Simulated gambling consumption mediation model (SGCMM): disentangling convergence with parallel mediation models

ABSTRACT

Simulated Internet gambling (SG) raises concerns, especially with regard to adolescents, because it may facilitate monetary gambling or problem gambling (PG). Only scarce research examined mechanisms via which SG prospectively impacts gambling onset or PG. The study fills this gap by using longitudinal survey-data (12 months; N = 1,178 pupils from Northern Germany; M = 13.6 years; 47.5% male). Parallel mediation models in different types of SG (via video games, apps, social networks, demo games) were applied to disaggregate bivariate associations of last year participation at the first stage of the survey with two outcomes: (1) PG, (2) gambling onset, both after 12 months. Mediating effects via patterns of consumption, cognition or other problematic online behaviors were examined. Both outcomes were impacted by different types of SG via quite different mechanisms: SG impacted PG mostly via indirect effects of gambling depth (maximum gambling frequency), irrational cognitions, and problematic Internet gaming (problematic Internet use revealed to be a PG decreasing mechanism). Onset was impacted via increased perception of advertising (only for SG in video games and social networks). Prospective parallel mediation models reveal relevant etiological pathways of SG on PG or gambling onset in a more exhaustive way than prior research.

KEYWORDS:

• Gambling
• video games
• addiction
• advertising
• behavior

Introduction

Technological progress, dissemination of Internet-based entertainment, and decreasing costs of personal computers and mobile devices dominated past decades. In this context, experts raised concerns about monetary and non-monetary forms of gambling available for the particularly vulnerable population of young people, for instance, in video games or demo games of real online casinos (King et al., 2010). Emerging concepts of non-monetary gambling include ‘social casino games’ and ‘simulated gambling on the Internet’ (Armstrong et al., 2018; Hayer et al., 2018, 2019; King, 2018). In contrast to monetary gambling, both do not require obligatory use of money, but are quite similar to classic forms of gambling (virtual currency, perceived outcome of chance). However, due to the lack of mandatory real-money stakes and wins, legal definitions of gambling are circumvented. Because of their similarity, both concepts of non-monetary gambling are termed simulated gambling (SG) in the following. Particular concerns (Armstrong et al., 2018; Hayer et al., 2018; King, 2018) are raised due to (1) unrestricted availability of gambling contents to minors, (2) targeted advertising for vulnerable customers (adolescents, actual or remitted problem gamblers), (3) cognitive distortions through unrealistic payout rates, (4) group pressure through interactive elements of social networks like dashboards, and (5) unnoticed entry of gambling contents and mechanics into everyday life (changing attitudes, normalization of gambling).

Initially, free-to-play games often generate turnover via a small group of strongly involved customers, who spend high amounts of money for virtual currency, virtual goods, shortening of waiting periods, or other in-game advantages. These phenomena are wide spread in SG activities (Armstrong et al., 2018; Hayer et al., 2018) as well as online gaming activities – herein also known as ‘predatory monetization’ schemes (King, 2018; King & Delfabbro, 2018, 2019). Prototypes of voluntary in-game purchases or ‘micro transactions’ are ‘loot boxes’, which utilize randomness or targeted payout schedules (e.g. conditioned by prior behavior) for virtual goods, which are progressively blurring the lines between gambling and gaming. Opportunities to trade and re-convert items into real money (King, 2018) or to place bets on eSports with virtual items (Greer et al., 2019; King, 2018; Macey & Hamari, 2018) exacerbate this issue and challenge efforts of adequate regulation.

A growing body of research in cross-sectional studies (Gainsbury et al., 2014; King et al., 2014; Veselka et al., 2018) shows bivariate associations of participation in SG and problem gambling (PG) as well as associations of PG and spending in SG (King et al., 2016) or on loot boxes (Zendle & Cairns, 2018). Moreover, several prospective studies showed a predictive value of specific SG activities for gambling onset among minors (Dussault et al., 2017; Hayer et al., 2018) and adults (particularly when paired with micro transactions: Kim et al., 2015) as well as a predictive value of problematic video gaming for PG (Molde et al., 2019).

Methodological challenges

Despite these achievements, research on SG and convergence of gambling and gaming can still be regarded as fragmentary. For example, Gainsbury (2019) recently criticized (1) the lack of representative samples, (2) sparse longitudinal research and an overexpansion of cross-sectional studies with causal interpretations that (3) fail to control the association of SG and PG for the biasing effect of interest in gambling or engagement in real gambling. Similar methodological issues emerged in current research on bivariate associations of (1) online gambling participation, (2) gambling intensity, and (3) PG (Gainsbury, 2015, 2019). In a literature review, Gainsbury (2015) showed that all three facets are highly associated and prospective empirical findings (Williams et al., 2015) demonstrated that several possible pathways emerge: (1) highly involved offline PG includes online gambling into an existing portfolio of risky products like electronic gaming machines, (2) online gambling proceeds PG and its structural characteristics alleviate or even cause PG, or (3) PG emerged first and high involvement in activities escalates online as well as offline. However, due to the complex interactions of environment, types of gambling, and individual gambling behavior (Blaszczynski, 2013; Shaffer & Martin, 2011), discussions about unique risk potentials of participation in online gambling remain controversial (Gainsbury, 2015; Gainsbury et al., 2019).

Against this background, different measures of individual gambling intensity emerged. One of them is the ‘breadth of involvement’ representing the number of used game types (LaPlante et al., 2011). Another construct is called ‘depth of involvement’ representing an aggregated measure of gambling intensity across all game types, for instance, stakes, bets, or active gambling days (LaPlante et al., 2014). When bivariate associations of participating in individual game types and PG are adjusted for the biasing effects of such measures of individual gambling intensity, prior positive associations may decrease substantially. Nevertheless, results only show that individual gambling intensity is an important component to consider in research on adverse relations of participating in specific game types and PG. But they do not answer the question how participation in a particular type explicitly impacts PG.

Explicating etiological pathways with multiple mediation models

Mediation analyses (Hayes, 2018) address this question of how with explicit estimates for effect sizes and confidence intervals of the indirect mediating mechanisms via which participation in particular game types impacts an outcome of PG. This information is of particular interest for prevention and regulation, because relevant indirect effects (i.e. suggested etiological mechanisms like advertising or micro transactions) can be addressed concisely to reduce potential risks associated with particular products of gambling, SG, or gaming. Recent research (Brosowski et al., 2020) on the ‘Gambling Consumption Mediation Model’ (GCMM) disentangled such bivariate associations of PG and participation in different types of gambling by explicating effect sizes and confidence intervals of plausible and simultaneously assessed etiological pathways (indirect effects or mediators). The applied mediating mechanisms were derived from current research on modeling gambling behavior: (1) demographic PG propensity, (2) number of gambling types gambled, (3) gambling frequency within and (4) beyond a type of interest as well as (5) usual spending within and (6) beyond a type of interest. Interestingly, the evaluated types of gambling showed two distinct profiles via which mediator they mostly impacted PG: Electronic gaming machines and scratch cards offline, live betting online, and poker both offline and online impacted PG mostly via gambling frequency within, whereas all other types mostly impacted via the number of game types played.

In comparison to common regression modeling, multiple mediation models generally provide a formalized framework to disentangle associations of game types and PG by explicating statistical impacts via indirect effects. Moreover, the applied analyses address methodological challenges of collinearity and circular reasoning (excessive involvement is not just a covariate to adjust for but part of the etiological model of PG).

The Social Gambling Consumption Mediation Model (SGCMM)

The core idea of the GCMM was to evaluate individual gambling products by (1) secondary data analyses of monitoring population surveys and (2) a formalized risk profile of mechanisms via which each type mostly impacts PG. Therefore, the GCMM included only demographic and behavioral variables which are mostly available in monitoring survey data. However, transferring the GCMM to SG (Social Gambling Consumption Mediation Model: SGCMM) has to incorporate established indirect behavioral paths of the GCMM via breadth of involvement (number of active game types in SG or gambling), depth of involvement (frequency within or beyond SG, maximum frequency across all gambling types) but also new suggested indirect behavioral paths via micro transactions within SG (Kim et al., 2015), as well as plausible cognitive mechanisms like the perception of advertising (Armstrong et al., 2018; Hayer et al., 2018), irrational cognitions about gambling (King, 2018), or other problematic online behaviors like excessive gaming or Internet use (Gainsbury, 2019; King, 2018; Molde et al., 2019). SG may increase gambling behavior via some or even all suggested mechanisms but only a simultaneous estimation of the indirect effects (mediators) facilitates evidence-based, precise regulation and prevention. For instance, increasing impacts on gambling via frequency or micro transactions within SG may be indicative of risk increasing game designs of SG, while increasing mechanisms of problematic online behavior may be indicative of a more general problem with excessive digital involvement. In contrast, irrational cognitions, perceived advertising and gambling behavior show a more specific reference to gambling products in a poorly regulated environment.

Aim of the study

First, representative estimates for last year participation of minors (12–17 years old) in different activities of SG in three cities in Northern Germany were presented to address the lack of representative estimates of SG-prevalence (Gainsbury, 2019). The major aim of the study was then to explicate plausible mechanisms via which SG may impact gambling onset or PG. For this, four SG types (in video games, in social networks, via apps, or via demo games), each differentiated by mode of access (stationary vs. mobile), were examined. Multiple mediation models were used to disentangle bivariate associations of last year participation in these game types at T0 with (1) an ascending level of PG symptoms after a period of 1 year (T1) or (2) gambling onset after a period of 1 year (T1).

Based on the literature review, the applied mediators at T0 included (1) measures of consumption intensity within and beyond (breadth and depth) the SG types of interest, (2) breadth and depth of involvement in real gambling at T0 (only for the outcome of PG), (3) indicators of problem behaviors other than gambling (Internet gaming, Internet use) as well as (4) other candidates of risk increasing mechanisms like irrational cognitions about gambling, micro transactions in SG or perception of advertising for gambling or simulated gambling. These results will add to a growing body of evidence about the convergence of gambling and gaming. They also address raised methodological concerns (Gainsbury, 2019) about a lack of representative samples as well as prospective studies that incorporate measures of gambling intensity and problem behavior other than gambling when analyzing the impact of SG activities on gambling onset or PG. In contrast to common regression models (e.g. Hayer et al., 2018), parallel multiple mediation models provide explicit information about size, direction, and confidence of indirect effects via which SG types of interest impact gambling onset or PG. This information is of particular interest for prevention and regulation, because indirect paths can be directly addressed (Brosowski et al., 2020).

Materials and methods

Sample and data collection

Questionnaires were administered to pupils in classes 6–10 as a classroom activity, which took approximately 30 minutes. Informed consent of parents was mandatory. Investigators read out instructions, distributed questionnaires, and remained in classrooms to answer comprehension questions. To avoid confusion over the terms SG and real gambling, precise definitions, examples, and core differences were given both in the standardized oral instructions as well as the written ones on the questionnaires. All procedures were in accordance with the 1964 Helsinki declaration and later amendments. Federal state supervisory school authorities approved the study. Data collection was done in two waves, with a time-lag of 12 months. Data were collected in 25 schools in the North German cities Hamburg, Bremen, and Lübeck. Schools were stratified by socioeconomic status and type of school and were selected randomly. Weightings for representative prevalence estimates were based on administrative statistics of the same year for age, gender, and school type. At T0, a total of N = 1,905 questionnaires (Hamburg: n = 435, Bremen: n = 964, Lübeck: n = 506) were included.

For mediation analyses, anonymized questionnaires were matched between T0 and T1 surveys by key variables (first two letters of the father’s and mother’s forenames, sex, class; in ambiguous cases by additionally comparing details of star sign, favorite club, and other preferences, including comparisons of handwriting for open-ended questions). Through this procedure, of the N = 1,905 valid questionnaires available from the first wave, it was possible to reliably match up N = 1,178 questionnaires from the second wave (match success rate: 61.80%). The matched sample included a total of n = 559 (47.45%) male pupils; 9.80% of all matched respondents showed a history of immigration (with a place of birth outside Germany). The mean age of pupils at T0 was calculated to be M = 13.60 years (SD = 1.40).

Scales and behavioral variables

Last year gambling was assessed in six types (online or offline): lotto, scratch tickets, sports betting, EGMs, poker, other card or dice games except poker.

Last year SG was assessed in four types: within video games, via apps, within social networks, via demo games. Each type was differentiated by mode of access (from home/stationary computer; on the way/via mobile device), resulting in a total of eight variables.

Gambling and SG were enquired on a five-step rating scale (never, less than once a month, 1–4 times a month, 5–8 times a month, more than 8 times a month).

There were strong correlations (r =.45–.68) between similar simulated games via different modes of access (see Appendix Table A3). Because of this, the two modes of access of each SG type were aggregated using the rounded mean function for the mediation analyses. Combining both modes of access reduced the total number of SG activities to four types. The five-step rating scale of last year SG participation was dichotomized (never; at least once) for prevalence estimates and to serve as a predictor in the following mediation models.

PG symptoms were assessed by a questionnaire for adolescents developed and validated in Germany (Fragebogen zu glücksspielbezogenen Problemen im Jugendalter FGP-J; Hayer et al., 2016) with a sum score across 19 items (categories: 0 = never, 1 = sometimes, 2 = often, 3 = very often; α = .91) and with ascending levels indicating larger problems. An item of the scale was for instance: ‘How often do you think about gambling?’

Problematic Internet gaming was assessed by a sum score across nine items (categories: 0 = not correct, 1 = hardly correct, 2 = rather correct, 3 = exactly; α = .85) recommended for Internet Gaming Disorder (IGD; Petry et al., 2014) and with ascending levels indicating larger problems. For example, the scale included the following item: ‘Do you spend much time thinking about gaming or planning to game?’

Problematic Internet use was assessed by a sum score across 14 items (Compulsive Internet Use Scale [CIUS]; Gürtler et al., 2015; categories: 0 = never, 1 = seldom, 2 = sometimes, 3 = often, 4 = very often; α = .90) and with ascending levels indicating larger problems. ‘How often do you continue to use the Internet despite your intention to stop?’ represented one item of this scale.

Perception of advertisements for gambling or SG was rated across seven items/situations like ‘via Facebook’ or ‘on TV’ (categories 0 = not at all, 1 = seldom, 2 = sometimes, 3 = often, 4 = very often; α = .79). All items were aggregated by mean; ascending levels indicating stronger perception.

Frequency of using one’s own money on micro transactions during the last year was assessed across three items/situations: (1) ‘for virtual currencies’, (2) ‘increasing the chance of winning’, (3) ‘shorten waiting periods’ (categories: 0 = never, 1 = sometimes, 2 = often, 3 = very often; α = .75). All items were aggregated by mean; ascending levels indicating more frequent spending.

Irrational cognitions about gambling were assessed by eight items from the Gambling-Related Cognition Scale (GRCS [Raylu & Oei, 2004]; referring to the control illusion subscale and the predictive control subscale, the latter reduced by two items; α = .78). An item of the scale was for instance: ‘Praying helps me win’.

Breadth of simulated gambling activities was assessed by the number of SG activities an individual was at least monthly involved in during the last year. Due to the aggregation of the different modes of access (see game types section), the score ranged from 0 to 4 types.

Frequency within a particular simulated type of gambling was operationalized by the original five-step rating scale of last year participation frequency, but was aggregated across both modes of access by mean (see above).

Frequency beyond a particular simulated type of gambling was operationalized by the maximum function across all other types of SG except the type of interest.

Breadth of gambling activities was assessed by the number of gambling activities an individual was involved in at least once during the last year. The score ranged from 0 to 6 types.

Depth of gambling activities was assessed by the maximum function across six gambling activities.

Data analyses

Data analyses were conducted with three different, successively reduced samples:

Last year participation prevalence rates for gambling and SG as well as Spearman-correlations among SG activities (Appendix Tables A1–A3) were estimated for the total T0 sample (N = 1,905).

In the matched prospective sample (n = 1,178), for each type of SG, a parallel multiple mediation model was computed to assess the impact of last year participation in a type of interest (yes, no) at T0 on the outcome of PG after 12 months (sum score, linear regression-based mediation).

In a subsample (n = 531) of the matched cases only individuals who did not participate in any type of gambling at T0 were included. The 531 cases split up into 140 starters who participated in any type of gambling at T1 and 391 constant abstainers who did not participate in any type of monetary gambling at T0 and also did not start monetary gambling at T1. In this subsample, for each type of SG activity at T0, a parallel multiple mediation model was computed to assess the impact of last year participation in the simulated game type of interest (yes, no) at T0 on the outcome of gambling onset after 12 months (binary outcome [yes, no]; logistic regression-based mediation). This subsample was already analyzed in prior research with regression models (Hayer et al., 2018), but multiple mediation analyses indeed provide a more exhaustive analytical approach than prior modeling.

The involved mediators of the models at T0 were:

(M1) Frequency within the type of interest,

(M2) Frequency beyond the type of interest,

(M3) Breadth of involvement in SG,

(M4) Perception of advertising,

(M5) Realization of micro transactions,

(M6) Irrational cognitions,

(M7) Score of compulsive Internet use (CIUS),

(M8) Score of Internet gaming disorder (IGD),

(M9) Breadth of involvement in gambling (only applied for PG, not onset),

(M10) Depth of involvement in gambling (only applied for PG, not onset).

A conceptual map of mediation models for each game type is illustrated in Figure 1. The a-paths (dashed lines) estimate the association of last year participation in a game type and proposed mediating mechanisms. The b-paths (permanent lines) estimate the association of the mediator with the outcome, while simultaneously controlling for all other variables (type, mediators, covariates) in the model.

Figure 1. Conceptual map of applied multiple mediator models with mediators (m) and predictor (x) assessed at T0 (SG = simulated gambling). Annotation: M9 an M10 (dashed boxes) were only included in the mediation models to predict problem gambling (y1 at T1), not to predict gambling onset (y2 at T1)

The product of a*b quantifies the indirect effect of participating in SG type x on PG via a specific mediator, for instance, micro transactions or depth of gambling involvement (single a- and b-paths are not presented because real mediation unfolds via the product a*b only). This quantification of particular indirect effects and their confidence intervals gives an explicit answer to the question via which etiological mechanism (mediator) participation in a game type of interest mostly impacts the outcome. The c’-path estimates the direct effect of participating in SG type x on PG when simultaneously considering all indirect effects of the mediators in the model. Furthermore, the total effect is the sum of the direct and all indirect effects on the outcome.

All data analyses were conducted with IBM SPSS Statistics Version 25. Multiple mediations were conducted with PROCESS v3.3 (Copyright [c] 2012–2019 by Andrew F. Hayes), a syntax oriented tool for mediation in SPSS (Hayes, 2018; model number 4). The number of bootstrap samples was set to 5,000 (seed random). Due to the exploratory nature of this study and the first time application of the SGCMM, p-levels were not adjusted for multiple comparisons. Missing values in the mediators (0–2.46%) were replaced by means.

Given the longitudinal matching-rate of 61.84% (1,178/1,905), the question arises as to whether nonparticipation gave rise to systematic distortion. Drop-out at T1 was significantly associated (p < 0.10; Chi Square) with the following five characteristics: (1) data collection in Hamburg (p < 0.001); (2) school type: Gymnasium (p < 0.001); (3) participation in SG in video games from home (p < 0.001); (4) immigrant background (p = 0.07), and (5) male sex (p = 0.07). Significant correlations with ordinal characteristics (Mann–Whitney) were: (6) school class (p < 0.001); (7) PG (p = 0.01); (8) video gaming breadth (p = 0.06), and (9) internet usage breadth (p = 0.098).

In a simultaneous model estimation of drop-out probability, only three significant predictors remained in the final model: (1) school class (OR = 1.49; 95% CI = 1.38–1.61); (2) data collection in Hamburg (OR = 2.39; 95% CI = 1.89–3.04) and (3) gymnasium (OR = 0.52; 95% CI = 0.42–0.65). This model correctly predicted drop-out in 66% of cases. Based on Ahern and Le Brocque (2005), drop-out probability served as covariate in all mediation models to control for attrition-bias. Moreover, the outcome of PG was associated with male gender and ascending grade. Therefore, both variables were additionally included as covariates. Gambling onset was not associated with any demographics and therefore no covariates were included into these mediation models.

Results

Participation in different game types at T0 (N = 1,905)

The weighted prevalence of gambling and SG, representative of the three cities, did not differ substantially from the non-weighted complete dataset (see Appendix Tables A1, A2). Therefore, only estimates of non-weighted total data are reported. In the non-weighted dataset 38.49% of the pupils reported of any gambling in the past year, with scratch tickets (18.72%), sports betting (14.78%), and card or dice games except poker (11.59%) being most prevalent. Poker (8.62%), EGMs (8.20%), and Lotto (5.31%) were less prevalent. Among the non-weighted data 50.29% reported any SG in the past year. In descending order, the prevalence estimates were SG in video games (40.00%), via apps (19.21%), in social networks (14.33%), and demo games (9.76%).

Predicting outcome 1 – PG score at T1 by participation and mechanisms at T0 (N = 1,178)

Total and direct effects

The results (unstandardized effect sizes of indirect effects a*b and 95%-confidence intervals) of predicting PG after one year (at T1) by T0 last year participation (yes, no) in the four different types of SG and their impacts via the applied mediators are illustrated in Table 1. It can be seen that the total effect (sum across direct and indirect effects) was only significant (95% confidence interval not including the value 0) for SG via apps. This effect was positive (increasing PG).

Table 1. Unstandardized effect sizes and 95% confidence intervals of mediation models to predict the sum score of problematic gambling after one year (N = 1,178)

Simulated gambling	Total effect	Direct effect c’	Mediators/Indirect effects a*b
Type			Frequency within SG	Frequency beyond SG	Breadth SG	Advertising	Micro transactions	Irrational cognitions	CIUS score	IGD score	Breadth Gambling	Depth Gambling
Video games	−0.105 (−0.796; 0.486)	0.64 (0.119; 1.162)	−0.603 (−1.379; 0.124)	0.099 (−0.012; 0.213)	0.123 (−0.172; 0.455)	−0.022 (−0.096; 0.039)	0.086 (−0.014; 0.263)	0.023 (0.002; 0.055)	−0.051 (−0.106; −0.008)	0.072 (−0.003; 0.157)	0.039 (−0.114; 0.173)	0.129 (0.008; 0.295)
Apps	1.082 (0.293; 1.969)	−0.239 (−0.953; 0.475)	0.753 (−0.232; 1.829)	0.019 (−0.194; 0.238)	−0.083 (−0.639; 0.483)	−0.014 (−0.071; 0.029)	0.079 (−0.016; 0.252)	0.057 (0.011; 0.120)	−0.062 (−0.129; −0.011)	0.07 (0.008; 0.158)	0.07 (−0.149; 0.287)	0.192 (0.003; 0.451)
Social network	1.262 (−0.085; 2.840)	−0.544 (−1.396; 0.309)	0.919 (−0.440; 2.586)	0.171 (−0.048; 0.433)	−0.166 (−0.663; 0.253)	−0.017 (−0.087; 0.035)	0.072 (−0.015; 0.262)	0.06 (0.011; 0.126)	−0.086 (−0.175; −0.015)	0.083 (0.002; 0.190)	0.067 (−0.150; 0.292)	0.158 (0.007; 0.368)
Demo games	1.146 (−0.739; 3.035)	−0.193 (−1.411; 1.025)	0.507 (−1.457; 2.433)	0.101 (−0.077; 0.292)	−0.014 (−0.442; 0.372)	−0.017 (−0.091; 0.030)	0.11 (−0.028; 0.403)	0.05 (0.007; 0.111)	−0.045 (−0.105; −0.003)	0.067 (0.003; 0.166)	0.093 (−0.242; 0.42)	0.295 (0.007; 0.694)

Annotation: Statistically relevant effects (95%-CI not including value 0) in bold type.

CIUS = sum score of Compulsive Internet Use Scale; IGD = sum score of 9 items of Internet Gaming Disorder.

A significant positive direct effect (impact of last year participation when simultaneously considering indirect effects) remained only for SG within video games. Insignificant direct effects of most SG types indicate that bivariate associations of SG and PG can be explained completely by indirect effects (different etiological mechanisms) of the applied mediation models. Moreover, insignificant total effects were attributable to many insignificant effects as well as to the opposite direction of PG increasing and decreasing indirect effects, which may cancel out each other.

Indirect effects

In order to evaluate the mechanisms via which participation in particular SG types strongly impacts PG, substantive indirect effects are the most interesting results. Frequency within and beyond SG, breadth of SG, breadth of gambling, perception of advertising, and micro transactions did not substantially mediate the associations of SG and PG. Nevertheless, impacts of all four types of SG on PG were significantly mediated via:

(1) irrational cognitions about gambling (positively),

(2) depth of real gambling (positively) and

(3) compulsive Internet use (negatively).

Problematic Internet gaming significantly impacted PG positively for three of four types (except SG via video games).

Comparisons of the effect sizes of the significant indirect effects are presented in Figure 2. Comparisons within a predictor of interest were warranted because the product of a*b canceled out scale effects of mediators. While comparisons showed that all four SG types influenced PG negatively via the CIUS score, there were two patterns for the positive influence of SG on PG. Participating in SG within video games impacted PG strongly and positively via (1) depth of gambling and to a lesser extent via (2) irrational cognitions. Participating via apps, social networks, and demo games all impacted PG positively mostly via (1) depth of gambling, but also via (2) IGD, and (3) irrational cognitions.

Figure 2. Unstandardized effect sizes of substantial mediators, grouped by game type. Annotation: CIUS = Compulsive Internet Use, IGD = Internet Gaming Disorder

Predicting outcome 2 – gambling onset at T1 by participation and mechanisms at T0 (N = 531)

Results of the mediation analyses to predict gambling onset after 1 year are presented in Table 2. Only participating in demo games at T0 showed a significant total effect on gambling onset after 1 year. The effect was negative and attributable to a risk decreasing impact of the mediator ‘frequency within’ (indirect effect). However, inspecting the data in detail showed that the effect was potentially an artifact: On the demo games last year participation level (frequency within) ‘never’, 26% of the participants (132/510) started to gamble, on the level ‘less than once a month’ 47% (8/17) started to gamble, and on the level ‘1–4 times a month’ 0% (0/4) started to gamble. No one participated on even higher levels of frequency within demo games. Therefore, the risk decreasing impact was possibly an effect of the few cases participating in demo games at T0 with higher levels of frequency within.

Table 2. Unstandardized effect sizes and 95%-confidence intervals of mediation models to predict onset of any gambling after one year (N = 531)

			Indirect a*b
Game type	Total	Direct c’	Frequency within	Frequency beyond	Breadth	Advertising	Micro transactions	Irrational cognitions	CIUS score	IGD Score
Video games	−0.372 (−1.518; 0.396)	0.697 (−0.228; 1.622)	−0.503 (−1.918; 0.483)	0.149 (−0.012; 0.34)	0.015 (−0.417;.576)	0.074 (0.01; 0.176)	−0.022 (−0.266; 0.069)	0.006 (−0.019; 0.046)	−0.038 (−0.144; 0.037)	−0.053 (−0.224; 0.087)
Apps	−0.206 (−1.834; 0.966)	0.893 (−0.362; 2.148)	−0.337 (−2.128; 1.008)	0.317 (−0.037; 0.721)	−0.173 (−0.84; 0.443)	0.065 (−0.006; 0.17)	−0.021 (−0.173; 0.042)	0.02 (−0.026; 0.095)	−0.037 (−0.173; 0.061)	−0.039 (−0.167; 0.04)
Social network	1.076 (−0.598; 4.284)	0.015 (−1.386; 1.416)	1.227 (−0.533; 4.614)	0.215 (−0.176; 0.671)	−0.308 (−1.106;.303)	0.119 (0.004; 0.307)	−0.062 (−0.479; 0.027)	0.037 (−0.057; 0.176)	−0.059 (−0.257; 0.084)	−0.093 (−0.331; 0.093)
Demo games	−105.403 (−128.456; −14.705)	89.647 (−6056.636; 6235.93)	−105.521 (−128.727; −14.783)	0.104 (−0.098; 0.4)	0.064 (−0.353;.445)	0.016 (−0.102; 0.141)	−0.02 (−0.322; 0.054)	−0.016 (−0.096; 0.045)	−0.006 (−0.09; 0.058)	−0.024 (−0.148; 0.041)

Annotation: Statistically relevant effects (95%-CI not including value 0) were in bold type.

CIUS = sum score of Compulsive Internet Use Scale; IGD = sum score of 9 items of Internet Gaming Disorder.

Meanwhile, SG in both (1) video games and (2) social networks positively impacted gambling onset via ascending levels of experiencing advertising.

Discussion

The current study revealed that in representative student samples (12–17 years old) of three German cities, diverse SG activities (via video games, apps, social networks, demo games) were common for half of the sample. Findings also hold for real gambling activities, with four in 10 individuals participating at least once in any type of gambling during the last year – although underage gambling is illegal in Germany.

Parallel multiple mediation models were applied with many mediating mechanisms suggested in the literature about the convergence of gambling and gaming. They sought to disentangle impacts of last year participation in different types of SG on (1) levels of PG (scale level outcome) as well as on (2) gambling onset (dichotomous level outcome) after 1 year. Extending the Gambling Consumption Mediation Model (GCMM) of Brosowski et al. (2020), which only applied demographic and behavioral mediators, mechanisms of the proposed simulated gambling consumption mediation model (SGCMM) comprised (1) behavioral attributes of consuming SG at T0 (SG frequency within type of interest, maximum SG frequency beyond type of interest, number of SG game types involved in at least monthly, ascending levels of micro transactions in SG), (2) cognitive attributes at T0 (irrational cognitions about gambling, perception of advertising for gambling or SG), (3) indicators of other online problem behavior at T0 (CIUS, IGD), as well as (4) two different proxies of real gambling intensity at T0 (only to predict PG at T1: number of used gambling types, maximum gambling involvement across all gambling types). For each of the four types of SG, effect sizes and percentile bootstrap confidence intervals for mediating mechanisms were calculated – a state-of-the-art-procedure to evaluate mediating mechanisms with a favorable trade-off between errors of Type I and II (Hayes, 2018). Mostly, total effects of SG participation (sum of direct and indirect effects) did not show substantial impacts on the applied outcomes, due to (1) many negligible mediators involved and (2) partially opposed effect sizes of the individual mechanisms. However, several results of this topical and explorative modeling approach provide useful insights.

Predicting gambling onset

Participation in hardly any of the simulated game types at T0 substantially impacted gambling onset via the suggested mediators, which of course could be a consequence of the reduced sample (n = 531) and a lack of power. Moreover, gambling participation is a quite complex and volatile phenomenon, comprising a large bundle of interacting attributes (Williams et al., 2015). Obviously, the SGCMM at hand has to be bolstered up by (1) larger samples and (2) additional mediators to adequately predict gambling onset (for instance, impulsivity [Kim et al., 2019] or other components suggested by King & Delfabbro, 2016 like parental monitoring or peers). In addition to aspects of power, collinearity of mediators may have destabilized estimates. Further research is needed to diminish these methodological issues. For instance, different operationalizations of the applied mediators may be useful expansions to predict gambling onset or PG. Some of the applied scales were only a combination of ad-hoc items without psychometric validation. However, frequency within demo games showed risk decreasing impacts on the onset of gambling after 1 year, but due to a lack of cases in the upper levels of involvement, this effect possibly represents an artifact. Further research is needed to substantiate this result and also to shed light on potential beneficial effects of SG activities (Gainsbury, 2019).

In contrast, last year participation in SG via video games or social networks at T0 positively impacted gambling onset after 1 year indirectly via increased perception of advertising for gambling or SG. Prior research on the applied prospective dataset (see Hayer et al., 2018) only showed associations of onset with SG in social networks and advertising. But the SGCMM provides more precise information for each type of SG via which etiological mechanism participation impacts gambling onset. Prior analyses lacked this kind of precision, even though it is of particular interest in gambling prevention and regulation (Brosowski et al., 2020). On the one hand, the pivotal role of advertising as an onset promoting mechanism substantiates prospective findings from real gambling (Yakovenko et al., 2016), which showed that cognitions temporally precede behavior. On the other hand, the increased self-reported perception of advertising in the current study does not indicate an objective level of exposure. Increased attention for gambling stimuli as well as advertising (salience) may also be indicative of increased individual interest in gambling themes (Gainsbury, 2019). Yet, exactly these changed attitudes or cognitions may be the ‘normalizing’ aspects of converged gambling and gaming in a digitalized environment (Armstrong et al., 2018; Gainsbury, 2019; Hayer et al., 2018; King, 2018). Further prospective research is needed to clarify this assumption in detail with more than two measurement points and less subjective (more objective) attributes of exposure to advertising (for instance, see: Smith et al., 2019; Vaipuna et al., 2020). Nevertheless, results of the SGCMM revealed that last year SG participation via (1) video games or (2) social networks positively impacted gambling onset via increased self-reported perception of advertising for gambling or simulated gambling.

Predicting problem gambling

For predicting levels of PG at T1, participating in most SG activities did not retain any substantial direct effect when simultaneously estimating indirect effects of the mediators in the models. Only SG within video games showed a substantial positive direct effect on PG that could not be explained by all other variables in the model. Interestingly, exactly this game type did not show a substantial mediating effect via problematic Internet gaming, which was a relevant PG increasing mediator for all other game types. Obviously, SG via video games on its own is part of problematic gaming behavior which substantially impacts PG. Moreover, there could be other mechanisms not included in the current model that may cancel out this direct path. Exploring this must be left to future research. Mechanisms beyond problematic Internet gaming may be the influence of gambling peers, impulsivity, anxiety, depression, parental monitoring, and traits like neuroticism or narcissism (King, 2018; King et al., 2019, in press; Salvarlı & Griffiths, 2019). Latest research on maladaptive player-game relationships in problematic gaming (King et al., 2019, in press) also suggests a demand to further differentiate the genres of video games in which SG may take place.

Looking at indirect effects beyond (1) problematic gaming, (2) depth of gambling involvement and (3) irrational cognitions at T0 were also quite important PG impacting mechanisms of all SG activities at T0. Consequently, both latter variables (cognitive and behavioral), which may influence each other in a maladaptive reciprocal way, also have to be prioritized by prevention tools to reduce future PG. Myths and unrealistic conceptions of stochastic gambling processes may be of particular interest to maintain gambling behavior and were invoked as risk increasing and reinforcing characteristics of SG (Hayer et al., 2018). Other behavioral components of SG (frequency within, beyond, breadth, micro transactions) as well as advertising and breadth of real gambling did not substantially mediate associations of SG participation at T0 and PG at T1.

Interestingly, both mediators of behavioral problems other than gambling (problematic Internet use, problematic Internet gaming disorder) showed quite different results. While ascending levels of problematic Internet use at T0 posed a substantial risk decreasing mechanism for PG at T1 across all types of SG, ascending levels of problematic Internet gaming at T0 posed a PG increasing mechanism of SG activities via apps, social networks and demo games. Possibly, problematic Internet use may be indicative of excessive levels of social communication, rather than more asocial immersion in any type of gambling or gaming. Effect sizes of mediating IGD scores even exceeded effects of irrational cognitions about gambling in these three types of SG. In contrast to problematic Internet use, PG and problematic Internet gaming may share common etiological components like challenge and competition (see also Karlsson et al., 2019; King, 2018) – a result in line with Molde et al. (2019). In contrast to the study at hand, Karlsson et al. (2019) revealed positive associations of PG and problematic Internet use. Further longitudinal research is needed to dissolve these inconsistencies and to test whether certain associations may be (1) explained in different populations only (moderating effects) or (2) may be explained by some higher order third variables, for instance, intro- or extraversion, impulsivity, anxiety, or depression (King, 2018). Including such moderators or higher order third variables may show some or even all of the revealed mechanisms as epiphenomenal. For instance, impaired control, poor parental monitoring or maladaptive coping strategies may be key-problem-drivers of most excessive digital behavior in adolescence. Further research has to address this issue in a more systematical and confirmatory way.

In summary, the versatile profiles of important mediating mechanisms of SG justify the more precise evaluation with parallel mediation models in longitudinal data. Differentiated results when including indirect effects bolster Gainsbury’s (2019) remark that studies on SG have to adjust for levels of gambling intensity to reveal unbiased findings about probable impacts of SG on PG. When individuals are already gambling with higher intensity at T0, future PG was not substantially impacted by different patterns of consuming SG (participation, breadth, frequency within or beyond). Here, SG becomes part of a broader portfolio of monetary or non-monetary gambling. However, these findings do not exonerate all SG activities at T0 from a putative role for future PG at T1, because effect sizes of frequency within or beyond SG as well as ascending levels of micro transactions were positive across several types. But when simultaneously considering gambling intensity as well as other attributes of online problem behavior, other mechanisms emerge as more important for future PG than SG. Those mechanisms include (1) depth of real gambling involvement, (2) irrational cognitions, and (3) problematic Internet gaming. Hence, simple bivariate associations of SG activities with PG may be partially explained by attracting individuals that also have problems with Internet gaming or individuals who already participate in gambling, and SG activities simply join into a broader spectrum of risky gambling and gaming behavior.

This methodological challenge has already occurred in the context of online gambling (Gainsbury, 2015) and can only be solved by further research with more than two sampling points and complex modeling. However, the complexity of the phenomenon must not serve as argument to decelerate regulation – an established strategy of lobbyism (Petticrew et al., 2017). Due to the large fraction of particularly vulnerable minors involved in SG, monitoring, prevention, and regulation are quite warranted to avoid harms at an early stage.

Conclusion

Both analyses with the SGCMM show that measures to reduce:

(1) advertising, (2) irrational cognitions, (3) gambling participation and (4) problematic Internet gaming will reduce gambling onset or PG. But these mechanisms are increasingly hard to address in converging online environments with targeted advertising, predatory monetization, and normalizing trends of gamblification. With this in mind, the methodological idealism of precisely disentangling mechanisms of the epidemiological interplay of environment, game/gambling types, and person (attributes and behavior) becomes increasingly challenging. The presented SGCMM constitutes a first step to disentangle such prospective mechanisms with precise information about effect sizes and confidence intervals of a first collection of plausible predictors, mediators, and outcomes. Further research is needed to unfold the complex and interactive associations of problematic online behaviors in progressively blurred digital environments, for instance, with cross-lagged-panels, because the revealed associations of SG and gambling attributes may also work bi-directional – a situation not explicitly tested in the current prospective mediation analyses.

Limitations

Several limitations need to be mentioned when interpreting the presented results. First of all, self report data were analyzed, with plausible biases by memory, motivation, or information management. Second, predictors and mediators were collected at the same time, not all with psychometrically validated scales. A future design should isolate the participation in SG at T0, probable mechanisms at T1, and PG at T2 to provide more valid results. Additionally, a serial mediation model may be applied instead of the parallel models of the analyses at hand. Third, a large number of mediators was exploratively included, which may have reduced the power of the analyses due to aspects of collinearity and sampling variance (Hayes, 2018), particularly in the smaller sample to predict the onset of gambling. More confirmatory research with larger samples should statistically contrast the effect sizes of mediating mechanisms exploratively revealed in this study. Fourth, the applied model estimates may be biased by mean imputation of mediators, skewed attributes, disregarded moderating effects, or nonlinearities, which should be addressed in further research. Fifth, the prospective period of 1 year may be too short to validly assess PG, which often manifests over the course of several years (Gainsbury, 2019).

Despite these limitations, the results of the SGCMM outline the potential of mediation analyses in prospective datasets to disentangle complex etiological networks of problem behavior in online environments, for instance, PG, gaming, or Internet use. Making use of sophisticated statistical models will fuel the scientific progress in the field of gambling research.

Conflict of interest

Tobias Turowski and Tim Brosowski declare they have no conflict of interest. Tobias Hayer currently receives grants for gambling research from the Federal Ministry of Health, the Ministry of the Interior of Lower Saxony, and the German Lotto and Toto Block.

Ethical approval

All procedures performed were in accordance with the 1964 Helsinki declaration and its later amendments. In each involved federal state, supervisory school authorities approved the study materials and design. A written declaration of agreement from parents was a general precondition of participation in the study. In addition and regardless of this, pupils were able at any time to refuse to participate in the survey. Matching of questionnaires from different stages for data analysis purposes relied on a coding system that guaranteed the anonymity of participants in all cases.

Re-use of data

Parts of this study re-analyze already published data (https://doi.10.1007/s10899-018-9755-1) in an extended manner and with advanced methods of data analysis in order to explicate etiological mechanisms more precisely than previous research did.

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Additional information

Funding

This study was developed with research funding from the Hamburg Department of Health and Consumer Protection. The funding was not subject to any restrictions or specific instructions with regard to data collection, analysis, interpretation, or to publication of the results.

Notes on contributors

Tim Brosowski

Tim Brosowski is a psychologist (Diploma) and data analyst with particular expertise in behavioral addictions, data mining and statistical modeling. He graduated in 2011 at the University of Bremen and is still working at the University of Bremen on several research projects about gambling as well as on his dissertation project about big data in gambling research.

Tobias Turowski

Tobias Turowski studied Psychology at the University of Bremen (bachelor’s degree) and at the University of Lübeck (master’s degree). He graduated in 2019. In 2018, he was a member of a research group that investigated the effectiveness and optimization of social concepts (including a newly established multi-venue exclusion program) in gaming halls in Hesse, Germany.

Tobias Hayer

Dr. Tobias Hayer has received his Ph.D. (University of Bremen) in 2012 by preparing a doctoral thesis on ‘Adolescents and gambling-related problems.’ For almost 20 years, his research focuses on various facets of gambling and gambling addiction. This includes, among other things, the prevention of gambling-related problems; the effectiveness of certain player and youth protection measures; the risk potential of various gambling forms as well as the interface of gaming and gambling.

Appendix 1

Table A1. 12-month-prevalence of online or offline gambling (n; % [95%-Confidence]). N = 1,905

Attribute	Total data set (not weighted)	Bremen (weighted) n = 964	Hamburg (weighted) n = 435	Lübeck (weighted) n = 506
Any gambling	722; 38.49% (36.28–40.69%)	365; 38.35% (35.26–41.44%)	173; 40.02% (35.41–44.64%)	191; 38.72% (34.43–43.01%)
Scratch tickets	349; 18.72% (16.95–20.49%)	188; 19.89% (17.35–22.44%)	60; 13.90% (10.62–17.17%)	99; 20.08% (16.54–23.62%)
Sport betting	276; 14.78% (13.17–16.39%)	145; 15.36% (13.07–17.66%)	65; 14.96% (11.60–18.33%)	71; 14.46% (11.35–17.56%)
Card or dice games except poker	216; 11.59% (10.13–13.04%)	104; 11.05% (9.05–13.05%)	59; 13.55% (10.32–16.77%)	59; 12.04% (9.17–14.91%)
Poker	161; 8.62% (7.35–9.89%)	68; 7.13% (5.49–8.77%)	38; 8.84% (6.16–11.52%)	48; 9.75% (7.14–12.37%)
EGMs	153; 8.20% (6.95–9.44%)	61; 6.49% (4.92–8.06%)	36; 8.34% (5.73–10.96%)	57; 11.60% (8.78–14.43%)
Lotto	99; 5.31% (4.29–6.33%)	59; 6.24% (4.70–7.78%)	26; 6.08% (3.83–8.34%)	22; 4.55% (2.71–6.39%)

Appendix 2

Table A2. 12-month-prevalence of simulated gambling activities from home or via mobile device (n; % [95%-Confidence]). N = 1,905

Attribute	Total data set (not weighted)	Bremen (weighted) n = 964	Hamburg (weighted) n = 435	Lübeck (weighted) n = 506
Any simulated gambling	958; 50.29% (48.04–52.53%)	475; 49.29% (46.14–52.45%)	226; 52.00% (47.31–56.70%)	290; 57.28% (52.97–61.59%)
From home
Within video games	526; 27.90% (25.88–29.93%)	285; 29.74% (26.85–32.64%)	124; 29.00% (24.70–33.30%)	167; 33.40% (29.26–37.54%)
Via apps	343; 18.20% (16.45–19.94%)	156; 16.36% (14.02–18.71%)	75; 17.42% (13.85–21.00%)	101; 20.32% (16.78–23.85%)
Within social networks	257; 13.64% (12.09–15.19%)	133; 13.86% (11.67–16.05%)	53; 12.39% (9.27–15.51%)	62; 12.45% (9.56–15.35%)
Demo games	169; 8.96% (7.67–10.24%)	88; 9.21% (7.38–11.04%)	37; 8.70% (6.04–11.36%)	51; 10.27% (7.61–12.93%)
Via mobile device
Within video games	547; 29.16% (27.10–31.21%)	264; 27.86% (25.00–30.71%)	142; 32.99% (28.55–37.43%)	158; 31.57% (27.50–35.65%)
Via apps	209; 11.13% (9.71–12.55%)	100; 10.60% (8.64–12.56%)	38; 8.88% (6.19–11.58%)	69; 13.85% (10.82–16.88%)
Within social networks	105; 5.59% (4.55–6.62%)	55; 5.74% (4.26–7.22%)	20; 4.67% (2.68–6.66%)	29; 5.89% (3.82–7.96%)
Demo games	77; 4.12% (3.22–5.02%)	46; 4.84% (3.47–6.21%)	12; 2.91% (1.31–4.50%)	26; 5.27% (3.30–7.23%)

Appendix 3

Table A3. Bivariate Spearman-correlations between individual simulated gambling activities. N = 1,905 (unweighted data set)

#	Type	2	3	4	5	6	7	8
	From home
1	Within video games	.31*	.27*	.21*	.45*	.24*	.15*	.13*
2	Via apps		.39*	.32*	.26*	.68*	.29*	.20*
3	Within social networks			.27*	.20*	.32*	.53*	.15*
4	Demo games				.17*	.25*	.19*	.51*
	Via mobile device
5	Within video games					.32*	.23*	.16*
6	Via apps						.39*	.26*
7	Within social networks							.23*
8	Demo games

Annotation: * p < 0.05. Associations of different modes of access of the same game type.