HEDG Working Paper 05/09
Sequential patterns of drug use initiation – can
we believe in the gateway theory?
Anne Line Bretteville-Jensen
Hans Olav Melberg
Andrew M Jones
October 2005
ISSN 1751-1976
york.ac.uk/res/herc/hedgwp
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Guidelines and tips
Prepare for your exams
Study with the several resources on Docsity
Earn points to download
Earn points by helping other students or get them with a premium plan
Secondary schools (high schools) in Canada do not need to go through the NCAA Eligibility. Center review process to establish a list of core courses or obtain a ...
Typology: Study notes
2021/2022
1 / 28
This page cannot be seen from the preview
Don't miss anything!
Related documents
Partial preview of the text
Download CANADIAN HIGH SCHOOL CODES and more Study notes Life Sciences in PDF only on Docsity!
HEDG Working Paper 05/
Sequential patterns of drug use initiation – can
we believe in the gateway theory?
Anne Line Bretteville-Jensen
Hans Olav Melberg
Andrew M Jones
October 2005
ISSN 1751-
york.ac.uk/res/herc/hedgwp
Sequential patterns of drug use initiation –
can we believe in the gateway theory?
Anne Line Bretteville-Jensen*
Hans Olav Melberg**
Andrew M. Jones***
Abstract The gateway, or stepping stone, hypothesis is important as it has had considerable influence on drug policy and legislation in many countries. The gateway hypothesis offers one possible explanation for young people's development of a serious drug problem. It simply states that the use of one drug increases the risk of starting to consume another, and possibly more harmful, drug later on and that the risk increases with frequency of use (dose-response). The empirical basis for the hypothesis is the common finding that most heavy drug users have started with less dangerous drugs first and that there seems to be a "staircase" from alcohol and insolvents via cannabis and tablets to amphetamine, cocaine and heroin. The core question is whether the sequential initiation pattern of drug use is best explained by the mechanisms substantiating the gateway hypothesis or whether the phenomenon is better understood by employing the concepts of accessibility and/or transition proneness? Based on a representative sample of 21-31 year olds in Oslo we have examined the possible gateway effect of both legal (alcohol) and illegal drugs (cannabis) on subsequent use of cannabis and hard drugs (amphetamine and cocaine). We use multivariate probit models that take account of unobservable individual-specific effects to reduce the possibility of a spurious causal effect of soft drug use on the onset of hard drug use. The gateway effects were greater when we did not take account of unobserved heterogeneity, but, although substantially reduced, they remained considerable also when unobserved factors were accounted for.
JEL: I12, I
Keywords: Gateway hypothesis, Stepping stone hypothesis, Substance abuse, Multivariate probit analysis, Unobserved heterogeneity, Amphetamine, Cocaine, Cannabis
*Norwegian Institute for Alcohol and Drug Research (SIRUS) and Health Economics Bergen (HEB) ** Norwegian Institute for Alcohol and Drug Research (SIRUS) ***Department of Economics and Related Studies, University of York and Health Economics Bergen (HEB)
The present paper aims at examining the gateway effect by analysing data collected among the general population of 21-31 year olds in Oslo. We employ models that take account of unobservable individual-specific effects to reduce the possibility of a spurious causal effect of soft drug use on the onset of hard drug use. Many youngsters seem to experiment with illicit drugs. The majority of problems related to drug use, however, are caused by regular users. Therefore, and in contrast to most studies in this field, we examine the gateway hypothesis studying “users” and separate people according to their frequency of use, not according to whether or not they report to have ever tried various drugs. The most policy-relevant question is not whether a soft drug makes it more likely that a person will just try a hard drug once at some later point in time, but whether having used a soft drug makes it more likely that the individual will progress to have a problematic use of a hard drugs later. Before presenting the methods, data and results, however, we look more into the gateway theory and give an overview of the relevant empirical literature.
2 The gateway theory
Kandel (1975) was one of the first to study the sequential pattern of drug use initiation based on longitudinal data. She found four stages in drug use with marijuana being the crucial step on the path to other illicit drugs. As is emphasised in Kandel et al. (1992), however, the authors state that entry into a particular stage is common and perhaps even necessary although not a sufficient prerequisite for entry into the next stage, i.e. they argue against a version of the stepping stone theory which claims that marijuana leads inexorably to the use of other illicit drugs. Goode (1972), on the other hand, is an early example of the assertion that there could be a casual link between different stages in drug use.
MacCoun and Reuter (2001) have a thorough discussion of the concept and Pudney (2003) lists three possible mechanisms that, on their own or in combination, might be the basis for a causal gateway effect in drug use:
- The consumption of soft drugs may create a psychological or physiological need for further and stronger experiences of the same type.
- The act of obtaining and using soft drugs may bring the user into contact with hard- drug users or suppliers whom they would not otherwise have met.
- Experience of the use of soft drugs with no obvious ill effects may appear to contradict and undermine the strong negative publicity directed against illicit drug use in general, so that advice against hard drugs becomes less persuasive.
In addition to these mechanisms, one could argue that for some individuals consuming a drug for the first time is like crossing a threshold and that the action makes it less costly to proceed into another "drug stage". Taking only one step at a time, when each step reduces the cost of the next, could increase the probability of ending up as a heavy drug user. Some people may not have started to consume, for instance heroin, if they were offered the drug without first having tried other illegal substances. Hence, despite increased legal sanctions and increased dangers associated with consumption of the various drugs along the path, some people proceed, and the claim is that there is a causal effect through reduced costs caused by the initial consumption of drugs at each stage.
Pacula (1997) approaches the possible causal relationship from a different angle. In contrast to reduced costs, she suggests that past consumption of any one drug will increase the marginal utility of consuming this drug and any other drug. Her model is a variation of the rational addiction framework developed by Becker and Murphy (1988) and it builds in reinforcement and tolerance effects of addictive goods. Pacula differs from Becker and Murphy in that she assumes that the "consumption capital" represents past consumption of many substances, and by that she opens the way for a possible gateway effect of drugs. She claims that young people start consuming the drug with the lowest marginal cost and then become more likely to initiate use of more costly substances as the marginal utility of using them rises. In Kenkel et al. (2001) the authors describe the data that would be required for a proper test of the rational addiction version of the gateway hypothesis.
If the use of soft drugs has a causal gateway effect on the transition to harder drugs, then restrictions on the use of soft drugs may be an effective policy tool to achieve the objective of reducing the use of hard drugs. However there are alternative explanations for the observed data that do not have the same policy implications. An alternative to the gateway hypothesis for explaining the observed sequential pattern of drug initiation is differences in accessibility for various age groups. With accessibility we mean physical availability, cultural acceptance, prevailing drug legislation and affordability (influenced by both individual income and drug prices). People may start to consume alcohol prior to cannabis and cannabis
The empirical literature shows contrasting results regarding a possible gateway effect. Some researchers report a strong and significant influence of previous drug consumption on current consumption of the drug, while others present results that do not support the gateway hypothesis. There are at least two possible explanations for the divergence in results. First of all, differences in data could obviously lead to different conclusions. When, for instance, van Ours (2003) concludes that data from the Netherlands in some cases support the gateway hypothesis this need not contradict Pudney’s (2003) finding of only a very small gateway effect in a sample of British youths. To some extent the differences may be caused by different rates of response, various sample selection criteria (e.g. age groups), the timing of the survey and a host of other rather mundane but still important factors that lead the researchers to different conclusions. Secondly, different results may be caused by different approaches and methods employed when analyzing the data. For example, different distributional assumptions give ample scope for two researchers with identical data to reach different conclusions. Although important, it is often difficult to test conjectures about the more practical data problems. The problems of how to best approach the data, however, can be explored more theoretically.
Testing the gateway hypothesis illustrates the classical problem of separating heterogeneity and causal effects. Simply documenting that most heavy drug users started with legal drugs and cannabis is not sufficient to establish a causal link. The problem can be illustrated as follows: assume that the probability of starting with e.g. amphetamine is estimated by ordinary regression analysis on the following equation:
(1) hit = α t (^) + β X (^) it + δ dit − 1 + ε it
in which hit is the risk of starting with amphetamine for a person (i) at a point in time (t); Xit is a vector of exogenous variables influencing the probability other than previous use of other drugs (e.g. gender, childhood experiences, peer influence and so on), dit-1 is a dummy representing previous use of other drugs and εit is the error term. In this context hit is the “outcome” of interest and dit-1 is the “treatment” to be evaluated. If the dummy for previous drug use turns out to be statistically significant, it's tempting to conclude that the gateway hypothesis is supported by the data. The problem, however, is that standard regression analysis on the equation will produce misleading results if potentially important variables that could explain amphetamine use are omitted. Moreover, some of these omitted variables may influence not only the probability of amphetamine initiation, but also initiation of other drugs.
In this case we will get biased estimates because the dummy variable for previous drug use will capture not only the “true” gateway effect, but also the effect of the omitted variables.
One obvious way to reduce the problem would be to include more variables in the model, on the assumption that selection into the different treatment regimes – in this case past drug use – is ignorable, after conditioning on all of these observable covariates. Yamaguch and Kandel (1984) and Fergusson and Horwood (2000) are two examples of studies that have included a wide range of variables assumed to influence drug use and deviant behaviour. A problem with this approach, however, is that one will never be sure that every relevant variable actually is included. This need not be due to ignorance on behalf of the researcher, but may be caused by lack of data or inherent problems in measuring some potentially important variables. Although the probability is reduced with more variables taken account of, the possibility for a spurious gateway effect still remains and the estimate of the casual effect may be biased by selection on unobservables.
Two possible approaches can be adopted to overcome the problem of selection on unobservables. Firstly, one can employ an instrumental variable (IV) technique that predicts the dit-1 on the basis of another variable, or a vector of variables, that are highly correlated to previous drug use but not to the error term in (1). The approach has been adopted by Pacula (1998) who uses past prices of alcohol as instruments for previous consumption of the drug and estimates a gateway effect of alcohol on current marijuana use. She uses data from the National Survey of Youth (NLSY) and reports that higher past alcohol prices are associated with lower likelihood of using marijuana. The same data set, but covering different years, is employed by DeSimone (1998) who uses information on individual characteristics and local prices as instruments. Also Beenstock and Rahav (2002) use variants of the IV approach to sequences of events when they employ prices by birth cohorts as instruments. The main problem with the IV-approach is finding good instruments. Alcohol and cigarette prices have been frequently used. They vary over time and between countries and states, but they cannot reflect contemporaneous individual differences in behaviour within the same area. Prices of illicit drugs are, in addition, hard to obtain. Credible instruments for previous consumption that are not based on prices are rare.
As a second alternative, one can employ models that take account of an unobserved factor that is possibly influencing both the dependent variable and dit-1. Some analysts have taken as their starting point that no study, despite survey method or level of details, will
3.1 Methods
We proceed in three steps: The first step is to examine whether the individuals in our sample start to use alcohol and illicit substances according to the gateway hypothesis, i.e. examine whether they start with alcohol before cannabis, cannabis before amphetamine and so on. The section employs tools from survival analysis and we estimate separately for each drug the probability of starting to use the drug at different ages given that they haven't used the drug previously (i.e using the Kaplan-Meier method to estimate hazard functions).
Second, we will estimate three separate single equation probit models to determine the statistical relationship between problematic (frequent) use of each substance (cannabis, amphetamine and cocaine) and the following independent variables: gender, social problems (problems with, parents, school, friends and police), attitudes towards free cannabis sale and previous use of other drugs_._ Previous drug use is included as a dummy. The three models serve as benchmarks against which we can judge the results that allow for unobservable heterogeneity. In line with the gateway hypothesis we have adopted the view that people first start with alcohol, then some proceed to cannabis and later on start with amphetamine and cocaine. This means that we test for previous alcohol use when we estimate the probability of starting with cannabis, only test for previous alcohol and cannabis use when estimating the probability for amphetamine use and test for alcohol, cannabis and amphetamine when cocaine is used as a dependent variable. As use of ecstasy had low prevalence when people in the oldest age groups were in the typical age for experimenting with this drug we have not included ecstasy in the analyses.
In the third step we estimate the three equations together using a multivariate probit specification. This model has been characterized as an "unfairly neglected procedure” (Lesaffre and Molenberghs 1991) in the context of medical statistics and a search of the economics literature indicates that it is no less true of economics. As mentioned previously, we may suspect that the single equations omit relevant variables which we may interpret as “unobserved heterogeneity”. The effect of this heterogeneity is captured by the error term in the single equations. The idea behind the multivariate probit model is to model the correlation between the error terms from the single equation models. If there is a systematic relationship between these, one may conclude that an important variable that affects all of the equations has been left out. One may then exploit this systematic relationship between the error terms in
the different equations to allow for the unobserved heterogeneity. Thus, by estimating all three equations at the same time, taking account of the cross-correlation in the error terms, one reduces the problems of unobserved heterogeneity which is a major problem when testing the gateway hypothesis (see Greene (2002) for more on estimation of the multivariate probit model and Contoyannis and Jones (2004) for a recent application that uses the multivariate probit model to estimate a recursive system similar to the one used here). The final element is to compare the results from the single probit models with the multivariate in addition to testing for potential problems. By comparing the results from the two steps we may examine the extent to which correcting for unobserved heterogeneity affects the sign and statistical significance of the estimated gateway effects.
3.2 Data
The data were collected through postal questionnaires sent to a representative sample of 21- year olds living in Oslo in 2002. The response rate was roughly 50 per cent with more women than men answering the questions (see Table 1). Only one reminder was sent and a total of 4561 questionnaires were registered. The respondents reported their experience with licit and illicit drugs in addition to socio-economic information on age, gender, education, income and possible childhood problems with parents, friends, school and police, and their attitude toward free sale of cannabis.
(Table 1 about here) As mentioned in the introduction, we focus on "users" in this study and employ a dummy based on frequency of drug use as the dependent variable for each drug in question. The frequency variables are set equal to one if the respondents report to ever have used the drug more than 25 times. Out of the 40 per cent reporting to have tried cannabis, about one third (13%) have used the drug on more than 25 occasions. The corresponding numbers for amphetamine are 11 per cent having ever tried the drug and 3 per cent are regular users. Ten per cent in the sample report to have ever tried cocaine while 2 per cent have used it on a regular basis. The percentage having ever used alcohol is high (93%) and only 1.5 per cent have ever tried heroin.
Based on a certain set of birth dates for the years 1972-1981 the sample was drawn from the national register. Larger birth cohorts in the first part of the 1970s and a higher
positive only when the individual has used a drug before the other drug. Hence, in the equation for amphetamine, the dummy for cannabis is zero for those who have never used cannabis and for those who have used amphetamine before cannabis since in that case cannabis could not be a gateway for amphetamine. Similarly, the dummies in all the equations were constructed to capture only whether the gateway drug had been used before the drug under consideration.
The data's representativeness is hard to assess. It is well known that in general surveys like the one used here, homeless and institutionalised people are under-represented as are people with many sorts of deviant behaviour. This selectivity is especially worrying when the topic of interest is illicit drug use and one may, perhaps, assume that the sample is representative only for the "normal", well-functioning fraction of the population. Reported income and educational achievements suggest that the sample is better off than the average of young people in Oslo. Still, the relatively high prevalence of illicit drug use in the present sample indicates that a large proportion of drug users respond to postal questionnaires.
Recall bias may be another problem, especially here where people are asked to recall the debut age of incidents that occurred, in some cases, more than a decade before. One may argue, however, that using an illicit drug for the first time is so unique that users will tend to remember it. In line with this, one recent study of response reliability in adolescent substance use progression suggests that the reported sequences were reported consistently when checked again three years after the first interview (Golub et al. 2000).
3.3 Results
To get a first impression of whether there is a gateway effect, it is useful to explore the order in which people have used various substances (see Table 2). The table confirms the general impression that “soft” drugs are used before “hard” drugs. For instance 10.9% of the total sample claims to have used cannabis before the other illicit drugs (amphetamine, cocaine, ecstasy and heroin), while only 1.5% claimed to have used one of these drugs without using cannabis first. Among the 503 amphetamine users in the sample only 77 per cent report to have first used cannabis and 14 per cent started to use both substances within the same year. The corresponding numbers for the 459 cocaine users were 89 and 7 per cent.
(Table 2 about here) In addition to giving an impression of the sequences of drug use that are most common, the table also helps to suggest which sequences it is worth testing for in the regression analysis. For instance, very few individuals used cocaine before cannabis and for this reason we have not included a dummy to test whether cocaine could be a stepping stone to cannabis use. Instead we have focused on the major pathways and the table indicates that the most common “stepping stone” is cannabis to some of the other drugs (amphetamine and cocaine).
The "staircase" in drug use initiation is illustrated in Figure 1 where the highest hazard rate for starting with alcohol peaks at an earlier age than the highest hazard rate for cannabis and use of amphetamine and cocaine. The hazard rates give the probabilities for various age groups of starting with a drug given that the person has not started up to that age.
(Figure 1 about here)
We also checked the hazard rate for heroin (n=67), and found that it deviates from that of the other substances by having a less uniform pattern with one peak corresponding to the age of 20 and one at the age of 22, but the small sample size is problematic
Univariate probit models
In Table 3 we present the results of separate estimates of univariate probit models for cannabis, amphetamine and cocaine. The dependent variables in these regressions are not whether the individual has used a substance, but whether the individual has used a substance frequently or not. As argued in the introduction, this is the most policy relevant variable since the justification for making softer drugs illegal is based on the dangers of developing a problematic use of another substance. By including dummies for previous use of drugs further down the staircase provide for preliminary evidence of possible gateway effects. Due to the nonlinearity of the probit function we have also calculated and displayed the partial effects for each of the estimations. These are based on the sample mean values of the regressors and indicate the absolute change in probabilities that occurs when the variable of interest changes by one unit (continuous variables) or when a dummy variable changes from zero to one in value. Unless otherwise stated, the parameters are statistically significant at a 5 per cent level.
The results for the cocaine equation reveal that both previous cannabis and amphetamine use are positively associated with for regular use of cocaine. The alcohol dummy, however, is negative and insignificant. There is no significant difference between the genders and, of the variables indicating childhood problems, only problems with the police obtain a significant estimate. The latter variable has a smaller influence on regular cocaine use than the variable had on regular cannabis and amphetamine use. The data does not indicate any statistically significant effect of the prevalence variable nor of the respondents’ attitudes towards free sale of cannabis.
Multivariate probit models
The interesting question now, however, is whether the substantial gateway effects found in the separate estimation of the three equations remain after we have taken account of unobserved heterogeneity. The results from the multivariate probit model are reported in Table 4.
(Table 4 about here)
The most striking result is the reduced values of the gateway effects in the equations for amphetamine and cocaine: For amphetamine the coefficient on the cannabis dummy has changed from 1.45 to 0.70 and for cocaine from 1.10 to a statistically insignificant value of 0.63. Further, the coefficient on the amphetamine dummy in the cocaine equation decreases from 1.56 to 0.51. The alcohol dummy remains insignificant for both substances. The sizeable correlation coefficients for the three equations are presented at the bottom of Table 4 and indicate the importance of estimating the equations as a system. All of the correlations are positive, consistent with the idea of a common unobservable propensity to substance abuse. In Table 4 the value of the freesale parameter for amphetamine is higher compared to the corresponding value in Table 3 whereas the estimates for the other variables in the amphetamine equation remain fairly unchanged. When comparing the univariate and the multivariate probit results for cocaine, we see that more coefficients are significant at a 10 per cent level in the multivariate probit model (parents and school). In addition, the estimates for childhood problems with parents, school and police and the freesale variable have increased values in Table 4. In contrast, there is hardly any difference between the parameter values in Table 3 and Table 4 for the cannabis equation, which means that taking account of
unobserved heterogeneity has not had any influence on the estimated gateway effect or on the other explanatory variables for this drug.
Table 5 presents the partial effects based on the coefficients from the multivariate probit model. They are computed at the sample means of the regressors for each of the three substances separately. For the dummy variables the partial effects show the difference in predicted probability of becoming a regular drug user when the dummy is 1 or 0. As the overall probabilities of frequent amphetamine and cocaine use are relatively small, we have also included the percentage changes in these predicted probabilities. The intention is to underline the quantitative importance of the various dummy variables. Statistically significant values are in bold, and we can see that the dummies for childhood problems with school and police, as well as the respondents' attitude towards free cannabis sale, are associated with a substantial change in the probabilities for frequent drug use across the three substances. For instance, given identical values of the other variables the probability of becoming a frequent amphetamine or cocaine user increases by 87 per cent in both cases if the individual has had problems with the police during childhood. The absolute partial effects are 0.08 and 0.02, respectively. Also the percentage increase in the probability of previous cannabis use on later amphetamine use and previous amphetamine use on later cocaine use are substantial. Hence, these variables are potentially of great importance even though the overall probability of becoming a frequent user of these drugs is relatively small.
In order to test sensitivity of the multivariate probit results reported in Tables 4 and 5, we re-ran the model with the cut-off point for frequency of use set to 11 occasions or more (previously it was set at 25 occasions). This means that more people were included as "frequent users". For cannabis, 225 people (12.4 percent of those reporting cannabis use) changed status, whereas the corresponding numbers for amphetamine and cocaine were 72 (14.3 per cent of all amphetamine users) and 69 persons (15.0 per cent of all cocaine users). The multivariate probit results with these new dependent variables were fairly similar to those presented above. The same set of coefficients were statistically significant and the signs were retained. The difference between the parameter values were not systematically positive or negative and not substantial for the statistically significant estimates. One may therefore infer from this that because many people report that they stop taking drugs after they have tested it once or twice, the cut off point of 25 occasions or more of using a substance seems not to influence the results substantially.
people reported to have had problems with friends, on the other hand, did not influence their probability of becoming frequent drug users.
"Freesale", measuring whether the respondents favour free sale of cannabis, was another type of variable included. As mentioned, we expected it to be highly correlated to own drug use as it may be interpreted as an indicator of attitudes towards drugs, although not every person claiming to be in favour of cannabis legalisation reported to have used illicit drugs themselves. The estimates show that the variable has a significant influence on frequent use of the three drugs with more influence on frequent cannabis use than on the other two. Although we cannot exclude the possibility that drug use itself subsequently changed individuals’ attitudes, the result may suggest that differences in personality and attitudes play an important role in regular drug use.
The last set of variables we included in the models was measures of drug use prevalences taken from annual survey data of 15-20 year olds in Oslo. We assumed that the percentage of youngsters reporting to have tried e.g. cannabis in a given year could be taken as an indicator of the general physical and cultural accessibility of the drug. Consequently, the accessibility was assumed to be generally higher for those aged 19 years in 2000 when the cannabis prevalence was 28.6 per cent than for those who were 19 in 1991 when the prevalence was 16.6 per cent. Only the cocaine prevalence employed in the cocaine equation did not seem to have a significant influence on regular use of the drug. This could be due to cocaine use being less common among 15-20 year olds and thereby the prevalence for this particular drug being a poorer indicator for cocaine use in general.
Including the above mentioned variables has proved important, but as the survey does not provide us with all the variables that could possibly influence frequent drug use, we have employed methods that take account of unobservable individual-specific effects to reduce the possibility of estimating a spurious gateway effect. The multivariate probit model is presented in section 3. The results shown in Tables 4 and 5 suggest that there are some gateway effects of previous drug use on subsequent frequent use of cannabis, amphetamine and cocaine, respectively. The effect of previous drug use was greater in estimates where we did not take account of unobserved heterogeneity (Table 3), but, although substantially reduced, they remain substantial when unobserved factors are accounted for. The tendency of reduced influence of the gateway variables after taking account of unobserved heterogeneity is in line with the findings of van Ours (2003) and Pudney (2003). They employ alternative methods
for taking unobservables into account. According to the current findings, alcohol is a gateway drug for cannabis, cannabis is a gateway drug for amphetamine and amphetamine is a gateway drug for cocaine.
Given that there is an effect of previous drug use on subsequent regular use of a drug further down the pathway to hard drug use, what are the policy implications of the finding? Should we ban alcohol to reduce later cannabis problems and will the recent policy change towards cannabis in the UK leads to an increased number of amphetamine users in the next few years? As discussed in section 2, there are at least five possible mechanisms that may explain an observed gateway effect and what drug policy to recommend will, among other things, depend on which of the mechanisms that actually operates. If the use of one drug creates a psychological or physiological need for further and stronger experiences of the same type, if the use on one drug reduces the costs of starting with another and more dangerous drug or if the use of one drug increases the utility of consuming another, then a strict drug policy may be the preferred option. An extensive cost-benefit analysis including various aspects of such a restrictive policy, however, is needed to determine the question.
On the other hand, if the act of obtaining a soft drug brings the user into contact with hard-drug users or suppliers whom he/she would not otherwise have met, then the Dutch option may seem more attractive. Separating the markets for soft and hard drugs by legalising consumption and sale of cannabis may then prove successful in reducing the rate of transition to hard drug use. Further, if people, after experiencing no obvious ill effects of soft drug use, have reduced confidence also in the strong negative publicity directed against hard drug use, the solution may be to make more distinct the differences between the various drugs, perhaps including more distinct differences in the legal sanctions directed against drug use and trade.
Acknowledgements
The authors would like to thank participants at the 24th Arne Ryde Symposium: The economics of substance use, Lund 13-14 August 2004 for useful comments to an earlier draft of the paper.