Early detection items and responsible gambling features for online gambling

Abstract

Early detection is an effective building block for the prevention of problem gambling. This study aims to identify communication-based indicators for gambling-related problems in the setting of online gambling. In the framework of a semi-structured interview, customer service employees of three online gambling operators were surveyed, to identify indicators in customer correspondence could be used as a predictor for gambling-related problems. In a confirmatory part of the study, we investigated to what degree these indicators are able to predict problem gambling in a prospective empirical design. An optimally parsimonious log-linear model, was able to correctly predict 76.6% of the cases. Discussed in the light of this evidence, communication-based indicators could constitute an effective component of early detection. Due to the fact that the internet offers optimal conditions for consistent monitoring and objective analysis, the suggested predictive model could be combined with other models, relying on the analysis of gambling behaviour.

Keywords:

• responsible gambling
• internet
• predictors
• harm-minimization
• screening instruments

Introduction

Online gambling

As an innovation of the late 1990s, online gambling quickly aroused both the interest and concerns of the scientific community. In a detailed treatise on the design characteristics of gambling products, Griffiths (1999) refers to the internet as a central factor that will lead to a sudden surge in the availability of gambling and concluded that it would therefore increase the number of pathological gamblers. This concern is again specifically expressed by Parke, Griffiths and Irwing (2004) as well as Orford (2005) in relation to the deregulation of online gambling under the UK Gambling Act.

In particular, these risks were derived from the easier accessibility and higher availability of gambling on the internet. The supposition that the higher availability of gambling services would necessarily lead to higher prevalence rates of disordered gambling was first expressed by Kallick, Suits, Dielman and Hybels (1976). It was supported by the meta-analysis conducted by Shaffer, Hall and Vander Bilt (1999), which analysed prevalence surveys conducted between 1975 and 1996. According to the authors of that study, the prevalence rate among adults between 1974 and 1997 rose continually with the increasing supply of lotteries, casinos and other forms of gambling (Volberg, 2004). The authors concluded that, if there are more opportunities to gamble, more people will gamble, so the prevalence rate of disordered gambling will increase (Meyer & Bachmann, 2005). At the same time, however, a number of studies failed to find a correlation between the availability of gambling products and the prevalence disordered gambling (see Volberg (2004) for an overview). The linear relation between availability and prevalence rate is questionable and might instead be moderated by effects like adaptation (LaPlante, Schumann, LaBrie & Shaffer, 2008). Volberg also hypothesizes that the greater availability of gambling products may have differential effects on different groups of gamblers. It might lead to increased exposure among moderate gamblers, whereas risk tends to decrease especially among frequent gamblers, if appropriate preventive measures are established in parallel (Volberg, 2001).

Responsible gambling

The term responsible gambling generally refers to concepts applied to a broad range of issues and policies, from individual behaviours and attitudes to public health (Dickson-Gillespie, Rugle, Rosenthal & Fong, 2008; IPART, 2004). Responsible gambling practices are part of the responsibility of the gambling industry and should be distinguished from treatment, which is aimed at gamblers whose gambling has already resulted in obvious consequences: ‘The treatment of gamblers who already have developed gambling-related harm remains the domain of specialists working in public health programmes, including counseling and other health services’ (Blaszczynski, Ladouceur & Shaffer, 2004, p. 308). Player protection seems almost to have become a competitive advantage within the gambling industry, especially on the internet, where players can easily move from one operator to another. A study by Parke, Rigbye, Parke and Williams (2007), in which over 10,000 gamblers were interviewed, revealed a very high level of acceptance for protective measures. Wood and Williams (2009) showed that the good general reputation of an online gambling operator is the main reason for choosing this operator over another. The most frequently considered responsible gambling measures relate to the display of messages, assisting in the process of informed choice (Blazczynski, Ladouceur, Nower & Shaffer, 2008), (self-) limitation, (self-) exclusion and the structural design of the games themselves (Parke & Griffiths, 2007).

All electronic forms of gambling – be it EGMs or gambling on the internet – can be modified to provide automated pop-up messages to the player, displaying responsible gambling messages or information about the total amount wagered or the duration of the gambling session. Informative pop-ups after specific session durations asking the players whether they wanted to continue was shown to have a small effect in decreasing the gambling duration and expenditure only for the sub-group of high-risk players (Schellinck & Schrans, 2002). On-screen messages presented to the player were able to correct irrational beliefs and erroneous perceptions about the independence of the games (Steenbergh, Whelan, Meyers, May & Floyd, 2004; Benhsain, Taillefer & Ladouceur, 2004; May, Whelan, Meyers & Steenbergh, 2005; Cloutier, Ladouceur & Sevigny, 2006). However no evidence for a change in gambling persistence behaviour could be provided. One possible explanation could be found in the message content. Monaghan and Blaszczynski (2010) demonstrate that messages, encouraging self-appraisal (e.g. ‘Have you spent more than you intended? Do you need to think about a break?’), resulted in a greater behavioural change than purely informative messages (e.g. ‘Your chances of winning the maximum price are generally no better than one in a million’). But also the timing of the pop-up messages could be a determining factor. Jardin and Wulfert (2009) show that even very simple informative pop-up messages are able to affect not only a player's attitude but also his gambling behaviour when displayed within the gambling session, instead of displaying it at the beginning.

Self-limitation is a type of voluntary agreement between the customer and the operator. For example, it is possible to limit the amount gamblers may deposit on their gambling accounts each week, which creates direct control over the potential losses incurred. It is, however, quite difficult to apply such protective measures in land-based gambling, since once the customer is inside the casino, it is impossible for the operator to control this gambler's further behaviour. Land-based gambling operators can monitor only the gamblers' frequency of visits, not their gambling behaviour itself. Smart-card systems might enable land-based venues to introduce self-limitation in the future, but the gambling industry is still hesitant about adapting this technology (Bernhard, Lucas & Jang, 2006; Parke, Rygbie & Parke, 2008) and to the author's knowledge, there is no venue yet that has introduced such a system for all its customers. In the online sector, in principle, all the characteristics of gambling behaviour (deposits, wagers, losses, duration of gambling, etc.) can be observed and therefore limited in real time. In practice, however, it is only deposit limits that have proven effective as a means of pre-commitment (Nelson et al., 2008), while imposed limits displayed very limited effectiveness (Broda et al., 2008). For other types of self-limitation, especially time limits, a thorough evaluation of effectiveness as well as possible side effects is still necessary. As pointed out in the thought experiment by Bernhard and Preston (2004) forms of time limitation could potentially trigger a frenzied gambling behaviour associated with a loss of control as the limit gets closer.

The option available to problem gamblers to voluntarily exclude themselves, or for the gambling operator to forcibly exclude them, is – in combination with other gambler protection measures – an effective means of gambler protection in land-based gambling (Ladouceur, Sylvain & Gosselin, 2007; Meyer & Bachmann, 2005). In the online gambling sector, too, voluntary or prescribed exclusion from gambling – in combination with other gambler protection measures – constitutes an effective means of gambler protection. Meyer and Hayer (2010) show that in the online setting compared to land-based gambling, self-exclusion is more frequently used as a preventive measure, before actual harm has occurred. Reasons for this difference could be the fact that self-exclusion on the internet is easily accessible and players are less daunted than they would be when having to personally contact a casino employee (Wood & Griffiths, 2007; Wood & Wood, 2009). Since the identity of the customer is known to the operator, it is possible to enforce the exclusion by deactivating the current account and disallowing the user to register a new account (TÜV, 2009). The following table shows player protection measures in land-based gambling vs online gambling. It is intended to clarify both the differences and commonalities of the two.

Table 1. Protective measures for gamblers in land-based and online gambling. Measures which are covered by self-regulatory best practice standards marked accordingly

Protective measure	Land-based gambling	Online gambling
Exclusion
Partial exclusion from single types of games	not possible	common practice
Self-exclusion	common practice	common practice
Prescribed exclusion	common practice	common practice
Limitation
Limit to gambling volume	not possible	common practice
Limit to gambling time	not possible^1	possible
Limit to gambling frequency	possible ^2	possible
Design of the gaminge structure
Pop-up messages	Possible (for EGMs)	Possible
Succinct presentation of the gambling time	possible ^3	common practice
Succinct presentation of the gambling volume	not possible	common practice
Succinct presentation of the gambling frequency	possible	common practice
Information offering
Awareness material and responsible gambling advice	common practice	common practice
Self tests	possible	common practice
Interactive Sself-help tools	not possible	possible
Contact with qualified support structure	common practice	common practice
Under-age protection
Access limitations	possible (but with many forms of land-based gambling not implemented)	common practice
Handling credit
No award of credit	common practice	common practice

Numerous findings argue against the concerns expressed in the past to the effect that the structural characteristics of online gambling make it an especially dangerous type of gambling. While in population-based prevalence studies online gambling is often associated with an increased risk (Griffiths, Wardle, Orford, Sproston & Erens, 2008; Wood & Williams, 2009), this finding might be the consequence of a statistical artefact. When analysed in a multivariate framework (Welte, Barnes, Tidwell & Hoffman, 2009) or corrected for the number of different types of games played (LaPlante, Nelson, LaBrie & Shaffer, 2009b), no additional risk related to online gambling remains. At the same time, analyses of the actual gambling behaviour have shown that the vast majority of gamblers exhibit a moderate gambling behaviour (LaBrie, LaPlante, Nelson, Schumann & Shaffer, 2007; LaBrie, Kaplan, LaPlante & Nelson, 2008; LaPlante, Kleschinsky, LaBrie, Nelson & Shaffer, 2009a). However the data for these analyses of gambling behaviour were derived from one gambling operator. It is therefore possible that some players could have several accounts with other gambling operators which were not considered (Meyer & Hayer, 2010, p. 65). The feasibility of player protection measures in the online sector (McMillen, 2003; Parke et al., 2007; Productivity Commission, 2010, pp. 15–22) might be an explanation why there is a discrepancy between the theoretically possible very risky behaviour and the moderate behaviour that is in fact observed.

Early detection of problem gambling

When gambling problems become evident, and a gambling operator can react to this, usually the player has lost control over his gambling and incurred a considerable amount of harm. Since players with manifest problems who have already lost control might not accept protective measures (Blaszczynski & Nower, 2008, p. 56) or might try to evade them by changing the operator they are playing with (Wood & Griffiths, 2007; Meyer & Hayer, 2010, p. 65), responsible gambling measures should not primarily focus on such a late state of problem development (Wood, 2010, p. 6). In contrast to a reactive approach, the early detection of incipient problems might enable gambling operators to prevent problematic developments via minimal interventions. If these are less invasive and come at an early point in time, when the player has not yet lost control over his gambling, the chance of compliance might be greater.

However, early identification of at-risk gamblers is connected with considerable challenges for gambling operators. Early detection implies that maladjusted developments can be systematically recorded and that firmly anchored structures are in place with which the observations can be continued, documented and analysed. The objective is to set up binding rules, develop standardized procedures and anchor the fundamental notions of early detection and early intervention in the corporate culture (Häfeli & Schneider, 2005; Häfeli & Lischer, 2010; Hafen, 2007; Meyer & Hayer, 2008). Against this backdrop, it comes as no surprise that research on the early detection of problem gamblers has gained more significance in recent years (Hancock, Schellinck & Schrans, 2008).

The methods of early detection in land-based gambling have been established through expert interviews conducted by Allcock (2002) as well as three field studies: Delfabbro, Osborn, Nevile, Skelt and McMillen (2007), Häfeli and Schneider (2005), and Schellink and Schrans (2004). The early detection indicators for land-based gambling established by the various authors are rather similar. Thus, they generally correspond to categories identified by Häfeli and Schneider (2005): ‘Frequency of visits and duration of visits’, ‘Sourcing money’, ‘Betting behaviour’, ‘Social behaviour’, ‘Gambling behaviour and reactions’, as well as ‘Outward appearance’. Häfeli and Lischer (2010) performed an empirical evaluation of the criteria currently applied by Swiss casino employees for the early detection of problem gamblers. The authors pointed out the importance of empirical validation for early detection systems, since many of the predictors used in practice displayed poor specificity.

In online gambling, in contrast, various factors related to problem gambling can be monitored and logged during the gambling activities. The following factors and trends over time could be used to indicate problem gambling. Duration and frequency of gambling activities, number and frequency of bets, size of the stakes, chasing of losses and lack of adaptation in gambling behaviour (TÜV, 2009). These indicators could enable the operator to record gambling behaviour at relatively little expense and to monitor it at the same time, which would provide operators with the means for the early detection of customers with problematic gambling behaviour. First empirically validated models, predicting future self-exclusion of online gamblers based on variables of gambling behavior, were proposed by Braverman and Shaffer (2010) and LaBrie and Shaffer (2010).

It is conventionally assumed that the indicators of gambling behaviour are more easily accessible online, yet in land-based gambling, because of the physical proximity of the gambler, indicators that are based on communication are more easily accessible. However, arguing against the latter assumption, communication also takes place in online gambling, at least in electronic form. Accordingly, it is to be investigated to what extent customer communication might also contain qualitative data about the problematic attitudes and behaviours of the gambler, which could in principle be used for the early detection of problem gamblers. On a daily basis customer service employees deal with a high quantity of customer requests and it is possible to assume that a fraction of these have their origin in gambling problems. To the author's knowledge, no empirical research has yet investigated the indicators for future gambling problems in communications with customer services. This exploratory study aims to identify communication-based predictors for the early detection of disordered gambling.

Method

Qualitative screening

In the framework of a semi-structured interview, eight senior staff members from three private internet gambling operators were surveyed, to determine which key words and indicators in customer communication can be used as an indicator for potentially problematic gambling behaviour. The aim of this screening was to generate an overview of how early detection of problem gambling is currently accomplished and to generate first hypotheses of which aspects of customer correspondence could serve as predictors. The interviewed employees were selected from different fields of competence. Besides management, operatives from first and second line customer service were interviewed. Each interview took between 45 and 60 minutes. The interviews were transcribed, evaluated and clustered, in order to build thematic groups of indicators.

All the online gambling operators interviewed had responsible gambling embedded into their business model. Therefore they have a dedicated responsible gambling backbone and standardized responsible gambling processes in relevant business areas (see EGBA Standards (European Gaming & Betting Association, 2009) or eCOGRA Generally Accepted Practices (e-Commerce & Online Gaming Regulation & Assurance, 2009) for more information). The organizational form of this responsible gambling structure and the communication processes built upon vary considerably – especially with respect to their degree of standardization – between the different operators consulted. While processes to escalate customer communication – potentially relating to problem gambling – with dedicated teams are in place, the structure and organizational embedding of the teams vary.

All interviewed employees agree that customer communication does contain indicators for future gambling problems, which cannot be solely based on discrete key words. Often, the problem is instead defined by the context of the complete communication. Assembling a list of key words might therefore be insufficient. Seven clusters are reported to be indicators for gambling related problems: Chasing losses, financial situation or financial requests, loss of control, family or social situation, heavy complaining about the results of the games, criminal activities or threats and health problems.

Quantitative prediction

The second part of the study should be understood as a confirmatory investigation with the aim of testing the research question, in how far the indicators identified in the qualitative part, are able to predict manifest gambling-related problems in a prospective longitudinal design. Furthermore it should be investigated how far additional predictors existing in the communication data are able to produce incremental validity and thus could improve early detection processes.

In line with recent predictive approaches (Braverman & Shaffer, 2010; LaBrie & Shaffer, 2010) self-exclusion was chosen as a criterion for problem gambling. This criterion does not completely match the clinical diagnosis of disordered gambling, but instead focuses on the player's own attitude towards his gambling. A predictive model based on this criterion will therefore not be able to predict a clinical diagnosis, but rather anticipate whether a player will be forced to change his gambling because of incipient problems.

Ethics

The design of the study was reviewed and approved by the research unit of Lucerne University of Applied Sciences and Arts. The usage of anonymous gambling data for research purposes is part of the terms and conditions of the gambling operator, who provided the data. All participating players have given their explicit consent to these terms and conditions.

Sample

The sample contains 150 self-excluders and 150 controls. The sample was drawn without taking into account the type of game played. The group of self-excluders is a random sample out of the population of German or English speaking players who opted to exclude themselves for problem gambling reasons during February 2007. Players who closed their account for other reasons (e.g. not satisfied) were not selected. The group of controls is a random sample out of the population of German or English speaking players, who played for real money at least once during this month but did not close their account. In relation to the gambler characteristics, the high percentage of men was particularly obvious: 93.0% (n = 279). The average age was 32.2 years (SD = 10.1).

Measures

The data provided for this study contains all e-mail correspondence since registration from both groups. All personal information about players was removed by the gambling operator prior to providing the data for research purposes. For the self-excluders, all communication received after the initiation of self-exclusion or relating to the actual process of self-exclusion was excluded. Then the e-mails were rated in an observed-blind design, meaning the coders did not know which group was the test group and which the control group. The data set contained 1008 e-mails to the operator written in English or German by the research subjects. The texts of the 1008 e-mails were investigated through text analysis by two independent coders using the seven thematic clusters of the qualitative screening as a first step towards developing classification categories. After a pre-analysis of the texts, performed independently by the two coders, additional categories – on top of those hypothesized to be risk indicators in the qualitative screening – were defined in order to be able to categorize the majority of e-mails: inquiries about increase or decrease of gambling limits, inquiries about blocking of specific games or the closure of the whole account, inquiries about account administration (e.g. lost password), technical problems or financial transactions (e.g. failed deposit) as well as requests for promotions or threats of account closure. The classification guideline was modified accordingly and the classification process was restarted. 163 of the e-mails (16.2%) did not contain any identifiable indicators (e.g. player thanking customer services), 747 e-mails (74.1%) contained exactly one indicator and 98 e-mails (9.7%) contained two indicators. Based on 1008 e-mails a total of 943 content indicators were found in the text sample. In addition, the tonality of the e-mail text was coded into the following categories: neutral, complaint and threats. The interrater reliability between the two coders amounts to κ = 0.78.

The codes of the 1008 e-mail texts were aggregated so that variables were available for the number of mentions of the risk indicators and of the tonality criteria and combined with the values of the socio-demographic parameters and other information into an aggregate data set for all research subjects. From the number of mentions of the risk indicators new variables were calculated, that measure for an individual player the proportion of the mentions of a certain risk indicator compared to the mentions of all risk indicators. This transformation to relative percentages of the mentions of the indicators was carried out in order to avoid problems with multicollinearity in the regression model.

As a measure of frequency, a proxy for the relative number of e-mail contacts per month per research subject was calculated. Since the distribution of the underlying variable that measures the period until the end of the customer contact was strongly skewed to the right, it was approximated to a normal distribution by means of a logarithmic transformation. Additionally, as a measure of urgency experienced by the player, we counted the number of immediate follow-ups, when a player consecutively sent several e-mails with an identical or similar request on the same day before an agent was able to reply.

Analyses

Besides evaluating the data with descriptive statistics, various bivariate analyses were performed between the group of self-excluders and the group of controls. In order to determine the impact of the hypothesized risk indicators and other variables on group membership (self-excluders vs controls) logistic regression was applied. Since the objective was to identify those variables among the hypothesized risk indicators that factually yield validity for the prediction of future self-exclusion, a stepwise procedure was selected. It should be taken into account that results can be heavily influenced by the characteristics of a given sample. A backward stepwise process was selected with an exclusion test, which was based on the probabilities of the likelihood quotient statistics based on conditional parameter estimators. The a priori significance level was increased to p = 0.1 for this analysis, in order to underscore the explorative character of the analysis. The objective was to detect possible new markers for early detection, which explains why an increased risk for type-I errors was accepted.

Results

In the sub-sample of self-excluders the percentage of men was 94.7% (n = 142), vs 91.3% in the control group (n = 137). There is no significant difference between the two percentages (χ² = 0.819, df = 1, p = 0.365). The mean age of the self-excluders was 31.5 (SD = 10.1), vs 32.9 in the control group (SD = 10.0). There is no significant difference between these two mean ages (F = 1.413, df = 1, p = 0.235). As a measure of exposure, the time since registration of the self-excluders ranged from 3 to 2534 days (median = 590), vs from 6 to 1827 days (median = 295) in the control group, resulting in an effect bordering to statistical significance (median test: χ² = 3.582, df = 1, p = 0.058). The percentage of research subjects who enter into communication, i.e. who make at least one e-mail contact, is 52.7% in the case of the self-excluders (n = 79) vs 39.3% in the control group (n = 59), which is a significant difference (z-test: p < 0.05). The distribution of the number of e-mail contacts per research subject is highly skewed to the right, so that the median was selected as a measure of centrality. In the self-excluders subset of the sample, the number of e-mail contacts per research subject ranged from 1 to 42 (median = 6.0), vs 1 to 20 in the control group (median = 2.0). The median of the self-excluders is significantly higher (median test: χ² = 17.054, df = 1, p < 0.001).

The distribution of frequencies (number of e-mail contacts per month per research subject) are strongly skewed to the right despite the logarithmic correction of the calculation, so that the median was selected as a measure of centrality. In the self-excluders subset of the sample, the frequency varied from 3.9 to 198.0 (median = 31.9) vs 4.6 and 102.9 (median = 12.5) in the control group. The median of the self-excluders is significantly higher (median test: χ² = 29.671, df = 1, p < 0.001). When analysing the distribution of customer communication over time for the self-excluders, the occurrence of customer contact cumulates during the final 6 months prior to self-exclusion. 36% of all mails are derived from this period. The remaining mail contact distributes uniformly over earlier stages (15%, 6–12 months prior; 21%, 13–24 months prior; 28%, more than 24 months prior).

The distributions of immediate follow-ups are strongly skewed to the right, so that the median was selected as a measure of centrality. In the self-excluders subset of the sample, the total varied between 0 and 14 (median = 2), vs 0 and 6 in the control group (median = 0). The median of the self-excluders is significantly higher (median test: χ² = 22.947, df = 1, p < 0.001).

Tables 2 and 3 show the mean values and standard deviations of the relative percentages of the mentions of the indicators. The three risk indicators with the highest relative percentages are doubts about results of games, account administration and financial transaction. The mean values of 5 of the total of 16 relative percentages are significantly different in the two groups (t-test: p < 0.05). As can be seen in Table 3, there is a significant difference in the value percentages of the tonality variable between the two groups in the categories neutral and threatening.

Table 2. Descriptive statistics of the hypothesized risk indicators

	Controls		Self-excluders
	(n = 59)		(n = 79)
Content predictors	Mean (%)	SD	Mean (%)	SD	t-test
Chasing losses					na
Financial problems			0.1	0.082	na
Loss of control					na
Social situation					na
Criminal acts					na
Health problems					na
Results of games/bets	11.0	20.5	16.9	22.4	ns
Increase limits	0.3	2.2	1.3	6.3	*
Lower limits			0.3	2.5	na
Partial blocking			0.3	1.3	na
Account reopening	2.7	14.5	8.6	19.3	***
Account administration	23.3	35.4	20.0	28.5	*
Technical problems	5.2	19.6	4.1	13.0	ns
Financial transaction	30.9	38.7	35.1	31.8	*
Request for bonus	7.9	20.1	5.0	14.8	*
Threat of account closure			0.5	2.0	na
Notes: The analysis includes only the players who communicated, and the sample size has thus been reduced from n = 150 to n = 59 and n = 79, respectively. Mean and standard deviation of relative percentages of the mentions of content predictors. Blanks entries are all nil.
*p < 0.05; **p < 0.01; ***p < 0.001; ns, not significant; na, not available.

Table 3. Descriptive statistics of the hypothesized risk indicators. Shares of tonality predictors

	Controls	Self-excluders
Tonality predictor	(n = 59) (%)	(n = 79) (%)	t-test
Neutral	61.3	44.5	*
Complaining	35.0	40.0	ns
Threatening	3.8	12.9	*
Other	0.0	2.6	na
Total	100	100
Note: The analysis includes only the players who communicated, and the sample size has thus been reduced from n = 150 to n = 59 and n = 79, respectively.
*p < 0.05; **p < 0.01; ***p < 0.001; ns, not significant; na, not available.

The relative number of e-mails per month, the absolute number of immediate follow-ups, the relative percentages of all indicators and the shares of all non-neutral tonality variables were entered into the stepwise predictor selection process, which converges into a significant model in eight steps (χ² (5;137) = 36.092, p < 0.001) with five significant predictors, i.e. in order of decreasing power of influence in the model (as measured by p-value). Frequency (number of e-mail contacts per month per research subject), the percentages of mentions of the hypothesized risk indicators ‘Request for account reopening’, ‘Financial transaction’ and ‘Account administration’ and the tonality ‘Threatening’. The results of the stepwise logistic regression are summarized in Table 4.

Table 4. Stepwise logistic regression regarding group membership (self-excluders vs controls)

			95% CI
Indicator	OR	p	Lower	Upper
Frequency	1.034	0.003	1.012	1.057
Account reopening	1.029	0.024	1.004	1.056
Financial transaction	1.012	0.057	1.000	1.025
Account administration	1.012	0.094	0.998	1.026
Threatening tonality	2.596	0.095	0.846	7.963
Note: Model ( = 36.092, p < 0.001);  = 0.311; overall classification: 76.6%.

The odds ratios (OR) reach levels between 1.012 and 2.596. The effect size of the explained variance is very high: Nagelkerke Pseudo-R ² = 0.311 (Cohen, 1992). The overall classification rate is 76.6%, resulting in a sensitivity of 78.2% and specificity of 74.6%. In two cases the confidence intervals (CI) for the ORs form an interval including the value 1.

Discussion

When considered as a mediating variable between the availability of gambling opportunities and the impact gambling has on public health, the development and evaluation of responsible gambling practices is increasingly receiving regulatory attention. The internet, while posing new challenges to player protection, also provides the technology for implementing innovative responsible gambling tools. However, as player protection thus becomes more differentiated, providing protective measures for different phases of prevention, the individualization of player protection becomes more important. By suggesting appropriate responsible gambling measures, tailored to the needs of the player, each player could be given the individual level of protection they need in order to gamble safely. As Schellinck, Schrans and Walsh (2000) found, there is a considerable delay between the emergence of problem gambling and the onset of manifest harm. Therefore, the earlier incipient gambling problems can be detected, the smaller the harm sustained and the extent of intervention necessary.

While there is growing interest in this new field, the number of empirically evaluated approaches is still small and they are fundamentally different for land-based and online gambling. It was assumed that, while quantitative indicators of gambling behaviour are more easily captured in online gambling, communication-based indicators were only accessible in land-based gambling. However, while the means of communication are different, there is a surprising quantity of dialogue between players and online gambling operators, using electronic means of communication – typically e-mail. This paper is the first empirical attempt to predict future gambling problems based upon this electronic communication behaviour, comparing it with the evidence of earlier studies that refer to land-based gambling.

In this paper, potential early detection indicators, derived from interviews with the staff of online gambling operators as well as directly derived out of customer correspondence, were investigated for their predictive power in relation to future self-exclusion of the player. Indicators, as the frequency of customer services contacts, the tonality of the mails and several predictors related to the content of the mail (e.g. e-mails about payments and financial transactions) could be shown to relate to the criterion. Applying multivariate methods, a predictive model could be generated, achieving an adequate classification rate of 76.6%. This underlines that customer correspondence is a useful predictor for the emergence of gambling-related problems on the internet as well.

When considered in greater detail, it can be seen that the predictors for problem gambling in land-based venues are based on a combination of the more easily observable indicators of gambling behaviour (e.g. duration or number of EGMs used) and of physical expressions of distress (e.g. aggression, sweating or groaning). In online gambling, these particular physical expressions of distress are not available for observation – all predictors have to be observable within written communication. Therefore, the explicit manifestation of problems – which DSM-derived indicators would be – is very rare. While it has to be assumed that, for the majority of the group of self-excluders investigated in the study, one or more DSM indicators would apply, they rarely communicate any information hinting at these indicators. In fact results show, contrary to our expectations after the screening interviews, that incipient problems often do manifest themselves in a more indirect way. While financial problems themselves are not directly communicated, financial transactions grow burdensome and involve a lot of communication. Issues that do not seem to indicate problems when taken alone, occur with considerably higher frequency and urgency. These indicators could be explained by the increased gambling involvement. Situations that moderate gamblers can easily handle on their own, like the administration of their gambling account, might no longer be manageable by these gamblers, resulting in a request to the operator. While moderate gamblers do not care about the fact that a request is not resolved within an hour, urgency becomes unbearable for those who will develop problems. In addition, there is a noticeable shift in tonality. Future self-excluders display a tendency to use threats and abusive language more frequently. This could possibly be an effect of the increasing psychological strain they are subjected to, and an indication that they are no longer able to cope with it.

Limitations

As a pilot, this study has certain limitations. First and foremost it is unclear how far the predictive model described in this paper can be generalized. In particular, the composition of the predictor set should be subject to further research, since it partially deviated from what we expected based on the very small sample of qualitative interviews. As communication might vary between different gambling operators, who offer different products or target different customers segments, the validity of the model could be limited to the gambling operators who provided communication data and could potentially not be generalizable over different languages or national markets. Therefore a follow-up cross-validation study should try to replicate the model with data from a different background.

To allow for a practical application of the model, a random sample of controls was drawn. The composition of the control group is therefore representative for the background against which future self-excluders should be identified. However this results in a potentially different exposure to gambling for self-excluders and controls. While the predictors have been selected so that this difference has no impact on the interpretation of the model, still this effect should be investigated in more detail. Among those players, used as controls in this paper, there could be some who have finally excluded themselves at a later point in time. Follow-up studies could apply matched samples, paring self-excluders and controls by their time at risk.

Furthermore the criterion of self-exclusion should be addressed. While it can be assumed that the majority of self-excluders have experienced gambling-related problems (Ladouceur, Jacques, Giroux, Ferland & Leblond, 2000), self-excluders only constitute a fraction of all problem gamblers (Nowatzki & Williams, 2002). While it is typical for predictive approaches that the criterion definition only covers manifest characteristics of the problem sought to be predicted, this aspect could be critical if self-excluding problem gamblers were to systematically differ in their communication-based indicators from non-excluding problem gamblers. It should therefore be investigated whether the indicators identified in this study can also be generalized not only to players who decide to self-exclude at different stages of their problem development – as done in this paper – but also to players who would not opt to self-exclude on their own at any point in time.

Conclusion

The findings of this paper demonstrate that, on the internet it is possible to detect future gambling problems not only based on actual gambling behaviour but also based on communication behaviour. The frequent speculation that there is a risk of a lack of contact between the gambler and the operator in online gambling is theoretical. When adapting the methods of monitoring to the properties of the new medium of communication, the alleged disadvantage vanishes. In fact, the internet offers a very effective framework for consistent monitoring reviewable over years, and a highly objective analysis of risk-indicators based on communication behaviour. Adding these new indicators for future gambling-related problems to existing policies could increase the hit-rate for early detection as well as the time interval in which emerging problems are identified in advance. This again might be used to trigger automated (e.g. displaying informative pop-up messages) or individualized interventions (e.g. actively approaching players identified to be at risk and arranging a pre-commitment for gambling expenses).

While this study points out the feasibility of detecting gambling problems before they arise through the analysis of customer communication, it also underlines the importance of regular and dedicated training for employees facing the customer (Meyer & Hayer, 2008). As our data showed, unmistakable indicators for problems are very rare and therefore reliance on single risk indicators, able to predict gambling-related problems with an acceptable level of specificity, necessarily leads to a low sensitivity of detection. Early detection policies should therefore not solely rely on ‘risk keywords’. Analogous to the findings of Schellinck and Schrans (2004) for the land-based setting, combining several observations, which are no predictor for future problems taken alone, can improve the validity of prediction and provide composite markers of adequate sensitivity and specificity. The practical implementation of objective early detection protocols however requires not only precise and structured records of past customer correspondence, but also a high level of expertise in customer services. As quality of judgment formation is also dependent on workload, creating dedicated teams – who exclusively deal with problem gambling cases – appears to be the most promising approach.

However, the detection of future gambling problems based on customer communication is only possible for those people who already communicate. As there seems to be a sub-group that does not communicate at all, a predictive model framework should always consist of several components. Effective prediction of gambling-related problems will have to monitor various aspects of player behaviour (e.g. gambling behaviour, communication behaviour or payment behaviour), collecting and aggregating information for qualified employees, allowing them to base their decisions on objective data.

Acknowledgements

This research was supported by a grant from the European Gaming and Betting Association. Funding bodies had no influence over the design and conduct of the study, and analysis and interpretation of the data. We would like to thank bwin Interactive Entertainment AG, PartyGaming Plc and Unibet Group Plc for supporting the research by providing access to anonymized customer communication data.