Identifying environmental and social factors predisposing to pathological gambling by standard logistic regression model and Logic Learning Machine

Motivation: In recent years supply of legal and commercial gambling has rapidly increased in western countries. Alongside this trend, an increased prevalence of behavioural disorders associated with gambling has emerged. Recently, the American Psychiatric Association has recognized pathological gambling as a real disease addiction, with a potentially tremendous psychological, social and economic impact. The onset of pathological gambling is commonly preceded by a borderline phase of behavioural disorders, named "problematic gambling" (PG). Recent studies have identified a large set of socio-demographic and behavioural factors associated to PG, namely: age, gender, education, marital status, social class, ethnicity, preferred type of gambling, alcohol consumption, drug dependency and some psychiatric conditions. However, the role of such variables in predisposing to PG is not completely understood. The present study is aimed at identifying the main factors associated to PG by combining information from standard statistical analysis and a recently proposed algorithm of supervised data mining. Methods: Logic Learning Machine (LLM) is an innovative method of supervised analysis based on an efficient implementation of the switching neural network model. LLM produces a set of simple intelligible rules whose accuracy has been demonstrated to be comparable or even superior to that of best machine learning techniques. The LLM classifier is obtained through Boolean function synthesis by using an aggregative policy, i.e., at any iteration some patterns belonging to the same output class are clustered to produce an intelligible rule. Suitable heuristic algorithms are employed to generate rules exhibiting the highest covering and the lowest error. In the present study LLM was applied to a dataset from a cross-sectional investigation carried out in Northern Italy between October and November 2014 aimed at identifying risk factors associated with PG. A set of 256 gamblers (245 men and 11 women) were successfully interviewed using a structured questionnaire. Information included: age, gender, marital status, employment, education, tobacco smoking, alcohol consumption, psychopharmacological treatments, use of illicit drugs, preferred type and place of gambling, frequency, average daily money spent in gambling, and age at the start of gambling. Questionnaire also included items about "magical thinking" (believe in lagging numbers) and having lent itself to bet on behalf of minors. Each subject also filled in a separate questionnaire based on two international standard scales for gambling evaluation: the South Oaks Gambling Screen (SOGS) and the Lie/bet scale. A gambler was classified as problematic if he/she tested positive to both scales. Each participant gave a written consent to participate into the study. The association between PG and the available putative risk factors was evaluated by binomial logistic regression analysis using a standard backward procedure for variables selection. Risk estimates were expressed as Odds Ratios (ORs) and 95% confidence intervals (95%CI) obtained by the Wald method. Relevant predictors of PG were identified and combined in intelligible threshold-based rules using LLM. Results: Ninety-seven individuals tested positive for PG. Predictors most strongly associated to PG in multivariable regression analysis were: daily gambling (OR = 7.7, 95%CI: 3.7-15.7), money expense per day (OR = 6.3, 95%CI: 2.7-14.6), intensive use of new slots (OR = 3.1, 95%CI: 1.3-7.3), believing in magical thinking (OR = 2.6, 95%CI: 1.2-5.5) and having bet for minors (OR =2.6, 95%CI: 1.3-5.4). LLM identified 14 simple classification rules, based on a small set of predictors. For instance, a high risk of PG was associated to: daily gambling, new slots, video lottery, virtual and sporting bets, magical thinking, foreign nationality, alcohol abuse, and a high daily money expense, whereas usually gambling in a betting office was inversely associated to PG. Classification accuracy of LLM, evaluated in leave-k-out cross-validation (k=10%) was 72%, sensitivity 63%, and specificity 86%. Ranking variables by their relevance in LLM confirmed daily gambling as the most important factor. Interestingly, contrarily to results from logistic regression analysis, the amount of money spent in gambling ranked only ninth as importance among PG predictors. A stratified reanalysis highlighted a strong association between money expense and PG among daily players, whereas among occasionally gamblers the association was smaller, indicating that money expense for gambling is a strong PG predictor only among daily players. Classification rules obtained by LLM can add precious information to results from standard statistical analysis providing a suitable method for the identification of problematic gamblers. In particular, LLM is able to identify interactions between risk factors that tend to escape standard methods of analysis.