Addictive consumption, imperfect substitutes and self control: A model and an application to slot machines

Deiana, Claudio; Dragone, Davide; Giua, Ludovica Working Paper Addictive consumption, imperfect substitutes and self control: A model and an application to slot machines Quaderni - Working Paper DSE, No. 1197 Provided in Cooperation with: University of Bologna, Department of Economics Suggested Citation: Deiana, Claudio; Dragone, Davide; Giua, Ludovica (2024) : Addictive consumption, imperfect substitutes and self control: A model and an application to slot machines, Quaderni - Working Paper DSE, No. 1197, Alma Mater Studiorum - Università di Bologna, Dipartimento di Scienze Economiche (DSE), Bologna, https://doi.org/10.6092/unibo/amsacta/8037 This Version is available at: https://hdl.handle.net/10419/306818 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc/4.0/ ISSN 2282-6483 Addictive Consumption, Imperfect Substitutes and Self Control: A Model and an Application to Slot Machines Claudio Deiana Davide Dragone Ludovica Giua Quaderni - Working Paper DSE N°1197 Addictive Consumption, Imperfect Substitutes and Self Control: A Model and an Application to Slot Machines * Claudio Deiana  Universit`a di Cagliari, CRENoS and IZA Davide Dragone  Universit`a di Bologna Ludovica Giua § Universit`a di Cagliari and CRENoS November 20, 2024 Abstract We propose a model of addictive consumption to study the demand for imperfect substitutes involving substances like alcohol, nicotine and opioids, as well as behavioral addictions like gambling and digital addiction. We study a 2017 Italian policy aimed at reducing gambling by limiting the number of available slot machines. Despite the reduction in slot machines, the policy produced an unintended 25% increase in net expenditure, particularly among low-wealth and low-educated individuals who also engage in other addictive behaviors. This result can be rationalized as the consequence of changes in self-control costs due to social contagion effects. Keywords: Addiction; Gambling; Horizontal differentiation; Self-control; Slot machines, Temptation. JEL codes: I18, L43, L83. * We thank Paolo Buonanno, Elias Carroni, Rocco d’Este, Marco Francesconi, Giovanni Mastrobuoni, Andrea Moro, Marco Nieddu, Matthias Parey, Giacomo Pasini, Giulio Zanella and participants at the SIEP Conference (Verona, 2023), the HERB workshop (Bologna, 2023), the EuHEA Conference (Vienna, 2024), the BHEPPE seminar (Bologna, 2023) for comments and suggestions. Any errors are the fault of the authors only. The authors declare no conflict of interest.  Universit`a di Cagliari, CRENoS and IZA, Department of Economics and Business, via S. Ignazio da Laconi 17, 09123 Cagliari (Italy); email: [email protected].  Universit`a di Bologna, Department of Economics, Piazza Scaravilli 2, 40126, Bologna, Italy; e-mail: [email protected]. § Corresponding author. Universit`a di Cagliari and CRENoS, Department of Economics and Business, via S. Ignazio da Laconi 17, 09123 Cagliari (Italy); email: ludo[email protected]. 1 Non-technical summary Addictive behaviors, including substance use disorders and behavioral addictions, pose significant challenges to policymakers due to their adverse effects on physical and mental health, interpersonal relationships, economic stability, and quality of life. Governments worldwide have implemented strategies such as taxation, prohibition, educational programs, and restrictions on access to mitigate these harms. These interventions often rely on the economic principle that increasing the cost of consumption reduces demand. However, the outcomes can deviate from expectations when such policies inadvertently alter self-control and temptation costs. This paper examines the consumption of addictive goods that are imperfect substitutes, combining theoretical modeling and empirical analysis. It integrates concepts from horizontal differentiation, rational addiction, and temptation costs to predict how addiction levels, availability of alternatives, and self-control constraints influence behavior. The theoretical framework addresses a broad range of addictive goods, including substances like nicotine, alcohol, and opioids, as well as behavioral addictions such as gambling and gaming disorders. Although behavioral addictions are less damaging to physical health than substance use disorders, they can significantly impair functioning across personal and professional domains. The empirical application focuses on gambling addiction, specifically assessing the effects of a 2017 Italian policy aimed at reducing gambling opportunities by limiting the availability of slot machines. Using municipality-level administrative data and household expenditure surveys, the study evaluates the policy’s impact through a difference-in-differences approach. The findings show that the number of slot machines declined by 21% and gambling venues by 11%. However, contrary to the policy’s intent, slot machine expenditures increased by 25%, driven largely by low-income and less-educated individuals. These unintended consequences indicate that the policy may have exacerbated harm for the very groups it sought to protect. Under standard economic assumptions, reduced access should discourage gambling by increasing transportation and opportunity costs. However, the findings suggest that reduced access also led to increased waiting times, which amplified cravings and temptation costs, especially among more addicted individuals. This dynamic likely increased gambling intensity, underscoring the importance of psychological factors in driving addictive behaviors 2 1 Introduction Substance use disorders and behavioral addictions raise policy concerns due to their impact on individual physical and mental health, personal relationships, economic and financial stability, and overall quality of life. To address these concerns, policymakers have adopted a range of strategies, including prohibition and taxation of addictive substances, education programs, age limits, and time curfews (see, e.g., Evans et al.,1999;Adda and Cornaglia,2006;Carpenter and Dobkin,2009;Anger et al.,2011;Cawley and Ruhm,2012;Hansen,2015;Ahammer et al., 2022;Moore and Morris,2024). The rationale underlying these forms of regulation is that consumption, even when addictive, is expected to decrease when its cost increases (Becker and Murphy,1988). However, when the policy intervention also affects self-control and temptation costs, this prediction may not hold. In this paper, we propose a theoretical and empirical investigation into the consumption of goods that are imperfect substitutes, addictive, and induce self-control costs. Building on the Hotelling (1929)’s model of horizontal differentiation, the Becker and Murphy (1988)’s theory of rational addiction, and the Gul and Pesendorfer (2001)’s model of temptation and self-control, we derive predictions about the extensive and intensive margins of addictive consumption as functions of the level of addiction, characteristics and availability of different addictive options, temptation costs, and access costs. The theoretical model provides a conceptual framework to understand consumption related to addictive substances such as, e.g., nicotine (Pesko and Warman,2022), alcohol (Calcott,2019) and opioids (Agrawal et al.,2023), and for behavioral addictions such as gambling and gaming disorders (WHO,2022).1 As an empirical application, we focus on gambling addiction. Specifically, we estimate the effects of a 2017 policy in Italy that mandated a large reduction in the number of available slot machines. Using novel Italian administrative data on various types of gambling activities at the municipality level and a representative survey data on household expenditure, we estimate the effects of the policy with a standard difference-in-differences approach. We find that in treated municipalities the number of slot machines decreases by 21%, and the number of venues hosting them decreases by 11%. However, the policy leads to unintended consequences, with net slot machine expenditure rising by 25% (30.20 EUR). This increase is driven by higher spending among players with low wealth and low education who also engage 1Despite being non-substance-related, and thus generally less harmful to health, behavioral addictions can also yield adverse consequences, impairing an individual’s ability to function across various life domains (Alavi et al.,2012;WHO,2018,2024). The ICD11 –the WHO (2022)’s latest International Classification of Diseases– describes gambling and gaming as disorders due to addictive behaviors that develop as a result of repetitive, rewarding activities. Other disorders, such as shopping disorder, exercise addiction (Berczik et al.,2012), and digital addictions related to internet use, including smartphone addiction (Allcott et al.,2022), are not included in the ICD-11. See Petry (2015) for a comprehensive overview. 3 in other addictive behaviors. Since slot machine players are predominantly from disadvantaged backgrounds (Mastrobattista et al.,2019;Resce et al.,2019), our results imply that the policy has harmed the very population it aimed to protect. The theoretical model allows to gain insights on the possible mechanisms driving this finding. In the absence of self-control and temptation costs, the reduction in slot machines should increase the transportation costs and reduce willingness to play because of, e.g., increased congestion and longer waiting times. This would unambiguously reduce the extensive and intensive margin of slot machines play. The observed evidence suggests that, on the contrary, temptation costs do play a role because waiting times can also increase cravings for playing, and the more so in more addicted individuals. When these cravings and the associated temptation costs are large enough, the intensive margin of playing and, thus total expenditure on slot machines, can eventually increase. Theoretically, our paper bridges three distinct streams of literature: addiction, horizontal product differentiation and self-control. The literature on addiction is typically focused on individual choices related to an addictive good that features habit-formation and self-control problems. The former is based on the assumption that current preferences depend on past consumption experiences (Pollak,1970), an idea that is typically formalized as accumulation of an addiction stock which creates tolerance and reinforcement (Becker and Murphy,1988). The latter has been studied as the consequence of regret (Orphanides and Zervos,1995), timeinconsistent behavior (Strotz,1955;Laibson,1997;Gruber and K¨oszegi,2001;Piccoli and Tiezzi,2021), costly temptation (Gul and Pesendorfer,2001,2007), projection bias (Loewenstein et al.,2003), dual-self models (Thaler and Shefrin,1981;Loewenstein and O’Donoghue, 2004;Ozdenoren et al.,2012), environmental cues (Bernheim and Rangel,2004), or a combinations of these factors (Allcott et al.,2022). Most existing contributions, however, focus on a single addictive good.2In this paper we extend the analysis to allow for imperfect substitutes and horizontal product differentiation `a la Hotelling (1929). This allows to take into account that people are heterogeneous, and they can satisfy their needs through goods that are imperfect substitutes, including the option of abstaining if the addictive alternatives are not good enough. Empirically, we contribute to understanding the effects of a policy aimed at regulating addictive behavior by restricting the availability to consumption opportunities. Specifically, we add to the economics literature on gambling behavior. By providing the first causal analysis of a policy aimed at reducing gambling access, our results complement those of Badji et al. (2023) and Baker et al. (2024), who analyze the effects of the opposite type of regulation—greater 2For exceptions see, e.g., Dockner and Feichtinger (1993); Bask and Melkersson (2004); Cawley and Dragone (2024). 4 gambling accessibility. Additionally, we contribute to empirical studies on gambling addiction and the substitutability between different types of gambling (Kearney,2005;Guryan and Kearney,2010). Finally, by emphasizing the possible roles of temptation and social contagion in gambling choices, we contribute to understanding behavioral drivers of gambling (Stetzka and Winter,2023), such as the “lucky store” effect (Guryan and Kearney,2008), the impact of unmet liquidity needs and financial constraints on gambling behavior (Herskowitz,2021), and the role of media exposure (De Paola and Scoppa,2014). The paper is organized as follows. In the next Section we present a model of addictive consumption with imperfect substitutes and self-control costs. In Section 3we describe the empirical application of the model to the case of the gambling industry in Italy. Section 4describes the empirical strategy and Section 5reports the empirical results. Section 6concludes. 2 Becker and Murphy meet Hotelling, Gul and Pesendorfer 2.1 A model of addictive consumption with imperfect substitutes and selfcontrol costs Consider a scenario where addictive consumption can occur at two venues v, denoted as A and Band located at the extremes of a [0,1] segment. Individuals are distributed along the [0,1] segment and pay a marginal transportation cost τv≥0 to reach venue v∈ {A, B}and consume. Once in a venue, individuals allocate their budget mbetween sunits of addictive consumption and a composite numeraire z. The cost of each addictive unit consumed at venue vis pv. In each period t, individuals first choose whether to reach a venue, then they choose the optimal amount of addictive consumption. The former choice informs about the extensive margin of addictive consumption, while the latter one describes the intensive margin. Addictive consumption contributes to building an individual level of addiction a≥0. Addiction evolves over time depending on past and current consumption, according to a(t+ 1) = γ(s(t) + a(t)) ,(1) where parameter γ∈[0,1) and a(0) = a0. After playing, individuals return to their initial location (“home”), the addiction level changes and a new period begins. Consumption of the composite good yields utility Z(z), with Zz>0,Zzz ≤0. If an individual attends a venue v, consuming sat venue v(denoted as sv) yields, for a given level of addiction a, the following utility: U(sv;a) (2) 5 The utility function (2) is strictly increasing in addictive consumption (Us>0). Consistent with Becker and Murphy (1988) and Becker et al. (1991), we assume that given levels of playing are less satisfying when past consumption has been greater (Ua<0) and that the more a person has played in the past, the more they like playing today (Usa >0). These two assumptions describe tolerance and reinforcement, which are typical properties of addictions.3 Utility from addictive consumption depends on the characteristics of the specific venue. These characteristics can literally describe features of the venue where addictive consumption occurs, such as the possible presence of amenities, the quality of the service, the presence of other consumers, but they can also be interpreted more broadly as features of the two available consumption goods. With the latter interpretation sAand sBdescribe imperfect substitutes for addictive consumption for which consumers have different tastes, conditional on consuming. This is the case of, e.g., vaping as an imperfect substitute for smoking, beer for wine, or heroin for other opioids. Addictive consumption typically features temptation and self-control costs. Specifically, we assume that addicted individuals are tempted to spend all available budget mon addictive consumption, disregarding the composite good z. In the spirit of Gul and Pesendorfer (2001, 2004), individuals can partially overrule such temptation, but at a cost. Specifically, an individual consuming svpays a temptation cost that depends on the distance between the tempting choice of spending all budget on the addictive good and the actual choice. For concreteness, we consider the following cost function (Allcott et al.,2022;Cawley and Dragone,2024): C(sv;a, σv)≡a·σv·(m−sv)≥0 (3) Expression (3) depends on ato account for the possibility that the temptation cost is higher when addiction is higher, and nil in case of no addiction, Parameter σv≥0 describes the marginal temptation cost at venue v, and it can depend on factors such as previous exertion of self-control (Muraven and Baumeister,2000;Loewenstein and O’Donoghue,2004), cognitive load (Shiv and Fedorikhin,1999), as well as environmental cues (Loewenstein,1996,2000; Bernheim and Rangel,2004), waiting times (Houser et al.,2018,2021), social contagion and peer-pressure (Lundborg,2006;Clark and Loh´eac,2007). Taking into account the temptation costs, the individual objective function is: U(sv, z;a, σv)≡ U (sv;a) + Z(z)− C (sv;a, σv) =U(sv;a) + Z(z) + aσv(sv−m) (4) 3The utility function is also assumed to be strictly concave. Note that partial derivatives of the utility function are denoted with subscripts, as in, e.g., Usa ≡∂2U(·) ∂s∂a . 6 It can easily be verified that the objective function (4) is increasing in consumption (Us>0), features reinforcement and tolerance (Usa >0, Ua<0), and that it decreases when the temptation parameter is higher (Uσ≤0). Individuals that do not attend either venue spend all budget on the composite good and obtain the reservation utility U(0) ≡U(0, m;a, σ0), which negatively depends on addiction a and the temptation parameter σ0. 2.2 Solving the model The problem is solved by backward induction. In the second stage of each period t, and conditional on being at venue v∈ {A, B}, an individual with addiction stock aoptimally chooses addictive consumption sand the amount of consumption of the composite good zthat solve:4 max sv,z U(sv, z;a, σv) (5) s.t. m=pvsv+z(6) The optimal amount of addictive consumption (s∗ v, z∗ v) exhausts the available budget and, assuming an interior solution, it satisfies the familiar condition where the marginal rate of substitution between sand zequals the relative price: (s∗ v, z∗ v) : Us(s∗ v, z∗ v) Uz(s∗ v, z∗ v)=pv(7) The participation choice in the first stage depends on the individual location ialong the unit line, and on the individual addiction level a. The former affects the transportation cost to reach the venues, the latter affects the utility from consumption and the marginal incentives to consume. An individual chooses to reach a specific venue, or the outside option of no addictive consumption, by comparing the reservation value U(0) and the maximized utility of consuming at either venue, i.e.: V(i, a, A)≡U(s∗ A, z∗ A;a, σA)−τA·i, (8) V(i, a, B)≡U(s∗ B, z∗ B;a, σB)−τB·(1 −i) (9) 4Unlike rational addiction models, which assume individuals account for how current choices impact future addiction, here we assume that individuals are myopic. Readers interested in forward-looking addiction models without horizontal product differentiation can refer to Becker and Murphy (1988), Chaloupka (1991), and Dragone and Raggi (2021) for time-consistent behavior, and to Piccoli and Tiezzi (2021) and Allcott et al. (2022) for models with time-inconsistent agents. 7 this finding is suggestive, it does not allow for causal conclusions about the effectiveness of the 2017 bill. This analysis will be conducted in Sections 4and 5. 3.2 Data sources We use two main datasets: administrative data on gambling at municipality level and survey data on expenditure at the household level. For additional insights on the empirical results, we use the ADM geolocalized data on the universe of tobacco shops in 2016 and the Italian Census 2011 data on socio-demographic characteristics. Gambling data The gambling data are sourced from ADM. They span from 2015 to 2019, they cover 7,877 Italian municipalities, and they contain the universe of legal gambling activities in Italy. ADM reports municipality-level yearly data on the counts of slot machines (both Newslots and Videolotteries), the counts of licensed venues where slot machines are installed, and the total expenditure net of winnings on various forms of gambling. Table 1presents the main summary statistics on the 2015 per capita net expenditure at the municipality level, categorized by gambling type: slot machines, sports bets, lotteries, bingo, scratch cards, and online betting. Slot machines account for the highest per capita expenditure, averaging EUR 121.35. This figure is largely driven by expenditure on Newslots, which alone accounts for EUR 101.83. Expenditure on other types of gambling is substantially lower. For instance, lotteries, which constitute the second-highest category, collect an expenditure that is approximately half of that spent on slot machines. Online betting collects only EUR 0.88 on average in 2015. It is noteworthy that, although this figure nearly doubles to EUR 1.42 by Mean Std. dev. Min Max Slot Machines 121.35 345.64 -1.00 28,416.85 — Newslots 101.83 218.18 -1.00 17,992.04 — VLT 19.51 140.43 0.00 10,424.82 Lotteries 60.77 213.05 -4,359.54 17,614.33 Scratch cards 7.01 41.86 -1,022.57 3,404.31 Sport bets 4.09 30.49 -22.65 2,575.94 Bingo 1.24 15.89 -70.48 625.04 Online 0.88 6.20 -1.26 491.62 Table 1: Per capita net expenditure in 2015, by type of gambling. Municipality-level per capita expenditure, net of winnings, in the year 2015. Sample: 7,877 municipalities. 14 (a) Newslots per capita (b) Newslots net expenditure per capita Figure 4: Newslots: per capita number and net expenditure in 2015. Geographical distribution of the per capita number of Newslots (panel a) and the per capita net expenditure on Newslots (panel b) by municipality in 2015. Black lines indicate regional borders (20 NUTS-2 areas). 2019, its overall importance in the Italian market during this period remains relatively small. The geographical distribution of per capita Newslots is positively correlated with per capita net expenditure, as shown in Figure 4for 2015.12 In line with the 35% reduction mandated by the 2017 bill, between 2016 and 2018 the number of Newslots decreases by about one third in each region. This is accompanied by a decrease in the number of venues (see Figure B.1 in Appendix B).13 The number of Videolotteries and their corresponding venues, which are not subject to the policy, remains substantially constant over time or slightly increased. Household expenditure data To examine the heterogeneous effects of addiction levels and socio-economic status on gambling expenditure, we use the Italian Household Expenditure Survey (HES). Published by the Italian National Institute of Statistics (ISTAT), the survey is representative of the Italian population and covers a broad range of household expenses, including gambling and alcohol consumption, as well as demographic characteristics of household members and self-reported economic conditions. We analyze data from the 2014 to 2019 survey waves, with an overall sample of over 100,000 households. The data includes total expenditure, and specific spending details on 12 The pairwise correlation is equal to 0.25 and is statistically significant at any conventional level. 13 Piedmont and Valle d’Aosta are an exception as a consequence of more stringent local policies. As shown in the robustness checks section, this does not drive the empirical results. 15 gambling and on alcoholic beverages, such as beer, cider, wine, liquors, and spirits. It covers regional information, household composition (number of members), and detailed characteristics of the household head, including gender, age, nationality, education level, employment status, occupation, economic sector, and type of employment contract. On average, total household monthly expenditure is around EUR 2,500. Households who report no expenditure on gambling amount to 87% of the sample. Figure B.2 shows the distribution of expenditure on gambling for the remaining 13%. For this subsample, average reported expenditure on gambling is EUR 23.77. 4 Empirical strategy 4.1 Conceptual framework We empirically assess the causal effects of the 2017 bill using a standard difference-in-differences approach. The predictions from the theoretical model presented in Section 2provide a natural framework for interpreting the empirical estimates. In fact, slot machine play can be entertaining (Conlisk,1993;Burger et al.,2020), although it may lead to negative consequences, such as adverse effects on mental health, including depression, anxiety, and stress (Muggleton et al.,2021;Wardle and McManus,2021;Badji et al.,2023), as well as loss of money and financial hardship. Furthermore, slot machines are designed to create addiction (Sch¨ull,2012), particularly due to visual and sensory features that encourage reinforcement through repeat play (Harrigan et al.,2010;James et al.,2016). Denoting the level of gambling addiction as a, and the amount of slot machine play–the number of spins– as s, the requirements for the utility function used in the Section 2are satisfied (Us>0, Ua<0, and Usa >0). Moreover, playing slot machines requires reaching the physical venue (v) where they are located, at a transportation cost (τv).14 Once at the venue, individuals choose how much to play (sv). The net expected cost of playing is p≡P−E(w)>0, where Pis the cost per play (by law, it is equal across venues) and E(w) is the expected monetary win.15 Denote venue Aas a treated unit, and venue Bas the control unit. Empirically, we aim at 14 Slot machines are imperfect substitutes because they are located in different venues. In this sense, the horizontal differentiation component of the model is literally understood in terms of location. However, transportation costs can also be interpreted as access or stigma costs for being a player. We will consider this alternative interpretation when discussing the empirical results in Section 5. 15 Taking literally the notion of playing, our approach is complementary to the approach based on riskloving preferences over monetary outcomes, which explains why people choose uncertain prospects even when the net expected gain is negative. See Friedman and Savage (1948); Hartley and Farrell (2002); Levitt (2004) for an alternative approach focused on the role of uncertainty in gambling choices. For an approach aimed at describing optimal stopping time for playing, based on cumulative earnings, see Lien and Zheng (2015). 16 estimating the following difference-in-differences (DiD) estimator: β=E′ A−EA−E′ B−EB= ∆EA−∆EB(11) where E′ vdenotes net expenditure at venue vafter the implementation of the policy, and ∆EA and ∆EBare the differences in net expenditure at venue Aand B, respectively. Using the terminology of the theoretical model, we conjecture that the reduction in the number of available slot machines produces different effects. First, a reduction in the number of available slot machines is likely to increase the transportation cost τAto reach the treated venue A. Second, the policy may reduce the utility derived from gambling at A, due to factors such as congestion or longer waiting times, which may make slot machine play at A less enjoyable. Both mechanisms imply that the expected effect of the bill is to reduce the extensive and intensive margins of slot machine play at A(hence ∆EA<0) and to increase the number of players at B(hence ∆EB>0). This is consistent with Badji et al. (2023), who show that people residing in close proximity to gambling venues are more likely to gamble and less likely to be happy. Third, it is possible that the bill influences the marginal temptation cost σAof playing at A, due to, e.g., longer waiting times that increase the desire to consume (Loewenstein,1987; Houser et al.,2018,2021) or the cost of exerting self-control (Muraven and Baumeister,2000; Vohs and Faber,2007;Hagger et al.,2010;Baumeister et al.,2018), or possible social contagion effects induced by higher concentration of players in the same venue (Rockloff and Dyer,2007; Rockloff et al.,2011,2012,2017;Hopfgartner et al.,2021). Since the model predicts that greater temptation leads to higher consumption (but also less consumers) at A, its effect on net expenditure can oppose those produced by increased transportation costs and reduced utility.16 The sign of the empirical estimate of βcan suggest which of the mechanisms above dominates. In the absence of changes in the temptation parameter σA, an increase in transportation costs and a reduction in the utility from playing is predicted to unambiguously decrease expenditure at A, while expenditure at Bshould increase (see Propositions 1and 2). In such a case, the difference-in-difference estimate βis predicted to be negative. If, however, the policy also produces an increase in the temptation parameter σA, or a reduction in the stigma cost associated to being a player, net expenditure at venue Amay increase (hence ∆EA>0). If 16 In principle, longer waiting times and increased crowding can reduce the utility of playing, but they may also raise temptation costs. Furthermore, a higher concentration of players in fewer venues could reduce the perceived stigma associated with being a player, which might be interpreted as a reduction in “transportation costs” (if τAis understood in a non-literal sense), leading to more players at venue A. We do not take an a-priori stance on which of these potential channels dominates. 17 such effect is large enough, βis positive (see Equation 11). 4.2 Treated units The 2017 bill required Newslots to be reduced in areas where they were more abundant and least profitable. Hence, we classify municipalities as treated units (the empirical counterparts of venue Ain the model) if two criteria are satisfied: (i) the number of slot machines per capita at the municipality level is above the national average as of December 2016, and (ii) total revenue per device at municipal level is below the 15th percentile of the regional distribution.17 Panel (a) of Figure 5illustrates the treatment assignment, using the Emilia-Romagna region as an example. Municipalities are first categorized based on whether the per capita number of Newslots in the municipality exceeds or falls below the national average. This corresponds to criterion (i) and is graphically represented by the red and pink curves in panel (a). Within each group, we identify municipalities where the average revenue per device in 2016 is below the 15th percentile This corresponds to criterion (ii) and by the municipalities to the left of the vertical line in panel (a) of Figure 5. Therefore, the treated municipalities are those that have a higher-than-average number of Newslots per capita (red curve) and lower-than-average (a) Municipalities by revenue per device, Emilia Romagna 0 5.000e-06 .00001 .000015 .00002 .000025 Density 0 50000 100000 150000 Revenue per device at municipal level Newslots per capita < national average Newslots per capita > national average (b) Treated municipalities, Italy Figure 5: Assignment of treated and control units. Panel (a) shows the distributions of municipalities based on the revenue per device in the Emilia-Romagna region in 2016. The pink and red curves correspond to the municipalities with a number of Newslots per capita below and above the national average, respectively. The vertical line indicates the 15th percentile of the regional distribution of the revenue per device. The dashed red curve denotes the treated units. Panel (b) shows the geographical distribution of the treated municipalities (221, in red). Black lines indicate regional borders (20 NUTS-2 areas). 17 The 15% threshold is motivated by the 15% reduction in slot machines mandated in 2017. The empirical results are robust to using different thresholds (see Appendix C). 18 revenues per device (left of the vertical line). These treated units are indicated by the dashed red curve. This procedure is repeated for each of the 20 Italian regions. Panel (b) displays the geographical distribution of the treated municipalities (221, 3% of the sample) across the 107 provinces and the 20 regions of Italy. Figure B.3 illustrates the time variation in the per capita number of Newslots and the per capita net expenditure on total slot machines across treated and control municipalities. Panel (a) shows the compliance of the treated municipalities with the 2017 bill. Before 2017, trends are parallel across treated and control units; afterwards, the gap narrows. Panel (b) displays the evolution of per capita net expenditure. Expenditure increases over time in treated units, while it remains substantially constant in control municipalities. 4.3 Empirical model We estimate the following empirical model: Ympt =α+βCm× 1 (t≥2017) + γmp +δtp +ϵmpt (12) where Ympt is the outcome of interest for municipality min province pand year t. Specifically, we consider the per capita number of Newslots, the per capita expenditure on slot machines, the per capita number of venues hosting Newslots, the per capita number of Videolotteries (i.e. slot machines not affected by the cut), and the expenditure on other types of gambling (scratch cards, lotteries, sports bets, and online bets). The coefficient of interest is β, as it captures the causal effect of reducing the number of slot machines (Cm) on the outcome. In the theoretical model, this corresponds to Equation 11. To account for time-varying heterogeneities at sub-regional level, the empirical model includes municipality fixed effects (γmp) and a set of year-province dummy variables (δtp). Regressions are population-weighted and standard errors are clustered at municipality level. To test for the possible existence of differentials between treatment and control groups in the pre-policy period, and to rule out that the reduction in Newslots is endogenously related to pre-treatment differentials in the outcomes, we consider the following event-study specification: Ympt =α+ 2019 X j=2015 βjCm× 1 [t=j] + γmp +δtp +ϵmpt (13) with year 2016 as the baseline. We also use this specification to examine potential heterogeneous effects based on the distance to tobacco shops, which serves as a proxy for the transportation costs discussed in Proposition 1. 19 To further support the causal interpretation of our results, we run additional tests as robustness checks. First, we perform sensitivity analyses on the definitions of treated and control groups. Second, we re-estimate the model considering different sub-samples. Third, we conduct several falsification exercises, including exercises focusing on unaffected outcomes and treatment randomization. The results are reported in Appendix C. Finally, we use household-level data from the Household Expenditure Survey (HES) to investigate the possible correlation between gambling and risky health behaviors, and to understand what parts of the population are most affected by the policy. This analysis is policyrelevant because the literature on risky health behaviors has consistently shown that individuals who engage in one risky behavior, such as drinking, are more likely to engage in others, such as smoking, substance use, and gambling (Cawley and Ruhm,2012). Moreover, Resce et al. (2019) note that in Italy, slot machines are predominantly used by individuals of lower socio-economic status, raising concerns about income-related inequalities in gambling, particularly among the most vulnerable populations. We estimate the following regression with Pseudo-Poisson Maximum Likelihood (PPML): Whrp = exp (ϑ+ϕThr × 1 (p≥2017) ×Ahrp +ψp+ζr+Xhrpη) + νhp (14) where Whrp is the monthly household expenditure of household hin region rand period p (month of the year in 2014m1-2019m12). We consider reported expenditure on gambling and total expenditure. Since the HES dataset only provides geographical information at the regional level, without specifying municipalities, we must adapt the identification strategy to the regional scale. Households are defined as treated (Thr = 1) if they reside in a region that is relatively more exposed to the reduction in Newslots in 2017 —that is, in a region where the share of the population living in treated municipalities is above the median of the distribution. Variable Airp identifies households who are in the top quartile of the distribution of expenditure in alcohol (beer, cider, wine, liquors and spirits). This proxies for individuals with high levels of addiction to alcohol and, possibly, of addiction in general. With this interpretation, the coefficient ϕcaptures the causal effect of the reduction in Newslots on monthly expenditure for households with higher addiction levels. The empirical model in Equation 14 includes period (ψp) and region fixed effects (ζr) to capture heterogeneities over time and across regions. Xhrp describes a set of household characteristics: number of components, household head’s gender, age, nationality, education, employment status, occupation, economic sector and type of contract.18 18 The categories for employment status are: employed, unemployed, homemaker/student, retired, or other; 20 5 Empirical results 5.1 Effects of the policy: Administrative data Table 2displays our main results, estimated using Equation 12. As shown in column 1, compliance with the bill was high, as the number of Newslots per capita in treated municipalities decreased by 21% after 2017. Column 2 of Table 2shows that also the number of venues hosting Newslots declined, by approximately 11%. The number of Newslots per venue decreased by a similar amount (column 3). The corresponding event-study analyses are illustrated in Figure 6. Panel (a) shows that the bill reduced the per capita number of Newslots by 0.0005 in 2017 and by to 0.0015 in 2018 and 2019. Given the 2015 average of 0.0054, this means that the reduction in the per capita number of Newslot devices reached 10% by the end of 2017 and 31% by the end of 2019. Panel (b) and (c) show that the reduction in the per capita number of venues was 4% in 2017 and 17% in 2019, while the number of Newslots per venue decreases by 7% in 2017 and by 14% in 2018 and 2019. Together, these results suggest that the policy likely increased transportation costs for players and led to higher crowding in the remaining venues. Contrary to the goal of the policy, per capita expenditure on slot machines significantly increases by 25% (EUR 30.20) in treated municipalities compared to control units. The results are shown in column 4 of Table 2. The corresponding event-study analysis, shown in panel (d) of (1) (2) (3) (4) Per capita Newslots Per capita venues Newslots per venue Per capita expenditure Treat ×Post -0.0012*** -0.0002** -0.4088*** 30.1966*** (0.0003) (0.0001) (0.1052) (9.3459) Observations 39,385 39,385 39,385 39,385 Municipalities 7,877 7,877 7,877 7,877 Mean outcome in 2015 0.0054 0.0015 3.6800 121.3451 Elasticity -21.42 -10.69 -11.11 24.88 Municipality FE ✓ ✓ ✓ ✓ Province ×Year ✓ ✓ ✓ ✓ Table 2: Effect on Newslots, venues and expenditure. The outcomes are the number of Newslots per capita (column 1), the number of venues hosting Newslots (column 2), the number of Newslots per venue (column 3) and the per capita net expenditure on slot machines (column 4). Elasticity is calculated and then multiplied by 100. The variable Treat is defined as in Section 4;Post represents the period after 2016. Population-weighted regressions are estimated based on Equation 12, with standard errors clustered at the municipality level. * p<.10 ** p<.05 *** p<.01. for occupation: executive/manager, administrative staff, manual workers, business owners/independent professionals, or self-employed; for the economic sector: agriculture, manufacturing, or services; and for the type of contract: full-time or part-time. 21 (a) Per capita number of Newslot devices -.002 -.0015 -.001 -.0005 0 .0005 2015 2016 2017 2018 2019 (b) Per capita number of Newslot venues -.0004 -.0003 -.0002 -.0001 0 2015 2016 2017 2018 2019 (c) Number of Newslots per venue -.8 -.6 -.4 -.2 0 .2 2015 2016 2017 2018 2019 (d) Per capita expenditure on slot machines -20 0 20 40 60 2015 2016 2017 2018 2019 Figure 6: Event-study: Effect on Newslots, expenditure, and crowding. Coefficients and corresponding 90% and 95% confidence intervals. Panel (a) shows the per capita number of Newslot devices, (b) shows per capita expenditure on slot machines, (c) shows the per capita number of venues hosting Newslots, and (d) shows the number of Newslots per venue. Regressions are population-weighted and include municipality and year-province fixed effects, with 2016 as the baseline year. Standard errors are clustered at the municipality level. Figure 6, displays a net expenditure increase of 10% in the first year, of 28% in 2018, and of 32% in 2019. These findings contradict the intended objective of the bill but they can be explained within our theoretical model as a consequence of increased temptation. In the absence of temptation costs, reducing the number of slot machines would be expected to decrease both the intensive and extensive margins of slot machine play due to higher transportation costs and reduced enjoyment. However, if the bill increases the cost of temptation, for example, through longer waiting times, crowding, or social contagion effects in treated municipalities, the opposite effect may occur (see Proposition 1). It should be noted that, since ADM did not release separate data for 2016, the per capita net expenditure on slot machines shown in column 4 of Table 2combines expenditure on both Newslots and Videolotteries. However, for the other years, separate expenditure data are 22 available, enabling us to re-run the analysis while distinguishing between Newslots and VLTs, except for 2016. This allows us to assess the direct impact of the policy on Newslots, as well as any spillover effects on Videolotteries, which were not directly affected by the policy. The results are shown in Table B.1 in the Appendix. Consistent with the previous findings, total expenditure on slot machines (Newslots and VLTs combined) significantly increases. As shown in column 3, the increase is primarily driven by increased spending on Newslots (EUR 22.27, corresponding to 70% of the overall increase). Notably, while the policy does not affect the number of Videolotteries (column 4), it also leads to an increase in expenditure on this type of device (EUR 9.72, column 5). This spillover effect is an additional unintended consequence of the bill. In Appendix Cwe show that the results presented above are robust to additional checks using (i) different definitions of treated and control groups, (ii) different sub-samples, and (iii) falsification exercises. 5.2 Closer venues, more gambling The previous Section has shown that, although the policy effectively reduced the number of slot machines and venues, it also produced unintended effects, with a 25% increase in slot machine expenditures. These results are not consistent with higher transportation costs or a decrease in the enjoyability of slot machine play in the treated municipalities. As stated in Propositions 1and 2, these drivers should instead lead to a reduction in the number of players and in the amount of play and, consequently, a reduction in slot machine expenditure. Expenditure in the treated municipalities could increase if the reduction in the number of slot machines and venues affects temptation costs. This may be the case if players concentrate in the remaining venues, potentially producing social contagion effects, peer pressure, and competition among players. More crowded venues may also reduce the stigma associated with gambling, making slot machine play more appealing.19 If these factors increase the temptation to play, and possibly the number of players, then expenditure in the treated units can increase after the policy. Notably, this effect would be more pronounced in places where the transportation costs are lower and among more addicted players (Proposition 1). To investigate the role of transportation costs, we create a measure of proximity to tobacco shops at the municipal level—since Newslots are often located in these venues—using data from the 2011 Census tracts and the 2016 list of tobacco shop addresses published by ADM. We compute the minimum distance between the centroid of each census tract and the nearest 19 The reduction in the stigma associated with being a slot machine player, due to an increase in the number of people playing slot machines in the same venue, can be formally described as a decrease in τA. As shown in Proposition 1, this would increase the extensive margin of slot machine play in the treated units. 23 Hagger, M. S., C. Wood, C. Stiff, and N. L. Chatzisarantis (2010). Ego depletion and the strength model of self-control: A meta-analysis. Psychological Bulletin 136(4), 495. Hansen, B. (2015). Punishment and deterrence: Evidence from drunk driving. American Economic Review 105(4), 1581–1617. Harrigan, K. A., K. Collins, M. J. Dixon, and J. Fugelsang (2010). Addictive gameplay: What casual game designers can learn from slot machine research. In Proceedings of the International Academic Conference on the Future of Game Design and Technology, pp. 127–133. Hartley, R. and L. Farrell (2002). Can expected utility theory explain gambling? American Economic Review 92(3), 613–624. Herskowitz, S. (2021). Gambling, saving, and lumpy liquidity needs. American Economic Journal: Applied Economics 13(1), 72–104. Hopfgartner, N., T. Santos, M. Auer, M. Griffiths, and D. Helic (2021). Social facilitation among gamblers: A large-scale study using account-based data. In Proceedings of the International AAAI Conference on Web and Social Media, Volume 15, pp. 185–195. Hotelling, H. (1929). Stability in competition. The Economic Journal 39(153), 41–57. Houser, D., J. Liu, D. H. Reiley, and M. B. Urbancic (2021). Checking out temptation: A natural experiment with purchases at the grocery register. Journal of Economic Behavior & Organization 191, 39–50. Houser, D., D. Schunk, J. Winter, and E. Xiao (2018). Temptation and commitment in the laboratory. Games and Economic Behavior 107, 329–344. James, R. J., C. O’Malley, and R. J. Tunney (2016). Why are some games more addictive than others: The effects of timing and payoff on perseverance in a slot machine game. Frontiers in Psychology 7, 46. Kearney, M. S. (2005). State lotteries and consumer behavior. Journal of Public Economics 89(11-12), 2269–2299. Laibson, D. (1997). Golden eggs and hyperbolic discounting. The Quarterly Journal of Economics 112(2), 443–478. Levitt, S. D. (2004). Why are gambling markets organised so differently from financial markets? The Economic Journal 114(495), 223–246. Lien, J. W. and J. Zheng (2015). Deciding when to quit: Reference-dependence over slot machine outcomes. American Economic Review 105(5), 366–370. Loewenstein, G. (1987). Anticipation and the valuation of delayed consumption. The Economic Journal 97(387), 666–684. Loewenstein, G. (1996). Out of control: Visceral influences on behavior. Organizational Behavior and Human Decision Processes 65(3), 272–292. 30 Loewenstein, G. (2000). Emotions in economic theory and economic behavior. American Economic Review 90(2), 426–432. Loewenstein, G. and T. O’Donoghue (2004). Animal spirits: Affective and deliberative processes in economic behavior. SSRN discussion paper 539843, Social Science Research Network. Loewenstein, G., T. O’Donoghue, and M. Rabin (2003). Projection bias in predicting future utility. The Quarterly Journal of Economics, 1209–1248. Lundborg, P. (2006). Having the wrong friends? Peer effects in adolescent substance use. Journal of Health Economics 25(2), 214–233. Mastrobattista, L., A. Minutillo, B. Genetti, A. Andreotti, R. Pacifici, and C. Mortali (2019). Gioco d’azzardo nella popolazione adulta: Studio epidemiologico trasversale di tipo osservazionale. In R. Pacifici, L. Mastrobattista, A. Minutillo, and C. Mortali (Eds.), Gioco d’azzardo in Italia: ricerca, formazione e informazione: Risultati di un progetto integrato, Rapporti ISTISAN 19/28, pp. 7–50. Roma: Istituto Superiore di Sanit`a. Moore, T. J. and T. Morris (2024). Shaping the habits of teen drivers. American Economic Journal: Economic Policy 16(3), 367–393. Muggleton, N., P. Parpart, P. Newall, D. Leake, J. Gathergood, and N. Stewart (2021). The association between gambling and financial, social and health outcomes in big financial data. Nature Human Behaviour 5(3), 319–326. Muraven, M. and R. F. Baumeister (2000). Self-regulation and depletion of limited resources: Does self-control resemble a muscle? Psychological Bulletin 126(2), 247. OECD (2018). Health at a glance: Europe 2018. pp. 26–32. Orphanides, A. and D. Zervos (1995). Rational addiction with learning and regret. Journal of Political Economy 103(4), 739–758. Ozdenoren, E., S. W. Salant, and D. Silverman (2012). Willpower and the optimal control of visceral urges. Journal of the European Economic Association 10(2), 342–368. Pesko, M. F. and C. Warman (2022). Re-exploring the early relationship between teenage cigarette and e-cigarette use using price and tax changes. Health Economics 31(1), 137–153. Petry, N. (2015). Behavioral addictions: DSM-5 ® and beyond. Oxford University Press. Piccoli, L. and S. Tiezzi (2021). Rational addiction and time-consistency: An empirical test. Journal of Health Economics 80, 102546. Pollak, R. A. (1970). Habit formation and dynamic demand functions. Journal of Political Economy 78(4, Part 1), 745–763. Resce, G., R. Lagravinese, E. Benedetti, and S. Molinaro (2019). Income-related inequality in gambling: Evidence from Italy. Review of Economics of the Household 17(4), 1107–1131. 31 Rockloff, M. J. and V. Dyer (2007). An experiment on the social facilitation of gambling behavior. Journal of Gambling Studies 23, 1–12. Rockloff, M. J., N. Greer, and L. G. Evans (2012). The effect of mere presence on electronic gaming machine gambling. Journal of Gambling Issues (27). Rockloff, M. J., N. Greer, and C. Fay (2011). The social contagion of gambling: How venue size contributes to player losses. Journal of Gambling Studies 27, 487–497. Rockloff, M. J., N. Moskovsky, H. Thorne, M. Browne, and G. M. Bryden (2017). Environmental factors in the choice of EGMs: A discrete choice experiment. Journal of Gambling Studies 33, 719–734. Santos Silva, J. and S. Tenreyro (2006). The log of gravity. The Review of Economics and Statistics 88(4), 641–658. Sch¨ull, N. D. (2012). Addiction by Design: Machine Gambling in Las Vegas. Princeton University Press. Shiv, B. and A. Fedorikhin (1999). Heart and mind in conflict: The interplay of affect and cognition in consumer decision making. Journal of Consumer Research 26(3), 278–292. Stetzka, R. M. and S. Winter (2023). How rational is gambling? Journal of Economic Surveys 37(4), 1432–1488. Strotz, R. H. (1955). Myopia and inconsistency in dynamic utility maximization. The Review of Economic Studies 23(3), 165–180. Thaler, R. H. and H. M. Shefrin (1981). An economic theory of self-control. Journal of Political Economy 89(2), 392–406. Vohs, K. D. and R. J. Faber (2007). Spent resources: Self-regulatory resource availability affects impulse buying. Journal of Consumer Research 33(4), 537–547. Wardle, H. and S. McManus (2021). Suicidality and gambling among young adults in Great Britain: Results from a cross-sectional online survey. The Lancet Public Health 6(1), e39– e49. WHO (2018). Inclusion of “gaming disorder” in ICD-11. https://www.who.int/news/item/ 14-09-2018-inclusion-of-gaming-disorder-in-icd-11, last accessed on 2024-09-20. WHO (2022). International Classification of Diseases for Mortality and Morbidity Statistics, Eleventh Revision (ICD-11). Geneva: World Health Organization, Geneva. WHO (2024). Addictive behavior. https://www.who.int/health-topics/ addictive-behaviour#tab=tab_1, last accessed on 2024-09-20. 32 A Appendix: Proofs of the theoretical model A.1 Solving the model The model is solved by backward induction. Once at a venue v∈ {A, B}, the individual problem to solve is max sv,z U(sv, z;a, σv) (15) s.t. m=pvsv+z(16) This is a standard consumer problem. The solution (s∗ v, z∗ v) at either venue satisfies: Us(s∗ v, z∗ v) Uz(s∗ v, z∗ v)=pv(17) and the budget constraint. In the following, assume that 0 < s∗ A< s∗ Bfor any given level of addiction. To simplify the notation, we use U(s∗ v) as a shorthand for U(s∗ v, z∗ v;a, σv), whenever it does not create confusion. At the participation stage, the choice depends on the comparison between the indirect utility levels that an individual at location iwith addition stock aobtains when consuming at either venue, or abstaining altogheter. Depending on the individual level of addiction a, the threshold locations that denote the indifferent player between Aand Bare iAB (a) = τB+U(s∗ A)−UB(s∗ B) τA+τB (18) The individuals indifferent between abstaining and consuming at venue vare iOA (a) = U(s∗ A)−U(0) τA ;iOB (a) = 1 −U(s∗ B)−U(0) τB (19) To justify the slope of the three loci for the indifferent consumers drawn in Figure 1, note that the threshold positions depend on the addiction level as follows: ∂iAB ∂a =1 τA+τB (Ua(s∗ A)−Ua(s∗ B)) <0,(20) ∂iOA ∂a =1 τA (Ua(s∗ A)−Ua(0)) >0,(21) ∂iOB ∂a =1 τB (Ua(0) −Ua(s∗ B)) <0.(22) Recall we assumed s∗ A< s∗ Band Usa >0 for all a, hence 0 > Ua(s∗ B)> Ua(s∗ A)> Ua(0), 33 Let aAdenote the intersection point of iAB with the vertical axis (it can be a finite value or infinity), and ass v=sss v·γ/(1 −γ) the steady state value of addiction at a venue v. Given the budget constraint, the maximum feasible consumption is sm=m/p for either v, which is associated to the addiction level am=sm·γ/(1 −γ). In equations 23 and 24, we consider the case in which the value of the addiction stock is bounded between 0 and am. This occurs when the budget constraint binds before reaching the (unconstrained) steady state, i.e. sss v> smfor either v. Formally, we assume am> aA>ˆa > 0. If the intersection point (ˆı, ˆa) is in the interior of the first quadrant, then the population is divided into consumers of sA,of sB,or abstainers, as shown in Figure 1. For given distribution function ϕ(i, a) of individual positions and levels of addiction in the population at time t, the shares of addictive consumers at either venue are: NA=ZZA ϕ(i, a) dadi=ZiOA 0Zˆa 0 ϕ(i, a) dadi+ZiAB 0ZaA ˆa ϕ(i, a) dadi(23) NB=ZZB ϕ(i, a) dadi=Z1 iOB Zˆa 0 ϕ(i, a) dadi+Z1 iAB ZaA ˆa ϕ(i, a) dadi+Z1 0Zam aA ϕ(i, a) dadi (24) A.2 Comparative statics Intensive margin of consumption. We now assess the effect of changes in the individual level of addiction, in marginal transportation costs, and in the marginal temptation costs reported in Proposition 1. By applying the implicit function theorem, the following holds: ∂s∗ v ∂a =−Usa (s∗ v) Uss (s∗ v)>0 (25) ∂s∗ A ∂τA =∂s∗ B ∂τA = 0 (26) ∂s∗ A ∂σA =−UsσA(s∗ A) Uss s∗ A>0; ∂s∗ B ∂σA = 0 (27) Note that the effect of a change in the marginal temptation cost at venue Ais higher for more addicted individuals if (omitting the arguments) ∂ ∂a ∂s∗ A ∂σA=1 U2 ss (UsσUssa −UsaσUss) = 1 U2 ss (aUssa − Uss)>0 (28) where the last equality follows from (4). In the linear-quadratic specification typically used in the rational addiction literature (see, e.g. Becker and Murphy,1988), expression (28) always holds because Ussa = 0. 34 To study the effects of a policy xthat reduces the enjoyability of addictive consumption at A, we assume that xreduces the utility and the marginal utility of addictive consumption at A, but not at B, i.e. Ux(s∗ A), Usx (s∗ A)<0, Ux(s∗ B) = 0. Then the introduction of the policy reduces consumption at A, while consumption at Bis unaffected: ∂s∗ A ∂x =−Usx (s∗ A) Uss s∗ A<0; ∂s∗ B ∂x = 0 (29) This explains the claim on the change at the intensive margin of Proposition 2. Suppose σA=σB=σ0=σ. A change in all marginal costs of temptation σimplies more play at either venue: ∂s∗ A ∂σ =−Usσ (s∗ A) Uss s∗ A>0; ∂s∗ B ∂σ =−Usσ (s∗ B) Uss s∗ A>0 (30) Extensive margin of consumption. Based on the expressions (23) and (24), we focus on changes at the extensive margins. Considering the share of consumers of sA, we are interested in the following expression: ∂NA ∂y =Zˆa 0ϕ(iOA (a), a)∂iOA ∂y dadi | {z } ny OA +Zm ˆaϕ(iAB (a), a)∂iAB ∂y dadi | {z } ny AB (31) Terms ny OA and ny AB describe the change in the mass of consumers along the OA and the AB margins, respectively, as a consequence of a change in y. Their sign depends on the sign of ∂iOA ∂y and of ∂iAB ∂y ,hence a movement to the left of either extensive margin of corresponds to a decrease in the share of consumers at A(see Figure 2for an example). Analogously, to study changes in the share of consumers of sB, we consider ∂NB ∂y =−Zˆa ˆaϕ(iAB (a), a)∂iAB ∂y dadi | {z } ny AB −Zˆa 0ϕ(iOB (a), a)∂iOB ∂y dadi | {z } ny OB (32) Consider an increase in the marginal transportation cost to reach A. Since ∂i0A ∂τA =−iOA τA <0; ∂iAB ∂τA =−iAB τA+τB <0; ∂iOB ∂τA = 0 (33) we conclude that NAdecreases, while NBand NOincrease. Considering that no effect is produced at the intensive margin of consumption, we conclude that an increase in the marginal transportation cost to reach Areduces expenditure at Aand increases expenditure at B. 35 An increase in the marginal temptation cost σAproduces analog results on the shares of consumers and abstainers, as ∂i0A ∂σA =UσA(s∗ A) τA <0; ∂iAB ∂σA =UσA(s∗ A) τA+τB <0; ∂iOB ∂σA = 0 (34) However, as shown before, consumption at Aincreases, while it remains unaffected at B. Hence, expenditure at Bunambiguously increases. The change in expenditure at A, instead, increases only if the increase at the intensive margin (30) more than offsets the decrease at the extensive margin (34). Suppose σA=σB=σ0=σ. A change in all marginal costs of temptation σ, has a different effect on the extensive margin, as ∂i0A ∂σ =Uσ(s∗ A)−Uσ(0) τA >0; ∂iAB ∂σ =Uσ(s∗ A)−Uσ(s∗ B) τA+τB <0; ∂iOB ∂σ =Uσ(0) −Uσ(s∗ B) τB <0 (35) Hence NBand EBincrease, while the share of abstainers NOdecreases. Changing the enjoyability of the good implies ∂i0A ∂x =Ux(s∗ A) τA <0; ∂iAB ∂x =Ux(s∗ A) τA+τB <0; ∂iOB ∂x = 0 (36) Hence NAdecreases, while NBand NOincrease when xreduces the enjoyability of the good. This implies that EAdecreases, while EBincreases. 36 B Appendix: Additional tables and figures (1) (2) (3) (4) (5) Total Newslots VLT expenditure Number Expenditure Number Expenditure Treat ×Post 31.8748*** -0.0011*** 22.2686*** 0.0000 9.7216** (8.8048) (0.0003) (6.5536) (0.0001) (3.8582) Observations 31,508 31,508 31,508 31,508 31,508 Municipalities 7,877 7,877 7,877 7,877 7,877 Mean outcome in 2015 121.3451 0.0054 101.8331 0.0004 19.5118 Elasticity 26.27 -20.45 21.87 3.69 49.82 Municipality FE ✓✓✓✓✓ Province ×Year FE ✓✓✓✓✓ Table B.1: Effect on expenditure by type of slot machine. The outcomes are total per capita expenditure on slot machines (column 1), the number of Newslots per capita (column 2), total per capita expenditure on Newslots (column 3), the number of Videolotteries per capita (column 4), and total per capita expenditure on Videolotteries (column 5). Elasticity is calculated and then multiplied by 100. The variable Treat is defined as in Section 4, and Post represents the period after 2016. The results are from population-weighted regressions based on Equation 12, with standard errors clustered at the municipality level. Due to data availability, year 2016 is excluded. * p<.10 ** p<.05 *** p<.01. Exposure Pop Density Old-age index Young with degree Unemployed NEET House Prices Households in Hardship TV licence (1) (2) (3) (4) (5) (6) (7) (8) Post ×Treat -0.9108 -2.9035 1.4838 4.3630 4.9784 -0.3034 7.9286 -2.9296 (7.0163) (9.2809) (7.2502) (11.7332) (11.4910) (6.9603) (10.6709) (7.1924) Post ×Treat ×Closest to tobacco shops 40.9709** 35.1430*** 35.3691*** 35.4921*** 35.6363*** 36.8134*** 37.4609*** 35.5365*** (17.2764) (11.9893) (12.4428) (13.0867) (12.9459) (13.2401) (12.6242) (13.5418) Post ×Treat ×High Exposure -12.6905 5.8661 -18.3797 -11.4608 -12.7358 -12.7674 -19.5852 1.2742 (18.8739) (14.3348) (14.6918) (17.1686) (16.8888) (15.9090) (16.5958) (16.3616) Observations 39,205 39,205 39,205 39,205 39,205 39,205 39,205 39,205 Municipalities 7,841 7,841 7,841 7,841 7,841 7,841 7,841 7,841 Municipality FE ✓✓✓✓✓✓✓✓ Province X Year FE ✓✓✓✓✓✓✓✓ Table B.2: Effect by distance from tobacco shops, robustness checks. The variable Closest to tobacco shops is a dummy equal to one when the median minimum distance between each census tract and the closest tobacco shop belong to the bottom terciles of the distance distribution. The variable High Exposure is a dummy equal to one when the municipalities belong to the top tercile of the distribution of: population density (column 1), over-65 to under-14 ratio (column 2), the share of 30-34yo with a degree (column 3), unemployment rate (column 4), NEET rate (column 5), housing prices (column 6), share of households in economic hardship (column 7), and the share of people paying public TV services (column 8). 36 municipalities are unmatched thus dropped. Coefficients associated to Post ×Closest to tobacco shops and Post ×High Exposure are not shown. Population-weighted regressions. Standard errors are clustered at the municipality level. * p<.10 ** p<.05 *** p<.01. 37 (a) Newslots -1 -.75 -.5 -.25 0 .25 .5 2016-2018 growth Veneto and Trentino Alto Adige Umbria Toscana Sicilia Sardegna Puglia, Basilicata and Molise Piemonte and Valle d'Aosta Marche Lombardia Liguria Lazio Friuli Venezia Giulia Emilia Romagna Campania Calabria Abruzzo (b) Venues hosting Newslots -1 -.75 -.5 -.25 0 .25 .5 2016-2018 growth Veneto and Trentino Alto Adige Umbria Toscana Sicilia Sardegna Puglia, Basilicata and Molise Piemonte and Valle d'Aosta Marche Lombardia Liguria Lazio Friuli Venezia Giulia Emilia Romagna Campania Calabria Abruzzo (c) VLT devices -1 -.75 -.5 -.25 0 .25 .5 2016-2018 growth Veneto and Trentino Alto Adige Umbria Toscana Sicilia Sardegna Puglia, Basilicata and Molise Piemonte and Valle d'Aosta Marche Lombardia Liguria Lazio Friuli Venezia Giulia Emilia Romagna Campania Calabria Abruzzo (d) Venues hosting VLT devices -1 -.75 -.5 -.25 0 .25 .5 2016-2018 growth Veneto and Trentino Alto Adige Umbria Toscana Sicilia Sardegna Puglia, Basilicata and Molise Piemonte and Valle d'Aosta Marche Lombardia Liguria Lazio Friuli Venezia Giulia Emilia Romagna Campania Calabria Abruzzo Figure B.1: Percentage change in the number of Newslots, VLTs, and venues from 2016 to 2018. The 20 Italian regions are aggregated into 15 regional areas, as reported in the ADM annual reports. Source: ADM annual reports. 0 .01 .02 .03 .04 Density 0 20 40 60 80 Household expenditure on gambling Figure B.2: Household expenditure on gambling Sample of households with positive expenditure on gambling (13% of the total sample) during the period 2014–2019. Expenditure is in EUR/month. 38 (a) Per capita number of Newslots 0 .0025 .005 .0075 .01 2015 2016 2017 2018 2019 Treated Controls (b) Per capita net expenditure on slots 0 50 100 150 200 2015 2016 2017 2018 2019 Treated Controls Figure B.3: Trends, raw averages. Municipality-level averages of the per capita count of Newslots (panel a) and per capita net expenditure on slot machines (panel b), by group. Red circles refer to treated units, pink diamonds to controls units. 39