تعیین کمی تأثیر دوستان و رفتار ضد اجتماعی در نوجوانان معتاد به مصرف حشیش با استفاده از مدل ZINB و داده کاوی

Abstract Cannabis is the most consumed illegal drug in Europe and its repercussions are more important when taken up at an early age. The aim of this study is to analyse and quantify the predictive value of different personal, family and environmental variables on the consumption of cannabis in adolescence. The sample is made up of 9284 adolescents (47.1% boys and 52.9% girls) with an average age of 15.59 years (SE = 1.17). The ZINB model highlights, as factors that increase the number of joints consumed per week, consumption by the peer group, nights out during the week, gender, the production of forbidden behaviour and the use of other substances, whereas the risk factors for the consumption of cannabis are consumption by friends, ease of access, production of forbidden behaviour and the use of other substances. Association rules highlight the relationship between cannabis consumption, ease of access, production of forbidden behaviour and tobacco consumption. Finally, decision trees enable us to predict cannabis consumption as well as the number of joints an adolescent will consume per week based on the production of forbidden behaviour, consumption of other substances and number of friends who consume cannabis. The results of this work have practical implications concerning the prevention of cannabis consumption in an adolescent population.

. Introduction Cannabis is the most consumed illegal drug in Europe and is the one that people begin to consume at the earliest age (European Monitoring Centre for Drugs and Drug Addition, 2008). The repercussions of the use of this substance have been well described and its effects are even more important when consumption is initiated in adolescence, involving deteriorated academic performance, consequences in long term social adaptation and greater odds of using other illegal drugs (Broman, 2009, Brook et al., 2008, Fergusson and Boden, 2008a, Fergusson and Boden, 2008b, Georgiades and Boyle, 2007, Hall, 2009, Jeynes, 2002, Lessem et al., 2006 and Zimmerman et al., 2005). The main aims of prevention are not only to decrease the prevalence of consumption, but also to reduce the quantity of consumption among consumers and, in order to achieve this, it is necessary to identify the risk factors that lead to greater consumption among adolescents who have already taken up consumption. In this sense, many studies have insisted on the existence of different variables related with the family environment as risk and protection factors involved in drug consumption by adolescents (Fernández et al., 2003 and Olsson et al., 2003). Thus, how fathers/mothers bring up their children, together with the type and degree of monitoring exercised by parents, directly influence vulnerability to consumption, in the sense that adolescents who perceive less parental monitoring will have greater odds of consuming addictive substances (Adalbjarnardottir and Hafsteinsson, 2001, Barrett and Turner, 2006 and DiClemente et al., 2001). Nevertheless, the relationship between parental practices and the use of substances seems to be mediated by the number of friends who are consumers (Simons-Morton, 2007). Indeed, one of the explanatory variables that stands out due to its influence on drug consumption in adolescence is the use of these same drugs by the peer group (Ciairano et al., 2008, Dick et al., 2007, Kokkevi, et al., 2007 and Kuntsche and Delgrande, 2006). Furthermore, certain personality factors, such as antisocial behaviour, thrill seeking or impulsivity have been related to the use of addictive substances at this age of development (Fergusson et al., 2007, Fothergill et al., 2009, Franken et al., 2006, Grant et al., 2010, Hayatbakhsh et al., 2008 and Wong et al., 2006). On the other hand, since Kandel posed the gateway theory (Kandel, 1975), that is, that the use of certain addictive substances like alcohol or tobacco exerts a causal influence on the use of other drugs like cannabis (Kandel, 2003), several studies have found a relationship between the use of cannabis and the consumption of other substances such as tobacco or alcohol (Agrawal et al., 2010, Korhonen et al., 2010 and Pérez et al., 2010). The relationship between the aforementioned risk factors and the consumption of substances such as cannabis has been analysed in several works in terms of a statistically significant relationship, odds of consumption or explained variability, but few studies have measured the power of these variables to increase or decrease, in quantitative terms, the number of joints consumed a week. In this sense, the appropriate model for analysing count data is the Poisson regression model (Long, 1997). Besides, there are even fewer studies which try to explain these relationships using tools that can provide new perspectives which go beyond traditional data analysis techniques. In this context, Data Mining technology, defined as the extraction of useful information from large databases, contains relatively little used tools to date which can be used to exploit information concerning substance consumption. Two popular examples of some descriptive and predictive tools included in this methodology are, respectively, association rules and decision trees (see Han and Kamber, 2006, Kantardzic, 2003, Larose, 2005, Larose, 2006, Witten and Frank, 2005 and Ye, 2003). The aim of this study is to analyse the predictive value of different personal, family and environmental variables on the consumption of cannabis in an adolescent population using appropriate modelling techniques and descriptive and predictive Data Mining tools capable of extracting interesting relationships from the data.

Results First of all, the prevalence of cannabis consumption in the sample is shown (Table 1). Table 1. Prevalence of adolescent consumers of cannabis in the sample (n = 9284). Current consumption of cannabis N % I have never consumed 5599 63.0 I have tried it a couple of times 1259 14.2 I used to, not any more 389 4.4 Occasionally 673 7.6 At weekends 404 4.5 During the week 215 2.4 Daily 346 3.9 Table options Due to the low frequency of cannabis consumption in the sample, and with the aim of being able to predict both cannabis consumption and lack of consumption, a sample equalling cannabis consumption was constructed. In this way, all the adolescents who reported consuming cannabis were selected (n = 1638) and, from the remaining non consumers, a random sample of adolescents who did not consume cannabis was chosen (N = 1638). As already mentioned, the Poisson regression model (PRM) is the appropriate model for analysing count data (Long, 1997) and was therefore used to measure the influence of the risk factors on the number of joints consumed a week. One of the basic assumptions of the PRM is that of equidispersion, which can be evaluated using the regression test proposed by Cameron and Trivedi (1990). After confirming the failure of this assumption, the models called Negative Binomial Regression Model (NBRM), Zero Inflated Poisson (ZIP) and Zero Inflated Negative Binomial (ZINB) were compared, and the ZINB model was observed to best fit the data ( Greene, 1994, Lambert, 1992 and Mullahy, 1986). The statistically significant variables are given in Table 2. The percentages that appear in this table provide information as to how much the number of joints consumed a week increases or decreases. It is possible to appreciate the influence of cannabis consumption in the group of friends on the consumption of this substance by an adolescent, which grows as the number of friends who are consumers rises. Thus, if all the friends consume cannabis the number of joints consumed per week increases 348.3%, and 108.2% if half of the friends consume cannabis, with respect to if no friend is a consumer. Likewise, for every night they go out between Monday and Thursday, 113% more joints are smoked per week. It is also worth noting that if the adolescent is a girl, she will consume 40.3% less per week. Finally, it is possible to observe the influence of the production of forbidden, illegal behaviour and the consumption of other substances, such as alcohol and tobacco, on the weekly consumption of cannabis. Table 2. ZINB model. Percentage change in expected count for cannabis users. Variable p % Gender < 0.001 − 40.3 All friends consume cannabis < 0.001 348.3 Most friends consume cannabis < 0.001 266.3 Half of friends consume cannabis 0.004 108.2 Nights I go out from Monday to Thursday 0.005 11.3 I do forbidden, illegal things < 0.001 107.3 Standard Drink Units (SDU) per week < 0.001 1.6 Number of cigarettes consumed a week < 0.001 0.8 Only statistically significant variables have been shown. Table options As well as providing information about the factors that increase and decrease the number of cigarettes consumed a week, the ZINB model calculates the risk and protection factors associated with the consumption of cigarettes. The risk factors that turned out to be statistically significant are shown in Table 3. The percentages that appear in the table provide information as to how much the odds of continuing being a smoker increases (or decreases). That is, as the number of friends who are smokers increases, the odds of continuing without consuming cannabis decrease, and if an adolescent perceives that it is easy to gain access to the substance, the odds of continuing as a non smoker decrease by 72.9%. The consumption of other substances and the production of forbidden, illegal behaviour also constitute risk factors for the consumption of cannabis. Table 3. ZINB model (binary equation). Factor change in odds of never using cannabis. Variable p % All friends consume cannabis < 0.001 − 98.2 Most friends consume cannabis < 0.001 − 97.0 Half of friends consume cannabis < 0.001 − 96.1 Few friends consume cannabis 0.003 − 88.4 Ease of access 0.005 − 72.9 I do forbidden, illegal things 0.005 − 55.9 Standard Drink Units (SDU) per week < 0.001 − 33.9 Number of cigarettes consumed a week 0.003 − 96.4 Only statistically significant variables have been shown. Table options Association rules (AR) (Agrawal, Imielinski, & Swami, 1993) originated with the study of transactions in department stores and their aim is to identify patterns without “a priori” knowledge of reality. ARs express relationships between items in the form of rules of the type “If Antecedent, then Consequent”. With the aim of finding interesting relationships between the data, the ARs were calculated using the classical a priori algorithm (Agrawal & Srikant, 1994). The interestingness of the ARs was evaluated using the best known indexes that express their degree of uncertainty: support, confidence and lift (Han & Kamber, 2006). The program finds 989,908 ARs, 4260 of which lead to the consumption of cannabis. Table 4 shows the ten ARs that have the best lift indexes. On the whole, the ARs obtained highlight the relationship between the consumption of cannabis, ease of access to the substance, the production of forbidden, illegal behaviour and tobacco consumption. In the first AR concerning the use of cannabis, the interpretation is the following: “In 11.81% of adolescents the three conditions are found together (he/she finds it easy to get cannabis, does forbidden, illegal things, does not break, burn or damage other people's property and smokes tobacco) and, what is more, 96% of these individuals also consume cannabis”. The lift, greater than 1, indicates that the AR found is useful. Table 4. Association rules with interestingness measurements (information generated with the arules package integrated in the freely distributed R program, version 2.10.1). Association rules for prediction of cannabis use Support Confidence Lift If he/she finds it easy to obtain cannabis, does forbidden, illegal things, does not break, burn or damage other people's property and consumes tobacco; then he/she consumes cannabis. 0.1181319 0.9602978 1.920596 If he/she finds it easy to obtain cannabis, is bored of always doing the same things, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1101954 0.9601064 1.920213 If he/she does extreme sports/activities, likes new, exciting experiences, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1007326 0.9593023 1.918605 If he/she finds it easy to obtain cannabis, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1413309 0.9585921 1.917184 If he/she finds it easy to obtain cannabis, likes new, exciting experiences, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1330891 0.9582418 1.916484 If he/she finds it easy to obtain cannabis, thinks he/she has good qualities, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1120269 0.9582245 1.916449 If he/she is self-satisfied, likes new, exciting experiences, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1028694 0.9573864 1.914773 If he/she finds it easy to obtain cannabis, can't wait to have things, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1095849 0.9573333 1.914667 If he/she is self-satisfied, does forbidden, illegal things and consumes tobacco; then he/she consumes cannabis. 0.1074481 0.9565217 1.913043 If he/she finds it easy to obtain cannabis, does forbidden, illegal things, consumes alcohol and consumes tobacco; then he/she consumes cannabis. 0.1272894 0.9564220 1.912844 Table options Fig. 1 shows a decision tree for the number of joints consumed per week (n = 3276) calculated using the Classification And Regression Trees (CART) algorithm ( Breiman, Friedman, Olshen, & Stone, 1984). The decision tree enables us to predict the number of joints an adolescent will consume a week based on the production of forbidden, illegal behaviour, alcohol consumption and number of friends who consume cannabis. Decision tree for weekly cannabis consumption generated using the CART algorithm ... Fig. 1. Decision tree for weekly cannabis consumption generated using the CART algorithm (Breiman et al., 1984). Figure options Fig. 2 shows a decision tree for cannabis consumption (n = 3276), obtained using the same algorithm as above, which enables us to predict, with a risk of 0.153, whether or not an adolescent will smoke joints based on the consumption of tobacco, the number of friends who are consumers in the peer group and the production of forbidden, illegal behaviour. Decision tree for any cannabis consumption generated using the CART algorithm ... Fig. 2. Decision tree for any cannabis consumption generated using the CART algorithm (Breiman et al., 1984).