Taking time into account. Neoinstitutional and social learning perspectives on policy diffusion

Transcription

1 Taking time into account. Neoinstitutional and social learning perspectives on policy diffusion MARK LUTTER Max Planck Institute for the Study of Societies Paulstr. 3, Cologne, Germany Paper presented at the workshop Microfoundations of Diffusion Research: Actors, Mechanisms, Institutions September 13-14, 2012 MPIfG, Cologne 1

2 Taking time into account. Neoinstitutional and social learning perspectives on policy diffusion ABSTRACT: Despite a large body of literature on the diffusion of policy innovations, scholars are unclear about what mechanisms drive diffusion. Using event history modeling on data about US state lottery adoptions between 1964 and 2008, this study disentangles social learning factors from sociological explanations that conceptualize diffusion based on increases in the cultural legitimacy of the policy over the course of time. Results show that social learning factors are indeed important drivers of diffusion, but that these factors lose their impact over the course of diffusion. The theoretical contribution of the article lies in the attempt to empirically disentangle efficiency-based factors from legitimacy-based accounts in policy diffusion. The study argues to take seriously the duration of a diffusion process and the possibility of time change in the explanatory models of policy diffusion. KEY WORDS: lotteries, social learning, neoinstitutionalism, policy diffusion, conditional diffusion Research on the diffusion of innovations between individuals, networks, organizations, societies and states has been a vibrant area in the social sciences (Dobbin, Simmons and Garrett 2007; Strang and Soule 1998). While early diffusion studies were interested in how and in what shape innovations diffuse across time and space (Rogers 2003), only recently have researchers begun to ask why a diffusion takes place (Karch 2007: 55). An especially innovative field in this respect is the policy diffusion literature (Berry and Berry 1990; Boehmke and Witmer 2004; Grossback, Nicholson-Crotty and Peterson 2004; Volden 2006). This literature generally builds broadly on two (mutually not necessarily incompatible) explanatory models (Berry and Berry 1990; Jensen 2004: 109; Walker 1969): (1) a model of internal factors which explains the diffusion of policies through specific, idiosyncratic characteristics of the adopting state, such as internal economic, political or social conditions; and (2) a model of external determination explaining diffusion through interaction processes with other (potential) adopters. In many studies, the model of external determination results 2

3 in the empirical finding that neighboring states influence each other in their likelihood to introduce a given reform, leading to patterns of regional diffusion (Berry 1988; Berry and Berry 1990; Boehmke and Witmer 2004). While the internal determinants are largely specific to the policy being studied (Boushey 2010: 105), scholars have developed two main explanations to explain externally driven diffusion processes. First, social learning theory argues that policy innovations diffuse because adopters learn from each other, either by perceiving the success of other policy adoptions and learning that the policy works or is politically feasible (Boehmke and Witmer 2004: 40; Volden 2006), or by facing direct economic competition (Baybeck, Berry and Siegel 2011; Berry and Baybeck 2005) and therefore learn that the adoption of this policy is the optimal solution to their economic situation (Gilardi 2010: 651f.; Meseguer 2005: 74-78). Second, sociological neoinstitutional theory, by contrast, claims that policies are adopted because the innovation gains cultural legitimacy with increases in the number of adoptions (Dobbin, Simmons and Garrett 2007: 450; Meyer et al. 1997: 150). Neoinstitutional theory generally argues that organizations do not necessarily seek for the most efficient solution, but strive to fit with what they believe society expects from them (Boxenbaum and Jonsson 2008: 78), which results in isomorphic structures, or convergence of forms (for an overview, see Powell and DiMaggio 1991). With regard to policy diffusion, a policy innovation is increasingly considered an appropriate, necessary, and legitimate policy as more governments introduce the reform. Each adoption causes a cultural shift in the perception of the acceptability of that reform, leading to convergence. As a consequence, jurisdictions tend to emulate each other s policy models independent of their success or outcome. From this perspective, the diffusion process itself is a cause of diffusion. 3

4 It is still far from clear which of these two theoretical explanations can actually account for an externally driven diffusion process (Fernández 2011: 6). Too often, to quote Dobbin, Simmons, and Garret (2007: 463), diffusion studies test only their own theory or simply show evidence of diffusion and impute that their favored mechanism is at work. One reason for this separation is that the two approaches stem from two relatively distinct disciplines. Social learning theory is mainly discussed in the political science policy diffusion literature; the other is discussed in the sociological neoinstitutional theory literature. This study will focus on these two approaches and empirically disentangle social learning factors from sociological neoinstitutional theory explanations. Although it is not easy to distinguish between these two theories (Meseguer and Gilardi 2009: 531) largely because of ambiguity in adequately operationalizing both concepts, the article tries to test both by taking the duration of the diffusion process into account. First, this study employs an outcome-focused approach on learning and uses revenue gains as a measure of social learning. Second, it is argued with Tolbert and Zucker (1983), Budros (2004), and Jensen (2003) that as more states adopt a given policy, the policy becomes taken for granted, gains cultural legitimacy, and spreads increasingly independent of its actual success. Social learning factors become less crucial as time goes by. In other words, the norms, ideas and interests that guide decision makers in evaluating a policy are shifting from efficiency to legitimacy concerns; or from rational adaption with economic environments to responses of conformity with ideological-institutional environments (Kennedy and Fiss 2009: 897; Kraatz and Zajak 1996: 813). While controlling for internal determinants of policy diffusion and other relevant variables, the results of this study confirm Berry and Baybeck s (2005) study that (1) social learning and interstate competition are indeed important drivers in explaining the likelihood that a policy innovation will be adopted, but that (2) those factors lose their impact over the 4

5 course of time, as more governments introduce the policy. In the later stages of the diffusion process, the revenue success of the policy innovation is not a significant factor anymore. With these explanations, the paper attempts to provide a more complex approach to the understanding of the adoption and diffusion of public policies. It adds a specific time component to the social learning perspective and the argument discussed by Berry and Baybeck (2005). The changing influence of learning factors over the course of diffusion demonstrates the role that sociological neoinstitutional factors can play in explaining the spread of policy innovations. The present study follows up on a classic study by Berry and Berry (1990) on the diffusion and adoption of state lotteries in the United States. We have updated and expanded the data of this study, using event history modeling techniques on data for all US states for the period from 1964 to We chose to use state lottery adoptions as our empirical setting for several reasons. First, several authors have used this data (Berry and Baybeck 2005; Boushey 2010; Grossback, Nicholson-Crotty and Peterson 2004; Nicholson-Crotty 2009) and made influential contributions to the policy diffusion literature. In fact, Berry and Berry s (1990) initial study on lottery adoptions pioneered the whole field by introducing event history analysis to the study of policy diffusion (Volden 2006: 294). Using the same data allows for a direct comparison to this literature. Second, lotteries represent an interesting case of policy diffusion because their spread started from zero and grew rapidly so that almost 90 percent of US states started to run lotteries within a 45-year period. Third, state lotteries are eminently well-suited for studying policy diffusion because of their organizational and political homogeneity. Each state offers identical games, uses similar financial structures, and imposes comparable political regulations of the lottery business namely, operating lotteries as state-driven monopolies. Lastly, as Berry and Berry (1990) have argued, the adoption of a state lottery provides a good case for studying the diffusion of (political) 5

6 innovations because of their novelty. Before 1964, lotteries were illegal in all US states. In fact, with a few exceptions (Austria and Italy), lotteries were banned in the whole western world for a period of more than 100 years starting in the mid-nineteenth century (Houtman-de Smedt 1997). The article is organized as follows: The next section briefly describes the history of the adoption process for state lotteries in the United States and presents basic information on their structure and organization. Section 3 discusses the research on policy diffusion and presents the two explanatory approaches used to derive testable hypotheses. The empirical section describes the data, variable definitions and the analytic strategy for testing the hypotheses, followed by a presentation of the results. The final part concludes with a discussion of the main findings and their implications. THE SPREAD OF STATE LOTTERIES IN THE UNITED STATES SINCE 1964 The diffusion of lotteries in the United States displays a clear pattern. States in the Northeast were the first to adopt lotteries as state monopolies in the 1960s and 1970s; this development spread to neighboring states, then to the central and western states, and finally to the Southwest (see Table 1). By 1975, 26 percent of all US states, covering approximately 35 percent of the entire population, had introduced a state lottery. In the mid-1980s, approximately 40 percent of all states ran a lottery, covering about 54 percent of the population. By the mid-1990s, thirty-six out of fifty states were operating a lottery. With the introduction in Arkansas in 2008, more than 90 percent of the US population lives in lottery states. INSERT TABLE 1 HERE The diffusion process can be divided into three phases. The first phase began with the introduction of lotteries in New Hampshire and New York in 1964 and 1967, respectively. 6

7 These lotteries were only moderately commercially successful, and the gaming scheme was relatively unattractive (Blakey 1979: 79). In New Hampshire, the lottery consisted of semiannual draws based on horse racing results (Rosecrance 1988: 45). Not having to fear competition, both states initiated their lotteries without effective marketing or efficient organization: As a novelty and monopoly it fared well, but it ultimately suffered from competition and poor marketing techniques (Blakey 1979: 80). The lack of competition meant that one did not have to optimize the organization of the lottery business. The second phase came with the introduction of the lottery in New Jersey in New Jersey established a more customer-friendly lottery, which initially featured monthly drawings, followed later by weekly draws. This meant that New Hampshire and New York experienced direct competition and revenue losses. From 1972 onwards competition was further aggravated as more states in the Northeast adopted lotteries. The third phase is characterized by intensified marketing and efforts to increase revenues. New variants of the game were introduced, such as lotto. Lotto, already a successful game in Europe (Zollinger 2006), soon became the most important game in the US lottery market, accounting in 1985 for 40 percent of sales (Mikesell 1990: 316). In contrast to the traditional lottery, lotto players can pick their own lucky numbers, which makes the game more attractive. Lottery operators made use of the jackpot a system which accumulates stakes over several drawings, resulting in very large sums of prize money in order to increase payouts and the popularity of the game (Hansen 2007: 5). Finally, they started to merge into a multistate lottery system, which is now known as Powerball. This game offers mega-jackpots that regularly amount to several hundred million dollars. With their growing popularity and increase in sales, lotteries produced revenues that the states increasingly began to rely on revenues were no longer regarded as a welcome surplus for the state treasury but were a constant item in the budget 7

8 (Blakey 1979: 80). In the 2007 fiscal year, lottery revenue totaled $51.48 billion, of which $17.69 billion went towards state budgets. 1 EXPLAINING POLICY DIFFUSION: SOCIAL LEARNING AND INSTITUTIONAL PERSPECTIVES There are two main approaches to explaining externally driven diffusion processes: (1) social learning theory and (2) sociological neoinstitutional theory. The most often articulated theory in the policy diffusion literature is social learning theory, which assumes that political decision-makers learn from the success or failure of others (Boehmke and Witmer 2004; Gilardi 2010; Karch 2007; Meseguer 2005; Mooney 2001; Mooney and Lee 1995; Walker 1969). Learning basically means that decision-makers change their beliefs about the possible outcome of a policy by recognizing the effects and outcomes of other adopters policies (Dobbin, Simmons and Garrett 2007: 460; Gilardi 2010: 651). The social learning model most often sees regional proximity as a driving factor. In an uncertain world, decision-makers cannot know what effect the adoption of a new policy will have; therefore, they learn from what their neighbors have done and consider their solutions. In fact, many studies take the number of neighboring states that previously adopted the innovation as an indication of social learning theory (Boehmke and Witmer 2004), and many find a positive correlation between diffusion and regional proximity (Berry and Berry 1990; Haider-Markel 2001). 2 The reasons why adopters could be more prone to consider the policy changes adopted by their direct neighbors can be diverse, including such factors as economic, demographic, political or ideological affinity between neighbors, same and faster communication channels, shared media, stronger commercial ties, and closer personal networks (Karch 2007: 57; Weyland 2005: 271). The question of what factors actually drive the regional diffusion process, however, is still very much unresolved (Karch 2007: 58). Only recently have researchers begun to find 8

9 ways to open this black box. Grossback, Nicholson-Crotty, and Peterson (2004) distinguish between social learning and ideological similarity among adopters. They theorize and conclude that the ideological position of previous adopters enhances the chances of learning. Another strand of research finds evidence that competition over resources among neighboring states drives the regional learning pattern. For instance, in the cases of state lottery adoptions and welfare benefit levels, Berry and Baybeck (2005) show that in the diffusion process it is economic competition that matters. Employing a sophisticated analysis using geographic information systems, their study finds that state lotteries diffuse regionally due to competition over fiscal revenue. Further, Boehmke and Wittmer (2004) find that economic competition is able to predict both policy diffusion and expansion. Others attach less relevance to regional proximity as the most important measure in social learning theory. States do not necessarily learn only from their direct neighbors, but perceive reforms also in distant places. Due to globalized channels of information, ideological similarities across borders, the rise of the Internet, or the role of inter-state associations (Balla 2001), close peers might not be the only role models. As a consequence, Volden (2006: 295) argues in favor of emphasizing the success of a policy as an indicator of social learning rather than the number of neighboring states that have already adopted the policy. States only learn and adopt policies if previous policies are found to be politically or economically successful. However, there has been no evidence to date of whether perceived success contributes to the diffusion of policies across states (Volden 2006: 295). In a recent attempt, using a study on unemployment benefits retrenchments, Gilardi (2010) demonstrates that decision-makers adopt successful policies but learn selectively from others namely, with respect to political ideology. The study concludes that similar political ideologies influence how and what is perceived as successful. 9

10 The second explanation of diffusion is drawn from the sociological neoinstitutionalist literature, which originates from organizational theory (DiMaggio and Powell 1983; Meyer and Rowan 1977; Meyer et al. 1997). From this perspective, organizations adopt innovations not primarily because they seek to solve efficiency problems, but because they seek to act in conformity to others. Hence, diffusion results from an adopter s social desire to increase legitimacy. According to DiMaggio and Powell (1983), this yields in isomorphic or convergent structures. There are three mechanisms by which isomorphic structures diffuse: coercive, mimetic, and normative isomorphism. With regard to voluntary policy reforms like state lottery adoptions, there is neither a coercive force to adopt this policy, nor a normative obligation that states have to have a state lottery. But there is a mimetic factor involved: Every adoption causes a small cultural shift in the perception of the moral acceptability of that policy, and the more governments introduce the innovation, the more the innovation becomes taken for granted over the course of time, resulting in an acceleration of the diffusion process through emulation. These assumptions are reflected in a number of empirical studies which find that previous adoptions increase the legitimacy of a given innovation. For instance, Frank et al. (2010) show that the worldwide number of reforms that include changes in the regulation of criminal sexual behavior is a function of the density of previous reforms of these laws. Hirsch (1986) argues that deviant innovations such as hostile takeovers rose rapidly as the practice became widespread and legitimate. Something that would have been unthinkable a few years ago has now turned into a legitimate form of policy or practice. Lotteries are, in fact, a similar case because gambling has always been a controversial issue in terms of morality (Olson et al. 2003). Adopting a morally contested innovation such as a lottery involves social costs at the beginning of the diffusion process. But as soon as more states have introduced a lottery, the perception declines that lotteries are sinful. 10

11 However, it is not clear which of the two main explanations matter to what degree. Is it learning-based efficiency or conformity-based legitimacy that matters? One approach to resolving this problem is to think of variations in the nature of diffusion across time. The vast majority of the policy diffusion literature has analyzed diffusion from a static perspective. 3 The organizational or neoinstitutional literature dealing with the diffusion of new forms and structures is no exception in this regard (Budros 2004: 355). However, a few studies have dealt with the possibility of change in the internal causes of adoption over the course of time (Budros 2004; Jensen 2003; Kraatz and Zajac 1996; Tolbert and Zucker 1983; Kennedy and Fiss 2009; Westphal et al. 1997). The most notable and prominent example is Tolbert and Zucker s (1983) two-stage approach in their study on the diffusion of civil service reforms. The authors argue that in those cases where there was no legal requirement to enforce a reform, the causes for adoption change as time goes by. Early adoptions, the study finds, were driven mostly by specific city characteristics which point to internal organizational requirements and reasons of organizational efficiency as a cause for adoption. On the contrary, later adoptions were driven by definitions of legitimacy and a desire to appear consistent with others. The factors causing adoption changed from efficiency to questions of legitimacy and conformity (Kennedy and Fiss 2009). While Tolbert and Zucker (1983) and follow up studies test the internal explanatory model of diffusion against the neoinstitutional perspective (see also Lutter 2010; 2011), this study aims to test the external, outcome-based social learning model against the neoinstitutional perspective. If the social learning via economic competition model holds true, then laggards will adopt a policy because they have learned from the economically successful enactment of the policy by others (controlled for ideological similarity). If this is true, then the effect of a policy s success on the likelihood of adoption should be a constant considerable effect over the course of the diffusion process (see H1 below). 11

12 On the other hand, if the taken-for-grantedness model holds true, then, as diffusion goes on, laggards should adopt a policy more and more independently of its concrete economic success. The policy should become adopted because it has achieved a certain level of cultural legitimacy, not necessarily a general level of acceptance. There might be groups opposing the policy, but at least among the political decision makers who adopt the policy, it has reached a degree of acceptance which goes beyond economic efficiency concerns. If this assumption holds true, the association between the success of a policy and the likelihood of adoption should fade over time, as more states having adopted this policy. This does not mean that economic reasoning plays no role anymore in the later stages of the diffusion process. It does mean, however, that it loses some of its effect on the probability to enact such a reform as compared to early stages of the diffusion process. H1 (Social learning approach): The success of previous adoptions of a policy innovation increases the likelihood that the policy will be adopted. H2 (Sociological institutional theory approach): The longer the duration of the diffusion process, the greater the cultural legitimacy. As a result, the relationship between previous success and likelihood of adoption should decrease with the duration of the diffusion process. EMPIRICAL ANALYSIS Data The data used for the analysis is an updated version of Berry and Berry s (1990) pioneering study on US state lottery adoptions. In particular, we extend their time frame to the year 2008, which enables us to capture a total of 86 percent of the US states which have instituted a state lottery. We also include Hawaii and Alaska in the analysis, resulting in a pooled crosssectional time series dataset on all 50 US states for the years 1964 to For each state- 12

13 year, our dependent variable is binary, with value 1 if a state has introduced a lottery and 0 otherwise. The recent data on the introduction of lotteries are taken from the World Lottery Almanac (La Fleur and La Fleur 2009). Table 2 provides descriptive statistics and data sources for all the variables in this study. For a correlation matrix, see Table A1 in the appendix. 4 INSERT TABLE 2 HERE In order to test the two hypotheses, we need a measure which captures the success of lottery adoptions. We assume that laggards learn that the policy adoption is successful when it increases state revenue (Baybeck, Berry and Siegel 2011; Berry and Baybeck 2005). Therefore, we use the yearly amount of net lottery revenue (for all states), lagged by one year, and spatially weighted by distances between states. Lottery revenue data are taken from the annual tax revenue reports in the Statistical Abstracts of the United States, for each year since We use spatial weights because in line with Berry and Braybeck (2005) and Baybeck, Berry and Siegel (2011), we believe that the economic success of a nearby lottery has much more impact on adoption than revenue gains in distant regions. Following Franzese and Hays (2007), the spatial weights have been calculated from row-standardized values of a matrix of geodesic distances between the population centers of US states, so that the yearly total lottery revenues from lotteries in nearby states are weighted higher than revenues from lotteries in distant states. We use distances (miles) between state population centers rather than distances between state capitals. A population center is a common geographical measure that locates the mean center point of a region s population. 5 Again following Berry and Braybeck (2005) and Baybeck, Berry and Siegel (2011), we believe that the distance between population centers is a more appropriate measure because under conditions of economic competition, population figures matter more than the formal criterion of whether a city is politically defined as a capital. As compared 13

14 with using distances between capital cities, for instance, the state of California has, in fact, a closer distance to Arizona and Nevada than to Oregon, because California s population center is located about 300 miles south of Sacramento, taking into account the large Los Angeles metropolitan area. In addition to this main predictor, we use several controls in order to see whether the results are affected by other relevant influences. Following Grossback et al. (2004), Gilardi (2010), and the discussions above, we assume that the learning effects of a policy s success are conditioned by the ideological similarity of potential adopters because policy-makers have prior beliefs which are structured by their political ideology. Therefore, they are more likely to recognize and learn from adopters with whom they share a common ideological basis. Not controlling for the ideological distance of the adopters would result in a biased test of our hypotheses. For this reason, we include a measure of ideological distance based on the government ideology index developed by Berry et al. (1998). Government ideology is an annual indexcapturing the political color, or ideological orientation of a state. It is measured for each year between 1960 and The index is based on several indicators including roll call voting scores of state congressional delegations, outcomes of state congressional elections, partisan division of state legislatures, the party affiliation of the governor, and several other assumptions on voters and state political elites (for details, see Berry et al. 1998). The measure takes into account the special ideological structure of the United States in which the same party label can have rather different meanings in different states and at different times. The index values vary between 0 (most conservative) and 100 (most liberal). In constructing the distance measure, we have exactly followed the approach proposed by Grossback et al. (2004: 528): for each state-year, we calculate the absolute distance between the average government ideology of all previous adopters and the focal, potential 14

15 adopter. This yields an annual distance measure in which higher values indicate a larger political difference between a potential adopting state and all previous adopters. In order to control for internal factors influencing the diffusion process, we use the following indicators. First, as noted by Berry and Berry (1990), the fiscal situation of a potential adopter is relevant in explaining lottery adoptions. Therefore, we take the fiscal health measure developed by Berry and Berry (1990), updated for the years until This indicator provides information on the fiscal situation of a state at the time point t and is calculated from the difference between total state revenue and total state spending, divided by total state spending. The index has been lagged by one year in order to correct for simultaneity bias (that is, revenue from recently established lotteries that flows into the calculation of the index and thus distorts the results of the analysis). The indicator takes on positive values in the case of fiscal surpluses and negative values with fiscal deficits. Political factors are operationalized by a dummy variable that measures whether a referendum exists in a state, as a means of direct democracy through which lotteries may be introduced. Furthermore, we use the government ideology index described above in order to control for state responsiveness to the adoption of new policies (Boushey 2010: ), assuming that liberal states tend to be more prone to adopting policy innovations, whereas conservative states are more likely to maintain the status quo. In order to capture the influences of religiously-motivated opposition to gambling (Ellison and Nybroten 1999: 358f.), we use the percentage of the state population which follows fundamentalist, conservative tradition of Protestantism. Since lottery introductions might depend on population size and economic wealth, all models control for the number of residents per state and per capita personal income. In addition, since states with more borders to other states might have higher chances of adopting lotteries because of economic pressures, we control for the number of directly adjacent states. In addition, we control for the 15

16 number of neighboring states that already adopted a lottery as well as the cumulative total number of previous adoptions. In order to achieve a better model fit, several predictors have been transformed to logged values. Analytical approach Due to the structure of our data, we make use of event history methods. In particular and in line with most policy diffusion research (Brooks 2007: 710; Shipan and Volden 2006: 832) we fit logit models to our data (Box-Steffensmeier and Jones 2004: 69ff.). To account for duration dependence, problems of heteroscedasticity and non-independence of observations (Beck, Katz and Tucker 1998; Buckley and Westerland 2004), all the models include a time counter (measured in years) and rely on robust standard errors with clustering by state. Since we also use the time counter to construct the time-conditional interaction effects, we chose not to use cubic splines. We begin with a baseline model which includes the lottery revenue variable and the main controls (see Model 1 in Table 3). We then gradually add the ideological distance measure and internal determinants in the two subsequent models. In order to analyze changes over time, Models 4-6 estimate the full model for three time intervals: an early, a middle, and a late adoption phase. The early phase covers the period from 1964 to 1983, and represents the first third of lottery introductions. During this period, 34% of all states introduced lotteries. The second time interval refers to the period between 1984 and 1991, during which another 33% of the states introduced a lottery. The third and final period includes the late introductory wave between 1992 and This approach makes it possible to recognize temporal changes in the coefficient of the lottery revenue variable. The assumption was that if the policy gains cultural legitimacy over the course of time, then this social learning effect should diminish with time. Therefore, 16

17 the variable should lose impact in later periods. Since the results might depend on the chosen timeframe, in the last column, Model 7, we specify an interaction effect between the duration variable and lottery revenue. In addition, we follow Brambor, Clark, and Golder (2006: 73) and display the entire range of values of the temporal interaction effect in a marginal effects diagram (see Figure 1).8 The standard coefficient of the multiplicative interaction term in the regression model does not reveal enough substantial insight, since its value depends on the values of the constitutive terms and only refers to one single scenario e.g., when one of the constitutive variables is zero. Displaying the marginal effect diagram allows us to examine how and with what statistical precision the effects change over time across the whole range of meaningful values. In order to test the robustness of our findings, we end the analysis by presenting the results of a few alternative model specifications and sensitivity analyses. Results Before considering the two hypotheses, it is important to discuss the effects of the statistical controls. First, normative or religious factors explain lottery adoptions. In line with Berry and Berry (1990), the proportion of fundamentalist Protestants in each state has a significant negative impact on the introduction of lotteries. As can be seen from Model 4-6, the coefficient declines slightly in later periods, but remains strong in the last period (Model 6). Although an expected increase in cultural legitimacy should also involve a decline in religiously motivated moral opposition over time, the coefficient remains strong because in later periods, only those states with high cultural or moral resistance remain non-adopters. Second, fiscal-economic conditions influence the introduction of lotteries. States with budget deficits are more likely to adopt this policy innovation. The size of the population is also positively related to lottery adoptions, which may speak for the attractiveness of a lottery as a 17

18 means of gaining fiscal revenue. The larger the population, the larger the potential tax revenue from lotteries, and the more likely it is that the state will introduce a lottery. Per capita income, however, does not affect lottery diffusion. The number of borders a state shares with lottery states correlates positively with the likelihood to adopt a lottery, at least in the first three models. Further, the total number of previous events imparts a negative effect on adoption likelihood. In line with Berry and Baybeck (2005), this finding can be interpreted as an indication that states do indeed compete for tax dollars when introducing lotteries. The effect is positive for nearby states but negative for far distant enactments. Lastly, political factors show mixed influence on lottery diffusion. The referendum dummy has no significant influence on adoptions, so does the political ideology variable. What matters is the politicalideological proximity between states. As can be seen from models 2, 3, and 7, the ideological distance measure is a significant factor in explaining adoptions. Identical to the results in Grossback et al. (2004), states with a larger ideological distance from previous adopters have the tendency not to follow lottery adopters, while states with a greater ideological similarity follow the diffusion process. Turning the focus to the main hypotheses test, the first assumption suggests that the adoption of lotteries results from the successful outcome of previous lotteries. As can be seen from Table 3, the coefficient of the lottery revenue variable supports this hypothesis. The more revenue is collected through previous adoptions, weighted by proximity, the more likely it is that a state lottery will be enacted. Hence, H1, the social learning explanation, is supported. As can be seen from Models 1-3, the effect remains robust when we include further controls such as ideological distance or internal socio-economic factors. In other words, the more successful the outcome of a policy innovation is, controlled for ideological similarity and other factors, the more likely it is that lotteries will be adopted. INSERT TABLE 3 HERE 18

19 Hypothesis 2, finally, assumes that the impact of lottery revenues will fade over time. The assumption was that if a policy gains cultural legitimacy over the course of time, then the social learning effect on adoption should diminish with time. Models 4-6 support this argument. While in the early period, revenue success from previous adoptions plays a significant role in shaping further adoptions, the effect becomes insignificant in later periods. The last column, Model 7, supports this result as well. It shows that the interaction effect between duration and lottery revenue is significant. Figure 1 displays the interaction effect over the whole range of values of the duration variable. The y-axis represents point estimates of the marginal effects from Model 7 for the social learning indicator as a function of the time duration of the diffusion process since 1964 (x-axis). The figure also shows the upper and lower limits of the 95 percent confidence interval. When the interval range lies outside of the y-zero line, the coefficient is significant at 95 percent. Figure 1 clearly shows the declining effect in the relation between success of previous lotteries and adoption likelihood. As can be seen, the effect is significant approximately after 8 years duration time and loses significance after 20 years of diffusion. This means that the effect between success of previous lotteries and adoption likelihood is significant during the early, initial diffusion period of the 1970s. During that period, precisely between 1972 and 1984, revenue success of previous lotteries is a significant factor explaining lottery adoptions. After 1984, however, previous revenue loses its impact. Lotteries then become adopted independent of the success of previous adoptions. According to these results, the learning explanation is supported for the initial diffusion period , but loses salience after the 1984, speaking for increases in cultural legitimacy as a driving factor and supporting H2. INSERT FIGURE 1 HERE In order to test the robustness of these results, we performed several sensitivity analyses. Table 4 reports the results. First, we replicate the full model from Table 3 using 19

20 different specifications of the time counter. In particular, we use a quadratic specification in Model 1 and cubic polynomials t, t 2, and t 3 in Model 2, as recommended by Carter and Signorino (2010). The results do not change substantially as compared to Table 3. The last three columns test the robustness of the interaction effect. Model 3 adopts a probit estimation of the full model, and Model 4 estimates a complementary log-log regression, as suggested by Buckley and Westerland (2004: 108) or Dobbin et al. (2011: 394). The interaction effect in both models does not change as compared to Model 7 in Table 3. Finally, Model 5 replicates the full model from Table 3 using an interaction effect with the cumulative number of previous adoptions instead of the time counter. Again, results do not differ. We conclude from this analysis that our findings remain robust and unaffected under different methodological approaches. INSERT TABLE 4 HERE DISCUSSION AND CONCLUSION Much of the recent literature on the diffusion of policy innovations has aimed to contribute to a better understanding of how the diffusion of policies works within the framework of social learning theory. While some scholars have tried to analyze the role of prior beliefs and conditional hypotheses on learning and ideology (Gilardi 2010), others have tried to disentangle learning from economic competition (Boehmke and Witmer 2004). However, diffusion theorists are still far from clear as to what mechanisms drive diffusion (Dobbin, Simmons and Garrett 2007: 463). This article has aimed to elucidate the patterns of policy diffusion by focusing on a time-conditional hypothesis. The contribution of this study is that it disentangles policy learning factors based on previous successful outcomes of the policy from explanations that conceptualize diffusion based on increases in the cultural legitimacy of the policy over the 20

21 course of time. In order to examine this, the article brought a sociological neoinstitutional perspective into the discussion of policy adoptions and tried to test this perspective empirically using data on state lottery adoptions, a dataset which has been used in other influential studies in the policy diffusion literature. The article applied event history modeling techniques to an updated and extended version of that data, covering the period from 1964 to Through analyzing interaction effects with time, the findings of the empirical analysis demonstrate changes in the effect of social learning on lottery adoptions over the course of diffusion. In later phases of the diffusion process, lotteries are increasingly introduced independent of social learning factors. As the number of established lotteries increases, the study argues, they gain legitimacy as an appropriate cultural model at least among statelevel decision makers, resulting in a process of institutional isomorphism. Hence, the study goes beyond Berry and Baybeck (2005) through adding the dimension of time change to their argument of interstate revenue competition which is framed here from a learning perspective. On the one hand, this study confirms Berry and Baybeck s (2005) finding that laggards, on average, and by taking the whole diffusion period into account, learn that adopting the policy is an optimal response to their environment. On the other hand it expands their finding by taking into account a duration-conditional perspective. Interstate revenue gains lose impact over the course of diffusion. We interpret this as a shift from responses of efficiency to responses of legitimacy, which is also reflected in studies that analyze the changing media tenor of lotteries from problem to promise-related characterizations, indicating an increase in the social acceptance of lotteries (Boudreau 2000). The study therefore contributes to the policy diffusion literature by advancing previous knowledge in the field by putting a sociological neoinstitutional argument into the foreground. However, there are limitations to the interpretation of the findings of this study. First, and probably most important, the context of adoption for each adopting state is certainly 21

22 much more nuanced than the current analysis suggests through a macro-quantitative design. 9 The fading impact of revenue gains gives evidence in an indirect manner, an approach which has been criticized for its lack of precision (Kennedy and Fiss 2009: 898). Kennedy and Fiss (2009) convincingly argue that such an indirect test does not discriminate between efficiency or legitimacy concerns at the cognitive-motivation level of the adopters. As their study suggests, both logics of evaluation could apply at early and later stages of the diffusion process. Early adopter decisions are governed by motivations to achieve gains, both economic and social gains social gains through increases in status by pioneering a new policy field. Later adoptions are determined by decisions to avoid losses, again both in terms of economic considerations and with regard to concerns of legitimacy and conformity. In the case of lottery adoptions, the introduction of new games such as lotto in the 1980s could be an indicator for economic or tax considerations relevant at later stages of the diffusion process, and hence, learning or outcome-related motivations to introduce a lottery. 10 If adopters would have introduced lotteries exclusively out of legitimacy reasons, they could have as well just instituted a basic lottery without extra games. Clearly they did not. However, in that case there should be no fading, but increasing impact of revenue on adoption likelihood. Findings from our indirect test demonstrate the opposite. Second, the advantage of analyzing a dataset which includes a full picture at the macro-level comes at the cost of reduced in-depth knowledge at the micro-level. A possible microfoundation, and an avenue for future research, could be an in-depth study at the decision-making level in order to flesh out the motivations and mechanisms relevant in the diffusion process. An in-depth study could reveal a detailed story of the possible multi-stage process of adoption and show how and in what aspects early and late adopters differ from each other. An institutional logics perspective could be an especially promising framework here (e.g. Thornton 2004; Thornton and Ocasio 2008; Greenwood et al. 2010). Within this 22

23 perspective, studies have yielded important insights into the mechanisms structuring organizational change and convergence, how multiple logics, ideologies, or categories as organizing principles overlap and influence the decision making process (see empirical examples in other fields, Fourcade 2012; Haedicke 2012; Hsu 2006; Smets et al. 2012). Applying this perspective to the case of lottery adoptions could reveal important insights into the mechanisms that are at work here in order to shed more light on how early and later policy adoptions are motivated. Lastly, while this study has focused on a single policy innovation, future research could also benefit from expanding this perspective to other areas of policy innovation, or to simultaneously include several different policies in one indicator (such as in Boushey 2010) and apply a time-conditional analysis. Moreover, it would be interesting to see whether there are time-conditional differences between various policy areas such as morality policies, economic policies, welfare issues or whether there are differences in the diffusion of different types of other innovations, such as organizational practices, new technologies or consumer products. Analyses of changes with time could also be fruitful to the social movement literature on policy diffusion (e.g. Fernández and Lutter 2013), as it can be argued with this study that in changing political outcomes, social movements might have their strongest impact in early and middle stages of a diffusion process, while at later stages, cultural acceptance of the reform renders obsolete the need for social mobilization. Clearly, such an expansion to other domains would contribute to a validation of the results and the approach taken here. 23

24 ENDNOTES 1 This makes up 0.02 percent of annual total US tax revenue (which was about $757 billion in 2007). However, among the small revenue sources, lottery revenue is a big source. For instance, total US liquor store revenue was about 5.7 billion in Hence, lottery revenue is about three times higher than revenue from liquor stores (all numbers from US Census Bureau, Government Finances 2007, see It should be noted that the lottery tax has always been a controversial issue since it represents a regressive form of taxation (for an overview, see Beckert and Lutter 2009; 2012). 2 However, see Mooney (2001) reviewing studies identifying negative relationships. 3 See, however, Boushey (2010) and Nicholson-Crotty (2009), analyzing diffusion processes at different speed levels. 4 Some variables are correlated, which is not uncommon in macro-quantitative research. However, multicollinearity is not a serious problem in regression analysis, since coefficients remain statistically unbiased (see Gujarati 2003: 350). 5 For a detailed explanation on how population centers are calculated, see: The data come from the US Census Bureau: 6 We use the latest version of the index, which was lastly updated in August The index and additional information can be accessed at Richard Fording s website at 7 Since we do not have a theoretical grounded guideline to divide the time frames, we explored alternative frames prior to this analysis. In particular, we employed , , and , , Results revealed no differences to the findings presented here (but can be made available upon request). 24

25 8 In order to produce these plots, we used the post-estimation commands margins and marginsplot in Stata For insightful case study research on lottery adoptions see Gribbin and Bean (2006), Ness and Mistretta (2009), Olson et al. (2003). 10 See Pierce and Miller (1999, 2004) for a study on the introduction of new games and on how the dedication of lottery money to special purposes changed the political discourse on gambling. 25

26 TABLES AND FIGURES TABLE 1: INTRODUCTION OF US LOTTERIES, BY REGIONS Year of adoption Northeast Midwest South West % US States % population 1964 New Hampshire New York New Jersey Connecticut Pennsylvania Massachusetts Michigan Maryland Rhode Island Maine Illinois Ohio Delaware Vermont Arizona Washington Colorado Oregon Iowa California West Virginia Missouri Montana South Dakota Kansas Florida Wisconsin Virginia Kentucky Idaho Indiana Minnesota Louisiana Texas Georgia Nebraska New Mexico South Carolina Tennessee North Dakota Oklahoma North Carolina Arkansas Notes: Year of adoption taken from LaFleur and LaFleur (2009); Regions taken from classification of the US Census Bureau; Population data based on 1990 census data, taken from Statistical Abstracts of the United States. 26