The author of this study accessed web-based anonymous gambler chat meeting data in Japan [29]. The meetings were conducted since September 2008 through web-based text chats with no pictures or sounds. A few offline meetings were conducted for a year in Osaka, Tokyo, or Sapporo. The web-based anonymous gambler chat meetings has no relationship with the Gamblers Anonymous group [28]. To post in the meetings, users had to be approved by the administrator in advance. Individuals who could participate in the meetings were those who were experiencing a gambling problem or their family members who were experiencing a gambling problem. The meetings were conducted by gamblers without any strict rules, and no outside experts were involved at all. The meetings were divided into those for users aiming for 1 week, 1 month, 3 months, 6 months, 1 year, 3 years, and more than 3 years of gambling abstinence, and users attended the meetings based on the number of days they had remained gambling-abstinent. However, these meetings had no strict rules, and a user aiming for 3-month gambling abstinence may attend a meeting for those aiming for 1-month gambling abstinence and vice versa. There were meetings for chatting as well as for providing daily reports and reports of memorial days (such as 100 abstinent days). Furthermore, these meetings were always open and free to join; therefore, the time and duration of participation varied widely among users (Table 1). Table 1 shows a comparison of the demographic variables, gambling history, gambling symptoms, participation forms, and characteristics of the utterances in web-based anonymous gambler chat meetings between abstinent and nonabstinent gamblers. In Table 1, abstinent gamblers refers to gamblers who have remained abstinent without relapse for at least three years. The utterance classifier in this study involves 6 clusters, but only participants’ change talk and sustain talk clusters were used in this study. The data size of the demographic variables and gambling histories was smaller than that of the utterances. The proportion of change talks was the number of change talks per the total number of change and sustain talks. Some users did not have any change and sustain talks; thus, their proportions were not calculated. When a single meeting exceeded 5000 posts, it would end and a new meeting would be started. These meetings spanned multiple days or months and were rarely completed within a day.

Research focusing on 3-year continuous abstinent periods is limited [27]; therefore, this study used 3-year-continuous abstinent smokers’ studies [25] because smokers and gamblers share similar biological and environmental factors [30]. In the previous study [25], 13.8% of smokers achieved 3-year abstinence with the help of treatment whereas 86.2% did not. Based on the abovementioned findings [25], the allocation rate between abstinent and nonabstinent gamblers was set as 0.16. To show satisfactory power (0.95) with alpha (.05) for small effect size (d=0.2), 2736 participants were required.

A total of 134 web-based anonymous gambler chat meetings were web scraped on March 10, 2020 (Figure 2) [29]. Among them, 35 meetings included private data and were thus excluded. From the 99 remaining meetings, 3967 users were identified based on their anonymous names. They posted Japanese texts in the meetings from September 8, 2008 to March 10, 2020. Among them, 1139 gamblers were excluded because their posts included lesser than 3 words or included only advertisements with “http.” Furthermore, 936 users were excluded because we could not confirm the gambling symptoms in their lifetime based on their posts (the evaluation of gambling symptoms is presented in the next section). The final participants were 1892 gamblers who experienced at least one gambling symptom in their lifetime. The number of study participants (n=1892) did not reach the ideal sample size (n=2736). To show satisfactory power (0.95) with alpha (.05) in the 1892 participants, the effect size must be over 0.24.

Users who registered as gamblers were provided with a personal counter that automatically counts the number of days wherein they ceased gambling. If they gamble, they report it to the administrator who resets the number of days on the counter. Those whose counters exceed 3 years are listed on a separate website and exemplified as models who have stopped gambling. According to the list on March 2020, 267 gamblers were regarded as abstinent gamblers without relapse for at least 3 years with 639,632 utterances; the other 1625 gamblers were regarded as nonabstinent gamblers with 996,231 utterances. These counters and lists were created by the administrator without the involvement of the author.

This study was approved by the ethics committee of a National University in Japan (Reception number 222, ethical review on September 1, 2020). Furthermore, all procedures were conducted in accordance with the guidelines for studies involving human participants, the ethical standards of the institutional research committee, and the revised 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

The gamblers examined in this study were primarily males (1125/1311, 85.8%) and aged around 35 years (Table 1). Although their sex, age, total amount of debt, and history of gambling were not significantly different, abstinent gamblers who had remained abstinent without relapse for at least three years experienced significantly more gambling symptoms (P=.01) than nonabstinent gamblers (Table 1). Although their gambling symptoms such as gambling tolerance, unsuccessful control over gambling, and reliance on others to provide money were not significantly different, abstinent gamblers had significantly higher preoccupation (P<.001) with gambling and lies for gambling compared with nonabstinent gamblers. The nonabstinent gamblers were significantly more experienced in illegal acts of gambling (P<.001) than abstinent gamblers. These findings indicated that abstinent gamblers experienced more gambling symptoms compared with nonabstinent gamblers although the effect size was small (d=0.179-0.261).

Abstinent gamblers participated significantly longer in the meetings and talked more than nonabstinent gamblers (Table 1, both P<.001, d=0.590 and d=0.462, respectively). The average probability of change talks was not significantly different between abstinent and nonabstinent gamblers; however, the average probability of sustain talks in nonabstinent gamblers was significantly higher than that in abstinent gamblers (Table 1, P<.001, d=–0.369). Hence, the proportion of change talks’ probabilities (average probability of change talk/average probabilities of change and sustain talks) in abstinent gamblers was significantly higher than that in nonabstinent gamblers (Table 1, P<.001, d=0.285). Similarly, the proportion of change talks (number of change talks/number of change and sustain talks) in abstinent gamblers was significantly higher than that in nonabstinent gamblers (Table 1, P<.001, d=0.280). These findings indicate that abstinent gamblers showed higher proportion of change talks in the meetings compared with nonabstinent gamblers. To clarify the proportion of change talks between abstinent and nonabstinent gamblers, dynamic differences in the utterances between abstinent and nonabstinent gamblers were compared with similar number of utterances. Figure 4A shows the accumulated numbers of change talks between abstinent (ID# 2790) and nonabstinent (ID# 317) gamblers. Figure 4A indicates that abstinent gamblers produced more change talks than nonabstinent gamblers. Figure 4B shows the accumulated numbers of sustain talks between them. In contrast with that seen in change talks, nonabstinent gamblers produced more sustain talks than abstinent gamblers. Figure 4 shows that the utterances of abstinent and nonabstinent gamblers in the meetings showed dynamic differences.

To visualize the dynamic differences between abstinent and nonabstinent gamblers, the change talk model for abstinence, involving the number of change and sustain talks as independent variables on x and y axes, respectively, classified 267 abstinent gamblers and 1625 nonabstinent gamblers (Figure 1A). Before classification, the correlations among the number of change talks, the number of sustain talks, and gambling abstinence were checked. The number of change and sustain talks were positively correlated with gambling abstinence (r=0.247, r=0.182, both were P<.001, n=1892). Moreover, the number of change talks was positively correlated with the number of sustain talks (r=0.676, P<.001, n=1892). These correlations indicated the positive relation of these talks with gambling abstinence.

Figure 1A shows abstinent and nonabstinent gamblers scattered on a map with the number of change and sustain talks on x and y axes, respectively. The abstinence likelihoods of gambling were also color-coded based on the SVM with RBF (accuracy was 0.8555, where cost of penalty and complexity of boundary area were 10¹ and 10^–1, respectively). The strong red color indicates low abstinence likelihood (negative evaluation scores) whereas the strong blue color indicates high abstinence likelihood (positive evaluation scores). Figure 1A shows that participants who increased the number of change and sustain talks were likely to move from the red zone to the blue zone, which indicated that they increased their abstinence likelihood of gambling.

To clarify the clinical use of our model, 1 beginner and 2 intermediate gamblers were plotted, and the individual evaluation values for abstinence as well as their ideal proportion of change talks were shown through the gradient descent method (Figure 1B). To clarify the meanings of the number of change and sustain talks, the numbers were unstandardized. The beginner gambler (ID# 2665) started at 1 change talk and 1 sustain talk with evaluation value –1.000 and finished at 1657.16 change talks and 351.31 sustain talks with an evaluation value of –0.588 after 500 trials. The beginner gambler’s (ID# 2665) evaluation value at the end remained negative, thereby indicating low abstinence likelihood. The ideal portion of change talks (number of change talks/number of change and sustain talks) for this gamble during the web-based anonymous gambler chat meetings was regarded as 0.8254. One intermediate gambler (ID# 2162) started at 2130 change talks and 441 sustain talks with evaluation value –0.1167 and finished at 3731.94 change talks and 813.04 sustain talks with evaluation value 2.507 after 500 trials. His evaluation value became positive at the end, thereby indicating high abstinence likelihood. The best portion of change talks for him during the meeting was considered 0.8115. The other intermediate gambler (ID# 1008) started at 3396 change talks and 259 sustain talks with evaluation value –0.063 and finished at 4290.42 change talks and 787.67 sustain talks with evaluation value +2.815 after 500 trials. This gambler’s evaluation value turned positive at the end, thereby indicating high abstinence likelihood. The best portion of change talks for this gambler during the meetings was considered 0.6285. These findings indicate that the evaluation values are personalized feedback for gamblers. Furthermore, high proportions of change talks (over 0.80) were required for most beginner gamblers, whereas the best proportions of change talks were different among intermediate gamblers.

The change talk classifier used in this study indicated 93% precision and analyzed 1.63 million utterances in web-based anonymous gambler chat meetings (Figure 3), thus leading to the development of the change talk model for gambling abstinence. The abstinent map with evaluation values showed the abstinence likelihoods of each gambler (Figure 1). Based on suggestions from a previous study on anonymous group meetings [4,5], this study confirmed that the abstinence process in web-based anonymous gambler chat meetings was similar to the process in other standardized therapies. Consistent with the results of previous studies regarding face-to-face motivational interviewing for gamblers [16,18,22,23], the proportion of change talks was also linked with improved outcome. In particular, the proportions at the initial stages were positively linked with 3-year continued gambling abstinence. This study advances previous findings regarding language classifiers in mental health fields [42,43] in 2 ways. First, the classifier used in previous studies classified the onset of mental disorders, whereas the classifier in this study classified recovery from a mental disorder. Second, the classifiers used in previous studies served as a screening tool for the early detection of mental disorders, whereas the classifier in this study is a therapeutic tool for identifying the recovery process. The 2D map with the abstinence likelihood slope used in this study was helpful for showing personalized evaluation values and ideal proportions of change talks, responding to every gambler’s utterance in the meetings.

This study involves 2 new methodologies: a change talk classifier and long-term data with dropout gamblers. First, this study used machine-learning methods and automatically classified gamblers’ utterances (Figure 3) [35]. Studies have utilized human expert resources to classify gamblers’ utterances [12,15,18-20,44-47], with limited data size (around 10³) and generalization of their findings. The change talk classifier used in this study increased data size (over 10⁶) similar to recent mental health research [8,42,43] and was applicable to other therapy processes focusing on change talks such as group and individual therapies for healthy dieting and prevention of substance abuse [16,23]. The utilization of the change talk classifier for these therapies can save human resources about change talk classification [12,19,20] and broaden the applicability of change talk classifications or the core scheme of motivational interviewing [14-16] for addictive disorders [7,21].

This study also involved long-term data (maximum participation length of 3872 days) with both dropout and ongoing data sets (Figure 2, Table 1). Inclusion of these data sets enabled extensive and detailed slope maps of abstinence likelihoods of gambling. Studies have primarily utilized the initial and final sessions’ scores of gamblers who completed the treatments. However, considering their high dropout rates [48,49], these data sets could bias gamblers’ characteristics. Moreover, the data set involving only the first and last sessions does not show the processes in between, and the model based on these data sets does not reveal the abstinence likelihood corresponding to the therapy processes [50]. Long-term data with dropout and ongoing data sets can be helpful for estimating changes during therapy, particularly for diseases that require extensive treatment periods [25-28].

This study has 5 limitations. First, the number of female participants in the study sample was limited. The abstinence process differs between female and male participants [51]; therefore, future studies must involve more female participants. Second, the current measurement of gamblers’ symptoms was based on their utterances in web-based anonymous gambler chat meetings. Hence, our measurement may underestimate several gambling symptoms because the measurement could not assess these symptoms that gamblers did not reveal in the meetings. To compensate for these limitations, questionnaires [32] or standardized interview [31] data are required in future studies. Third, this study identified gamblers through their anonymous names, but it is possible that 2 different gamblers used the same anonymous name or that one gambler used multiple anonymous names. Their identities could not be checked because of the web-based anonymity settings. Another data set such as physical meetings [24,49] would be helpful for validating the findings of this study. Fourth, the small number of long-term participants is also one of the study limitations. This study model shows the highest likelihood area for gambling abstinence (Figure 1A, top right most blue area); this implies that individuals in this area will not benefit from participating in web-based anonymous gambler chat meetings. This point must be re-examined by collecting more data regarding long-term participants. Fifth, the internal validity of this study was limited because abstinent and nonabstinent gamblers freely interacted and the proportion of their change talks was not manipulated. To increase the internal validity, future studies must consider an experimental design.

Despite these limitations, this study developed the change talk model for gambling abstinence in web-based anonymous gambler chat meetings with the change talk classifier. The change talk classifier used in this study increased the data size of psychotherapies [6,12,13,18,44-47] and broadened the applicability of language analysis on social media for mental health services [42,43]. The change talk model for abstinence used in this study can provide personalized evaluation scores for abstinence and an ideal proportion of change talks to the gamblers responding to every utterance during the meetings. In clinical settings, using this model, participants can determine their abstinence likelihood because the evaluation values are determined based on their previous utterances before they participate in the meetings (Figure 1B). The ideal proportion of change talks can be estimated using this model (Figure 1B) and they can be motivated to comment at the meetings based on that proportion. After the meeting, gamblers can also observe changes in their evaluation values during the meeting, thus allowing them to objectively evaluate their individual progress of each meeting (Figure 1B). These factors are considered to provide personalized feedback for participants [14,17] and can likely lead to improved outcomes [38]. Providing personalized evaluation values and the optimal proportion of change talks, which were also validated in millions of data sets, can increase the abstinence likelihood of gamblers. This may help to prevent severe mental, social, and financial problems caused by the gambling disorder [1-3].