Regional Data for the State of Baden-Wuerttemberg

For our analysis, we focused on the state of Baden-Wuerttemberg (BW), one of 16 federal states in Germany. The state of BW lies in southwestern Germany and is bordered by the states of Rhineland-Palatine to the northwest, Hessen to the north, and Bavaria to the east; it shares borders with two countries, France to the west and Switzerland to the south (see Figure 1). BW is the third largest of the federal states, with an area of 35,751 square kilometers and a population of approximately 10.8 million people.

Because our ecological study design used aggregate-level data, as our unit of spatial analysis, we chose local communities (Gemeinden), which correspond to the lowest administrative tier in BW. In total, there are 1,101 local communities in BW, but we considered only communities with populations over 10,000, leaving us with 244 local communities for our analysis. Our 244 communities comprised approximately 7.3 million inhabitants, which amounts to approximately 68% of the total population of BW.

The data originated from three sources. The first data set concerned the EGM locations in BW and was compiled by Trümper and Heimann (2012), who restricted their analysis to communities with a population of at least 10,000. For each of these 244 communities, they gathered information regarding the granted gambling concessions and the number of gambling establishments and EGMs from the respective public order offices (Ordnungsamt; effective date: January 1, 2012). A total of approximately 31,000 EGMs were included in our analysis.

The community data sets for our socioeconomic variables originated from the publicly available 2011 European Union Census and from data files for 2012 that were made available to us by the Statistical Office of BW upon our request.

Socioeconomic Risk Factors for Germany

We used the PAGE study to derive valid socioeconomic risk factors for our regression models (Meyer et al., 2011). With 15,023 participants, the PAGE study included the largest community-based sample in Germany. It used computer-assisted telephone interviews with a stratified and clustered sampling design conducted from January 2010 to March 2011, and it included individuals between 14 and 64 years of age. Of the interviews with the 15,023 participants, 14,022 were based on landline telephone numbers and 1,001 were based on mobile phone numbers, with the aim of including disadvantaged populations that did not have landline telephones. For the landline telephone survey, the response rate was calculated at 44.5% and the cooperation rate at 54.6%. For the mobile phone survey, the response rate stood at 36.8% and the cooperation rate at 57.3%. Using the guidelines of the American Psychiatric Association, the PAGE study categorized gambling severity according to the fourth edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV; American Psychiatric Association, 1994). The researchers categorized the severity of gambling problems into two subthreshold categories of gambling problems: one with one or two symptoms and one with three or four symptoms from the DSM-IV. Pathological gambling was defined by the presence of five to 10 symptoms from the DSM-IV criteria. In total, the 12-month prevalence rates for pathological gambling stood at 0.3% in comparison to subthreshold gambling problems with one or two criteria at 1.4% and three to four criteria at 0.3% (Meyer et al., 2011). As expected, the lifetime prevalence rates were substantially higher, standing at 1.0% for pathological gambling, 5.5% for subthreshold gambling with one to two criteria, and 1.4% for subthreshold gambling with three to four criteria. Meyer et al. (2011) used a multinomial logistic regression to quantify the relative risk for pathological gambling in relation to given sociodemographic risk factors; Table 1 summarizes their results with respect to the diagnosis of lifetime pathological gambling.

For the PAGE study, Table 1 shows that the odds ratios associated with younger age, male gender, less school education, migration background, and unemployment status are statistically significant at the 5% level (p < .05).

Ecological Analysis

Selecting variables from the PAGE study and using them in an ecological analysis makes sense because a community-based analysis offers a good starting point to assess the significance of risk factors at the ecological level. In contrast, the indiscriminate choice of explanatory variables may lead to spurious correlations and unfounded statistical associations at the ecological level.

We needed to translate the variables from Table 1 into operable aggregate-level variables. For our model, we defined younger age as the share of 15- to 29-year-olds in each community population. “Male gender” referred to the share of males in each community, and the share of migrants and unemployed individuals in the communities represented the risk factors “migration background” and “unemployment status” from the PAGE study. “School education” was captured by the share of high school graduates in the community, with high school graduation referring to finishing the 12 to 13 required years of secondary education in Germany (Abitur). At the individual level, marital status was not a predictor for a lifetime diagnosis of pathological gambling in the PAGE study (p = .87, cf. Table 1). Nevertheless, we used the share of married persons in our ecological study to assess whether this variable was associated with EGM density at an aggregate level. All of our socioeconomic risk factors were expressed per 1,000 individuals to remove variations arising from differences in population size across communities. Table 2 summarizes information regarding the variables that we used in our ecological analysis.

For the 244 communities in BW, Table 3 displays the mean, standard deviation, minimum, and maximum of our variables.

Statistical Methods

To analyze our data, we used the ordinary least squares method in a multiple linear regression model. Because all of our data were continuous, we assessed the statistical significance using Student’s two-tailed t test, with the significance level set to .05. We did not adjust alpha levels to compensate for the familywise error rate, as this would have compromised the statistical power of our analysis. Effect sizes were quantified by the unstandardized coefficients of the regression model, as well as by the unadjusted and adjusted coefficients of determination. We confirmed the validity of the regression model, but those results are not shown here; that is, we tested for the absence of heteroscedasticity and multicollinearity, the independence and normality of the error terms, and the proper model specification. The statistical program STATA 13 was used for the analyses.

Practical Conclusions

Starting from an evaluation of socioeconomic risk factors in the PAGE study, we used those individually based variables at the group level and assessed their usefulness in an exploratory analysis. For the first time in Germany, we used an ecological study design to investigate the associations between EGM density and socioeconomic deprivation at the community level.

Our analysis demonstrates a statistically significant association between EGM density and the three socioeconomic risk factors of unemployment status, migration background, and high school graduation. Judging from the effect size indices, only unemployment status is also of practical interest. With an average density of four EGMs per 1,000 individuals in the 244 communities of BW (cf. Table 3), the increase in the quantity of EGMs per additional unemployed individual (b = 0.140) is of considerable practical significance. The importance of unemployment status at the aggregate level is in line with findings from other ecological correlation studies (Pearce et al., 2008; Wardle et al., 2014; Wheeler et al., 2006). The results of our study provide preliminary indications that gambling providers place EGMs at geographical locations with relatively high unemployment rates. Further research should investigate whether this relationship also holds true for other geographical areas in Germany.

We cannot establish statistically or practically significant ecological correlations between EGM density and the share of married, male, or younger people. This is in contrast to other community-based epidemiological studies that have consistently associated gambling disorders at the individual level with younger age and male gender (e.g., Blanco, Hasin, Petry, Stinson, & Grant, 2006; Kessler et al., 2008). However, these individually based risk factors do not necessarily hold at an ecological level. The atomistic fallacy in epidemiological research describes situations in which associations between variables at the individual level may differ from associations between variables measured at the ecological level (Diez Roux, 2002). Therefore, in some cases, we may not be able to predict the behavior of groups from the behavior of individuals.

One possible solution is to reexamine statistical associations at a finer scale of ecological analysis. In future studies, we can investigate whether statistical associations exist at the city level, similar to the study of Gilliland and Ross (2005), which assessed the spatial distribution of EGMs and the socioeconomic conditions for the municipalities of Montreal and Laval. For example, Stuttgart, the state capital of BW, would provide an ideal field for carrying out an ecological gambling study. If the socioeconomic risk factors of male gender, marital status, and younger age remain nonsignificant at the city level, we can approve their nonsignificance at the ecological level with greater confidence.

Our analysis is only a starting point for investigating the complex set of interactions between socioeconomic area characteristics and EGM density. Caution is advised when moving from statistical associations at an ecological level to causal interpretations and relationships. In order to strengthen the evidence base for a relationship between socioeconomic deprivation and environmental exposure to EGMs, further longitudinal or repeated cross-sectional studies need to assess how changes in the socioeconomic environment affect the distribution of EGMs. An extension of our ecological model includes time as an important constituent of a time-geographical model. Spatial interdependencies likely vary over space and time, and it is important to identify the most crucial places and environmental contexts over different times to help estimate gambling exposure most effectively (Cummins, 2007).

Studies that examine how participation rates and gambling problems evolve over time provide another valuable research direction. International studies showed that an expansion of EGMs and increased exposure to gambling lead to rapid increases in gambling participation and problem gambling at first, followed by an adaptation period with gradual reductions in gambling participation and problem gambling rates (LaPlante & Shaffer, 2007; Shaffer, Labrie, & LaPlante, 2004; Storer et al., 2009). Factors that scholars believe influence the gradual adaptation of society to gambling include increased public awareness of problem gambling, informal social controls, expansion of treatment and self-help organizations, and regulatory and public health measures (Abbott, Stone, Billi, & Yeung, 2015). In a meta-analysis of 34 surveys in Australia and New Zealand, Storer et al. (2009) demonstrated that increased problem gambling prevalence rates are linked to higher EGM density. However, in a multivariable regression analysis, they also showed how problem gambling prevalence decreased over time at a rate of 0.09% per annum. In a meta-analysis of 202 studies from 1975 to 2012, Williams, Volberg, and Stevens (2012) affirmed the decrease in gambling participation and problem gambling rates. The downward trend began in the late 1990s in North America and in the early 2000s in Australia and Europe.

The authors of two recent studies in Sweden and Australia only partially affirmed the adaptation hypothesis. In comparing the results of the Swedish Gambling Study in 1997–1998 and the Swedish Longitudinal Gambling Study in 2008–2009, Abbott, Romild, and Volberg (2014) showed that past-year gambling participation declined from 88.0% to 71.6%, whereas past-30 days participation in EGMs increased from 3.0% to 4.0%. The authors found, contrary to the prediction of the adaptation hypothesis, a statistically significant increase in the prevalence of lifetime probable pathological gambling from 1997–1998 to 2008–2009, with no significant changes in the prevalence of lifetime problem gambling and of 12-month probable pathological or problem gambling.

For Australia, Abbott et al. (2015) compared the results of the 2003 Victorian Longitudinal Community Attitudes Survey and the 2008 Victorian Gambling Study and found significant reductions in gambling participation rates for the previous 12 months, as well as for the monthly and weekly participation frequencies. Gambling participation on EGMs for the previous 12 months decreased from 33.4% in 2003 to 21.5% in 2008, affirming the adaptation hypothesis. Here again, in contrast to the predictions of the adaptation hypothesis, there is no evidence of prevalence rate reductions in problem and moderate-risk gambling.

For Germany, the BZgA studies offer insight into the processes of exposure and adaptation (BZgA, 2014). The repeated cross-sectional analyses allow us to monitor patterns of change over time. Gambling participation rates in the 12 months preceding the surveys decreased from 55.0% in 2007 to 44.9% in 2013, affirming the adaptation hypothesis. Contrary to the general trend of decreasing participation rates in nearly all gambling activities, the 12-month prevalence of EGM gambling steadily increased from 2.2% in 2007 to 3.0% in 2013. The increases in 12-month prevalence rates of EGM gambling were most evident in 18- to 20-year-old males and 21- to 25-year old males. In the first group, the prevalence rate quadrupled from 5.8% in 2007 to 23.5% in 2013, and in the second group, it more than doubled from 5.1% in 2007 to 12.8% in 2013. It is interesting to note that the prevalence rate for the male age group of 26- to 35-year-olds increased only slightly from 5.8% in 2007 to 7.7% in 2013. How differential exposures and vulnerabilities to EGMs lead to gender and age disparities merits a systematic analysis. Does increased EGM availability explain the rise in prevalence rates among males aged 18 to 20 and 21 to 25 years? Is the male age group of 26- to 35-year-olds not vulnerable to increased EGM exposure, or has the process of adaptation decreased prevalence rates? It is also possible that the lower prevalence rate for the 26- to 35-year-old males is biased downward because the age bracket covers a 10-year range, whereas the 18- to 20-year bracket covers only a 3-year range and the 21- to 25-year bracket covers only a 5-year range. In epidemiological terms, we need to exhibit the changes in prevalence rates over time by age-period-cohort modeling strategies. In decomposing age, period, and cohort effects, we might be able to determine why certain age groups are more susceptible to increased EGM availability and which age groups are less affected or more able to adapt to the negative effects of gambling by an accumulation of exposure experience.

As far as problematic and pathological gambling behavior is concerned, the adaptation hypothesis seems not to hold from a viewpoint of statistical significance, although a trend is observed toward lower rates. The 12-month prevalence of pathological gambling in the BZgA studies decreased from 0.45% in 2009 and 0.49% in 2011 to 0.38% in 2013. The 12-month prevalence of problematical gambling decreased from 0.64% in 2009 to 0.51% in 2011 to 0.45% in 2013. The decreases in the proportion of problem and pathological gamblers in both male and female respondents were statistically nonsignificant at the 5% level.

With the 2012 Interstate Treaty on Gambling, legislators plan to implement harm minimization policies based on geographical criteria. The minimum distance regulation in BW obliges gaming halls to keep a straight-line minimum distance of 500 m from each other, as well as from child and youth facilities. These provisions will come into effect after a transitional period that ends in July 2017, with some experts expecting a 55% reduction in the number of EGMs and a 68% reduction in the number of gaming halls (Becker & Heinze, 2015; Vieweg, 2013). Area-based policies and locally targeted interventions offer promising possibilities to combat gambling problems; on the basis of our results, public authorities should focus on decreasing EGM density in communities with relatively high unemployment rates.

To support our policy recommendation, we refer to a principle that originated from German national environmental law during the 1970s (Myers, 2002): The “principle of precaution,” translated from the German word Vorsorgeprinzip, found its way into international law in the fields of environmental policy, natural resource management, and biodiversity conservation and more recently emerged as a principle of international law in international treaties and national policy statements (Andorno, 2004). The precautionary principle calls for preventive measures in public health affairs whenever potential harm to human health or the environment may be anticipated. This principle of practical decision making provides policymakers with the guidance to forestall potential adverse effects in the face of uncertainty of impact and causality.

We have reasonable grounds for concern that increased availability of EGMs raises the threat of gambling-related harm to human health. With the precautionary principle, we err on the side of caution, shifting the burden of proof about absence of harm from the public to the gambling industry. Policy makers should not wait until evidence of gambling-related harm from increased EGM availability is established beyond all reasonable doubt before taking preventive measures. Instead, they are permitted to act on the basis of evidence that is not conclusive.

Directions for Future Research

Future research should focus on two extensions to our model: multilevel and spatial regression models (Anselin, 1988; Diez Roux, 2002). Multilevel methods consider the concomitant inclusion and analysis of individual and ecological variables within a single model. In a multilevel analysis of gambling, the effect and influences of the neighborhood are a separate contextual level that acts on the individual gambling activity among the local population. For instance, living in a privileged socioeconomic environment might offer a protective effect with regard to at-risk gambling because migrant populations are less likely to reside in these neighborhoods and therefore the local population is less likely to adopt the behaviors of these vulnerable populations. In addition to capturing the simultaneous effects of individual and group-level variables on individual-level outcomes, multilevel models can incorporate interactions across levels. For example, the interaction between EGM density and socioeconomic risk factors can be reciprocal in nature: Neighborhoods with a high density of EGMs might be unattractive to populations of higher socioeconomic status, and when those populations avoid living in those neighborhoods, rental costs and property prices can decrease, which in turn attracts individuals with low socioeconomic status. By focusing on the nested structure of the data, multilevel analysis helps explain the ways that higher-level environments affect the decisions of individuals.

Whereas multilevel models consider only the correlation within neighborhoods, spatial regression models focus on the inter-neighborhood correlation structure. We expect an interaction between neighboring units that depends on geographical distance or shared borders. Future research should focus on the use of regression models with spatially varying parameters to capture the local variations over different geographical areas and to improve the understanding of local relationships and spatial variations in gambling behavior.