Our Hypotheses

We argue that the geographical concentration of criminal violence in surrounding areas of the studied infrastructures is explained by DTOs’ ambition for monopolistic control of them that can lead to logistic, tactical, and economic advantages (Dell Reference Dell2015; Fuerte Celis et al. Reference Fuerte Celis, Pérez Lujan and Cordova Ponce2019). Previous studies have highlighted the importance of airports and ports for the shipping of large amounts of drugs (Atkinson et al. Reference Atkinson, Kress and Szetchman2017). Similarly, DTOs may consider highways important for the distribution of drugs to nearby cities or specific domestic markets. Furthermore, free passage of drugs to key drug markets like passage through the border with the United States offers a considerable advantage to any DTO, especially if such an arrangement excludes other DTOs from doing so (Fuerte Celis et al. Reference Fuerte Celis, Pérez Lujan and Cordova Ponce2019). Finally, (illegal) extraction of oil or gas from pipelines provides additional financial resources that can help DTOs increase their income (Berbotto and Chainey Reference Berbotto and Chainey2021; Jones and Sullivan Reference Jones and Sullivan2019). As DTOs mainly seek economic gains (Phillips Reference Phillips2015), these potential logistical, tactical, and economic advantages may incite DTOs to fight one another for monopolistic control.

Such infrastructures, however, do not produce similar levels of violence in surrounding territories. To explain these variations, we break down the impact of these infrastructures on violence into two effects. The first is the effect on violent competition, measured by the number of DTOs that enter the area to compete through violent means for monopolistic control. As violence is always costly, if DTOs decide to employ it, the benefits must outweigh its costs. ^{Footnote 5} We expect that, as the benefits of monopolistic control over the infrastructure grow, the number of DTOs deciding to compete (by employing violence) for the control of the infrastructure will increase. And, as current research shows, increases in the number of DTOs leads to an increase in violence (Atuesta and Ponce Reference Atuesta and Ponce2017). In sum, the first effect considers the composition of violent competition for the control of the infrastructure.

The second effect is determined by certain characteristics of the analyzed infrastructures that may make DTOs competition more violent, regardless of the number of additional rivals. We argue that the costs of excluding other existing DTOs (for those that decided to compete for control of the infrastructure) from using the infrastructures also explain the level of violence among DTOs. The greater the cost, the less inclined a DTO may be in fighting to ensure monopolistic control. We posit that the costs to exclude DTOs from using the infrastructure beyond the costs of fighting are of two principal types: 1) monitoring costs of learning the location of rival DTOs when using the infrastructure; and 2) costs of bribing and monitoring state officials for excluding other rival DTOs from using the infrastructure. As these costs increase, DTOs might avoid trying to exclude other DTOs from using the infrastructures. For instance, as the size of an infrastructure grows, monitoring costs increase because rivals can maintain their anonymity more effectively and it becomes more costly to deploy more resources to track other DTOs. Bribing officials also becomes more costly as enforcing compliance of corrupt agreements gets harder. Following this reasoning, DTO violence may be negatively correlated with monitoring and bribing costs. The rest of this section presents the hypotheses based on these two effects.

Empirical Analysis

To test our hypotheses, we built a novel database with information at the locality-year level from 2006 to 2011. Localities are spatial units below the municipal level that are employed by the National Institute of Statistics, Geography, and Information (INEGI)—the Mexican statistics office—to collect census data. Specifically, a locality is defined as any place that is occupied by one or more households that may or may not be inhabited (INEGI 2010). Since we rely on a geocoded database of criminal violence, we need spatial polygons to assign events to a spatial unit. As such, we only use those localities for which INEGI had defined polygons. INEGI draws spatial polygons only for localities that have at least one well-defined block. The rest are registered as points and are mostly rural localities—i.e., with less than 2,500 inhabitants, according to INEGI. Unfortunately, we cannot include them in the sample as there are not pre-defined spatial polygons to link them with the events of violence. ^{Footnote 9} This might be a potential drawback because we might be truncating events to only urban areas, whereas some violence might occur in overlooked rural areas—we acknowledge this limitation. However, 90% (45,428 out of 49,981) of the localities included in the sample are rural. In addition, we are including 90% (32,704 out of 36,378) of violent events from the geocoded database (Atuesta, Siordia, and Lajous Reference Atuesta, Siordia and Madrazo Lajous2018).

By employing localities as the level of analysis we can more accurately capture local dynamics of violence, in contrast to studies using more aggregated units, like the municipal level. Moreover, we ensure there is at least one inhabitant in the localities we employ. Without any inhabitants, transportation facilities, or extraction sites, localities are likely to be irrelevant to explain violence in the empirical analysis. ^{Footnote 10} If there is a bias produced by omitting such areas, we do not anticipate it to be significant as the level of violence in these localities is marginal.

To measure our dependent variable, we rely on the geocoded version of the CIDE-PPD database (Atuesta, Siordia, and Lajous Reference Atuesta, Siordia and Madrazo Lajous2018) and matched each point to the spatial polygons of localities. We measure the dependent variable in two ways. First, the variable Total Violence counts the number of homicides by locality and year regardless of the type of event and the victim. Second, the variable Inter-DTO Violence counts deaths registered only from inter-cartel violence (number of homicides excluding security officials killed in such events). ^{Footnote 11} Even though our focus is on violent interactions between DTOs, these groups might also expand violence when fighting other actors such as officials or civilians, which are included in the measure of Total Violence.

Key independent variables are measured as the distance of each locality to key segments of the US border, airports, highways, non-military seaports, and pipelines. Except for the distance to the US border, the rest of our measures indicate the distance of each locality to the closest object of interest throughout Mexican territory (natural resource or transportation facility). For the case of the US border, we employ 100 km segments of the border for each port of entry (50 kms on each side). This helps us consider areas where DTOs are more likely to both develop illegal networks with customs officials and build tunnels to connect the transportation of drugs to highways and storage facilities on the US side. ^{Footnote 12} We calculate Great Circle distances to account for the fact that the earth is an oblate spheroid (see Keele and Titiunik Reference Keele and Titiunik2015). All distances are coded in kilometers and logged to account for differences in scale. ^{Footnote 13}

The location and installation of the studied infrastructures are plausibly exogenous to other factors such as governmental security policies—which are a major driver of drug violence (Lessing Reference Lessing2017; Durán-Martínez Reference Durán-Martínez2018). Most infrastructures were built during the administration of the Institutional Revolutionary Party (Partido Revolucionario Institucional, PRI) and their construction depended on economic drivers rather than on factors linked to security. Likewise, the US-Mexico border was drawn in the nineteenth century. Considering these facts, it becomes reasonable to assume that these infrastructures are exogenous to changes in violence levels. This enhances confidence on the right causal path between resources and the levels of violence: from infrastructures to violence. ^{Footnote 14}

We control for covariates that the literature on criminal violence identifies, related to both the number of competing DTOs and criminal violence. Some scholars suggest that military strategies increase violence and fragmentation of DTOs—that the capturing or removal of key DTO personnel can in turn incite internal fights for DTO leadership (breaking up single DTOs into multiple factions) or may incite violent challenges from more powerful DTOs upon those weakened by the state (Williams Reference Williams2012; Phillips Reference Phillips2015; Atuesta and Ponce Reference Atuesta and Ponce2017). As such, we control for the effect of state enforcement policies with the number of killed DTO members in confrontations with security agencies during the previous year.

We also employ binary variables to account for the presence of each DTO in each locality. Durán-Martínez (Reference Durán-Martínez2018) points out that drug violence increases when DTOs organize their security agencies with external recruits. Since DTOs use different recruitment strategies, we account for this factor and other intrinsic characteristics of DTOs through these dichotomous variables. We consider the major Mexican DTOs identified by Coscia and Ríos (Reference Ríos2012). ^{Footnote 15} Furthermore, previous scholarly work suggests that the partisan identity of local governments can shape the levels of violence both because some parties use more crackdowns and because opposition parties are unprotected by the federal government (Dell Reference Dell2015; Trejo and Ley Reference Trejo and Ley2016). We thus include the partisan identity of the municipality and the state governments.

Moreover, we control for socio-economic variables at the locality level. We include the natural logarithm of the population as in more populated places the incidence of violence might be higher. We also incorporate a measure of social development as this factor could explain citizens’ involvement in DTOs. This measure captures the number of deficiencies in key development areas such as health, public services, and spaces between dwellings (CONEVAL 2010). Moreover, we incorporate the percentage of people over 15 years old without basic education to account for schooling, which might explain easier recruitment by DTOs (Osorio Reference Osorio2015). Finally, we estimate all the models with fixed effects by year to account for unobserved heterogeneity of each year that might drive change in violence and number of DTOs. ^{Footnote 16}

To test our hypotheses, we employ mediation analysis through a structural equations model. This allows us to break down the effect of geographical distances on violence into two effects: 1) the direct effect of such distances on the number of homicides; and 2) the indirect effect (or mediated effect) of spatial distances on violence occurring because of changes in the number of DTOs engaging in violence ^{Footnote 17} (Preacher, Zyphur, and Zhang Reference Preacher, Zyphur and Zhang2010). While the indirect effect evaluates our “A” Hypotheses, the direct effect assesses the validity of our “B” Hypotheses. Following this description, we employ the following equations:

(1) $$violenc{e_{lt}} = \alpha + {\theta _k}distanc{e_l} + \gamma Number\;of\;DTO{s_{tl}} + X^\prime {\delta _k} + {\phi _t} + {\varepsilon _l}$$

(2) $$Number\;of\;DTO{s_{lt}} = \alpha + {\theta _k}distance + Z^\prime {\xi _k} + \;{\phi _t} + {\varepsilon _l}$$

In Equation (1) we estimate the direct effect of being closer to any of the infrastructures considered strategic for DTOs on levels of violence in locality $l$ at time $t$ . In Equation (2) we estimate the indirect effect of the distance of localities to the infrastructures on the number of DTOs of the criminal market inside each locality $l$ at time $t$ . ${\phi _t}$ is the year fixed effect. $Z'$ and $X'$ are k x n control matrices for each equation. We estimate both equations using a negative binomial model as the dependent variables are counts with variances higher than the mean, rendering a Poisson model inappropriate.

Note that we include different control matrices given that both processes—violence and the number of DTOs—depend on different factors. Matrix $Z'$ includes a one-year lagged term of the number of groups to account for serial correlation and all controls except for which major DTO had presence in each municipality. Matrix $X'$ , by contrast, includes a one-year lagged term for the levels of violence and the variable indicating which group had presence in each municipality.

Criminal violence has temporal dynamics that are important to consider. Our current strategy is based on estimating the models with all locality-year units pooled. We account for such temporal dynamics of the dependent variable using year fixed effects as we built our dataset with a panel structure. Moreover, in the Online Appendix we include models using a cross-section design summing up the total number of deaths per locality from 2006 to 2011. This model helps show our estimation is robust to both data structures and to the fact that our key independent variables do not vary over time.

We also acknowledge that DTOs might be able to anticipate some of the costs involved in excluding rival DTOs from using the infrastructures. However, we pose that such expectations are imperfect due to several informational problems. First, DTOs are uncertain about the duration of the conflict against rival DTOs. Second, DTOs might not have complete information on the number of DTOs already competing for control of the infrastructures. Third, it is also hard to anticipate how many other DTOs might enter the area to compete for control of the infrastructure. We clarify that the absence of perfect information does not preclude DTOs from estimating some costs and making decisions on whether to compete. If DTOs enter to compete, our “A” Hypotheses (first effect) account for it. While controlling for the number of DTOs competing (measured by the indirect effect), the second effect, employed to elaborate our “B” Hypotheses, focuses on the costs of excluding competitors given certain level of competition. Finally, we do not employ estimates of the benefits or costs of excluding other DTOs from using the infrastructures to support our causal argument. We acknowledge this is a limitation of our research design. Although our results may convincingly point in the direction of a causal link, further research could attempt to provide estimations of the components of DTOs’ finances.

Results

Table 1 reports the results of our estimations. As our independent variables are distances, a negative sign indicates that localities further away from a given resource have less incidence of violence than those closer; but this also implies the contrary, i.e., localities closer to a given infrastructure have more violence than those further away. Panel A shows the estimations of Equation (1) where we assess the effect of the location on violence. Panel B shows the estimations that explain the number of DTOs as shown in Equation (2). Columns 1 and 2 show that localities closer to strategic US border segments, to ports, and to airports register more DTOs than localities further away from these infrastructures.

Table 1. Mediation Negative Binomial Estimates

Note: Standard errors in parentheses; #p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

To analyze the effects of each infrastructure on violence, in Tables 2 and 3 we report the exponentiated coefficients (IRR) of the estimation of the direct and the indirect effects calculated from the structural model. It is noteworthy that IRR values below 1 reflect a negative impact, which means that violence decreases as distance to the infrastructure increases (or violence increases as distance to the infrastructure decreases). In contrast, IRR values above 1 reflect a positive impact; that means that violence increases as distance to the infrastructure increases. Table 3 shows the same calculations as in Table 2 but using the number of inter-DTO homicides as the dependent variable, and the results are similar. In Table 2 we employ the first model specification (from Table 1) to calculate the impacts of the distance to infrastructures on the total number of homicides. Results indicate that that an increase of one percentage point in the proximity to the US border, drug-related homicides increase by 46% (IRR=0.546) due to the increased number of DTOs, supporting Hypothesis 1A, whereas homicides increase by 33% (IRR=0.662) regardless of more groups, supporting Hypothesis 1B.

Table 2. Direct, Indirect, and Total Effects on Violence

Note: Exponentiated coefficients; standard errors in parentheses

#p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

Table 3. Direct, Indirect, and Total Effects on Violence

Note: Exponentiated coefficients; standard errors in parentheses

#p < 0.10, *p < 0.05, **p < 0.01, ***p < 0.001.

Column 3 in Table 2 indicates that an increase of one percentage point in the proximity of localities to ports, implies an increase of 17% (IRR=0.833) in the number of total homicides due to a higher number of DTOs employing violence, but of only an increase of 10% (IRR=0.898) as a direct effect, which supports Hypothesis 2B. By contrast, although an increase of one percentage point in the proximity of localities to airports is associated with 10% (IRR=0.896) more deaths as an indirect effect, the same change in the distance sees a reduction in the number of homicides by 21% as a direct effect, rejecting Hypothesis 2B (IRR=1.212), but only for airports.

Highways show a different picture. When localities are closer to highways, they experience less violence than those further away, although the impact is only 7% on average (IRR=1.068) as a direct effect, which supports Hypothesis 3B. In addition, the indirect effect is non-significant, which also shows support for Hypothesis 3A.

Finally, for an increase of one percentage point in the proximity to a pipeline, localities have, on average, 40% (IRR=1.390) fewer drug-related homicides due to the reduced number of groups, supporting the validity of Hypothesis 4A, but only 28% (IRR=1.275) fewer homicides regardless of the number of groups, supporting arguments in Hypothesis 4B.

In sum, Tables 2 and 3 confirm that closeness to ports and the US border positively correlates with increased violence, occurring through the effect on the number of DTOs competing violently and through mechanisms suggested in our “B” hypotheses (indirect and direct effects). The effect of closeness to airports is mixed. Only the indirect effect (through the effect on additional number of DTOs competing violently) seems to be positively correlated with greater violence. However, this effect is not strong enough to produce significantly greater violence when we observe the total effect indicator. These results suggest that some DTOs seem likely to achieve monopolistic control of airports. Nevertheless, since the costs of monitoring/excluding other DTOs are high, DTOs ultimately tend to share (relatively) peacefully with other DTOs their access to airports. Finally, violence is not higher in localities close to highways and pipelines. This suggests DTOs are not interested in competing violently for the monopolistic control of these infrastructures or excluding other DTOs from using it. ^{Footnote 18}

To show the substantive magnitude of our estimations, Figures 1 and 2 show the expected means of killings as the distances to different infrastructures vary, based on the first model shown in Table 1. The graphs report how the expected mean of killings change between localities being in the fifth percentile in distance to a particular infrastructure and localities in the ninety-fifth percentile of the distribution, with the rest of the covariates held at their means. Some infrastructures are more comparable to each other given their size; therefore, we first compare in Figure 1 the expected values of violence related to the largest infrastructures: highways, pipelines, and the US border. As displayed, localities closer to the US border have significantly higher levels of violence, whereas localities closer to highways and pipelines have less violence than those further away. For instance, when the average locality is close to an oil pipeline (1 km away) it would experience a third of the killings of a similar locality far from it (215 km away), whereas a locality close to the US border (1 km away) will have two times more killings than one far away (215 km away). Interestingly, localities close to a highway (1 km away) have almost 75% more killings than those close to a pipeline. This suggests that the differences between the US border and the other two infrastructures likely respond to greater competition and the relatively low costs of excluding rivals. ^{Footnote 19}

Figure 1. Predicted Mean of Killings.

Figure 2. Predicted Mean of Killings by Distance to Resources.

The two other infrastructures, ports and airports, are smaller in nature. We thus show in Figure 2 the expected mean of killings as distances to these sites vary. In this case, the direct effect seems to explain more convincingly differences in the expected means of killings between these two types of infrastructures. ^{Footnote 20} For example, when the average locality is located 1 km away from a port (in the fifth percentile), it would experience 24% more killings than a similar locality located 320 km away. In contrast, an average locality close to an airport (1 km away) would have an average of killings almost 38% smaller than one further away (320 km away). ^{Footnote 21}