The relationship between problematic gambling severity and engagement with gambling products: Longitudinal analysis of the Emerging Adults Gambling Survey

Aims To measure the association between problem gambling severity and 19 different gambling activities among emerging adults (aged 16–26). Design An online non-probability longitudinal survey collecting data in two waves: wave 1, July/August 2019; wave 2, July/October 2020. Setting Great Britain Participants A total of 2080 young adults participating in both waves. Measurements Problem gambling scores were collected using the Problem Gambling Severity Index (PGSI). Binary variables recorded past year participation in 19 different gambling forms, ranging from lotteries to online casino and gambling-like practices within digital games (e.g. loot box purchase, skin betting). Controls included socio-demographic/economic characteristics, the Eysenck Impulsivity Scale and the number of gambling activities undertaken. Findings Zero inflated negative binomial model lacked evidence of an effect between past year participation in any individual activities and subsequent PGSI scores. However, negative binomial random effects models for current gamblers (n = 497) showed that skin betting (incidence-rate ratio [IRR] = 2.32; 95% CI = 1.69–3.19), fixed odd betting terminals (IRR = 2.21, 95% CI = 1.61–3.05), slot/fruit machines (IRR = 1.43, 95% CI = 1.07–1.91), online betting on horse/dog races (IRR = 1.53, 95% CI = 1.17–2.00) and online betting on non-sports events (IRR = 1.44, 95% CI = 1.11–1.89) were associated with increased PGSI scores. Online casino gambling had a significant interaction by wave; the impact of online casino betting in wave 2 on PGSI scores increased by a factor of 1.61. Conclusions Past year participation of young adults (aged 16–26) in certain forms of gambling does not appear to be associated with future Problem Gambling Severity Index scores. Among young adults who are current gamblers, past year participation in certain land-based (e.g. electronic gaming machines) and online forms (e.g. skin betting) of gambling appears to be strongly associated with elevated Problem Gambling Severity Index scores.


Severity Index (PGSI). Binary variables recorded past year participation in 19 different
gambling forms, ranging from lotteries to online casino and gambling-like practices within digital games (e.g. loot box purchase, skin betting). Controls included socio-demographic/economic characteristics, the Eysenck Impulsivity Scale and the number of gambling activities undertaken.

Conclusions:
Past year participation of young adults (aged [16][17][18][19][20][21][22][23][24][25][26] in certain forms of gambling does not appear to be associated with future Problem Gambling Severity Index scores. Among young adults who are current gamblers, past year participation in certain land-based (e.g. electronic gaming machines) and online forms (e.g. skin betting) of gambling appears to be strongly associated with elevated Problem Gambling Severity Index scores.

K E Y W O R D S
Covid-19, emerging/young adults, gambling, harms, longitudinal, products

INTRODUCTION
Understanding the relationship between certain forms of gambling and gambling harms is a critical policy consideration. In Britain, legal forms of gambling range from lotteries to electronic gaming machines (EGMs) and online casino products. As each form of gambling is associated with a different set of structural characteristics, and are provided in different contexts, debate has focused on the extent to which each type may be more (or less) harmful and their consequent association with problematic gambling [1][2][3][4]. Engagement in continuous forms of gambling activities, like EGMs or online equivalents, has been consistently associated with gambling problems [3].
A range of potentially interlinked explanations have been suggested for these patterns. These include selection effects, whereby those with greater vulnerability gravitate to certain forms of gambling; exposure effects, whereby those exposed to certain forms of gambling are at greater risk because of the structural characteristics of the gambling format and broader commercial effects, whereby the commercial actions of the industry, especially pertaining to access, availability, marketing and promotion, increase risk propensity [3,5,6].
Conversely, some have argued that the range and breadth of gambling involvement, rather than engagement in specific activities, are more useful predictors of harms (termed the involvement hypothesis) [4].
Yet few studies have focused on the relationship between engagement in certain forms of gambling and problematic gambling among the 'emerging adult' age cohort specifically, although associations have been noted between online gambling and problematic gambling for young men and scratchcard purchase and problematic gambling [7][8][9]. Given the changing landscape of gambling provision in Britain, and continuing debate about the role of gambling-adjacent activities (e.g. loot boxes) in the production of gambling harms [10][11][12], this warrants attention to map key associations and empirically test some of the suggested mechanisms underpinning observed associations (such as selection effects etc).
Emerging adults (those approximately 18-24 years old) have been identified as a group at heightened risk of experiencing problem gambling. Forrest and McHale [13] showed that rates of problem gambling increased significantly between the ages of 17 and 21, leading them to suggest that extra measures could be warranted to protect emerging adults from harms during this period of increased vulnerability.
This was a key question posed by the British government in their review of the 2005 Gambling Act [14]. Furthermore, according to Arnett [15], who coined the term 'emerging adult', this age group are demographically distinct with a greater propensity for risk-taking behaviour, including impulsivity, and engaging in sensation-seeking experimentation before settling into adult roles and responsibilities [15]. These are known risk factors for the experience of problem gambling.
This study explores how gambling activities, but also newly emerging gambling-adjacent activities within digital games, are associated with problem gambling severity. Loot boxes are one example of 'gambling-adjacent' activities, described as psychologically akin to gambling [16]. They are items that may be bought for real-world money containing randomized contents whose value is uncertain at the point of purchase [17]. The betting of skins is another example.
Skins are decorative items bought or won within digital games that can be traded or bet through a series of connected marketplaces [18].
Because skins can be used as collateral for betting on various websites and because loot boxes mimic variable reward mechanisms has led to wide-scale debate about whether these represent new forms of gambling [18]. Few studies have measured both gambling-adjacent and for-money gambling activities concurrently, although researchers have increasingly argued they should be considered in parallel [19].
Using a longitudinal survey of emerging adults (aged 16-26) this paper conducts exploratory analysis to: a. explore the relationship between problem gambling scores and engagement in different forms of gambling (including gamblingadjacent activities: loot box purchase; skin betting), while controlling for demographic/socioeconomic status and gambling involvement, and, b. examine if and how these associations changed over time.

Design
The Emerging Adults Gambling Survey is a longitudinal study of young people aged 16 to 26 living in Britain. The study's primary aim was to examine individual gambling trajectories over time; sample size calculations were based on being sufficient to estimate change in gambling behaviours between waves. Assuming a between wave correlation of 0.5, the study was designed to be able to detect changes in problem gambling behaviours of ±0.3 percentage points (at 80% power). The survey protocol was pre-registered [20]. Data analysed here include participants from wave 1 (n = 3549) and wave 2 (n = 2080). Participants were drawn from YouGov's online panel of over 1 million people living in Britain [21,22]. Participants were eligible if they were aged 16 to 24 (at wave 1), living in Britain and had not participated in another YouGov study on gambling in the past year at baseline. E-mail invitations were sent by YouGov to selected panellists inviting participation in a survey, without advertising its content, and to click through to the bespoke study. The first page of the bespoke survey described the study's aims and objectives, including that this was a longitudinal survey and that we would be wishing to recontact them 1 year later and obtained consent. In wave 1, 93% of people who accessed this page went on to complete the survey. Participants received YouGov points (equivalent to 50p in value) for taking part.

Controls
Impulsivity was measured using a shortened form of the Eysenck Impulsivity Scale validated for use among adolescents [24][25][26].
Responses to seven impulsivity statements were recorded on a five-point scale with response options ranging from very true (1) to not at all true (5). Impulsivity scores were computed as the average of the seven questions. Impulsivity was collected at wave 1 only and scores used as fixed effects in the model (wave 1 mean Responses were grouped by whether at least one parent had a degree or higher or whether both parent's qualifications were lower than degree level. Marital status at wave 2 was asked and grouped into: lives with a partner; has partner, but does not live with them; single.
Finally, personal gross income at wave 2 was obtained and grouped by: <£5000; £5000-<£20 000; £20 000 or more. Missing data for income was high and coded as a dummy category.

Analyses
Frequencies described the characteristics of the sample and participants gambling behaviours (Tables 1 and 2). The relationship between gambling activity and PGSI score was first explored using zero-inflated negative binomial models ( Table 3). The model uses wave 2 PGSI score as the dependent variable and a set of binary indicators showing participation in each gambling activity (including skin betting and loot box purchase) in the past 12-months at wave 1 as predictors.
The model indicates whether participation in a specific activity at wave 1 is associated with the PGSI score at wave 2, when controlling for gambling involvement (measured by the number of activities undertaken); other demographic/socioeconomic controls and PGSI score at wave 1.
The relationship between participation in each gambling activity and PGSI score was further explored for those who gambled at both waves using negative binomial models. This approach looks at the associations between current gambling activities and current PGSI scores. It also describes whether the nature of this relationship, in terms of direction and magnitude, changed over time. Unlike the zeroinflated models, the PGSI score used as the dependent variable in this model is for gamblers only and does not contain excess zeros, hence the first step of the modelling-the logit model used to estimate whether or not a zero PGSI score was because of an absence of gambling-is not required. The remaining count model is a negative binomial model. This less complex approach allows random effects to be included in the model to account for the panel structure of the data, incorporating a longitudinal element into the analysis. *Exceptions to this were National Lottery and Scratchcards, which at the time of data collection were legally available to those ages 16 and over.  [27,28].
Random effects models account for the longitudinal data design with repeated measures taken from the same individual over time.
The model assumes independence exists between the different individuals in the sample, but not between time points for the same individual and that the over-dispersion in the dependent variable varies between individual, but is consistent within individuals over time.
These assumptions were tested using a likelihood-ratio test. A significant test result (P < 0.05) confirmed these assumptions.
Two regression models were run. The first (Table 4a and 4b) included information on all 19 gambling activities ( Table 2). Interaction terms between each activity and survey wave were included in the regression (Table 4a). The second model (Table 5a and  To test the sensitivity of the models, data were split at random into two halves and the modelling repeated. Despite the low prevalence of some gambling activities and smaller sample sizes, the IRRs T A B L E 2 Gambling behaviour.  Tables 1 and 2   This pattern was true regardless of whether number of gambling activities was included as a control or not (results available at https:// osf.io/6sem8).

Multi-variate analyses
Looking at current gamblers only, Table 4a shows that past year gambling on fixed odd betting terminals (FOBTs), skin betting, online betting on horses/dog, online casino/slot gambling, playing poker at a pub/club; purchasing loot boxes or playing fruit/slot machines were each significantly related to having a higher PGSI score across both waves. Table 4a were significant (P < 0.05). These were lotteries and playing poker in a pub/ club, suggesting that the relationship between these specific activities and PGSI scores changed between the two waves. With respects to lotteries, the P value for the main effect was 0.705, (IRR = 1.07, 95% CI = 0.76-1.49), yet the interaction term was significant (P = 0.001).

Two interactions terms for individual activities in
The IRR for the interaction term was 0.51 (95% CI = 0.34-0.76) suggesting that impact of playing the lotteries on PGSI scores decreased at wave 2 by nearly half. By contrast, the main effect for poker was that playing poker at wave 1 increased PGSI scores by a factor of  Notably, analyses for current gamblers took into account broader gambling involvement and a range of socioeconomic and demographic vulnerabilities as well as impulsivity. This suggests that factors other than gambling involvement or vulnerable groups being attracted to these forms may explain the association between these gambling forms and problem gambling severity. This might include the structural characteristics of these gambling formats and/or commercial or regulatory practices governing their provision and promotion.
With respect to changing behaviours over time, the significant interaction term for online casino/slot style games is notable.
Although associations between online casino/slot engagement and PGSI scores were statistically inconclusive at wave 1, the impact of playing in online casinos in wave 2 on PGSI scores increased by a factor of 1.61, indicating that betting on online casinos has a larger impact on PGSI scores in wave 2 than wave 1. Wave 2 data was collected in July-October 2020, during the COVID-19 pandemic. During this time, online gambling firms, particular online casino/slot games reported growth in the number of active players and in revenues [34].
Concern was raised about the potential for some people to engage more problematically with these products during this time [35]. The observed interaction in this study may reflect these broader processes.
Finally, the largest association was observed for skin betting.
To date, much attention has been given to the relationship between loot boxes and problem gambling scores, with less focus on other gambling-like mechanics within the digital game ecosystem [11,12,16,17]. Skin betting involves using items from digital games as collateral to wager. Our results show that emerging adult gamblers who engaged in these practices increased their PGSI scores by a fac- This study has a number of limitations. First, the YouGov panel is a non-probability sample with attendant issues of generalisability.
Nevertheless, compared with other sample frames, it has good sample coverage, including young people both in and out with full time education (unlike sample drawn from Higher Education Institutes or the Postcode Address File, which excludes those living in halls of residences). Studies have shown that although online non-probability methods should not be used for prevalence estimates, they can perform better (although still not without some issues) when focusing on the relationship between variables, which this study does [36].
Second, attrition between waves was high, although commensurate with other longitudinal studies of young people [13]. Relatedly, because of attrition, the sample size was smaller than hoped. Engagement in some of the forms of gambling reported here are relatively rare. Therefore, non-association should not be taken to mean that these things are not related, but rather that the study was underpowered to examine these. Finally, this analysis focuses on whether people had participated in gambling in the past year or not. It would be useful to explore how changes in gambling frequency for each activity also relate to problem gambling scores. Because of changes to the wave 2 questionnaire necessitated by the COVID-19 pandemic, it was not possible to include this here. Future studies should assess this.

CONCLUSION
Among current gamblers, both land-based and online forms of gambling were strongly associated with elevated PGSI scores among young adults. Skin betting, rather than the purchase of loot boxes, emerged as one of the strongest predictors of elevated PGSI scores.
Although policy attention should focus on online gambling and gambling-adjacent forms, EGMs in land-based venues should continue to command attention.