Gambling, Saving, and Lumpy Liquidity Needs

Sylvan Herskowitz∗

March 22, 2018Link to most recent version

Abstract

I present evidence that unmet liquidity needs for indivisible, “lumpy”, expendituresin the presence of financial constraints create demand for betting as a second-bestmethod of liquidity generation. With a sample of 1,842 sports bettors in Kampala,Uganda, I conduct a natural experiment on expenditures following wins, a randomizedsavings treatment, and two lab-in-the field experiments. The savings treatment reducesdemand for betting by 28% while two lab-in-the field experiments show that unmetliquidity needs and saving ability are important driving mechanisms behind the result.These results cannot be explained by betting as a purely normal good.

Keywords: Gambling, Betting, Saving, Liquidity NeedsJEL Codes: D14, D81, O12, O16, L83

∗Markets, Trade, and Institutions Division, International Food Policy Research Institute. I am deeplygrateful to Jeremy Magruder and Elisabeth Sadoulet. I would also like to thank Michael Anderson, PierreBachas, Patrick Baylis, Peter Berck, Lauren Bergquist, Josh Blonz, Fiona Burlig, Alain de Janvry, AlumaDembo, Seth Garz, Tarek Ghani, Marieke Kleemans, Ken Lee, David Levine, Ethan Ligon, Aprajit Mahajan,Craig McIntosh, Ted Miguel, Leo Simon, Chris Udry, Liam Wren-Lewis, and Bruno Yawe for their feedbackand contributions. I would also like to thank Antoine Guilhin, Jackline Namubiru, Amon Natukwatsa,and Innovations for Poverty Action for excellent work implementing the study. This project would nothave been possible without generous funding support from a Doctoral Dissertation Research Grant fromthe National Science Foundation, CEGA’s EASST Collaborative, the Rocca Center at UC Berkeley, theAAEA, and the Institute for Money, Technology, and Financial Inclusion at UC Irvine. The project hadIRB approval from the University of California’s Institutional Review Board: 2014-03-6147. The originalstudy design was registered at the Social Science Registry: AEARCTR-0001025. All errors are my own. Themost recent version and online appendix can be found at www.sylvanherskowitz.com/betting.html. Email:s.herskowitz@cgiar.org.

1

1. Introduction

Gambling has been popular for millennia.1 Today, it is a global industry with revenuesestimated at nearly half a trillion dollars.2 Over the past decade, sports betting has emergedas one of the fastest growing forms of gambling, itself frequently valued over a hundred billiondollars.3 While the world’s largest betting markets have historically been in major developedcountries such as the United Kingdom, United States, Japan, Australia, and China, newtechnologies have enabled international companies to enter previously untouched markets.4

As a result, growth of sports betting has been fastest across Africa and in many developingcountries over the past decade.5

Demand for gambles, particularly those with negative expected returns, has presented along-standing puzzle for economists. Theoretical work points to a wide range of explana-tions for its broad and persistent popularity. Proposed factors include addiction (Becker andMurphy 1988), misperception of odds (Barberis 2013; Bordalo et al. 2012), and fun (Conlisk1993) among others. These prevailing explanations contribute to an image of gambling aseither an indulgence or error on the part of the participant. However, seminal work by Mil-ton Friedman and Leonard Savage (1948) posited a source of rational demand for gamblesresulting from non-concavities in peoples’ indirect utility curves. Kwang (1965) then showedthat non-concavities could result from underlying demand for indivisible, “lumpy”, expen-ditures and their accompanying liquidity needs. Further work then argued that access tocredit or ability to save should reduce this motivation for gambling, by affording alternativeways to satisfy unmet liquidity needs (Bailey et al. 1980). Efforts to test these relation-ships empirically have been limited. This paper is one of the first to present evidence thatfinancial constraints can push people towards gambling as a second-best method of liquiditygeneration.

Uganda is, in many ways, an ideal setting to test these linkages. Sports betting hasexploded in Uganda over the past decade. As of 2016 there were over twenty-five legallyregistered betting companies with more than 1,000 separate branches operating throughoutthe country. A recent report estimated that in Kampala, the nation’s capital, 36.7% of males

1Schwartz (2013)2https://www.statista.com/statistics/253416/global-gambling-market-gross-win/3http://www.statista.com/topics/1740/sports-betting/http://www.bbc.com/sport/0/football/24354124

4Industry report from H2 Gambling Capital (2015)5Beck and Cull (2014)PricewaterhouseCoopers (2014)MORSS Global Finance: http://www.morssglobalfinance.com/the-global-economics-of-gambling/

2

above the age of 18 had participated in sports betting over the past year with mean bettingexpenditures among participants reported as 12% of monthly income (Ahaibwe et al. 2016).Furthermore, availability of affordable credit and services to facilitate effective savings arelimited in Uganda.6 Sports bettors in Kampala therefore present a population for whom theappeal of betting may plausibly be driven by unmet liquidity needs in the face of imperfectalternatives for obtaining lumpy sums of liquidity.

The full study sample consists of 1,842 sports bettors in Kampala. 1,003 men wereincluded in a full study consisting of five bi-weekly visits, creating a unique, high-frequencypanel of reported betting behavior, business performance, and expenditure data. This wasthen supplemented with an additional sample of 839 participants in a condensed single-visit study. Using these two samples, the resulting analysis includes a randomized fieldexperiment, a natural experiment, and two lab-in-the-field experiments to generate fourpieces of empirical evidence in support of the theory that unmet liquidity needs along withfinancial constraints create demand for betting.

After presenting descriptive evidence that bettors both perceive and use sports bettingas a method of liquidity generation in pursuit of large expenditures, I outline a brief model.Building on work by Crossley et al. 2016, I show that demand for liquidity can generatedemand for gambles while improved ability to save will reduce this demand. I then usea natural experiment to show that expenditure behavior is consistent with respondents’stated motivations for betting as a way to make lumpy purchases. Next, I use a randomizedfield experiment to show that introducing a commitment saving technology lowers bettingdemand as measured in an offer of cash or betting tickets administered by the research team.I then conduct a pair of lab-in-the-field experiments to directly test whether people demandbets as a mode of liquidity generation and in response to low saving ability. Betting demandincreases after experimentally inducing greater salience of a desired lumpy expenditure. And,betting demand decreases following a budgeting exercise for participants who learn that theircapacity to save was better than previously believed. These findings cannot be explainedby the primary alternative hypothesis that betting demand is purely driven by its value asa consumption good. I conclude with a set of back-of-the-envelope calculations, showingthat the expected returns for betting and saving may be similar for many people within areasonable range of patience levels and returns to saving.

6See Dupas et al. (2016) for discussion of specific saving challenges in Uganda and Karlan et al. (2014) fora general review of the literature on saving challenges in developing country settings. African DevelopmentBank (2011) and Beck and Cull (2014) provide information on the high cost and limited availability of creditin Uganda.

3

Bettors both view and report to demand betting as a means of liquidity generation. 79%of respondents said that their primary reason for betting was as a way to get money (funwas cited by just 15%) while betting was the second most frequently cited likely sourceof liquidity for a desired expenditure (following saving). If these responses reflect theirtrue motivation, then winnings should increase lumpy expenditures. While the size andlikelihood of winnings are not random, being chosen by bettors themselves, winnings shouldbe effectively random after conditioning on the number and types of bets an individualmakes. Implementing this selection on observables design, I test for the impact of winningson lumpy expenditures and find that winnings increase both the size and likelihood of makinglumpy expenditures. Results are strongest among respondents categorized as having a lowability to save, consistent with demand for betting being higher for those whose alternativeavailable methods of liquidity generation are limited.

In the second result, I test one of the model’s main predictions: Improved ability to savereduces betting demand. Randomly selected participants were offered a wooden saving box,similar to a piggy bank, to assist them in their ability to save. At the end of the study, onemonth later, participants were offered a choice between cash and betting tickets in a revealedpreference measure of betting demand. Recipients of the saving box were 28% less likely todemand the full amount of tickets offered.

I then use two lab-in-the-field experiments in order to isolate the role of betting as amethod of liquidity generation, the key feature that distinguishes it from other normal goods.The third result uses a randomized dialog in conjunction with the betting ticket offer. In-terviewers asked randomly selected respondents a set of questions related to a previouslyidentified and desired expense, priming them with increased salience of this expenditure.Respondents who received this prime before the betting ticket offer were 15.7% more likelyto demand the maximum number of betting tickets. This response suggests that many studyparticipants view betting as a means of acquiring liquidity for their lumpy expense. If bet-ting demand were purely driven by consumption, increased salience of a lumpy expenditureshould not have increased betting demand, and may have led to a reduction if respondentsanticipated using the offered cash for saving toward this expense instead. For those catego-rized with low ability to save, the prime had a 27% effect size, four times greater than thosewith high saving ability. This parallel dimension of heterogeneity with post-win expendituresprovides further evidence that both ex-ante betting demand and ex-post winning usage arelinked with liquidity needs for lumpy expenditures among this group of bettors.

The final empirical result shows that a positive update on perceived saving ability de-

4

creases demand for bets. Before the betting ticket offer, randomly selected respondents wereguided through a brief budgeting exercise that assisted them in making realistic assessmentsof their weekly saving potential. Others were lead through this exercise after the bettingticket offer. The results show that respondents who did the budgeting exercise before thebetting ticket offer and learned that they had more capacity to save than previously believedwere 16-31% less likely to demand the maximum number of tickets for the median sized up-date. If betting were purely a consumption good, new information revealing the availabilityof additional disposable income should have led to an increase in demand for all normalgoods, including betting, and not the decrease that we observe.

Together, these results tell a consistent story: sports betting is in competition with savingas a mode of liquidity generation. However, expected losses of 30-50% per dollar spentsuggest that sports betting as a liquidity generation strategy comes at a very high price.Given these expected losses, can betting demand really be rational? This depends largely onthe available alternatives. Limited access to affordable credit and demand for non-creditableexpenditures make borrowing infeasible or unappealing to most people in the sample. Back-of-the-envelope calculations suggest that, after accounting for future discounting and theimpact of inflation, exposure to temptation, social pressures, and risk of loss or theft onpeoples’ return to saving, betting may offer a better return than cash saving, or possiblyother available forms of saving as well, for many people in the sample.

Local media and political figures in Uganda have expressed increasing concern aboutthe social effects of sports betting, including crowding out of scarce household resources,dis-saving, domestic violence, and even suicide, concerns that are echoed and shared byothers throughout the world wherever gambling and betting are popular.7 If policy-makerswant to reduce betting, this paper suggests that improving financial services and alternativestrategies of liquidity generation for vulnerable populations can be effective. Even the simplesaving interventions in this study, focused on budgeting and commitment saving, loweredbetting demand. More ambitious initiatives like low-cost secure banking and mobile savingservices or broadened access to affordable credit could plausibly have larger impacts.

This paper contributes to multiple areas of economic research. While there has beensome theoretical and descriptive work on the linkages between demand for gambles, lumpyliquidity needs, and financial constraints, empirical work showing these causal relationships

7http://allafrica.com/stories/201603150296.htmlhttp://www.monitor.co.ug/Business/Prosper/The-price-of-betting-on-Ugandans/-/688616/2107602/-

/k7i4bh/-/index.htmlhttp://www.monitor.co.ug/News/National/Soccer-fan-kills-self-over-Arsenal-s-loss-to-Monaco/-

/688334/2639990/-/dn6tkoz/-/index.html

5

is limited.8 The few papers that do exist present mixed evidence and are based on samplesfrom developed countries where financial services are typically better functioning. Snowbergand Wolfers (2010) examine American horse betting and find that misperceptions of oddsdrive the well-established long-shot bias more than demand for high payouts, although theydo not rule out liquidity needs as a contributing factor. Two other papers focus on usageof winnings: Imbens et al. (2001) show that lottery winners purchase large durable goodsfollowing wins. Crossley et al. (2016) present similar evidence in the United Kingdom, goinga step further to show that credit-constrained people who buy lottery tickets use infusions ofliquidity from inheritances to make lumpy expenditures, suggestive that bettors face bindingliquidity constraints. While consistent with financial constraints leading to gambling demand(and similar to the first result in this paper), these studies are not able to show that thisex-post behavior is a driver of ex-ante demand.

Second, this paper links closely to a number of themes in the development literature,including work on the financial lives of the poor, saving constraints, and temptation goods.Work by Collins et al. (2009) and Banerjee and Duflo (2007) have shown how financial man-agement strategies of the poor often take unusual forms in response to the constraints theyface in their economic and non-economic lives. This paper shows that betting is being used asa second-best strategy of liquidity generation in the presence of financial constraints. In par-ticular, this paper speaks to work on saving behavior and the impact of saving constraints.9

A related paper from Casaburi and Macchiavello (2016) shows that Kenyan dairy farmersare willing to sacrifice a portion of their income in return for less frequent payments fromupstream buyers, a very different manifestation of saving constraints pushing people towardssecond-best strategies of liquidity generation. Additionally, this paper links to literature ontemptation goods, defined and summarized by Banerjee and Mullainathan (2010). Whilethe consumptive value of betting may fall under this umbrella, this paper shows that thefinancial asset component of betting makes it distinct from other temptation goods and canhave a rational underpinning.

There has also been some recent work examining disproportionately high valuations oflottery linked incentives (relative to flat payments) in both developing and developed countrysettings, with a particular focus on lottery linked savings products.10 Dizon and Lybbert

8See Ariyabuddhiphongs (2011), Bruce (2013), and Grote and Matheson (2013) for recent reviews of theliterature.

9Karlan et al. (2014) provide an excellent summary of this overall literature, with major contributionson the effects of saving constraints on financial investments and resiliency to shocks from Brune et al. (2015)and Dupas and Robinson (2013a,b).

10Kearney et al. (2011) reviews the literature on lottery linked savings in the United States. See examples

6

(2018) and Cole et al. (2017) both show that these products are particularly appealing forpopulations facing other financial constraints. This heterogeneity is consistent with themodel in this paper (and parallels the third empirical result), however it is, in itself unableto establish a causal relationship between these constraints and betting demand.

Finally, the choice of setting is itself a contribution. Ariyabuddhiphongs (2011) conducteda comprehensive review of over 100 gambling studies and found only three based in developingcountries (although his review preceded the work on lottery linked savings cited above).11

Ultimately, an increased focus on how betting relates to financial constraints in developingcountries is extremely important, especially in Africa, where gambling is growing fastestand financial services are lacking. The insights from this work will extend to disadvantagedpopulations wherever existing financial services are failing to meet peoples’ liquidity needs.

This paper proceeds as follows. Section 2 provides further background and details onsports betting and the specific context of Kampala, Uganda. Section 3 motivates the researchquestion with an illustrative model of rational betting behavior. Section 4 provides detailson the research design and data collection. Section 5 presents descriptive evidence of demandfor betting and betting behavior in the sample. Section 6 discusses the main empirical resultsof the project. Section 7 presents back-of-the-envelope calculations comparing betting andsaving in Uganda. Section 8 concludes.

2. Sports Betting Background

2.1. Global Expansion of Sports Betting

Gambling has been popular for millennia. For the entirety of recorded human history, peoplehave sought out opportunities to make wagers and bets, taking on risk for the chance to winmoney or some other reward in return. Dice used for gambling have been found dating backbefore 1300 B.C. while modern gambling encompasses a wide range of activities includinglotteries, bingo, board games, card games, casinos, as well as betting on any of a nearlylimitless range of possible events and outcomes (Schwartz 2013). The gambling industryhas huge global reach and scale with billions of people participating each year. Globalgross gambling yield is estimated at nearly half a trillion dollars with 120% growth since

in developing countries from Cole et al. (2017), Dizon and Lybbert (2018), and Gertler et al. (2018). Brune(2016) shows a related phenomenon using labor supply instead of savings deposits in a lottery linked workincentive scheme.

11I was able to find only one additional paper directly addressing gambling in Africa from Abel et al.(2015), that focuses on experiential learning about compound probabilities in South Africa.

7

the beginning of the century.12 This estimate does not even incorporate unofficial or illegalgambling where data is not available but whose scale may be even bigger.

Estimates of the size of the global sports betting industry vary widely but are frequentlybetween tens and hundreds of billions of dollars.13 Growth has been accelerating over thelast ten years with a recent report from the European Gaming and Betting Associationestimating that regulated gambling markets in Europe grew 19% between 2007 and 2012and projected an additional 20% by 2016.14 In the US, where most sports betting is stillillegal, monetized fantasy sports became a multi-billion dollar industry led by companies likeFan Duel and Draft Kings before recent regulations curbed their expansion.15

Growth has been even faster in many developing countries within Africa. Adaptation ofonline betting technology in the form of internet-linked, vendor-operated betting consolesand betting shops has broadened access to new betting products with higher payoffs and awider range of betting options than have previously been available. These platforms haveallowed new investors to enter into previously unprofitable markets by using internationallycalibrated odds in local betting shops. A 2009 consultant’s report by MORSS Global Financeestimates that between 1999 and 2007 Africa experienced a 114% increase in betting revenues,a faster rate of growth than any other region.16 A 2014 PricewaterhouseCoopers reportestimated that sports betting in South Africa quintupled between 2009 and 2013, from 15.8to nearly 80 million USD in gross revenues.17

Scarcity of reliable data makes it difficult to precisely estimate the size of the sportsbetting industry across much of the African continent. However, what is known is thatinternational companies are rapidly entering and expanding into African and other developingcountry markets.18 In Kenya, the next wave of expansion and innovation is already takingplace with mobile betting technologies reportedly driving the expansion of M-Pesa, Kenya’smobile money platform.19 Regulation varies widely by country, but the appeal of new tax

12https://www.statista.com/statistics/253416/global-gambling-market-gross-win/13www.statista.com/topics/1740/sports-betting/www.bbc.com/sport/0/football/24354124www.ibisworld.com/industry-trends/global-industry-reports/other-community-social-personal-service-

activities/sports-betting-lotteries.html14European Gaming and Betting Association (2014)15http://fortune.com/2015/04/06/draftkings-and-fanduel-close-in-on-massive-new-investments/16www.morssglobalfinance.com/the-global-economics-of-gambling/17PricewaterhouseCoopers (2014)18Recent media articles from Ghana, Nigeria, Senegal, Malawi, Sierra Leone, Tanzania, Liberia, Zim-

babwe, and Kenya all observe a sharp rise in sports betting in their respective countries. Click on thecountry name for a linked article.

19www.techweez.com/2016/05/10/m-pesa-sports-betting/ and www.bloomberg.com/news/articles/2016-05-09/vodafone-mobile-money-volumes-boosted-by-sports-betting-in-kenya

8

revenue streams is a strong incentive for local governments to permit continued entry andgrowth, making the expansion of sports betting across the continent likely to continue.

2.2. Sports Betting in Uganda

Sports betting is a legal, large, and rapidly expanding industry in Uganda. ThroughoutKampala, Uganda’s capital, nearly every commercial center features one or more bettingshops. A typical betting shop in Kampala can accommodate more than forty customers at atime, but at peak hours they often overflow with customers.20 Gambling in different formshas long been a part of Ugandan culture, but this format and ensuing explosion in popularityare new. The arrival and expansion of international betting companies throughout the coun-try began less than ten years ago. As of June 2015, there were 23 licensed betting companiesoperating in Uganda with over 1,000 betting outlets spanning the country (Ahaibwe et al.2016).

A local policy report recently analyzed a representative sample of Kampala residents andfound high participation rates among young men aged 18-40 (Ahaibwe et al. 2016). 36% ofmen in Kampala had gambled at some point in the last year, devoting 12% of their incometo betting on average. Men in the lowest income quintile spend the largest share of theirearnings on betting. The impacts of sports betting have not been rigorously identified, butsurvey responses suggest that betting displaces household expenditures and investments.Meanwhile, local media coverage has reported cases of bankruptcy, loss of school fees, andsuicide as a result of accrued debt and shame attributed to betting.21

The format of betting in Uganda is the same as that spreading across the rest of thecontinent and available on most online sports betting sites. First, a bettor chooses whichmatches to include on his betting ticket from a list of options, typically featuring over 100games. He then predicts a result or outcome for each of these matches such as “Sevilla FCdefeats Manchester United”. Predicting less-likely outcomes or adding additional games toa ticket is rewarded with a higher possible payout should the ticket win. If every predictedoutcome on the ticket occurs, it can be redeemed for its payout value. If any single outcome isincorrect, the ticket is worth nothing. Even by local standards, the minimum cost of placing

20http://www.monitor.co.ug/Business/Prosper/The-price-of-betting-on-Ugandans/-/688616/2107602/-/k7i4bh/-/index.html

21http://allafrica.com/stories/201603150296.htmlwww.monitor.co.ug/Business/Prosper/The-price-of-betting-on-Ugandans/-/688616/2107602/-

/k7i4bh/-/index.htmlwww.monitor.co.ug/News/National/Soccer-fan-kills-self-over-Arsenal-s-loss-to-Monaco/-

/688334/2639990/-/dn6tkoz/-/index.html

9

a bet is relatively low, at just 0.18 USD per ticket. While bettors can target extremely largepayouts if they choose, companies often cap the maximum payout at around 2000 USD, andmost bettors target amounts much lower, around 50 USD. For most participants the cost,payouts, and odds of betting relative to their incomes is more similar to that of Americanscratch tickets than national lotteries or Megabucks. While the payouts and expected returnto a given betting ticket depend on a number of factors, my best estimates suggest thatfor a “typical” betting ticket targeting the median payout from the sample (55 USD), on astandard sized bet of 0.35 USD, and including 7-8 match predictions, has an expected returnof roughly 55% won back per dollar spent, or 45% expected losses. Additional details on thestructure of sports betting in Uganda are contained in Online Appendix B.

3. Conceptual Framework

In this section, I sketch a model of demand for betting as a method of liquidity generationand alternative to saving. The model is an extension and generalization of work by Crossleyet al. (2016), allowing for the flexible form of gambles offered by sports betting operatorsin Uganda and in order to accommodate saving behavior. This section sets up the modeland provides intuition for its predictions in order to fix ideas and motivate the paper’sempirical tests. A more complete treatment is contained in Appendix C. Ultimately, themodel generates four main insights. First, demand for lumpy expenditures creates demandfor gambles. Second, increased valuation of a lumpy expenditure will increase demand forgambles among people who cannot afford to make the lumpy expenditure. Third, demandfor lumpy expenditures also creates demand for saving in an overlapping range of incomelevels as those who demand betting. Finally, improvement in saving ability will decreasedemand for gambles as a method of liquidity generation.

Betting is a bundled good. It includes direct enjoyment from the fun and excitement ofthe activity of gambling and it also serves as a financial asset with the possibility of monetarypayouts. While acknowledging that enjoyment is undoubtedly an important component ofbetting demand, the model focuses on this latter feature. It is the possibility of a payout thatmakes betting distinct from other consumption goods and makes it a plausible (althoughnot necessarily optimal) method of liquidity generation.

A rational agent wants to maximize expected utility subject to a budget constraint ina single time period. His income is Y and utility is derived from the consumption of onedivisible good, D, and the possible purchase of a single unit of a lumpy, indivisible, good:

10

L ∈ {0, 1}. Consumption of the divisible good yields utility u(D), where u′(D) > 0 andu′′(D) < 0. Purchase and consumption of the lumpy good yields a discrete utility payoff, η,and costs a price, P . An agent’s utility is therefore: v() = u(D) + ηL.

In this single period setting, an individual must assess whether the payoff from thelumpy good, η, justifies the reduced expenditure and consumption of the divisible good (byan amount equal to the cost of the lumpy good, P . He must choose between utility with andwithout the lumpy expenditure. Figure 1a illustrates the two pieces of his indirect utility.On the left, without purchase of the lumpy good, he has simple concave utility from pureconsumption of D. The segment on the right shows his utility while consuming the lumpyexpenditure, an upward shift in utility by η and a rightward shift in the curve by P , signifyingthat only Y − P remains to be spent on the divisible good. Given an individual’s incomelevel, he will choose the higher segment and make the associated choice of whether or not topurchase the lumpy good. Those with income levels above Y ∗ will make the purchase whilethose with income below Y ∗ will not. The indirect utility of income is therefore the envelopeof these two curves.

The possibility of consuming a lumpy good has created a non-concavity in the indirectutility function around Y ∗. Some individuals with income in the area of Y ∗ would, in fact,be willing to make certain fair bets (or gambles) risking an amount B for a chance, σ, ofwinning W . For simplicity, we begin by assuming that bets are actuarially fair so thatexpected gains are equal to expected losses, σW = (1− σ)B, such that the likelihood ofwinning is therefore σfair = B

B+W . For those who purchase the lumpy good if and only ifthey win their bet, expected utility from betting will be the probability weighted utilitieswhen they win or lose:

E[vb(Y )] = σ[u(Y − P +W ) + η]︸ ︷︷ ︸If W in

+ (1− σ)u(Y −B)︸ ︷︷ ︸If Lose

An optimal pair of betting outcomes for a fair bet of an individual with starting incomeY can be shown as Y −B∗ and Y +W ∗, with associated utilities of A and C respectively.This is illustrated in Figure 1b. While the amounts of betting and winnings targeted changefor individuals with different income levels, the model shows that these are the optimalpost-realization outcomes for fair bets of people whose income level is between Y −B∗ andY +W ∗. Expected utility for people in this range of incomes who do bet is the convexcombination of utility following a loss or win. For an individual with income Y , Figure 1bshows that this would result in an increase in expected utility from point D to point E. In

11

fact, given the gap between non-betting (D) and betting (E) expected utility, there will alsobe demand for gambles with unfair payouts, whose likelihood of winning given a certain riskand reward pair leads to a loss in expected income but an increase in expected utility. Thisminimum acceptable winning likelihood for an individual with income Y is illustrated inFigure 1c.

For a given rate of return on betting, there is a resulting range of income levels whereindividuals demand bets in pursuit of liquidity. This range will always contain Y ∗. Anincrease in the valuation (anticipated utility) of the lumpy good will increase demand forgambles, both on the extensive and intensive margins. Appendix C.2 contains more details.

Shifting to saving, I expand the model to allow for a second time period. Now, in the firstperiod, individuals can choose to set aside some portion of their income, S, and transfer it tothe next time period with a rate of return on saving of γ. Although γ could be greater thanone in settings with positive interest on saving, participants in this study are more likely tohave γ ≤ 1 because of a variety of well-documented factors including inflation, risk of theftor loss, exposure to temptation, and social pressure from others.22 The discount factor ofthe next period’s utility is δ ≤ 1.

Individuals will only save, transferring income from the first to the second period, if theyare then able and willing to make the lumpy expenditure. If not, then shifting income fromthe first to the second period will lower average two-period utility because of diminishingreturns to consumption of the divisible good, even before accounting for future discount-ing or anticipated losses from saving. For some, the benefits to saving and purchasing thelumpy expenditure (if even possible) still do not outweigh the losses from reduced currentconsumption of the divisible good. Individuals who would rather make the lumpy expen-diture immediately in the first period than save will behave similarly in the second period,removing their incentive to save. However, for a range of income levels in the area of Y ∗,agents will be willing to save and purchase the lumpy expenditure only in the second period.For them, indirect utility from saving will be:

vs(Y ) = u(Y − S) + δ[u(Y + γS − P ) + η]

Optimal saving will be defined as the level of saving, S∗, that maximizes expected utility. Arange of income levels can now be defined for which saving and making the lumpy expenditureis preferred to not making the lumpy expenditure. If a range of individuals who demandsaving exists (because δ and γ are not prohibitively low), it will necessarily include Y ∗.

22See Karlan et al. (2014)’s survey of this literature on the barriers to effective saving.

12

Because the range of income levels at which individuals demand bets also included Y ∗, thesetwo ranges (if they both exist) will overlap. An individual’s preference for either betting orsaving will therefore be determined by the parameters of the model.

Ultimately, betting demand will be reduced in response to improved saving by two sep-arate channels. The model shows that improved saving ability reduces the relative appealof betting as a way to generate liquidity. In addition, the consumptive portion of bettingdemand, contained as part of D, also decreases as income is shifted away from all presentconsumption towards the future. See Appendix C for a more complete treatment of themodel.

4. Experimental Design and Data

4.1. Overview

Field work for the project was conducted over eleven months between September 2015 andJuly 2016, involving three phases of data collection and including 1,842 participants. First,between October and December of 2015, a set of 483 participants were identified and in-cluded in Wave 1 of the study. These respondents were visited and interviewed in personfive times, once every two weeks. A second group of 520 participants were identified and in-cluded in Wave 2, between April and June 2016, following similar protocols. I refer to these1,003 participants as being part of the “full study”. To further explore the link betweensaving ability and demand for lumpy expenditures, a complementary “condensed” study wasconducted with 839 additional respondents over three weeks in July 2016, with all activitiesand interviews contained in a single visit.23

4.2. Listing/Targeting

The study targeted young men between the age of 18-40, self-employed in small micro-enterprises or services, with weekly incomes below $50. Piloting, as well as existing assess-ments of betting in Uganda have suggested that this group has a high incidence and intensityof betting along with unmet liquidity needs (Ahaibwe et al. 2016; Ssengooba and Yawe 2014).This demographic is itself of inherent interest, constituting a significant portion of Uganda’s

23The full size of this condensed study was actually 1,293. However, 581 participants were randomlyassigned to a different treatment group testing hypotheses unrelated to those in this paper about psychological“hot states” and the impact on betting demand. They are excluded from from all analyses presented in thispaper.

13

informal economy and holding important positions in their households.Each survey round began with a listing exercise in selected parishes around Kampala in

order to identify suitable respondents and invite them to participate in the study.24 Par-ticipants were identified at their place of work and asked a short set of initial questions todetermine whether they met the targeting criteria of gender, age, employment, and income.

Listing from both waves of the full study included 5,522 people. Their characteristicsare consistent with piloting, policy papers, and media coverage. Sports betting is extremelypopular in this demographic: 32% reported to bet in most weeks.25 After completing thelisting, a randomized selection of respondents was chosen among those who bet regularly.The full study was launched at the beginning of October. The condensed study was con-ducted in July 2016 and followed the same criteria for eligibility, however respondents wereinterviewed immediately upon identification instead of returning later. Additional details onfield protocols are included in Appendix D.

4.3. Data Collection

Full study participants were interviewed in person five times, once every two weeks. In ad-dition, brief phone check-ins were conducted on the weeks between visits. Surveys captureda wide range of background characteristics and information, including household composi-tion, education, numeracy, literacy, savings background, credit background, and risk andtime preferences. For these topics, unexpected to change over the study period, questionswere asked only once. For responses that were likely to vary from week to week, recurrentsurvey modules were conducted at each in-person interview to record household expendi-tures, household shocks, investments, earnings, transfers, expenses, betting expenditures,and winnings. Phone check-ins were restricted to the noisiest and most important recurrentvariables: weekly earnings, major expenditures, and betting participation.

During the third in-person visit (one month after the initial visit and one month beforethe final visit), members of the research team gave wooden saving boxes to randomly se-lected respondents. These boxes served as a simple soft commitment savings device similarto piggy banks. This basic technology contains features common to many saving products: acomponent of ex-ante commitment to save and a reduction in exposure to spending pressureand temptation. During the final visit of the full study, as well as the first visit for partic-

24In Wave 1, parishes were randomly chosen from the full set of parishes in Kampala that had viablecommercial centers where the target population was likely to be found. In Wave 2, parishes closer to thecity center were targeted due to logistical challenges and budget constraints.

25Appendix Table A.2 summarizes the listing data.

14

ipants in Wave 2, field team members conducted a revealed preference measure of bettingdemand detailed in the next subsection. Randomized primes were conducted in conjunctionwith these betting ticket offers with further details of these activities included in Section 6.

Additional randomized treatments unrelated to the hypotheses in this paper were alsoconducted during the second and fourth visits of the project.26 All treatments were random-ized and included as controls in final estimating regressions.

The condensed study was designed to test a number of hypotheses that could not beincluded in the full study. In particular, it expanded on the priming experiment with abrief budgeting exercise designed to test the effect of perceived savings ability on demandfor betting. It was conducted over three weeks following the conclusion of the full study.Details on this budgeting activity will be provided in the discussion section.

4.4. Measuring Betting Demand

Field team members collected a revealed preference measure of betting demand in the final(fifth) in-person visit for all participants in the full study, as well as in the first visit forparticipants in Wave 2 of the full study. It was also included at the end of the condensedstudy and is an important outcome variable for three of the four empirical results.

Respondents were offered the choice between pre-filled betting tickets and a designatedamount of cash. They were told the amount spent on the ticket as well as the approximatesize of the payout should the ticket win, but they were not permitted to see the actualoutcomes predicted on the ticket.27 The amount of cash offered was less than the priceof the ticket, preventing respondents from taking the money and purchasing a new ticketthemselves, but it was similar to the expected value of the ticket. The cost of the ticket(or stake) was 1,000 Ugandan Shillings (UGX, approximately 0.35 USD), which is the mostcommon value for bets in Uganda. The bets were placed with well-known and trusted bettingcompanies, familiar to all respondents.

Respondents were then asked how many units of cash or betting tickets they would liketo choose. Participants in the full study could select up to four, whereas participants in the

26The second round contained a randomized offer of a wallet with which respondents were encouraged toset aside money and budget for betting. The fourth round contained a randomized information treatmentwhereby selected respondents were given a detailed accounting and aggregating of their betting expensesand winnings up to that point in the study. The endline also included a randomized short video prime offootball highlights. However, technical glitches hindered the ensuing analysis.

27The decision not to show them the tickets was made because of participants’ strong beliefs about theoutcomes of matches, such that they might value a betting ticket at zero if it chose an outcome at odds withtheir opinions.

15

condensed study were limited to two.28 The analyses in Section 6.2., 6.3., and 6.4. use thebinary outcome of “maximum tickets demanded” out of the total number offered.29 Thischoice was made because maximum ticket demand is the highest powered outcome withjust over 40% of respondents having demanded the full number. Results using alternativecontinuous measures of the outcome variable are provided in the appendix.

5. Descriptive Evidence

5.1. Background Characteristics

Descriptive statistics from the survey provide context on the financial situation and con-straints shaping peoples’ betting, saving, and expenditure decisions. Panel (a) in Table 1shows considerable heterogeneity of income levels, betting intensity, household situation,age, and education. It also shows that the full and condensed study samples are broadlysimilar along most of these dimensions.30 For the full study, weekly income and bettingexpenditures were calculated as weekly averages of reported betting and income over thecourse of the study. In the condensed study, respondents were asked how much they spent ina “normal” week. Although the condensed study sample appears wealthier on average thanthe full study sample, they otherwise look similar.

Overall, individual and household incomes are low. Adjusting for children in the house-hold, median per capita income is beneath the two dollar per day poverty line. Meanwhile,betting intensity is high, as the median bettor in both samples spent (or reported to spend)roughly 8% of his weekly income on betting with mean expenditures of 12%, indicating thatsome people in the sample bet at very high levels of intensity.

Survey responses also highlight a number of obstacles people face in their financial lives.Risk of theft (49/59%), pressure to spend (27/34%), and existing debt (43/23%) are all cited

28The basic setup of the betting ticket offer was the same across all groups; however, there were twoadditional differences between the full and condensed study. First, during the full study, participants weregiven the additional choice of whether they wanted tickets that targeted low, medium, or high payouts,whereas in the condensed study the payout size was always medium. Second, the amount of money theycould choose instead of betting ticket was held fixed during the full study but was experimentally variedduring the condensed study. All of these varying factors are controlled for with time and price fixed effectsin the analysis.

29Appendix Table A.1 shows that there is a positive and significant relationship between this measure ofbetting demand and respondents’ reported levels of betting.

30Differences between the full and condensed study samples are not a point of primary concern. The twosamples were drawn from different communities and therefore are likely to differ along certain dimensions.In addition, randomization was conducted by survey round and, therefore, these differences do not threatenthe identification strategies implemented.

16

with high frequency by participants in both samples (full/condensed). Although roughly 90%have mobile money, only 41% have bank accounts and less than 50% felt they would be ableto get a loan from a bank.31

For betting and saving to be relevant strategies for meeting liquidity needs, pre-existingaccess to liquidity should be low. Table 1 Panel (b) confirms this. In particular, the variable“Available Liquidity” is respondents’ answer to the question, “What is the biggest expenseyou could make without needing to borrow?” The majority of participants could not affordan expense greater than 1.5 times their normal weekly income without borrowing money.The table also shows that the distribution of targeted betting payouts is above the level ofrespondents’ available liquidity. Correlations between the log of the targeted win amountand the log of available liquidity or mean income are 0.082 and 0.213, respectively andhighly statistically significant (p < 0.01), providing further suggestive evidence that financialcapabilities relate to betting behavior.

Finally, respondents were asked directly, “what is the biggest reason that you bet?” Over79% of respondents said that their primary motivation for betting was “a way to get money”.The second most common answer was simply “fun”, cited by just 15% of respondents. Theseresponses suggest that bettors themselves consider the possibility of a financial payouts themost important feature of sports betting in determining their participation. While surveyresponses should always be treated with healthy skepticism, the dominance of this responsesuggests that it should also be taken seriously.

5.2. Lumpy Expenditures and Liquidity Generation

A primary assumption of this paper is that bettors want to make lumpy expenditures thatthey cannot currently afford (or are unwilling to pay for). Without this, utility will nothave a non-concavity of the form outlined in Section 3, removing this source of demandfor gambles. In the full study, interviewers asked respondents about three categories ofpotential desired lumpy expenditures: business investments, household expenditures, andpersonal expenditures. Enumerators explained that these should be indivisible expenses andthat they should be realistically attainable in order to avoid purely aspirational reportedtargets. Table 2, Panel (a) shows their responses. Most notably, the majority of respondentscould readily list an expense for each of the three categories. Only 5.8% were unable toidentify a desired expenditure from any of the three categories.

During the condensed study, after being asked to identify a desired large expenditure,

31Appendix Table A.3 summarizes these factors.

17

interviewers asked respondents how they thought that they were most likely to get fundsfor this purchase. Table 2, Panel (b) shows these responses. Although saving is viewed asthe most likely source (by a considerable margin), betting is second. 24.8% of respondentsconsidered betting to be a likely source of funding, almost as much as family and friends,bank loans, or money lenders combined.

5.3. Availability of Credit

The data provide evidence of why credit is viewed so unfavorably in this sample. First andforemost, the cost of bank credit is high. Respondents report that they expect to pay 20-25%interest on a six-month bank loan. In terms of income losses, this would be the equivalentof a betting ticket that returned 80-83 cents per dollar spent. Lack of access is also an issue.Only 41% of respondents have a bank account, a pre-requisite for most loans. Of those whodo not have an account, one out of three said they were deterred by high usage fees, while15% said that they did not have the required documents or had been refused in the past.When respondents were asked directly about their prospects of getting a loan, just 48%thought that bank credit was a possibility.

The previous section also showed that people demand many different types of lumpyexpenditures. While business expenses may be eligible for bank loans, banks do not typicallygive credit for furniture, repairs, phones, or clothes, all frequently cited in this population.85% of respondents had a non-business expenditure that they were eager to make in thecoming months.

In contrast to banks, local money lenders do not restrict how borrowers use their loans.However, despite their availability and low barriers to borrowing, they are not viewed favor-ably; just 2.3% of respondents cited money lenders as the likely source of money for theirdesired expense, and they were the only source cited less frequently than bank loans. Thestandard money lender rate in Kampala is 50% interest on a six-month loan. This is equiva-lent to 33% expected losses. Although this is still slightly better than betting, after factoringin the possibility of default, penalties, and risk of losing collateral, the expected losses frommoney lender credit are comparable with betting.32 Credit from money lenders is likely so

32Missing a scheduled repayment to a money lender typically results in penalties substantial penalties. Ata minimum, money lenders typically double the interest fees for missed monthly repayments. Someone whoborrows 60 USD is expected to pay back 15 USD per month (10 USD toward the principal and 5 USD towardthe interest) for six months, a total repayment of 90 USD. A single missed payment raises the repaymenttotal to 95 USD, dropping the rate of return from 67% to 63%, a rate that is now in the same range asbetting. The greater risk is that the money lender decides not to be lenient, and keeps whatever collateralis being held as leverage for repayment. Even a low risk of missing a repayment and losing one’s collateral

18

unpopular precisely because it is not an improvement over betting.

5.4. Saving Ability

Saving ability is the key dimension of heterogeneity used in this paper. In particular, it isused in the analysis of the impact of winnings on lumpy expenditures and in the responseof betting demand to an increase in the salience of a desired lumpy expenditure. Table 1Panel (b) shows “saving ability” as the response to the question, “How much money do youthink you could save in a normal week without straining your regular household finances?”Median saving ability relative to income in the sample is between 20 and 30% of weeklyincome.

Simply claiming that everyone below the median in the proportion of income they couldset to savings is a low ability saver is appealing but problematic. This measure is highlycorrelated with income. In order to remove this correlation, individuals’ are instead onlycompared to others within the same income decile of their survey group (wave 1, wave 2,or condensed), removing 95% of the correlation between this indicator of saving ability andmean income from 0.3234 to 0.0155. Although this measure may still correlate with othercharacteristics, it serves as a reasonable first approximation of people who are relatively moreor less able to use saving as a way to generate liquidity.

This section has provided some descriptive credence to the hypotheses and assumptionsof the theory presented in this paper. People have limited liquidity available, face a high costof credit, have considerable constraints on their ability to save, view betting as a means ofacquiring liquidity, and consider it a likely source of liquidity for a set of desired expenditures.Taking these responses seriously suggests that we should revisit and test theories of rationaldemand for gambles as a financial asset.

6. Results

This paper contains four main empirical results. The first result looks at expenditure behav-ior and shows that winnings increase the size and frequency of lumpy purchases, consistentwith the motivations for betting articulated in the previous section. Second, I test the effectof a change in saving ability on betting demand, using the randomized offer of a simple com-mitment saving technology, and find a significant reduction in response to the treatment. Inorder to isolate the role of betting as a means of liquidity generation, I then conduct two

will push the expected value of credit below 60% and into the range estimated for betting.

19

lab-in-the-field experiments. I use a randomized priming treatment to show that increasedsalience of a desired lumpy expenditure increases betting demand. And finally, I use a ran-domized budgeting exercise to update participants’ perceptions of their own ability to saveand find that those who receive a positive update reduce their demand for betting.33

6.1. Effect of Winnings on Lumpy Expenditures

The previous section showed that respondents view betting as a likely source of liquidityfor desired lumpy expenditures; that they claim to bet because it is a way to get money;and that the amounts that they target correlate with their income levels and the amount ofliquidity to which they have access. However, if winnings do not increase lumpy expenditures,this stated motivation could just be cheap talk. This first result confirms that winnings doincrease lumpy expenditures, and that this response is driven primarily by people with lowability to generate savings, consistent with the hypothesis that betting is a strategy forliquidity generation most appealing to those with limited alternatives.

Figure 2 shows the biggest win observed in the data for each participant in the full study,scaled by his mean income. Over 60% of the population won some amount over the course ofthe study, with many having won substantial sums relative to weekly earnings. If winningswere randomly assigned, we could look directly at the impact of these wins on expenditurebehavior. However, the amount that one can win depends on the number and types ofbets that he places. To identify a causal effect of winnings on expenditures, I implement aselection on observables approach. This is done by characterizing every individual’s bets ineach time period of the study. Knowing the amount spent on betting, the number of ticketspurchased, the payoff they were targeting (on average), and the number of independentmatches included on the betting tickets, I can also infer the bookmakers’ assessment of thelikelihood that a bet will win. I then characterize the full distribution of potential bettingrealizations for each bettor in each time period by their moments. Controlling for thesemoments and their respective higher order terms, I consider the realized amount of winningsto be effectively random.34 Additional details about the structure of betting in Uganda arecontained in Appendix B1, while Appendix B3 provides details on the conversion of reportedbets into the moments of a betting portfolio.

33Analysis samples in this section differ by result as summarized in Appendix Table A.4. This is becausedifferent identification strategies have different data requirements that are not available for all groups ofparticipants. I use the largest sample possible for each analysis.

34This approach is similar to that utilized by Anderson (2016) in his analysis of the impact of collegesports success on fund-raising ability, where he argues that, conditional on bookmaker spreads, winning isuncorrelated with potential outcomes.

20

Using this identification strategy, I look at the effect of winnings on the size of an indi-vidual’s largest lumpy expenditure in each time period. The estimating equation used forthe analysis is:

Yi,t = β0 + β1Wi,t +4∑

m=1

3∑b=1

BetMomentsbi,m,t + λXi,t + γi + δt + ψs + εi,t

Yi,t is the size of individual i’s largest expense in that time period (scaled by mean incomethroughout the study) in period t. Wi,t is the amount of reported winning in a given two-weekperiod (also scaled by mean income). BetMomentsi,t is the calculated moments, m, (mean,variance, skewness, and kurtosis) and higher-order terms (linear, quadratic, and cubic) ofthe betting portfolio for individual, i, in period, t, and Xi,t are time-varying covariates whichinclude betting expenditures and income. Individual fixed effects, survey round fixed effects,and time fixed effects are also included. Standard errors are clustered at the individual level.

Table 3 presents these results. Column (1) shows the raw correlations between biggestexpenditures and winnings. Column (2) adds in the full set of fixed effects and covariates.Column (3) also includes the moments of peoples’ bets in that week and is the preferredspecification. In general, we see a significant and positive relationship between wins andlarge expenditures. The estimate in column (3) suggests that an additional dollar of winningswould lead to a 22.4 cent rise in the size of an individual’s biggest expenditure in that timeperiod (both the independent and dependent variables are scaled by mean income). If lumpyexpenditures are not made immediately, this may be an underestimate.

However, we expect this effect to be greatest among people whose ability to use savingas an alternative method of liquidity generation is limited. Columns (4) and (5) split thesample between respondents with low and high saving ability as discussed in Section 5.4.The size of the estimated effect for low ability savers has increased by more than 50% andis more than four times larger than for high ability savers. Column (6) tests this differencein response to winnings and confirms that this difference is highly significant (p = 0.035).

One concern with a linear specification is that outliers may be driving the result. Columns(7) and (8) convert both winnings and expenditures using the inverse hyperbolic sine trans-formation to address skewness of these raw variables without losing observations with valuesof zero. While the magnitudes fall, the results remain highly significant with the effect ofwinnings on respondents with low saving ability significantly larger than for those with ahigher ability to save.

An alternative way to assess the impact of winnings on lumpy expenditures is to look at

21

the effect of winnings on the likelihood of making large lumpy expenditures in that period.To do this, I adjust the outcome variable to an indicator for making a purchase above a giventhreshold. I repeat the estimation for a range of different threshold sizes between zero andtwo times an individual’s income. This range of thresholds, while arbitrary, is in line withthe sizes of actual wins observed in the data and where we may have statistical power todetect an effect. Figure 3a shows these regression results graphically. In the figure, the x-axisrepresents the threshold for the biggest expenditure in that time period, while the y-axis isthe estimated coefficient on the win amount. This figure represents the regression resultsshown in Appendix Table A.5. Figure 3c rescales Figure 3a by the mean of the outcomevariable so that it can be interpreted as the proportional increase in likelihood of a purchaseof a given size as a result of additional winnings.

Figure 3b recreates the results in Figure 3a after splitting the sample between low andhigh ability savers. The effect for high ability savers is drawn in red, while low ability saversare shown in blue, with the 95% confidence interval represented with blue dashed lines. Fornearly all of these thresholds, additional winnings have a positive and significant effect onlow ability savers, always larger than for those with high saving ability for whom the effect isnot distinguishable from zero at any threshold. Appendix Table A.6 shows these separatelyestimated results in regression form.

Regardless of specification, additional winnings impact both the size and likelihood ofmaking large lumpy expenditures. This effect is consistently and significantly stronger forthose with low saving ability than those who are more able to generate liquidity throughsavings. Individuals likely anticipate their own future consumption in deciding whetheror not to bet and although ex-post usage of winnings does not definitively confirm ex-antemotivation for betting, this evidence suggests that lumpy expenditures and betting are tightlylinked in this context and corroborate respondents’ stated motivation.

6.2. Commitment-Savings Treatment

The model in Section 3 predicts that an improvement in saving ability should decrease de-mand for betting. To test this causal relationship, I use a randomized field experiment tointroduce a new saving technology and create exogenous variation in saving ability. Ran-domly selected participants were chosen to receive a soft commitment-savings device in theform of a wooden savings box. These boxes are nailed closed and have a small slit in the topso that, like a piggy bank, money can be easily deposited but cannot be retrieved withoutbreaking it open. These boxes are commonly found in Ugandan markets and are therefore

22

familiar to the study participants. At the end of the third visit, field team members gaverandomly selected respondents a saving box and assisted them in writing down their savingtarget on the outside.

In Wave 1, 25% of respondents were selected to receive the boxes, whereas 50% of par-ticipants in Wave 2 were selected.35 Panels (a) and (b) of Table 4 show baseline balancesby wave. Despite random assignment, the endline lumpy good prime was administered toa larger portion of participants who received the saving box treatment than those who didnot. To address this potentially confounding correlation, all analyses control for the effectof the endline lumpy good prime and additional robustness checks are conducted to ensurethat the observed effect is not being driven by an interaction between treatments.36

At the endline, one month after the savings boxes were distributed, interviewers askedparticipants if they had used a savings box at any time in the preceding month. People in thetreatment group were 53 percentage points more likely to have used a saving box comparedto a control group mean of 16%. Although low and high ability savers have slightly differentpropensity to use saving boxes in the absence of the intervention, both responded similarlyto the treatment. Treatment on the treated estimations use random assignment to thetreatment group as an instrument for saving box use.37

I estimate the effect of the saving box treatment using a difference in differences strategywith participants in Wave 2 for whom I have both a baseline and endline measure of bettingdemand. I use the following equation:

Bi,t = β0 + β1SaveBoxi,t + λXi,t + γi + δt + εi,t

Bi,t is an indicator of whether the respondent chose the maximum number of betting ticketsfrom the offer. SaveBoxi,t is an indicator of whether or not individual, i, had been offered asaving box by time t. Xi,t are time-varying covariates for individual i. Individual fixed effectsand time fixed effects are also included. Standard errors are clustered at the individual level.

Table 5 shows these estimation results. Using within variation and controlling for other

35In Wave 1 another 25% were selected to have help setting up formal bank accounts. The tight timeline ofthe study made this extremely difficult and we were unable to successfully open accounts for most respondentsin this group. Assignment to this group is controlled for in the analysis.

36The effect of these treatments is in opposite directions. Therefore, the positive correlation works againstfinding a measurable effect for either result. Both results survive these robustness checks are are included inthe appendices.

37Appendix Table A.7 shows these takeup rates. Control group usage of saving boxes is actually slightlyhigher for people with low saving ability than those with high ability, suggesting that those with low savingability may be aware of their own saving challenges.

23

treatments and weekly income, Column (3) shows that receiving a savings box reduced thelikelihood of demanding the maximum number of tickets by 14.13 percentage points from acontrol group average of 49.84 percent, a reduction of 28% (p-value = 0.011). Column (4)shows that the magnitude of this effect is nearly three times larger for low ability savers thanhigh ability savers, although the difference does not reach statistical levels of significance.Column (7) shows that the treatment on the treated estimate is nearly double the size of theaverage treatment resulting in a 27 percentage point decrease in the likelihood of demandingthe maximum number of betting tickets.38

These are large magnitudes. One potential concern is that, despite randomization, thebaseline measure of betting demand for people with low saving ability in the treatment groupwas significantly higher than for those in the control group. The difference-in-differencesspecification used above should appropriately adjust for these baseline differences and stillidentify a valid causal effect. However, to ensure that the estimated effect is not purely theresult of a baseline irregularity, I re-estimate the effects of the saving box using a cross-sectional analysis of all participants in the full study, ignoring the baseline measure. Thisalternative estimation strategy results in an average treatment effect of a 8.28 percentagepoint (20%) reduction in betting demand (p-value = 0.017) although there is no evidence ofheterogeneous response. These results are shown in Appendix Table A.10.39

Regardless of specification, the analysis shows a large and significant reduction in bettingdemand, between 20 and 28% in the cross-sectional and difference-in-differences estimations,respectively. There is no clear heterogeneity by saving ability as categorized using poten-tial savings. However, a treatment designed to reduce exposure to temptation and socialpressures on money designated for saving is not likely well-suited for a group of people char-acterized by limitations on the share of income that they feel able to designate for saving.Other saving products may be better tailored to these “low ability” savers.

The model suggested that improved saving ability should affect demand for betting38Switching to a continuous measure of betting tickets demanded reveals consistent results although

significance is lost in some regressions from using a lower powered outcome variable. These results are shownin Appendix Table A.8. The average treatment effect (as in Column (3)) shows a reduction of 0.32 ticketsdemanded, down from an average of 2.5, a reduction of 13% (p-value = 0.068).

39Appendix Table A.11 confirms that these results are not driven by correlation with the lumpy expen-diture prime. Additionally they show that they are not resulting from an interaction between the lumpyprime and the saving box treatment. Interacting the saving box treatment with the lumpy prime increasesthe magnitude and significance of the overall average treatment effect. Interestingly, the effect now seemsstronger among people with high saving ability, which might be because they have more discretionary incometo put toward saving. This could result from people feeling that responsible saving decisions in one part oftheir weekly budget frees them to take riskier choices with the remainder of their discretionary spending.Appendix Table A.10 shows that switching the outcome to the count of tickets demanded, a less poweredoutcome, reduces statistical power and loses significance, although the sign of the effect remains the same.

24

through two main mechanisms: crowding out normal goods and undermining the finan-cial appeal of betting. Appendix Table A.13 suggests that the saving box treatment did notaffect available on-hand liquidity (Panel (a) uses a threshold of 3 USD while Panel (b) usesa threshold of 15 USD). Meanwhile, Appendix Table A.12 shows that there is not a consid-erable increase in available liquidity on average resulting from the saving box treatment.40

Regardless of mechanism, the reduction in betting demand of between 20-28% demonstratesan important link between improved saving ability and reduced betting demand. The finaltwo results use lab-in-the-field experiments in order to isolate the importance of the financialmotive for betting.

6.3. Prime on Lumpy Good

The third result uses a lab-in-the-field experiment to show that increasing the salience ofa desired lumpy expenditure increases demand for betting. During the baseline of the fullstudy, interviewers asked respondents to identify a large business, household, or personalexpense they wanted to make in the next few months. During the condensed study, thesequestions were asked at the beginning of the survey. For randomly selected respondents,interviewers went through a dialog referring to these desired expenditures just before thebetting ticket offer. They stated, “Earlier, you mentioned that you wanted to buy .How much would it cost? How much more money do you think you would need in order tobe able to make that expense? Do you know where you would go to purchase it?”41 Thesequestions were designed to make the respondent reflect on this expenditure, increasing itssalience, before measuring betting demand. Right after these primes, respondents wereoffered the choice between betting tickets and cash. Respondents in the control group wereasked these same questions immediately after the betting ticket offer.

To the extent that the appeal of betting results from its potential as a strategy of liquiditygeneration, the model predicts that an increase in salience should increase betting demand.However, demand for betting based on consumption and enjoyment may fall following theprime if the respondent sees the cash offered as potential savings that they could put towards

40There is strong differential impact of the saving box treatment between low and high ability saverswhereby those with low saving ability have significantly more liquidity available to them as a result of thesaving box treatment while those with high ability have a significant decrease. This would be consistent withlow ability savers also having greater reductions in their betting demand, but the instability of this resultwhen switching to the cross-sectional estimation makes this claim tenuous.

41For respondents in the full study, the enumerators first checked to see whether the large expenditurehad already been made and, if so, whether they needed to make that expense again (as in the case of rentor school fees). These answers were controlled for in the analysis. If a respondent said that he already hadthe money necessary to make the purchase readily available to him, he was dropped from the analysis.

25

their desired expenditure. These predictions push demand in opposite directions and it isan empirical question whether one or the other effect will dominate.

The estimating equation for this analysis is:

Bi = β0 + β1LumpyPrimei + λXi + εi

Bi is again whether the maximum number of tickets were demanded. β1 is the coefficientof interest and measures the effect of having received the lumpy prime before the bettingticket offer, where LumpyPrimei = 1 for those in the treatment group and equal to zero forthose in the control group. Xi is a set of covariates for individual i including an indicator forthe time period in which that person was offered the betting tickets. The full specificationalso includes the amount of cash offered and treatment status for the other randomizedtreatments in the study. Robust standard errors are used to adjust for heteroskedasticity inthe error term. Randomization balance is shown in Table 4 Panel (c).42

Because LumpyPrimei is randomly assigned, the estimate of β1 can be interpreted asthe causal effect of the prime on betting demand. Table 6 shows the results. The first fourcolumns show that, regardless of which covariates are included, the lumpy good prime hasa highly significant, stable, and positive effect on the likelihood that a respondent chose themaximum number of tickets. The preferred full specification in Column (3) shows an effectsize of 6.6 percentage points relative to a control group mean of 41.9, a treatment effect of15.7% (p < 0.01). Again, I expect that betting will be particularly appealing as a modeof liquidity generation for people with limited ability to save. Therefore, I split the samplein columns (4) and (5) by the same dimension of heterogeneity as used in the analysis ofwinning usage. The effect size for those with low saving ability is nearly four times largerthan those with relatively high saving ability. The effect of the prime on people with lowsaving ability climbs to 11 percentage points, an effect size of 27% relative to the mean ofthe relevant control group whereas the effect falls to 2.8 percentage points for those withhigh saving ability, statistically indistinguishable from zero. The difference between thesetwo groups is tested in Column (6), showing marginal significance (p = 0.074).43

42None of the 22 variables shown are imbalanced, an unusual degree of balance across treatment andcontrol group statuses. This is because 503 participants in the second wave of the full study were giventhe priming experiment twice (once at baseline and once at endline) and included once in the control groupand once in the treatment group, creating considerably more balance across observable characteristics thanwould occur if each individual were randomly allocated to one or the other treatment status.

43Appendix Table A.14 shows results using the continuous outcome of proportion of tickets demanded.The results look qualitatively similar with somewhat less statistical significance although the test for differ-ential responses is stronger (p = 0.033). Appendix Table A.15 shows that these results remain strong even

26

An alternative possibility is that increased salience of a desired expenditure makes peopleless patient and therefore pushes them towards betting in the hope of getting the neededliquidity more quickly. While it is not possible to experimentally vary respondents’ levels ofpatience, I test for whether people with different measures of baseline patience are more orless responsive to the prime in Appendix Table A.16. These results do not show significantdifferential responses to the prime by survey-based measures of Beta or Delta patience,providing support for the claim that the prime is not purely influencing betting demandamong people with lower levels of patience and low self-control.

Overall, these results show that an increase in the salience of a lumpy expenditure in-creased betting demand. They corroborate a story that betting is perceived and demandedas a credible alternative strategy of attaining desired liquidity for a large expenditure. Ad-ditionally, the results exhibited a parallel heterogeneous response to that observed in theanalysis of the effect of winnings on lumpy expenditures. In the group of people for whomsaving is least likely to be a feasible strategy of liquidity generation, winning usage andbetting demand both appear linked to lumpy expenditures and demand for liquidity.

6.4. Budgeting Exercise for Savings

The final result uses another lab-in-the-field research design to identify the causal effectof a change in perceived ability to save on betting demand. Changing perceived abilityto save should also affect the relative appeal of saving to betting so long as the update iscredible. To do this, interviewers guided randomly selected respondents from the condensedstudy through a brief budgeting exercise. This exercise was nested inside the lumpy primetreatment and aimed at assessing ability to save.44 Table 4, Panel (d) shows the balance bycovariates.45

Early in the survey, interviewers asked respondents about their typical weekly earningsand essential expenditures on food, transportation, and rent. They were asked to identifya large lumpy expenditure that they hoped to make within the next few months and howmuch money they thought they could save per week without putting excessive strain on

while interacting the lumpy prime treatment with the saving box treatment, which we saw in the previoussection, was correlated with the lumpy good prime despite randomization.

44This group was omitted during the lumpy good prime analysis above.45Baseline proportionate saving ability is significantly different across treatment despite randomization.

That some imbalance emerges is to be expected, having checked across 20 variables. What matters mostfor this result is that saving updates are balanced across treatment status. We do not see any statisticaldifferences by raw saving ability, raw saving update amount, and proportionate update amount. Baselinesaving ability levels are included in all regressions to account for this baseline imbalance.

27

their personal or household finances. At the end of the survey, after the lumpy good prime,interviewers guided randomly selected respondents in the treatment group through a briefbudgeting activity before the betting ticket offer. Those in the control group went throughthese same questions immediately following the offer, including those who did not receive thelumpy good prime before the ticket offer. In the budgeting dialog, respondents were told,“Earlier in this interview you said that you earn UGX in a typical week. You also saidyou spent on food, on transportation, and on rent in a normal week. Thisleaves you with UGX per week. How much money do you think you could realisticallysave per week?” Tablets used for data collection automatically calculated and filled in theblanks based on earlier responses. Respondents were unconstrained in their answers to thisfinal question and were free to ignore this information.46 Interviewers did not make anyreference to the respondent’s earlier estimate of saving ability.

The anticipated effect of this treatment depends on whether a respondent’s updatedestimate is above or below his original “naively” estimated saving ability, reported at thebeginning of the survey. In other words, the expected sign of the treatment depends onwhether the respondent learns encouraging or discouraging information about his abilityto save. In the data, 48% of respondents decreased their estimated saving ability, 27%did not change their estimate after the discussion, and 25% increased their estimate. Themedian raw positive update was 15,000 Ugandan shillings (approximately 4.25 USD) andthe median proportionate update was 10% of income. The median raw negative update was17,000 Ugandan shillings (approximately 4.85 USD). The median negative proportionateupdate was 17% of income. Figure 4a shows the raw update size (in thousands of Ugandanshillings) and Figure 4b shows the update scaled relative to mean income. By having botha naive and budgeted estimate of saving ability for each person in the sample, I can assessthe impact of receiving this update on betting demand while controlling for the appropriatecounterfactual of someone who would have gotten that same update.

Because the content of the saving ability update influences the predictions for betting de-mand, I anticipate heterogeneous treatment effects, estimated using the following regression

46After the respondent gave an answer, the enumerator said, “At that rate of saving, it would take youweeks/months to have enough money to make that expense.” Analysis similar to that detailed below

was also conducted with respect to the time update. Learning that saving would take twice as long aspreviously anticipated does increase demand for the maximum number of tickets by 11.6%, although it isonly marginally significant. This is consistent with betting and saving being competing strategies of liquiditygeneration. There is no clear heterogeneity by measures of Beta and Delta discounting. These results are inAppendix Table A.19.

28

equation:

Bi = β0 + β1LumpyPrimei + β2Budgeti + β3(Budget×Update)i + β4Updatei + λXi + εi

Bi is the outcome measure of betting demand from the betting ticket offer. LumpyPrimei

indicates whether the individual received the lumpy prime. Budgeti is an indicator forbeing assigned to the budgeting treatment group and doing the budgeting activity beforethe ticket offer. β2 is the effect of the budgeting activity independent of the update. β3

is the effect of the content of the update. Updatei is calculated as the difference betweena respondent’s new, assisted estimate of saving ability and his original, naive estimate,regardless of whether the budgeting exercise took place before or after the betting ticketoffer. It is positive if their estimated ability to save has increased following the budgetingexercise. In other specifications, the update is scaled by mean income, coded as a binary(increase or decrease) update, and split above and below zero. β4 captures differences inpeople with different update sizes independent of whether they did the budgeting exercisebefore or after the ticket offer. Xi is a set of covariates for individual i. In all specifications Icontrol for the amount of cash offered, as well as all other treatments included in the study.In addition, because this analysis relies on a smaller sample (N=688) than the primingexperiment, I use a wider range of covariates than in the lumpy prime, including amountof available liquidity, whether the respondent lives with others, education levels, scores on amath test, measures of risk aversion, hypothetical demand for gambles, and risk aversion.47

Robust standard errors are used to adjust for heteroskedasticity in the error term.Table 7 shows the results. Column (1) shows that the budgeting exercise had a negative

but insignificant effect on demand for betting tickets. The estimates in Columns (2) and(3) show that, regardless of whether the updating is estimated in raw local currency orconverted to proportion of income, improving perceived saving ability lowers betting ticketdemand. Column (3) suggests that, when an individual learns that he can save 10% moreof his income than previously thought, the median sized positive update, the likelihood thathe will demand the maximum number of betting tickets decreases by 4.8 percentage pointsor approximately 13%.48

47Results are qualitatively similar even using minimal covariates and can be seen in Appendix TableAppendix Table A.17.

48This approach is valid if the betting ticket offer did not affect peoples’ responses in the budgetingexercise which could create systematically different updates by treatment status and lead to invalid controlgroup comparisons. Appendix Table A.18 shows that the betting ticket offer did not significantly affect theraw update size, update size relative to income, or likelihood of a positive, negative, or zero update.

29

As discussed, we expect different responses depending on the content of the update. Col-umn (4) codes the saving updates as positive or negative (zero is the omitted category). Wesee large and negative effects for people receiving positive information about their savingability and an insignificant effect for those learning negative information. A positive up-date causes a 28.2 percentage point reduction in the likelihood of demanding the maximumnumber of betting tickets. The effect for the negative update is indistinguishable from zero.Columns (5) and (6) test for a linear relationship between the update amount and bettingdemand, again splitting the treatment at zero. The raw update in Column (5) estimates asimilar treatment effect on both sides of zero, although the proportionate measure only showsthe significant effect for the positive update. A median positive update of 15,000 shillingswould cause a six percentage point decrease in the likelihood of demanding the maximumnumber of betting tickets, as shown in Column (5), or approximately 16% of the controlmean. Column (6) suggest that a median positive proportional update of 0.10 causes an11.5 percentage point decrease in betting demand, or just over 31% of the control mean.

Imposing linearity on the estimates may be too rigid, so I also conduct a non-parametricestimation of the treatment effect of the budgeting exercise with the proportionate updatemeasure. Figure 5 shows the non-parametric lowess regression of the saving update, scaled byweekly income, on demand for the maximum number of tickets. A linear model with a splineat zero is included for reference. These non-parametric estimates suggest that there is noclear effect of budgeting exercise on people learning negative information, whereas positiveinformation decreases demand for betting tickets with bigger effects for larger update sizes.

Attributing the effect of the saving box treatment to a change in the relative appeal ofsaving and betting as competing methods of liquidity generation was confounded by otherfactors. In particular, I couldn’t rule out that the effect was coming from crowding outall current expenditures on normal goods. However, an update revealing that a person hasmore disposable income available for saving does not face the same challenges. Learningthat you have more available liquidity should increase demand for bets if it is exclusively aconsumption good. These results show the opposite. The reduction in betting demand forpeople receiving positive saving updates suggests that improved perception of the feasibilityof saving as a liquidity generation strategy undercut that source of appeal for betting.

30

7. Comparing Betting and Saving

Section 3 framed saving as the primary alternative to betting, consistent with survey re-sponses discussed in Section 5.2. The results in Section 6 showed expenditure behavioras well as shifts in betting demand consistent with participants valuing betting as a modeof liquidity generation and in competition with saving. This section looks at how bettingperforms relative to saving as a pure financial asset.

Stashing money under ones mattress may be better than 35-50% expected losses frombetting, but it may not. Participants reported a number of challenges impeding their abilityto save effectively, citing risk of theft, pressure from family or friends, and personal tempta-tion. These all contribute to expected “losses” when money is set aside for saving. Inflationlowers this effective interest rate further. Ultimately, given his level of patience, an individ-ual has to consider this expected return on saving and compare it to the return on bettingwhen deciding whether to allocate his money to one or the other strategy.

Acknowledging that different saving strategies have different cost and benefit structures, Iuse the simplest form of saving, cash savings, as a benchmark to compare saving and bettingstrategies. From what is known about the structure of betting, I calculate the expectedpayoff to a person using betting to pursue a desired expenditure. I similarly calculate theexpected payoff to that same person if he instead pursues a (cash) saving strategy. I can thenidentify the return on saving which makes a person indifferent between saving and bettingfor his given level of patience over a reasonable range patience levels.

To do this, I expand the model to allow for more than two time periods and make someparameter assumptions. I set the amount to be either saved or bet at 10% of weekly income(between the mean and median betting expenditures in the population) and the payout of awin equal to the price of the lumpy expenditure at twice the size of the individual’s income(approximately 200,000 UGX or 60 USD). I also must make assumptions about the size offorgone utility from reduced consumption of the divisible good while pursuing the lumpyexpenditure (either from saving or betting) and the payoff to the large expenditure, η, fromthe original model. Forgone utility is treated essentially as a numeraire, and I try a widerange of utility payoffs to the lumpy expenditure; results are not overly sensitive to thisparameter choice.49 The return on betting is captured in the likelihood of winning the bet,p, and estimated based on knowledge of the structure of betting in Uganda for a payout ofthis size detailed in Appendix B2.

49The ratio between forgone utility and the utility payoff to the lumpy good captures the individual’s riskaversion by defining the steepness of an individual’s utility function at that income level.

31

Research in neighboring Kenya by Mbiti and Weil (2013) estimates a yearly discountfactor of 0.64, suggesting a weekly discount factor of approximately 0.9915. With thesevalues, I can examine a range of reasonable weekly discount factors, δ, and solve for the“tipping point” rate of return on saving, γ∗, that leaves people indifferent between bettingand saving. People with γ ≥ γ∗ prefer to save and those with γ < γ∗ prefer to bet. Additionaldetails and derivation are contained in Appendix D1.

Figure 6 traces the locus of weekly discount factors and threshold return to saving. Lesspatient people are willing to tolerate less expected losses from saving than people withgreater patience. For Mbiti and Weil’s estimated discount factor, threshold gamma is 0.716,suggesting that people with this discount factor would prefer betting if they expect morethan 28.4% losses from money set aside for saving. Of course, people who entirely discountthe future, on the left side of the diagram, will always prefer betting over saving. For peoplewith γ < 0.63 who expect to lose 37% of each dollar saved, then even perfect patience (δ = 1)will not be sufficient to sustain saving, a convergence to the expected return on betting.

A comprehensive estimate of γ should incorporate at least five main components: (1)Inflation, (2) Expenditures on Temptation Goods, (3) Social Pressures, (4) Theft/Loss, and(5) Transaction Costs. For each component, I estimate a reasonable range of discount sub-factors γ1−5 and take their product in order to construct a range of return to saving such thatγ = Π5

1γi. Table 8 summarizes the range of sub-discount factors resulting from my estimates.This exercise suggests that the rate of return to saving likely falls between 0.6415 and 0.9215for people in this population. Details of the assumptions and sources of these estimates areprovided in Appendix D2. Comparing this range with the “tipping point” values of returnon saving calculated earlier suggests that, for a substantial portion of people, betting will bethe preferred strategy.

Of course, people save in many ways and each technology or saving strategy has itsown benefits and drawbacks. Some alternative strategies may significantly improve thereturn on saving estimated above for cash savings. For many people, they likely do offeran improvement, or else rates of betting would be far higher and rates of saving far lowerthan even those seen in this select population. Bank accounts do exist in Kampala, andmany respondents have them. But ATM locations are inconvenient and overcrowded leadingto considerable lost time for either work or leisure. Focus group discussions revealed thatdeposit fees cost approximately 3% of the median respondent’s weekly income and manyaccounts also charge withdrawal fees. All of these reduce the return and appeal of a banksaving strategy.

32

Rotating saving groups are also very common in Uganda and can greatly reduce theamount of time before a participant has access to his liquidity (Anderson and Baland 2002).As both a benefit and a risk of group participation, social pressure acts as a commitmentdevice for people who may not always follow through with their saving goals. However, thesegroups have their own limitations and rely on the efficacy of social sanctions among theirmembers for the group to survive (Anderson et al. 2009). The rigidity of payment structuresand the threat of social sanctions impose their own risks. Additionally, the amount ofthe contributions and payout are typically the result of a bargaining process among groupmembers and may not correspond with an individual’s personally optimal amounts or needs.

Finally, mobile money is becoming increasingly widespread in Uganda. However, sincethe time of the study, mobile betting services have followed quickly behind (Ssengooba andYawe 2014). Money put into a mobile money wallet that was previously shielded fromtemptation may now be even more exposed to impulsive betting.

There are undoubtedly many factors leading to such high demand for betting in this set-ting and with this population. Financial motivations may have seemed initially implausibleor far-fetched, but given the constraints and challenges that many people face in their effortsto save, the gap between a return on betting and saving may be very small for many, if itexists at all.

8. Conclusion

Using four empirical results, this paper presented evidence that unmet liquidity needs andfinancial constraints are significant drivers of demand for betting among 1,842 study par-ticipants in Kampala, Uganda. Corroborating their own stated motivations, bettors usewinnings in order to make large lumpy purchases. This effect is driven by those in the pop-ulation with a limited ability to use saving as an alternative method of liquidity generation.Turning to experimental results on betting demand, a randomized savings box interventionreduced demand for betting. I could not, however, disentangle two plausible mechanisms:that improved saving crowds out all current expenditures and that the intervention reducedthe relative appeal of betting as an alternative form of liquidity generation. I thereforeuse two lab-in-the field experiments to provide further evidence in support of the lattermechanism. An experimental prime aimed at increasing the salience of a desired lumpy ex-penditure significantly increased demand for bets, particularly among those with low savingability, while learning positive information about saving ability through a budgeting exercise

33

decreased betting demand.The findings in this paper contribute to a broader theoretical literature on gambling

and to a limited existing empirical literature showing the causal links between demand forgambles, unmet liquidity needs, and financial constraints. It additionally speaks directly toa large literature in development economics on constraints leading to second-best financialstrategies. The losses from betting are large and meaningful for a population living at orbelow the poverty line and the pervasiveness of demand for bets in pursuit of liquidity belieshow unappealing the existing alternatives are for many people in the study.

Saving is a struggle for people across the globe, but in developing countries these chal-lenges may be particularly severe. And yet, even the simple interventions tested in this studyreduced betting demand. More ambitious interventions, such as lowering the cost of securesaving or expanding access to affordable credit may have stronger effects. Broadly, if policy-makers are interested in deterring gambling, marginalized populations need better financialservices and alternative ways to generate liquidity in order to undermine its financial appealand avoid its high associated costs as a mode of liquidity generation.

While the choice of setting and sample in Uganda was done with the intention of testingthese linkages, the behavior of the participants documented in this paper is likely relevantto populations outside of Uganda as well. Gambling and betting are, of course, not uniqueto Uganda, with problem gambling in the United States and the United Kingdom knownto be highest in marginalized populations (Crossley et al. 2016; Welte et al. 2008). Unmetliquidity needs and financial constraints are not unique to Uganda either and the insightsof this paper are likely to speak to the experiences and motivations of billions of peoplethroughout the world.

9. ReferencesAbel, Martin, Shawn Cole, and Bilal Zia (Feb. 2015). “Debiasing on a Roll: Changing Gam-

bling Behavior through Experiential Learning.” In: Policy Research Working Paper 7195,pp. 187–202.

African Development Bank (Dec. 2011). “Institutions for Credit Markets in Africa.” In: AfricaCapacity Development Brief 2 (5).

Ahaibwe, Gemma et al. (2016). “Socio Economic Effects of Gambling: Evidence from Kam-pala City, Uganda.” In: Economic Policy Research Centre (126).

34

Anderson, Michael (2016). “The Benefits of College Athletic Success: An Application ofthe Propensity Score Design with Instrumental Variables.” In: Review of Economics andStatistics (Accepted).

Anderson, Siwan and Jean-Marie Baland (Aug. 2002). “The Economics of Roscas and In-trahousehold Resource Allocation.” In: The Quarterly Journal of Economics, pp. 963–995.

Anderson, Siwan, Jean-Marie Baland, and Karl Ove Moene (2009). “Enforcement in InformalSaving Groups.” In: Journal of Development Economics 90, pp. 14–23.

Ariyabuddhiphongs, V. (2011). “Lottery Gambling: A Review.” In: Journal of GamblingStudies 27, pp. 15–33.

Bailey, Martin J., Mancur Olson, and Pau. Wonnacott (1980). “The Marginal Utility ofWealth Does Not Increase: Borrowing, Lending, and Friedman-Savage Gambles.” In: TheAmerican Economic Review (703), pp. 372–9.

Banerjee, Abhijit and Esther Duflo (2007). “The Economic Lives of the Poor.” In: Journalof Economic Perspectives (21(1)), pp. 141–168.

Banerjee, Abhijit and Sendhil Mullainathan (May 2010). “The Shape of Temptation: Impli-cations for the Economic Lives of the Poor.” In: NBER Working Paper Series 15973.

Barberis, Nicholas (2013). “The Psychology of Tail Events: Progress and Challenges.” In:American Economic Review: Papers % Proceedings 103(3), pp. 611–616.

Beck, Thorsten and Robert Cull (July 2014). “Small- and Medium-Sized Enterprise Financein Africa.” In: Africa Growth Initative at Brookings - Working Paper 16.

Becker, Gary S. and Kevin M. Murphy (Aug. 1988). “A Theory of Rational Addiction.” In:Journal of Political Economy 96 (4), pp. 675–700.

Bordalo, Pedro, Nicola Gennaioli, and Andre Shleifer (2012). “Salience Theory of ChoiceUnder Risk.” In: The Quarterly Journal of Economics, pp. 1243–1285.

Bruce, Alistair (2013). “Betting Motivation and Behavior.” In: ed. by Leighton Vaughan-Williams and Donald S. Siegel. Chap. The Oxford Handbook of the Economics of Gam-bling, pp. 451–473.

Brune, Lasse (2016). “The Effect of Lottery-Incentives on Labor Supply: Evidence from aFirm Experiment in Malawi.” In: Unpublished.

Brune, Lasse. et al. (Nov. 2015). “Commitments to Save: A Field Experiment in RuralMalawi.” In: Economic Development and Cultural Change, pp. 187–220.

Capital, H2 Gambling (2015). “Top 20 Sports Betting Nations 2014.” In: SportBusiness inNumbers 8, pp. 90–91.

35

Casaburi, Lorenzo and Rocco Macchiavello (2016). “Firm and Market Resonse to SavingConstraints: Evidence from the Kenyan Dairy Industry.” In: Working Paper.

Cole, Shawn Allen, Benjamin Charles Iverson, and Peter Tufano (2017). “Can Gambling In-crease Savings? Empirical Evidence on Prize-Linked Savings Accounts.” In: Said BusinessSchool Working Paper 2014-10.

Collins, Daryl et al. (2009). Portfolios of the Poor: How the World’s Poor Live on TwoDollars a Day. Princeton University Press.

Conlisk, John (1993). “The Utility of Gambling.” In: Journal of Risk and Uncertainty (6),pp. 255–275.

Crossley, Thomas F., Hamish Low, and Sarah Smith (2016). “Do Consumers Gamble toConvexify?” In: Journal of Economic Behavior & Organization (131), pp. 276–291.

Dizon, Felipe and Travis J. Lybbert (2018). “Leveraging the Lottery for Financial Inclusion:Lotto-Linked Savings Accounts in Haiti.” In: Working Paper.

Dupas, Pascaline and Jonathan Robinson (Jan. 2013a). “Savings Constraints and Microenter-prise Development: Evidence from a Field Experiment in Kenya.” In: American EconomicJournal: Applied Economics 5.1, pp. 163–92.

— (2013b). “Why Don’t the Poor Save More? Evidence from Health Savings Experiments.”In: American Economic Review 103 (4), pp. 1138–1171.

Dupas, Pascaline et al. (2016). “Banking the Unbanked? Evidence from Three Countries.”In: NBER Working Paper 22463.

European Gaming and Betting Association (2014). Sports Betting: Commercial and IntegrityIssues. http://www.egba.eu/media/Sports-Betting-Report-FINAL.pdf. Accessed: 2016-05-05.

Friedman, Milton and L. J. Savage (Aug. 1948). “The Utility Analysis of Choices InvolvingRisk.” In: The Journal of Political Economy 56.4, pp. 279–304.

Gertler, Paul et al. (2018). “The Long-Term Effects of Temporary Incentives to Save: Evidenefrom a Prize-Linked Savings Field Experiment.” In: Working Paper.

Grote, Kent and Victor A. Matheson (2013). “The Economics of Lotteries: A Survey ofthe Literature.” In: ed. by Leighton Vaughan-Williams and Donald S. Siegel. Chap. TheOxford Handbook of the Economics of Gambling, pp. 670–691.

Imbens, G., D.B. Rubin, and B.I. Sacerdote (2001). “Estimating the Effect of UnearnedIncome on Labor Earnings, Savings, and Consumption: Evidence from a Survey of LotteryPlayers.” In: The American Economic Review (91(4)), pp. 778–794.

36

Jakiela, Pamela and Owen Ozier (2015). “Does Africa Need a Rotten Kin Theorem? Ex-perimental Evidence from Village Economies.” In: Review of Economic Studies 83(1),pp. 231–268.

Karlan, Dean, Aishwarya Lakshmi Ratan, and Jonathan Zinman (Mar. 2014). “Savings Byand For the Poor: A Research Review and Agenda.” In: The Review of Income and Wealth60(1), pp. 611–616.

Kearney, Melissa S. et al. (2011). “Making Savings Fun: An Overview of Prize-Linked Sav-ings.” In: ed. by Olivia Mitchell and Ammamaria Lusardi. Oxford University Press.Chap. Financial Literacy: Implications for Retirement Security and the Financial Mar-ketplace.

Kwang, Ng Yew (Oct. 1965). “Why do People Buy Lottery Tickets? Choices Involving Riskand the Indivisibility of Expenditure.” In: Journal of Political Economy 73.5, pp. 530–535.

Mbiti, Isaac and David N. Weil (2013). “The Home Economics of E-Money: Velocity, CashManagement, and Discount Rates of M-Pesa Users.” In: American Economic Review:Papers and Proceedings (103(3)), pp. 369–374.

PricewaterhouseCoopers (2014). Raising the Stakes in Africa, Gambling Outlook: 2014-2018.http://www.pwc.co.za/en_ZA/za/assets/pdf/gambling-outlook-2014.pdf.

Schwartz, David G. (2013). Roll the Bones: The History of Gambling. Winchester Books.Snowberg, Erik and Justin Wolfers (Aug. 2010). “Explaining the Favorite-Longshot Bias: Is

It Risk-Love or Misperceptions?” In: Journal of Political Economy 118.4.Ssengooba, Kizito and Bruno L. Yawe (2014). “Gambling and Mobile Money Payments: A

Case Study of Sports Betting in Uganda.” In: IMTFI - Working Paper.Welte, JW et al. (June 2008). “The Prevalence of Problem Gabmling among U.S. Adolescents

and Young Adults: Results from a National Survey.” In: Journal of Gambling Studies(24(2)), pp. 119–33.

37

Figures

Figure 1: Indirect Utility with Lumpy Good and Demand for Gambles

𝑌

𝑉

𝑌∗𝑃

𝜂

𝑢 𝑌 − 𝑃 + 𝜂

𝑢 𝑌

No Purchase

Purchase

(a) Indirect utility with a lumpy good

𝑌𝑌�𝑌𝑌�𝑌𝑌 − 𝐵𝐵∗

𝜎𝜎𝑓𝑓𝑓𝑓𝑓𝑓𝑓𝑓 =𝐵𝐵

𝐵𝐵 + 𝑊𝑊

V

�𝑌𝑌 + 𝑊𝑊∗

A

C

E

D

(b) Fair Gambles

𝑌 𝑌

𝜎𝑚𝑖𝑛 =𝐵 − 𝐻

𝐵 + 𝑊

𝐻

V

DF

(c) Unfair Gambles

Notes: Panel (a) shows indirect utility from income. For income levels Y > Y ∗, people will pay a price Pto consume L with a utility payoff of η. Maximized utility is determined by income endowment and will bethe envelope of the two pieces of the utility function. Panel (b) shows that someone with income level Ywill demand a fair gamble that risks reducing his income by B∗ for a chance to win W ∗ with a likelihood ofwinning, σ. Expected utility from the gamble is at E. Panel (c) shows that there is also demand for unfairgambles with the same loss and win amounts but win likelihood as low as σmin.

38

Figure 2: Biggest Win During Study Period

0

.5

1

1.5

2

2.5

Den

sity

0 2 4 6Biggest Win during Study / Mean Income

Notes: This figure shows the biggest recorded win for each respondent in the full study over the course ofthe nine weeks of participation. 10 people had wins bigger than six times their weekly income.

39

Figure 3: Effect of Winnings on Expenditure Thresholds

0

.02

.04

.06

.08

.1

Reg

Coe

ff on

Win

Am

t / M

ean

Inc

0 .5 1 1.5 2Expenditure / Mean Income Threshold

(a) All - Regression Raw Coefficients

0

.02

.04

.06

.08

.1

Reg

Coe

ff on

Win

Am

t / M

ean

Inc

0 .5 1 1.5 2Expenditure / Mean Income Threshold

(b) Split - Regression Raw Coefficients

0

.1

.2

.3

.4

.5

.6

Res

cale

d R

eg C

oeff

on W

in A

mt /

Mea

n In

c

0 .5 1 1.5 2Expenditure / Mean Income Threshold

(c) All - Rescaled Regression Raw Coefficients

0

.1

.2

.3

.4

.5

.6

Res

cale

d R

eg C

oeff

on W

in A

mt /

Mea

n In

c

0 .5 1 1.5 2Expenditure / Mean Income Threshold

(d) Split - Rescaled Regression Raw Coefficients

Notes: Each graph shows the coefficient estimates for β1 from a regression of Yi,t = β0 + β1Wi,t +∑4m=1

∑3b=1 BetMomentsb

i,m,t + λXi,t + γi + δt + ψs + εi,t where the outcome variable is making a pur-chase above a threshold (indicated on the x-axis) in that time period. The magnitude of the estimate forβ1 is captured on the y-axis. BetMomentsi,m,t are the moments and higher order terms of an individual’sbets for that time period. Xi,t are time-varying individual covariates. I include time, survey round, andindividual fixed effects in all regressions. Standard errors are clustered at the individual level. Panels (a)and (c) are the estimates for all respondents together with the 95% confidence interval dotted around theestimates. Panels (b) and (d) split the sample by people with relatively low and high ability to save. Lowsaving ability is the top solid line in blue in both sub-figures. Confidence intervals are only included forpeople with low saving ability. Panels (a) and (b) show the raw regression coefficients, whereas Panels (c)and (d) re-scale the coefficient by the mean of the outcome variable so that the re-scaled coefficient is theproportion change in likelihood of this purchase.

40

Figure 4: Budgeting Exercise Saving Updates

0

.02

.04

.06

Den

sity

-100 -50 0 50 100Updated - Naive Saving Ability

(a) Raw Saving Update

0

2

4

6

Den

sity

-.6 -.4 -.2 0 .2 .4(Updated - Naive Save Ability) / Mean Income

(b) Saving Update / Mean Income

Notes: This figure shows the update size in perceived saving ability resulting from the budgeting exercise.This is calculated as the amount participants felt they could save in a typical week after the budgetingexercise, minus the amount they estimated naively at the beginning of the survey. Panel (a) is the rawupdate size in thousands of Ugandan Shillings (3,500UGX ≈ 1 USD). Panel (b) converts this update sizerelative to mean income.

41

Figure 5: Effect of Savings Ability Update on Max Ticket Demand - Lowess

-.4

-.2

0

.2

Effe

ct o

f Upd

ate

-.2 -.1 0 .1 .2Saving Ability Update / Weekly Income

CI-95% Linear FitLowess Estimate

Top and Bottom 1% of x-var and updates trimmed. Update = Assisted - Naive

Notes: This figure shows the non-parametric, lowess, estimate of the effect of different saving ability updatesfrom the budgeting exercise on betting demand. The update is the difference between the newly estimatedamount an individual can save minus their original naive estimate, scaled relative to mean income. Themedian negative update was -0.15 and the median positive update was 0.1. The y-axis is the likelihoodof demanding the maximum number of betting tickets offered during the revealed preference measure ofbetting demand following the budgeting activity for people in the treatment group. The background is alinear regression with a linear spline at zero and its associated 95% confidence interval.

42

Figure 6: Separating Saving and Betting Preference by Saving Return Thresholds and Patience

SAVE

BET.6

.7

.8

.9

1

Thre

shol

d G

amm

a

0.2 (0.97) 0.4 (0.98) 0.6 (0.99) 0.8 (0.996) 1 (1)

Discount Factor - Yearly (Weekly) Delta

Notes: This graph shows the threshold level of saving ability needed to sustain a saving strategy for eachlevel of patience, based on calculations from the model developed in Section III and extended in Section VII.People with a given level of future discounting, δ, with a return on saving above the traced locus will bewilling to save in pursuit of a lumpy expenditure. People whose saving ability is below that threshold levelof γ will switch to betting. People will begin switching to betting as saving ability worsens until expectedlosses from saving approach 35%, at which point even the most patient people will switch to betting. Mbitiand Weil (2013) find a reasonable yearly discount factor of 0.64 in neighboring Kenya. This would implya threshold γ of 0.716 at which people facing expected losses above 29% for money set aside to saving willprefer to bet.

43

Tables

Table 1: Summary Statistics: Background and Household

Full Study (N=1,003) Condensed Study, (N=839)Panel (a) Mean p25 p50 p75 Mean p25 p50 p75Weekly Income (USD) 28.70 17.5 24.6 35.1 36.83 22.9 34.3 45.7Betting Expenditures (USD) 3.37 1.3 2.2 4.0 3.91 1.4 2.9 5.7% of Income Spent on Betting 14.29 3.4 8.0 16.5 12.17 4.5 8.3 15.0Household Size 3.09 1.0 3.0 4.0 2.87 1.0 2.0 4.0% Contribution of HH Finances 74.49 50.0 75.0 100.0 75.57 50.0 100.0 100.0Weekly HH Inc. Per Cap. (USD) 14.29 6.4 10.6 17.6 25.43 11.4 20.0 34.3Weekly HH Inc. Per Cap. Adj. (USD) 18.25 8.6 12.5 21.0 NA NA NA NAAge 27.12 23.0 27.0 30.0 26.57 22.0 25.0 30.0Primary 0.83 - - - 0.84 - - -Junior Secondary (O-Level) 0.45 - - - 0.44 - - -Senior Secondary (A-Level) 0.17 - - - 0.19 - - -

Panel (b) Mean p25 p50 p75 Mean p25 p50 p75Available Liquidity (USD) 96.45 14.3 28.6 85.7 99.79 8.6 17.1 57.1Available Liquidity/Mean Inc 2.94 0.5 1.3 3.5 2.34 0.3 0.6 1.5Saving Ability Per Week (USD) 12.42 3.4 8.0 14.3 8.43 2.9 5.7 11.4Saving Ability/Mean Inc 0.50 0.1 0.3 0.6 0.25 0.1 0.2 0.3Win Target (USD) 136.39 24.0 45.9 98.4 353.06 22.9 57.1 200.0Winning Target/Mean Income 5.87 0.9 1.9 4.4 11.37 0.7 1.9 5.0Winning Target/Avail. Liq. 8.86 0.5 1.5 4.9 43.26 1.0 3.0 10.0Winning Item Cost (USD) 361.13 22.9 57.1 171.4 NA NA NA NA

Notes: HH income only calculated for 97% in full and 92% in condensed study who contributed tohousehold expenses. Condensed study did not ask about targeted item for winnings or number ofchildren and adjusted per capita income could not be calculated.

44

Table 2: Lumpy Expenditures and Source of Liquidity

Panel (a): Desired lumpy expenditures by category.Good Business Household Personal#1 Working capital-19% Furniture-17% Clothes-31%#2 Improve worksite-13% Entertainment-17% Phone-11%#3 Motorcycle-13% Build/Repair-9% Vehicle-4%#4 Tools-12% Appliance-5% Entertainment-4%#5 New venture-2% School fees-5% Jewelry-3%

Other 10% 20% 9%None 33% 27% 38%Price $285.6 $114.3 $42.8P rice

Mean Inc 12.9 4.1 1.8

Panel (b): Likely source of liquidity for desired expenditure.Source Most Likely Likely

Saving 85.5% 96.2%Betting 7.1% 24.8%Credit from Family/Friend 2.8% 14.4%Credit from Bank/Loan Organization 4.5% 11.1%Credit from Money Lender 0.1% 2.3%Any Credit Source 7.4% 27.8%

Notes: Panel (a) shows responses to the question “Is there a large expenditure that you are hoping tomake in the next few months?” They were asked to name something in each of the three categories.Interviewers were instructed to ensure that the item or expense named was in fact non-divisible (workingcapital would mean a bulk purchase) and they were additionally instructed to make sure that theseexpenditures were realistic and not simply something they would like to have as a dream. The full setof lumpy expenditures was only asked for participants in the full study. Panel (b) shows the follow-up question conducted during the condensed study, typically following the identification of a businessexpense. Panel (b) suggests that betting is considered the second most likely source of money for theirdesired expense, after saving, and was cited almost as often as all different sources of credit combined.

45

Table 3: Effect of Winnings on Value of Biggest Expenditure

(1) (2) (3) (4) (5) (6) (7) (8)All All All LSA HSA All IHST IHST

Win Amount / Mean Income 0.341 0.250 0.224 0.338 0.080 0.080 0.066 0.016(0.119) (0.085) (0.071) (0.122) (0.091) (0.088) (0.024) (0.033)

LSA (Win Amt/Mean Inc) 0.292 0.100(0.138) (0.046)

Two Week Income 0.040 0.047 0.026 0.090 0.048 0.037 0.037(0.088) (0.089) (0.114) (0.136) (0.089) (0.017) (0.017)

Bet Amt 0.696 1.403 1.677 0.332 1.175 0.264 0.255(0.373) (0.608) (0.677) (0.470) (0.495) (0.081) (0.077)

Mean Y 1.0562 1.0562 1.0562 1.0027 1.1101 1.0562 0.7564 0.7564Fixed Effects No Yes Yes Yes Yes Yes Yes YesBet Moments No No Yes Yes Yes Yes Yes YesStandard Errors Indiv Indiv Indiv Indiv Indiv Indiv Indiv IndivNum Obs 4675 4675 4675 2340 2333 4675 4675 4675Num Inds 957 957 483 474 957 957 957R2 0.025 0.322 0.327 0.358 0.311 0.330 0.470 0.471

Notes: This table shows the un-trimmed regression results of the largest expenditure in a given period on the size of an individual’s biggest win inthat period. All specifications include individual, time, and survey round fixed effects. In columns 1-4 these regressions are unadjusted. Column(1)-(3) include the full population. Column (4) is restricted to people with relatively low saving ability and Column (5) has those with high savingability. Column (6) tests the difference between these two groups. To address concern of outliers driving these results, columns (7) and (8) convert theoutcome using the inverse hyperbolic sine transformation (IHST). The IHST conversion is: IHST (x) = log(x+ (x2 + 1).5). The estimating equationin column (3) is Yi,t = β0 + β1Wi,t +

∑4m=1

∑3b=1 BetMomentsb

i,m,t + λXi,t + γi + δt + ψs + εi,t

46

Table 4: Treatment Assignment Balance Tables

Saving Box Lumpy Good Prime Budgeting ExercisePanel (a) Panel (b) Panel (c) Panel (d)

Wave 1: N=465 Wave 2: N=501 N=1803 N=698Control Treat P-Val Control Treat P-Val Control Treat P-Val Control Treat P-Val

Mean Weekly Inc (USD) 30.988 27.967 0.37 31.335 28.285 0.20 31.315 31.071 0.77 36.851 39.301 0.15Bet Expenditures / Normal Inc 0.171 0.136 0.36 0.126 0.150 0.11 0.122 0.122 0.98 0.123 0.124 0.89Save Ability (USD) 12.309 12.186 0.91 7.257 7.013 0.65 8.573 8.322 0.52 8.327 7.531 0.11Save Ability / Mean Inc 0.495 0.494 0.98 0.311 0.307 0.87 0.325 0.314 0.42 0.248 0.218 0.02Liquidity Available / Mean Inc 3.367 3.311 0.92 2.568 2.579 0.97 3.008 2.418 0.31 2.640 1.441 0.45Lumpy Good Purchased NA NA NA NA NA NA 0.110 0.099 0.46 NA NA NACost of Lumpy Good / Mean Inc NA NA NA NA NA NA 10.529 10.675 0.91 4.487 5.431 0.44Completed Primary 0.820 0.836 0.69 0.837 0.833 0.92 0.834 0.841 0.72 0.842 0.837 0.87Completed O-Level 0.424 0.457 0.53 0.471 0.480 0.83 0.458 0.469 0.63 0.448 0.473 0.56Completed A-Level 0.155 0.129 0.50 0.187 0.194 0.83 0.192 0.189 0.87 0.208 0.167 0.22Math Score 7.850 7.922 0.74 8.292 8.075 0.17 6.732 6.684 0.70 3.501 3.488 0.90Live Alone 0.266 0.267 0.98 0.245 0.266 0.59 0.288 0.279 0.69 0.356 0.360 0.92Household Size 2.974 3.216 0.24 3.206 3.012 0.26 3.085 3.042 0.65 2.943 2.650 0.08Beta Discounting 0.486 0.476 0.74 0.495 0.475 0.47 0.466 0.460 0.73 0.397 0.404 0.80Delta Discounting 0.639 0.630 0.80 0.601 0.566 0.20 0.586 0.577 0.56 0.543 0.549 0.85Risk Aversion. 0=Low 6=High 2.097 2.190 0.65 2.156 2.107 0.75 2.320 2.284 0.69 2.657 2.803 0.44Upside Seeker - Low 0.163 0.164 0.99 0.171 0.171 0.99 0.056 0.064 0.43 0.067 0.059 0.71Upside Seeker - Mid 0.219 0.172 0.28 0.241 0.222 0.61 0.071 0.072 0.90 0.077 0.108 0.18Upside Seeker - High 0.263 0.224 0.40 0.268 0.230 0.32 0.077 0.082 0.69 0.079 0.089 0.67Endline Lumpy Prime 0.241 0.267 0.57 0.440 0.528 0.05 0.00 1.00 NA 1.000 1.000 1.00Treatment - Saving Box 0.000 1.000 NA 0.000 1.000 NA 0.171 0.183 0.50 NA NA NATreatment - Wallet 0.490 0.517 0.61 0.315 0.306 0.82 0.166 0.147 0.28 NA NA NATreatment - Bet Info 0.482 0.500 0.74 0.304 0.310 0.88 0.143 0.162 0.26 NA NA NAN 349 116 252 249 918 885 495 203

Notes: Values marked as “NA” were either not collected for that round or, in the case of the saving box treatment in Panels (a) and (b) was apotential outcome of the treatment. Saving update was only included as part of the budgeting exercise. No alternative treatments are relevantto the budgeting exercise because these participants were exclusively in the condensed study so that they had not been eligible for treatments.The 520 participants in Wave 2 participated in the lumpy prime experiment twice, in the first and final round of the study. Each was giventhe lumpy prime once and kept as a control once and thus are included in both groups.

47

Table 5: Effect of Savings Box Treatment on Demanding Maximum Betting Tickets - Difference in DifferencesEstimation

(1) (2) (3) (4) (5) (6) (7)FE FE FE LSA HSA FE IV

Savings Box (SB) -0.1340 -0.1410 -0.1413 -0.2031 -0.0736 -0.1093 -0.2732(0.0552) (0.0552) (0.0552) (0.0757) (0.0813) (0.0738) (0.1084)

SB x Low Save Ability -0.0633(0.0847)

Log Income -0.0150 -0.0259 -0.0112 -0.0152 -0.0178(0.0277) (0.0355) (0.0438) (0.0278) (0.0280)

Time FEs Yes Yes Yes Yes Yes Yes YesIndividual FEs Yes Yes Yes Yes Yes Yes YesOther Treats No Yes Yes Yes Yes Yes YesMean Dep Var 0.4628 0.4628 0.4628 0.4833 0.4444 0.4628 0.4628Adj Control Mean 0.4984 0.4984 0.4984 0.6264 0.4030 0.4984 0.4984Num Obs 968 968 968 478 486 968 968Num Indivs 484 484 484 239 243 484 484R2 0.6331 0.6376 0.6378 0.6726 0.6140 0.6383 0.6270Adj R2 0.2578 0.2607 0.2596 0.3180 0.1965 0.2590 0.2374

Notes: Results from regression of Bi,t = β0 + β1SaveBoxi,t + λXi,t + γi + δt + εi,t. Dependent variableis an indicator for demanding the maximum number of tickets (4) in the betting ticket offer. SaveBoxi,tis an indicator for an individual having received the saving box in time t. Individual fixed effects, timefixed effects, amount of cash offered, and a set of background controls are included, as well as time-varying covariates, Xi,t. LSA= Low saving ability, HSA= High saving ability. Robust standard errorsare used to adjust for heteroskedasticity in the error term. IV estimation in column (7) uses randomizedassignment to treatment as an instrument for respondents reporting that they have used a saving boxover the preceding four weeks. Standard errors are clustered at the individual level.

48

Table 6: Effect of Lumpy Prime on Demand of Maximum Tickets Offered

(1) (2) (3) (4) (5) (6)All All All LSA HSA All

Lumpy Good Prime 0.066 0.067 0.066 0.110 0.028 0.026(0.024) (0.023) (0.023) (0.034) (0.033) (0.033)

Prime x Low Save Ability 0.084(0.047)

Low Saving Ability -0.062(0.038)

Save Ability / Mean Weekly Inc -0.090 -0.100 -0.107 -0.114(0.043) (0.200) (0.063) (0.056)

Mean Weekly Inc 0.002 0.002 0.001 0.002 0.002(0.001) (0.001) (0.001) (0.001) (0.001)

Mean Dep Var 0.4515 0.4515 0.4515 0.4587 0.4446 0.4515Mean Y-Control 0.4194 0.4194 0.4194 0.4074 0.4300 0.4194Full Set of Covariates No No Yes Yes Yes YesNum Obs 1803 1803 1803 883 920 1803R2 0.0290 0.0325 0.0392 0.0614 0.0508 0.0412Adj R2 0.0164 0.0194 0.0240 0.0306 0.0209 0.0250

Notes: Results from regression of Bi = β0 + β1LumpyPrimei + λXi + εi. Dependent variable is anindicator for demanding the maximum number of tickets in the betting ticket offer. LumpyPrime isan indicator for going through the lumpy prime dialog prior to the ticket offer. All regressions controlfor status of other treatments in the study and the amount of cash offered instead of tickets. Columns(3)-(6) also control for the price of the desired expenditure as well as whether it was purchased sincethe baseline for respondents in the full study. LSA= Low saving ability, HSA= High saving ability.Robust standard errors are used to adjust for heteroskedasticity in the error term. Columns (1)-(4)show stability of the estimated treatment effect regardless of specification. Columns (5)-(7) showheterogeneity of response between people with low and high saving ability.

49

Table 7: Effect of Budgeting Exercise on Demanding Maximum Betting Tickets

(1) (2) (3) (4) (5) (6)UGX UGX Prop Both UGX Prop

Lumpy Good Prime 0.090 0.090 0.088 0.088 0.089 0.090(0.043) (0.043) (0.043) (0.043) (0.043) (0.043)

Budgeting Exercise (BE) -0.028 -0.006 -0.055 0.084 -0.031 -0.009(0.045) (0.047) (0.046) (0.081) (0.055) (0.058)

BE x Update -0.004 -0.479(0.002) (0.236)

Update 0.000 0.024(0.001) (0.141)

BE x (Update > 0) -0.282(0.104)

BE x (Update < 0) -0.068(0.099)

Update > 0 0.090(0.060)

Update < 0 0.034(0.054)

BE x Positive Update Amount -0.004 -1.149(0.002) (0.488)

BE x Negative Update Amount 0.003 0.188(0.003) (0.342)

Positive Update Amount 0.001 0.210(0.001) (0.221)

Negative Update Amount -0.000 0.080(0.002) (0.209)

N 688 688 688 688 688 688Mean Dep Var 0.4142 0.4142 0.4142 0.4142 0.4142 0.4142Control Group Mean 0.3704 0.3704 0.3704 0.3704 0.3704 0.3704Full Set of Covariates Yes Yes Yes Yes Yes YesR2 0.1334 0.1377 0.1388 0.1437 0.1394 0.1411Adj R2 0.1021 0.1038 0.1049 0.1073 0.1028 0.1046Prop columns scaled to respondent income, UGX in 1,000s. Update = Assisted - Naive Estimate

Notes: Results from regression of Bi = β0 + β1LumpyPrimei + β2Budgeti + β3(Budget×Update)i +β4Updatei + λXi + εi. Dependent variable is an indicator for demanding the maximum number oftickets in the betting ticket offer. LumpyPrimei is an indicator for doing the lumpy prime dialogbefore the betting ticket offer. Budgeti is an indicator for doing the budgeting exercise before thebetting ticket offer. Updatei is the assisted estimate of the amount that an individual can save fromthe budgeting exercise minus the naive estimate. UGX columns use the raw measure of the updatein thousands of shillings. Prop columns rescale this update relative to an individual’s mean income.Individual covariates include background education, household characteristics, and preference variablesas well as controls for other treatments during the study and the amount of cash offered instead oftickets. Robust standard errors are used to adjust for heteroskedasticity in the error term.

50

Table 8: Estimating Return on Cash Saving, γ

γ Source Estimateγ1 Inflation 0.8998 - 0.9844γ2 Temptation 0.784 - 0.9361γ3 Social Pressure 0.92 - 1γ4 Theft/Loss 0.9885 - 1γ5 Transaction Costs 1

Notes: γ is an individual’s return on saving. This can be thought of as the portion of each dollar setaside for saving that he expects to actually be converted into an eventual expenditure on his desiredsaving target. γ is divided into sub-components in order to create a reasonable estimate of valuesfor this population. γ is assumed to be the product of these sub-components. These components areapproximated for cash savings as a benchmark, acknowledging that estimates would likely differ whenusing different saving instruments or strategies.

• γ1 is based on the range of inflation rates in Uganda over the five years leading up to the study,2011-2016.

• γ2 is estimated from the consumption modules in the survey, categorizing certain expenditures astemptation goods (alcohol, video hall tickets, betting) and assuming that people regret between25-50% of these expenses.

• γ3 captures expenditures that are made out of obligation or as a result of inter- or intra- householdpressure. This estimate is from a recent study by Jakiela and Ozier (2015).

• γ4 was based on survey responses estimating the frequency of theft.

• γ5 is assumed to be one for cash savings, but would be lower for most other saving technologiesthat require usage fees, coordination with others, or effort for either deposits or withdrawals.

These estimates imply a reasonable range for γ between 0.6415 and 0.9215. This range of γs fallsconsiderably below the range of threshold γ∗s illustrated in Figure 6, suggesting that there may be aconsiderable portion of the population for whom betting is a rationally preferred strategy to saving,given their levels of patience and return on saving.

51