Scarcity Without Leviathan - The Violent Effects Pt 2

3.2 Heterogeneity by proximity to the US We now investigate the prediction in Proposition 1 that the effect of cocaine shortages on Mexican violence is greater in municipalities that are more important for the cocaine trade. We use distance from US entry points as a proxy for a municipality’s importance for trafficking. We do this because, unlike heroin or cannabis, cocaine is not produced in Mexico, and internal consumption of cocaine in Mexico is minor when compared to revenue from exports to the US (see Kilmer et al., 2010). Thus, municipalities close to the US border have a major strategic advantage in trafficking. As shown in table 1, drug-related violence and cartel presence are concentrated in the north of Mexico, which suggests that these are the places that command a greater share of the drug trade. To illustrate the heterogeneity in the response to seizures in Colombia as a function of distance to the US, we estimate the time series equation (3) for groups of municipalities in 901 moving windows of distance to the US, each one of which comprises 10 percent of the municipalities. Figure 3 plots the estimated effects. Near the US border, a 10% increase in cocaine seizures causes a 1.6% increase in violence. The effect vanishes gradually as we move away from the US frontier; in the south of Mexico, for example, cocaine shortages have no impact on violence. We next examine the evidence in more detail using a regression framework. Let pi denote proximity to the US border—the negative distance in kilometers from municipality i to the US nearest entry point. We are interested in explaining the homicide rate in municipality i during month t, hi,t , as a function of municipality fixed effects, time effects, and the interaction between proximity to the US and cocaine seizures in Colombia pi × ln St . We present two types of estimates. First, we estimate a Poisson fixed effects model as in Hausman et al. (1984) and Wooldridge (1999), which assumes the following conditional expectation for the homicide rate:25 Ehi,t = δi αt exp (βpi × ln St + pi × Ft (γ)) . (4) Second, we estimate the following simpler linear model ln(hi,t + r) = δi + αt + βpi × ln St + pi × Ft (γ) + ǫi,t , (5) where the left-hand side is a monotone transformation of the homicide rate that deals with the large dispersion in homicide rates and is well defined at zero. The constant r > 0 is the 90th percentile of the homicide rate in our sample. We choose this particular value for r because the estimated residuals resemble a normal distribution, suggesting that the transformation adequately removes the skewness of homicide rates.26 The transformed value is approximately equal to hi,t +r hi,t ln hi,t ; thus, the estimates can be interpreted as elasticities after multiplying them by the sample average of hi,t +r . hi,t To ease the interpretation of our findings, we report estimates obtained after this normalization. In both models, the dummies αt control for monthly changes that are common across Mexico and δi control for municipality invariant characteristics. The term pi × Ft (γ) flexibly controls for differential trends in violence in the North of Mexico (recall that Ft (γ) includes a cubic polynomial in time and year dummies). Finally, β captures the additional impact of cocaine seizures in Colombia on violence in municipalities that are located 100 km closer 25 The Poisson model has several advantages (see Wooldridge, 2010, chapter 18): First, it is designed for applications with non-negative dependent variables with large dispersion, like homicide rates. Second, it is consistent under very general distributional assumptions. Third, it does not suffer from the incidental parameters problem, allowing us to control directly for municipality fixed effects. We also experimented with negative binomial models and obtained similar results. In this alternative exercise, instead of controlling for fixed effects, we use a wide range of municipality characteristics as controls. 26 Experimenting with different values of r, we obtained similar results. We set r equal to different moments of the homicide rates distribution, including its minimum positive value, as well as its 10th, and 50th percentile. We also obtained similar results when we used the level of the homicide rate as the dependent variable. To save space, these results are not presented here. to the US border. This effect is identified by comparing how high-frequency changes in cocaine seizures by Colombian authorities relate to spikes or dips in violence in Mexican municipalities at different proximities to the US. Regarding inference, we compute standard errors that are robust to heteroskedasticity and serial correlation within municipalities over time. For the estimates presented in the main text, we assume there is no spatial correlation left in the error term, ǫi,t . Time effects already take into account aggregate shocks, and although there may be correlated shocks, we believe that the differential trends by distance account for the bulk of the spatial correlation in violence. The Online Appendix shows that we obtain similar results when we compute standard errors for our estimates of equation (5) that allow for several forms of spatial correlation, as in Conley (1999) and Conley (2008), and when we allow for correlated shocks (other than aggregate) across municipalities within a month and compute standard errors with two-way clustering by municipality and month, as in Cameron et al. (2011). Table 3 presents our results, both for municipalities in all of Mexico (columns 1 to 4) and for those in distance quintiles 1 and 2 (columns 5 to 8). The dependent variable in the top panel is the homicide rate. According to the Poisson estimates from column 1, being 100 km closer to the US implies an additional 0.31% (s.e=0.11) increase in homicide rates in response to a 10% increase in seizures in Colombia. 27 In column 2 (and the remaining even columns) we control for differential time trends in municipalities that have different levels of historical PAN support. These capture the stronger deployment of Calderón’s anti-drug policies in areas where the PAN had more political power. The inclusion of these controls does not change our results.28 27 The Poisson model drops municipalities that had no homicides during our period of analysis, which is the reason why in all panels we have fewer observations in columns 1 and 2 than in columns 3 and 4. 28 We obtain very similar results (not shown here) using the presence of a PAN mayor as a control. Although this control is endogenous, it shows that our estimates are not driven by the election of PAN mayors during Calderón’s administration. Because our measure of PAN support is missing In columns 3 and 4 we estimate the model in equation (5). Being 100 km closer to the US implies an additional 0.14% (s.e=0.04) increase in the homicide rate in response to a 10% increase in seizures in Colombia. In columns 5 to 8, we estimate a larger interaction when we restrict the sample to municipalities in quintiles 1 and 2 of distance to the US. The second panel presents results using the drug-related homicide rate as the dependent variable. We find again that seizures in Colombia have a stronger effect on Mexican drug-related violence as one moves closer to US entry points. The bottom panels show that, consistent with our time series results from Table 2, the differential effect on drug-related homicides in the north of Mexico is driven by assassinations of cartel members and confrontation deaths. As before, we find no differential effects of scarcity on the assassination of government officials. Our results indicate that scarcity created by seizures in Colombia contributed to the sharp increase in violence in the north of Mexico relative to the rest of the country. The estimate in column 5 of the top panel implies that scarcity created by Colombian policies from 2006 to 2009 could explain up to 31 log points of the differential increase in homicides in the north of Mexico relative to the average municipality. This corresponds to 25% of the 128 log point differential increase in violence in the north of Mexico during this period. 3.3 The role of cartels and Mexican policies In this section we investigate if the presence of cartels and Mexican anti-drug policies exacerbate the negative consequences of scarcity in cocaine markets. For our analysis, we use the INEGI homicide rate as our preferred measure of violence. We begin by estimating equation (3) separately for the homicide rate in municipalities with no cartels, one cartel, and at least two cartels. The top panel of Figure 4—constructed in the same way as Figure 2—depicts the estimated change in the homicide rate in municipalities with different cartel presence following a 10% increase in cocaine seizures in Colombia. In line with our model, cocaine seizures in Colombia do not have an effect on violence in Mexican for some municipalities, the models that control for these variables have fewer observations. municipalities with zero or a single cartel. In contrast, at the time of the cocaine shortage, homicides spike in Mexican municipalities that have at least two cartels present. We also estimate equation (3) for municipalities with an historical PAN vote share above and below 28%, which we refer to as high and low PAN support municipalities, respectively. The 28% threshold corresponds to the average PAN vote share in local elections between 1980 and 2000, but our results are robust to using different thresholds. The bottom panel of Figure 4 depicts the estimated change over time in the homicide rate of municipalities with different PAN support following a 10% increase in cocaine seizures in Colombia. Scarcity has a small effect on violence in the sample of Mexican municipalities that have low PAN support, which is not significant. In contrast, we find that scarcity has a larger effect on violence in the sample of Mexican municipalities that have high PAN support (although the effect is not precisely estimated). Table 4 summarizes the results from Figure 4. The left panel (columns 1 to 6) presents our time-series estimates for the homicide rate in Mexican municipalities that have different numbers of cartels. The right panel (columns 7 to 10) presents our time-series estimates for the homicide rate in Mexican municipalities that have different levels of historical PAN support. For each set of municipalities—indicated at the column headers—, we present two estimates of equation (3): our baseline specification and one that controls for the dynamics of homicide rates by adding lags of the dependent variable (presented in the even columns). To test whether the differences in the time-series responses uncovered above are jointly significant, we estimate extensions of equations (4) and (5) that also allow cocaine seizures in Colombia to affect the homicide rate differently in Mexican municipalities that have none, one or multiple cartels, and in municipalities that have a low or high historical PAN support. Let DT Oiy be a vector of cartel presence variables in municipality i during year y, including a dummy for the presence of at least one cartel and a dummy for multiple cartels. Also, let P ANi be a dummy for high historical PAN support (defined as in Figure 4). We estimate equations (4) and (5) by adding the interaction terms ln St × DT Oiy and ln St × P ANi . To ensure that we only identify these interactions from the high-frequency variation in cocaine shortages, we also include a full set of terms of the form DT Oi,y ×Ft (γ1 ), and P ANi ×Ft (γ3 ), which control for time trends in municipalities with a different numbers of cartels, or in municipalities with different historical support for the PAN party. Columns 1 and 4 of Table 5 present the results from this exercise. Columns 1 to 3 present Poisson estimates of equation (4), while columns 4 to 6 present OLS estimates of equation (5). The top panel presents results using all Mexican municipalities, while the bottom panel restricts the sample to those in the first two quintiles of distance to the US. The results in columns 1 and 4 show that cocaine shortages have a larger effect in municipalities in which there is high PAN support. For instance, the estimates in column 4 of the top panel show that a 10% increase in seizures in Colombia causes an additional increase in homicides of 0.36% in municipalities that have high PAN support relative to other municipalities. Although they are not precisely estimated, these coefficients also suggest that cocaine shortages only cause a differential increase in violence when there are at least two cartels operating in a municipality. These estimates are less precise in part because drug cartels locate in municipalities near the US border (see Table 1). As a consequence, it is hard to distinguish the effect of the interaction of scarcity with proximity to the US border from its interaction with the presence of multiple cartels. The estimates in column 4 of the top panel show that a 10% increase in seizures in Colombia causes an additional increase in homicides of 0.47% in municipalities that have multiple cartels relative to other municipalities. In line with our model, we find no difference in the effect of scarcity between municipalities that have a single cartel and those that have none. Our models in columns 1 and 4 estimate the interaction of cocaine shortages with cartel presence and PAN support for the average municipality. One limitation of this analysis is that conflict among cartels and PAN policies may play a more important role in municipalities near the US, where the cocaine trade is booming. In fact, our model suggests that there should be a triple interaction between shortages, proximity to the US, and other factors that worsen the conflict among cartels. To explore this possibility, we include in columns 2 and 5 triple interactions of the form ln St × pi × DT Oiy and ln St × pi × P ANi , which allow us to test whether cartels and PAN policies had a differential effect in areas close to the US border.29 In columns 3 and 6 we control for ln St × p2i to ensure that our triple interactions are not capturing non-linearities in the relationship between scarcity and proximity to the US border. To ease the interpretation of all these models, the main effects of the double interactions are now evaluated at border municipalities with multiple cartels and high PAN support. In line with our model and the above conjecture, coefficients on triple interactions involving high PAN support and the presence of multiple cartels in columns 2 and 5 are all large and positive, and most of them are statistically significant (except for the triple interaction with high PAN support in column 2, which is not precisely estimated). These triple interactions are quantitatively important. To see this, we can go back to the double interactions between cocaine seizures in Colombia and cartel presence, which are now evaluated at border municipalities where the cocaine trade is booming. The estimates in column 5 of the top panel show that when we focus on municipalities at the US border, a 10% increase in seizures in Colombia causes an additional increase in homicides of 2.42% in municipalities that have multiple cartels relative to those that have one or none. The fact that this estimate is five times larger than its counterpart in column 4 illustrates the quantitative importance of the triple interaction term for multiple cartels. The triple interaction term in column 5 indicates that the 2.42% gap narrows by 0.33% when we move 100 km further away from the US frontier, which implies that cartels had a more detrimental effect on the response of violence to scarcity in the North of Mexico. As before, the estimates in columns 2 and 5 show that there is no differential effect of scarcity on violence in municipalities controlled by a single cartel (relative to municipalities that have no cartels) even in the very North of Mexico (the triple interaction term is a precisely estimated zero). Likewise, we can go back to the double interaction between cocaine seizures in Colombia 29 To guarantee that we only exploit the high-frequency variation in cocaine shortages and homi- cides to identify these interactions, we also include a full set of differential time trends of the form pi × DT Oiy × Ft (γ2 ), pi × P ANi × Ft (γ4 ). and high PAN support, which are now evaluated at border municipalities. The estimates in column 5 of the top panel show that when we focus on municipalities at the US border, a 10% increase in seizures in Colombia causes an additional 1.64% increase in homicides in municipalities that have high PAN support relative to those that have low support (about five times the corresponding estimate in column 4). The triple interaction term indicates that this 1.64% gap narrows by 0.19% as we get 100 km further away from the US frontier. This triple interaction indicates that PAN policies had a more detrimental effect on the response of violence to scarcity in the North of Mexico.30 The interpretation that emerges from Table 5 is that Calderón’s policies exacerbated the negative consequences of supply shortages, but this happened mostly in areas near the US border where the cocaine trade boomed. The evidence in this section supports the possibility that starting in 2006, Calderón’s policies and the expansion of Mexican cartels provided a dangerous blend that amplified the response of violence in Mexico to cocaine shortages generated in Colombia. From 2006 to 2010, the percentage of municipalities that had multiple cartels increased from 5% to 17%. Before 2006, there were fewer and more isolated cartels, and informal arrangements among cartel leaders presumably kept violence and rapacity low or prevented them altogether. After 2006, the environment was very different: fragmented cartels contested multiple key locations, especially in the North of Mexico, and the policy of beheading cartel leaders may have fueled conflict and division among cartels.31 Our model 30 To gauge the quantitative implications of these results, consider the case of two border mu- nicipalities near US entry points with different historical levels of PAN support: Ciudad Juarez (Chihuahua) and Nueva Laredo (Tamaulipas). The historical voting for PAN in Ciudad Juarez was about 44%, whereas in Nueva Laredo it was 14%. A reduction in the supply of cocaine of 95.6 log points (observed between 2006-2009), would cause a differential increase of 15 to 30 log points in homicides in Ciudad Juarez relative to Nueva Laredo (using the estimates in column 2 or 5 of the top panel). During this period, homicides fell in Nueva Laredo by 23.4 log points and increased in Ciudad Juarez by 236.9 log points. The interaction between the decline in the net supply of cocaine and greater PAN support in Ciudad Juarez may explain 5% to 10% of this differential change. 31 In line with this view, when we re-estimate equation (3) for all the period 2003-2010, we find and evidence indicate that it is in this environment that the scarcity generated by Colombian anti-drug policies contributed to the increase of violence in Mexico. 3.4 Is scarcity (and the price effect it brings) the right channel? We have documented that cocaine seizures in Colombia are associated with large spikes of violence in Mexico at a monthly frequency. We argue in our theoretical model that this occurs through an increase in cocaine revenues, which take place so long as demand is inelastic. Our interpretation relies on three key assumptions, which we now discuss. The first assumption is that cocaine seizures by Colombian authorities are associated with periods of scarcity in transit markets. This assumption would not hold if seizures were simply increasing during periods of high cocaine productivity in Colombia or during periods of high cocaine demand in downstream markets. In both of these cases, cocaine seizures would be associated with abundance, not scarcity. We do not believe that these concerns confound our interpretation. As Figure 5 shows, estimates of variants of equation (3) reveal that there is no high-frequency relationship between seizures and other anti-drug policies in Colombia. Seizures of chemical precursors, interdiction of solid inputs, and the destruction of cocaine labs do not increase during months in which the Colombian government seizes large quantities of cocaine. This finding suggests that the high-frequency variation in cocaine seizures in Colombia does not simply reflect changes in cocaine production (driven by demand or productivity). If that were the case, all measures of interdiction policies would respond together and would be correlated.32 We also estimate a large and negative elasticity of cocaine seizures in Mexico with respect to high-frequency changes in cocaine seizures in Colombia. Although it is not precisely estimated, this elasticity is telling: a 10% increase in cocaine seizures in Colombia during one month reduces cocaine seizures in Mexico by 6.8% during the same month (standard that the effect of negative supply shocks on violence is driven by Calderón’s term. 32 Unfortunately, there are only yearly measures of gross cocaine production and productivity in Colombia. Thus, we cannot directly test if months of high seizures correspond to months of higher temporal productivity or cocaine production. error=4.6)—an elasticity that roughly corresponds to the share of Andean cocaine supplied by Colombia. This finding suggests that during months when the supply of Colombian cocaine turns out to be lower than expected, Mexican cartels traffic less cocaine to the US market. Cartels do not appear to offset the lack of Colombian cocaine by relying on stored drugs or cocaine from other suppliers like Peru and Bolivia. This high-frequency relationship is in line with the anecdotal evidence that suggests that once a kilogram of cocaine leaves Colombia, it takes Mexican cartels 4 to 6 weeks to smuggle it and sell it in the US market (see Gaviria and Mejia, eds, 2016). Our estimates and the anecdotal evidence support the key assumption that, even at a monthly frequency, cocaine seizures in Colombia disrupt the cocaine trade downstream and generate significant scarcity. The second assumption is that any coincidence in timing between policies or market conditions in Mexico and Colombia breaks down at higher frequencies once we flexibly control for time trends. Otherwise, changes in Mexican policies could confound our estimates. This assumption could fail if there were complementarities in enforcement or cooperation between authorities across countries. Likewise, changes in enforcement or market conditions in Mexico could raise the demand for Colombian cocaine and, hence, cocaine seizures in Colombia. We do not believe such alternatives explain our findings. The timing of the estimated effects suggests that seizures in Colombia do not react to past violent upsurges in Mexico, which supports our view that seizures in Colombia are not driven by market conditions in Mexico. If anti-drug policies in Mexico drove up both violence in Mexico and the demand for Colombian cocaine, one would observe a positive correlation between seizures in both countries, but this is not present in the data. If anything, the opposite holds as shown in figure 5. Contrary to these alternative explanations, there is no correlation between highfrequency changes in seizures in Colombia and the seizure rate of cocaine in Mexico, nor is there any correlation with other anti-drug policies in Mexico, such as heroin seizures.33 33 We construct the cocaine seizure rate in Mexico as total seizures over the net supply of cocaine flowing out of Colombia. Because we do not observe monthly cocaine production, we interpolate annual production figures to obtain monthly estimates. We then subtract monthly seizures in Additionally, the effects we uncovered are specific to municipalities that have two or more cartels and high PAN support. If market conditions in Mexico affected seizures in Colombia, one would also observe a spurious correlation between seizures in Colombia and violence in Mexican municipalities that have low PAN support. If there were complementarities in enforcement across countries, one would also observe a spurious correlation between seizures in Colombia and violence in Mexican municipalities that have a single cartel, and this would be so because operations against this cartel would also induce more seizures in Colombia. Thus, the heterogeneity of the effects we uncovered is consistent with the model and it rules out other potential explanations of our findings. The third and final key assumption is that cocaine shortages increase revenue, which is theoretically the case so long as demand for cocaine is inelastic (as the existing evidence suggests). Unfortunately, we cannot measure prices or the amount of cocaine revenue in Mexico on a monthly basis, and this precludes any direct test of this assumption. Instead, we provide some reality checks that suggest that cocaine shortages in Colombia seem to increase cocaine revenues downstream. Annual price data from STRIDE indicates that, from 2007 to 2010, cocaine prices increased dramatically as the supply of Colombian cocaine declined. While net cocaine production in Colombia went down from about 520 MT per year in 2006 to about 200 MT in 2009, the average price per pure gram of cocaine on US streets went up from about $114 in 2006 to about $180 in 2009, and the wholesale price went from $40 to $68 during the same period (the Appendix summarizes the available price data).34 Taking into account net production from Peru and Bolivia—which increased from 340 metric tons to 390 metric tons—the total gross drug trade revenue measured using wholesale prices in the US increased from 34 billion to 40 billion dollars—a 15% increase. Using retail prices, we find that between 2006 and 2009, cocaine revenue increased from 98 to 106 billion dollars—a Colombia to obtain an estimate of the cocaine flowing out of Colombia. 34 There is no reliable data on cocaine prices at higher frequencies. Although quarterly prices can be constructed using the raw STRIDE data, some have expressed doubts about the quality of quarterly variation in these data (see Horowitz, 2001; Arkes et al., 2008). 7% increase. These trends support the possibility that cartels face an inelastic demand. 4 Concluding remarks Focusing on the case of the Mexican cocaine trade, this paper investigates the role of scarcity as a determinant of violence in markets without a centralized authority. A standard conflict model in which cartels dispute unprotected revenue suggests that scarcity increases violence if the demand for cocaine is inelastic. The increase in violence should be larger in municipalities that are more important for the cocaine trade, where two or more competing cartels are present, and where, because of policies that lead to the arrest and killing of cartel leaders, cartels behave more shortsightedly and do not cooperate. We explore the link between scarcity and violence using monthly data from Mexican municipalities from December 2006 to December 2010. In line with the predictions of our model, we find that scarcity generated by large and temporary increases in cocaine seizures in Colombia increased violence in Mexico, especially at locations close to US entry points. Most of the effect was driven by municipalities that had at least two competing cartels, while there were no effects in municipalities that had one or no cartels. Our evidence supports the view that President Calderón and the PAN’s strategy of beheading cartels amplified the effect of scarcity in areas that had high PAN support. Our results indicate that greater cocaine seizures in Colombia after 2006 raised prices and contributed to the increase in Mexican violence. The sharp decline in the supply of Colombian cocaine from 2006 to 2009 of 95.6 log points may account for 10%-14% of the increase in Mexican violence during this period, and 25% of the differential increase of violence in the North of Mexico. Our paper documents spillovers from anti-drug policies that transcend national frontiers. By generating scarcity, anti-drug policies in Latin American countries may be fostering violence—even if the burden is borne by other countries. Throughout this paper, we have emphasized a particular mechanism that suggests that cartels use violence to secure the increases in revenue generated by scarcity. There remains the possibility that this is not the sole mechanism that explains our findings. For instance, when cocaine shortages occur, cartels might increase their efforts in an attempt to achieve a target level of sales, and this could increase violence. Moreover, if cartels cannot perfectly observe seizures, they may confound shortages with wrongdoing on behalf of other cartels or their members, and this could potentially trigger cycles of violence.