Avoiding gambling harm: An evidence-based set of safe gambling practices for consumers

Prior studies have identified self-regulatory strategies that are infrequently used by problem-gamblers, but which might be protective if used. However, guidelines with evidence-based safe gambling practices (SGPs) that prevent gambling-related harm are lacking. This study aimed to: 1) identify a parsimonious set of evidence-based SGPs that best predict non-harmful gambling amongst gamblers who are otherwise most susceptible to experiencing gambling harm; 2) examine how widely are they used; and 3) assess whether their use differs by gambler characteristics. A sample of 1,174 regular gamblers in Alberta Canada completed an online survey measuring uptake of 43 potential SGPs, gambling harms and numerous risk factors for harmful gambling. Elastic net regression identified a sub-sample of 577 gamblers most susceptible to gambling harm and therefore most likely to benefit from the uptake of SGPs. A second elastic net predicted gambling harm scores in the sub-sample, using the SGPs as candidate predictors. Nine SGPs best predicted non-harmful gambling amongst this sub-sample. The behaviour most strongly associated with increased harm was using credit to gamble. The behaviour most strongly associated with reduced harm was ‘If I’m not having fun gambling, I stop’. These SGPs form the basis of evidence-based safe gambling guidelines which can be: 1) promoted to consumers, 2) form the basis of self-assessment tests, 3) used to measure safe gambling at a population level, and 4) inform supportive changes to policy and practice. The guidelines advise gamblers to: stop if they are not having fun, keep a household budget, keep a dedicated gambling budget, have a fixed amount they can spend, engage in other leisure activities, avoid gambling when upset or depressed, not use credit for gambling, avoid gambling to make money, and not think that strategies can help you win. These guidelines are a promising initiative to help reduce gambling-related harm.

Introduction A substantial minority of gamblers experience problem or at-risk gambling, leading to harmful consequences and reductions in health-related quality of life [1][2][3]. However, harm minimisation efforts to date have been criticised for focusing most attention on the small minority of gamblers with clinically significant gambling problems and for failing to reduce gamblingrelated harm [4][5][6][7][8]. Consistent with a public health approach, harm minimisation efforts need to extend beyond just reducing problem gambling, to also prevent harm amongst lower risk gamblers [9][10][11][12].
A widely used harm minimisation strategy is to advise all consumers to use various safe gambling practices (SGPs), but practices that are currently promoted are inconsistent and lack scientific evidence for their efficacy [13]. Some commonly promoted strategies have good face validity as they are a symptom of problem gambling and gambling disorder [14][15] (e.g., don't chase your losses); however, others have conflicting research evidence for their effectiveness (e.g., don't gamble alone [16]); or arguably provide little practical behavioural advice about how to implement the strategy (e.g., ensure your gambling does not cause harm for yourself or others). Evidence-based, directly actionable practices are needed to inform consumers how to keep their gambling safe.
Descriptive studies of practices used by different gambler risk groups dominate the literature on gambling self-regulatory strategies [17]. These studies have identified practices whose use is associated with lower-risk gambling, but have not yet examined them with gambling harm as, arguably, the most relevant outcome. Also, many are correlates of safer gambling, rather than behavioural strategies that could be implemented as a proactive measure. Wood and Griffiths [16] compared 1,484 'positive players' and 209 problem gamblers. The former were more likely to engage in several non-gambling leisure activities; work out what they could afford to spend and set expenditure and time limits before gambling; and take only a predetermined amount of money and not take ATM cards when going to gambling venues. They placed less importance on feeling excited and feeling relaxed to enjoy a gambling session; and were less likely to gamble when bored, depressed or upset, and were less likely to gamble with friends and family. Amongst 860 regular gamblers, Hing, Sproston, Tran and Russell [18] found several practices associated with lower-risk gambling: setting a money limit before gambling; balancing gambling with other activities; being more motivated to gamble for pleasure and entertainment, and less for money, challenge or mood regulation. Wood, Wohl, Tabri and Philander [19] identified a pool of potential practices, using factor analysis in two samples of gamblers, to develop the Positive Play Scale. All four subscales of the Positive Play Scale (honesty and control, precommitment, personal responsibility, gambling literacy) correlated negatively with PGSI score [15].
The most comprehensive categorisation to date has identified 99 behaviour change strategies used by gamblers [20]. This study involving 489 gamblers, including 333 problem gamblers, factor analysed these strategies into 15 categories: cognitive, well-being, consumption control, behavioral substitution, financial management, urge management, self-monitoring, information seeking, spiritual, avoidance, social support, exclusion, planning, feedback, and limit finances. While differences in the use of these strategies was not reported by PGSI group, problem gamblers reported greater usefulness of all strategy categories than low and moderate risk gamblers, except for planning, limiting, finances, and consumption control for which no significant differences were found. An in-venue study with a 30-day follow-up recruited 104 participants from 11 gaming machine venues who completed the 30-item Gambling In-Venue Strategies Checklist [21]. Compared to problem gamblers, low risk/non-problem gamblers more frequently avoided chasing losses, set cues to keep track of time, used only the money brought into the venue, planned their spending in advance, and viewed gambling as entertainment. Numerous other studies have examined the uptake of more limited sets of self-regulatory strategies, generally finding less use amongst higher-risk gamblers (see [13]).
These studies provide useful insights into what SGPs lower-risk gamblers tend to use more than higher-risk gamblers. However, simply examining practices that correlate with PGSI group does not necessarily identify those that best protect against gambling harm. This is because lower-risk gamblers may not use some practices simply because they have little or no need to do so. For example, they would see little need to leave their bank cards at home as they can control their spending; whereas higher risk gamblers may be more likely to use this practice as a protective strategy. Use of this practice, while potentially protective, would therefore correlate with higher-risk rather than lower-risk gambling.
To avoid this confounding issue, we first identified a sample of gamblers who are susceptible to experiencing gambling harm based on the presence of known risk factors for gambling problems. We then compared the use of SGPs amongst those who either were, or were not, experiencing gambling harm. Unlike past research, we excluded people who are not susceptible to harm, since they might not use SGPs simply because they have no need. In recognition that SGPs that are currently promoted are inconsistent and lack scientific evidence for their efficacy in protecting against gambling-related harm, this study aims to: 1) identify a parsimonious set of evidence-based SGPs that best predict non-harmful gambling amongst gamblers who are most susceptible to experiencing gambling harm; 2) examine how widely are they used; and 3) assess whether their use differs by gambler characteristics.

Formative research
Our formative research identified the (potential) SGPs tested in this study. Several methods were used to generate a comprehensive pool of potential SGPs for testing, as described elsewhere [13]. This process commenced with a systematic literature search of major online databases and the grey literature using a wide range of relevant search terms (e.g., responsibl � , gambl � , self control, self limit � , self moderat � , self help, self regulat � , harm minimis � , harm reduc � , consumption and protect � ), supplemented with a second round targeted search using search terms based on specific SGPs (e.g., precommit � , limit-set � , gambling budget, gambling motiv � ). This search located 3,707 unique publications, with 96 directly relevant to safe gambling. Of these, 26 focused on safe gambling consumption and together identified 57 SGPs.
We then conducted a content analysis of gambling-related websites as these typically provide the most comprehensive consumer advice on safe gambling, and often replicate consumer information available in print (e.g., brochures, posters). Thirty websites were purposively selected as having a comprehensive suite of safe gambling information. Because this formative research was conducted in Australia, 25 Australian websites were analysed, as well as five international websites with particularly comprehensive information. They comprised six government, 10 industry and 14 help service websites. The content analysis identified 88 additional SGPs that these websites recommended for consumers.
Several SGPs from the literature review (57) and the content analysis (88) overlapped, and we collapsed these 145 practices to 61 items. A sample of 107 gambling research, treatment, training and policy professionals were then recruited by email from the research team's professional contacts and from members of Gambling Issues International, a mailing list forum restricted to professionals who work with gambling issues. Using their professional judgment, the online survey asked these respondents to rate the importance of the 61 SGPs in helping people to gamble safely (on a 5-point scale from 'not at all important' to 'extremely important'). They were also asked to identify any other SGPs that might be important, in an open-ended question. No other SGPs were identified that did not overlap with those already included in the survey. Ten items with mean ratings below the mid-point of the scale ('moderately important') were then discarded. The remaining 51 items were considered an appropriate foundation for the current research.

Participants and procedure
A market research company, Qualtrics, recruited participants for an online survey in November and December 2017, and compensated them with points exchangeable for rewards according to their internal protocols. Inclusion criteria were: residing in Alberta Canada (location of the funding body); aged 18 years+; and at least monthly gambling (in aggregate) during the past 12 months on VLTs/slots, casino games, bingo, instant win tickets, race betting, sports betting, keno, eSports and fantasy sports. A total of 2,041 people started the survey, however 391 did not fully complete the survey and 476 failed one or more of the attention checks implemented throughout the survey. In total, 1,174 people completed the survey and met all inclusion criteria. We later subsampled from this group those most susceptible to gambling harm (n = 577). Table 1 presents demographic information for both samples.
The survey included 65 questions, although many of these were multi-item scales and question sets. Participants were advised that the survey would take approximately 20 minutes to complete.

Measures
Safe gambling practices (SGPs). The 51 practices from the formative work were condensed to 43 items by discarding five items relating to help-seeking (considered relevant only to problem gamblers), and removing three items that were similar to others from a behaviour standpoint; despite having being retained as distinct within the prior study. We operationalised the 43 items as clear statements reflecting discrete practices to which respondents could respond 'yes' or 'no' to using them within the past 12 months.
Outcome measures. Short Gambling Harms Screen (SGHS; [22]): This screen requires yes/no responses to ten gambling harm items (e.g. 'felt like a failure'), framed as whether they were experienced as a result of one's own gambling in the past 12 months. 'Yes' responses are summed. Higher scores indicate more gambling-related harm. It deliberately measures only consequential harms from gambling, and does not assess cognitions and behaviours associated with disordered gambling that are not directly harm-related. The SGHS is the only published validated instrument that exclusively measures gambling harm.
Risk factors. We reviewed the literature to identify risk factors for problematic gambling with most empirical support, along with appropriate measures. Those measured in the study were: An evidence-based set of safe gambling practices for consumers • Demographic characteristics: (see Table 1) • Importance of spirituality/religion: 5-point scale ('not at all important' to 'extremely important').
• Carer status: whether 1) primary carer for another adult; 2) dependent on another adult for primary care • Early gambling experiences: 1) age started gambling; 2) frequency of adults in household gambling when growing up; 3) frequency of gambling with or accompanying parents when they gambled; 4) whether any adults in the household had a gambling problem when growing up.
• Frequency of gambling in the past 12 months: 1) on nine different gambling activities; 2) alone; 3) online; 4) with a dependent; 5) with a carer.
• Highest spend gambling activity in the past 12 months: single question.
• Number of the friends who gamble: single question.
• Self-reported previous gambling problem: 1) prior to the past year; 2) in past two years.
• Mental disorder diagnosis from a professional: Yes/no if ever received.
• Currently consume tobacco products: yes/no • Gambling Outcomes Expectancies Scale (GOES; [23]): 18 items rated on a 6-point scale ('strongly disagree' to 'strongly agree'). Total scores are generated for five domains of gambling motivation (social, money, excitement, escape, ego enhancement). Higher scores indicate greater strength of motivations.
• Gambling Urge Scale (GUS; [24]): six items rated on a 7-point scale ('strongly disagree' to 'strongly agree' to measure thoughts and feelings about gambling urges (e.g. 'I crave a gamble right now'). Higher total scores indicate stronger urges.
• Gambling Fallacies Measure (GFM; [25]): ten items examining cognitive errors in gambling (e.g. 'a positive attitude or doing good deeds increases your likelihood of winning money when gambling'). Correct responses are coded as 1. Higher total scores reflect greater resistance to gambling fallacies.
• Brief Perceived Social Support (BPSS; [26]): six items (e.g. 'I receive a lot of understanding and security from others') measured on a 5-point scale ('does not apply at all', to 'exactly applicable'). Higher scores indicate greater perceived social support.
• Kessler Psychological Distress Scale-Brief (K6; [27]): six items pertaining to the past 30 days (e.g. 'during the last 30 days how often did you feel nervous'), measured on a 5-point scale ('none of the time' to 'all of the time'). Higher scores indicate greater psychological distress.

Statistical analysis
Our analysis aimed to evaluate the candidate SGPs in the sub-sample of gamblers who could potentially benefit from their use. This involved two stages: (1) identifying the population of gamblers most susceptible to gambling harm, and (2) evaluating the SGPs in this population. Both stages relied on a robust form of regression, 'elastic net'. Elastic net regression. In situations involving numerous potentially correlated and multicollinear predictors, ordinary least squares (OLS) regression can perform poorly in prediction and interpretation [29]. The large number of degrees of freedom, i.e. unconstrained beta coefficients, can lead to overfitting of the true effects. Interpretation is also problematic and nonintuitive, both because of the sheer number of free parameters, and also due to a phenomenon whereby beta estimates are highly interdependent. The estimated value of one beta can depend largely on the estimates of other beta coefficients, meaning that beta values can change substantially if one or more predictors are excluded from the model. This is inconsistent with the natural and desired interpretation of regression coefficients as a set of distinct and largely independent effects.
Classically, large candidate sets of predictors have been handled via different algorithms for selecting a smaller subset of predictors, including stepwise variable selection techniques. However, these methods are extremely sensitive to the peculiarities of any one dataset because of the inherent discreteness and tendency to find local optima [30]. A stable and robust alternative is to introduce an additional penalty term to the standard OLS criterion, which is to minimise the sum of squared errors (SSE). Ridge regression minimises not only SSE, but also the L 2 norm (i.e. sum of squares) of the beta coefficients themselves [31]. Similarly, the 'lasso' [32] penalises the L 1 norm, which is the summed absolute value of the coefficients. Whilst ridge regression tends to penalise overly large beta coefficients, the lasso tends to drive less useful coefficients to zero-essentially performing variable selection. Both methods encourage 'efficiency' in beta coefficients, as the estimator balances the dual criteria of maximising both predictive performance and model parsimony.
The elastic net method (R package elasticnet) incorporates advantages of both ridge regression and the lasso, incorporating both L 1 and L 2 norms in the penalty term [29]. Two metaparameters determine the amount of penalisation, and the L 1 versus L 2 balance, which are estimated via cross-validation. The practical advantages over OLS regression in this context are: 1) overfitting is largely prevented as model complexity is intrinsically constrained by ability to generalise; 2) many potential candidate predictors can be considered; 3) beta coefficients tend to reflect uncorrelated and unique effects, improving interpretation; and 4) less useful coefficients are driven to zero, yielding a robust form of variable selection. All these features are useful in both stages of analysis. Given elastic net regression is a robust procedure that automatically handles multicollinearity, and given all binary predictors (i.e. use of SGPS) and reasonably large sample size, the method only assumes that the response is a continuous variable.
Identifying the population. Elastic net regression was used to create an operational definition of gamblers most susceptible to experiencing gambling-related harm. The predictor variable set comprised all risk factors for gambling-related harm, and the outcome was SGHS score [22]. The predicted scores of this model represent a measure of vulnerability in that they reflect the expected value of harm, integrating information from all available risk factors. We defined this population as those having an expected value of 1+ harms, regardless of whether or not they had actual reported harms. These gamblers do not necessarily experience harm but still experience risk; consequently, this group also included some unharmed gamblers. The elastic net allowed us to incorporate a large number of correlated risk factors in making the estimation of an expected value of 1+ harms, whilst preventing overfitting to the data.
Evaluating the SGPs. The second analysis step was to evaluate the SGPs with respect to the restricted set of gamblers most susceptible to experiencing gambling-related harm. We excluded gamblers who were not susceptible to harm from this analysis, since some people may not use SGPs simply because they do not need to. Elastic net regression was also employed in evaluating the SGPs, using all candidate SGPs as predictors, and the SGHS score again as the predicted outcome amongst the subset of vulnerable gamblers. Multivariate regression is intrinsically geared towards identifying predictors with unique explanatory power. However, as described above, the elastic net variant provides an additional advantage in that it accomplishes implicit variable selection. That is, it identifies the smallest set of SGPs that are instrumental in affecting the outcome. Negative parameters indicate that use of a SGP is associated with a reduction in harms; positive coefficients with an increase in harms.

Ethics
The study procedures were carried out in accordance with the Declaration of Helsinki. The Institutional Review Board of the University of Alberta approved the study. All subjects were informed about the study via a participant information sheet that detailed the study's aims, investigators, survey topics, survey length, voluntary participation, that they could withdraw at any time, that the survey was anonymous, the security of data storage, publication of aggregated results, ethics approval number and contact details of the approving ethics office. All participants provided informed consent by clicking 'yes' to confirm that they were 18 years or over and 'yes' to confirm that they were providing informed consent to participate in the study. Table 2 presents the standardised elastic net regression coefficients for the risk model predicting harm from all risk factors as the first step of the analysis: identifying the restricted set of gamblers most susceptible to experiencing gambling-related harm. Note that standard errors / p-values usually associated with regression models are not applicable to elastic nets. However, the coefficients themselves indicate relative variable importance, in the context of all other predictors. The risk model explained 40.1% of the variance in harm scores. The most important predictor of (expected) current gambling harm was the existence of gambling problems prior to the past two years (b = .77), followed by the presence of gambling urges (.63), impulsivity (.27), and when as a child, adults in the household had a gambling problem (.26).

Results
The expected number of harms for each respondent, given knowledge of their risk factors, was generated from the risk model. Of the 1,174 cases analysed, 577 cases had an expected harms score equal to or greater than one. For the second step in the analysis, a second elastic net again predicted harm scores for these 577 gamblers, using the SGPs as candidate predictors. Table 3 provides the SGPs associated with higher or lower degrees of harm amongst these gamblers. Negative coefficients indicate SGPs that are associated with less gambling harm; positive coefficients indicate those that are associated with more gambling harm. The behaviour most strongly associated with increased harm (b = 2.08) was using credit card cash advances to gamble. The behaviour most strongly associated with reduced harm was 'If I'm not having fun gambling, I stop' (b = -1.07).
The top portion of Table 3 identifies the most effective SGPs, addressing the first aim of the study. Our primary criteria for selection was efficacy in independently predicting gambling related harm. However, importantly, two SGPS with strong effect sizes were manually  An evidence-based set of safe gambling practices for consumers An evidence-based set of safe gambling practices for consumers excluded (last two rows in Table 3) from the list of effective practices, based on item content. We screened the top performing SGPs based on whether they could be framed as positive, general advice to gamblers. For example, the positive coefficient for 'I often talk about gambling with my friends and/or family' suggests that this is an indicator of a problematic preoccupation with gambling. However, it would be clearly unhelpful to advise gamblers not to discuss their gambling with their family. The other excluded item assumes that the respondent is in a primary carer relationship, and therefore is unsuitable as general advice. Of the selected items, 1-6 were most strongly associated with reduced gambling harm and therefore also represent the most likely efficacious practices to use for gambling to be nonharmful. Items 7-9 were most strongly associated with increased gambling harm, and represent the most evident practices to avoid. Our cut-off point of nine SGPs is somewhat arbitrary, although it was also a choice-point informed by their appropriateness for consumer messaging and guidelines. An expanded set of effective SGPs might be useful for other purposes.
To address the study's second aim, we examined how widely the nine most evidently important SGPs were used by gamblers who are most susceptible to experiencing gamblingrelated harm (Table 4).
Each SGP associated with reduced gambling harm was used by over 70% of these gamblers, except for 'I have a dedicated budget to spend on gambling' and 'If I'm feeling depressed or upset, I don't gamble', each used by only 45% of this group. Approximately three-quarters of these vulnerable gamblers reported not using each SGP associated with increased gambling harm, except for 'I use gambling to make money / supplement my income' which 65.5% did not use.
The third aim was to assess whether use of SGPs differs by gambler characteristics. A total SGP score was calculated by scoring +1 for use of SGPs items 1-6 and -1 for items 7-9. Nonparametric tests examined relationships between total SGP score and the independent personvariable predictors. A Mann-Whitney U test examined gender and SGP scores. Spearman's correlations were examined between total SGP score and (in-turn): age, PGSI score, K6 score and BIS score. Kruskall-Wallis tests examined differences in SGP scores and (in-turn): PGSI group, gambling frequency, and highest spend gambling activity. Where significant differences were found, Mann-Whitney U tests, with Bonferroni corrections, were performed as post-hoc analyses.

Discussion
Previous research has identified self-regulatory strategies used more by lower-risk than higher-risk gamblers [16,[18][19]. Our study extends this research by identifying a set of nine safe gambling practices that best prevent gambling-related harm amongst those most susceptible to experiencing harmful consequences from their gambling. It is important to note that these nine SGPs are not the only practices that can help to protect against harmful gambling, and that other practices promoted on gambling-related websites, in player information, in the broader media and by treatment professionals can also be useful. The nine SGPs are those that were the most protective amongst the much larger group of change strategies that gamblers can use to self-regulate their gambling, and therefore can provide the basis of an evidencebased set of guidelines.
Our parsimonious set of nine SGPs can be expressed as the following safe gambling guidelines: • If you're not having fun gambling, stop.
• Keep a household budget.
• If you gamble, have a dedicated budget for your gambling.
• Engage in other leisure activities, hobbies, social activities or sports.
• Do not gamble if you're feeling depressed or upset.
• When you gamble, always set aside a fixed amount you can spend.
• Do not use credit, or cash advances on your credit card, to gamble.
• Do not use gambling to make money or supplement your income.
• Do not think that systems or strategies will ensure your success at gambling.
While the nine SGPs may sound simplistic, their effective implementation requires gamblers to enact several broader cognitive-behavioural change strategies that can be used to selfregulate gambling [20]. These include strategies relating to limiting finances (keep a household budget), controlling consumption (having a dedicated gambling budget, setting aside a fixed amount to spend, not using credit to gamble), avoidance of certain behaviours (not gambling if depressed or upset, stop gambling if not having fun), behaviour substitution (engage in other leisure activities), and cognitive strategies (not thinking systems or strategies will help you win, not using gambling to make money). The strategies relate both specifically to gambling (e.g., not gambling when upset or depressed) and to practices that are not specific to gambling (e.g., keep a household budget, engage in other leisure activities), so their effective implementation requires changes beyond gambling behaviour alone. Importantly, the nine SGPs encompass both distal (pre-gambling) and proximal (during gambling) strategies [33]. Distal strategies include, for example, keeping a household budget, setting a dedicated budget for gambling, allocating a fixed amount that one can gamble before commencing gambling, and engaging in other leisure activities. Proximal strategies, such as stopping gambling if not having fun, require gamblers to take actions during a gambling session, which is likely to be more difficult than adhering to distal strategies. Many gamblers find it difficult to limit their gambling during play when they may feel excited, frustrated, emotional, dissociated, vulnerable to erroneous beliefs, subject to peer pressure, and tempted to chase losses [34][35], and effective strategies to manage gambling urges appear to be particularly challenging [20].
Effectively implementing behaviour change strategies, such as the nine SGPs, requires adequate action and coping planning between intentions and behaviour in order to realise the behavioural goal [36][37]. The process requires pre-decisional strategies to form intentions to achieve a desired goal (e.g., reducing or quitting gambling), pre-actional strategies by using planning to initiate the intention or goal, actional strategies to implement the behaviour, and post-actional strategies to evaluate it outcomes [38]. Coping planning is also needed to identify situations and barriers where goals may be undermined [38]. Additional research is needed to understand how action and coping planning can best support the implementation of the nine SGPs.
Currently, each of the nine SGPs is promoted on various gambling help, government and gambling industry websites, but these websites often do not include all of these SGPs and often instead include practices with lower demonstrated efficacy in protecting against gamblingrelated harm. In contrast, the guidelines developed in this study could be consistently promoted on gambling-related websites and apps, in public health materials, and in gambling venues. Market testing might optimise wording to ensure resonance and comprehension. These guidelines can also form a consumer self-assessment test, ideally with automated personalised feedback that identifies practices to make an individual's gambling safer. Gambling treatment providers might also use the guidelines to provide practical advice to clients on cognitive-behavioural change.
The use of SGPs can also be measured at a population level. Prevalence studies rely on problem gambling screens to track changes in maladaptive gambling behaviour. However, the prevalence of problem gambling is too low to reliably detect changes between assessment periods. Instead, prevalence studies could measure the use of SGPs to detect changes in safe gambling behaviour, which would be more reliable, given the much greater prevalence of SGP use in a population. Such assessments would be particularly useful to evaluate the efficacy of new harm minimisation initiatives, as well as changes in policy and practice that might be expected to impact on harmful gambling.
This study can inform harm minimisation efforts in Alberta, as well as across wider locations. Certain evidently helpful SGPs were practised by only a minority of gamblers who are susceptible to experiencing gambling-related harm; specifically having a dedicated budget to spend on gambling, and not gambling when feeling depressed or upset. Public health messaging promoting these practices may help to increase their uptake. Male gamblers, more frequent gamblers especially on VLTs/slots, and gamblers with higher impulsivity, psychological distress and PGSI scores are most likely to experience gambling-related harm, but are less likely to use SGPs. This knowledge can inform public health communications which can be tailored accordingly in terms of target audiences, appropriate messages, and use of relevant media.
Given that public health messaging is rarely sufficient on its own to change behaviour [39], the SGPs should also be used to change policy and practice. In addition to promoting the guidelines, gambling regulators and operators could facilitate use of the SGPs. They could provide budgeting tools to encourage gamblers to calculate an affordable gambling budget in the context of their overall household budget. Operators could provide pre-commitment systems to facilitate limit-setting prior to gambling. They could avoid extending credit for gambling and prevent customers using credit cards to gamble. Operators and regulators should ensure that gambling advertising does not encourage faulty cognitions, such as suggesting that certain systems or strategies will enhance the chances of winning. Identifying these specific changes to policy and practice that directly relate to the SGPs does not preclude the need for additional reforms to overcome the limitations of responsible gambling [4][5][40][41], but discussion of these reforms is beyond the scope of the current paper.

Limitations and future research
Data were collected only in Alberta with a modest sample size. While reasonably balanced by gender, age and other demographic characteristics, our convenience sample was unlikely to be representative of the population of gamblers. Replicating the study in other locations and with larger and more representative samples is needed to confirm the results. Some variation in the uptake of SGPs may be expected in different locations, given that socio-economic characteristics, cultural norms, legal gambling forms, their accessibility and marketing vary. Further, some comprehensive studies of self-regulatory practices used by gamblers have been published since the survey was conducted for the current research, and should be considered in future research to inform the set of SGPs tested [20][21]33]. As noted earlier, the safe gambling guidelines would benefit from market testing to optimise wording to ensure resonance and comprehension. For example, ' . . .have a dedicated budget for your gambling' might be clearer as ' . . .have a dedicated budget for your gambling and stick to it'. As noted earlier, research is needed to understand how action and coping planning can best support the implementation of the nine SGPs. Finally, evaluation studies could examine the efficacy of the guidelines across different forms of gambling, and over time in longitudinal designs.

Conclusion
To our knowledge, this study has developed the first evidence-based set of safe gambling practices whose use predicts the absence of gambling-related harm amongst gamblers who might otherwise be expected to experience harm. As safe gambling guidelines, they provide practical direction for consumers on how to avoid harmful gambling behaviours and consequences. They can be further used to measure the prevalence of safe gambling and changes over time at the population level, and to inform supportive changes in gambling policy and practice.