Predicting future harm from gambling over a five-year period in a general population sample: a survival analysis

The Leisure, Lifestyle, and Lifecycle Project (LLLP), described in detail in other sources [22, 38, 39], was a prospective five-year panel study of 1808 adolescents and adults living in rural and urban Alberta. Briefly, data were collected over four waves (covering the years 2006 to 2011) on multiple factors theoretically linked to the etiology and natural progression of gambling habits. Random digit dialing (RDD) recruited participants from the general population in Alberta as well as a proportion of individuals (n = 524; 29%) who were likely to develop gambling problems during the longitudinal follow-up period (individuals who were above the 70th percentile in gambling expenditure or frequency based on national population data). Participants completed a battery of self-report and administered tests covering gambling, substance use, personality, intelligence, mental health, life events, and social environment. Fourteen months separated each period of data collection. The current study sample consisted of adults who reported gambling at time 1 and had valid data for gambling activity and harms for at least one post-baseline assessment (N = 780). The rate of attrition between the first and last waves of data collection was 24%.

The Quinte Longitudinal Study (QLS) was initiated at the same time as the LLLP [35]. It recruited 4123 Ontario adults from the Quinte Region in southeastern Ontario, Canada. The time frame (also 2006 to 2011), goals, and content of the baseline and follow-up assessments were very similar to the LLLP. Sampling was also done via RDD within the Quinte region. A similar proportion (26%) of adults with at-risk gambling characteristics (defined as spending at least 10 CAD per month on gambling in the past year or engaging in slot machines or horse racing) was recruited. Both studies employed a similar rate of remuneration for participants (QLS: 50CAD vs. LLLP: 75CAD for initial assessment). The rate of attrition between Time 1 and 5 in the QLS was 4%, much lower than the LLLP. Assessment intervals within the QLS were separated by 12 months. The study sample comprised 3054 adults who reported gambling at time 1 and had valid data for the measures of interest for at least one post-baseline assessment. Both studies used a combination of in-person and online data formats to collect the data used in the present analysis. Most of the gambling measures were collected using web surveys, a method thought to enhance the honesty of participant answers to sensitive questions such as gambling losses and psychosocial harms [40].

The QLS and LLLP used the same measures for assessing intensity and breadth of gambling habits and for gambling problems. The core questions derived from the Canadian Problem Gambling Index (CPGI) [41], a self-report survey designed to collect descriptive information on gambling habits in population studies. The CPGI collects detailed information on participant engagement with the most common games of chance available in Canada including: lottery tickets; instant win tickets; electronic gambling machines (EGM); casino table games; games of skill for money against other people (cards, pool, etc.); sports betting; horse or dog racing, and; other forms of gambling including online games. In terms of gambling expenditures, QLS and LLLP participants were asked to estimate over the past year the amount spent on each form of gambling in a typical month. The question wording conformed to the recommended standard for producing the most reliable estimate of actual expenditure [42]. The questions used in the present study asked people “Roughly how much money do you spend on [gambling type] in a typical month?” (‘spend’ means how much you are ahead or behind, or your net win or loss in an average month in the past 12 months). The question is repeated for all types of gambling reported by the participant in the past year, a method that is far superior than global estimates for all gambling activities.

The total expenditure on all forms of gambling was estimated by summing the expenditures for the individual gambling formats. Both self-reported losses and wins were considered in the calculation of total expenditure. Due to the presence of several extreme outliers, monthly expenditure was winsorized; values exceeding the 99th percentile for the distribution were replaced with the next lowest value (3700CAD per month). The percent of income spent on gambling was calculated by dividing the total expenditure for the month by the participant’s gross monthly household income (to a maximum of 100%).

Frequency of gambling was also assessed separately for each gambling format. The 7-point categorical scale [41] used in each study was converted to a quantitative scale to estimate number of gambling days each month. For the LLLP, the conversion factor was: 1–5 times/year = 0.25 days; 6–11 times/year = 0.5 days; 1 time/month = 1 day; 2–3 times/month = 2.5 days; once per week = 4 days; 2–6 times/week = 16 days, or; daily = 30 days. The conversion factor used in the QLS was: less than once a month = 0.5 days; once a month = 1 day; 2–3 times a month = 2.5 days; once a week = 4 days; 2–3 times a week = 10 days; 4 or more times a week = 16 days. Overall frequency of gambling was calculated by summing the frequency values for the individual gambling formats resulting in a value ranging from 0 to 30 times per month.

Gambling activity was assessed at each assessment interval. The measures of gambling intensity—frequency of gambling, amount spent per month in Canadian dollars, and expenditure on gambling as a proportion of family income—were converted into dichotomous variables by applying by the low-risk gambling limits established through previous research [36]. Gamblers who exceeded any of the low-risk limits at time 1 (frequency ≥ 8 times per month; expenditure ≥75CAD per month; percent of income ≥1.7%) were deemed to be gambling above the low-risk threshold. Although these three dimensions of gambling activity are correlated, each independently predicts harm from gambling [43]. Each participant was assessed on the total number of low-risk limits exceeded at time 1 (range 0 to 3).

Gambling related harm was the primary outcome of interest. All harms were assessed using the Problem Gambling Severity Index (PGSI), a nine-item scale from the CPGI that assesses consequences and behavioural symptoms of problem gambling in the past 12 months [41]. The PGSI has well-established psychometric properties [44]. At the time of these studies, there was no validated measure of gambling-related harm. Although the PGSI is a measure of problem gambling severity, it assesses harms as well as behavioural symptoms. Using the Problem Gambling Severity Index (PGSI) we defined two levels of harm, both scored dichotomously. Moderate level of harm was defined as reporting at least two consequences from the PGSI items addressing feeling guilty, betting more than one can afford, recognition of a problem, health problems, financial problems, being criticized by others, and borrowing money to gamble. This is the same definition of harm employed in several investigations on low-risk gambling limits [11, 12, 22]. In previous work we found this definition of harm to have the best psychometric properties (highest area under the curve, sensitivity and specificity values) compared to alternative harm definitions [43]. We also studied problem gambling as an outcome, defined as scoring five or higher on the full PGSI scale Although eight was the original cut-off for identifying problem gambling, research has shown that use of this cut-off has good correspondence to clinically assessed problem gamblers in treatment, but poor correspondence to clinically assessed problem gamblers in the general population [45, 46]. More recent studies indicate score of five or higher demonstrate high sensitivity, specificity, and overall classification accuracy in detecting problem gamblers compared to clinician assessments [45, 47, 48]. The PGSI was normed on a small group of treatment-seeking problem gamblers [41] who tend to have a more pervasive and severe set of problems compared to problem gamblers in the general population. Lowering the PGSI threshold to ≥5 has been shown to successfully capture both treatment-seeking and non-treatment seeking problem gamblers [45].

Because our results are intended to inform prevention initiatives, we only included modifiable risk factors in the modelling. Covariates were selected based on being a changeable behaviour or a treatable comorbidity and a strong relationship with problem gambling. For the latter criteria we selected variables that were shown to be predictors of future problem gambling in the multivariate model of etiology developed from the LLLP and QLS datasets [34, 35]. We also used the results of a recently completed meta-analysis of problem gambling risk factors [49]. Covariates included in the model consisted of the presence of a comorbid mental illness and substance use disorders (SUD), number of stressful life events in the past year, participation in continuous types of gambling (EGMs or casino table games), and number of gambling cognitive fallacies. To assess comorbidities, both studies included a structured diagnostic interview used extensively in population research [50] to assess the presence or absence of DSM-defined major depression, generalized anxiety disorder, panic disorder, obsessive-compulsive disorder, alcohol use disorder and drug dependence in the past 12 months. The Life Events Questionnaire (LEQ) [51] assessed the number of significant life events (e.g., loss of employment) that may have occurred in participants in the past 12 months. A total of 58 different life events across nine categories, including relationships, work, and finances are assessed by the LEQ. The total score provides a general measure of number of stressful events in the past year. The 10-item Gambling Fallacies Measure (GFM) [52] was used to assess common cognitive distortions about gambling such as misunderstanding the random nature of games and believing that one can win by using a system. Higher scores on the GFM reflect fewer cognitive gambling distortions.

In addition to the above time-fixed covariates, we also included in the Cox model two time-dependent variables that were measured at each wave. Because gambling behaviour was expected to vary over time, we included number of different game formats played at each time point (range 1 to 8) as a predictor. Playing EGMs or casino games elevates an individual’s risk of gambling problems [22, 53]. Therefore, playing EGMs or casino games between assessment intervals was another time-dependent covariate. Although internet gambling has also emerged as a high-risk gambling format [54], the proportion of the LLLP and QLS samples who reported engaging in internet forms of gambling was too small (< 5%) to warrant inclusion as a predictor in the models.

Because of differences between the LLLP and QLS in both the assessment interval spacing (14 months vs. 12 months) and number of follow-up waves (4 vs. 5) it was not possible to merge the datasets for the Cox hazards models. Although sampling weights were available for the LLLP, the QLS had no weights therefore all analyses were conducted on unweighted data. Censored cases were managed the same way in both samples. Both datasets contained left and right-censored data on gambling harms. Because our interest was predicting the incidence of new gambling-related harm, participants who were assessed as problem gamblers at time 1 (left-censored data), were excluded from the samples. Right censored cases consisted of participants who dropped out and participants who remained free of harm at the last assessment wave. In keeping with standard procedures for survival analysis, right censored case were retained by coding dropouts with the last available data point. Missing data among the remaining variables was minimal (< 1% of cases).

Separate Cox proportional hazards models were run on the two outcome variables of interest: moderate harm, consisting of reporting two or more consequences from the PGSI, and new onset of problem gambling (PGSI score ≥ 5). Sequential models were run with the presence of a mental disorder, presence of a SUD, GFM total, LEQ total, EGM or casino game play at each time period, and number of different gambling formats played at each time period entered as a block first. The main predictor of interest, number of low-risk gambling limits exceeded at time 1 (range 0 to 3), was entered last to assess the unique importance of gambling above the recommended limits after controlled for other, modifiable risk factors. All analyses were conducted with SPSS Version 25. Assumptions of proportionality of hazards and non-linearity were tested for each model and were found to be within acceptable limits. Among the continuous covariates, LEQ scores displayed a modest positive skewness (1.76). Because transforming the variable did not significantly improve the distribution the analysis was conducted on the original data.

Table 1 compares the demographic characteristics of the LLLP and QLS samples. As previously reported [36] the gambling characteristics of the two samples were very similar. The most notable differences in the samples were age and marital status. The LLLP sample was on average 6 years younger and had a much higher proportion of single individuals. The 12-month prevalence rates of mental health disorders was high in both samples—18 and 33% for the QLS and LLLP, respectively. Similarly, the proportions meeting diagnostic criteria for a substance use disorder was 8 and 11% in the QLS and LLLP samples, respectively. Forty-three percent of QLS participants were below all low-risk gambling limits at time 1 compared to 63% of the LLLP sample. Exceeding 1, 2, and 3 low-risk limits was observed in 23, 19, and 14% of QLS gamblers, respectively. Within the LLLP sample, 14, 16, and 6% of gamblers exceeded 1, 2, and 3 low-risk limits, respectively.

The results of the Cox regression models are shown in Table 2. Using moderate harm as the outcome, the model on the QLS sample was highly significant before the addition of exceeding the low-risk gambling limits as a predictor (χ² = 118.59; p < .001). The addition of number of low-risk limits exceeded as a predictor resulted in a significant change in the model χ² value (12.59; p < .001). In the final model, mental health problems, stressful life events in the past year, number of game types played each year, playing EGMs or casino games, more cognitive distortions, and exceeding the low-risk limits all were independently predictive of future gambling harm. Each additional low-risk limit exceeded increased the cumulative probability of harm by 24%. Figure 1 displays the difference in the longitudinal survival proportions for gambles who exceeded 0, 1, 2, and all low-risk limits. The five-year cumulative survival rate for QLS gamblers who stayed below all low-risk limits was 95%. For gamblers who exceeded all the low-risk limits the cumulative survival rate was 83% at the last assessment.

The same pattern of results was evident when problem gambling (PGSI ≥5) was used as the outcome. The model without exceeding the low-risk gambling limits was significant (χ² = 107.07; p < .001). Adding the low-risk gambling limits predictor produced a significant change in the model χ² value (11.58; p < .001). For every low-risk limit exceeded the probability of meeting criteria for problem gambling increased by 40%. In total, 54 gamblers become problem gamblers (1.9% of the baseline sample) during one of the four follow-up time points. As shown in Fig. 1, exceeding two low-risk limits had virtually the same survival pattern as exceeding three limits. At the end of the study period, the cumulative survival rate for avoiding problem gambling was 99% among individuals who adhered to all low-risk limits and 93% in gambling who exceeded all three limits.

Modelling conducted on the LLLP sample yielded similar results. The model predicting moderate harm was significant before the addition of exceeding the low-risk limits as a predictor (χ² = 55.5, p < .001). The change in the model χ² was significant (5.17; p < .05) after adding the low-risk limits factor. Significant predictors of moderate harm in the final model were experiencing a mental disorder, endorsing more cognitive distortions, more stressful live events, and exceeding the low-risk gambling limits. For each additional low-risk limit exceeded there was an increase in the cumulative probability of harm by 27%. The five-year cumulative survival rate among individuals who stayed below all the low-risk limits was 96% compared to 85% among gamblers who exceeded all limits.

Figure 2 illustrates how participants with an anxiety or depressive disorder at time 1 took less time to experience moderate problems from their gambling compared to persons without a mental disorder. The Cox regression model with problem gambling as the outcome measure did not show a significant incremental change in χ² after adding the low-risk gambling limits factor (χ² = 3.45; p > .05). In the overall model, the predictors of future harm were baseline mental illness, playing EGMs or casino games, and cognitive distortions. A smaller number of incident cases of problem gambling appeared in time periods 2 to 4 in the LLLP sample (n = 38; 4.9% of baseline sample).

This study has several limitations that require acknowledgement. Foremost, the two study samples were drawn from specific areas of Canada. The results may not generalize to all provinces or to other countries. The higher attrition rate of the LLLP led to a smaller longitudinal sample and less robust results compared to the QLS. The diminished statistical power and fewer number of participants who developed serious problems by study end contributed to the lack of significance for low-risk limits within the model predicting problem gambling. We included in the Cox predictive models only a subset of variables that may be associated with gambling problems and did not include demographic factors. Our rationale was to focus on modifiable risk factors with variables pre-screened to have a strong correlation of gambling-related harms, with the intention to provide information useful for prevention efforts related to responsible gambling initiatives. Nonetheless, there may be several unmeasured factors that could have as much or even greater importance in shaping an individual’s gambling behaviours and risk for future harm. Although it has been subject to extensive validation studies [44, 45], the PGSI, is nonetheless a self-report screening measure for problem gambling. The PGSI is not a clinical tool and we have no means to confirm if all individuals scoring above our cut-off of five would be diagnosed with DSM-defined gambling disorder based on the gold standard of a clinical interview.

Exceeding the low-risk gambling limits significantly predicted both moderate harm and problem gambling, however the overall effect size in terms of the odds ratios and survival rates was small. In the predictive models, having a mental disorder had a stronger influence on future harm than the number of low-risk gambling limits exceeded. The oversampling of at-risk gamblers in both studies may account for this finding in addition to the high prevalence of mental disorders at baseline. For gamblers who exceeded all low-risk limits the five-year cumulative survival rate for avoiding moderate harm was 83%. When problem gambling was the outcome, the survival rate is 93%. Most gamblers who exceed the low-risk limits don’t develop problem gambling or experience moderate levels of harm. This is not an uncommon finding in predictive models involving health behaviours. Most individuals who exceed the low-risk drinking limits do not develop serious health consequences [59]. However, a much larger number of individuals who are moderate alcohol consumers experience harm compared to problem drinkers, a phenomenon that has been labelled the ‘prevention paradox’. Research conducted on a population samples in the United Kingdom and Australia has found evidence of the prevention paradox in gambling. Specifically, the total number of people who report harm at moderate levels of gambling far exceed the number of problem gamblers [25, 60]. Responsible gambling approaches that adopt a population strategy (e.g., public health campaigns that aim to correct erroneous believes about the odds of winning; promotion of safe spending limits) to reduce the average level of consumption in low to moderate risk gamblers are more likely to impact a larger portion of the population than approaches that target highly active, problem gamblers (e.g., casino self-exclusion programs, pop-up messages on EGMs that only appear after an hour of continuous play).

The low-risk gambling limits, like the Canadian low-risk limits for drinking, are intended to educate gamblers on the relative risk of exceeding certain thresholds. The comparison group is gamblers who remain below the limit (e.g., spend less than 75CAD per month) rather than individuals who abstain or gamble at very low levels. There is debate among researchers on the shape of the gambling dose-response curve with some advocating the true shape is linear or r-shaped suggesting that there is no basis for setting a discrete cut-off—any level of gambling above total abstinence is harmful [61]. In that case, an absolute risk approach might be better for setting a limit. With an absolute risk approach, a jurisdiction first must define how much risk would be considered acceptable to the general public. This is the approach taken by Australia for their version of the low-risk drinking limits. The drink thresholds are based on the risk of premature death from alcohol-related diseases being 1 in 100 or greater. A criticism of the absolute risk approach is that the determination of what is considered an acceptable level of risk is subjective. There are no standards for setting absolute risk levels for addictive behaviours [62, 63]. It also appears the public has a higher tolerance for voluntary behaviours such as drinking and gambling, than involuntary risks such as exposure to radiation or contaminants in air or water supplies [64] where the typical absolute risk threshold is 1 in a million.

The present study should help to inform the expansion of responsible gambling advice to include quantitative thresholds. Canada’s work in developing low-risk gambling limits has since expanded to include many international collaborative partners who have expressed the desire to adopt the limits for their own country [14]. With endorsement by multiple stakeholders including treatment providers, responsible gambling advocacy groups, and government regulators, low-risk gambling limits will eventually enjoy broad circulation with other public health initiatives.