Explaining the Consumption of Illicit Cigarettes

Abstract

Objectives

What drives the consumption of illicit cigarettes? While criminology has rarely addressed the divers of the illicit cigarette markets from a theoretical and empirical perspective, studies from other disciplines point to two classes of causes. Some studies stress the impact of cigarette prices and taxes on the market for illicit cigarette; others emphasize the importance of different non-price factors, including informal economy and corruption. This study tests the influence of both price and non-price factors on the illicit cigarette market.

Methods

Multilevel growth curve analysis—three-level MLM for longitudinal measures—of the illicit cigarette market at the subnational level in the European Union. The analysis focuses on 247 regions in the EU between 2007 and 2013.

Results

This study shows that both price and non-price factors influence illicit cigarette consumption. Lower affordability of legal products, proximity to sources of cheap cigarettes, higher national income inequality, greater population density, and the levels of illicit cigarettes in neighboring regions are associated with higher illicit consumption. On the contrary, there is no empirical evidence of the role of two ‘usual suspects’: corruption and shadow economy. The paper also shows that the market for illegal cigarettes is shaped by both demand and supply factors.

Conclusions

The geographic concentration of illicit consumption and smuggling calls for the creation of anti-illicit-trade units in most densely populated areas or custom task forces at the most sensitive borders. The disproportionate relevance of illicit flows from eastern non-EU countries suggests to increase the political pressure on these source countries. Finally, given the importance of the demand side in determining the size of the illicit market, price increases should be matched with consumer awareness campaigns. These campaigns should focus on the societal consequences of purchasing illicit cigarettes together with illustrating the harm of consuming tobacco products.

Appendices

Appendix 1: Missing Data Handling

In consideration of the possible relevance of missing observation for the proposed analysis and the exploited econometric techniques (Fig. 5), we imputed missing data by relying on two different techniques. Then, we exploited both the estimates to empirically verify our research hypothesis. The rest of Annex A presents both the single imputation and the multiple imputation techniques we performed.

Fig. 5

Missing map of the entire database, 2007–2013. Note: the figure provides a visualization of the amount of missing data as a matrix where columns are the variables used in the analysis and row are each region-and-year values. Red cells correspond to missing values, blue cells correspond to available data. The rounded percentage of missing values overall is in the legend. Whenever Eurostat (2017b) does not provide estimates of the share of population at risk of poverty or social (242 missing observations, 14.1% of the total) exclusion or of the population-passenger rate (448 missing observations, 26.1% of the total) at the NUTS2 level, we used the available NUTS1 or NUTS0 data

The single imputation approach follows Transcrime (2015), and combines Mean imputation and Using information from related observations as by Gelman and Hill (2007). In particular, we adopted a naïve version of the K Nearest Neighbors-based method (Troyanskaya et al. 2001) in which neighbors’ distance is evaluated only in terms of physical and temporal (k = 1) terms. In particular, whenever data are missing only for some years—i.e., 29 regions—, the missing estimate was imputed by taking the average of the neighboring regions and adjusting it by the ratio between the data in the region of interest and in its neighboring regions in the years in which both the data were available. In the 10 regions for which data are missing for the entire period, annual estimates are based on the average of the corresponding values of the neighboring regions. The resulting completed data set was then used for inference.

The single imputation approach relies on two main arguments: (1) illicit consumption of cigarettes tends to cluster geographically (Calderoni et al. 2017); (2) region with missing values are far from hot/cold-spots. As an example, considering the last available year in the dataset, the global Moran’s Index for the variable ICR, excluding the missing regions, is equal to .625 (p value = .000) indicating a high and positive tendency to clustering of similar values (Bailey and Gatrell 1995). Nevertheless, the 15 regions with missing data in 2013 are mainly located away from the main geographical clusters of high or low values of the ICR identified using the G_i* statistic (Getis and Ord 1992) (Fig. 6).

Fig. 6

Geographical clusters of high or low values of the ICR (2013)

The relatively low number of missing values (i.e., 11.1%) and the characteristics of the regions with missing data support the production of ad hoc single imputations. Nonetheless, the use of a single imputation strategy may lead to downwards-biased estimates of the standard errors. This is because a substantial uncertainty characterizes the missing values, but by choosing a single imputation, we ignore it (Gelman and Hill 2007). Due to this possible source of concern, we also performed a multiple imputation which is unbiased under both missing completely at random (MCAR) and MAR missingness mechanisms and allows for valid frequency inference (Newman 2014; Rubin 1996).

The multiple imputation method is structured in three steps: (1) creation of plausible complete versions of the incomplete data, (2) multiple statistical analysis, (3) pooling of the different outcomes into an overall statistical analysis (Rubin 1987). Multiple imputations for the set of missing values are multiple sets of plausible values draws from the posterior predictive distribution of the missing values under a Bayesian model for both the data and the missing-data mechanism (Rubin 1996).

Specifically, using a multivariate normal regression and an iterative Markov chain Monte Carlo (MCMC) method (von Hippel 2013), we obtained 40 complete data sets. Given our actual missingness rate, 40 imputations should provide the necessary information to get estimates of the standard errors that accurately reflect the uncertainty about the missing values (Graham et al. 2007). To create the complete datasets, we included all independent variables in the prediction equations, together with dummy variables for regions and country effects (Reiter et al. 2006). Then, each complete dataset was exploited in multilevel growth curve model. Finally, inferences were combined across datasets. By replacing each missing observation with several imputed values the resulting standard errors and significance tests reflect our uncertainty about our imputation model (Gelman and Hill 2007).

Appendix 2: Collinearity Diagnostics

The analysis of the collinearity diagnostics indicates the collinearity of the independent variables is likely to be inconsequential. Variance inflation factors (VIFs) and tolerance are below the most common thresholds values (see Table 6) (Hair et al. 2010). Yet, the simultaneous inclusion of potentially highly correlated variables (see Table 7) might nonetheless cause the results to suffer of specification and linear dependency problems leading to a failure of significance for the main effects (Greene 2011; O’brien 2007). This potential issue is addressed produced parsimonious models that rely on a reduced number of variables to show the stability of the results obtained in richer models and we did not simultaneously include highly correlated variables in any of the models—e.g., Corruption control and Shadow economy.

Table 6 Collinearity diagnostics

Table 7 Independent variables correlation matrix, 2007–2013