the crimes associated with drug tourism to map the geographical distribution. Their study used nationwide crime data in combination with interviews of local informants, municipal-, and police officials with knowledge on the soft drug market.
Their sample included 31 coffee shop municipalities dispersed over 11 judicial regions across the Netherlands. Their findings revealed strong geographical variation in the crimes associated with drug tourism and coffee shops. Municipalities close to the Southern border area ranked the levels of soft drug tourism, coffee shop nuisance, and illegal soft drug dealing as moderately to high levels. Five municipalities in the Southern border regions that restricted coffee shop access to non-residents resulted into drug runners and street dealers recruited tourists for illegal soft drug purchases. This was reflected in the high number of police reported soft drug incidents and prosecutions of soft drug crimes.
van Ooyen-Houben et al. (2014) on the relationship between coffee shops, drug tourism, nui-sance, and the illegal sales of soft drugs. They did an in depth survey in five coffee shop municipal-ities with three of them located within the Southern provinces. Their survey results revealed that residents in the three Southern coffee shop municipalities experienced nuisance from traffic, sound, loitering, and ”shady characters”. According to their survey, 40 - 50% of respondents answered that traffic and parking were significant problems.
Large portion has seen drug usage on the street but only a small percentage (7%) experienced it as nuisance. Respondents in three Southern coffee shop municipalities experienced the problem of drug runners, though only a small percentage of the respondents experienced them as a significant nuisance. Dealers were also frequently sighted on the street and a larger percentage of respondents experienced these as nuisance. In total, coffee shop and soft drug tourists were experienced as a nuisance by 4 and 8% by respondents.
Therefore, while the policy failed its intended purpose in the year of introduction, drug tourism within Southern coffee shop municipalities had been significantly reduced it came with a short-term cost of increased crime and an expanded drug market. However, it should be noted that the majority of studies on this topic are qualitative in nature. The adoption of an empirical approach approach in determining the relationship between the act of restricting soft drug sales within Southern coffee shop municipalities and whether the subsequent reduction in drug tourism had a significant impact on the number of drug crimes committed within Southern coffee shop municipalities would provide further evidence on the effectiveness of this policy.
The implication of this study could be significant in a political landscape where a larger num-ber of countries are adopting the perspective of harm reduction for drug users and the general public, understanding whether specific policies can adequately address new issues that arise due to changing to drug prohibition or legalization. Thus, the possible findings of this chapter could have significant theoretical and practical importance for further policy makers about to dealing with the nuisance of drug tourism.
to as the treatment group and a control group which acts as the counter-factual and does not undergo or experience the proposed intervention. In addition, there is available yearly data on the necessary outcome variables to satisfy the requirement to have several years before and after the implementation of the policy. Based on these reasons, I adopt the DID approach and use the following DID specification:
yit=α+δDi xdt+µi+λt+ωi+it, (3.1) whereyitis the outcome variable of interest of coffee shop municipalityiin yeart,αis the constant, D is the treatment variable (1 for being affected by the policy in 2012 and 0 if not affected by the policy), d is the time period (1 for post-treatment after 2012 and 0 for pre-treatment period which is before 2012), µi is the municipality fixed effect,λtis the time fixed effect,ωi is the linear time trend that de-trends for municipality specific drug trends, and is the error term. The parameter of interest is δwhich captures the average impact of the policy on the municipal drug crime rate. The municipality fixed effect term captures time invariant differences in the coffee shop municipalities. The time effect controls for time-varying characteristics that affect all coffee shop municipalities and the municipality specific linear time trend de-trends for particular drug crime trends. Therefore, the DID model captures potential issues of observed and unobserved heterogeneity.
The validity of the DID model depends on whether it satisfies the parallel time trend where the change in the outcome variable of the control and treatment group in the pre-treatment period is not statistically significant different. The parallel trend assumption can be checked by including yearly difference-in-difference estimator. I follow Autor (2003) and estimate the parallel trend assumption through the following DID specification:
yit=
2
X
k=1
βkDi xdt+k+
2
X
k=0
β−kDi xdt,−k+µi+λt+ωi+it, (3.2) Equation 3.2 consists out of two treatment variables whereDi xdt+k represents the lead dummies which captures the outcome of interest in the pre-treatment period and equals 1 before the policy is implemented, 0 otherwise. Dixdt−k represents the lag dummies which captures the outcome of interest in the post-treatment period and equals 1 for coffee shop municipalities after the policy is implemented, 0 otherwise. The policy was implemented for the Southern provinces in 2012 and the database contains yearly information from 2009 to 2014 on the municipal drug crime rate.
Equation 3.2 therefore estimated using two interaction terms that captures the parallel trend of the outcome variable of interest in the pre-treatment period (2010, 2011), one interaction term that captures the impact of the policy during the implementation year (2012), and two interaction terms that capture the post treatment effects in the post-treatment period (2013, 2014).
The DID estimator of Equation 3.2 would be considered biased if the coefficients ofDi xdt−k are reported as being statistically significant which would indicate that there a difference in the trend of the municipal drug crime rate before the policy is implemented. As a result, the sig-nificant differences in the pre-treatment period would result in the treatment effect being under-or overstated in the post-treatment period. However, besides evaluating whether the DID esti-mator satisfies the key assumption of parallel assumption, Equation 3.2 is also used to determine whether the impact of the policy had a significant effect on the municipal drug crime rate and the subsequent fade-out of the treatment effect after 2012.
The DID model specified in Equation 3.1 controls for the unobservable time invariant and time-varying characteristics through municipality-, time fixed effects, and a linear time trend. However, the DID model does not control for initial observed differences in the coffee shop municipality characteristics that may cause a biased DID estimator. One method to overcome the issue of a biased DID estimator is to create a counterfactual group of control coffee shop municipalities that more closely resembles the group of treated coffee shop municipalities in terms of their observed municipal characteristics. However, as pointed out by Caliendo and Kopeining (2008), the higher the number of characteristics used to match treated and control observations would lead to a significant reduction in suitable matches that have similar values ofX.
Rosenbaum and Rubin (1983) propose a statistical matching technique referred to as propensity score matching (henceforth abbreviated as PSM) to overcome the issue associated with a high dimensional vector. The PSM technique uses a single index named the propensity score which
captures the probability of an observation to be selected for treatment based on their observed characteristics. Let theP(Xi be the propensity score which is defined asP(Xi) =P(Di|Xi).
The validity of the PSM technique relies on two assumptions. The first assumption is the conditional independence assumption which is expressed asyi0, yi1`
Di|Xi where`
indicates that the treatment assignment is independent of the outcome variable of interest and solely based on the set of observed characteristics. Fulfilling the conditional independence assumption eliminates the bias attributed due to differences in the observable characteristics between treated- and control coffee shop municipalities. The second assumptions is the common overlap which states that coffee shop municipalities with a similarX values have an equal probability of being both treated-and control coffee shop municipalities which is necessary to form valid counterfactuals for treated observations.
The propensity score is estimated with either a probit- or logit model using a set of observed characteristics before the start of the treatment and then matches treated- and control coffee shop municipalities on the proximity of their propensity score. There are several alternative matching procedures available with their own positives and negatives. Existing PSM studies often include a maximum distance between the propensity score of treated- and control observations.
One issue of the PSM technique is that the specification is dependent on the set of observed characteristics. According to Rosenbaum (2002), Gangl (2004) and Caliendo and Kopeining (2008), it is important to check whether the probability of the treatment assignment is vulnerable and can be altered by unobservable factors.6 Letπi be the treatment assignment,Xibe the set of observed characteristics,γis the effect of unobserved characteristics on the treatment assignment probability, and let ui be the unobserved component where πi = P r(Di = 1|Xi = F(βX+i+γui) which states that the probability of treatment assignment is determined by observed and unobserved characteristics.
Parameterγcaptures the effect of the unobserved characteristics on the probability that coffee shop municipalities are selected as either treated- or control observations. If there is no unob-servable bias influencing the probability, then γ would be zero and πi = P r(Di = 1|Xi is solely determined by the observed characteristics. Existing literature such as Duvendack and Jones (2011) argue that in the case of social sciences, aγ value of 2.0 would indicate strong insensitivity to unobserved characteristics influencing the probability of treatment assignment.
I used the municipal drug crime rate as the outcome variable of interest with the following variables as regressors: the number of residents within a coffee shop municipality that have em-ployment, the number of residents within a coffee shop municipality that receive poverty assistance from government institutions, the number of residents within a coffee shop municipality that are following a university degree, the median household income of households within a coffee shop municipality, the total value of real estate located within a coffee shop municipality, the number of restaurants located within the domain of coffee shop municipality, the number of establishments categorized as coffee shops within a coffee shop municipality due to possessing a permit to sell soft drugs, and the change in the number of arrested suspects within a coffee shop municipalities.
Several different matching algorithms were used to check the consistency of the matching re-sults such as pair matching with and without replacement, kernel matching, and nearest-neighbor matching up to 2, 4, and 6 neighbors.7 To improve on matching accuracy, I set a maximum dis-tance in terms of the propensity score between treated- and control coffee shop municipalities. I followed the recommendation of Austin (2011) which stated to use a caliper size of 25% of the pooled standard deviation of a logit model.
Matching quality was assessed by comparing the bias percentage before and after the matching procedure, the percentage reduction in the bias, and whether the balance test reports a signifi-cant difference between the treated- and control coffee shop municipalities for the set of observed characteristics. The results presented in the empirical analysis are based on the best matching results which was the nearest-neighbor matching algorithm with 4 neighbors as it provided the best balance between treated- and control coastal municipalities with the set of control variables.
The combination of propensity score matching with difference-in-differences (henceforth referred as PSM-DID) controls for observable differences between treated- and control coffee shop munici-palities by creating a counterfactual group that more closely resembles the treatment group. The
6Sensitivity of the propensity score specification was estimated through the Stata user module -rbounds- by Gangl (2004).
7Matching procedure and assessment of balancing was done through the Stata user module -psmatch2- by Leuven and Sianeti (2003).
PSM-DID estimator also controls for unobserved time-invariant and time-varying characteristics through municipal- and time fixed effects.