4 Gambling Disorder and Semicontinuous Data

“If the gambling establishment cannot persuade a patron to turn over money with no return, it may achieve the same effect by returning part of the patron’s money on a variable-ratio schedule.”

B. F. Skinner, Science and Human Behavior (1953)

The methodological issues investigated in this thesis are applied to two studies on gambling disorder. In one study, we investigate the agreement between a collateral and the person that gambles on the amount of gambling losses, and in the other we perform an RCT to evaluate an internet-delivered support program for the concerned significant others of problem gamblers. Following is an overview of both gambling disorder and interventions aimed at CSOs, which is followed by an introduction to the problem of analyzing gambling expenditures as a treatment outcome; a problem dealt with in Study II and Study III.

4.1 Gambling Disorder

Gambling disorder is generally categorized into problem gambling and pathological gambling. While pathological gambling is defined in both DSM-5 and ICD-11, problem gambling is a broader term used to describe a less severe form of gambling problems. Problem gambling is characterized by the inability to control time spent and money wagered on gambling, despite having a negative impact on the gambler’s economy, emotional well-being, and social relations. Problem gambling is often associated with trying to win back money lost on gambling, using gambling to cope with depressive feelings, having to loan money to pay expenses or to gamble more, or lying about the time and money spent on gambling (Hodgins, Stea, and Grant 2011; Petry and Weiss 2009).

About 2% of the Swedish population between 16 and 85 years of age experience problems caused by gambling (Abbott, Romild, and Volberg 2018). Prevalence estimates for gambling problems vary across countries, from 0.2% in Norway to 5.3% in Hong Kong (Hodgins, Stea, and Grant 2011). This is generally attributed to differences in accessibility and availability of gambling, but also to differences in survey methods such as screening techniques, timeframe, administration, and response rates (Hodgins, Stea, and Grant 2011). However, results from different prevalence studies indicate that gambling problems in Sweden seems to be relatively stable across time (Abbott, Romild, and Volberg 2018; Volberg et al. 2001). Gambling problems are unevenly distributed in Sweden, with a higher prevalence among men, people born outside of Sweden, people with a low eduction, and people 18 to 24 years old (Abbott, Romild, and Volberg 2014).

Problem gambling is associated with psychological distress (Barry et al. 2011). Psychiatric comorbidity is common, especially substance-related disorder, as well as anxiety and affective disorders (Williams, Volberg, and Stevens 2012; Håkansson, Karlsson, and Widinghoff 2018; Håkansson, Mårdhed, and Zaar 2017). Furthermore, suicide attempts and suicide mortality are more common among people with gambling problems (Newman and Thompson 2007; Karlsson and Håkansson 2018).

Psychological interventions, such as cognitive-behavioral therapy is the most well-supported treatment for problem gambling (Cowlishaw et al. 2012). However, evaluating gambling treatments is challenging. Many studies of gambling treatments are faced with a substantial relapse rate (Echeburúa, Fernández-Montalvo, and Báez 2001), and high attrition rates (Westphal 2007), severely affecting what conclusions are possible to draw. Moreover, there seems to be large non-specific effects associated with just deciding to seek treatment. Placebo, or non-specific response to gambling treatment, has been discussed by several authors (Carlbring et al. 2010; Toneatto and Ladoceur 2003; Westphal 2008). Some empirical evidence for non-specific effects exists in the literature. For instance, it has been found that problem gamblers respond to very minimal interventions, such as reading a 30-page booklet on CBT, receiving one session of motivational interviewing, receiving just a clinical interview or even being put on a waiting list (Diskin and Hodgins 2009; Hodgins, Currie, and el-Guebaly 2001; Petry et al. 2008). These non-specific effects, combined with high attrition rates, make it hard to draw conclusions about the long-term effects of gambling treatments.

4.1.1 Concerned Significant Others

Gambling can not only be devastating for the gambler, it can also have serious negative effects on the lives of the concerned significant others (CSOs; Langham et al. 2016). A large portion (18%) of the adult Swedish population sees themselves as CSOs of problem gamblers (Svensson, Romild, and Shepherdson 2013). CSOs of problem gamblers tend to report worse physical and psychological health, and that the relationship to the gambler is harmed (Kalischuk et al. 2006; Volberg et al. 2001). In a representative sample in Norway, Wenzel, Øren, and Bakken (2008) found that 63% of the CSOs reported that the gambler had worsened the family’s financial situation, and 65% reported that the gambling had led to conflicts in the family. Many CSOs report that they are often left feeling isolated and unsupported (Krishnan and Orford 2002). However, CSOs of gamblers can play an essential role in recovery. For instance, as many as 50% of problem gamblers report that they rely on informal help provided by their CSO (Clarke et al. 2007), and gamblers report concerns for CSOs as an important reason for entering treatment (Bertrand et al. 2008).

There is evidence that shame and stigma are the main barriers for CSOs in seeking help (Hing et al. 2013; Valentine and Hughes 2010), and that CSOs typically turn to self-help, online or telephone support before seeking professional help (Hing et al. 2013). Thus, it is possible that an internet-delivered treatment could seem attractive to CSOs.

Despite the long list of gambling-related negative consequences that CSOs suffer, support for CSOs has been limited (Hodgins et al. 2007). A mere handful of studies have evaluated interventions for CSOs of problem gamblers. The types of interventions available for CSOs of addicts can broadly be categorized into three categories: 1) working with the CSO to motivate the addict to enter treatment, 2) involving a CSO in the treatment of the addict, and 3) working with the CSO’s needs in their own right (Copello, Velleman, and Templeton 2005).

4.1.2 CSOs and Problem Gamblers’ Motivation to Seek Treatment

Only about 5% of the problem gamblers seek professional help (Cunningham 2005; Statens folkhälsoinstitut 2010). Numerous researchers have suggested that CSOs can play a crucial role in getting the gambler to enter treatment, and they have highlighted the need to equip CSOs better to handle the problem gambling (Clarke et al. 2007; Dickson-Swift, James, and Kippen 2005; Gomes and Pascual-Leone 2009; Ingle et al. 2008; Petry and Weiss 2009). Even though financial concerns are often the main reason that gamblers seek help (Bellringer et al. 2008), many gamblers report concerns for CSOs as an important reason for entering treatment (Hing et al. 2013; Hodgins and el-Guebaly 2000).

4.1.3 Community Reinforcement and Family Training

Research on training-programs aimed at CSOs of substance misusers has shown promising results in getting treatment-refusing substance misusers into treatment. The approach with the most substantial empirical support is community reinforcement and family training (CRAFT; Copello, Velleman, and Templeton 2005; Fernandez, Begley, and Marlatt 2006; Meis et al. 2013). The CRAFT-model is based on cognitive-behavioral therapy (CBT)-principles and has three main goals: 1) motivate to the substance misuser to seek treatment, 2) decrease the substance misuse, and 3) increase the CSOs quality of life. A meta-analysis of CRAFT-studies found that, overall, 66% of the CSOs managed to get their loved ones to enter treatment (Roozen, Waart, and Kroft 2010), whereas the corresponding numbers were 18% for Al/Nar-anon and 30% for Johnson Intervention. Also, CRAFT was found to improve CSO functioning in terms of depression, anger, family conflicts, and relationship happiness.

The CRAFT-model is founded on principles from CBT, especially operant conditioning. The CRAFT-method employs six overall concepts: 1) functional analysis of the substance misuse, 2) communication training, 3) positive reinforcement of sober behavior, 4) the use of natural negative consequences, 5) helping the CSO enrich their own lives, and 6) teaching the CSO when and how to invite the substance misuser to enter treatment.

4.1.4 CRAFT and Problem Gambling

The CRAFT approach has been modified and tested with CSOs of problem gamblers in three studies. Makarchuk, Hodgins, and Peden (2002) first evaluated CRAFT for gambling in a pilot RCT, where a 45-page self-help manual was developed and evaluated. The study compared a group that received the CRAFT-manual to a control condition that received a standard information packet. Both groups displayed significant improvements, but there was no difference in treatment engagement. However, the CRAFT-group reported a greater reduction in gambling, a greater amount of satisfaction with the program, and having their needs met to a larger extent than the control condition. The same research group proceeded with evaluating the CRAFT-program in a larger RCT (n = 186; Hodgins et al. 2007). In this study, they added a CRAFT-condition that received minimal telephone support (1 to 2 calls). Unfortunately, the second study yielded essentially the same results as the pilot study—i.e., inconclusive results regarding any difference in treatment entry between the groups, but a significant difference in favor of CRAFT on days gambling, CSOs’ program satisfaction and experiences of having their needs met. The authors concluded that the approach was promising, but that it is likely that CSOs are in need of additional support in order to successfully implement the CRAFT-techniques. Nayoski and Hodgins (2016), therefore, tested the CRAFT approach in a small study (n = 32), where they compared a face-to-face treatment to a workbook-only group. No conclusive results where found, but effect sizes were in the same direction as previous studies. CSOs in the individual treatment reported greater functioning, and that their gambler spent less money and time on gambling compared to the workbook-only group.

4.1.5 Working with the CSOs in Their Own Right

Few studies have evaluated interventions that focus on working with CSOs of problem gamblers in their own right. My review of the literature only identified one such study. Rychtarik and McGillicuddy (2006) performed a preliminary evaluation of a coping skills training (CST) program for CSOs of pathological gamblers. They found a large reduction in depression and anxiety in the CST-group relative to a wait-list control. However, they could draw no conclusions regarding differences between the groups on partner gambling or treatment entry.

4.2 Semicontinuous Gambling Data

“All models are wrong but some are useful.”

George E. P. Box, 1979

An important aim of gambling treatments is to reduce gambling losses and help prevent relapses. In a consensus statement regarding the reporting of outcomes from problem gambling trials, it was proposed that measures of gambling behavior should focus on net expenditure and days gambled (Walker et al. 2006). In the published literature, these outcomes are often analyzed as if they were normally distributed, or by log transforming the outcome. Net expenditure on gambling is typically heavily skewed with some participants losing much more money than the rest. Adding further complexity, many participants stop gambling or gamble on very few days when they enter a treatment trial, resulting in data with a lot of zeros (no expenditure). Data on the daily losses on gambling, or the daily number of drinks has typically been collected by retrospective reports, such as the timeline follow-back (TLFB) method. However, the internet and smartphones have made electronic collection much more feasible. More intensive data collection methods, such as diary methods, or ecological momentary assessment, are gaining in popularity. Moreover, gambling research might be unique in the possibilities offered by behavioral tracking of online gambling. By collaborating with gambling operators, researchers get access to ecologically valid data on a transactional level. These research opportunities will likely increase; for instance, an increasing amount of gambling in Sweden is performed online (Folkhälsomyndigheten 2019). Behavioral tracking is also increasing; all gambling on Svenska spel’s land-based and online products are tracked. There will be a need to evaluate the responsible gambling tools offered by the operators, and the data generated will require sophisticated statistical methods to gain insight into the gambling behavior of the consumers.

4.2.1 Similar Problems in Other Research Fields

Other addiction sub-fields face similar issues. Studies that collect drinks per day or cigarettes per day also include a lot of zeros (Atkins et al. 2013; Bandyopadhyay et al. 2011). To analyze these data methodologists have mostly proposed different count models, most commonly zero-inflated or hurdle models. Both zero-inflated and hurdle models split the model into two parts: a Bernoulli part for modeling abstinence, and a count distribution (e.g., Poisson or negative binomial) for the number of drinks on drinking days. In the case of hurdle models, the count process is truncated at zero, whereas zero-inflated models allow both parts of the model to contribute zeros. The two types of zeros in a zero-inflated model are often called “structural” and “sampling” zeros (He et al. 2014). In addiction research, individuals that are not at risk of using, e.g., non-smokers, are viewed as structural zeros, whereas smokers that happened to not smoke during the sampling period are sampling zeros, i.e., they are at risk but did not smoke. In a hurdle model, only the Bernoulli part contribute zeros, so all individuals are conceptualized as being at-risk; thus, zeros are sampling zeros. Hurdle models are also described by a two-step decision process, where individuals first decide if they should drink or not, and once this “hurdle” is crossed a user then decides on how much to drink. Participants in clinical trials are often defined as at-risk simply due to the study’s inclusion criteria, and the hurdle model is often preferred. However, a zero-inflated model might still be useful if there are reasons to suspect two distinct processes that generate zero-observations.

DeSantis et al. (2013) found that a hurdle-Poisson model worked well to evaluate treatment effects from high-resolution drinking data. They also found that placing the hurdle at a “low-risk”-cutoff of 4 to 5 for the number of drinks per day, fit the data better than a hurdle at zero. Xing et al. (2015) proposed a two-part Bayesian random-effects model with a skewed distribution to model dependency symptoms data. Atkins et al. (2013) proposed the hurdle Poisson model for count drinking data in treatment research, and Bandyopadhyay et al. (2011) also found that a hurdle binomial model best fit their data. However, these are mostly count models, and their findings are unlikely to generalize to net losses from gambling.

Continuous dependent variables with excess zeros have a long history in the econometrical literature. Similar to the count models described earlier, the data-generating process is characterized by a two-part economic decision process. One process governs participation (binary part), and another governs the amount of money to spend (continuous part). For instance, studies on health service use typically contain a non-trivial proportion of individuals that did not use health services, and thus did not have any health expenditure during the study period. In these settings, commonly used models are the Tobit, Heckman sample selection model, and two-part model (Basu and Manning 2009; Mihaylova et al. 2011). The main difference between the models is the assumption about how zeros arise (Neelon, O’Malley, and Smith 2016). In the Tobit model, zeros are censored normal observations, and predictors are assumed to have the same influence on the decision to participate and the intensity. Whereas, both Heckman and two-part models separate the decision to participate from how much to spend (Wooldridge 2010). Conceptually, zeros in the Heckman model represents censored positive values, that could have been observed under ideal circumstances, whereas the two-part model treats zeros as actual zero expenditures. Duan et al. (1983) noted that when zeros represent actual zero expenditure, two-part models are easier to interpret. Moreover, two-part models are more numerically stable (Min and Agresti 2002). Selection models can be used to model actual outcomes, but are typically used when zeros represent missing values, hence the name sample selection models (Madden 2008). Additionally, as an alternative to two-part models, Deb and Trivedi (2002) proposed finite-mixture models, to capture e.g., “frequent” and “infrequent” use.

These models have been discussed in gambling research, mostly related to lottery participation, e.g., Humphreys, Lee, and Soebbing (2010) used them to study consumer behavior in lotteries and found that the hurdle model best fit their data. Similar results were found by Rude, Surry, and Kron (2014), who studied Swedish gambling expenditure, and Jaunky and Ramchurn (2014) when modeling consumer behaviors on scratch card markets. Economic models of gambling expenditure have mostly used Tobit or two-part models (Abdel‐Ghany and Sharpe 2001; Crowley, Eakins, and Jordan 2012; Farrell and Walker 1999; Sawkins and Dickie 2002), possibly favoring two-part models. For instance, Stranahan and Borg (1998) made the point that the decision to gamble should be statistically separated from the decision on how much to spend gambling. However, all these are cross-sectional models applied to non-clinical data, mostly applied to population expenditure and participation in lotteries or scratch cards.

4.2.2 Longitudinal Extensions

Two-part models have been extended to longitudinal analyses by incorporating random effects into each part (Olsen and Schafer 2001; Tooze, Grunwald, and Jones 2002). In clinical research, it is highly likely that the two parts of the model are correlated, e.g., it is possible that individuals that are more likely to participate in gambling, also are likely to gamble for more money. With longitudinal data it is also possible to have correlated random slopes between the two parts of the model, meaning that change in the probability of participation over time, is correlated with the change in expenditure over time. Tooze, Grunwald, and Jones (2002) described how this could be achieved by letting the two parts of the model be correlated via their random effects. Independence between the two parts are often assumed due to computational reasons; however, Su, Tom, and Farewell (2009) have shown that ignoring the correlation will bias the results.

4.2.3 Appropriate Treatment Effect Estimands

An issue with two-part models is that they lead to two treatment effects, one for each part of the model—one effect on reporting a zero, and one effect on the impact of the treatment on the non-zero values. It is possible to average over the two parts to get a “marginal” treatment effect; however, due to the nonlinear transformations, this marginal effect will be heterogeneous over the random effects (Smith et al. 2015). This property of two-part random effects models seems to have been largely overlooked in the addiction treatment studies that use them. However, instead of modeling expenditure conditional on it being non-zero, it is possible to directly model the overall expenditure in the continuous part of the model, by solving for the marginal mean (Smith et al. 2015). These marginalized two-part random effects models will lead to treatment effects on the overall expenditure that are homogenous over the random effects. Yet, as noted by Zhang, Liu, and Hu (2018), these models are not truly marginal models, in the sense that they estimate population-average estimates, but instead estimate treatment effects that are conditional on subject-specific random effects (Diggle et al. 2002). However, this conditional property of the model applies to all (generalized) linear mixed-effects models, and it is not necessarily something negative. Choosing between population-average or subject-specific effects depends on the research question. However, subject-specific models are probably most useful for clinical research, and population-average in public health research (Fitzmaurice, Laird, and Ware 2012).