When saving is gambling

ARTICLE IN PRESS

JID: FINEC

[m3Gdc;April 19, 2018;13:41]

Journal of Financial Economics 0 0 0 (2018) 1–22

Contents lists available at ScienceDirect

Journal of Financial Economics journal homepage: www.elsevier.com/locate/jfec

When saving is gamblingR J. Anthony Cookson Leeds School of Business, University of Colorado at Boulder, Campus Box 419, Boulder, CO 80309, USA

a r t i c l e

i n f o

Article history: Received 20 December 2016 Revised 13 February 2017 Accepted 14 March 2017 Available online xxx JEL classification: G21 D14 L83

a b s t r a c t Prize-linked savings (PLS) accounts, which allocate interest using lottery payments rather than fixed interest, encourage savings by appealing to households’ gambling preferences. I introduce new data on casino cash withdrawals to measure gambling, and examine how individual gambling expenditures respond to the introduction of PLS in Nebraska using a difference-in-differences design. After PLS is introduced, individuals who live in counties that offer PLS reduce gambling by at least 3% more than unaffected individuals. The substitution effect is stronger in low-frills gambling environments, which most resemble PLS, indicating that these accounts fulfill the desire to gamble. © 2018 Elsevier B.V. All rights reserved.

Keywords: Prize linked savings Gambling Risk aversion Financial literacy Credit unions

1. Introduction

R

The author gratefully acknowledges helpful comments from conference and seminar participants at the 2015 Financial Research Association Conference, the 2015 European Finance Association Conference, the 2015 UBC Summer Conference, the 2015 Oregon Summer Conference, the 2015 Helsinki Finance Summit, the 2015 NBER-SI Workshop on Household Finance, the 2015 Finance Down Under Conference, the 2015 FTC Microeconomics Conference, the 2015 Northern Finance Association Conference, the 2015 Israel Behavioral Finance Conference, the 2015 Boulder Summer Conference on Consumer Financial Decision Making, the 2016 Chicago Quantitative Alliance Spring Meeting, the Federal Reserve Bank of Cleveland, Colorado State University, University of Nevada – Las Vegas, University of Manitoba, Montana State University, and the University of Colorado. In addition, the paper owes much to helpful suggestions from an anonymous referee, Asaf Bernstein, Brad Barber, Jamie Brown, Bruce Carlin, Shanna Cookson, Rob Dam, Shaun Davies, Daniel Dorn (discussant), Cary Frydman (discussant), Rawley Heimer, Ben Iverson (discussant), Petri Jylha (discussant), Chi Liao, Carsten Murowski (discussant), Sridhar Narayanan (discussant), John Lynch, Genevieve Melford (discussant), Marina Niessner (discussant), Charlie Sprenger (discussant), Sheridan Titman (discussant), Carly Urban, Ed Van Wesep, Toni Whited (the editor), and Fernando Zapatero. All remaining errors are my own. E-mail address: [email protected]

“Why do people play the lottery or why do people gamble, period? You know, it is with the hope of winning something more. There is a sense that this (prize-linked savings) actually makes savings fun.” – Derek Kilmer, Washington State Senator, PBS Newshour, 23 November 2013 Some households prefer financial products with high variance and skew, despite earning low return (Dorn and Sengmueller, 2009; Boyer and Vorkink, 2014). Gambling in financial markets is typically viewed as problematic, but catering to a gambling preference could also encourage households to engage more with financial markets. From this standpoint, financial products that cater to gambling preferences could be beneficial, especially given the low level of engagement with household savings products (Lusardi et al., 2011). In this spirit, policymakers have proposed Prize-Linked Savings (PLS) accounts, which reward savings by offering randomly drawn lottery payoffs in lieu of fixed interest payments. PLS accounts are a potentially

https://doi.org/10.1016/j.jfineco.2018.04.001 0304-405X/© 2018 Elsevier B.V. All rights reserved.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 2

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

appealing way to encourage savings if people with low savings rates also like to gamble. Despite the intuitive appeal, we still know little about the effects of PLS because these programs have only recently gained academic attention (e.g., Filiz-Ozbay et al., 2015; Cole et al., 2017). In this context, I empirically examine how household casino gambling responds to the introduction of PLS, and I find that PLS substitutes strongly for gambling. Showing that PLS substitutes for gambling in other markets is of general interest for at least three reasons. First, this finding shows that households prefer to take on some kinds of risk in a financial context, which contrasts with the traditional view that individuals require compensation for exposures to risk (e.g., Savor and Wilson, 2014). Indeed, exposure to these risks appears to incentivize saving, rather than discourage it. This aspect of my findings is similar to recent work using the callable options market to study catering to behavioral preferences (Li et al., 2018). Second, in my analysis, households substitute between gambles in different domains (i.e., savings lotteries versus traditional gambling), which indicates a general preference for gambling that is not naturally explained by narrow framing (e.g., Barberis et al., 2006). In this way, my findings provide novel evidence on individual preferences for skewness (e.g., Boyer and Vorkink, 2014; Viva et al., 2017). Third, as policies introduce lottery-like elements to financial products, it is important to understand their effects, not just for personal savings rates, but for other household behaviors as well. In finding significant effects on gambling in other markets, my work suggests that these broader effects are important. The savings lottery accounts I study—Save-to-Win (STW) accounts—became available to members of select credit unions in Nebraska in January 2012. My empirical analysis uses proprietary transaction-level data on casino cash withdrawals to measure how households in affected counties change their casino gambling activity differently from households in nearby counties unaffected by the new accounts. Using this difference-in-difference approach, I find robust evidence that the introduction of savings lotteries reduces the amount of casino gambling. That is, households’ newfound opportunity to gamble while saving in STW accounts is a strong substitute for gambling at commercial casinos. My estimates indicate that the introduction of prizelinked savings led to economically substantial reductions in gambling activity. Relating cash withdrawals to overall casino demand, this estimated effect of STW on cash withdrawals translates into a 3.7–10.2% reduction in the amount of gambling for the average affected county. In aggregate, I estimate that the introduction of prizelinked savings reduced gambling by between $175,0 0 0 and $396,0 0 0, which is a substantial fraction of the approximately $2 million in additional savings at participating credit unions. This effect arises primarily because of less visitation as individuals who were exposed to savings lotteries were 15% less likely to visit the casino in the postperiod. These results are robust to a battery of alternative specifications, different subsamples, and accounting for pre-trends in the amount of casino gambling. Moreover, I find similar substitution effects away from scratch ticket lottery sales, which shows that my findings are not an ar-

tifact of the casino cash data, and indicates a broad change in gambling behavior in the wake of the introduction of prize-linked savings. If the substitution effect works through gambling preferences, savings lotteries and casino gambling ought to be weaker substitutes when they are more differentiated along other dimensions. Consistent with gambling preferences, I present three heterogeneity tests that show substitution among similar gambles, and minimal substitution when savings lotteries and casino gambling are more differentiated. First, savings lotteries are a strong substitute for local gambling, but not for destination gambling when the trip is part of the enjoyment. Second, as the date of the savings raffle draws nearer, savings lotteries and casino gambling become stronger substitutes, consistent with the gambling payoffs becoming more similar in terms of their immediacy. Third, savings lotteries are a strong substitute for gambling at casinos without nightlife, but not for casinos with nightlife, which are better differentiated from savings lotteries. The evidence on heterogeneity in the substitution effect helps rule out a broad class of alternative interpretations. In particular, the finding that substitution is strongest when savings lotteries and casino gambling are most similar contrasts with an interpretation that STW reduced gambling through an attention-grabbing effect, or through effective advertising that was targeted toward Nebraska consumers in served counties (Becker and Murphy, 1993; Barber and Odean, 2008; Hastings et al., 2017). Attentiongrabbing effects and advertising cannot explain why gambling activity at local casinos is more sensitive to the introduction of savings lotteries, why the substitution effect is stronger among casinos without nightlife, nor why late-inthe-month casino transactions are more sensitive to savings lotteries.1 In a similar spirit, the findings cannot be explained by a blanket commitment to spend less because savings lotteries affect some, but not all types of gambles. The fact that substitution is strongest when casino gambling is most similar to the experience of a savings lottery strongly suggests that the substitution effect reflects gambling preferences. My findings can be viewed as a partial empirical validation of the Barberis (2012) model of casino gambling in which patrons exhibit different degrees of self-control at the casino. In the model, high self-control patrons receive a lottery-like payoff from casino gambling (right-skewed with a few high positive outliers), whereas low self-control gamblers do not receive a right-skewed payoff profile from casino gambling because they cannot commit to stopping while ahead. In this way, the Barberis (2012) model motivates why some gamblers would view lotteries and casino gambling as substitutes. This intuition provides the testable prediction that the substitution effect should be stronger among high self-control patrons than it is for low self-control patrons. Indeed, when I examine how the substitution effect interacts with proxies for patron self1 Moreover, treated and control counties are likely to be located in the same advertising markets, which is the level at which advertising is determined. See Spenkuch and Toniatti (2016) for a detailed discussion of advertising markets.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

control (e.g., low propensity to request unavailable funds on debit transactions, or use a credit card to access cash), I find stronger substitution among patrons with greater self-control. Consistent with the model, low self-control individuals do not substitute at all, and these findings are robust to controlling for differences in economic conditions and education. My work contributes to the literature on household saving programs, which has gained significant recent attention (Beshears et al., 2015; Brown et al., 2017). Prior work has recognized that households exhibit significant inertia in savings decisions, making it difficult for policies to influence savings rates directly (Madrian and Shea, 20 01; Thaler and Benartzi, 20 04; Chetty et al., 2014; Choi, 2015). Within this literature, my work is most closely related to a small number of studies on prize-linked savings accounts (Tufano, 2008; Kearney et al., 2011). For example, Cole et al. (2017) study South African savings lottery accounts (specifically, Million-a-Month-Account; MaMA), showing their efficacy for increasing savings, and FilizOzbay et al. (2015) provide laboratory evidence on lottery savings as a mechanism unto itself to encourage greater saving. By analyzing actual gambling patterns, my work complements these analyses by showing that savings lotteries lead to significant substitution away from casino gambling, which is an important consideration given the broad growth in gambling highlighted by Kearney (2005). My findings are also relevant to the literature on the relation between financial literacy and household financial decisions (Behrman et al., 2012; Lusardi and Mitchell, 2014). In this area, prior work links more financial education to better personal financial outcomes (Cole et al., 2014; Brown et al., 2016), whereas others argue that personal characteristics are important for financial behavior (Fernandes et al., 2014; Kuhnen and Miu, 2017). Viewed through this lens, my analysis suggests that the design of innovative savings programs ought to carefully consider how personal characteristics and financial education interact with the program’s incentives. For example, my evidence on the introduction of Save-to-Win indicates that the STW program was less effective for the sub-population of individuals with low self-control who arguably could benefit the most from saving. On the other hand, this finding also suggests that improvements to personal characteristics like self-control (e.g., through targeted financial education interventions) could complement innovative savings programs in a broader program to address low personal savings rates. My findings also contribute to the literature on the role of gambling and sensation seeking in financial markets. In this growing body of research, researchers have examined how sensation seeking by individual investors leads to excessive trading to the detriment of retail investors’ portfolios (Grinblatt and Keloharju, 2009; Barber et al., 2009). More recently, research has noted how sensation seeking and differences in sophistication have led to incentives to engage in financial gambling (Barberis and Huang, 20 08; Kumar, 20 09; Liao, 2015; Andrikogiannopoulou and Papakonstantinou, 2016; Doran et al., 2012). Within this literature, some authors have studied the interaction between traditional lotteries and trading in stocks with

3

lottery-like features. For example, Dorn et al. (2014) find a negative relation between traditional lottery prize amounts and retail stock market activity for both the U.S. and Germany. In a similar vein, Gao and Lin (2015) find a negative spillover effect in Taiwan. Although these studies show that investors appear to exhibit gambling preferences in the construction of their investment portfolios, they have not quantified the degree to which gambling in financial contexts replaces gambling in other contexts. By providing evidence on the substitution away from casino gambling, my work suggests that financial alternatives to gambling may lead to welfare improvements for individuals who replace commercial gambling, which yields negative expected returns, with financial gambling, which tends to yield positive returns.2 More generally, the study of gambling has been a subject of curiosity in the economics literature since the development of the expected utility hypothesis (Friedman and Savage, 1948; Rosett, 1965). Because this paper links formal gambling and gambling through the use of a financial product (savings lotteries), my findings relate to the general study of gambling behavior, which occurs well beyond gambling markets (Becker et al., 2005). In this broader literature, my findings should be of interest to scholars who study consumer finance outcomes, gambling motives in corporate financial decisions, and gambling motives in occupational choice (Schneider and Spalt, 2017; Peng and Thibodeau, 2017; Zhang, 2017). 2. Setting To provide context for the empirical analysis, this section describes the historical background of prize-linked savings accounts, and the introduction of prize-linked savings to Nebraska via participating credit unions. 2.1. Background on prize-linked savings Though uncommon in the United States, lottery savings accounts and related financial products exist in other countries, and in some cases, have existed for a long time. For example, since 1956, investors in the United Kingdom have been able to invest in premium bonds issued by the National Savings and Investments agency. These bonds pay interest, but instead of the interest being allocated to individual accounts, a lottery determines the set of accountholders who receive the interest payments as well as the payment amounts. In an analysis of U.K. premium bonds, Tufano (2008) notes that investors treat lotterystyle accounts as part savings, part gambling, and he conjectures that the gambling feature of these accounts could be used to increase savings rates. Similar products are offered throughout the world in at least 18 countries according to Kearney et al. (2011). In a product similar to the recently introduced savings lottery products in the United 2 This is an underappreciated perspective in the finance literature, which tends to compare financial gambling to textbook investing. On the other hand, investment losses exceed gambling losses in some contexts. For example, Barber et al. (2009) present a comparison of lottery losses to individual trading losses.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 4

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

States, First National Bank in South Africa implemented a prize-linked savings program as well (studied in Cole et al., 2017). 2.1.1. Prize-linked savings in the United States In the United States, lottery savings accounts first became available in Michigan in 2009 with the launch of the Save-to-Win (STW) program in select Michigan locales. Subsequently, similar STW programs were introduced in Nebraska (2012, described in detail below), North Carolina (2013), and Washington (2013). During this period, savings lotteries were unavailable to consumers in other states, largely because federal regulation against gambling prohibited banks from offering similar products until 2015. Of these introductions of prize-linked savings, my empirical work focuses on the Nebraska STW rollout because the casino gambling data are available before and after Nebraska first offered savings lotteries, which enables a more careful empirical design. Under the STW program, participating credit unions offer one-year certificates of deposit (CDs) in which each $25 deposit qualifies the accountholder for an entry in monthly raffles as well as the raffle for the annual grand prize. The amount of the grand prize varies by implementation, year, and jurisdiction, but it is generally a significant amount of money. For example, Michigan’s grand prize in 2009 was $10 0,0 0 0, while Nebraska’s grand prize in 2012 was $25,0 0 0. STW accounts also earn a nominal rate of interest, but the salient feature of these accounts is the raffle component. The fixed interest component of the accounts is typically below comparable products, which helps to offset the cost of administering the prize pool and marketing of the program. The participating credit unions pay an upfront fee to Doorways to Dreams to administer the drawings, and to provide marketing and branding for the STW product (e.g., direct mailings to members, posters at credit union locations, banner advertising on credit union websites). As Doorways to Dreams offered a turnkey marketing package to participating credit unions, the marketing of the accounts was relatively standardized, and similar to the kinds of statements made in casino advertising (e.g., emphasizing the large prize amounts, calling the entry in the savings raffle “free money”). This advertising almost exclusively focuses on the benefits of winning the raffle, and an interested consumer needs to read the fine print to learn that there is a base fixed rate. Moreover, like other kinds of lotteries, the amounts of the large prizes are salient, but the odds of winning these prizes are not. Given this framing and the difficulty in organizing the required information for an apples-to-apples comparison, it is reasonable that individuals would perceive and value the introduction of these accounts as a new lottery. 2.1.2. Prize-linked savings accounts in Nebraska The Nebraska STW program was launched in January 2012 by the Doorways to Dreams Fund in collaboration with nine participating credit unions.3 As with STW pro3 A tenth credit union joined the program during 2012. For the empirical analysis, I focus on the nine initial participating credit unions. The

grams in other states, members of participating credit unions can open a one-year balance-building certificate of deposit (CD) with a minimum of $25 on deposit, and each $25 contributed to the account per month qualifies the individual for an entry into the raffle (up to ten entries per month). As the accounts are designed to encourage savings, there is a $25 fee for the first early withdrawal (before the 12-month term), and a second withdrawal from the account closes the account and disqualifies the individual from winning prizes from subsequent raffles. This fee structure is similar to other certificates of deposit, which can be liquidated in the event of an emergency, but are intended to be held to term. Beyond retaining the principal and receiving a nominal amount of interest, each entry qualifies the accountholder for an entry into various savings raffles—a monthly credit-union-specific drawing, statewide quarterly drawings, and the annual drawing for the grand prize. In 2012, the monthly drawings were for amounts up to $1500 (with many smaller amounts possible), and the grand prize drawing was $25,0 0 0, paid to one winner. According to the Doorways to Dreams Fund, nearly 1600 Nebraska credit union members opened STW accounts by the end of 2012 (out of around 20 0,0 0 0 eligible members). More than a quarter of accountholders (445) won a prize of some amount through the program, and the total dollar value of prizes distributed was $51,375, which is 2.7% of the $1.9 million saved by STW accountholders over the year. This effective interest rate is generous, but there were notable fixed rate alternatives at the time that were competitive with this rate. For example, according to an archived website of Mutual 1st Federal Credit Union (one of the participating credit unions in the sample), a rewards checking account that offered 3.4% interest APY was available. Although the large prizes receive the most media attention, the average prize amount was $115, conditional on receiving a prize. The average accountholder held $857 by July 2012, $1163 by year-end 2012, and $1641 by year-end 2013, indicating that there was significant and persistent growth of the program in its initial couple of years. 2.1.3. Characteristics of STW accountholders The stated goal of the STW program is to encourage low-to-moderate income individuals to save, and to harness household gambling demand to achieve this end. In line with this motivation, recent surveys of low savings households indicate a strong gambling preference. For example, according to a 2016 survey by the Consumer Federation of America, “more than one-fifth of Americans (21%) – 38% of those with incomes below $25,0 0 0 – think that winning the lottery represents the most practical way for them to accumulate several hundred thousand dollars.” Specific to saving at credit unions, the Doorways to Dreams Fund has conducted several surveys of STW accountholders about their motivations to participate in savings lotteries, and about their experiences with gambling. Information from these surveys suggests that

possibility that the other credit union affects additional (control) counties later in the year likely biases against finding any effect of the program.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

ARTICLE IN PRESS

JID: FINEC

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

5

Table 1 Details of the Nebraska Save-to-Win (STW) program. This table summarizes the rollout of Save-to-Win savings lottery accounts in Nebraska. Panel A provides summary statistics compiled from various surveys reported by the Doorways to Dreams Fund. Source: Doorways to Dreams Fund White Papers. Panel B provides the names of the nine Save-to-Win participating credit unions, the town of their main branch, and the counties the credit union serves. Panel A: Save-to-Win account growth Total amount saved July 2012 January 2013 January 2014

Average account balance

$1.1 million $1.9 million $2.4 million

$857 $1163 $1641

Panel B: Save-to-Win program characteristics Percent of accountholders who reported ... Never baving before ... Less than $50 0 0 in financial assets ... Inability to pay for 3 months expenses ... Playing the lottery (MI 2011) ... Visiting casinos/racetracks (MI 2009) Percent of Nebraska population in participating credit unions

43% 31% 50% 63% 38% 20%

Panel C: Save-to-Win participating credit unions Credit union name

Main branch location

Counties served

Centris FCU Gallup FCU Kearney FCU KEE FCU Liberty First FCU MembersOwn Mutual 1st FCU Omaha Police FCU

Omaha Omaha Kearney Kearney Lincoln Lincoln Omaha Omaha

SAC FCU

Bellevue

Douglas, Hall, Lincoln, Sarpy Douglas Buffalo, Dawson Buffalo, Dawson Lancaster Gage, Lancaster Cass, Douglas, Sarpy, Washington Cass, Dodge, Douglas, Lancaster, Sarpy, Saunders, Washington Cass, Douglas, Sarpy, Saunders, Washington

the program appeals to low savings individuals with high propensity to gamble. Panel B of Table 1 reports summary information on the characteristics of accountholders. For example, 43% of Nebraska accountholders report never having saved before, and a significant fraction report having very little in savings and an inability to pay three months of expenses from savings. Doorways to Dreams did not systematically survey Nebraska STW participants about gambling behavior, and they do not report a standard set of questions across their surveys of accountholders conducted in different states. Thus, I supplement the description of the Nebraska STW program with information from surveys of STW participants in Michigan to give a broad sense of the propensity to gamble among STW accountholders. In general, STW accountholders report an abnormally high propensity to gamble. In a survey conducted in 2011, 63% of Michigan accountholders report having played the lottery in the past six months. Specific to casinos, from a separate survey conducted in 2009, 38% of Michigan STW accountholders reported visiting casinos or racetracks. By comparison, only 29% of households in the broadly administered consumer expenditure (CEX) survey report gambling expenditures of any type during a year (Li, 2012). Piecing this survey evidence together, STW participants are at least twice as likely to gamble as the population at large. Although these comparisons give a strong indication that STW participants are frequent gamblers, the survey information from Doorways to Dreams comes from informal surveys of stigma be-

havior, which suffer from well-known underreporting biases (Meyer et al., 2009; Li, 2012). Thus, the true amount of gambling among STW accountholders is likely higher than reported to Doorways to Dreams. 2.2. Credit union membership and Nebraska Given that the STW program is only available through credit unions, it is important to understand the scope of the program relative to the population of Nebraska. The Nebraska STW program launched with the participation of nine credit unions, which serve ten of Nebraska’s 93 counties. Panel C of Table 1 lists the credit unions that were enlisted by Doorways to Dreams to offer STW to their members, as well as their main branch locations and the counties served by each credit union. Further, Fig. 1 presents a map that highlights the geographic distribution of affected and unaffected counties. By year-end 2012, there were over 20 0,0 0 0 members of participating credit unions, a sizable fraction of the 1.4 million individuals in Nebraska counties covered by my sample (approximately one million gambling-age adults aged 21 and older). Because credit union membership in Nebraska is so common, my findings should speak more directly to the typical resident’s experience than if the study were conducted in another state with lower credit union membership. More than leading to a simple substitution from one account type to another, the introduction of prize-linked savings to Nebraska credit unions significantly increased ag-

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 6

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Fig. 1. Map of treatment and control counties. This map portrays the geographic distribution of treatment and control counties used in the empirical analysis. The treatment counties are counties served by a credit union participating in the Save-to-Win program. Within-Nebraska control counties are Nebraska counties that were not served by a credit union participating in the Save-to-Win program. Adjacent-state control counties are counties within 20 miles of the Nebraska border. For inclusion in the sample, a county must have at least five transactions in more than ten months of the sample period. The excluded counties do not satisfy these sample inclusion criteria.

gregate deposits at the credit unions that offered STW. To quantify the increase in desposits from offering STW, I estimate the following specification using credit union call report data on deposits from the National Credit Union Administration (NCUA):

log (depositsit ) =γi + γt + β1 postt × ST Wi + it ,

(1)

where depositsit equals the amount of deposits at credit union i during date (year-quarter) t, STWi is an indicator variable for whether credit union i offers prize-linked savings accounts after 2012, and postt is an indicator for whether the date is after the January 2012 introduction of prize-linked savings accounts. The coefficient of interest in this specification is the difference-in-differences coefficient β 1 , which quantifies how much participating credit unions (ST Wi = 1) increase deposits relative to non-participating credit unions (ST Wi = 0) in the period after prize-linked savings is made available. Standard errors are clustered by credit union to account for serial correlation in deposits over time (Bertrand et al., 2004). Table 2 presents the estimates from this specification, indicating that deposits increased by 5.7% at STW-participating credit unions, relative to deposits in non-participating credit unions. This increase is significant, amounting to $2.36 million additional deposits in absolute dollar terms. Columns 3 and 4 restrict the sample to relatively large credit unions ( > $5 million in deposits), and find quantitatively similar results (4.4% increase in deposits, corresponding to an increase of approximately $2.4 million in deposits), which indicates that the estimated change in deposits is not driven by small credit unions. This aggregate increase is slightly larger than $1.9 million in deposits in STW accounts reported by Doorways to Dreams, suggesting that individuals deposited additional money into ordinary accounts that were not explicitly part of the STW program. Given the constraints on the number

of qualifying deposits per month, the increase in engagement with ordinary accounts is reasonable. Although this evidence is suggestive and indirect, the aggregate increase in engagement with STW credit unions is consistent with individual-level evidence in Cole et al. (2017), which shows that depositors who participate in prize-linked savings tend to open additional accounts beyond those with the prize-linked savings feature.4 2.3. The casino gambling industry The casino gambling industry has experienced significant growth over the past 25 years. Fig. 2 presents a timeline for important events in the growth of the casino industry. As late as the 1950s, casino gambling in the United States was exclusively confined to Nevada. The casino gambling industry grew in fits and starts throughout the 1970s and 1980s with the start of casino gambling in Atlantic City (1978) and the construction of high-stakes bingo parlors and Indian casinos on a number of prominent American Indian reservations (Seminole (FL) 1979; Cabazon (CA) 1980; Foxwoods Casino (CT) 1986). After the Indian Gaming Regulatory Act of 1989, the casino gambling industry grew dramatically in the 1990s, both from Indian casinos that were explicitly authorized under the Act, but also from independent authorizations of riverboat gaming, special casino districts, and casinos across a large span of states from the Midwest to the Gulf region.

4 Although this evidence suggests that it is unlikely that the new deposits in STW accounts merely displace deposits in other accounts, it does not rule out the possibility that the additional savings occurs because of higher debt (e.g., credit card balances). This kind of substitution between savings and debt is less plausible given that there is virtually no substitution across savings accounts, but it is not possible to speak to this possibility without detailed data on credit card balances.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

7

Fig. 2. Timeline of notable events and growth in the American casino industry. This figure provides a timeline of notable changes to the casino industry during the recent casino era, as discussed in detail by Rose (1991). The statistics on entry and extent of the casino gambling market post-2003 are taken from the Gambling Business Directory (Casino City Press, 2012) as computed in Cookson (2017). Table 2 Credit union deposits after the introduction of prize-linked savings. Using deposit information from the call reports of the National Credit Union Administration (NCUA), this table presents results from estimating the difference-in-difference specification in Eq. (1), where the dependent variable is the natural logarithm of depositsjq , the dollar amount of deposits at the credit union j during quarter q (for years 2010–2012). The variable postq is an indicator that equals one for dates after the introduction of Save-to-Win in January 2012, PLSj is an indicator for whether credit union j began offering prize-linked savings in January 2012. The lottery ticket sales data are restricted only to games that had sales before and after January 2012 (for the years 2011 and 2012). Each specification includes credit union fixed effects, and the specifications in Columns 2 and 4 include date fixed effects (year-quarter level). To mitigate concern that the change is driven by very small credit unions, Columns 3 and 4 focus on the subsample of credit unions with $5 million or more in deposits. Standard errors are clustered by credit union, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. > $5m in deposits

All credit unions

Post × # of participating CUs Credit union FE Year-quarter FE R2 # of credit unions # of quarters N

(1)

(2)

(3)

(4)

0.057∗∗∗ (0.021)

0.057∗∗∗ (0.022)

0.044∗∗ (0.021)

0.044∗∗ (0.022)

x

x x

x

x x

0.9985 72 12 831

0.9987 72 12 831

0.9985 53 12 614

0.9989 53 12 614

From the 20 0 0s, the industry continued to expand. As of August 2012, there were casinos in 41 states, and the traditional centroid of the casino gambling industry (Nevada) only comprised approximately 20% of the industry’s casinos (and a similar percentage of the industry revenues). Using an industry database, Cookson (2016) reports that there were 134 plans to enter the casino industry from March 2003 through August 2012. Although not all of these plans successfully resulted in a new casino, recent casino industry growth has been sizable. Casino industry growth has reached the point that some observers have expressed concern that the industry is reaching a saturation point, and a number of prominent casino closings (rare for the industry over the past 20 years) have been taken as evidence for this (Calvert and Kamp, 2014). Given the geographic dispersion of casinos in the United States, willing patrons no longer need to travel long distances to visit one. Consequently, casino visitation is a part of the expenditure profile of consumers in every state. Nebraska is no exception. Over the course of the 26 months covered in the proprietary cash withdrawal data set (May 2010–June 2012), Nebraska patrons withdrew a total of $19.3 million in cash from American casinos. The pervasiveness of casino gambling in Nebraska sets a quantitatively important backdrop for studying the

effect of prize-linked savings as a substitute for casino gambling. Despite there being considerable growth in the industry at large, comprehensive casino industry data described in Cookson (2016) indicate that there was no casino entry in the greater Nebraska region during the sample period (2010–2012). Cookson (2016) shows that these entries into the casino industry predominantly affect patrons who live local to the casino. The fact that there is no entry in the Nebraska region alleviates the potential concern that the effects are driven by differential exposure to entering casinos in treatment and control counties. 3. Data and measurement This section describes the main data source on cash withdrawals at casinos, and outlines the difference-indifferences empirical strategy that exploits the introduction of prize-linked savings. 3.1. Casino cash withdrawal data The data set on casino gambling activity is a proprietary transaction-level database of cash transactions at casinos

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 8

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

throughout the United States.5 For each transaction in the database between May 2010 and June 2012, the data provide a casino identifier, a patron identifier, a timestamp for the transaction, the type of transaction (withdrawal, inquiry, deposit), the form of transaction (credit, debit, etc.), casino characteristics, and characteristics of the patron such as a gender and home ZIP code. The empirical analysis focuses on cash withdrawals by gambling patrons whose home ZIP code is in Nebraska or in counties that are within 20 miles of a Nebraska county (Collard-Wexler, 2014). Thus, out of the 22-million-record database, this paper utilizes 54,375 cash withdrawal records between May 2010 and June 2012. Beyond measuring overall gambling activity, the data also provide insight into the nature of gambling activity. For example, around 12% of cash withdrawal attempts request more funds than available in the bank account (insufficient funds). In addition, 42% of the transactions in the overall sample are transactions where the patron uses a credit card to obtain cash, which is a particularly highfee method of obtaining cash. Beyond the types of transactions, the precise timestamp can be used to compute the fraction of daytime and weekend transactions, which helps assess the balance of attributes and compositional changes to the sample. Finally, the demographic characteristics of the counties covered by the sample indicate that the regions in Nebraska and adjacent states are quite similar in both overall population and average per capita income, which is useful empirically. 3.2. Representativeness of casino cash data Most research on gambling uses data from surveys in which it is likely that individuals under-report their amount of gambling activity due to social stigma (Meyer et al., 2009). The most precise surveys on gambling activity of households, which attempt to address these wellknown survey biases, are two prominent gambling-specific surveys that were conducted by the National Opinion Research Center (NORC)—the Gambling Impact and Behavior Study (1999), and the 2006 California Problem Gambling Prevalence Survey. Unfortunately, these surveys cover gambling activity from outside of my sample window, and thus, cannot be used for the analysis. The only other largescale data set with gambling expenditure information at the individual or household level is the consumer expenditure survey (CEX). As an analysis of the gambling expenditure data by Li (2012) describes in detail, the CEX likely understates “the prevalence of gambling at both the extensive and intensive margins.” For example, the NORC surveys report that the fraction of individuals who gamble is approximately 60%, whereas the analogous calculation from the CEX reports that 29% of households gamble in a given year, a figure that is half of the NORC surveys. By comparison to these gambling surveys, the casino cash data set does not suffer from biases induced by the 5 DISCLAIMER: All information, data, reports, and other information used herein was provided to the author in a proprietary, confidential, and anonymous manner with respect to the identification of any particular casino or patron.

natural urge to under-report gambling activity. Nevertheless, it is important to understand the degree to which cash accessed at the casino is representative of overall gambling expenditure. According to information shared by the data provider, the cash access data are representative of the dollars wagered at casinos in the United States, and have coverage that is stable across US geographies. The data come from approximately three-quarters of casinos throughout the United States, and correspond closely to revenue fluctuations at the casino-month level. The data provider reports that, for casinos that make their revenue data available for validation, a single regression of monthly casino revenues on aggregated cash withdrawals has an R-squared exceeding 90%. Thus, aggregated withdrawal amounts constitute a reasonable proxy for changes in gambling activity across regions and across time, though cash withdrawals do not correspond one-for-one with casino revenues. At the individual level, the number of patrons who access cash at casinos is 15–20% of overall visitation based on headcounts. Nonetheless, when the data are paired with casino revenue and visitation data, the individuals covered in the cash access data account for more than 50% of casino revenues. Thus, the sample is representative of frequent casino patrons who account for the majority of gambling expenditures at casinos. Indeed, this observation is consistent with the magnitudes of the summary statistics in the casino cash data. The average amount per transaction was slightly larger than $750, and over the 26-month sample, the average patron withdrew approximately $3,500, which implies that the average annual cash withdrawals equal to $1615 for the patrons covered in the casino cash data. As a comparison, Li (2012) reports that frequent gamblers (those who report gambling expenditure on all four quarterly surveys of the CEX) have an average annual gambling expenditure of $633 in 2010 dollars using the Consumer Price Index to account for inflation. Given evidence that the CEX expenditures are under-reported by half relative to other surveys of gambling, the magnitude of cash withdrawals for frequent gamblers is of the correct order of magnitude to be representative of the gambling expenditure of frequent gamblers, absent survey biases. This observation is consistent with the data provider’s internal calculations. An analysis of a sample representative of frequent casino gamblers is an interesting exercise unto itself. After all, these individuals account for more than half of casino gambling revenues. Thus, frequent casino gamblers are an important subpopulation to understand. Beyond frequent casino gamblers, it would be useful if the response of these patrons to the introduction of PLS is similar to the responses of other types of gamblers. In Section 4.2, I present an analysis of scratch ticket sales, which speaks to the degree to which the sample represents the behavior of other types of gamblers. Despite the scratch ticket sample coming from a broader set of gamblers, I find a similar magnitude substitution effect to what I find using the casino cash data. Thus, although frequent casino gamblers differ on some dimensions from other individuals, the conclusions drawn from this sample likely extend beyond the behavior of frequent gamblers.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

ARTICLE IN PRESS

JID: FINEC

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

9

Table 3 Summary of county-level data by treated and non-treated counties. This table reports averages of key variables in the pre-treatment period, based on cash withdrawal data using counties in which a participating Save-to-Win (STW) credit union operates to separate treated counties from non-treated counties in Nebraska, as well as the difference and conditional difference. The comparisons in this table are based entirely on within-state differences between treatment and control counties. The conditional difference is computed as the coefficient on a treatment indicator in a regression that also controls for whether the county has a population greater than 10 0,0 0 0. Standard errors are two-way clustered by county and month– year. ∗∗∗ , ∗∗ , and ∗ denote statistically significant differences at the 1%, 5%, and 10% levels.

# # # ... ...

of Counties of Months of Observations Before After

Treated

Not treated

10 26 207 159 48

44 26 1183 907 276

Pre-treatment characteristics (all control counties) Mean transaction amount ($) 537.40 Mean # of transactions 63.85 % Male 58.44 % Not sufficient funds 14.45 % Use credit card for cash 54.96 % Daytime transactions 34.22 % Weekend transactions 46.41 Per capita personal income ($10 0 0s) 41.54 Population (10 0 0s) 122.25 % with population > 10 0,0 0 0 30.15

3.3. Empirical approach The implementation of Nebraska STW provides quasiexperimental variation in exposure to savings lotteries because credit union membership is based on whether the individual works or lives in a credit union’s counties. Moreover, the counties where savings lotteries were introduced are near state borders, which provides a natural set of counties in adjacent states that can be used as control counties in the empirical analysis. To exploit this variation in exposure of savings lotteries around their 2012 introduction, I adopt a difference-in-difference strategy, which identifies the effect of savings lotteries on gambling by contrasting whether patrons in treated counties (where STW is available) respond differently to the introduction of STW than patrons in nearby control counties. My main sample is an aggregated county-month panel covering 54 Nebraska and adjacent counties across 26 months from May 2010 to June 2012. To ensure the reliability of the aggregated gambling measure, the final sample retains month-county observations with at least five transactions, and counties that have five transactions for at least ten months of the sample period. Table 3 presents a summary of observation counts for both treated and control counties. The monthly frequency of the county-month panel enables an analysis of differential pre-trends in the difference-in-difference specification. This treatment-control structure is likely to produce a conservative estimate of the effects of savings lotteries. This is because if nearby households in adjacent regions are affected by the policy (i.e., because they work in the credit union’s county), their inclusion in the control counties will dampen the difference between treatment and control. In support of this idea, I complement my countymonth tests using patron-level location information, which enables finer measurement of exposure to savings lotteries

453.97 25.27 55.10 11.53 40.18 35.91 48.54 39.91 25.47 4.61

Difference

Conditional difference

80.43 38.58 3.44 2.31∗ 14.78∗∗∗ −1.69 −2.13 1.63 96.78∗∗ 25.54∗

76.75 −7.56 2.99 1.80 14.76∗∗∗ −1.31 −2.48∗∗ 2.47∗ – –

using the distance of the patron’s home ZIP code from branches of participating credit unions. In these specifications, which are presented in the online appendix, the estimated effects using proximity to credit unions are slightly stronger, which matches the intuition that the treatmentcontrol approach should conservatively estimate the effect. My empirical work also employs a sample of observations at the patron-by-{pre, post} level. As this sample is useful to condition the effects of savings lotteries on patron characteristics and to explore heterogeneous effects of savings lotteries, I employ patron-level tests as a complement to the main county-month analysis. Nonetheless, because there are few observations per patron, patron-level tests must be conducted on the patron-{pre, post} level rather than a more granular patron-month panel. As such, the patron-level tests cannot account for trends. In this way, it is natural to conduct the main tests on countymonth aggregates, and supplement with patron-level information. 3.4. Balance of attributes and parallel trends As the geographic pattern in Fig. 1 indicates, counties where prize-linked savings accounts are available are geographically close to counties where prize-linked savings accounts are unavailable, even for control counties in adjacent states. The treatment and control counties are well balanced across casino gambling attributes. Table 3 presents means of pre-treatment gambling and demographic characteristics, and tests of the difference in means. The treatment and control counties are well matched on the amount of gambling, the number of transactions, the fraction of male patrons, the fraction of daytime transactions, and the fraction of weekend transactions. The only gambling characteristics that differ significantly between treatment

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 10

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Fig. 3. Testing for differential trends in gambling by treatment and control counties. This figure presents a graphical depiction of the pre-trend in the dependent variable logged cash withdrawals after residualizing by county and month–year fixed effects, plotted for the years 2011 and 2012 in the sample. The difference in slopes with respect to the Date variable equals −0.189 (with a p-value of 0.514). This effect is economically insignificant as well in that it can only explain an effect size of 0.0158 going from December 2011 to December 2012, whereas the jump is an order of magnitude larger.

and control are the propensity of patrons to use a credit card for cash and to request a transaction amount for which there is not sufficient funds. Upon conditioning the difference-in-means test on whether the county has greater than 10 0,0 0 0 residents, the only gambling characteristic that differs across treatment and control is the propensity to use credit cards to obtain cash. In comparing the demographic characteristics between treatment and control counties, population is significantly greater in treatment counties, suggesting that the STW program was targeted to more urban areas. Indeed, 30.15% of treated counties have greater than 10 0,0 0 0 residents, while only 4.61% of control counties do. In addition, per capita income in treated counties is slightly greater, though only significant when conditioning on the large population dummy. Thus, it is important to control for these characteristics when assessing whether the availability of lottery savings accounts reduced the amount of gambling in treatment counties relative to control counties. Including these controls mitigates the concern that an omitted factor related to population or income is responsible for the observed effects. For this reason, all of the empirical specifications control for population and per capita income, though the results do not depend upon including these controls. Fig. 3 presents a plot to evaluate the parallel trends assumption for the response variable—the logged amount of cash withdrawn. There is a negligible difference between the slopes of treatment versus control leading up to the introduction of savings lotteries at the beginning of 2012. Indeed, the p-value on the test that the slopes are different equals 0.514, and the magnitude of the slope is not significant relative to the jump at the introduction of savings lotteries. On this basis and the fact that treatment and control counties are well balanced, the difference-indifference analysis is informative of the causal effect of savings lotteries on casino gambling. It is useful to note

that, beyond checking the conditions for the difference-indifference analysis, I also present a battery of robustness tests (see Table 6), which further demonstrate the robustness of my main results to differential pre-trends, and potential time-varying confounding factors that influence the desirability of gambling. 4. Main findings This section presents the main findings on substitution away from casino gambling and lottery ticket sales after the introduction of Save-to-Win. 4.1. Substitution of cash withdrawals for savings lotteries The main specification exploits the fact that savings lotteries are available in some counties rather than others using the difference-in-difference specification:

log(1 + cashwdit ) =

γi + γt + β1CUi + β2 postt + β3CU i × postt + ξ Xit + i , (2)

where cashwdit is the total amount of cash withdrawn by patrons living in county i during month–year t,6 γ i and γ t are county and month-year fixed effects, CUi is a count of the number of participating credit unions in county i,7 6 There are a few county-month observations for which patrons had cash withdrawals of zero. As log (0) is undefined, adding one enables estimating an approximate log specification. Because the added one is small relative to the typical amount of withdrawals, it has a small impact on the usual percentage change interpretation. Using the inverse hyperbolic sine transformation, which is also defined for negative values and also admits an approximate log (x) interpretation, the results are nearly identical. 7 I also present specifications where CUi is replaced by a treatment dummy variable that indicates whether a participating credit union serves county i. The main specifications use CUi because it also contains information about the intensity of treatment. The online appendix presents two other measurement schemes to capture intensity of treatment—

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

11

Table 4 The effect of the availability of prize-linked savings on gambling demand. This table presents results from estimating the difference-in-difference specification in Eq. (2), where the dependent variable is the natural logarithm of one plus the total amount of cash withdrawn at casinos by individuals in county i during month–year t. In the specification, posti is an indicator that equals one for dates after the introduction of Save-to-Win in January 2012, CU_treatedi is the number of credit unions that offer STW deposits and accounts, or in some specifications, an indicator variable for whether there is a credit union in county i that offers STW. The variable cashwdit is winsorized at the 99th percentile to reduce sensitivities to extreme observations. The vector of control variables X includes logged population and per capita income measures at the county-year level from the Bureau of Economic Analysis. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels.

Post × # of participating CUs # of participating CUs

(1)

(2)

−0.188∗∗∗ (0.047) −0.132 (0.097)

−0.197∗∗∗ (0.049)

Post × STW accounts available STW accounts available Month–year FE County FE R2 # of counties # of months N

(3)

(4)

−0.567∗∗∗ (0.173) −0.757∗∗∗ (0.270)

−0.594∗∗∗ (0.182)

x

x x

x

x x

0.488 54 26 1390

0.675 54 26 1390

0.511 54 26 1390

0.676 54 26 1390

postt is an indicator that equals one after the Nebraska STW program was rolled out in January 2012, and Xit is a vector of controls for demographic characteristics (logged population and per capita income) from the Bureau of Economic Analysis. In this difference-in-difference specification, the coefficient on the CUi × postt interaction, β 3 , is the coefficient of interest. Standard errors are clustered by county to account for serial correlation in the difference-in-differences analysis (Bertrand et al., 2004).8 As an alternative specification to account for serial correlation, I follow the alternative recommendation of Bertrand et al. (2004) to ignore time series information by collapsing units into pre-treatment and post-treatment observations. Specifically, I present an analysis of data collapsed to the patron × {pre, post} level in Section 5. The magnitudes and the statistical significance are consistent with one another across types of analyses,

indicating that choice of aggregation and technique to account for serial correlation are not consequential for my ultimate conclusions. Table 4 reports the results from estimating Eq. (2) using Ordinary Least Squares with county-clustered standard errors. The difference-in-difference estimate is highly significant across specifications, regardless of whether I employ a treatment dummy or CUi to capture the intensity of treatment. According to Column 2, an additional participating credit union decreases casino demand in county i by 17.88%,9 an effect that is statistically significant at the 1% level. The typical difference between treatment and control counties is stark: The treatment county experiences a reduction of 44.8% in the amount of casino gambling from the introduction of Nebraska STW (Column 4). This effect is significant at the 1% level and robust to the inclusion of county and month–year fixed effects. 4.2. Evidence from scratch tickets

CUi /populationi on the county-month data set, and distance to credit union (nearest branch, headquarters, or ATM) using the patron-{pre, post} sample. The quantitative estimates are similar across measurement schemes. 8 For the number of counties in my analysis (N = 54), the Monte Carlo simulations in Bertrand et al. (2004) show that clustering by county accounts for the serial correlation problem as well as the alternative of collapsing into pre-treatment and post-treatment observations (for N = 50, Type I error rates of 5.3% for collapsing and 6.3% for clustering by county; see their Tables VI and VIII). When the number of clusters is small (N = 10 or smaller), ignoring time series information by collapsing the data into county x {pre, post} observations produces more accurate Type I error rates, but according to Bertrand et al. (2004), the downside of aggregation as a solution to the serial correlation problem is that the solution suffers from lower power than clustering. Consistent with the observation of low power, I find that upon collapsing to the county x {pre, post} level, I obtain a similar magnitude estimate (-0.142), but the statistical significance is somewhat lower (at the 10% level). See Table 6 for details on this and other robustness exercises.

In this section, I complement the analysis of cash withdrawals using an analysis of scratch lottery ticket sales using data from the Nebraska lottery at the ZIP codegame-month level. The analysis of scratch ticket sales helps validate the casino cash data on three dimensions. First, using scratch ticket sales data alleviates concerns about representativeness because the data represent all scratch ticket sales, which speaks to a broader population than the casino cash data do. Second, the lottery results can help confirm that the substitution effect in the previous section is genuine substitution across gambles, rather than a 9 This calculation employs the exact formula provided by Wooldridge (2003) for calculating the percentage effect with a logged dependent variˆ able, e f f ect = eβ − 1.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 12

ARTICLE IN PRESS J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

change in withdrawal behavior. Third, because the lottery data are available at a longer horizon (until January 2013), the analysis can help speak to the persistence of the effect of STW.10 To analyze the responsiveness of scratch ticket sales to the introduction of savings lotteries,11 I use data on lottery sales by ZIP code, month, and type of scratch lottery game from the Nebraska Lottery. Specifically, I estimate the specification:

l og(sal esigt ) =

[m3Gdc;April 19, 2018;13:41]

γi + γgt + β1CU i × postt + i ,

(3)

where salesigt is the dollar-value of lottery ticket sales for ZIP code i, scratch game g, and month t. As in the main specifications on cash withdrawals, CUi is the number of participating credit unions in the county (ZIP code i is mapped to its surrounding county) to capture the intensity of treatment. The coefficient of interest is β 1 , which gives the difference-in-difference effect of how scratch ticket gambling in treated counties (those with credit unions participating in STW) differ from the baseline control counties. To maintain a consistent sample throughout, the sample is restricted to scratch ticket games that were offered both before and after the introduction of STW in Nebraska in January 2012, and the data run from January 2011 to January 2013. Each specification includes game × month − year fixed effects to allow for each scratch lottery game to have an arbitrarily different time trend, as well as county (or ZIP code) fixed effects to capture differences across geographic units. To account for serial correlation within county and over time, standard errors are clustered at the county level. Table 5 presents the results from estimating the difference-in-difference specification in (3). The estimates imply a 1.8–2.5% reduction in scratch lottery ticket sales for each additional participating credit union after the introduction of savings lotteries, and the effect is statistically significant at the 1% level. In addition to showing this significant average effect (pre versus post), Fig. 4 presents a leads and lags plot that also shows the dynamics of the effect over the sample period. The leads and lags plot serves two purposes. First, similar to the casino cash evidence in Fig. 3, it shows an absence of pre-trends for the lottery data. Second, the estimates suggest that the effect of STW is relatively long-lived. Rather than exhibiting a 10 Scratch ticket games are changed over time with turnover from year to year. As such, the most expansive post-period that includes games offered before STW runs until January 2013. As scratch ticket sales are only available at the ZIP-month-game level, the casino cash access data enable a more detailed analysis of the substitution effect because I can construct measures of sophistication and similarity from the transactions data, and examine whether savings lotteries changed the nature of gambling. These tests—presented in Section 6—cannot be run using aggregated lottery data. 11 The reason to focus on scratch games instead of jackpot games is that jackpot games in Nebraska (Mega Millions and Powerball) had notable changes in early 2012. First, Powerball doubled the initial size of its progressive jackpot from $20 million to $40 million in early 2012, and that this led to a corresponding increase in jackpot ticket sales. Second, the Mega Millions lottery had a record $656 million jackpot in March of 2012, also boosting sales. To account for these changes to the jackpot games, specifications in the online appendix control for jackpot lottery ticket sales in the difference-in-difference specifications. The substitution effect becomes slightly stronger when controlling for jackpot lottery sales.

Table 5 The effect of savings lotteries on scratch ticket lottery sales. Using data on game-month-ZIP code lottery ticket sales from the Nebraska Lottery, this table presents results from estimating the differencein-difference specification in Eq. (3), where the dependent variable is the natural logarithm of salesigt , the dollar amount of scratch ticket lottery sales in ZIP code i, for scratch ticket game g during month–year t. In the specification, posti is an indicator that equals one for dates after the introduction of Save-to-Win in January 2012, CU_treatedi is the number of credit unions that offer STW deposits and accounts. The lottery ticket sales data are restricted only to games that had sales before and after January 2012 (for the years 2011 and 2012). Each specification includes game by month–year fixed effects to allow for arbitrary game-specific trends over time. The variable salesigt is winsorized at the 99th percentile to reduce sensitivities to extreme observations. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. (1) Post × # of participating CUs Game × month–year FE County FE ZIP code FE R2 N

(2) ∗∗∗

−0.025 (0.009)

−0.018∗∗∗ (0.006)

x x

x

0.556 2006

0.714 2006

x

very strong effect early in the year that evaporates late in the year (i.e., a novelty effect), the effect is slightly stronger in the second half of the year. The online appendix presents additional evidence that the substitution effect for casino gambling does not appear to be driven by a spike in novelty. Consistent with the evidence from the longer post-period using the lottery data, the estimated effect on casino gambling is of a similar magnitude during the April–June post-period months as it is for the January– March post-period months. In the end, the results cannot rule out an intermediate-term novelty effect (lasting more than a year), and thus, one should exercise caution when drawing longer-term policy implications from the findings (e.g., for long-run savings rates). These findings suggest that the broad conclusions using the cash access data are not primarily reflecting changes in how patrons access cash, but rather appear to reflect a change in gambling behavior. Beyond the evidence in this section, my findings in Section 6.1 that substitution is stronger when savings lotteries are more like gambling further supports the notion that the results reflect substitution. 4.3. Magnitudes and robustness The findings in the main specification are not sensitive to using alternative sample cuts, controlling for differential time trends by urban areas and high unemployment areas, or alternative techniques to estimate the effect of savings lotteries on cash withdrawals (i.e., quantile regression, patron-level analysis, using distance to credit union branches to measure treatment). The robustness tests in this section address two classes of related concerns: (i) uniformity of treatment (e.g., are the findings driven by a particular part of the sample?), and (ii) robustness to alternative specifications, measurement, and omitted factors.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

[m3Gdc;April 19, 2018;13:41] 13

Fig. 4. Persistence of the effect of STW on scratch ticket sales. This figure presents a leads and lags graph to illustrate the difference-in-difference effect of STW on logged scratch ticket sales. Each point in the plot is an estimated β 1τ coefficient from the following specification    log salesigt = γi + γgt + 12 τ =−5 β1τ Ii{lag=τ } × CUi + i jt , which allows for the effect of STW to vary depending on the lead or lag (−5 months to +12 months) relative to the January 2012 introduction (t = 0 ) of STW in Nebraska. In the underlying regression, the period used to benchmark pre-treatment effects (i.e., where all of the lead and lag indicators equal zero) is January 2011–June 2011. The dashed horizontal lines on the plot represent the mean effect pre-treatment and post-treatment.

4.3.1. Uniformity of treatment Because gambling is a potentially abnormal behavior, one might worry that a few extreme and anomalous individuals in the sample could drive the pattern of substitution. In this case, there would be a substitution effect, but it would not be relevant to the typical gambler. I present two complementary analyses to address this point. First, I present evidence using quantile regression to evaluate heterogeneity in the substitution effect across individuals. Second, I conduct a systematic subsampling analysis, dropping treatment counties one, two, or three at a time. Specifically, I estimate the difference-in-difference effect using quantile regression, both for the main countymonth data set as well as the patron-{pre, post} data analyzed in Section 5. Together with other robustness measures, the quantile regression results for the 20th and 80th percentiles are presented in Panel A of Table 6. Regardless of the percentile of the distribution considered, there is a significant and economically meaningful effect of the introduction of STW. For example, the smallest magnitude estimate is found using the 20th percentile of county-month gambling amounts. The estimated effect is 9.7%, which is statistically significant at the 5% level. The online appendix presents the full results using quantile regression. An alternative concern about sensitivity to a few observations is that the results are potentially driven by individuals from a few treatment counties, rather than an anomalous part of the distribution of cash withdrawals. To ad-

dress this concern, I undertake a systematic subsampling analysis, which is an example of the specification curve analysis recently proposed in Simonsohn et al. (2015).12 In this spirit, I construct all subsamples across three subsampling choices: (i) dropping one, two, or three of the ten treatment counties, (ii) using within-state, out-of-state, or all control observations, and (iii) different stances on seasonality (all month–year observations, dropping all July through December observations). For each subsample, I estimate the effect of introduction of STW (60 0 0 specifications in total). The empirical cumulative distribution function of these estimates is the specification curve, which gives a comprehensive view of how sensitive the result is to these subsampling choices. Fig. 5 plots the specification curve, plotted against the 5th, 50th, and 95th percentiles of 500 bootstrapped specification curves. Regardless of the subsample, I obtain a negative point estimate, and 79.7% of these estimates are 12 To construct a specification curve, the researcher identifies a set of defensible specifications over which the researcher would usually choose a subset to present in tabular form. Rather than running a selective sample of these specifications, the researcher runs all of them and reports the distribution of coefficient estimates. The properties of the distribution of tests across specifications give a comprehensive, and ex ante, balanced view of robustness to a set of choices. The goal in my application of the specification curve is to present a comprehensive view of the robustness of the findings to subsampling choices, but the method can be applied broadly to evaluate robustness to a wider set of specification choices. Simonsohn et al. (2015) provide several examples of robust and non-robust published results.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

ARTICLE IN PRESS

JID: FINEC 14

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Table 6 Alternative specifications and robustness checks. Panel A presents results from estimating alternative specifications and sample cuts. Each entry in this table reports a separate robustness check in which month-year and county fixed effects are employed. In the case of patron-level specifications, ZIP code fixed effects are employed. The coefficient reported for each test is the coefficient estimate on the post × # o f part icipat ing CUs term. Standard errors are clustered by county. Full regression detail is provided in the online appendix. Panel B presents summary information on the systematic subsampling analysis (specification curve) in which I estimate the model on all subsamples that (i) exclude one, two, or three of the ten treatment counties, (ii) drop within-state control counties, drop adjacent-state control counties, or include all control counties, (iii) drop data from months July–December, or retain these dates (accounting for seasonality, or not). There are 60 0 0 such subsamples. To evaluate statistical significance, I follow the bootstrapping procedure in Simonsohn et al. (2015) with 500 replications to generate 500 resampled specification curves. ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the1%, 5%, and 10% levels. Panel A: Summary of alternative specifications and robustness exercises Robustness test

Estimated effect

Alternative sample cuts Main specification Within-Nebraska controls only Adjacent-to-Nebraska controls only January–June Obs (seasonality) April–June post period (persistence) Quantile regression specifications 20th pctile on county-month data 80th pctile on county-month data 20th pctile on patron-level data 80th pctile on patron-level data

Robustness test

Estimated effect

Controlling for differential time trends Difference relative to 2011 trend

−0.197∗∗ (0.049) −0.135∗∗ (0.062) −0.185∗∗∗ (0.055) −0.137∗∗∗ (0.049) −0.194∗∗∗ (0.056)

−0.104∗∗ (0.049) −0.133∗∗ (0.058) −0.164∗∗ (0.064) −0.195∗∗ (0.058) −0.176∗∗ (0.050)

Differential trend by > 50K residents Differential trend by > 100K residents Differential trend by > median unemp. Differential trend by > 90th pctile unemp. Other controls and measurement Controlling for Jackpot Lottery Sales

−0.097∗∗ (0.045) −0.273∗∗ (0.048) −0.148∗∗∗ (0.032) −0.187∗∗∗ (0.034)

(−1)∗ Log distance to branch (patron-level) Credit unions per capita (Z) Patron × {pre, post} County × {pre, post} Inverse hyperbolic sine of cashwd as dependent variable

−0.204∗∗ (0.048) −0.148∗∗ (0.051) −0.143∗∗∗ (0.053) −0.156∗∗∗ (0.055) −0.143∗ (0.082) −0.199∗∗∗ (0.047)

Panel B: Robustness to systematic subsampling (specification curve) Percentiles of the specification curve % of Significant results (α = 0.10)

5th

25th

50th

75th

95th

Characteristics of specification curve Estimate 0.797∗∗ p-value (0.012)

−0.225∗ (0.064)

−0.189∗∗∗ (0.0 0 0)

−0.148∗∗∗ (0.004)

−0.120∗∗∗ (0.004)

−0.076∗∗∗ (0.008)

Bootstrap summary statistics Mean Median Std Dev

−0.070 −0.064 0.067

−0.030 −0.028 0.057

0.003 0.004 0.053

0.035 0.034 0.054

0.074 0.072 0.059

0.110 0.028 0.166

statistically significant at the 10% level. Using the bootstrap procedure suggested by Simonsohn et al. (2015), only 1.2% of the bootstrap specification curves yield higher rates of statistical significance, indicating that it is unlikely that the degree of uniformity of the effect would appear by chance. Panel B of Table 6 summarizes the distribution of the effects across the 5th, 25th, 50th, 75th, and 95th percentiles of the specification curve. The magnitude of the estimated effects across these percentiles ranges from approximately −7.6% to −22.5%, and all of these percentiles are significantly significant according to the Simonsohn et al. (2015) bootstrap procedure. Taken together, these results indicate that substitution between gambling and savings lotteries is a pervasive phe-

nomenon throughout the sample. It does not appear that the effects I document reflect abnormal or anomalous behavior of a few individuals or regions. 4.3.2. Additional robustness and quantifying magnitudes I also present a series of tests to evaluate whether the results are sensitive to other specification choices about measurement, aggregation, or how to account for pretrends. Panel A of Table 6 summarizes this set of robustness exercises, presenting estimates for the differencein-difference effect across each of the alternative specifications I have considered. The online appendix presents more detail on each of these tests. Across specifications, I find savings lotteries significantly reduce casino cash with-

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

[m3Gdc;April 19, 2018;13:41] 15

Fig. 5. Specification curve analysis for the effect of prize-linked savings. This figure presents the empirical cumulative distribution function (ECDF) for coefficient estimates across 6,0 0 0 subsamples that (i) drop one, two, or three out of ten treatment counties, (ii) use either all control observations, withinstate control counties only, or adjacent state control counties only, and (iii) use all year–month observations, or drop July–December observations to account for seasonality. The plotted ECDF is compared with 5th, 50th, and 95th percentiles of bootstrapped specification curves from a bootstrapping procedure suggested by Simonsohn et al. (2015) that bootstraps the specification curve itself. The dashed lines represent 90% confidence bands from the bootstrapping procedure.

drawals, with a percentage effect that ranges from 9.2% to 23.9%. For the average control county with 2.4 participating credit unions, the substitution effect ranges from 20.8% to 48.1%.,13 These estimated magnitudes are large as a percentage of cash withdrawals, but are moderate relative to overall gambling. To quantify the effect on overall gambling, note that there is approximately $45 billion annually wagered at casinos in the data set, and approximately $10 billion withdrawn annually in the data.14 Thus, each dollar withdrawn corresponds to approximately $4.50 of overall gambling. If gamblers reduce gambling by first reducing cash withdrawals, and subsequently wagering less outside money, the percentage decline in cash withdrawals is 4.5 times larger than the percentage decline in gambling. Using this link between withdrawals and gambling, the 95th percentile of the specification curve corresponds to a 3.71% decline in casino gambling. This substitution effect is similar to the percentage increase in deposits at Save-to-Win credit unions relative to non-participating credit unions (4.4–5.7%, Table 2). In addition, the substitution effect captured in the casino cash data is not

13 For the estimated β 3 from Panel A of Table 6, this calculation employs the exact formula provided by Wooldridge (2003) for calculating the percentage effect with a logged dependent variable, ˆ e f f ect = eβ − 1. 14 The $45 billion figure is reported as an aggregate summary statistic from the casino cash data provider.

dramatically larger than the substitution effect shown in the lottery results (1.8–2.5%), which reflect substitution among all lottery gamblers. Given that the casino cash data reflect the behavior of frequent gamblers (see discussion in Section 3.2), these somewhat larger substitution magnitudes are plausible. In addition, Table 7 presents several back-of-theenvelope calculations for the substitution effect in terms of dollar values, giving reasonable magnitudes from that standpoint as well. For example, multiplying the percentage reduction by the average withdrawal amount over six months ($794.86), we obtain dollar-for-dollar substitution effects. By this calculation, the conservative estimate gives an estimated $132.64 reduction in cash withdrawals. To compute the aggregate substitution effect, multiply the average dollar-value effects by the number of treated patrons (1323 patrons from treated counties made withdrawals in the post-period). In aggregate, this calculation implies a substitution away from casino gambling of $175,689 on the conservative side, and $396,443 from the main specification. In comparison with the approximately $2 million saved in participating credit unions over 2012 (both from the STW white papers, and from the credit union deposits regression evidence), these are reasonable magnitudes. Beyond these comparisons, there are at least two reasons to expect a large effect relative to the relatively small number of prize-linked savings accounts opened. First, treatment is not necessarily the opening of an account, but

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

ARTICLE IN PRESS

JID: FINEC 16

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Table 7 Relating cash-withdrawal effects to gambling demand. This table presents back-of-the-envelope calculations for several scenarios to relate the documented substitution effect on cash withdrawals to dollar value and percentage substitution effects on overall casino gambling. The “Effect for average patron” case is constructed using the estimated treatment effect .73 × 6). The “Conservative from Table 3, Column 4, and the $794.86 average amount withdrawn every six months by patrons in the sample (794.86 = 3448 26 effect for average patron” case is constructed using the 95th percentile of the specification curve, summarized in Panel B of Table 6, which equals −7.6 log ˆ − 1). Finally, points, scaled by the average number of credit unions in treatment counties (2.402), and then applying the percentage effect formula (exp{β} the effect from the median patron is taken by computing the percentage effect from the median regression (reported in the online appendix), and applying this effect to the amount of cash withdrawn by the median patron over six months ($184.56, computed from the summary statistics). The “% of Casino gambling” is computed assuming that initially $3.5 of cash are brought per one dollar of cash accessed, a figure that matches aggregate casino revenue and withdrawal statistics. All the calculations in this table maintain the assumption that the entire dollar-value substitution effect comes from a decrease in withdrawals (and not a decrease in the amount of cash brought to the casino).

Average patron Average patron (conservative) Median patron

Per credit union effect (%)

% of cash withdrawn

% of gambling

$ Withdrawn (per 6 mo.)

Effect in $

−17.9 −7.3 −22.7

−37.7 −16.7 −46.1

−9.28 −3.71 −10.24

794.86 794.86 184.56

−299.66 −132.64 −85.01

the option to open an account. Even for individuals who plan to use this option later, the introduction of PLS could have led the individual to substitute away from gambling. Second, gambling is a social activity in which peer effects are likely. Encouraging one individual to forsake gambling is likely to lead to a reduction of gambling by that individual’s friends, amplifying the original effect.

4.4. Evidence on the nature of gambling Using the detail on the frequency and timing of transactions from the cash access data, I deepen the analysis by examining how the introduction of savings lotteries affected the nature of gambling. Specifically, I can evaluate whether the decline in gambling was due to fewer transactions to the casino, or withdrawing less money per transaction. In addition, I can examine whether the nature of casino gambling changes along other dimensions—timing during the day (morning, afternoon, evening), timing during the week, types of transactions (credit card, debit card, mistakes), and for a subsample of observations, gender. To evaluate these potential changes to the nature of gambling, I estimate specifications analogous to Eq. (2), separately using different outcomes related to the nature of gambling as response variables. Panel A of Table 8 illustrates how casino gambling behavior changed after the introduction of savings lotteries. The overall reduction in demand comes most robustly from a reduction in the number of transactions. Average transaction size declines by approximately the same magnitude as the number of transactions, but the effect is estimated imprecisely. These findings hint at extensive versus intensive margin effects, which are more readily analyzed in the patron-{pre, post} sample. On the nature of gambling, the frequency of insufficient funds transactions and frequency of credit-card-for-cash transactions do not exhibit statistically significant changes. Although females appear to substitute away from casino gambling more than males (owing to the slight rise in %Male in the post-period), Panel B of Table 8 indicates similar time-of-day and day-of-week patterns of gambling. In sum, the introduction of savings lotteries influences the amount of gambling (somewhat more robustly through visitation), but it does not appear to influence the qualitative nature of gambling.

5. Patron-level evidence The main empirical tests employ county-month aggregated measures for which there is enough data to use the transaction data to reliably represent casino demand for that county. A potential concern is that this sample construction could bias the estimated substitution effect. To address this possibility, I analyze data at the patronby-{pre, post} level using all patrons in the region. This data set has two observations per patron, one observation aggregated over the patron’s pre-treatment transactions, and the other aggregated over post-treatment transactions. When aggregating the data in this way, approximately 30% patrons make transactions both before and after the introduction of savings lotteries in January 2012. Using this patron-level data set, I estimate a difference-indifference specification that is analogous to the differencein-difference specification for the county-month panel:

Ykt =

γz∗ + β1∗ postt + β2∗CU _treatedi × postt + γ ∗ Xi + kt∗ , (4)

where the outcome variable Ykt reflects one of three outcomes: (i) the logged amount withdrawn by patron k in period t ∈ {pre, post}, (ii) an indicator for whether patron k had positive withdrawals in post-treatment period, and (iii) the logged amount of cash access fees paid by patron k in period t ∈ {pre, post}. As in the county-month data specifications, the coefficient of interest is the differencein-difference coefficient, β2∗ , which captures the effect of being treated by greater exposure to participating credit unions that offer savings lotteries. To evaluate the intensive margin of gambling, I focus on the persistent sample of patrons who are observed in both the pre-period and the post-period. To the extent that treatment reduces the propensity of the patron to gamble at all, these estimates understate the true effect. Columns 1 and 2 of Table 9 present the results from estimating Eq. (4) with logged cash withdrawn as the dependent variable. According to these specifications, an additional participating credit union reduces the amount of gambling in the post-period by 14.4%. This is a similar estimate to the estimate from the county-month aggregated sample, and the effects are significant at the 1% level using standard errors clustered by county. To evaluate the extensive margin of gambling, I focus on the sample of patrons who were observed in the

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

ARTICLE IN PRESS

JID: FINEC

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

17

Table 8 The effect of savings lotteries on the extent and nature of gambling. This table presents results from estimating the difference-in-difference specification in Eq. (2), where the dependent variable is either a different measure of gambling activity or a patron characteristic. In Panel A, the dependent variable is a type of observed gambling activity—either the logged number of transactions, the logged average transaction size (in dollars), the fraction of transactions where the patron had insufficient funds for the cash withdrawal (%NSF), or the fraction of transactions where the patron used a credit card for cash in county i and month–year t. In Panel B, the dependent variable is a characteristic of the patron or timing of gambling—either the fraction of weekend transactions, the fraction of daytime transactions, or the fraction of transactions by males in county i and month–year t. As for the right-hand side variables, posti is an indicator that equals one for dates after the introduction of Save-to-Win in January 2012, CU_treatedi is the number of credit unions that offer STW deposits and accounts, or in some specifications, an indicator variable for whether there is a credit union in county i that offers STW. The vector of control variables X includes logged population and per capita income measures at the county-year level from the Bureau of Economic Analysis. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. Panel A: Changes in gambling behavior

Post × # of participating CUs

Log(#t ransact ions )

Log(transsize)

% NSF

% credit card

−0.089∗∗∗ (0.024)

−0.117 (0.072)

0.004 (0.006)

0.011 (0.010)

x x

x x

x x

x x

0.836 54 26 1390

0.352 54 26 1390

0.193 54 26 1390

0.447 54 26 1390

% daytime

% weekend

% male

0.0 0 0 (0.005)

−0.008 (0.005)

0.038∗ (0.023)

x x

x x

x x

0.088 54 26 1390

0.112 54 26 1390

0.339 54 26 1091

Month–year FE County FE R2 # of counties # of months N

Panel B: Changes in sample composition of patrons

Post × # of participating CUs Month–year FE County FE R2 # of counties # of months N

Table 9 The effect of savings lotteries on gambling, propensity to gamble, and cash access fees. Each observation in this table is a patron × (before, after) treatment. This table presents results from estimating the difference-in-difference specification in Eq. (4) where the dependent variable is either one plus the total amount of cash withdrawn at casinos by the patron k during treatment period t (transactions are aggregated before versus after treatment) or an indicator for whether the patron had any withdrawals during the treatment period t, γ z are ZIP code fixed effects, posti equals one post-treatment observations that were aggregated for transactions after the introduction of Save-to-Win in January 2012, CU_treatedi is the number of credit unions that offer STW deposits and accounts, or in some specifications, an indicator variable for whether there is a credit union in county i that offers STW. For each specification, the cashwdk variable is winsorized at the 99th percentile to reduce sensitivity to extreme observations. The estimates for the logged withdrawal amount are constructed using the persistent subsample, which contains only patron × treatment for which I observe pre- and post-observations. The estimates for the indicator for no withdrawals are constructed using a sample of patron transactions for which the patron was observed in the pre-period. The vector of control variables X includes logged population and per capita income measures at the county-year level from the Bureau of Economic Analysis interacted with the post-period. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. Log(cash withdrawn) (1) Post × # of participating CUs

−0.156∗∗∗ (0.055)

2

R # of ZIP Codes N

No withdrawals dummy (3)

(4)

0.036∗∗∗ (0.007) −0.674∗∗∗ (0.223)

Post × STW accounts available ZIP code FE

(2)

Log(fees) (5)

(6)

0.007 (0.016) 0.154∗∗∗ (0.026)

0.004 (0.074)

x

x

x

x

x

x

0.149 482 7262

0.150 482 7262

0.479 654 18730

0.479 654 18730

0.404 482 7262

0.404 482 7262

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 18

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

pre-period and use the no_gamblingkt indicator as the dependent variable. In this context, the difference-indifference estimate reflects the differential propensity to avoid gambling in the post-period, given that the patron was observed to withdraw cash in the pre-period. Columns 3 and 4 of Table 9 present the results from this estimation. These estimates highlight that the introduction of savings lotteries was important to the extensive margin of gambling as well. An additional participating credit union increases the probability of not gambling in the post-period by 3.6 percentage points. In aggregate, being in a county that has at least one participating credit union reduces the propensity to gamble in the post-period significantly. Patrons in affected counties are 15.4 percentage points more likely to not gamble at all in the post-period. It is also informative to analyze the effect of the introduction of savings lotteries on the amount of cash access fees paid because it indirectly indicates the amount of sophistication among those who substitute. The results in Columns 5 and 6 show that there is no change in the fees paid by affected patrons relative to unaffected patrons. This finding suggests that the customers who substitute away from casino gambling are those who pay minimal or no fees to access cash. That is, the substitution effect appears to most strongly affect individuals who exhibit more sophisticated cash-access behavior. This finding foreshadows the analysis of the heterogeneity in how patrons respond to the introduction of savings lotteries. 6. Heterogeneity This section provides greater insight into why consumers substitute from casino gambling by examining heterogeneous effects. The findings are consistent with genuine substitution that is stronger when the gambling products are more similar, giving a strong indication that the substitution effect reflects gambling preferences. In addition, the pattern of substitution across different levels of patron self-control is consistent with a leading behavioral model of casino gambling (Barberis, 2012), further validating the empirical findings by grounding them in theory. 6.1. Substitution across similar gambles If consumers substitute between savings lotteries and casino gambling on account of a gambling preference, the substitution effect should be stronger when savings lotteries are more similar to casino gambling, ex ante. Using information on casino attributes and the location and timing of transactions, this section presents three subsampling tests that show stronger substitution among similar gambles. Notably, the substitution effect is (i) stronger for local gambling than for destination gambling, (ii) stronger for more immediate lottery payoffs than those more distant in time, and (iii) stronger for low-frills casinos than casinos with nightlife. These findings on heterogeneity are consistent with substitution among similar products, and it cannot be jointly explained by a broad attention-grabbing effect (Barber and Odean, 2008), the effects of advertising (Becker and Murphy, 1993), nor a blanket commitment to spend less.

Table 10 The effect of savings lotteries is greater when savings lotteries and casino gambling are similar. This table presents results from estimating the main difference-indifference specification in Eq. (2), but aggregated among similar transactions versus dissimilar transactions for three similar/not cuts of the sample of transactions. For each specification, the cashdit variable is winsorized at the 99th percentile within each subsample. For the construction of subsamples, (i) close is the sample of transactions by patrons within 120 miles of their home ZIP code, while far is the sample of transactions where the patron is more than 120 from his home ZIP code, (ii) short time until lottery is for transactions that occur after the 21st of the month, while long time until lottery are transactions that occur before the 7th of each month, and (iii) casinos with/without nightlife are transactions that occur at casinos that do / do not offer nightlife. The vector of control variables X includes logged population and per capita income measures at the county-year level from the Bureau of Economic Analysis. Full results are reported in the online appendix. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. Sample split Similar Close transactions (within 120 miles) Short time until lottery (week 4 transactions) Casinos without nightlife Differentiated Far transactions (outside of 120 miles) Long time until lottery (week 1 transactions) Casinos with nightlife

Post × # of participating CUs −0.222∗∗ (0.101) −0.239∗∗∗ (0.083) −0.291∗∗∗ (0.098) −0.063 (0.086) −0.157 (0.113) 0.044 (0.098)

6.1.1. Local versus destination gambling According to the substitution-among-similar products hypothesis, savings lotteries should be more substitutable with local casino activity if destination gambling involves other elements of entertainment, such as visiting a new area, whereas gambling at local casinos is more easily substituted using the gambling feature of savings lotteries. Consistent with substitution among similar attributes, Table 10 demonstrates that close transactions (within 120 miles) are much more responsive to the introduction of savings lotteries than far transactions (farther than 120 miles). An additional STW-participating credit union reduces the amount of local casino gambling by approximately 20% for patrons in affected counties, but for transactions far away from the patron’s home ZIP code, the introduction of savings lotteries does not seem to matter. The finding that local gambling is more substitutable with savings lotteries suggests a sensible pattern of substitution across gambling products. In theory, a similar effect could arise mechanically if gambling trips are planned in advance and locked in, rather than being about attribute substitution as I argue. If the effect is mechanical due to locked-in plans, the substitution effect will eventually show up in far regions later in the sample when patrons are not locked into their destination trip. The online appendix addresses this possibility by splitting the post-period into two subsets: the first three months and the last three months. In these specifications, there

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

is no significant substitution effect in any part of the post-period for far transactions, a finding that supports the attribute-substitution interpretation. In fact, the statistically insignificant effects for destination gambling appear to diminish even further in the last three months of the post-period for non-local gambling. 6.1.2. Immediacy of payoff and substitution with casino gambling The timing of monthly drawings also provides useful variation in how substitutable the savings lotteries are with casino gambling. As much of the thrill of casino gambling is the immediate payoff, lotteries with more immediate payoffs should intuitively exhibit greater substitutability with casino gambling. Because Nebraska STW has monthly drawings for which accountholders qualify with a deposit during that month, payoffs from the savings lottery become more immediate closer to the end of the month. According to this logic, the substitution effect should be magnified later in the month relative to earlier in the month. Table 10 presents results from estimating Eq. (2), where the dependent variable only aggregates cash withdrawals over part of the month. Consistent with the close-versusfar results, the time-until-lottery results indicate a greater effect on casino cash withdrawals that occur when savings lotteries are more substitutable with casino gambling. For example, the reduction in week-one gambling due to an additional credit union is 15.7 log points and statistically insignificant, while the reduction in week-four gambling (22nd and after) is 23.9 log points and significant at the 1% level. Moreover, the effect becomes more statistically significant and generally larger in magnitude as the time period shifts from week one to week four. This pattern of results is consistent with the idea that the immediacy of payoff is an important factor behind casino demand, which suggests that there is more to the substitution away from casino gambling than offering an attention-grabbing alternative or merely advertising the STW programs to members of the credit union. 6.1.3. Product differentiation: casinos with and without nightlife Beyond information about transactions, I also utilize information about casino attributes. If the effects are due to attribute substitution, transactions at casinos that are differentiated from pure gambling should be less affected than transactions at relatively unadorned casinos. One dimension that distinguishes casinos from each other is whether the casino has nightlife (i.e., a bar, dance club, etc.). We should expect stronger substitution away from transactions at casinos without nightlife because those casinos are less differentiated from the experience of a savings lottery. Table 10 presents the findings separately for transactions at casinos with nightlife and for transactions at non-nightlife casinos. Consistent with the other attribute-based substitution findings, I find that the substitution effect is particularly strong for transactions that occur at casinos without nightlife (25.2% for each additional participating credit union), whereas casinos with nightlife

[m3Gdc;April 19, 2018;13:41] 19

exhibit no statistically significant change in withdrawals post-introduction. Taken together, these results on the similarity of casino gambling and savings lotteries are difficult to reconcile with rationales unrelated to the gambling feature of the STW product. In particular, if the accounts merely grabbed investor attention, there would be no reason to expect similar gambles to substitute more strongly. Similarly, it would be difficult to explain the stronger substitutability among similar gambles if the accounts were simply attractive for their non-gambling features (e.g., high expected interest relative to other accounts). The findings are also inconsistent with savings lotteries presenting merely a broad commitment to spend less on unplanned expenditures. Speaking to this alternative, my finding of no substitution away from casinos with nightlife is most relevant. Expenditures are just as likely to be unplanned at casinos with nightlife versus without nightlife, but casinos without nightlife are less differentiated from gambling in savings lotteries. Thus, the findings are most naturally explained by consumers substituting across products that are less differentiated in attribute space. 6.2. Substitution and self-control The Barberis (2012) model of casino gambling implies that individuals with high self-control tend to substitute more strongly away from casino gambling under competition from lotteries than low self-control individuals. In the Barberis (2012) model, individuals with high self-control who gamble at casinos can commit to a stopping rule at which to leave the casino, but low self-control individuals cannot. As a result, the casino gambling payoff profile for high self-control individuals looks similar to a lottery payoff (right-skewed with a few large positive outliers). By contrast, low self-control gamblers in the Barberis (2012) model cannot commit to stopping while ahead, and thus do not receive a lottery-like payoff profile from gambling at a casino (eventually losing their maximum stake). Based on the similarity of their payoff profiles, lotteries and casino gambling will appear similar to high selfcontrol individuals, but not to low self-control individuals. In light of the Barberis (2012) model intuition, finding that savings lotteries substitute for casino gambling already provides indirect evidence that gamblers, on average, exhibit some degree of self-control. Yet, beyond an average effect, it would be reassuring from the standpoint of the Barberis (2012) model to find that relatively low selfcontrol gamblers exhibit weaker substitution between savings lotteries and casino gambling. To speak to this point, the transaction-level detail in the cash access data provide direct insight into gambling patrons’ self-control based on the nature of their cash access behavior. For example, a patron likely lacks self-control if he or she uses a credit card to obtain cash at the casino, incurring the high cash advance fees. Similarly, a patron who—at the casino ATM— requests funds in excess of the account balance (insufficient funds) likely has low self-control. As both the frequency of credit card use (to obtain cash at casinos) and insufficient funds transactions are behaviors related to low self-control, I employ the frequency of

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 20

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Table 11 Heterogeneity by self-control and education. The results in this table come from estimating the triple-difference specification modification of Eq. (2), where the difference-in-difference effect is interacted with a characteristic of interest, characteristici . In Panel A, characteristici is one of two indicators of short-run patron self-control: (1) fraction of credit card transactions in the county-month, and (2) fraction of insufficient funds transactions. In Panel B, characteristici is one of two indicators of educational attainment from the US Census: (1) fraction of county residents who attended at least some college in 2010, and (2) fraction of county residents with a high school diploma in 2010. The vector of control variables X includes logged population and per capita income measures at the county-year level from the Bureau of Economic Analysis, as well as interactions between the effect and the employment-topopulation ratio. For ease of interpretation, this table reports separately the estimated difference-in-difference effect for a high level of these characteristics (+1 sd) versus a low level of these characteristics (−1 sd). Full regression results—including estimates for interaction terms—are reported in the online appendix. Standard errors are clustered by county, and ∗∗∗ , ∗∗ , and ∗ indicate statistical significance at the 1%, 5%, and 10% levels. Panel A: Differential effects by proxies for self-control Characteristic High self-control Infrequent use of credit card for cash Infrequently requesting unavailable funds Low self-control Frequent use of credit card for cash Frequently requesting unavailable funds

Post × # of participating CUs −0.329∗∗ (0.076) −0.353∗∗∗ (0.066) −0.017 (0.076) −0.039 (0.066)

Panel B: Differential effects by education Characteristic High education Large % attending college Large % high school graduates Low education Small % attending college Small % high school graduates

Post × # of participating CUs −0.186∗∗ (0.082) −0.170∗ (0.102) −0.210∗∗ (0.082) −0.170∗ (0.102)

these behaviors as proxies for low self-control in my empirical tests. Specifically, I estimate triple-difference specifications that interact the effect of savings lotteries with these self-control proxies to allow the substitution effect to differ by high self-control versus low self-control. Based on these triple-difference specifications, Panel A of Table 11 reports the estimated substitution effect for high selfcontrol versus low self-control individuals. Consistent with the Barberis (2012) model, high self-control individuals respond more to the introduction of savings lotteries than low self-control individuals, and the difference is striking. The substitution effect ranges from -21.9% to -29.7% for high self-control individuals, whereas the effect is an insignificant −1.7% to −3.8% for low self-control individuals. This heterogeneity does not arise because of differing economic conditions because the base specification accounts for population and per capita income, and

the effect is robust to controlling for a triple-difference term allowing for heterogeneity by the employment-topopulation ratio in the county (emp_to_ pop × postt × CUi ). In these richer econometric specifications, indicators of greater sophistication continue to amplify the substitution effect of savings lotteries beyond the terms that control for relatively stronger local economies. Indeed, the findings on sophistication are quite robust in magnitude and significance (see the online appendix for details).15 Moreover, the stronger effect among high self-control individuals is consistent with my finding of a negligible substitution effect on fees in Section 5. The individuals who substitute away from the casino pay little or no cash access fees, indicative of greater self-control. Taken together, these results indicate that patrons who have relatively more self-control–as exhibited by their cash access behavior at the casino—tend to substitute more strongly away from casino gambling. In addition, the heterogeneity across proxies for selfcontrol does not reflect differences in educational attainment, broadly measured. Panel B of Table 11 reports how higher education interacts with the substitution effect between PLS and gambling. These specifications indicate that educational levels are largely unrelated to the substitution effect. Beyond the specifications included in the online appendix, the findings on heterogeneity by self-control are robust even after including triple-difference terms for educational attainment. Viewed through the lens of recent work showing that education promotes financial literacy (Cole et al., 2014; Brown et al., 2016), these findings suggest that self-control might not be influenced by general education or even financial education, which can work through other channels to affect financial behavior (e.g., better numeracy or general financial knowledge). Rather, as a personal characteristic, self-control and its effects appear to be relatively unrelated to educational attainment. Finally, one interpretation of the self-control proxies is that they may broadly proxy for the generic quality of financial behaviors. Intuitively, individuals who exhibit better financial behavior ought to respond to the introduction of a reasonable gamble.16 From the standpoint of parsimony, the self-control view not only provides a rationale for the heterogeneity in the response, but also, it provides clear intuition for why savings lotteries would substitute casino gambling in the first place. The view that the findings are driven by generic quality of financial be-

15 Interpreting the coefficient estimate on the employment-topopulation ratio from the online appendix, I find that regions with better economic conditions exhibit stronger substitution due to the introduction of savings lotteries. This finding complements the main finding in that it highlights that the effects of savings lotteries are less pronounced among financially sensitive individuals (low financial sophistication, worse economic conditions). 16 For example, individuals who do not frequently request unavailable funds may also have greater awareness of financial alternatives, which may also contribute to the heterogeneous effects. Unlike self-control in the Barberis (2012) model, financial awareness has potentially an ambiguous relationship with the substitution effect. On one hand, highawareness patrons may see that savings lotteries and casino gambling are kindred ways to gamble, whereas low-awareness gamblers may not see the connection. On the other hand, low-awareness gamblers could overreact to the savings lottery by mistaking how much of a gamble has been staked (just the interest, not the whole deposit).

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC

ARTICLE IN PRESS J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

havior does not provide intuition for the substitution effect, and would need a separate motivation for why the two types of gambling are substitutes. Thus, although some of the heterogeneous effects can be explained by better quality financial behavior, self-control is a more parsimonious and unifying explanation for the substitution effect between savings lotteries and casino gambling. 7. Conclusion Saving whilst gambling is a promising idea to influence household savings rates (Tufano, 2008). This paper introduces new transaction-level data on casino cash withdrawals to study how those affected by the introduction of PLS change their gambling behavior. Because I am able to evaluate gambling decisions by individuals, this paper offers a new perspective not just about the likely effects of PLS, but also about how individuals evaluate gambling across different contexts. The lottery component of lotterylinked savings accounts genuinely substitutes for consumer gambling, and the effects are large, amounting to at least 3% of overall gambling. These findings suggest that offering innovative savings products like Save-to-Win have under-appreciated effects beyond influencing savings rates. In showing that significant gambling resources were redeployed when savings lotteries become available, my findings imply that there are important effects of prize-linked savings on actual gambling expenditures. These effects ought to be considered as policymakers consider prize-linked savings as a tool to improve consumer financial outcomes (Filiz-Ozbay et al., 2015; Cole et al., 2017). On the other hand, prizelinked savings accounts are not a panacea for improving household savings and gambling behavior. An important subpopulation of casino patrons—those with low selfcontrol—do not respond to the introduction of savings lotteries. Although this finding implies that prize-linked savings accounts are not a full replacement for other savings enhancement programs, the significant effects that I find for those who exhibit greater self-control suggest that PLS accounts and other programs can be complements. Under the right conditions, savings lotteries can dramatically reshape consumption patterns, thereby increasing saving. References Andrikogiannopoulou, A., Papakonstantinou, F., 2016. History-dependent risk preferences: evidence from individual choices and implications for the disposition effect. Unpublished working paper.. London School of Economics and Imperial College. Barber, B.M., Lee, Y.-T., Liu, Y.-J., Odean, T., 2009. Just how much do individual investors lose by trading? Rev. Financ. Stud. 22 (2), 609–632. Barber, B.M., Odean, T., 2008. All that glitters: the effect of attention and news on buying behavior of individual and institutional investors. Rev. Financ. Stud. 21, 787–818. Barberis, N., 2012. A model of casino gambling. Manag. Sci. 58, 35–51. Barberis, N., Huang, M., 2008. Stocks as lotteries: the implications of probability weighting for security prices. Am. Econ. Rev. 98, 2066–2100. Barberis, N., Huang, M., Thaler, R.H., 2006. Individual preferences, monetary gambles, and stock market participation: a case for narrow framing. Am. Econ. Rev. 96 (4), 1069–1090. Becker, G.S., Murphy, K.M., 1993. A simple theory of advertising as a good or bad. Q. J. Econ. 108, 941–964. Becker, G.S., Murphy, K.M., Werning, I., 2005. The equilibrium distribution of income and the market for status. Journal of Political Economy 113, 282–310.

[m3Gdc;April 19, 2018;13:41] 21

Behrman, J.R., Mitchell, O.S., Soo, C.K., Bravo, D., 2012. How financial literacy affects household wealth accumulation. Am. Econ. Rev. Papers Proc. 102 (3), 300–304. Bertrand, M., Duflo, E., Mullainathan, S., 2004. How much should we trust differences-in-differences estimates? Q. J. Econ. 119 (1), 249–275. Beshears, J., Choi, J., Laibson, D., Madrian, B.C., Milkman, K.L., 2015. The effect of providing peer information on retirement savings decisions. J. Finance 70 (3), 1161–1201. Boyer, B.H., Vorkink, K., 2014. Stock options as lotteries. J. Finance 69 (4), 1485–1527. Brown, D.C., Cederburg, S., O’Doherty, M.S., 2017. Tax uncertainty and retirement savings diversification. J. Financ. Econ. 126, 689–712. Brown, M., Grigsby, J., van der Klaauw, W., Wen, J., Zafar, B., 2016. Financial education and the debt behavior of the young. Rev. Financ. Stud. 29 (9), 2490–2522. Calvert, S., Kamp, J., 2014. Casino boom pinches Northeastern states. Wall Street J https://www.wsj.com/articles/ casino- boom- pinches- northeastern- states- 1403220428. Casino City Press, T., 2012. Gaming business directory. Online. http:// online.casinocity.com Chetty, R., Friedman, J.N., Leth-Petersen, S., Nielsen, R.H., Olsen, T., 2014. Active vs. passive decisions and crowd-out in retirement savings accounts: evidence from Denmark. Q. J. Econ. 129 (3), 1141–1219. Choi, J., 2015. Contributions to defined contribution pension plans. Ann. Rev. Financ. Econ. 7, 161–178. Cole, S., Iverson, B., Tufano, P., 2017. Can gambling increase savings? Empirical evidence on prize-linked savings accounts. Unpublished Working Paper.. Harvard Business School, Brigham Young University, and University of Oxford. Cole, S., Paulson, A., Shastry, G.K., 2014. Smart money: the effect of education on financial outcomes. Rev. Financ. Stud. 27 (7), 2022–2051. Collard-Wexler, A., 2014. Adjacent county data. Technical Report. New York University, Stern School of Business. available online http:// pages.stern.nyu.edu/∼acollard/Data.html Cookson, J.A., 2016. Anticipated entry and entry deterrence: evidence from the American casino industry. Manag. Sci. Forthcoming https: //pubsonline.informs.org/doi/abs/10.1287/mnsc.2017.2730. Cookson, J.A., 2017. Leverage and strategic preemption: lessons from entry plans and incumbent investments. J. Financ. Econ. 123 (2), 292–312. Doran, J., Jiang, D., Peterson, D., 2012. Gambling preference and the New Year effect of assets with lottery features. Rev. Finance 16 (3), 684–731. Dorn, A.J., Dorn, D., Sengmueller, P., 2014. Trading as gambling. Manag. Sci. 61 (10), 2376–2393. Dorn, D., Sengmueller, P., 2009. Trading as entertainment? Manag. Sci. 55 (4), 591–603. Fernandes, D., Lynch, J.G.J., Netemeyer, R.G., 2014. Financial literacy, financial education, and downstream financial behaviors. Manag. Sci. 60 (8), 1861–1883. Filiz-Ozbay, E., Guryan, J., Hyndman, K., Kearney, M.S., Ozbay, E.Y., 2015. Do lottery payments induce savings behavior? Evidence from the lab. J. Pub. Econ. 126 (1), 1–24. Friedman, M., Savage, L.J., 1948. The utility analysis of choices involving risk. J. Pol. Econ. 56, 279–304. Gao, X., Lin, T.-C., 2015. Do individual investors trade stocks as gambling? Evidence from repeated natural experiments. Rev. Financ. Stud. 28 (7), 2128–2166. Grinblatt, M., Keloharju, M., 2009. Sensation seeking, overconfidence, and trading activity. J. Finance 64 (2), 549–578. Hastings, J., Hortacsu, A., Syverson, C., 2017. Advertising and competition in privatized social security: the case of Mexico. Econometrica 85 (6). Kearney, M.S., 2005. State lotteries and consumer behavior. J. Publ. Econ. 89, 2269–2299. Kearney, M.S., Tufano, P., Guryan, J., Hurst, E., 2011. Making savers winners: an overview of prize-linked savings products. In: Mitchell, C., Lusardi, A. (Eds.), Financial Literacy: Implications for Retirement Security and the Financial Marketplace. Oxford University Press, New York, pp. 218–240. Kuhnen, C.M., Miu, A.C., 2017. Socioeconomic status and learning from financial information. J. Financ. Econ. 124, 349–372. Kumar, A., 2009. Who gambles in the stock market? J. Finance 64, 1889–1933. Li, G., 2012. Gamblers as personal finance activists. Working Paper 2012-18. FEDS. Li, X., Subrahmanyam, A., Yang, X., 2018. Can financial innovation succeed by catering to behavioral preferences? Evidence from a callable options market. J. Financ. Econ. 128 (1), 38–65. Liao, C.M., 2015. Risk taking begets risk taking: evidence from casino openings and investor profiles. Working Paper. SSRN 2639700

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001

JID: FINEC 22

ARTICLE IN PRESS

[m3Gdc;April 19, 2018;13:41]

J.A. Cookson / Journal of Financial Economics 000 (2018) 1–22

Lusardi, A., Mitchell, O.S., 2014. The economic importance of financial literacy: theory and evidence. J. Econ. Lit. 52 (1), 5–44. Lusardi, A., Schneider, D., Tufano, P., 2011. Financially fragile households: evidence and implications. Brookings Papers on Economic Activity. Spring, pp. 83–134. Madrian, B.C., Shea, D.F., 2001. The power of suggestion: Inertia in 401(k) participation and savings behavior. Q. J. Econ. 116 (4), 1149–1187. Meyer, B.D., Mok, W.K., Sullivan, J.X., 2009. The under-reporting of transfers in household surveys: its nature and consequences. Working Paper No. 15181. NBER. Peng, L., Thibodeau, T.G., 2017. Idiosyncratic risk of house prices: evidence from 26 million home sales. Real Estate Econ. 45, 340–375. Rose, I.N., 1991. The rise and fall of the third wave: gambling will be outlawed in forty years. In: Eadington, W., Cornelius, J. (Eds.), Gambling and Public Policy: International Perspectives. University of Nevada Press, Reno, NV, p. 70. Rosett, R.N., 1965. Gambling and rationality. J. Pol. Econ. 73, 595–607. Savor, P., Wilson, M., 2014. Asset pricing: A tale of two days. J. Financ. Econ. 113 (2), 171–201.

Schneider, C., Spalt, O.G., 2017. Acquisitions as lotteries: Do managerial gambling attitudes influence takeover decisions? Crit. Finance Rev. 6 (1), 77–132. Simonsohn, U., Simmons, J.P., Nelson, L.D., 2015. Specification curve: descriptive and inferential statistics on all reasonable specifications. Working Paper SSRN 2294998. Spenkuch, J.L., Toniatti, D., 2016. Political advertising and election outcomes. Working Paper. Kilts Center for Marketing at Chicago Booth. Nielsen Dataset Paper Series 1-046 Thaler, R.H., Benartzi, S., 2004. Save more tomorrow: using behavioral economics to increase employee saving. J. Pol. Econ. 112, S164–S187. Tufano, P., 2008. Saving whilst gambling: An empirical analysis of U.K. premium bonds. Am. Econ. Rev. Papers Proc. 98, 321–326. Viva, L.D., Kasanen, E., Trigeorgis, L., 2017. Real options, idiosyncratic skewness, and diversification. J. Financ. Quant. Anal. 52 (1), 215–241. Wooldridge, J.M., 2003. Introductory Econometrics: A Modern Approach, second ed. Southwestern Publishing, Thompson Learning, Mason, OH. Zhang, H., 2017. Pre-matching gambles. Unpublished Working Paper.. Michigan State University.

Please cite this article as: J.A. Cookson, When saving is gambling, Journal of Financial Economics (2018), https://doi.org/10.1016/j.jfineco.2018.04.001