Introduction

The relationship between problem gambling and criminal behavior has been examined extensively (see reviews by Adolphe et al. 2019; Banks 2017; Campbell and Marshall 2007). Most broadly, research shows that problem gambling is positively related to a range of criminogenic outcomes (Clark and Walker 2009; Grinols 2017; National Opinion Research Center 1999; Walker et al. 2010), including financially-motivated (Rudd and Thomas 2015) and violent (Laursen et al. 2016) crime. These patterns have been observed among general (e.g., Folino and Abait 2009) and incarcerated (e.g., Abbott et al. 2005; Turner et al. 2009) populations.

The extent to which problem gambling is linked to criminal behavior is grounded in Agnew’s (1992, 2001) general strain theory, which contends that crime can be a coping mechanism in response to strains—or economic and noneconomic stressors that result in negative emotions (see reviews by Eitle and Taylor 2010; Greco and Curci 2017). As this relates to problem gambling, any financial, familial, or social-psychological strains resulting from gambling might increase one’s desire for criminal behavior as a means to cope with such negative circumstances (Abbott and McKenna 2005; Adolphe et al. 2019; Blaszczynski and McConaghy 1994; Sakurai and Smith 2003; Turner et al. 2017). Indeed, among a sample of 332 male inmates in a Japanese prison, Yokotani et al. (2019) find that problem gambling increases unwanted gambling urges (i.e., strains), which, in turn, increases income-generating crimes. Moreover, based on their analysis of 90 convicted men with pathological gambling and/or antisocial personality disorders, Pastwa-Wojciechowska (2011: 675) concludes, “the crimes committed by persons who gamble result from the[ir] personal and financial problems.”

Yet, despite the theoretically sound relationship between problem gambling and criminal behavior, and the extensive number of studies showing a positive association between the two, questions regarding causality remain (Adolphe et al. 2019). For instance, some contend that offending precedes problem gambling (Abbott and McKenna 2005; Jun et al. 2019). Alternatively, as Perrone et al. (2013) note, the association between problem gambling and involvement in crime may be co-symptomatic, wherein common underlying factors might account for both gambling and crime. In other words, problem gamblers may already be prone to criminal behavior because the background characteristics that predict problem gambling also predict crime.

Indeed, research using national surveys shows that men, young adults, racial and ethnic minorities, and socioeconomically disadvantaged populations report more problem gambling behaviors compared to their respective counterparts (Welte et al. 2011, 2017). Importantly, each of these sociodemographic characteristics is also a correlate of criminality (DeLisi and Vaughn 2016; Sumter et al. 2018). Traits such as low self-control (e.g., impulsivity and irrational decision-making) and depression are also associated with both problem gambling (Bergen et al. 2012; Welte et al. 2017) and crime (DeLisi and Vaughn 2016; Gottfredson and Hirschi 1990). Furthermore, social environmental factors, such as affiliating with delinquent peers and low attachment to parents and school, are related to problem gambling (Dowling et al. 2017; Magoon and Ingersoll 2006; McComb and Sabiston 2010; Welte et al. 2017) and crime (Akers 1973; Gottfredson 2006; Haynie and Osgood 2005; Hoeve et al. 2012). Employment during adolescence and poor academic performance are also associated with both problem gambling (Canale et al. 2016; Ladouceur et al. 1999; Winters et al. 1993) and criminal behavior (Katsiyannis et al. 2008; Uggen and Wakefield 2008). Additionally, studies suggest that religiosity and involvement in religious activities are negatively associated with both problem gambling (Hoffmann 2000; Welte et al. 2017) and crime (Baier and Wright 2001; Miller and Vuolo 2018). Finally, substance abuse and being arrested are related to problem gambling (Barnes et al. 2015; Ladouceur et al. 1999; Martins et al. 2014; Winters et al. 1993; Zhai et al. 2020) as well as subsequent criminality (Bernburg and Krohn 2003; D’Amico et al. 2008).

In light of these common correlates, research has examined whether some of the aforementioned characteristics do in fact account for the relationship between gambling and crime. For instance, Mishra et al. (2011), using a sample of 180 male students, find that controlling for traits indicative of low self-control accounts for a significant proportion of the relationship between problem gambling and criminal behavior. Moreover, Vitaro et al.’s (2001) examination of 717 adolescent males suggests that earlier delinquency and substance use account for the association between problem gambling and subsequent delinquency. These findings have led some researchers to conclude that the relationship between problem gambling and crime is merely a product of shared predictors that fall under the “generality of deviance” umbrella (Hirschi and Gottfredson 1994; Jones and Quisenberry 2004; Osgood et al. 1988).

Past studies supporting the “generality of deviance” framework have provided valuable insight into the causal association between problem gambling and criminal behavior. Yet, given the relatively small, and sometimes non-random nature of the samples used in these studies, the conclusions are not generalizable to the broader population (Mishra et al. 2017). Moreover, findings from existing studies are often cross-sectional (e.g., Mishra et al. 2011), meaning that issues of temporal ordering persist. With this in mind, the present study assesses the role of confounding bias in the relationship between problem gambling and criminal behavior using a nationally representative, longitudinal sample. We also employ a series of propensity score weighting and matching techniques to account for pre-existing differences in several background characteristics between problem gamblers and non-problem gamblers (e.g., sociodemographic characteristics, low self-control, delinquent peers, parental attachment, and substance use) that may explain the relationship between problem gambling and crime.

Method

Participants

We use data from the National Longitudinal Study of Adolescent to Adult Health (Add Health), which began in 1994–1995, when over 90,000 middle and high school students from 80 schools completed an in-school questionnaire. Approximately 21,000 adolescents were selected for in-home interviews (Wave I). One year later, 14,838 of the original in-home respondents were re-interviewed (Wave II). In 2001–2002, when respondents were ages 18 to 26 years, 15,197 respondents were interviewed again (Wave III). In 2008–2009, 15,701 of the original Wave I respondents were re-interviewed (Wave IV). Although higher response rates were observed for female, white, and native-born respondents at Wave IV, the non-response bias is negligible, and Wave IV respondents represent the Wave I sample (see Brownstein et al. 2010).Footnote 1

Our analysis includes 12,227 respondents from Waves I, III, and IV. The majority of covariates have minimal missing data (i.e., less than 1%), with the smallest amount of missing data on our measures of juvenile arrest and religious activities. Some exceptions are the measures of public assistance and low self-control, where approximately 13% and 16% are missing, respectively. Testing for the missing completely at random assumption (see Little 1988) and assessing the potential role of data that are not missing at random (see Yuan 2014) suggest that our data satisfy the missing at random assumption. We therefore use multiple imputation to address missing data, and results from 10 imputed datasets are combined following Rubin’s (1987) method that accounts for variation within and between imputations.Footnote 2

Measures and Covariates

Dependent Variables: Crime Types (Wave IV)

Respondents reported whether and how often (in the past 12 months) they engaged in criminal behavior, including deliberately damaging property, stealing, using a weapon to get something from someone, selling drugs, and getting into fights. We use this information to create three dependent variables. We first create a binary measure of overall crime, which is coded as (1) for those who reported involvement in any criminal acts and (0) for those who reported no involvement. We also create an indicator of financial or instrumental crime, which is coded as (1) for those who reported engaging in crimes with a financial motive, including stealing, selling stolen property, deliberately writing bad checks, or using someone else’s credit card without their knowledge (referencing those who reported no involvement in instrumental crime). Finally, we create a binary measure indicating involvement in violent crime (coded as (1)), including fighting, hurting someone badly enough that they needed medical care, and shooting or stabbing someone (referencing no involvement in violent crime, which is coded as (0)).Footnote 3

Independent Variable: Problem Gambling (Wave III)

Add Health asked eight gambling-related questions that were developed based on those used in clinical diagnostic tools (e.g., DSM-IV-TR; the South Oak Gambling Screen) for assessing problem-gambling (Clark et al. 2013; Jun et al. 2019; see also Feigelman et al. (2006) for a discussion of the gambling-related questions used in Add Health). First, all respondents answered three questions about their lifetime gambling participation (i.e., whether they ever (1) bought lottery tickets; (2) played casino tables or video games for money; or (3) played any other games, such as cards or bingo, for money, or bet on horse races or sporting events, or taken part in any other kinds of gambling for money). Among those who reported ever gambling, respondents answered one question that asked about the largest amount of money they had ever lost from gambling in a single year. From there, respondents who reported ever being behind by more than $500 in a given year were asked the remaining four questions tapping additional gambling problems (i.e., whether (1) they ever experienced a period of two or more weeks spending a lot of time thinking about/planning gambling; (2) they ever gambled to relieve uncomfortable feelings; (3) they ever lost money gambling and returned to get even; or (4) gambling ever caused serious or repeated relationship problems).

Approximately 3% (i.e., 422 respondents) of our sample reported ever being behind by more than $500 in a given year, and these respondents were asked the four follow-up questions related to other gambling problems. The prevalence of additional gambling problems was sparse, as only about half of the 422 respondents answered affirmatively to experiencing at least one of the four other gambling problems. Thus, in order to retain the largest number of respondents exhibiting problem gambling behaviors, we code problem gamblers as (1) for those who reported being behind by more than $500 in a given year, referencing those who either did not report any gambling participation or who had never been behind by more than $500 in any given year (coded as (0)).

To ensure that our indicator of problem gambling was not masking variation in the overall pattern of results, we conducted several sensitivity analyses utilizing alternative measures for problem gambling (see Uecker and Stokes 2016). These included four measures indicating whether respondents were ever behind by $500 over a year and (1) experienced a two week period thinking about gambling; (2) gambled to relieve uncomfortable feelings; (3) returned to gambling to get even; or (4) experienced serious relationship problems due to gambling. We also used these problem gambling indicators to create measures assessing the number of gambling problems reported by respondents. Across all of these operationalization strategies, the results from these sensitivity analyses (available upon request) were identical to those presented here.

Covariates: Background Controls (Wave I)

Demographic measures include self-reported gender (1 = male; 0 = female), race/ethnicity (dichotomous indicators for non-Hispanic White, non-Hispanic Black, Hispanic, non-Hispanic Asian, and non-Hispanic other race), and age. Familial background controls for socioeconomic disadvantage include a binary indicator for whether respondents lived with both biological parents during adolescence (1 = lived with both parents; 0 = lived in any other family structure). Additionally, family socioeconomic status ranges from 1 to 10 and is based on a combination of parents’ educational attainment and occupational status (see Ford et al. 1999). Neighborhood disadvantage is the average of four census tract measures, including the percentage of adults unemployed, families below poverty, households receiving public assistance, and households headed by a single mother (Cronbach’s α = 0.931). Moreover, a dichotomous control for parents’ receipt of public assistance such as welfare is also included (1 = received public assistance; 0 = otherwise).

Delinquency is a count (ranging from 0 to 9) of respondents’ involvement in nine different behaviors in the past year (e.g., stealing, getting into physical fights, selling drugs, using a weapon to get something from someone). Delinquent peers is a count (ranging from 0 to 9) of how many of the respondents’ three best friends smoked cigarettes daily, smoked pot monthly, or drank alcohol monthly (Cronbach’s α = 0.758). Drug use is a count (ranging from 0 to 5) of respondents’ use of (1) marijuana, (2) cocaine, (3) inhalants, (4) intravenous drugs, or (5) other illegal drugs in the 30 days preceding the interview. Juvenile arrest is based on retrospective questions from Wave IV and indicates whether respondents ever experienced an arrest before their 18th birthday (1 = juvenile arrest; 0 = otherwise).

We control for respondents’ most recent grade point average (GPA), which is the average of respondents’ self-reported grades in English, history, math, and science. We also control for Add Health Picture Vocabulary Test Scores (AHPVT; ranging from 14 to 146), which is a 78-item abbreviated version of the Peabody Picture Vocabulary Test-revised. School attachment (ranging from 1 to 5) assesses respondents’ average level of agreement that they feel close to people at their school, feel a part of their school, and are happy to be at their school (Cronbach’s α = 0.775; see Johnson et al. 2006). Parental attachment is a count (ranging from 0 to 6) of whether respondents (1) talked about schoolwork or grades, (2) worked on a project for school, or (3) talked about other school-related topics with each of their parents in the past four weeks (Cronbach’s α = 0.709). Following Beaver et al. (2009), low self-control is a sum of 23 items that assess respondents’ impulsivity and decision-making processes (e.g., “you usually go with your “gut feeling” without thinking too much about the consequences of each alternative;” “when making decisions, you generally use a systematic method for judging and comparing alternatives;” Cronbach’s α = 0.765). We also control for whether adolescents work 20+ hours per week (1 = work 20+ hours; 0 = otherwise). Depression in the past week is a summed scale of the frequency (never or rarely to most of the time or all of the time) of experiencing ten indicators from the Center for Epidemiological Studies Depression Scale, such as “you felt that you could not shake off the blues” and “you felt sad” (Cronbach’s α = 0.685). Finally, we include several dichotomous measures assessing religious importance (1 = religion is very important; 0 = otherwise), weekly religious service attendance in the past year (1 = attend religious services weekly or more; 0 = otherwise), daily prayer (1 = pray daily; 0 = otherwise), and weekly youth group participation in the past year (1 = participate in youth group activities, Bible classes, or choir weekly or more; 0 = otherwise). All descriptive statistics (i.e., mean, standard deviation, and minimum and maximum values) are presented in Table 1.

Table 1 Weighted descritpive statistics for crime types, problem gambler, and all covariates
Full size table

Analytic Strategy

Propensity score methods are used to statistically balance differences in background characteristics between problem gamblers (the treatment group) and non-problem gamblers (the control group) so that we may better assess the direct relationship between problem gambling and criminal behavior, net of confounding bias (Guo and Fraser 2015). This is accomplished by first regressing our binary indicator of problem gambling on the set of background controls via logistic regression analysis and retaining the predicted probabilities. The predicted probabilities—or propensity scores—range from 0 to 1, where higher values reflect a higher likelihood that a respondent reports problem gambling behaviors.Footnote 4

Using the propensity scores, we then weight controlled respondents in a manner so that they look statistically similar to treated respondents, so as to estimate the average treatment effect on the treated (ATT). That is, treated respondents receive a weight equivalent to:

$${\text{Problem Gambler }} = \, 1$$

whereas controlled respondents receive a weight equivalent to:

$${\text{Non - Problem Gambler = }}\frac{{p({\text{Problem Gambler}})}}{{1 - p({\text{Problem Gambler}})}}$$

where p is the predicted probability generated from the logistic regression analysis. We assess whether differences in the background controls in our unweighted and weighted samples are balanced with the weights by examining the standardized differences in the means and proportions between treated and controlled respondents. A standardized difference greater than |0.10| suggests imbalance (Austin 2009). Once balance is achieved, we examine the relationship between problem gambling and criminal behavior in our weighted sample.

In addition to propensity score weighting, we also utilize propensity score matching. Rather than creating a counterfactual sample with weights, we instead match treated respondents with controlled respondents who have a similar propensity for problem gambling. We then re-examine the relationships between problem gambling and criminal behavior within our matched sample. Our results conclude by testing the robustness of our findings via alternative propensity score matching algorithms. The propensity score weighting analysis was conducted in SAS 9.4 (Lanehart et al. 2012), and the propensity score matching analyses were conducted in Stata 15 via the “psmatch2” procedure (Leuven and Sianesi 2003).

Results

Table 2 shows the means and proportions of all background controls by problem gambling status in the unweighted and weighted samples. As expected in the unweighted sample, problem gamblers differ considerably on a number of background controls compared to non-problem gamblers. For instance, problem gamblers report more delinquency, delinquent peers, and drug use. After applying the propensity score weights, however, all differences in the background controls between problem gamblers and non-problem gamblers are balanced, as evidenced by the standardized differences falling below |0.10| in the weighted sample.

Table 2 Bivariate statistics and standardized differences in background characteristics between problem gamblers and non-problem gamblers
Full size table

Table 3 shows the coefficients from several logistic regression analyses, where problem gambling is used to predict criminal behavior. Coefficients can be exponentiated (i.e., exp(bk)) in order to obtain the odds ratios. The first set of coefficients shows the estimates associated with problem gambling in the unweighted sample, which are all positive and statistically significant (p < 0.001). That is, in the unweighted sample, problem gambling increases the odds of overall crime, instrumental crime, and violent crime by a factor of 2.108, 2.029, and 2.522, respectively. In the weighted sample, however, all of the estimates associated with problem gambling are not only attenuated in size, but they all fall out of statistical significance. That is, after accounting for differences in the background characteristics between problem gamblers and non-problem gamblers, the crime-inducing effects of problem gambling are null.

Table 3 Logistic regression estimates of problem gambling on crime types: propensity score weighting
Full size table

To further investigate the role of confounding bias in the association between problem gambling and crime, a one-to-one nearest neighbor matching algorithm (without replacement) is used in the following analysis. A caliper of 0.01 is used in the matching procedure, meaning problem gamblers are matched with a non-problem gambler whose propensity score deviates by no more than 0.01 from their match. All treated respondents are successfully matched with a controlled respondent. The bivariate associations shown in Table 4 suggest that this matching procedure results in sufficient balance in the background controls between problem gamblers and non-problem gamblers, as all of the standardized differences are below |0.10|.

Table 4 Bivariate statistics and standardized differences in background characteristics between problem gamblers and non-problem gamblers in the matched sample (N = 421 matched pairs)
Full size table

Table 5 reports the coefficients from logistic regression analyses, where problem gambling is used to predict the odds of criminal behavior in young adulthood in the matched sample. For ease of comparison, we repeat the estimates from these associations in the original, unmatched sample in Table 5, which are all positive and significantly associated with increases in the odds of offending. Once again, however, these associations are no longer statistically significant after we account for pre-existing differences between problem and non-problem gamblers in our matched sample.

Table 5 Logistic regression estimates of problem gambling on crime types: propensity score matching
Full size table

To further aid interpretation, Fig. 1 presents the predicted probabilities for each of the crime outcomes for problem gamblers and non-problem gamblers in both the unmatched and matched samples. The predicted probabilities for all of the crime outcomes are significantly higher for problem gamblers in the unmatched sample; however, there are no significant differences observed in the matched sample.

Fig. 1
figure 1

Predicted Probabilities of Crime by Problem Gambling Status in the Unmatched and Matched Samples

Full size image

As a final assessment, Table 6 reports the logistic regression coefficients associated with problem gambling using three additional propensity score matching algorithms. The first employs one-to-one nearest neighbor matching with replacement, wherein controlled respondents can be used in more than one match. The second uses a one-to-three nearest neighbor matching algorithm, which matches each problem gambler with three non-problem gamblers. The third matching algorithm utilizes a kernel-based matching estimator, which matches all problem gamblers with their closest controlled match but gives greater weight to closer matches and less weight to more distant matches. Across all of the various matching algorithms and outcomes, problem gambling is not significantly associated with criminal behavior.

Table 6 Logistic regression estimates of problem gambling predicting crime types: alternative propensity score matching techniques
Full size table

Discussion and Conclusions

Using a longitudinal, nationally representative dataset and several propensity score weighting and matching techniques, this study examined the role of confounding bias in the relationship between problem gambling and criminal behavior in young adulthood. Consistent with the wealth of prior research, we initially observed a positive and statistically significant relationship between problem gambling and crime. However, once we accounted for the pre-existing differences between problem gamblers and non-problem gamblers, we found no meaningful association between problem gambling and crime.

Scholars have long debated the causal association between problem gambling and crime (Adolphe et al. 2019). Whereas some suggest that gambling is a precursor to crime (Clark and Walker 2009; Laursen et al. 2016), others suggest that crime is a precursor to problem gambling (Abbott and McKenna 2005; Jun et al. 2019). Further, some contend that problem gambling and crime are merely different manifestations of a common set of characteristics that underlie the “generality of deviance” (Barnes et al. 2005; Jones and Quisenberry 2004; Mishra et al. 2011, 2017). Our results parallel the latter, as problem gamblers exhibited social-psychological and behavioral characteristics that criminological research has long speculated to be associated with crime (e.g., Akers 1973; Gottfredson and Hirschi 1990).

The implications of our findings suggest that curtailing problem gambling, and the criminogenic consequences that may arise from problem gambling, involves implementing early prevention/intervention efforts targeted at those with a propensity for general deviance. That is, it may be effective to implement strategies aimed at reducing risky behavior in general, as opposed to strategies specific to gambling behaviors (Mishra et al. 2011, 2017; Stinchfield 2004). For example, treatment and prevention programs can attempt to strengthen mental health and decision-making abilities in at-risk youth and young adults (Na and Paternoster 2012). In addition, policies can further seek to provide opportunities to engage in structured activities that encourage school and adult interaction with at-risk youth, such as having after-school programs and activities (Nofziger and Rosen 2016). Moreover, such programs could target high-risk areas, such as more disadvantaged neighborhoods or neighborhoods with high concentrations of gambling (Welte et al. 2017). Criminologists have suggested that early intervention among those with behavioral problems, such as those mentioned above, can reduce the likelihood of involvement in crime over the life course (Moffitt 1993). Given the overlapping predictors of problem gambling and crime, such practices may simultaneously reduce a variety of problem behaviors (Barnes et al. 2005; Vitaro et al. 2001).

Although it is relatively uncommon among existing studies on problem gambling (see Gainsbury et al. 2016), we relied on propensity score methods since these quasi-experimental analyses improve our assessment of the direct effects of problem gambling on crime net of potential confounders. Indeed, randomized controlled trials are the gold standard insofar as treatment effects are concerned (Austin 2011). Yet, in the social and behavioral sciences, where assigning “treatments” like problem gambling is impractical, the estimation of such effects can be difficult. As our results show, propensity score methods share many of the appealing aspects of randomized controlled trials, and future studies examining the consequences associated with problem gambling may benefit from implementing such methods.

Some study limitations should be noted. First, because Add Health is a school-based sample, it is plausible that those most susceptible to gambling and criminal behavior (e.g., dropouts or absentees) in young adulthood are missing. Moreover, although the prevalence (3.451%) of problem gambling based on our operationalization (i.e., ever being behind by more than $500 in a given year due to gambling) is close to the national average (Welte et al. 2015), we recognize that our measure is fairly broad. As aforementioned, sensitivity analyses using alternative definitions of problem gambling confirm the patterns presented here. Moreover, Uecker and Stokes (2016), also using Add Health, examined several different operationalization strategies for problem gambling and found that the indicator for being behind by $500 had the greatest influence on the results. Thus, we are confident that our approach to distinguish problem gamblers from non-problem gamblers was sufficient; however, we do encourage future research to test the robustness of these findings with other data.

Notwithstanding these limitations, our use of a nationally representative, longitudinal U.S. sample, along with propensity score methods, make this study an important contribution to the broader literature on the relationship between problem gambling and criminal behavior. Given the stark differences in crime between problem gamblers and non-problem gamblers in our original sample, coupled with the insignificant differences in our propensity score weighted and matched samples, our results suggest that problem gambling is part of a larger—and more general—set of behaviors that predict both gambling and subsequent crime. It is hoped that these findings shed light on the long-standing debate regarding the causal link between problem gambling and criminal behavior.