Addictive Behaviors 32 (2007) 1441 – 1450
Cannabis use when it's legal Jan C. van Ours Department of Economics, Tilburg University, P.O. Box 90153, NL-5000 LE Tilburg, The Netherlands
Abstract This paper addresses the question of whether alcohol and tobacco are “gateways” for cannabis use. To investigate this the relationships between the starting rates in the use of alcohol, tobacco, and cannabis are analyzed. The starting rate for cannabis use appears to be higher for smokers and lower for users of alcohol. Indeed, tobacco use seems to be a gateway for cannabis use. The main policy conclusion is that measures that reduce smoking will also reduce the incidence of cannabis use. © 2006 Elsevier Ltd. All rights reserved. Keywords: Alcohol; Tobacco; Cannabis; Gateway
1. Introduction The concept of progression in drug use has a long history and became known in the 1930s as the steppingstone theory (Kandel, 2004). The stepping-stone theory considered the progression in drug involvement to be inexorable with the use of one drug invariably leading to the use of the next drug. The Gateway Hypothesis also considers the progression in the use from drug to the next but emphasizes that although the use of certain drugs precedes the use of others, progression is not inevitable. In other words, the use of one drug is a necessary but is not sufficient condition for entry into the use of the next drug (Kandel, 2004). Typically, the sequence of drug use starts with the legal drugs, alcohol and tobacco and via cannabis use it proceeds to illicit drugs like cocaine or heroine. This paper focuses on the first part of the sequence and considers the use of alcohol, tobacco and cannabis. Relationships between alcohol, tobacco and cannabis have been studied before both in a contemporaneous and in an intertemporal setting. Concerning the contemporaneous relationship – the relationship in the use at the same moment in time – there seems E-mail address:
[email protected]. 0306-4603/$ - see front matter © 2006 Elsevier Ltd. All rights reserved. doi:10.1016/j.addbeh.2006.10.006
1442
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
be consensus that tobacco and cannabis are complements. Factors that stimulate tobacco consumption like a low tobacco price also stimulate cannabis consumption. Circumstances that stimulate cannabis consumption like decriminilization of cannabis use also have a positive effect on tobacco consumption. Pacula (1998a) and Farrelly, Bray, Zarkin, and Wendling (2001) found complementarity between tobacco and cannabis use for the U.S. while Cameron and Williams (2000) found in for Australia. The contemporaneous relationship between alcohol and cannabis is more puzzling. Chaloupka and Laixuthai (1997) and DiNardo and Lemieux (2001) for the U.S. and Cameron and Williams (2000) for Australia concluded that cannabis and alcohol are substitutes but Saffer and Chaloupka (1999) for the U.S. and Williams and Mahmoudi (2004) for Australia found that cannabis and alcohol are complements. There are also a few intertemporal demand studies that consider how the use of one drug at one point in time influences the use of another drug at a later point in time. Pacula (1998b) found that prior use of alcohol and tobacco increased the likelihood of later cannabis use. From this she concluded that both alcohol use and tobacco use are a gateway to cannabis. Beenstock and Rahav (2002) using data from Israel found a causal gateway effect from cigarettes to cannabis. This paper investigates intertemporal relationships in drug use with a particular interest in the question whether alcohol and tobacco are gateways for cannabis use. According to Kandel (2004) the validity of the Gateway Hypothesis is based on three criteria. First, there has to be a sequence of initiation of use between drug. Second, there has to be an association in the use of drugs, such that use of a drug lower in the sequence increases the risk of using drugs higher in the sequence. Third, the association in the use of drugs should not be through other determinants. Association implies causation if all possibilities for spurious associations have been eliminated. So, it is not easy to establish the empirical relevance of the gateway theory because of an identification problem. The fact that many cannabis users have previously consumed alcohol or tobacco does not necessarily imply that there is a causal relationship. An alternative explanation is that the use of the three drugs is correlated because certain characteristics make some individuals more susceptible to alcohol, tobacco and cannabis. There have been epidemiological, sociological, psychiatric and a few econometric studies investigating the empirical relevance of the Gateway Hypothesis. However, most of the studies do not take the identification problem into account and do not distinguish correlation from causality (Beenstock & Rahav, 2002). The current paper presents an econometric analysis of the dynamics of drug consumption with a focus on the question whether or not alcohol and tobacco are gateway drugs leading to cannabis. The analysis is based on Amsterdam data and investigates the relationship between individual starting rates with respect to alcohol, tobacco and cannabis. The use of Amsterdam data is particularly interesting because in the Netherlands the use of cannabis and to some extent the sale of cannabis is legal. If an individual wants to use cannabis it is not particularly difficult to buy the stuff. If an individual is asked whether he or she uses cannabis there is no particular reason for that individual to deviate from the truth. The econometric analysis used makes it possible to distinguish between causality and association. 2. Method 2.1. The data Amsterdam has a population of 700,000 inhabitants and around 300 recognized, so-called “coffeeshops” where cannabis can be purchased. The sale of small quantities of cannabis is technically an
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
1443
offence, but prosecution proceedings are only instituted if the operator or owner of the shop does not meet certain criteria. These criteria are that no more than 5 g per person may be sold in any one transaction, no hard drugs may be sold, drugs may not be advertised, the coffee-shop must not cause any nuisance, no drugs may be sold to persons under the age of 18, which may not be admitted to the premises. The mayor may order a coffee-shop to be closed. Cannabis is smoked with the most popular technique being to make a kind of cigarette – a “joint” because consumption is usually communally. Joints in the Netherlands typically combine tobacco and a small amount of high potency cannabis. Dutch coffee houses offer prerolled joints containing about 0.1 g of cannabis while street joints average around 0.25 g.1 This paper uses data on drug use collected by CEDRO, the Center for Drug Research of the University of Amsterdam in 1994, 1997 and 2001 (see Abraham, Kaal, and Cohen (2003) for a detailed description). There are some differences between these surveys, but the information used in this paper is collected consistently through time. The data on drug use are based on self-reported information, which is the norm for analyses of drug consumption. The survey population is defined as all persons in the Municipal Population Registry of Amsterdam. In 1994 two interview methods were used, a written and a computerassisted version (using laptop computers where the interviewer typed in the answers directly). The sample was randomly subdivided into two equal sized samples. It turns out that the interview method did not affect the answers to the questions. The 1997 survey was fully computer-assisted. The 2001 survey was based on a mixture of methods. Respondents could choose between a paper questionnaire, a computer-assisted faceto-face interview, an interview by telephone, via their own computer on the Internet or on a floppy disk (by mail). The non-response in 1994 was 49.2%, in 1997 48.1%, and in 2001 60%. The available data refer to all inhabitants of Amsterdam of 12 years and older. The sample was reduced using a number of criteria. Because the focus of the paper is on the age of onset, only individuals aged from 26 to 50 are considered. Because some studies find individuals from ethnic minority groups to under-report drug consumption, the focus is on individuals born in the Netherlands with a Dutch nationality. After removing observations with incomplete information, the net samples contain 2227 females and 1970 males. 2.2. Measures In the analysis the following variables were used – Male: Dummy variable with a value of 1 if the individual is male and a value of 0 otherwise; Age: Age of individuals at the time of the survey; Secondary education: Dummy variable with a value of 1 if the individual attended secondary general or vocational education, and a value of 0 otherwise. Secondary education refers to intermediate vocational or secondary general education; Higher education: Dummy variable with a value of 1 if the individual attended higher vocational or academic education, and a value of 0 group consists of individuals with basic or primary education; Birth year: Year of birth, calculated as (year of survey – age – 1950)/10; Cannabis use parents: Dummy variable with a value of 1 if one or both parents have ever used cannabis and a value of 0 otherwise. Table 1 presents some characteristics of the dataset used in the analysis. As shown the share of males is about 47% while the average age is 36.7 years. About 26% of the sample has secondary education while 48.5% has higher education. The birth year variable indicates that individuals were born between 1944 and 1975. About 7.1% of the individuals had parents that used cannabis. 1
In the U.K. cannabis is also used mostly in combination with tobacco and a joint contains about 0.15–0.35 g of cannabis. In contrast in the U.S. tobacco is rarely used while the cannabis has a low potency. On average an American joint contains about 0.5 g of cannabis (United Nations, 2006).
1444
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
Table 1 Characteristics of the dataset
Male Age Secondary education Higher education Birth year Cannabis parents Age of onset alcohol Age of onset tobacco Age of onset cannabis
Mean
Minimum
Maximum
N
0.469 36.7 0.264 0.485 1.044 0.071 16.1 16.4 19.5
0 26 0 0 −0.6 0 10 10 10
1 50 1 1 2.5 1 49 47 48
4197 4197 4197 4197 4197 4197 3931 3173 2120
Note that ‘birth year’ is calculated as 0.1 ⁎ (year of survey – age – 1950).
Table 1 also provides information about the age of onset – for those that started using a particular drug. For alcohol this is 16.1, for tobacco 16.4 and for cannabis 19.5 years. So the age of onset for cannabis is on average about 3 to 3.5 years higher than for alcohol and tobacco. From the numbers of individuals with information about the age of onset one can derive that at the time they were surveyed 93.7% of the individuals in the sample had used alcohol, 75.6% has used tobacco and 50.5% had used cannabis. Based on the age of onset empirical starting rates can be calculated indicating the probability to start using at a particular age conditional on not having started consuming up to that age. Fig. 1 shows the empirical starting rates for alcohol, tobacco and cannabis use. The starting rates are calculated on the assumption that an individual who did not use alcohol, tobacco or cannabis but is below age 50 has an incomplete duration of non-use. There are clear spikes at age 16, 18, 20, 25 and 30. In terms of smoking and drinking nothing much happens after age 20; the same holds for cannabis, but then after at 25. The cumulative starting probability for alcohol reaches its maximum right after age 20 at a level of about 95%. This indicates that it there is a group of about 5% that will not use alcohol at all. For smoking the group of abstainers is about 25%, for cannabis this is about 50%.
Fig. 1. Starting rates alcohol, tobacco and cannabis by age.
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
1445
2.3. Statistical analysis The relationship between the use of alcohol, tobacco and cannabis is studied using a multivariate mixed proportional hazard model with a flexible baseline hazard (see Van Ours (2003) for a similar approach on the use of cannabis and cocaine). This type of models is used in labor market studies to establish the impact of policy interventions on unemployment durations. Treatment effects – the effects of policy interventions – can be estimated in the context of a multivariate duration model. A major advantage of this approach is that identification of the treatment effect does not rely on a conditional independence assumption neither is it necessary to have a valid instrument. Given that economic theory does not suggest a natural instrument, this is a particularly useful feature of this approach. Abbring and Van den Berg (2003) provided a formal proof of the non-parametrical identification of the treatment effect in a bivariate duration model. They show that in this duration model framework the ‘timing-of-events’ allows identification without the usual restrictions. Omitting a subscript for individual the starting rate for alcohol, tobacco, and cannabis, at age t (t ≥ 10) conditional on observed characteristics x and unobserved characteristics v is specified as: hj ðtjx; vj Þ ¼ kj ðtÞexpðx Vbj þ vj Þ for j ¼ a; b
ð1Þ
hc ðtjx; vc ; ta ; tb Þ ¼ kc ðtÞexpðx Vbc þ da Iðta btÞ þ db Iðtb btÞ þ vc Þ
ð2Þ
where λ(t) represents individual age dependence, v represents individual specific unobserved heterogeneity, the subscripts a represents alcohol, b tobacco, and c cannabis. The δ's indicate whether or not alcohol and tobacco influence the starting rate for cannabis use. We model flexible age dependence by using a step function: kj ðtÞ ¼ expðRk kjk Ik ðtÞÞ for j ¼ a; b; c
ð3Þ
where k (=1,.., 21) is a subscript for age-interval and the Ijk (t) are time-varying every year is represented separately. The last interval refers to age over 30. Because a constant term is also estimated, λj1 is normalized to 0. The conditional density functions of the completed durations of non-use can be written as Z tj hj ðsjx; vj ÞdsÞ for j ¼ a; b ð4Þ fj ðtj jx; vj Þ ¼ hj ðtj jx; vj Þexpð− 0
Z
tc
fc ðtc jx; vc ; ta ; tb Þ ¼ hc ðtc jx; vc ; ta ; tb Þexpð−
hc ðsjx; vc ; ta ; tb ÞdsÞ
ð5Þ
0
The possible correlation between the unobserved components is taken into account by specifying the joint density function of the three durations of non-use ta, tb, and tc conditional on x as Z Z Z f ðta ; tb ; tc jxÞ ¼
fa ðta jx; va Þfb ðtb jx; vb Þfc ðtc jx; ta ; tb ; vc ÞdGðva ; vb ; vc Þ va
vb
vc
ð6Þ
1446
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
Table 2 Parameter estimates independent starting ratesa)
Male Secondary education Higher education Birth year Cannabis parents δalcohol δtobacco Unobserved heterogeneity α1 Prob. pos rate (%) – Loglikelihood No unobserved heterogeneity – Loglikelihood
Alcohol
Tobacco
0.27 (6.4)⁎⁎ 0.30 (5.6)⁎⁎ 0.57 (11.4)⁎⁎ 0.33 (11.8)⁎⁎ 0.42 (4.3)⁎⁎ – –
− 0.13 (3.0)⁎⁎ − 0.27 (4.3)⁎⁎ − 0.47 (8.4)⁎⁎ 0.07 (2.5)⁎⁎ 0.16 (1.8)⁎ – –
0.37 (6.6)⁎⁎ 0.65 (7.8)⁎⁎ 0.84 (11.1)⁎⁎ 0.30 (6.3)⁎⁎ 1.36 (13.9)⁎⁎ −0.00 (0.1) 0.95 (16.5)⁎⁎
2.69 (42.2)⁎⁎ 93.6 8944.2
1.13 (31.31) 75.6 9889.7
0.87 (21.2)⁎⁎ 70.4 8027.6
9178.5
10124.5
Cannabis
8043.2
2227 females and 1970 males; all estimates contains age-specific dummy variables and dummy variables for survey years; a ⁎⁎ (⁎) indicates a parameter estimate significant at the 5% (10%) level.
a)
G(va, vb, vc) is assumed to follow a discrete distribution with 8 points of support (v1a, v1b, v1c), (v2a, v1b, v1c), (v1a, v2b, v1c), (v2a, v2b, v1c), (v1a, v1b, v2c), (v2a, v1b, v2c), (v1a, v2b, v2c), (v2a, v2b, v2c) and nÞ , with n = 1, .., 8 and associated probabilities p1 to p8 are specified as a multinomial logit, so pn = Rexpða n expðan Þ α8 = 0. To allow for the possibility that some individuals will never start using a particular drug the second mass points are imposed to be v2a = v2b = v2c = −∞. 3. Results In the estimates observations of individuals that did not start to consume alcohol, tobacco or cannabis are considered to be right censored durations, i.e. durations that are still incomplete at the time of the survey. The parameters are estimated using maximum likelihood. The analysis starts with estimates of independent starting rates in which the unobserved heterogeneity is assumed to be uncorrelated. Then the likelihood factorizes and each starting rate can be estimated separately. The parameter estimates are presented in Table 2. As shown males and higher educated individuals have higher starting rates for alcohol and cannabis and lower for tobacco than females and lower educated individuals.2 Birth year and parental cannabis use have positive effects on all starting rates. The fact that recent cohorts are more likely to consume may have to do with an income effect (alcohol/tobacco) or increased availability and lower prices (cannabis). The effect of parental cannabis use might indicate that children who grow up in particular families are more oriented towards the use of any kind of drug. Conditional on the observed characteristics the starting rates are influenced by unobserved individual characteristics. For alcohol 93.6% has a positive starting rate while the remaining 6.4% has a zero starting
2
An analysis in which the models were estimated separately for females and males did not generate additional insights into the starting rate processes.
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
1447
Table 3 Parameter estimates correlated starting ratesa)
Male Secondary education Higher education Birth year Cannabis parents δalcohol δtobacco Unobserved heterogeneity α1 α2 α3 α4 α5 α6 α7 – Loglikelihood
Alcohol
Tobacco
Cannabis
0.38 (5.0)⁎⁎ 0.71 (10.0)⁎⁎ 0.35 (8.7)⁎⁎ 0.43 (3.3)⁎⁎ – –
−0.35(4.0)⁎⁎ −0.52 (6.8)⁎⁎ 0.10 (2.4)⁎⁎ 0.16 (1.3)⁎⁎ – –
0.89 (6.8)⁎⁎ 1.19 (9.7)⁎⁎ 0.37 (3.4)⁎⁎ 1.32 (6.6)⁎⁎ −0.25 (2.4)⁎⁎ 0.35 (3.4)⁎⁎
2.98 (29.2)⁎⁎ −0.70 (3.7)⁎⁎ 0.67 (5.3)⁎⁎ −2.19 (6.4)⁎⁎ 1.80 (13.0)⁎⁎ −0.40 (2.5)⁎⁎ 1.73 (16.8)⁎⁎ 26,613.6
See footnote Table 1; the probabilities p1 to p8 implied by the estimates of α1 to α7 are (in %): 55.3, 1.4, 5.5, 0.3, 17.0, 1.9, 15.8, and 2.8. a)
rate. For tobacco about 75% has a positive starting rate with the remaining 25% being the group of abstainers. About 70% has a positive starting probability for cannabis. The last column of Table 2 shows that alcohol use has no effect on the starting rate for cannabis use but tobacco use has a significant positive effect. Table 3 shows the parameter estimates if unobserved heterogeneity is allowed to be correlated. As shown, it is possible to identify 8 mass points indicating that there are 8 types of individuals.3 By far the largest group consists of individuals (55.3%) that have a positive starting rate for alcohol, tobacco and cannabis. The second group (15.8%) only has a positive starting rate for alcohol. And, there is a group of 2.8% of the females that consists of abstainers from alcohol, tobacco and cannabis. Most of the parameters in Table 2 are very similar to the estimates shown in Table 3. However, the effect of alcohol is now significantly negative. The effect of tobacco is positive and significant but substantially smaller than before. Whereas alcohol dampens cannabis use tobacco use is a gateway for cannabis use. To give some indication of the size of the estimated effects the parameter estimates of Table 3 are used for some simulations shown in Table 4. The first column shows the simulation results for a female reference person who has secondary education, is born in 1950, is not confronted with parental cannabis use and never used alcohol or tobacco. As shown 21.1% of the females with these characteristics have started using cannabis at age 20, 32.2% at age 30, and 33.2% at age 40. The other columns of Table 4 show the effect of alcohol and tobacco on cannabis use. Females that start drinking alcohol at age 15 are less likely to start using cannabis, females that start smoking tobacco are more likely to start using cannabis. It is clear that the age
3
The likelihood ratio test statistics comparing the sum of the loglikelihoods in columns 1 to 3 of Table 2 with the loglikelihood in Table 3 is 495.8. This is highly significant since at a 1% level of significance the χ2 statistic for 4 degrees of freedom (4 additional probability parameters) equals 13.3.
1448
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
Table 4 Simulation results cumulative cannabis use by age (%)a) Age
20 30 40 a)
Reference person
Alcohol
Tobacco
Tobacco
Age 15
Age 15
Age 25
21.1 32.3 33.2
17.2 26.8 27.6
27.7 41.0 42.1
21.1 34.0 35.2
Reference person: Female, secondary education, born in 1950, no parental cannabis use, no use of alcohol or tobacco.
of onset of smoking tobacco is also important. Females that start smoking at age 25 at much less likely to start using cannabis than females that start smoking at age 15. 4. Discussion This paper uses information about the age of onset of alcohol, tobacco, and cannabis use of prime age individuals living in Amsterdam to study starting rates for these three drugs. It appears that adoption of alcohol and tobacco usually occurs up to an age of 20 while adoption of cannabis usually occurs up to an age of 25. If an individual did not use drugs up to that age then (s)he is very unlikely to do so at a later age. Starting rates are also affected by gender, education, birth year, and parental cannabis use. Furthermore, the starting rate for cannabis use is higher for smokers and lower for individuals who have started using alcohol. Kandel and Jessor (2004) conclude from an overview of many studies that there is no support for the proposition about causality in the effect of the use of one drug on the use of the next drug. According to them “there is no compelling evidence that the use of a drug earlier in the sequence, in and of itself, causes the use of a drug later in the sequence or, for that matter, that it causes the use of any other drug or, indeed, any other behavior.” Kandel and Jessor (2004) indicate that when an association remains after controlling a large variety of relevant variables, the residual association might represent its direct causal influence. However, they think such claims need to be tempered by the limitations of the existing evidence. According to them all the relevant variables cannot be controlled in any single study. It is true that association does not establish causation because this requires that all reasonable alternative inferences are rejected. Indeed this study shows that part of the association in the use of alcohol, tobacco and cannabis is through factors that influence the use of the first two drugs and the use of the third drug. However, the introduction of unobserved determinants which are allowed to be correlated across the different drugs accounts for any remaining spurious association. Therefore, what remains is the causal effect of alcohol use and tobacco use on the starting rate of cannabis use. The finding that alcohol use lowers the probability of cannabis use is difficult to explain, but it is consistent with the finding in some studies that alcohol and cannabis are substitutes in use. Indeed, if an individual has started using alcohol (s)he apparently is less eager to try cannabis. The finding that tobacco is a gateway to cannabis is more easy to explain since in the Netherlands cannabis is usually consumed in combination with tobacco. Persons that already smoke will find it easy to add some cannabis to their tobacco while persons that never smoked face a “double barrier”. If they want to consume cannabis they also have to start smoking tobacco. Nevertheless, it is difficult to draw general conclusions from the current study. The finding that tobacco is a gateway drug may be typical for
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
1449
countries in which cannabis is used in combination with tobacco. However, even for countries where this is not the case, like in the U.S., tobacco use and cannabis use are found to be complementary. So, it may be that also in the U.S. tobacco is a gateway for cannabis. It is difficult to indicate to what extent tobacco use is a general gateway to cannabis use as other circumstances of cannabis use are also different. In the Netherlands for example cannabis is much easier to obtain than in other countries. Furthermore, cannabis prices in the Netherlands are lower than in many other countries. For example, in 2004 the cannabis price was estimated to be 5.8 $ per g in the Netherlands while it was 15.0 $ per g in the U.S. (United Nations, 2006). The policy implications from the analysis presented in this paper are clear. Since tobacco and cannabis are complements policies that reduce smoking will reduce cannabis use. For those that worry about cannabis use it is reassuring to know that policy measures that reduce smoking will also reduce cannabis use. Furthermore, the analysis shows that policy measures should focus on young individuals; if not to prevent them from starting to smoke, then to increase the age of onset of smoking. It is clear that individuals that start smoking at a higher age are less likely to start using cannabis than individuals that start smoking very young. Acknowledgement The author thanks CEDRO, the Center for Drug Research, and the SCO-Kohnstamm Institute of the University of Amsterdam for making their data available.
References Abbring, J. H., & Van den Berg, G. J. (2003). The nonparametric identification of treatment effects in duration models. Econometrica, 71, 1491−1517. Abraham, M. D., Kaal, H. L., & Cohen, P. D. A. (2003). Licit and illicit drug use in Amsterdam: 1987–2001. Amsterdam: CEDRO, University of Amsterdam. Beenstock, M., & Rahav, G. (2002). Testing gateway theory: Do cigarette prices affect illicit drug use? Journal of Health Economics, 21, 679−698. Cameron, L., & Williams, J. (2000). Cannabis, alcohol and cigarettes: Substitutes or complements? Economic Record, 77, 19−34. Chaloupka, F. J., & Laixuthai, A. (1997). Do youths substitute alcohol for marijuana? Some econometric evidence. Eastern Economic Journal, 23, 253−276. DiNardo, J., & Lemieux, T. (2001). Alcohol, marijuana, and American youth: The unintended consequences of government regulation. Journal of Health Economics, 20, 991−1011. Farrelly, M. C., Bray, J. M., Zarkin, G. A., & Wendling, B. W. (2001). The joint demand for cigarettes and marijuana: Evidence from the National Household Surveys on Drug Abuse. Journal of Health Economics, 20, 51−68. Kandel, D. B. (2004). Examining the gateway hypothesis: Stages and pathways of drug involvement. In D. B. Kandel (Ed.), Stages and pathways of drug in-volvement; examining the gateway hypothesis (pp. 3−13). Cambridge: Cambridge University Press. Kandel, D. B., & Jessor, R. (2004). The gateway hypothesis revisited. In D. B. Kandel (Ed.), Stages and pathways of drug involvement; examining the gateway hypothesis (pp. 365−372). Cambridge: Cambridge University Press. Pacula, R. L. (1998a). Does increasing the beer tax reduce marijuana consumption? Journal of Health Economics, 17, 557−585. Pacula, R. L. (1998b). Adolescent alcohol and marijuana consumption: Is there really a gateway effect? NBER working paper series, Vol. 6348.
1450
J.C. van Ours / Addictive Behaviors 32 (2007) 1441–1450
Saffer, H., & Chaloupka, F. J. (1999). Demographic differentials in the demand for alcohol and illicit drugs. In F. J. Chaloupka, M. Grossman, W. K. Bickel, & H. Saffer (Eds.), The economics analysis of substance use and abuse (pp. 187−211). Chicago: University of Chicago. United Nations (2006). World drugs report 2006. Vienna: Vienna International Centre. Van Ours, J. C. (2003). Is cannabis a stepping-stone for cocaine? Journal of Health Economics, 22, 539−554. Williams, J., & Mahmoudi, P. (2004). Economic relationship between alcohol and cannabis revisited. Economic Record, 80, 36−38.