and lowest-information deciles in Figure 7. EITC and Wage Earnings Distributions: the year before their first child is born. Panel B Changes around Childbirth replicates panel A using data from the year in which the first child is born. There are no differ- Notes: This figure reproduces Figure 6 in Chetty, Friedman, and Saez 2013 . The sample includes wage earners living ences in the distribution of wage earnings prior ( ) to childbirth across areas, but as soon as the first in the highest and lowest self-employment sharp bunching deciles, i.e. areas with the highest and lowest levels of infor- child is born, the number of individuals in the mation about the EITC. Self-employed sharp bunching is EITC refund-maximizing plateau region rises in defined as the percentage of EITC claimants with children in high-information areas relative to low-informa- the ZIP-3-by-year cell who report total earnings within $500 tion areas. Apparently, people are more likely to of the first EITC kink and have nonzero self-employment continue to work and maintain earnings between income. Panel A plots W-2 wage earnings distributions in the year before child birth. Panel B replicates these distributions $9,000–$16,000 after they have a child in areas for the year of child birth. Each distribution is a histogram with better knowledge about the EITC’s incen- showing the percentage of observations in $1,000 bins. The tive structure. dashed vertical lines depict the plateau region of the EITC Building on this approach, CFS show that the schedule for single individuals with one child, where indi- EITC primarily induces increases in earnings in viduals obtain the largest EITC refund amounts. the phase-in region rather than reductions in the phase-out region. They therefore conclude that the EITC is quite effective in increasing labor focus on short-run changes in behavior around supply, as intended. The responses to the EITC policy reforms. These studies may have detected are largest in areas with dense EITC popula- extensive-margin participation responses tions, where knowledge is more likely to spread. because knowledge (about the higher) return to In addition to explaining the spatial variation working diffused more quickly than knowledge in the effects of the EITC, information diffusion about how to optimize on the intensive margin. can also explain findings from the prior liter- Indeed, surveys show that the knowledge that ature on the EITC. Most studies of the EITC working can yield a large tax refund—which 18 AEA PAPERS AND PROCEEDINGS MAY 2015 is all one needs to know to respond along the therefore be used to construct new comparison extensive margin—is much more widespread groups to identify treatment effects. than knowledge about the nonlinear marginal incentives created by the EITC e.g., Liebman C. Providing Information about the EITC 1998; Romich and Weisner 2002( . This pattern of knowledge diffusion is consistent) with Given the preceding evidence, a natural ques- a model of rational information acquisition, as tion is whether we can increase the impacts of reoptimizing in response to a tax reform on the the EITC by providing more information about extensive margin has first-order large bene- the program. That is, can one use the insight that fits, whereas reoptimizing on the intensive( ) mar- knowledge mediates the effects of the EITC to gin has second-order small benefits Chetty develop new policy tools n as in Section II 2012 . ( ) ( rather than just predict the effects( of the existing) CFS’s) analysis illustrates two lessons regard- policies more precisely? ing the pragmatic value of behavioral econom- Recent studies have investigated this ques- ics for public policy that can be translated to tion using experiments that provide information other applications. First, incorporating behav- about the EITC. Chetty and Saez 2013 report ioral features into the model in this case, dif- results of an experiment with 43,000( EITC) cli- ferences in knowledge helps( us better predict ents of H&R Block, in which one-half of the tax the impacts of existing )policies in this case, the filers were randomly selected to receive infor- effects of the EITC on income reporting( behav- mation from their tax preparer about the mar- ior . Second, behavioral models can be used to ginal incentive structure of the EITC. Chetty and generate) new counterfactuals to estimate pol- Saez find that this intervention had no effect on icy impacts that would otherwise be difficult to earnings in the subsequent year on average.20 identify, such as the effect of the EITC on wage This finding suggests that it is difficult to manip- earnings. Similar approaches can be applied ulate information about marginal incentives to identify reduced-form treatment effects in through policy even though knowledge about many other contexts. For example, recent stud- the EITC affects behavioral responses to the ies have shown that individuals exhibit iner- program. This could be because information tia in choosing health insurance plans Handel from tax preparers has much smaller effects on 2013; Ericson 2014 . Such inertia creates( dif- individuals’ perceptions than information pro- ferences in the health) insurance plans that indi- vided on a more regular basis by trusted friends. viduals have depending upon what plans were Given the apparent challenge in informing indi- offered when they joined their current company. viduals about the EITC, an alternative approach Under the plausible and potentially testable is to include the EITC directly in individuals’ assumption that individuals’( underlying health) paychecks as an automatic wage subsidy. For does not vary at a high frequency across entry instance, if individuals were quoted an hourly cohorts within a company, one could exploit wage rate of $14 per hour instead of $10 per the cross-cohort variation arising from differ- hour by their employers, they would not have to ences in plan availability to identify the impacts think about the EITC when making labor supply of insurance plans on health care spending and decisions at all, and might respond more to the health outcomes. As another example, Gallagher higher wage rate.21 and Muehlegger 2011 show that tax rebates Bhargava and Manoli 2014 conduct an to buy energy-efficient( ) hybrid cars have much experiment involving 35,000( individuals) who larger effects on hybrid car sales if they are framed as sales tax rebates given at the point 20 of purchase rather than income tax rebates paid Chetty and Saez 2013 find evidence of heterogeneity in treatment effects (across) tax preparers, with some tax when individuals file their income tax returns. preparers inducing larger earnings responses than others. By comparing the subsequent behavior of indi- They interpret this finding as evidence that persuasion by tax viduals who get tax rebates framed in different preparers may matter more than raw information about the ways, one may be able to evaluate the causal EITC’s parameters. 21 A practical complication in implementing this proposal effects of owning a hybrid car on driving behav- is that EITC amounts are currently based on annual house- ior. The general point is that behavioral models hold income, and hence the marginal subsidy is not known offer new insights into selection models, and can until a household’s annual income is fully determined. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 19 were eligible for the EITC but did not file the predicting the effects of policies on behavior. tax forms needed to claim it. Approximately Though such predictions are a key input into 25 percent of EITC-eligible individuals do not economic analysis, understanding the effects of file the paperwork needed to take up the credit. policies on social welfare is equally important. Bhargava and Manoli 2014 find that mailing This section turns to the welfare implications eligible individuals simplified( ) information about of behavioral models. I illustrate these impli- the EITC raises EITC take-up rates significantly. cations using an application to neighborhood One potential explanation for why providing effects and housing voucher policies. I begin information increases EITC take-up rates but by summarizing a set of empirical results on appears to have little mean impact on earn- neighborhood effects and then discuss neoclas- ings responses is that take-up generates larger sical and behavioral models that fit these facts. net utility gains than changing labor supply. I then discuss optimal policy in neoclassical Individuals may rationally pay more attention versus behavioral models, focusing on recent to information that they have left money on the work that develops nonpaternalistic methods of table which can be claimed at little or no cost welfare analysis in behavioral models. Finally, relative( to information that their marginal wage) I consider implications for optimal policy when differs from what they thought which requires we are uncertain about whether the underlying real work to generate gains, and (thus yields sec- positive model is neoclassical or behavioral. ond-order benefits . Testing this explanation and developing new models) of when and how knowl- A. Three Facts about Neighborhood Effects edge can be manipulated through policy would be a very useful direction for future research. One of the most important decisions families In determining whether it is desirable to pro- make is where to live. A large body of research vide more information about the EITC, it is in sociology and economics has investigated also important to consider general equilibrium the consequences of neighborhood environ- effects, as in neoclassical models. Leigh 2010 mental conditions on children’s and adults’ and Rothstein 2010 present evidence that( part) outcomes e.g., Jencks and Mayer 1990; Cutler of the benefits( of the) EITC accrue to employers, and Glaeser( 1997; Sampson, Morenoff, and who reduce wage rates in equilibrium given the Gannon-Rowley 2002 . Recent work has used outward shift in the labor supply curve induced newly available administrative) data to identify by the EITC. Making the EITC more salient— three empirical results about the causal effects especially by including it in individuals’ pay- of neighborhoods that motivate the analysis in checks as discussed above—could potentially this section. further reduce wage rates in equilibrium, reduc- First, children’s long-term outcomes vary ing the redistributive value of the program. significantly across neighborhoods conditional Hence, there may be a trade-off between increas- on parent income. Using data from population ing labor supply and providing redistribution in tax records covering all children born in the choosing how to inform individuals about the United States between 1980–1985, Chetty et al. program’s incentives. More generally, incorpo- 2014b study how children’s prospects of mov- rating firm responses and equilibrium effects ing( up in) the income distribution relative to their when predicting the effects of policy changes in parents vary across areas of the United States. behavioral models is an important area for fur- Chetty et al. 2014b divide the United States ther research.22 into 741 commuting( ) zones CZ , geographic units that are analogous to metro( )areas but pro- IV. Welfare Analysis: Neighborhood Choice vide a complete partition of the United States based on commuting patterns, including rural Thus far, we have focused on the positive areas. Figure 8 presents a heat map of a simple implications of behavioral economics, i.e., measure of upward mobility by CZ: the probability that a child born to parents in the bottom quintile of the US income distribution reaches 22 See DellaVigna and Malmendier 2004 ; Gabaix and the top quintile of the US income distribution. ( ) Laibson 2006 ; and Ko˝szegi 2014 for some examples of The map is constructed by dividing commuting research (in this) vein. ( ) zones into deciles based on this probability, with 20 AEA PAPERS AND PROCEEDINGS MAY 2015

16.8% > 12.9–16.8% 11. 3 –12.9% 9.9–11.3% 9.0–9.9% 8.1–9.0% 7. 1 –8.1% 6.1–7.1% 4.8–6.1% 4.8% < Insufficient data

Figure 8. The Geography of Upward Income Mobility in the United States

Notes: This figure reproduces Appendix Figure VIb in Chetty et al. 2014b . It presents a heat map of upward income mobility based on de-identified federal income tax records for all children( in the) 1980–1985 birth cohorts in the United States. The map is divided into 741 commuting zones. In each commuting zone, upward income mobility is measured by the probability that a child reaches the top quintile of the national income distribution for his birth cohort conditional on having parents in the bottom quintile of the national income distribution. Children are assigned( to commuting) zones based on where they grew up i.e., where they were first claimed as a dependent on a tax return , irrespective of where they live as adults. The commuting zones( are divided into deciles based on their rate of upward mobility,) with lighter colored areas representing areas with higher rates of upward mobility.

lighter colored areas representing areas with mobility is nearly 13 percent, almost three times higher levels of upward mobility.23 Children’s larger.24 chances of realizing the “American Dream” Most of the geographic variation in out- vary substantially across areas. In some areas, comes in Figure 8 appears to be driven by such as Atlanta or Indianapolis, less than 5 per- causal effects of place rather than differences cent of children born to parents in the bottom in the type of people living in different places. quintile reach the top quintile. In others, such as Chetty and Hendren 2015 study 8 mil- Salt Lake City and San Jose, the rate of upward lion families who move( across) areas and use quasi-experimental methods—sibling comparisons, exogenous displacement shocks, and a set 23 Children are assigned to commuting zones based on of placebo tests—to show that neighborhoods the location of their parents when the child was claimed as have causal effects on children’s outcomes. a dependent , irrespective of( where they live as adults. The ) In particular, they find that spending more of income quintiles for children are based on their household one’s childhood in an area with higher rates income in 2011–2012, when they are around age 30, while parents’ incomes are based on mean household income between 1996–2000. Children are ranked relative to other children in their birth cohort and parents are ranked relative 24 In a society where parent income has no influence at all to other parents when constructing income quintiles. The on children’s outcomes, we would expect 20 percent of chil- quintiles are defined based on thenational income distribu- dren growing up in families in the bottom quintile to reach tion and hence do not vary across areas. See Chetty et al. the top quintile. The variation in rates of upward mobility 2014b for further details on how income and other vari- across areas is quite substantial given that the largest plausi- ables( are) measured. ble value of the statistic is 20 percent. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 21 of upward mobility i.e., a lighter-colored B. Neoclassical versus Behavioral Models of area in Figure 8 leads( to higher earnings in Neighborhood Choice adulthood. Chetty,) Hendren, and Katz 2015 revisit the Moving to Opportunity (MTO) Neoclassical models of neighborhood choice experiment, which offered families living( in) posit that families choose to live in the area housing projects subsidized housing vouch- that maximizes their utility e.g., Tiebout 1956; ers to move to lower-poverty neighborhoods Epple and Sieg 1999; Bayer,( Ferreira, and via a randomized lottery. They find that mov- McMillan 2007 . Such models offer two expla- ing to a lower poverty neighborhood signifi- nations for why) families do not move to areas cantly improves college attendance rates and where their children do better. First, families’ earnings for children who were young below current neighborhoods may have advantages age 13 when their families moved, consistent( such as lower commuting costs or proximity with the) quasi-experimental results of Chetty to friends that offset the gains from moving. and Hendren 2015 . The treatment effects of Second, parents may have high discount rates moving are substantial:( ) children whose fami- or place low weight on children’s long-term lies take up an experimental voucher to move outcomes. Hence, it is perfectly plausible that to a lower-poverty area when they are less than low-income families rationally choose to stay 13 years old have an annual income that is 31 in high-poverty environments, and that doing so percent higher relative to the control group maximizes their experienced utility. in their mid-twenties. Importantly, the moves Theories from behavioral economics suggest induced by the MTO experiment are across several different explanations for why families short distances, often less than ten miles. The stay in areas that ultimately harm their children. MTO evidence therefore shows that there is I consider four such explanations here. First, substantial variation in neighborhoods’ causal models of present bias e.g., Laibson 1997 effects on children’s long-term outcomes even suggest that parents may not( move because the) at fine geographies e.g., census tracts , not long-term gains for children are realized only 10 just at the broad commuting( zone level shown) or 20 years after the point of the move, but the in Figure 8. costs of moving must be paid up front.25 Such The second fact about neighborhood effects present bias may be a particularly strong deter- that emerges from recent work is that mov- rent to moving because the marginal loss from ing to a lower-poverty neighborhood has little delaying a move at any given time is small, as or no impact on adults’ earnings. In particu- children’s outcomes improve smoothly in pro- lar, the MTO experiment had little effect on portion to their exposure to a better environment the earnings or employment rates of adults Chetty and Hendren 2015 . Since there is no Sanbonmatsu et al. 2011; Chetty, Hendren, discrete( deadline by which )one has to move in and( Katz 2015 . Hence, parents do not incur order to reap the gains from a better neighbor- a personal cost )in terms of lost earnings when hood, even small fixed costs of moving can lead moving to an area where their children do a present-biased agent to procrastinate in mov- better. ing despite the large potential gains from doing Third, many low-income families live near so Carroll et al. 2009 . areas that would offer better outcomes for their Second,( low-income) parents may lack infor- children without significantly higher house mation about neighborhoods’ causal effects on prices or rents than their current neighborhood. children. Consistent with this view, Hastings In particular, Chetty and Hendren 2015 show and Weinstein 2008 present evidence that that the correlation between the causal( effect) of low-income parents( are) less likely to choose a county on children’s outcomes and local rents good schools as measured by students’ test or house prices is less than 0.2 within commut- ( ing zones. Together, these three facts raise a simple 25 Present bias differs from a neoclassical model with question: why don’t parents move to affordable high discount rates because present-biased agents place low weight on the future in their decision utility but not their neighborhoods where their children would do experienced utility, whereas neoclassical agents with high better? The next subsection discusses a set of discount rates place low weight on the future in their expe- models that can answer this question. rienced utility. 22 AEA PAPERS AND PROCEEDINGS MAY 2015 scores than high-income parents when they are alternative explanations by examining new pre- offered) a choice between schools in their area. dictions would be a very useful direction for Hastings and Weinstein 2008 show that pro- future work because the neoclassical and behav- viding simplified information( )about the relative ioral models have quite different implications quality of schools substantially changes the for optimal policy, which I discuss in the next choices made by low-income parents, suggest- subsection. ing that they choose worse schools not because of intrinsic preferences but rather because of a C. Welfare Analysis in Behavioral Models lack of information. Third, models of projection bias suggest that I now turn to the normative implications of individuals may not accurately predict how their the models of neighborhood choice discussed tastes will evolve when they move to a new above, focusing on whether policymakers should neighborhood Loewenstein, O’Donoghue, and seek to influence where low-income families Rabin 2003 . (For instance, individuals might live.26 For example, the US federal government overweight )the lost utility from moving away currently provides subsidized Section 8 hous- from nearby friends, not fully recognizing ing vouchers to 2.2 million low-income( ) fami- that they may make new friends in their new lies at a cost of approximately $20 billion US neighborhoods. Department of Housing and Urban Development( Finally, recent models of scarcity in cogni- 2014 . Are such policies desirable? tive capacity suggest that poverty can amplify The) neoclassical model says that policy individuals’ focus on immediate needs Shah, interventions that alter neighborhood choices Mullainathan, and Shafir 2012 . At a physiolog( - decrease social welfare unless neighborhood ical level, the stress induced by) living in pov- choices have externalities that families do not erty has been shown to elevate cortisol levels, take into account when choosing where to which in turn raises individuals’ discount rates live. Such externalities include the benefits to and amplifies present bias Haushofer and Fehr other citizens from having better outcomes for 2014 . More generally, individuals( have lim- children—such as reduced rates of crime—as ited )bandwidth to make complex decisions, well as fiscal externalities such as the increased and living in extreme poverty may focus atten- tax revenue obtained from children who earn tion on immediate-term needs—such as having more as adults. They could also include inter- enough food to last through the end of the month generational externalities that arise if parents Shapiro 2005 —rather than searching for infor- underinvest in children relative to the weight mation( and making) the longer-term plans needed the social planner places on children’s utilities to find an apartment in a better neighborhood. Lazear 1983 . As we shall see below, the wel- Note that all of these behavioral models are fare( implications) of intergenerational external- consistent with the fact that moving to a different ities are very similar to those that emerge from neighborhood has large causal effects on chil- behavioral models. dren’s long-term outcomes but not adults’ cur- In contrast with the neoclassical model, the rent incomes. A higher level of current income behavioral models described above all imply has an immediate payoff, eliminating discount- that encouraging families to move to areas ing and projection biases. Moreover, individu- where children do better e.g., lower-poverty als are presumably more likely to know about areas will increase their( own private wel- available jobs in nearby areas than the causal fare and) hence is desirable even ignoring any effects of an area on their child’s outcomes several years later. Hence, individuals who could 26 immediately obtain a higher salary by moving to Like much of the existing literature in behavioral welfare economics, this application focuses on a case where a nearby neighborhood would presumably have agents’ decision utilities differ from their experienced util- already done so even in the absence of a housing ities because of behavioral biases. As discussed in Section voucher encouraging them to make such a move. II, behavioral models can generate new welfare implications In summary, the three facts on neighborhood even when agents maximize their experienced utility if they have nonstandard preferences. For example, Rabin 1993 effects discussed in Section IVA are consistent discusses welfare implications in a model where agents( have) with both neoclassical models and a variety tastes for fairness. The implications of such nonstandard of behavioral models. Testing between these preferences for optimal policy deserve further exploration. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 23

externalities. Behavioral models thus call for quality q D p v −1 p lies below their true ( ) = ′ ( ) using either traditional policy tools e.g., subsi- willingness to pay u′ q , as shown in Figure 9. dies or nudges e.g., counseling and( assistance Given a price of p ( , )the individual chooses ) ( 0 in finding a new apartment to influence neigh- q 0 units of neighborhood quality, below the borhood choice to some degree.) utility-maximizing choice of q ∗ where the mar- To formalize and quantify the implications ginal experienced utility of additional neighbor- of the two types of models for optimal pol- hood quality equals the price. The lost surplus icy, consider a special case of the framework from underconsumption—shown by the shaded in Section I in which individuals make two triangle in Figure 9—is analogous to the dead- choices: where to live and how much to spend weight loss that arises from a positive consump- on other consumption goods y. To eliminate tion externality such as an intergenerational the complexities that arise from( )discrete choice externality in the( neoclassical model, where ) of neighborhoods, assume that there is a con- u′ q would be social marginal welfare, i.e., pri- tinuum of neighborhoods that differ in their vate( ) marginal utility v q plus the externality ′( ) impacts on children’s long-term outcomes, benefit of consumptionu ′ q v ′ q . As in the which I refer to as neighborhood “quality” q. case of Pigouvian taxes to( correct) − ( externalities,) For example, one may think of q as measuring the optimal policy to correct the “internality” the local poverty rate or school expenditures in depicted in Figure 9 depends upon the differ- an area. Let p denote the price of one unit of ence u q∗ v q∗ . neighborhood quality and normalize the price Identifying′ ( ) − the′( optimal) policy—e.g., the opti- of y to 1. Letting Z denote the consumer’s mal size of housing voucher subsidies—requires wealth, we can write y Z pq. Assume for an assessment of how individuals’ experi- simplicity that utility is linear= − in y. The individ- enced utilities u q differ from their decisions D ′( ) ual’s experienced utility as a function of neigh- q p v′ q . This issue lies at the heart of borhood quality is the( common) = ( concern) that behavioral economics can lead to paternalism, as policymakers’ u q Z pq perceptions of individuals’ experienced utility ( ) + − u′ q could be given priority over individuals’ and his decision utility is own( ) choices q D p .28 Why do policymakers necessarily have a( better) sense of where families v q Z pq, should live than they themselves do? ( ) + − The pragmatic approach to addressing these where u q and v q are smooth, concave func- concerns about paternalism is to measure u′ q tions. The( )agent chooses( ) q to maximize his deci- empirically without leaving it as a free param( )- 27 sion utility, setting v′ q p. eter at the discretion of policymakers. Recent This simple framework( ) = nests the neoclassical research has developed three nonpaternalistic and behavioral models described above. In the methods of identifying experienced utility in neoclassical model, u q v q . In all of the behavioral models that resemble methods used behavioral models described( ) = (above,) individu- to identify the magnitude of externalities in als underestimate the benefits of neighborhood neoclassical models. Each of these approaches quality relative to their true willingness to pay has certain advantages and drawbacks, which I when deciding where to live: v q u q . As a describe in turn. result, their observed demand for( ) neighborhood< ( ) Method 1: Subjective Well-Being.—The first approach is to measure experienced utility 27 I do not restrict u′ q 0 or v′ q 0. Living in an directly using data on self-reported happiness area that is better for children( ) > might have( ) costs> such as lower amenities for parents that drives the marginal utility of moving to an area that is better for children below 0 beyond some level of q . This case may be empirically relevant because as 28 This problem does not arise in the neoclassical model, noted above, in some areas, living in a neighborhood that where v q u q , because q Dp coincides with the produces better outcomes for children does not appear to schedule (of) marginal= ( ) utilities by assumption( ) and hence will- have a significant monetary cost, i.e.,p 0. The only way ingness to pay can be recovered directly from the observed to explain why demand for q remains finite= whenp 0 is demand curve. This revealed preference approach no longer if v q 0 for some q. = works when decision utility differs from experienced utility. ′( ) < 24 AEA PAPERS AND PROCEEDINGS MAY 2015

Marginal experienced utility u q true willingness to pay ′( ) =

Lost surplus from failure to optimize Analogous to deadweight loss from externality ) p Observed demand qD p ( ) marginal decision utility v q = ′( ) Price ( p 0 ( )

q 0 q ( ) * Quantity q ( )

Figure 9. Welfare Analysis in Behavioral Models

Notes: This diagram illustrates welfare analysis in behavior models and its connection to welfare analysis of externalities. The lower curve shows the demand function q D p that one would observe in the data as a function of price, which coincides with the agents( ) marginal decision utility v q . The upper curve shows the agent’s marginal experienced utility u , which is the agent’s′( ) true willingness to pay for the good. In neoclassical models, these ′two(q) curves are identical. In behavioral models, the two curves may differ, and the lost surplus from failing to choose the level of consumption that maximizes experienced utility is depicted by the shaded triangle. This triangle is analogous to the dead-weight loss that arises from positive consumption externalities, where social marginal utility the equivalent of u q here is higher than private marginal utility v q . The challenge for welfare( analysis in behavioral′( ) ) models, as in models with externalities,( ′ (is ))identifying u q empirically. ′( )

Diener 2000; Kahneman and Sugden 2005 . with the presence of behavioral biases in neigh- This( approach—which is analogous to the) borhood choice. use of contingent valuation methods to assess The subjective well-being approach suffers externalities Diamond and Hausman 1994 — from shortcomings analogous to those faced is attractive in( its simplicity and versatility,) as in the contingent valuation literature on individuals can be surveyed about their hypo- externalities, which are discussed at length thetical happiness in many settings. Indeed, in Diamond and Hausman 1994 . Self- in the context of the Moving to Opportunity reported measures of happiness( )can be experiment, adults who received an experi- systematically distorted by transient contextual mental voucher to move to a lower-poverty factors, are affected by selective memory area report significantly higher subjective and projection bias, and do not have a clear well-being after moving Ludwig et al. 2012 . cardinal interpretation. These problems are This finding suggests that( experienced utility) not necessarily insurmountable. For example, increased after individuals moved, consistent researchers have made progress on recall bias VOL. 105 NO. 5 RICHARD T. ELY LECTURE 25 by eliciting measures of well-being in real prices of goods—showing the price of the good time Stone, Shiffman, and DeVries 1999 and inclusive of sales tax—at a large grocery store by having( individuals reconstruct their )daily and estimate the impact of this intervention on activities and recall their feelings during each demand. Under the assumption that individu- episode Kahneman et al. 2004 . More recently, als maximize experienced utility when prices Bernheim( et al. 2013 propose) a method include taxes, Chetty, Looney, and Kroft 2009 of combining choice( data) with subjective recover experienced marginal utilities( from) preferences to form predictions about observed demand when prices include taxes. preferences that remove systematic biases. They use these estimates to calculate the dead- Further work is needed to determine whether weight cost of commodity taxes in a representa- and how subjective well-being metrics can be tive-agent model. used to reliably measure experienced utility, Alcott and Taubinsky 2015 use a similar but they appear to offer at least some qualitative approach to recover individuals’( ) true will- information on ex post preferences than can ingness to pay for energy-efficient compact help mitigate concerns about paternalism in fluorescent CFL light-bulbs in a model that behavioral welfare economics. permits heterogeneity( ) across agents in behavioral biases and preferences. They give con- Method 2: Sufficient Statistics.—The second sumers information about the true costs and method of identifying u′ q is to return to choice benefits of CFL bulbs relative to standard data and use revealed preference( ) in an environ- incandescent bulbs. They then estimate each ment z where agents are known to maximize individual’s demand curve for CFL bulbs both experienced utility, i.e., an environment z such with and without this information treatment by that v q z u q . Intuitively, if we can find varying the price of CFL bulbs experimentally. a setting( | where) = we( ) can “trust” agents’ choices Under the assumption that their information as reflecting their true experienced utilities, then treatment eliminates all behavioral biases, we can back out u q simply from the observed the demand curve post-information coincides D ′( ) 1 demand curve q p z v − p z . This strategy with marginal experienced utilities, u′ q in closely parallels “sufficient( | ) = ′ statistic”( | ) approaches Figure 9. They use these estimates to (derive) to optimal policy in public economics, which seek the optimal subsidy for compact fluorescent to identify optimal policy based on reduced-form light-bulbs and the welfare gain from correct- elasticities rather than deep structural parame- ing the internality. ters Chetty 2009 .29 This approach is easiest to Like the subjective well-being methodology, understand( in the )context of some examples. the sufficient statistic approach does not Chetty, Looney, and Kroft 2009 implement require specifying the exact behavioral model this approach in the context (of sales) taxes and that describes agents’ choices. This is attractive commodity purchases in a grocery store. Their because there are many behavioral models which analysis is motivated by the observation that could generate differences between decision individuals might not account for sales taxes utility and experienced utility, as illustrated which are not included in posted prices in the by the neighborhood choice application. A United( States when they make consumption common criticism of behavioral economics is decisions. To )recover the true willingness to that it does not offer a single unified framework pay for these goods, they post tax-inclusive as an alternative to the neoclassical model. The sufficient statistic approach provides a method of handling this problem in the context of 29 This approach is sometimes called a “choice-based” normative analysis: if one can find a domain approach to welfare analysis Bernheim and Rangel where agents optimize, one can make robust 2009 or a “reduced-form” approach( to welfare analysis )Mullainathan, Schwartzstein, and Congdon 2012 . statements about optimal policy that are valid Bernheim( and Rangel 2009 discuss a more general version) irrespective of the underlying behavioral model. of this approach in which( one) has choice data from settings For example, in the context of neighborhood with various ancillary conditions d, following the notation effects, if the predominant source of bias is that used in Section II. They show that( one) can derive bounds on experienced utility from observed choices even if one does individuals are uninformed about the benefits not observe a setting where decision utility perfectly coin- of living in better areas for their children, one cides with experienced utility. could identify experienced utility by estimating 26 AEA PAPERS AND PROCEEDINGS MAY 2015 demand after providing complete information future utility relative to current utility because about the consequences of growing up in of present-bias. Laibson 1997 assumes that ( ) different neighborhoods.30 individuals discount future payoffs exponen- The drawback of the sufficient statistic tially in their experienced utility i.e., 1. ( β = ) approach is that one may not always be able If this is the true model of behavior, one can to find an environmentz where behavioral identify experienced utility in two steps. First, biases do in fact vanish. For example, Bordalo, one estimates the structural parameters and (β Gennaioli, and Shleifer 2015 propose a model using variation in the stream of payoffs over δ) of salience effects in which( a) surprise display time. Then one simply sets 1 to identify β = of tax-inclusive prices as in Chetty, Looney, and experienced utility and derive welfare impli- Kroft’s 2009 application causes consumers to cations. For example, Angeletos et al. 2001 ( ) overreact( to taxes,) thereby leading to mis-es- calibrate Laibson’s model using con- β − δ timation of experienced utility. Similarly, in sumption data and show that individuals exhibit Alcott and Taubinsky’s 2015 application, one substantial present-bias when making savings may be concerned that consumers( ) did not fully decisions, calling for policies to increase sav- understand and pay attention to the information ings rates such as those discussed in Section II. they were provided on different light-bulbs and Similarly, Paserman 2008 estimates a model ( ) hence were not fully “debiased” even after the of job search with hyperbolic discounting and information treatment. More generally, if there uses it to predict the effects of unemployment are many behavioral factors at play—not just benefit policies. inattention but also present bias, cognitive lim- Other behavioral models can be estimated itations, etc.—it may be very difficult to identify using a similar structural approach. DellaVigna settings where all biases are removed. et al. 2014 estimate a model of job search with reference-dependent( ) preferences and use Method 3: Structural Modeling.—The third it to make predictions about the optimal time approach to welfare analysis is to specify and path of unemployment benefits. A similar struc- estimate the structural parameters of a behav- tural approach could be used in the context of ioral model. The logic here is to identify how neighborhood choice by picking one or more demand varies as a function of the degree of of the behavioral biases discussed above—e.g., behavioral bias and then extrapolate to the case present bias or cognitive limitations—and esti- with no bias to infer experienced utility. This mating the parameters of such a model using approach is analogous to estimating the struc- the techniques applied by Bayer, Ferreira, and tural parameters of the production function for McMillan 2007 in neoclassical models. externalities. The strength( of) the structural approach is that Perhaps the most well-known application of it allows us to infer experienced utility from the structural approach in behavioral models is choice data even if we cannot find domains Laibson’s 1997 quasi-hyperbolic discount- where all behavioral biases vanish. The draw- ing formulation( )of present-bias. In Laibson’s back—like structural estimation in neoclassical life-cycle model, individuals making decisions economics—is that it relies on strong model- t 31 at time t 0 maximize u c 0 t u c t , ing assumptions. This is well illustrated by = ( ) + β ∑ δ ( ) Paserman 2008 and work in DellaVigna et where c t denotes consumption in period ( ) t , u c t is the flow utility from consumption in period( ) t , denotes the agent’s discount factor, and δ1 represents the underweighting of 31 In the working paper version of their paper on tax β < salience, Chetty, Looney, and Kroft 2007 take a structural approach to welfare analysis by developing( ) a bounded rationality model of inattention to taxes. The formulas for 30 As this example illustrates, one typically needs to place the efficiency costs of taxation obtained from this approach some structure on the behavioral model to understand what illustrate the trade-offs between sufficient statistic and struc- conditions will produce unbiased choices. However, one tural approaches to welfare analysis. The sufficient-statistic may not need to fully specify and parametrize the positive formulas are more general, but the structural formulas can model. For instance, one does not need to specify exactly be estimated even without data from an environment where why individuals are uninformed about neighborhood effects agents maximize experience utility and yield additional in order to recover experienced utility when they are given predictions about the determinants of the welfare costs of full information. taxation. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 27 al. 2014 on unemployment benefit policy in D. Model Uncertainty and Optimal Policy behavioral( ) models. Paserman 2008 allows for present bias but rules out other( behavioral) fea- In many situations, one may have to make tures that appear to affect job search behavior, policy decisions without full information about such as reference dependence and biased beliefs the underlying positive model. For instance, as Spinnewijn 2015 . DellaVigna et al. 2014 discussed in Section IVB, both neoclassical and permit( present bias) and reference dependence,( ) behavioral models can fit available evidence but do not allow for biased beliefs. Moreover, about neighborhood effects.33 While future they require the reference point to be determined research will hopefully shed light on the degree by past unemployment benefit levels rather than of behavioral biases in this application—i.e., D past consumption or expectations of future con- the difference between q p and u′ q in sumption, which would be an equally plausible Figure 9—what is the optimal( )policy given( ) the specification Ko˝szegi and Rabin 2006 . Ideally available data? one would use( more flexible specifications) in When faced with uncertainty about the true each of these models, but this is often theoret- model, economists are naturally inclined to use ically and empirically intractable.32 the neoclassical model as the default. A more Although much remains to be learned about principled approach is to explicitly account for normative analysis in behavioral models, the model uncertainty when determining the opti- three approaches discussed above demonstrate mal policy. This approach once again follows one can make progress on characterizing experi- naturally from established methodological tra- enced utility and optimal policy in a disciplined, ditions in neoclassical economics, such as the nonpaternalistic manner. Importantly, the wel- literature on robust control Hansen and Sargent fare implications obtained from neoclassical 2007 . Although a full treatment( of optimal pol- models will generally be incorrect if agents’ icy with) model uncertainty is outside the scope experienced utility differs from their decision of this paper, some simple examples show that utility. Hence, the challenges in identifying the neoclassical model should not necessar- experienced utility do not provide a rationale ily be given priority in the presence of model for adhering to the neoclassical model. Further uncertainty. research on welfare analysis in behavioral models is particularly critical because many policy Nudges with Model Uncertainty.—Consider debates are already motivated by presumed a model with two states of the world. Families behavioral biases. For example, policy proposals either optimize as in the neoclassical model such as mandated retirement and unemployment when choosing neighborhoods or are biased savings accounts e.g., Feldstein 1998; Feldstein toward staying in worse areas because of the and Altman 2007( or consumer protection reg- behavioral biases described in Section IVB. ulations that limit) consumers’ choice sets are Suppose optimizers are insensitive to nudges typically justified by behavioral arguments. If such as counseling on the relative benefits of economists do not contribute to these debates different neighborhoods. However, behavioral by analyzing the welfare consequences of these agents are influenced by the way in which dif- policies in behavioral models, such policies may ferent neighborhoods are framed. The intuition be implemented based on paternalistic assump- underlying this assumption is that behavioral tions rather than empirical evidence. biases are positively correlated. If agents optimize perfectly in choosing neighborhoods, then they are also likely to be insensitive to fram- 32 The challenges in identifying experienced utility grow ing. If they suffer from behavioral biases when larger when one allows for heterogeneity across individuals in preferences and the degree of behavioral biases. Each of the approaches described above can accommodate heterogeneity, but the data requirements grow much larger. See 33 Behavioral and neoclassical models can produce differ- Allcott and Taubinsky 2015 and Goldin and Reck 2014 ent normative implications even though they produce similar for methods of identifying( the) distribution of preferences( ) positive predictions within a given domain. Of course, there using a sufficient statistic approach and Carroll et al. 2009 will always be further tests one could run to discriminate for an analysis of optimal default policy using a structural( ) between the two models; if there is no domain in which the model of present bias and procrastination with heteroge- two models generate different behavior, then they would neous agents. effectively imply the same underlying decision model. 28 AEA PAPERS AND PROCEEDINGS MAY 2015 choosing where to live, then they may be influ- optima to begin with but yields a first-order enced by how choices are framed as well. gain for behavioral agents) since they under-in- In this environment, the optimal policy is to vest in neighborhood quality( in the absence of follow the prescriptions of the behavioral model the subsidy by assumption . Hence, once again and nudge agents toward moving to better e.g., the optimal policy is not to) strictly follow the lower-poverty areas for children. Nudging( is prescriptions of the neoclassical model but optimal irrespective) of one’s prior beliefs on rather to put some weight on the possibility of the two models because it is a weakly domi- behavioral biases. However, unlike with nudges, nant policy. If families turn out to be optimizers, it is not optimal to put full weight on the behav- there is no loss from nudging families toward ioral model, because the subsidy imposes a dis- certain neighborhoods because the nudges will tortionary cost on neoclassical agents. be ignored.34 But if families turn out to have An interesting implication of these examples behavioral biases, then policymakers increase is that model uncertainty can provide a new jus- welfare by nudging families to neighborhoods tification for using nudges instead of conven- where their children will do better. tional policy tools such as subsidies. Nudges The result that one should exactly follow the are typically justified based on the principle of prescriptions of the behavioral model rests on libertarian paternalism, as they influence choice the strong assumption that the nudge has liter- without limiting any individual’s options Thaler ally no cost if agents optimize.35 But this simple and Sunstein 2003 . The examples above( sug- example illustrates that the neoclassical model gest an alternative) rationale for nudges: they should not necessarily be given priority when maximize expected welfare in the presence of we are uncertain about the true model. Indeed, model uncertainty. Nudges work when agents there may be good reasons to prioritize behav- make behavioral mistakes, but have no impact ioral models instead. when they do not, thus providing a more robust way to correct internalities than tax incentives.36 Subsidies with Model Uncertainty.—Now Studying the optimal combination of nudges and consider an alternative policy tool in the same price incentives in a more general setting with two-state example above: subsidized housing model uncertainty would be a fruitful direc- vouchers that are financed through a lump- tion for future research in behavioral welfare sum tax. Housing vouchers affect the behavior economics. of both optimizing and behavioral agents. If families optimize, the optimal subsidy is zero V. Conclusion ignoring externalities , as any subsidy distorts families’( choices relative) to their true prefer- The central message of this paper is that the ences. If families underestimate the benefits of decision to include behavioral factors in eco- neighborhood quality, the optimal subsidy is nomic models should be viewed as a pragmatic positive. If the policymaker seeks to maximize rather than philosophical choice. In some appli- expected utility given uncertainty about the true cations, insights from psychology or other social model, the optimal housing voucher subsidy sciences can help us develop better answers by is strictly positive, a result that follows from identifying new policy tools, offering better O’Donoghue and Rabin’s 2006 analysis of predictions of the effects of existing policies, optimal sin taxes. The reason( is that) introducing or producing new welfare implications. In a small subsidy has a second-order cost for opti- other applications, one may be able to safely mizing agents because they are already at their ignore behavioral factors and use neoclassical ( economic models. Decisions about whether to include behavioral factors in a model should be 34 This assumes that the marginal costs of the nudge e.g., treated like other standard modeling decisions, counseling are negligible. ( 35 Moreover,) this simple analysis ignores the equilibrium effects of nudges, which could have important wel- 36 Moreover, as illustrated by the results on retirement fare consequences. For instance, Handel 2013 shows that saving in Section IIA, subsidies may be ignored by agents nudges that improve individuals’ choices in( health) insurance who suffer from behavioral biases such as inattention, markets could reduce social welfare by exacerbating adverse which may make them particularly ill-suited for correcting selection. internalities. VOL. 105 NO. 5 RICHARD T. ELY LECTURE 29 such as whether to assume quasilinear or time economics can offer more accurate and robust separable utility. These assumptions are conve- prescriptions for optimal policy, again building nient analytical simplifications in some cases; in on familiar methods used to recover individuals’ others, relaxing these assumptions is essential preferences dating to work on externalities by to capture key features of the problem. In this Pigou 1920 . Beyond these methodological jus- sense, behavioral economics is better viewed as tifications,( )perhaps the most important argument a part of all economists’ toolkit like other tools for such a pragmatic perspective is that it can in applied theory rather than as( a separate sub- help us answer some of the critical policy ques- field. Dividing )our field into “behavioral” and tions of our time, from childhood to retirement. “neoclassical” economics is akin to distinguishing “time separable” economists from others. Although this article has focused on applica- REFERENCES tions of behavioral economics, this pragmatic perspective also has lessons for researchers Aguiar, Mark, and Erik Hurst. 2005. “Consump- who develop and test new behavioral theories. tion versus Expenditure.” Journal of Political Behavioral economists sometimes focus on the Economy 113 5 : 919–48. goal of rejecting the assumptions of neoclas- Allcott, Hunt. 2011.( ) “Social Norms and Energy sical models by searching for applications to Conservation.” Journal of Public Economics demonstrate behavioral anomalies. At this point, 95 9–10 : 1082–95. it may be more productive to instead ask how Allcott,( Hunt.) 2014. “Site Selection Bias in Pro- behavioral models can contribute to answering gram Evaluation.” https://files.nyu.edu/ha32/ core economic questions. In this vein, it would public/research/Allcott_SiteSelectionBias.pdf. be useful for experts in psychology and eco- Allcott, Hunt, and Dmitry Taubinsky. 2015 “Eval- nomics to distill the list of behavioral anoma- uating Behaviorally-Motivated Policy: Experi- lies into those that are most relevant in common mental Evidence from the Lightbulb Market.” applications. One of the challenges practitioners https://files.nyu.edu/ha32/public/research/ face in incorporating behavioral insights is that AllcottandTaubinsky_Lightbulbs.pdf. there are myriad factors to consider, with little Altonji, Joseph G., and Christina H. Paxson. guidance about which factors are most import- 1992. “Labor Supply, Hours Constraints, and ant. The same problem arises in the neoclassi- Job Mobility.” Journal of Human Resources 27 cal model, but economists have developed a 2 : 256–78. set of conventions that streamline applications. Angeletos,( ) George-Marios, David Laibson, Andrea For instance, economists have learned that Repetto, Jeremy Tobacman, and Stephen Wein- Giffen goods goods whose demand increases berg. 2001. “The Hyperbolic Consumption with price are( a theoretical possibility but are Model: Calibration, Simulation, and Empirical not relevant) in most real-world applications. Evaluation.” Journal of Economic Perspectives Identifying which behavioral anomalies are like 15 3 : 47–68. Giffen goods would simplify the models under Ashraf,( )Nava, Colin F. Camerer, and George Loe- consideration and may ultimately increase the wenstein. 2005. “Adam Smith, Behavioral application of behavioral economics. Economist.” Journal of Economic Perspectives Why should economists adopt the pragmatic 19 3 : 131–45. approach to behavioral economics advocated Atkinson,( ) Anthony B., and Joseph E. Stiglitz. here? One argument is that it follows naturally 1976. “The Design of Tax Structure: Direct from widely accepted methodological traditions versus Indirect Taxation.” Journal of Public in our profession. From a positive perspective, Economics 6 1–2 : 55–75. the applications reviewed in this paper show Bayer, Patrick, (Fernando) Ferreira, and Robert that an updated reading of the “as if” approach McMillan. 2007. “A Unified Framework for to economic modeling advocated by Friedman Measuring Preferences for Schools and Neigh- 1953 —traditionally one of the main argu- borhoods.” Journal of Political Economy 115 ments( ) used to defend neoclassical models— 4 : 588–638. calls for incorporating behavioral economics Bernheim,( ) B. Douglas. 2009. “Behavioral Welfare into the analysis of important economic ques- Economics.” Journal of the European Eco- tions. From a normative perspective, behavioral nomic Association 7 2–3 : 267–319. ( ) 30 AEA PAPERS AND PROCEEDINGS MAY 2015

Bernheim, B. Douglas, and Antonio Ran- Chetty, Raj, John N. Friedman, and Emmanuel gel. 2009. “Beyond Revealed Preference: Saez. 2013. “Using Differences in Knowledge Choice-Theoretic Foundations for Behavioral across Neighborhoods to Uncover the Welfare Economics.” Quarterly Journal of Impacts of the EITC on Earnings.” American Economics 124 1 : 51–104. Economic Review 103 7 : 2683–2721. Bernheim, B. Douglas,( ) Daniel Bjorkegren, Jef- Chetty, Raj, and Nathan( ) Hendren. 2015. frey Naecker, and Antonio Rangel. 2013. “Do “The Effects of Neighborhoods on Chil- Hypothetical Choices and Non-Choice Ratings dren’s Long-Term Outcomes: Quasi-Exper- Reveal Preferences?” National Bureau of Eco- imental Estimates for the United States.” nomic Research Working Paper 19269. Unpublished. Besley, Timothy, and Stephen Coate. 1992. “Work- Chetty, Raj, Nathan Hendren, and Lawrence fare versus Welfare: Incentive Arguments for Katz. 2015. “The Effects of Exposure to Bet- Work Requirements in Poverty-Alleviation ter Neighborhoods on Children: New Evidence Programs.” American Economic Review 82 from the Moving to Opportunity Experiment.” 1 : 249–61. Unpublished. Bettinger,( ) Eric. P., Bridget Terry Long, Philip Chetty, Raj, Nathaniel Hendren, Patrick Kline, Oreopoulos, and Lisa Sanbonmatsu. 2012. and Emmanuel Saez. 2014b. “Where Is the “The Role of Application Assistance and Land of Opportunity? The Geography of Inter- Information in College Decisions: Results generational Mobility in the United States.” from the H&R Block FAFSA Experiment.” Quarterly Journal of Economics 129 4 : Quarterly Journal of Economics 127 3 : 1553–1623. ( ) 1205–42. ( ) Chetty, Raj, Adam Looney, and Kory Kroft. 2007. Bhargava, Saurabh, and Dayanand Manoli. 2014. “Salience and Taxation: Theory and Evidence.” “Why Are Benefits Left on the Table? Assess- National Bureau of Economic Research Work- ing the Role of Information, Complexity, and ing Paper 13330. Stigma on Take-up with and IRS Field Experi- Chetty, Raj, Adam Looney, and Kory Kroft. ment.” Unpublished. 2009. “Salience and Taxation: Theory and Bhargava, Saurabh, George Loewenstein, and Evidence.” American Economic Review 99 Justin Sydnor. 2014. “Health Insurance and 4 : 1145–77. Sub-Optimal Choice.” Unpublished. Chetty,( ) Raj, and Emmanuel Saez. 2013. “Teach- Bordalo, Pedro, Nicola Gennaioli, and Andrei ing the Tax Code: Earnings Responses to an Shleifer. 2015. “Memory, Attention and Experiment with EITC Recipients.” American Choice.” http://scholar.harvard.edu/files/ Economic Journal: Applied Economics 5 1 : shleifer/files/evokedsets_march18_0.pdf. 1–31. ( ) Carroll, Gabriel D., James J. Choi, David Laib- Choi, James J., David Laibson, Brigitte C. son, Brigitte C. Madrian, and Andrew Metrick. Madrian, and Andrew Metrick. 2002. 2009. “Optimal Defaults and Active Deci- “Defined Contribution Pensions: Plan Rules, sions.” Quarterly Journal of Economics 124 Participant Decisions, and the Path of 4 : 1639–74. Least Resistance.” In Tax Policy and the Chetty,( ) Raj. 2009. “Sufficient Statistics for Wel- Economy, vol. 16, edited by James M. fare Analysis: A Bridge between Structural and Poterba, 67–114. Cambridge, MA: MIT Reduced-Form Methods.” Annual Review of Press. Economics 1: 451–88. Congdon, William J., Jeffrey R. Kling, and Send- Chetty, Raj. 2012. “Bounds on Elasticities with hil Mullainathan. 2011. Policy and Choice: Optimization Frictions: A Synthesis of Micro Public Finance through the Lens of Behavioral and Macro Evidence on Labor Supply.” Econo- Economics. Washington, DC: Brookings Insti- metrica 80 3 : 969–1018. tution Press. Chetty, Raj, John( ) N. Friedman, Soren Leth-Pe- Conlisk, John. 1996. “Why Bounded Rational- terson, Torben Heien Nielsen, and Tore Olsen. ity?” Journal of Economic Literature 34 2 : 2014a. “Active vs. Passive Decisions and 669–700. ( ) Crowd-out in Retirement Savings Accounts: Cutler, David M., and Edward L. Glaeser. 1997. Evidence from Denmark.” Quarterly Journal “Are Ghettos Good or Bad?” Quarterly Jour- of Economics 129 3 : 1141–1219. nal of Economics 112 3 : 827–72. ( ) ( ) VOL. 105 NO. 5 RICHARD T. ELY LECTURE 31

DellaVigna, Stefano. 2009. “Psychology and Teacher Incentives through Loss Aversion: A Economics: Evidence from the Field.” Journal Field Experiment.” National Bureau of Eco- of Economic Literature 47 2 : 315–72. nomic Research Working Paper 18237. DellaVigna, Stefano, Attila( )Lindner, Balázs Gabaix, Xavier, and David Laibson. 2006. Reizer, and Johannes F. Schmieder. 2014. “Shrouded Attributes, Consumer Myopia, and “Reference-Dependent Job Search: Evi- Information Suppression in Competitive Mar- dence from Hungary.” http://eml.berkeley. kets.” Quarterly Journal of Economics 121 2 : edu/~sdellavi/wp/HungaryUIRD_ 505–40. ( ) wp_2014_07_20.pdf. Gallagher, Kelly Sims, and Erich Muehlegger. DellaVigna, Stefano, and Ulrike Malmendier. 2011. “Giving Green to Get Green? Incentives 2004. “Contract Design and Self-Control: The- and Consumer Adoption of Hybrid Vehicle ory and Evidence.” Quarterly Journal of Eco- Technology.” Journal of Environmental Eco- nomics 119 2 : 353–402. nomics and Management 61 1 : 1–15. Diamond, Peter( A.,) and Jerry A. Hausman. 1994. Goldin, Jacob, and Daniel H. Reck.( ) 2014. “Prefer- “Contingent Valuation: Is Some Number Bet- ence Identification under Inconsistent Choice: ter than No Number?” Journal of Economic A Reduced-Form Approach.” Unpublished. Perspectives 8 4 : 45–64. Hamermesh, Daniel S. 2013. “Six Decades of Diener, Ed. 2000.( “Subjective) Well-Being: The Top Economics Publishing: Who and How?” Science of Happiness and a Proposal for a Journal of Economic Literature 51 1 : National Index.” American Psychologist 55 162–72. ( ) 1 : 34–43. Handel, Benjamin R. 2013. “Adverse Selection Eissa,( ) Nada, and Hilary W. Hoynes. 2006. “Behav- and Inertia in Health Insurance Markets: When ioral Responses to Taxes: Lessons from the Nudging Hurts.” American Economic Review EITC and Labor Supply.” In Tax Policy and 103 7 : 2643–82. the Economy, vol. 20, edited by James Poterba, Hansen,( Lars) Peter, and Thomas J. Sargent. 2007. 73–110. Cambridge, MA: MIT Press. Robustness. Princeton, NJ: Princeton Univer- Engen, Eric M., William G. Gale, and John Karl sity Press. Scholz. 1996. “The Illusory Effects of Saving Hastings, Justine S., and Jeffrey M. Weinstein. Incentives on Saving.” Journal of Economic 2008. “Information, School Choice, and Aca- Perspectives 10 4 : 113–38. demic Achievement: Evidence from Two Epple, Dennis, and( Holger) Sieg. 1999. “Estimat- Experiments.” Quarterly Journal of Econom- ing Equilibrium Models of Local Jurisdiction.” ics 123 4 : 1373–1414. Journal of Political Economy 107 4 : 645–81. Haushofer,( )Johannes, and Ernst Fehr. 2014. Ericson, Keith M. 2014. “Consumer( Inertia) and “On the Psychology of Poverty.” Science 344 Firm Pricing in the Medicare Part D Prescrip- 6186 : 862–67. tion Drug Insurance Exchange.” American Eco- Hoxby,( )Caroline, and Sarah Turner. 2014. nomic Journal: Economic Policy 6 1 : 38–64. “Expanding College Opportunities for High- Farber, Henry S. 2014. “Why You Can’t( ) Find a Achieving, Low Income Students.” SIEPR Taxi in the Rain and Other Labor Supply Les- Discussion Paper 12-014. sons from Cab Drivers.” National Bureau of Internal Revenue Service. 1996. “Federal Tax Economic Research Working Paper 20604. Compliance Research: Individual Income Tax Feldstein, Martin, ed. 1998. Privatizing Social Gap Estimates for 1985, 1988, and 1992.” Pub- Security. Chicago: University of Chicago lication 1415. Washington, DC. Press. Internal Revenue Service. 2012. “Statistics of Feldstein, Martin, and Daniel Altman. 2007. Income: Individual Income Tax Returns 2012.” “Unemployment Insurance Savings Accounts.” Publication 1304. Washington, DC. In Tax Policy and the Economy, vol. 21, edited Jencks, Christopher, and Susan E. Mayer. 1990. by James M. Poterba, 35–64. Cambridge, MA: “The Social Consequences of Growing Up in MIT Press. a Poor Neighborhood.” In Inner-city Poverty Friedman, Milton. 1953. Essays in Positive Eco- in the United States, 111–86. Washington, DC: nomics. Chicago: University of Chicago Press. National Academies Press. Fryer, Roland G., Steven D. Levitt, John List, and Joint Committee on Taxation. 2012. “Estimates Sally Sadoff. 2012. “Enhancing the Efficacy of of Federal Tax Expenditures for Fiscal Years 32 AEA PAPERS AND PROCEEDINGS MAY 2015

2011–2015.” Washington, DC: US Govern- Income Distribution.” In Tax Policy and the ment Printing Office. Economy, vol. 12, edited by James M. Poterba, Kahneman, Daniel, Alan B. Krueger, David A. 83–120. Cambridge, MA: MIT Press. Schkade, Norbert Schwarz, and Arthur A. List, John A. 2004. “Neoclassical Theory versus Stone. 2004. “A Survey Method for Char- Prospect Theory: Evidence from the Market- acterizing Daily Life Experience: the Day place.” Econometrica 72 2 : 615–25. Reconstruction Method.” Science 306 5702 : Loewenstein, George, Ted O’Donoghue,( ) and Mat- 1776–80. ( ) thew Rabin. 2003. “Projection Bias in Pre- Kahneman, Daniel, and Robert Sugden. 2005. dicting Future Utility.” Quarterly Journal of “Experienced Utility as a Standard of Policy Economics 118 4 : 1209–48. Evaluation.” Environmental & Resource Eco- Ludwig, Jens, Greg( J.) Duncan, Lisa A. Gennetian, nomics 32 1 : 161–81. Lawrence F. Katz, Ronald C. Kessler, Jeffrey R. Kahneman, Daniel,( ) and Amos Tversky. 1979. Kling, and Lisa Sanbonmatsu. 2012. “Neigh- “Prospect Theory: An Analysis of Decision borhood Effects on the Long-Term Well-Being under Risk.” Econometrica 47 2 : 263–92. of Low-Income Adults.” Science 337 6101 : Kahneman, Daniel, Peter P. Wakker,( ) and Rakesh 1505–10. ( ) Sarin. 1997. “Back to Bentham? Explorations Madrian, Brigitte C. 2014. “Applying Insights of Experienced Utility.” Quarterly Journal of from Behavioral Economics to Policy Economics 112 2 : 375–405. Design.” Annual Review of Economics 6 1 : Kanbur, Ravi, Jukka( )Pirttilä, and Matti Tuomala. 663–88. ( ) 2006. “Non-Welfarist Optimal Taxation and Madrian, Brigitte C., and Dennis F. Shea. 2001. Behavioural Public Economics.” Journal of “The Power of Suggestion: Inertia in 401 k Economic Surveys 20 5 : 849–68. Participation and Savings Behavior.” Quar( )- Keller-Allen, Chandra, and( ) Rose Maria Li. 2013. terly Journal of Economics 116 4 : 1149–87. “Psychological Science and Behavioral Eco- Meyer, Bruce D. 2010. “The Effects( of) the Earned nomics in the Service of Public Policy.” Meet- Income Tax Credit and Recent Reforms.” In ing Summary, National Institute of Health, Tax Policy and the Economy, vol. 24, edited by National Institute on Aging, May 22, 2013. Jeffrey R. Brown, 153–80. Chicago: Univer- Washington, DC. sity of Chicago Press. Ko˝szegi, Botond. 2014. “Behavioral Contract Mirrlees, James A. 1971. “An Exploration in Theory.” Journal of Economic Literature 52 the Theory of Optimum Income Taxation.” 4 : 1075–1118. Review of Economic Studies 38 2 : 175– Ko˝szegi,( ) Botond, and Matthew Rabin. 2006. “A 208. ( ) Model of Reference-Dependent Preferences.” Mullainathan, Sendhil, Joshua Schwartzstein, Quarterly Journal of Economics 121 4 : and William J. Congdon. 2012. “A Reduced- 1133–65. ( ) Form Approach to Behavioral Public Finance.” Laibson, David. 1997. “Golden Eggs and Hyper- Annual Review of Economics 4: 511–40. bolic Discounting.” Quarterly Journal of Eco- O’Donoghue, Ted, and Matthew Rabin. 2006. nomics 112 2 : 443–78. “Optimal Sin Taxes.” Journal of Public Eco- Lazear, Edward.( )1983. “Intergenerational Exter- nomics 90 10–11 : 1825–49. nalities.” Canadian Journal of Economics 16 Paserman, M.( Daniele.) 2008. “Job Search and 2 : 212–28. Hyperbolic Discounting: Structural Estimation Leigh,( ) Andrew. 2010. “Who Benefits from the and Policy Evaluation.” Economic Journal 118 Earned Income Tax Credit? Incidence among 531 : 1418–52. Recipients, Coworkers, and Firms.” B. E. Jour- Pigou,( Arthur) C. 1920. Economics of Welfare. nal of Economic Analysis & Policy 10 1 . London: Macmillan. Levitt, Steven D., and John A. List. 2007.( “What) Poterba, James M. 2014. “Retirement Security Do Laboratory Experiments Measuring Social in an Aging Population.” American Economic Preferences Reveal about the Real World?” Review 104 5 : 1–30. Journal of Economic Perspectives 21 2 : 153– Poterba, James( M.,) Steven F. Venti, and David 74. ( ) A. Wise. 1996. “How Retirement Saving Pro- Liebman, Jeffrey. 1998. “The Impact of the grams Increase Saving.” Journal of Economic Earned Income Tax Credit on Incentives and Perspectives 10 4 : 91–112. ( ) VOL. 105 NO. 5 RICHARD T. ELY LECTURE 33

Rabin, Matthew. 1993. “Incorporating Fairness Shleifer, Andrei, and Robert W. Vishny. 1997. into Game Theory and Economics.” American “The Limits of Arbitrage.” Journal of Finance Economic Review 83 5 : 1281–1302. 52 1 : 35–55. Rabin, Matthew. 1998. (“Psychology) and Eco- Simon,( Herbert) A. 1955. “A Behavioral Model of nomics.” Journal of Economic Literature 36 Rational Choice.” Quarterly Journal of Eco- 1 : 11–46. nomics 69 1 : 99–118. Ramsey,( ) Frank P. 1927. “A Contribution to the Smeeding, Timothy( ) M., Katherin Ross Phillips, Theory of Taxation.” Economic Journal 37 and Michael O’Connor. 2002. “The EITC: 145 : 47–61. Expectation, Knowledge, Use, and Economic Romich,( ) Jennifer L., and Thomas S. Weisner. and Social Mobility.” National Tax Journal 53 2002. “How Families View and Use the Earned 4 : 1187–1210. Income Tax Credit: Advance Payment versus Spinnewijn,( ) Johannes. 2015. “Unemployed but Lump-Sum Delivery.” In Making Work Pay: Optimistic: Optimal Insurance Design with The Earned Income Tax Credit and Its Impact Biased Beliefs.” Journal of the European Eco- on America’s Families, edited by Bruce D. nomic Association 13 1 : 130–67. Meyer and Douglas Holtz-Eakin, 366–92. Stone, Arthur A., Saul S.( Shiffman,) and Marten New York: Russell Sage Foundation. W. DeVries. 1999. “Ecological Momentary Rothstein, Jesse. 2010. “Is the EITC as Good as Assessment.” In Well-Being: The Foun- an NIT? Conditional Cash Transfers and Tax dations of Hedonic Psychology, edited by Incidence.” American Economic Journal: Eco- Daniel Kahneman, Ed Diener, and Norbert nomic Policy 2 1 : 177–208. Schwarz, 26–39. New York: Russell Sage Saez, Emmanuel. 2010.( ) “Do Taxpayers Bunch at Foundation. Kink Points?” American Economic Journal: Thaler, Richard. 1980. “Toward a Positive Theory Economic Policy 2 3 : 180–212. of Consumer Choice.” Journal of Economic Sampson, Robert J., (Jeffrey) D. Morenoff, and Behavior & Organization 1 1 : 39–60. Thomas Gannon-Rowley. 2002. “Assessing Thaler, Richard H., and Shlomo( )Benartzi. 2004. ‘Neighborhood Effects’: Social Processes and “Save More Tomorrow: Using Behavioral New Directions in Research.” Annual Review Economics to Increase Employee Saving.” of Sociology 28 1 : 443–78. Journal of Political Economy 112 1 : 164–87. Sanbonmatsu, Lisa,( )Jens Ludwig, Lawrence F. Thaler, Richard H., and Cass R. Sunstein.( ) 2003. Katz, Lisa A. Gennetian, Greg J. Duncan, “Libertarian Paternalism.” American Eco- Ronald C. Kessler, Emma Adam, Thomas W. nomic Review 93 2 : 175–79. McDade, and Stacy Tessler Lindau. 2011. Mov- Thaler, Richard H., (and) Cass R. Sunstein. 2008. ing to Opportunity for Fair Housing Demon- Nudge: Improving Decisions About Health, stration Program: Final Impacts Evaluation. Wealth, and Happiness. New York: Penguin Washington, DC: US Department of Housing Books. and Urban Development. Tiebout, Charles M. 1956. “A Pure Theory of Sen, Amartya. 1985. Commodities and Capabili- Local Expenditures.” Journal of Political ties. Amsterdam: North-Holland. Economy 64 5 : 416–24. Shah, Anuj K., Sendhil Mullainathan, and Eldar US Department( of) Housing and Urban Develop- Shafir. 2012. “Some Consequences of Having ment. 2014. “Budget of the United States Gov- Too Little.” Science 338 6107 : 682–85. ernment.” Washington, DC: US Government Shapiro, Jesse M. 2005. “Is( There) a Daily Dis- Printing Office. http://www.appropriations. count Rate? Evidence from the Food Stamp senate.gov/sites/default/files/hearings/ Nutrition Cycle.” Journal of Public Economics HUD%20-%20FY%202015%20Budget%20 89 2–3 : 303–25. Testimony%20Senate%20FINAL.pdf. ( ) This article has been cited by:

1. Julia Brüggemann, Kilian Bizer. 2016. Laboratory experiments in innovation research: a methodological overview and a review of the current literature. Journal of Innovation and Entrepreneurship 5:1. . [CrossRef] 2. Richard H. Thaler. 2016. Behavioral Economics: Past, Present, and Future. American Economic Review 106:7, 1577-1600. [Citation] [View PDF article] [PDF with links] 3. Johannes Lohse, Timo Goeschl, Johannes H. Diederich. 2016. Giving is a Question of Time: Response Times and Contributions to an Environmental Public Good. Environmental and Resource Economics . [CrossRef] 4. Balázs Király, András Simonovits. 2016. Megtakarítás és adózás egy önkéntes nyugdíjrendszerben – ágensalapú modellezés. Közgazdasági Szemle 63:5, 473-500. [CrossRef] 5. Raj Chetty, Nathaniel Hendren, Lawrence F. Katz. 2016. The Effects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment. American Economic Review 106:4, 855-902. [Abstract] [View PDF article] [PDF with links] 6. Jason F. Shogren, Linda Thunström. 2016. Do We Need a New Behavioral Benchmark for BCA?. Journal of Benefit-Cost Analysis 7:01, 92-106. [CrossRef] 7. Lisa A. Robinson. 2016. [Ir]rationality, Happiness, and Benefit-Cost Analysis: Introduction to the Special Issue. Journal of Benefit-Cost Analysis 7:01, 1-11. [CrossRef] 8. Matthew D. Adler. 2016. Behavioral Economics, Happiness Surveys, and Public Policy. Journal of Benefit-Cost Analysis 7:01, 196-219. [CrossRef] 9. W. Kip Viscusi, Ted Gayer. 2016. Rational Benefit Assessment for an Irrational World: Toward a Behavioral Transfer Test. Journal of Benefit-Cost Analysis 7:01, 69-91. [CrossRef] 10. Cass R. Sunstein. 2016. Cost-Benefit Analysis, Who’s Your Daddy?. Journal of Benefit-Cost Analysis 7:01, 107-120. [CrossRef] 11. Carol Graham. 2016. Unequal Life Chances and Choices: How Subjective Well-Being Metrics Can Inform Benefit-Cost Analysis. Journal of Benefit-Cost Analysis 7:01, 121-146. [CrossRef] 12. A.M. Lavecchia, H. Liu, P. OreopoulosBehavioral Economics of Education 1-74. [CrossRef] 13. Oren Bar-Gill, Cass R. Sunstein. 2015. Regulation as Delegation. Journal of Legal Analysis 7:1, 1-36. [CrossRef] 14. Ricardo Perez-Truglia. 2015. A Samuelsonian validation test for happiness data. Journal of Economic Psychology 49, 74. [CrossRef]