Mexican Cartel Wars: Fighting for the U.S.

Opioid Market
Fernanda Sobrino∗†

This version: May 21, 2020

Click here for the latest version

Abstract
The number of major Drug Trafficking Organizations (known as cartels) in Mexico
increased from four to nine over the last two decades. This was accompanied by an
increase in drug trade related violence. This paper examines the relationship between
violence and competition for market share among cartels. To measure cartel presence,
a difficult to measure phenomenon, I construct a novel data set of cartel presence
across Mexican municipalities by scraping Google News and using natural language
processing. To study how market size and structure interact with violence, I exploit
two empirical strategies using within municipality variation. First, I interact heroin
prices with agro-climatic conditions to grow opium poppy, using exogenous variation in
demand for heroin from the 2010 OxyContin reformulation. This reformulation made
OxyContin harder to abuse and led some opioid abusers to switch to heroin. Second,
I exploit variation in the timing of cartel entry in a municipality. Cartel presence
increases substantially after 2010 in municipalities well-suited to grow opium poppy.
As more cartels enter a market, homicide rates increase. These results suggest that
substantial part of the increase in violence that Mexico experienced in the last fifteen
years is due to criminal groups fighting for market share of heroin, not only due to
changes in government enforcement.

∗
I am very grateful to Thomas Fujiwara, Micaela Sviatschi and Esteban Rossi-Hansberg for their in-
valuable support and guidance. For helpful comments, I thank Janet Curie, Anne Karing, Alicia Adsera,
Leonard Wantchekon, Swati Bhatt, Eduardo Morales, Matteo Bobba, and other visitors and participants
to the Health and Development Student Seminar Series. I am particularly grateful to Will Lowe for helping
me with the deep learning in this paper. I also thank Patrick Signorent, Bogdan Popescu, Nikita Melnikov,
Faizaan Kisat, Patrick Agte, and all the other visitors and participants to the weekly Development Tea.
Thanks to Hannah Rubinton, Jenny Shen, Chrissy Ostrowsky, Mauricio Matsumoto and Fabiola Alba for
continuous support and encouragement. Comments are welcomed and all errors are my own.
†
Department of Economics, Princeton University: [email protected]

1
1 Introduction
They understand the prescription drug issue here, and that is one of the major reasons
why you are seeing the expansion of poppy production.1
-Jack Riley, former DEA special agent in charge of the Chicago field office, when asked
about cartel expansion and the opioid crisis in the U.S.

Very much like any corporation. They judge the market demand and they shift accord-
ingly, and I would have to say the cartels shift much more efficiently and quickly than any
major corporation, because they don’t have to deal with the bureaucracy.2
-Mike Vigil, former chief of international operations of the DEA, when asked about Mexi-
can cartels business strategies.

Drug trafficking is the second most lucrative illegal activity with an estimated global
revenue of $539 billion dollars each year.3 Mexican Drug Trafficking Organizations (know
as cartels) are notorious criminal groups; they are the largest foreign suppliers of heroin,
marijuana, methamphetamines, and cocaine to the United States. The number of major
drug cartels in Mexico increased from four in the early 2000s to nine organizations by
2016.4 The increase in the number of cartels has been accompanied by a peak in violence
across Mexico with 250,547 homicides, as well as 330,000 displaced, and 37,400 missing
persons that can be directly attributed to organized crime.
Billions of dollars5 are spent every year to reduce the presence of organized crime
and the negative externalities associated with it. However, there is limited evidence on
the underlying determinants of these externalities. This paper argues that the key to
understanding the externalities is the relationship between market structure, institutional
context, and violence.
In this paper, I seek to answer two questions. First, how do cartels react to an exter-
nal demand shock for heroin. Second, what is the relationship between market structure
and violence in the illegal drug market. I answer these question with three contributions.
First, I create a novel data set on cartel presence across Mexican municipalities. Second,
I use an external demand shock to the heroin market, the 2010 OxyContin reformulation,
to estimate the effect of this market shift on cartel entry. Then, I estimate the relation-

1
Ahmed (August 2015).
2
Woody (November 2017).
3
Counterfeiting is the most lucrative illegal activity with a yearly revenue of approximate $928 billion dollars
Mavrelli (2017).
4
Through this paper I will use the word cartel and Drug Trafficking Organizations as interchangeable.
The United States Department of Justice define a major Drug Trafficking Organization as a "complex
organization with highly defined command and control structures to produce, transport, and or distribute
large quantities of drugs."
5
The US government spends around 15.8 billion dollars each year in law enforcement related to illegal drugs.
https://www.drugwarfacts.org/chapter/economics

2
ship between cartel entry, exit, and violence. I find that the heroin market became less
concentrated and that violence increases with cartel entry into a municipality.
The context for this analysis is the increase in demand and supply for legal opioids in
the United States between 1996 and 2010. OxyContin became popular for recreational use
and abuse because the drug offered more of the active ingredient, oxycodone. The pills
were easily manipulated by crushing them so oxycode is released all at once instead of
slowly. To reduce the misuse of legal opioids, in 2010 the FDA approved the reformulation
of OxyContin. The new pill was harder to crush and dissolve. This change made the drug
less appealing to some users who changed to its illegal substitute, heroin. It is estimated
that around 80% of heroin users started with a legal opioid (Muhuri et al. 2013). This
resulted in higher heroin prices in the United States and a subsequent increase in heroin
production by Mexican cartels. Mexican Drug Trafficking Organizations have historically
produced some opium for the US market but the percentage of heroin of Mexican origin
seized by the DEA increased from 19% in the early 2000s to around 90% by 2016 (DEA
2018).
This shock allows me to use two different sources of exogenous variation, agro-climatic
conditions to cultivate opium poppy6 and the change of heroin prices in the United States.
I use this variation to measure the effects of the demand shock on cartel presence and
violence across Mexican municipalities. Data on illegal economic activities, such as drug
trafficking are sparse, so this analysis is subject to significant data restrictions. Official
data does not exists on the cultivation or distribution of drugs nor does a complete panel
of which trafficking organizations operate in each territory. To deal with this problem I
use several machine learning techniques that allow me to approximate measures for illegal
drug trafficking.
To the best of my knowledge, data detailing the presence of different cartels in different
Mexican municipalities after 2010 is not available. Following Coscia and Rios (2017), I use
web-content to obtain information of an otherwise difficult to measure phenomena. The
idea of using web content, in particular Google-News, to generate full panel data of cartel
presence by municipality is motivated by the assumption that local and national media
outlets contain regular, detailed, and systematic coverage of when and where criminal
organizations are operating.
I use a web crawler, an automated script which methodically browses the web, to
extract articles related to a municipality and cartel pair. Then, I use natural language
processing to validate whether an article is actually discussing a cartel being active in that
municipality. I use a semi-supervised7 convolutional neural network (CNN) to achieve this.
A CNN is a series of algorithms primarily use to classify images but that have been proven
to achieve good performance for sentence classification (Kim 2014). A CNN allows more
6
Opium poppy cultivation is illegal in Mexico so any production goes to the illegal market and opium is
the main precursor for morphine, codeine, heroin, and oxycodone.
7
Semi-supervised learning uses label and unlabeled data to gain more understanding of the population
structure.

3
complex relationships between words in a sentence than a simple bag of words algorithm.8
I trained the CNN by manually classifying 5,000 sentences as either cartel presence or not.
The resulting data set is highly correlated with Coscia and Rios (2017).9 Their data set use
the same web crawling technique and then classify the presence of the cartels based on the
relative number of results extracted from Google. The technique proposed here extends
this by looking into the articles and applying state-of-the-art natural language processing
techniques to classify the articles. The data is also highly correlated with two data sets
collected by hand from local newspapers (Sánchez Valdés 2015, 2017)10 these data sets are
snapshots for two different states in two different periods. The data is also highly correlated
to State level data from the DEA.
Second, I build an opium suitability index for Mexican municipalities. A suitability
index measures an area comparative advantage for crop cultivation base on geographic and
climatic characteristics. Ideally I would like to observe opium yields from each Mexican
municipality. Unfortunately Mexico is a relatively new player in the mass production
of opium. As a result, historical data on opium yields does not exist by municipality.
Hence, I use yields from Afghanistan and a rich set of agro-climatic conditions to build a
suitability index using an elastic net. An elastic net is a penalized OLS that helps reducing a
model dimensionality by generating zero-valued coefficients.11 . The agro-climatic variables
chosen through the optimal elastic net were then employed to build the suitability index
for Mexican municipalities.12 Across municipalities over time, the index is correlated with
the occurrence of seizures and eradication of poppy crops by the Mexican Military. Time
variation comes from the price of heroin in the United States. These two variables together
define a municipality time specific shock which leads to differential exposure from the
increased demand in the heroin market.
This paper provides four sets of results. First, I show that the increase in demand
for heroin encouraged Drug Trafficking Organization to expand their operations into high-
suitable municipalities. Estimates imply that when the price of heroin doubles a munici-
pality in the 75 percentile of suitability will have an additional 5% probability of having
more than one active cartels each year after the OxyContin reformulation compared to a
municipality in the 25 percentile of suitability.13 The result for the number of cartels is
similar with the entry of 0.22 more cartels in a municipality in the 75 percentile of suitabil-
ity compared with one in the 25 percentile. These results are in line with the intuition that
Drug Trafficking Organizations adapt and react to external market pressures. In particu-
lar, the cartels will enter new valuable markets, or expand production with the increase in
8
A full explanation of the algorithms and the techniques used can be found in Section 3 and Appendix A.2.
9
These data set covers cartel presence between 1990 and 2010, the one built here is from 1990 to 2016.
10
These two data sets can be find here: Sánchez Valdés (2015, 2017)
11
Afghanistan is the world largest producer of opium and the United Nations have data on yields and
hectares cultivated since 1990. A detailed explanation of the particular elastic net used and how it was
chosen can be found in the Appendix A.2
12
The suitability index construct here is similar to the ones use in Bounadi (2018) and Gehring et al. (2018).
13
Assuming that all the other characteristics of the municipalities are the same.

4
demand. These results are robust to different sets of controls and fixed effects.
Second, I find that the entry of a second, third, fourth, and fifth cartel increases the
homicide rate per 100,000 inhabitants. For example, the entry of a second cartel in a
municipality increases the murder rate by 34.3%, the pre-entry average is 15.38 homicides.
The highest increase in the homicide rate is when the fifth cartel enters a municipality
with a 94.8% increase in the homicide rate after the entry. The pre-entry homicide rate
per 100,000 inhabitants is 17.31. There does not seem to be any effect after the seventh
cartel enters a municipality, but there are too few municipalities with more than six cartels
to say this with confidence. The entry of the first cartel in a municipality does not increase
violence. The exit of cartels from a location has a significant effect when a municipality
goes from two cartels to one, this exit decreases the murder rate by 34%. These results
suggests that it is not the presence of illegal activities that generates violence but the
presence of more than one criminal organizations fighting for scare resources.
Third, I ask how the demand shock affected different socio-economic outcomes in the
exposed municipalities. I find that the demand shock leads to a decrease in population
and average years of education. The percentage of households with women as a head
increased after the shock. These results suggest that there was outmigration from the
exposed municipalities of the more educated and wealthier members. I also show that
the percentage of households without dirt floor and the percentage of households with
basic service: electricity, water and, sewage increases. These two results suggest that there
is a cash flow from the opium market that is probably being capitalized by the poorest
households through investments in their homes.
Lastly, I observe that the military is eradicating more opium and less marijuana in
these municipalities, consistent with the value of opium going up as the relative value of
marijuana decreases.
Finally, I identify which cartels expand to new markets and the fragmentation of existing
cartels. Before 2010 the production of heroin was highly concentrated. Only the Sinaloa
Cartel was present in municipalities where opium poppy was eradicated by the military. By
2016 nine organizations are present in opium producing municipalities. I am able to identify
two expanding cartels, Sinaloa and Los Zetas, two newly created cartels that immediately
entered the heroin market, Jalisco New Generation14 and the Knights Templar15 and one
contracting cartel, La Familia Michoacana. The two newly generated cartels splintered from
existing organizations and immediately started fighting, with their parent organizations
for highly suitable land to cultivate opium. Suggesting that market pressures could have
influenced cartel fragmentation. These results are consistent with information from the
Mexican Prosecutors Office and the DEA.
These results provide novel evidence that external policies that shift demand or supply
of illegal drugs have direct effects on criminal organization’s activity and subsequent effect

14
Jalisco New Generation Cartel splinter from the Sinaloa Cartel.
15
The Knights Templar splinter from La Familia Michoacana.

5
on violence and other outcomes. This paper speaks to four strands of literatures. First,
it complements the literature studying the recent increase in violence in Mexico. Most of
this literature examines how law enforcement strategies and political alliances increased
violence (Rios 2013; Dell 2015; Osorio 2015; Phillips 2015; Atuesta and Ponce 2017). This
study provides evidence of an alternative channel, an increase in heroin demand, that might
explain some of the increasing violence in Mexico. This paper is the first to try to casually
isolate the effect of the OxyContin reformulation on drug cartel activity in Mexico.
This paper confirms the predictions of models where increasing value of territory leads
to turf wars (Mesquita 2018; Castillo and Kronick 2019). It also confirms the result from
Biderman et al. (2018) that estimates that a single gang having monopoly power in Brazil
decreases violence.
These results also relate with the literature that study the effects of external demand
and supply shocks to illegal markets. Millán-Quijano (2019) and Mejía and Restrepo
(2013) use cocaine prices and link them to increased violence in Colombia, while Gehring
et al. (2018) use heroin prices and their relationship to conflict in Afghanistan. My paper
confirms the results from these literature in the Mexican context and ads the reaction on
cartel competition. Furthermore, contributes to the literature on illegal markets and its
effects on different outcomes (Dube et al. 2016; Sviatschi 2018; Dell et al. Forthcoming),
by showing how the number of competing cartels affect socioeconomic outcomes across
municipalities.
Finally this paper contributes to the increasing social science literature that uses text as
data to measure otherwise hard to quantify phenomena. To the best of my knowledge this
is the first paper in economics that uses a semi-supervised convolutional neural network.
Previous papers using text as data, for example those surveyed by Gentzkow et al. (2017),
use methods based on counting the number of times a word appears. The algorithm used
here goes further by using the meaning of the word and the relative position of it on the
sentence. The techniques presented here have the ability to help measure other difficult to
quantify variables.
The rest of the paper is organized as follows. In the next section, I provide the institu-
tional background for my analysis. In section 3, I present the data, in Section 4, I provide
a theoretical model and the main econometric specification. Sections 5 and 6 present the
results and section 7 concludes the paper.

2 Background
2.1 The Opioid Crises
From 1996 to 2010, consumption of legal opioids increased in the United States, leading
to the opioid epidemic. One of the opioids that was particularly abuse was OxyContin.
OxyContin is a narcotic analgesic that due to its time-released formula contained more
milligrams of its active ingredient than other similar opioids. Opioid abusers would crush

6
the tablets to snort or inject them, undoing the time-release mechanism so that all the active
ingredient is absorbed immediately by the body. In 2010 physician organizations pressured
the government to make it harder to accumulate pills and to abuse them. As a result, the
federal and state governments started to crack down on pill mills, and Purdue Pharma
introduce an abused deterrent version of OxyContin. This new pill was harder to crush or
dissolve, thus deterring the most-dangerous methods of abuse by injection or inhalation.
The restriction in the legal supply drove some users to switch to its illegal counterpart,
heroin. Abby et al. (2018) use cross sectional variation in OxyContin availability across
states and find a relationship between the reformulation of OxyContin and the increase
in heroin deaths. It has been estimated that about 80% of people who use heroin first
misused some prescription opioid (Muhuri et al. 2013).

2.2 Mexican Cartels and the Heroin Market

Historically, Mexico has produced some opium and exported it to the United States
since the beginning of the twentieth century.16 Heroin is an illegal highly addictive drug
processed from opium. Mexican cartels exported some heroin but it was never their main
activity.The heroin of Mexican origin seized by the DEA went from 22% before 2010,
to 51% by 2011 and reached around 90% by 2016 (DEA 2018). Until 2006 there was
just one cartel, the Sinaloa Cartel; in areas where opium poppy was eradicated by the
Mexican government. By 2016 the nine main organizations can be found in municipalities
where opium poppy was eradicated by the military. This paper considers nine main Drug
Trafficking Organizations recognized by the Mexican prosecutor’s office and the DEA.17

3 Data
The main goal of this paper is to document how Mexican Drug Trafficking Organizations
reacted to the opioid crisis and disentangle the relationship between market structure and
violence in illegal markets. To achieve these two goals ideally I would like to observe when
each cartel enters a municipality, raw opium production by cartel, and violence related to
drug trafficking. Due to the illegality of the market these data either do not exist or are
reported in aggregate geographical levels and across sparse periods of time.18 To create a
data set for cartel presence and opium suitability, I use web scraping and machine learning.
The data sets and how they were built are described below.

16
A brief history of the drug trade in Mexico and the war on drugs can be found in Appendix A.1.1 and
Appendix A.1.2.
17
A description of these nine cartels can be found in Appendix A.1.3.
18
The Mexican prosecutor’s office, military, and federal police have their own data on presence, unfortu-
nately they are only available at the state level and they are not published every year. The same is true
for the DEA datasets.

7
3.1 Cartel Presence
In order to measure cartel presence I need to construct a novel data set. These data set
tracks for every Mexican municipality in each year in the 1990-2016 period whether or not
each of the nine major Drug Trafficking Organizations were operating in that municipality.
The algorithm I used to create this data set is the following one.
First, I use a web crawler19 to scrape Google News Mexico. Google News is a news
aggregator that watches more than 50,000 news sources worldwide. Google does not provide
the number of sites it tracks for the Mexican version but through the scrape I was able
to identified 770 local, 33 national, and 83 international media outlets that report in
Spanish. This web crawler looked for any articles between 1990 and 2016 that contained a
municipality-cartel pair.
The web crawler collected every article whose main body20 mentions: i) a Mexican
municipality and ii) the name of one of the nine major Drug Trafficking Organizations
in Mexico. The number of articles found by the crawler are 1,201,483. I used a sentence
extractor to keep the sentences from this articles that included a municipality-cartel pair.21
If the article contained a year I assign the event to that year, if no year was assigned I use
the publication year. The number of sentences that I analyzed are 2,802,224.
Next, I manually classified 5, 000 sentences as either presence or not to train a semi-
supervised CNN.22 I use 80% of this sample as the training set and 20% as the test set.
A CNN23 is a deep learning set of algorithms usually used for image classification but
had proven effective in text classification (Kim 2014; Young et al. 2017). The CNN works
as follows. Sentences are first broken into words, then transformed into a word embed-
ding matrix.24 Then several filters are applied that constitute of different word window
sizes that go over each sentence. This is followed by a discretization operations that re-
duce dimensionality of the output. This produces the final sentence representation that is
classified. The particular CNN used here has an out of sample classification accuracy of
0.86.25
In order to validate this data set for cartel presence, I use two data sets that also use
news articles to measure cartel presence and DEA aggregated data by state.

19
A bot that systematically browses the world wide web.
20
Some media web pages include links to other articles in the side of the main content. If this side content
includes the name of a cartel or the municipality sometimes the search engine will return this as a hit.
Appendix A.2 includes examples of these type of media pages.
21
In further iterations of this project I will use the whole article as the input and not just the sentences.
22
Examples of this sentences can be found in Appendix A.2
23
I also use a simple bag of words with logistic classifier the comparison between these two methods can be
found in Appendix A.2
24
A word embedding is a learned representation of text where words that have the same meaning have
similar representation.
25
A more technical explanation of the algorithm can be found in Appendix A.2

8
3.1.1 Validation with other data sets created from news articles:
These two data sets are: Coscia and Rios (2017) time series and data collected by Pro-
fessor Victor Manuel Sanchez from the Autonomous University of Coahuila, Mexico. The
correlation between these two data sets and mine is positive and statistically significant.
Coscia and Rios (2017) tracked the presence of the same nine DTOs at the municipal level
between 1990 and 2010. They use a web crawler and then code a cartel as being present
if the frequency of the hits for a particular municipality-cartel exceeds certain threshold.
The correlation between their data and the one constructed here between 1990 and 2010
is 0.38.26 The fact that this correlation is not larger can be explained by the change in
the Google algorithm between 2012 and 2019. The low correlation can also be attributed
to the difference in validation techniques, relative number of hits versus natural language
processing.
The second data set was generated by Professor Sanchez: he manually recorded data
from local and national newspapers for the state of Michoacán between 2011 and 2013 and
for the Mexico City Metropolitan Area between 2014 and 2017. The correlation between
this data set and the one used here is 0.71.

3.1.2 External validation:

The DEA has published a biannual map of cartel presence across Mexican states, since
2009. In 2009, they published a complete map of cartel presence which turns to be highly
correlated with my data set at 0.69. For the years 2011, 2013 and 2015 the DEA just
recorded the presence of dominant cartels and the correlation with my data sets goes down
to 0.35. The DEA documents do not explain what they mean by dominant presence, and
the algorithm used here does not aim to classify dominance only presence.
The high correlation between my data set and the manually collected one provide
evidence that the algorithm used here has a high performance compared to a human clas-
sification of the articles. Despite the fact that probably the articles manually coded were
different than the ones found by the web crawler. The deep learning techniques use to
generate this data set can be use elsewhere to measure similar hard to observe outcomes.
The high correlation between the DEA data and my data set shows that using news ar-
ticles as a source of illegal activity data might be as good as using data from the authorities.
The advantage of using news is that they are reported at a more disaggregated geograph-
ical level and with higher periodicity. This is particularly important in measuring drug
related criminal activity. These groups continuously change their territorial dominance
and knowing these fluctuations is key to understand them better.
Figure 1 shows the evolution of cartel presence between 2004 and 2016. Figure 1a shows
the mean average number of active cartels between 2004 and 2009 and Figure 1b shows
the mean number of active cartels between 2010 and 2016. The maps show the increase

26
All hints before 2000 come from a Google project that digitized newspapers.

9
in the number of municipalities with a cartel active and also the increase in the number of
municipalities with multiple cartels present.

3.2 Agro-ecological data

The geographic variation in opium poppy suitability is drawn from multiple data sets.
A suitability index measures an area comparative advantage for crop cultivation base on
geographic and climatic characteristics. While these indexes exist for almost all legal
crops, they do not exist for papaver somniferum commonly know as opium poppy. The
only available measure is the FAO overall characteristics for papaver somniferum to sur-
vive.27 Suitability indexes are built using agro-climatic characteristics and crop yields.
Unfortunately, there does not exist historical data on opium production in Mexico.28
I built a suitability index using Afghanistan’s output data. Afghanistan is the main
world opium producer and the UN has been collecting data (through surveys and satellite
images) on opium yields and hectares cultivated since the early 1990s.29 I use yields by
district between 2000 and 2018 plus 45 agro-climatic characteristics and all their interac-
tions.30 The main agro-climatic variables used are temperature, precipitation, elevation,
terrain ruggedness, soil quality, and river density. A full description of these variables and
the data sets they came from can be found in Appendix A.3.
The suitability index is build by regressing log productivity of opium on the geo-climatic
characteristics by district and year. The number of total regressors is 903. I use an
elastic net, a penalized OLS regression, Zou and Hastie (2005) to reduce the number of
regressors. A 10 fold cross-validation is use to validate the model.31 This model is then
used to predict log productivity of opium in Mexican municipalities given its agro-climatic
characteristics.32 This measure is then standardized and its range set between 0 and 1 for
interpretation purposes, so that 1 means perfectly suitable and 0 means not suitable to
grow opium poppy.
Figure 2 shows the suitability index for each municipality. The darker the tone shown
the more suitable the area is to grow opium poppy. This map together with the maps in
Figure 1 imply that cartel presence is related to opium suitability particularly after 2010.
Mostly all the states in the Pacific coast; Sinaloa, Nayarit, Jalisco, Colima, Michoacán, and
Guerrero, are highly suitable to grow opium and the number of active cartels increased there

27
http://ecocrop.fao.org/ecocrop/srv/en/dataSheet?id=8296
28
UNDOC just started collecting these data in 2014 but they are still unable to provide accurate measures
for opium yields at the municipality level.
29
https://www.unodc.org/unodc/en/crop-monitoring/index.html?tag=Afghanistan
30
The suitability index built here is similar to the ones used by Kienberger et al. (2017), Sonin et al. (2019),
Bounadi (2018) and Gehring et al. (2018).
31
k-fold cross validation generally results in a less biased and less optimistic estimate. It consist on splitting
the data into k groups and use this to train and test the model.
32
The shrinking and mixing parameters used can be found in Appendix A.3.

10
(a) Mean Cartel Presence 2004-2009

(b) Mean Cartel Presence 2010-2016

Figure 1: Cartel Presence

11
after 2010. The correlation between the index and opium poppy eradication data from the
Mexican military between 2000 and 2005 is 0.45. The correlation between mean presence
before 2010 and the suitability index 0.18 and the correlation after 2010 is 0.47.
Table 1 shows the relationship between the suitability index and mean hectares of
opium poppy eradicated by the Mexican military. The coefficients show are the result of
regressing the suitability index on the mean opium poppy eradication before and after 2010.
The specification controls for police and military presence and includes municipality fixed
effects. Highly suitable municipalities have eradications which imply that some opium is
being produce there.

Table 1: Suitability Index and Eradication

Post2010 Pre2010 All Years

SuitIndex 401.025∗∗∗ 386.471∗∗∗ 788.409∗∗∗
(137.447) (194.678) (283.476)

Constant −85.514 −75.265 −160.993

(59.360) (69.883) (101.758)

Observations 2,455 2,455 2,455

R2 0.003 0.002 0.003
Adjusted R2 0.003 0.001 0.003
Residual Std. Error 642.414 (df = 2453) 908.534 (df = 2441) 1,322.943 (df = 2441)
F Statistic 8.513∗∗∗ (df = 1; 2453) 3.941∗∗ (df = 1; 2441) 7.735∗∗∗ (df = 1; 2441)

Notes: This table presents the results of a ordinary least square regression where the dependent
variable is the mean eradication of opium poppy, measured in hectares. Controls for police, military
presence and includes municipality fixed effects.∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

3.3 Homicide and Demographic Data

There does not exist a full panel of homicides related to organized crime. I use the total
number of homicides by municipality and year from INEGI. This data set includes all the
deaths in the country for which a death certificate was generated, then they classified the
cause of death and one of the classifications is intentional deaths.
Data on population, average years of education, percentage of indigenous population,
illiterate population, unemployment, basic characteristics of the households that includes
houses with dirt floors, percentage of houses without electricity, sewage or running water
and percentage of houses with TV, refrigerator, washer machine, phone, Internet and car

12
 Figure 2: Opium Poppy Suitability Index

come from INEGI 2000, 2005, 2010, and 2015 censuses and intercensal surveys. I also use
the CONEVAL data set that reports an index of how marginalize a municipality is and
data on the affiliated political parties of mayors and governors are drawn from CIDAC and
INAFED.

3.4 Drugs Time Series

The Mexican National Defense Office publishes data each year on how many hectares
of opium and marijuana are eradicated by municipality and the amount of drugs seized by
the military and marines. The data on heroin prices come from the UNDOC data set; data
on heroin overdose deaths from the national CDC wonder data set, and data on heroin
users each year from the National Survey on Drugs and Health.

13
4 The effect of external demand shocks on cartel presence
In this section I examine the casual effect of a plausible external demand shock to the
heroin market on the presence of Drug Trafficking Organizations across Mexico. First, I
describe the theoretical framework for the mechanism behind the relationship between the
increase in the value of certain territories, entry, and violence. Second, I introduce the main
econometric specification and present the effect of the shock on the probability of having
more than one cartel active and in the number of active Drug Trafficking Organizations
per municipality. Last, I address possible threats to my identification strategy.

4.1 Theoretical Framework

In this section I provide a theoretical model that relates cartel entry, investment in mili-
tary capacity, and violence. The mechanism behind the analysis is the following: a positive
demand shock to the heroin market increases the value of controlling drug production and
drug trafficking routes. In the absence of conditions to reach peaceful agreements to share
the market, drug traffickers will fight each other to gain as much market shares as possible.
The following model is a two-stage entry game, a variant of the endogenous sunk cost
models in (Schmalensee 1992), with the addition of a success function that determines
the proportion of the total shares of rents that each organization will win. To simplify
the model, I assume that each municipality is an independent production site. I also
assume that the government actions are completely known by the cartels and that they
have already internalized any possible seizures or extra costs that government interventions
might generate.33 Finally, the only way a Drug Trafficking Organization can gain market
power is through investing in military capacity.
There are N potential cartels that want to enter a production site, each of them invest
Mi in military capacity, and the total military presence in the municipality will be M =
PN ∗ ∗
i=1 Mi , where N is the number of cartels that entered the municipality. The potential
revenue of site j is given by Rj , this potential will depend on how suitable municipality j
is.
In the first stage, cartels decide whether to enter or not a site and pay fixed cost F.
Cartels need to make sure the site is productive and also know who else is active in that
territory. In the second stage, they decide how much to invest in military capacity.
The following success function will determine the market share each cartel will have as
a function of its own investment and the investment of all the other competitors:

Mη
si = PN ∗ i η
j=1 Mj
In the data I observe multiple Drug Trafficking Organizations in one site, so this success
function can be interpreted as the proportion of the total revenue that each cartel gets from
33
For a model on how the government interdictions affect violence see Castillo and Kronick (2019).

14
its investment. η describes the returns to military efforts, if η ≤ 1 there are decreasing
returns to effort if η > 1 the returns are increasing.
The profit for a cartel entering municipality j is given by:

πi = Rj si − Mi − F
where Rj is the exogenous total revenue from the production site j.
Violence will be a function of the total military investment of all the cartels V ( N
P
i=1 Mi ).
The only assumption about this function is that is zero if military capacity is zero and that
it increases with total expenditure.
As long as η ≤ 2 there will be well-behaved symmetric Nash equilibria in Mi with
non-negative profits.34
The Cournot symmetric NE in military expenditure in site j will be:

N −1
M = Rj η
N2
and the free entry condition is :

F
N ∗2 + N ∗ (η − 1) − η = 0
Rj
The parameter η will determine the relationship between the total revenue from a site
and the number of cartels that enter:

• If η < 1 then when Rj → ∞ so does N → ∞. There are decreasing returns to

military effort.
R
1/2
• If η = 1 then N ∗ = Fj In this case the military expenditure per cartel needs to
grow as the market grows in order to keep the free entry condition active.
η
• If η ∈ (1, 2] then when Rj → ∞ the number of cartels is bounded by N ∗∗ = η−1 here
competition is tough and military expenditure increases more than in the previous
cases.
These three results are summarized in Proposition 1:

Proposition 1. Suppose that the return to military efforts is η ≤ 2. Then a positive shock
to the total revenue from a production site, Rj , will have the following effects:
1. Total military expenditure M ∗ in the municipality will increase.
2. The number of active cartels N ∗ will also increase.
3. Violence will increase with them.
34
All the calculations can be found in Appendix A.4.

15
4.2 Empirical Strategy
4.2.1 Baseline Econometric Specification
To estimate the causal effect of competition between Drug Trafficking Organizations
on violence. I ideally need a policy change or external shock that shifts demand or supply
but that does not directly affect violence. This disqualifies any shift that might happen
through law enforcement, as this kind of shock will have two different effects on violence: a
direct one from the change in law enforcement and an indirect one from market pressures.
Throughout this paper I use the reformulation of OxyContin in 2010, which was followed
by an increase in heroin overdose deaths and a spike in heroin prices in the United States.
This seems like a plausible exogenous shock to the heroin market that might affected
entry decisions into new territories of Drug Trafficking Organizations. I exploit within-
municipality variation from combining the suitability index and heroin prices in the United
States. The main specification is an event study analysis, where the relevant event is the
OxyContin reformulation in 2010.
2016
X
Ymt = βt Suitm ∗ P riceHert + ψXmt + αm + γt + σs t + mt (1)
t=2004

Suitm is a measure of opium poppy suitability of municipality m, this measure is

between 0 and 1, where 1 means perfectly suitable and 0 not suitable. P riceHert is the
retail price of a milligram of heroin in the US adjust by purity and inflation and normalized
to 2002 dollars. The Ymt are different measures of cartel presence: the probability of having
more than one cartel present and the number of active cartels. The αm are municipality
fixed effects, γt year fixed effects and σs t state specific time trends. Fixed effects controls
for all individual municipality characteristics that are fixed over time. The controls include
police presence, military presence, a dummy if the mayor is from the conservative party
PAN and a dummy if both the state governor and the mayor are from the conservative
party.35 The only difference between this specification and an standard event study is that
here the treatment is a continuous variable.36
The main results are presented by plotting the coefficients βt to show the evolution of
the outcome variables relative to the reformulation in 2010. The year 2009 was normal-
ized to zero and the plots show two different coefficients: the simple post average from
the event study coefficients and the difference-in-difference coefficient. These difference-
in-difference results are estimated using the same specification as equation (1), but with
the event dummy years replace by P ostRefmt = 1[t >= 2010]. These dummies then are
interacted with the suitability and heroin prices. Standard errors are clustered using Con-

35
Dell (2015) finds that drug related violence increases after close elections of PAN mayors. There is also
evidence that PAN mayors and governors cooperate more easily with the federal government on the efforts
against drug dealers.
36
Strategy commonly use to estimate the effect of commodity shocks (Dube and Vargas 2013).

16
ley (1999) with a radius of 500 Km. This is used to address the fact that there is spatial
correlation in the suitability index, and clustering by municipality might be a too small
geographic variation. The results using standard errors clustered at the municipality level
can be found in Appendix A.6.1.

4.2.2 Results
The section below shows the cartel presence reaction to the 2010 shock. First, I show
how the probability of having more than one cartel active changes after the shock and then
how the number of cartels changed after the shock. The results are robust to including the
set of covariates described in the section above; all the fixed effects and also the baseline
characteristics interacted with fixed effects.

Cartel Activity: Figure 3a shows the increase in the number of municipalities with more
than one cartel active after 2010. The maps from Figure 1 and Figure 2 show a relationship
between the suitability index and increase cartel presence, specially after 2010. To test the
relationship between suitability and the increase in cartel presence I used specification (1).
Figure 3a shows the event study coefficients when the dependent variable is one if the
municipality has more than one cartel and zero otherwise. The graph shows that the pre-
shock probability of having more than one cartel present is 0.013. The event study post
event average is 0.079 and the difference-in-difference coefficient is 0.108. To understand
the magnitude of these coefficients consider the effect when the heroin price doubles in the
United States. The probability of having more than one cartel present will be 5% higher
in a municipality in the 75 percentile of suitability compared with one in the 25 percentile.
Figure 3b shows the coefficients from specification (1), where Ymt is the number of
active cartels in municipality m at year t. As in the case for the probability of having
more than one cartel present, these coefficients are close to zero before the shock and
start increasing after 2010. The average post-shock event study coefficient is 0.310, the
difference-in-difference coefficient is 0.466, and the pre-shock average number of cartels is
0.05 cartels. To understand the magnitude of these coefficients, lets compare a municipality
in the 75 percentile with one in the 25 percentile when the heroin price doubles. The number
of active cartels will be 0.22 more cartels each year after the shock in the municipality in
the 75 percentile compared with one in the 25 one.
Robustness Checks: Table 2 presents a series of robustness checks. The first column
uses the specification in equation includes municipality and year fixed effects. The second
column includes the same set of fixed effects plus a set of controls that are police presence,
military presence, a dummy if the mayor is from the conservative party PAN and a dummy
if both the mayor and the governor are from the conservative party. These controls for
potential biases coming from the war on drugs acting differently across municipalities with
different political alliances or municipalities with a higher presence of authorities. Columns

17
 Post(Average): 0.079 Post(Average): 0.310
0.3
Post(DD): 0.108*** Post(DD): 0.466***
Pre−shock mean,dep.var.: .013 Pre−shock mean,dep.var.: 0.057
0.8
More than One Active Cartel

0.2

Number of Cartels
0.1 0.4

0.0
0.0

−0.1
2004 2006 2008 2010 2012 2014 2016 2004 2006 2008 2010 2012 2014 2016
Year Year

(a) More than one cartel active (b) Number of active cartels

Figure 3: Event Study on cartel presence

3 and 4 add a set of baseline time trends. These baseline characteristics are population,
years of education, poverty index, hectares of drugs cultivated and kilos of drugs-seized all
of these variables are from 2000 and are interacted with year fixed effects. Column 4 just
adds the same set of controls as Column 2 to the set of baseline controls. The estimates
are robust to all these controls. Finally, Column 5 includes linear time trends by state and
all the covariates from the previous columns. The results for having more than one cartel
active and the number of cartels are robust to all these controls and fixed effects. The
standard errors are clustered using Conley (1999) with a radius of 500 Km.

In sum, patterns of cartel entry are consistent with Proposition 1. The increase in the
value of some production sites encourage more cartel entry. When the prices of the final
product is high more Drug Trafficking Organizations will try to enter the more suitable
production sites. These results are also consistent with anecdotal evidence of increased
cartel presence in municipalities were cartels where not active before and that now pro-
duce most of the opium Ahmed (August 2015). The results above do not show a sharp
jump in 2010, this is consistent with cartels slowly realizing that these sites are now more
valuable. The production process from plating the opium to getting heroin takes at least
four months, the time the crop need to grow. This supports the casual mechanism used
in this paper. An explanation of how opium is transformed into heroin can be found in
Appendix A.3.

18
 Table 2: The effect of the reformulation on cartel activity

Cartel Activity
More than one Number of Cartels
Average effect 0.108∗∗∗ 0.121∗∗∗ 0.064∗∗∗ 0.044∗∗∗ 0.049 ∗∗∗ 0.466∗∗∗ 0.537∗∗∗ 0.292∗∗∗ 0.227∗∗∗ 0.241∗∗∗
post-2010 (0.011) (0.014) (0.012) (0.012) (0.012) (0.002) (0.049) (0.038) (0.039) (0.039)

Observations 32,202 31,752 32,202 31,752 31,752 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 0.012 0.012 0.012 0.012 0.012 0.057 0.057 0.057 0.057 0.057
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Notes: This table presents the results of the difference-in-difference model for the dependent vari-
ables: more than one cartel active and number of cartels. These results are estimated using the the
same specification as Equation 1, but with the event dummy years replace by a post shock dummy
interacted with suitability and heroin prices. The first four columns show the results for the prob-
ability of having more than two cartels active. The next four columns show the coefficients for the
number of cartels. Columns (1) and (6) present the results with municipality and year fixed effects.
Columns (2) and (7) add controls for police presence, military presence, and political party of the
mayor and the governor. Columns (3) and (8) control for baseline characteristics interacted by
year such as population, drug activity in 2000 and marginalization rate. Columns (4) and (9) add
to the baseline trends, the set of controls from Column (2) & (7). Columns (5) and (10) add state
specific time trends. Standard errors are clustered using a 500 Km radius. Significant at ∗ p<0.1;
∗∗
p<0.05; ∗∗∗ p<0.01

19
4.2.3 Threats to Identification
Here I address some potential concerns regarding the main identification strategy. First,
I provide evidence that measuring cartel presence using news articles is not biased towards
any particular cartel. Second, I discuss the validity of the reformulation as an external
demand shock. Finally, I present a series of robustness checks that address that violence
and cartel activity might be completely driven by the war on drugs and not by market
pressures.
The main concern is measurement error in the cartel presence variable. I already showed
in the data section that this data set is highly correlated with other data sets built through
news articles. The data set is also highly correlated with aggregate data from the DEA.
Despite this, it could be that the measure is biased. Mexico is known for being one of the
most dangerous countries in the world to be a journalist. More than 200 media workers
have disappeared or been killed since 2000. Anecdotal evidence from other journalists
shows that some of them had stop reporting cartel activities that are not first reported by
the police or the military. To train the neural network used to validate the data I read
5,000 sentences and classify them as either presence or not. This sub-sample confirms that
most of the journalists are either not reporting names of particular cartel members, just
the bigger organizations, or that they are just reporting using official data from local and
federal police or the military.37 Examples of these sentences can be found in Appendix
A.2. From the 1,201,483 articles analyzed here that contained a cartel-municipality pair in
the core of the article 44% of them do not have a byline. This means that the news outlet
is the author of the article and not a particular journalist, this can also help journalists
to avoid violence from the cartels.The data set that includes the location, date and media
where these journalists either were killed or disappeared is public.38
I use this data set to test if there is any particular bias towards journalist being killed
by a particular cartel. In order to do this I regress the number of killed or disappeared
journalists on each of the nine cartels presence. The coefficients from these regressions can
be found in Appendix A.5. The specifications controls for police and military presence,
political party at the municipality and state level and they also include municipality and
year fixed effects and state specific time trends. Standard errors are clustered at the
municipality level. None of these coefficients are significant and all of them are near zero.
Correlations between the journalist data set and each cartel presence are low. The highest
one is for Los Zetas and is 0.17. These provides evidence that there does not seem to be
any particular bias towards reporting or misreporting on a particular organization. Figure
4 shows the map of where media workers have either disappeared or being killed between
2004 and 2016. There is not a particular area where they are concentrated. The correlation
between this map and the mean number of active cartels during this period is 0.45. This

37
Police reports are not public so they cannot be use to create a cartel presence data set.
38
The data set of missing and killed journalist can be found here https://en.wikipedia.org/wiki/
List_of_journalists_and_media_workers_killed_in_Mexico.

20
confirms that all the journalists where attacked in a municipality with at least one cartel
active but that there does not seem to be a particular cartel perpetuating all the attacks
against the press.

Figure 4: Disappeared or Killed Media Workers between 2004-2016

Despite this, the dataset I built might underestimate actual cartel presence. Journalists
and police reports are made when something actually happens in a municipality related to
a particular organization. Therefor, I am probably just observing when there is a violent
event, an arrest or an interdiction. This implies that when a cartel is active but the police
or the military do not know about it this data set would not observe it either. Then
all the results were the dependent variable is cartel presence will not be biased by this
measurement error. The results when the independent variable is cartel activity will be
biased towards zero. Thus all the results from this paper can be considered as a lower
bound of the effect that cartels have on violence.
The second potential concern with this identification strategy is that Drug Trafficking
Organizations had a direct impact in the price of retail heroin in the United States. Then
the price will not be exogenously changing but will be reacting to supply changes. The
sharp increase in the heroin of Mexican origin seized by the DEA in the United States

21
from 19% in 2009 to 51% in 2011 and 86% by 2016, suggests that the cartels adapted after
the increase in demand in the United States. The Drug Trafficking Organizations reacted
by producing more heroin as the demand increased. Figure 5 shows the increase in heroin
users and heroin prices in 2010. Abby et al. (2018) and Muhuri et al. (2013) have shown
that the reformulation drove up the demand for heroin and the overdose deaths from it.
Then, it is reasonable to assume that the increase in demand and the subsequent increase
in prices attracted more drug traffickers to the heroin market.

Figure 5: Heroin Users and Heroin prices in the United States

560
900000

Heroin Users

Heroin Prices

540
Retail Heroin Price(2016 dollars)
Heroin Users in the US

520
700000

480 500
500000

460
440
300000

2000 2005 2010 2015

Year

Note: Heroin prices are from UNDOC data set and heroin users come from National Survey on Drugs and

Health.

Finally, one last concern is that the increase in violence and the multiplication of
Drug Trafficking Organizations is a direct effect from the war on drugs. Former Mexican
President Felipe Calderón, in office between 2006-2012 from the conservative party PAN,
declared war on drugs in 2006. The main government strategy was to behead criminal
organizations by targeting high rank kingpings. This strategy generated a lot of instability
inside the cartels which led to more violence (Jones 2013; Calderón et al. 2015; Phillips
2015; Signoret 2018). Despite this being true for the overall trend in violence, none of the
strategies of the government will generate entry to suitable municipalities to grow opium
poppy or a sharp shift in the heroin market around 2010. To address any possible concern

22
regarding government enforcement I use several controls that include, police presence,
military bases, garrisons, and ports, if the mayor of the municipality is from the conservative
party PAN, if the mayor and governor are both from the PAN party. I also add baseline
controls for the municipality characteristics interacted with year fixed effects. I also add
state-specific time trends to control for any other government policies that might affect
cartel activity. The results seem to be robust to all of these different specifications.
One last identification concern is marijuana legalization across the United States. Even
though, medicinal marijuana has been legal in some states since the mid-1990s, the legal-
ization of recreational marijuana started in 2012. This does not seem to be a big problem
and in any case both of these external shocks increase the value of opium and drive cartels
to switch to the heroin market and leave the less profitable marijuana market.

5 Entry, Exit and Violence

5.1 Econometric Specification
This section provides evidence of the relationship between cartel entry, exit and the
homicide rate. I use an event study specification that exploits heterogeneity on the time
of entry and exit across municipalities. The relevant event of study is the entry or exit of
the cartels. The relationship between the number of cartels and violence is not obvious.
The presence of illegal activities does not necessarily breeds violence.Snyder and Duran-
Martinez (2009) and Castillo and Kronick (2019) show that the ability to reach peaceful
agreements to share profits between criminal organizations relies on how strong national
institutions are. Mexico during this period of time lacked such conditions so I expect to
see the criminal organizations fighting over potential profits. 39
The main econometric specification used in this section is the following one:

5
βτ 1{τ = t − em } + ψXmt + αm + γt + σe t + mt
X
Vmt = (2)
τ =−5

The time relative to the entry of the cartel is indexed by τ . The variable em denotes
the calendar year in which municipality m experience the entry of cartel number n, so
1{τ = t − em } is an indicator of municipality m in year t having experience the entry of
the cartel number n τ years ago. In the summation τ = −5 (or τ = 5) term includes all
the years greater than or equal to five years before (or five years after) the entry of the
first cartel. This specification normalizes β−1 to zero. I control for a vector of municipality

39
Snyder and Duran-Martinez (2009) and Castillo and Kronick (2019) argue that the government plays a
key element on how criminal organizations react to increasing profits. The none linearities between the
number of active cartels and the homicide rate are probably related to government action. Incorporating
these effects beyond political party and military and police presence are left for further research.

23
fixed effects αm , calendar year fixed effects γt , the same controls as before Xmt and linear
state time trends σe t.
The results are presented by plotting the βτ coefficients, to show the within munici-
pality evolution of the homicide rate relative to the event of the entry or exit of the nth
cartel. The graphs also show difference-in-difference results which are estimated with the
same specification as equation (2) but with the event time year dummies replaced by a
dummy variable P ostEntrymt (indicating that in year t municipality m has the entry of
the nth cartel). The standard errors are clustered using Conley (1999) error with a radius
of 500Km.

5.2 Results
This section shows three main results. First, the presence of a single cartel does not
increases violence. Second, there seems to be a non linear relationship between the number
of cartels and the increase in the homicide rate. Last, the only significant effect from the
exit is when a municipality goes from two to one cartel.
Figure 6a presents the coefficients from equation (2) where the relevant event is the
entry of the first cartel to a municipality. There does not seem to be a significant effect
from this entry. This result is consistent with Biderman et al. (2018), that find a decrease
in violence in São Paulo (Brazil) when a single gang gains monopoly power. This result
suggests that is not criminal activity alone that generates violence but is the interaction
between different actors that increases violence. It might indicate that as long as a single
group can maintain monopoly power in a location there should not be an increase in
violence regardless of the illegality of their activity. This result is robust when adding
controls for police and military presence, political party affiliations, municipality baseline
characteristics interacted with time trends, and state specific time trends.
The second result from this analysis is the increase in the homicide rate from cartel
entry. Figure 6b shows that when a municipality goes from one to two cartels the homicide
rate per 100,000 inhabitants increases by 7 homicides. The increase peaks at the year the
second cartel enters and stays above pre-entry levels for the next five years. The entry of
the third and fourth cartel increase the homicide rate by 13 homicides each one.
From Figure 7a there is an sharp increase the year the third cartel enters and the num-
ber of homicides remain high for the next five years. Figure 7b shows the effect from the
entry of the fourth cartel. This graph shows a pre-trend before the event, a peak a year
after the fourth cartel enters, and a slightly decline after the second year. The entry of the
fifth cartel increases the homicide rate by 17 homicides. This is the biggest increase and
Figure 8a shows that the biggest increase in the homicide rate occurs the year after the
40
All the coefficients graphs show through the paper have the event study coefficients for the basic speci-
fication with municipality and year fixed effects. The difference-in-difference results for all the different
specifications can be found in the Appendix. Standard errors are clustered using a 500Km radius.

24
 15 Post(Average):1.57 15 Post(Average):6.64
Post(DD):1.64 Post(DD): 7.08***
Pre−shock mean,dep.var.: 15.43 Pre−shock mean,dep.var.: 15.38
10 10
Homicides

Homicides
5 5

0 0

−5 −5

−10 −10
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5
Years Since Entry Years Since Entry

(a) First Cartel Entry (b) Second Cartel Entry

Figure 6: Cartel Entry40

fifth cartel enters and not simultaneously with the entry. There is not as sharp an increase
as there is for the entry of the second and third cartel. The coefficients are increasing
and then get noisier. The entry of the sixth cartel increases the homicide rate by 10 extra
homicides. There is not significant effect after the sixth cartel, but the number of munic-
ipalities with more than 6 cartels is only 14. These sets of results are robust to including
controls for police presence, military presence, political party, baseline trends and state
specific linear time trends. The tables that show the difference-in-difference coefficients
from specification (2) can be found in the Appendix A.6.2.

The last result is related to the exit of cartels from a municipality. There does not
seems to be any significant effect when a cartel leaves a municipality. The only clear effect
is when a municipality goes from two to one cartels. Figure 9 shows the coefficients from
specification (2) when the relevant event is a municipality with two cartels that experiences
the exit of one of them. This graph shows a sharp drop of the homicides after the exit
of the cartel and the homicides keep going down for a couple of years. The homicide rate
per 100,000 inhabitants decreases by 6 homicides after the exit. This result is robust to
including the different controls and fixed effects. The three results presented in this section
are consistent with the theoretical model from section 4. The model predicts that when
there exists a single criminal organization, violence should be zero and that violence is an
increasing function of the number of cartels. The first and last results from this section
suggest that when a single organization has monopoly power over a territory violence
should be low. It is not the presence of drug traffickers or their illegal activities that
generate violence but the competition in which these organizations engage to win market
power that generates the increase in violence.

25
 30 Post(Average):9.20 30
Post(DD):13.46***
Pre−shock mean,dep.var.: 16.58
20 20

10 10
Homicides

Homicides
0 0

−10 −10
Post(Average):11.92
Post(DD):13.71**
−20 −20 Pre−shock mean,dep.var.: 19.36
−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5
Years Since Entry Years Since Entry

(a) Third Cartel Entry (b) Fourth Cartel Entry

Figure 7: Cartel Entry

60 Post(Average):14.49 60 Post(Average):14.51
Post(DD):17.53*** Post(DD):10.81*
Pre−shock mean,dep.var.: 17.52 Pre−shock mean,dep.var.:18.90

40 40
Homicides

Homicides

20 20

0 0

−20 −20

−5 −4 −3 −2 −1 0 1 2 3 4 5 −5 −4 −3 −2 −1 0 1 2 3 4 5
Years Since Entry Years Since Entry

(a) Fifth Cartel Entry (b) Sixth Cartel Entry

Figure 8: Cartel Entry

26
 Figure 9: Effect of one cartel exit in a municipality with two previously present cartels
Post(Average):−6.80
Post(DD):−6.07**
Pre−shock mean,dep.var.: 17.75
5

Homicides

−5

−10

−15
−5 −4 −3 −2 −1 0 1 2 3 4 5
Years Since Exit

Figure 10 shows the difference-in-difference coefficients from equation (2) but with the
event time year dummies replace by a dummy variable P ostEventmt using just municipality
and calendar year fixed effects. The coefficients and error bars in red, show the effects of the
entry of the first, second and so on up to the ninth cartel entry. From this graph there is a
non-linear relationship between the entry of the cartels and the homicide rate per 100,000
inhabitants. The standard errors of these estimates are bigger as the number of active
cartels increases in a municipality. This is because there are not as many municipalities
with more than five cartels active in their territory. The coefficients and error bars in blue
account for the exit of cartels. The coefficient in zero measures the effect in the homicide
rate in municipalities that go from having one to zero cartels. The coefficient in -1, is the
effect of a cartel leaving a municipality that previously had two cartels. The other three
coefficients are the effects from a municipality that experience the exit of one cartel and
previously had three of them, -2. The effect of a municipality that goes from four to three
cartels, -3, and the first bar shows the coefficient and standard errors from municipalities
that had five cartels and experience the exit of one of them. This figure summarizes the
relationship between cartel entry, exit and the homicide rate. The significant effects occur
when a municipality goes from one to two or two to one. This confirms that when there is
a single criminal organization in one place homicides should not increase. The rest of the
graph shows that there are not significant effects of exit if there are at least two cartels left
and that violence appears to be increasing in the number of cartels.

27
 Figure 10: Entry and Exit of Cartels

200

100
Homicides

−100

−200
−4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9
Effect of Cartel Entry (+) or Exit(−)

5.3 Effect of Demand Shock on Homicides, Migration, and Household Char-

acteristics
In this section I analyze the effect that the external demand shock had on other out-
comes. I use the first specification (1) where the relevant event happens in 2010. First, I
show that there was an increase in the homicide rate after 2010. Second, I ask how demo-
graphics such as population, average years of education and the percentage of households
with dirt floor change after the shock. Finally, I see how the government data on eradica-
tion confirms that after 2010, the suitable municipalities were producing more opium than
before. These results show that the shock had a direct impact on other outcomes across
Mexican municipalities.
Homicide Rate per 100,000 inhabitants: Figure 11 shows the coefficients from the event
study (1) where Ymt is the homicide rate per a 100,000 inhabitants. There is an increase
in the number of homicides after 2010 and it stays high after the shock. The difference-
in-difference coefficient is an increase of 24 homicides per 100,000 inhabitants. A high-
suitable municipality will have a homicide rate with 24 more homicides than a non-suitable
municipality. Table 3 shows the results for the different specifications. The results are

28
robust to all the different controls and fixed effects. The model that includes all the
controls and state specific time trends has an increase of homicides per 100,000 of 11 extra
homicides after 2010.
Figure 11: Homicides per 100,000 inhabitants
40

20
Homicides

−20

−40 Post(Average):14.49
Post(DD):24.53***
Pre−shock mean,dep.var.:12.06
2004 2006 2008 2010 2012 2014 2016
Year

Change in Demographics: This section analyses the effect of the shock on different
demographic outcomes. First, I estimate the effect of the shock on the population and
households composition. The variables used are log of the population, mean years of ed-
ucation, and percentage of households with women as the head of the household.41 The
second set of variables suggest that there was probably an income boost in the munici-
palities exposed to the shock. These variables are the percentage of households without
dirt floor and percentage of households with access to basic services (water, electricity and
sewer).42 Figure 12a shows an steady decrease in the log of the population after 2010 and
Figure 12b has a similar steady decrease in the average years of education. Figure 13,
shows an increasing share of households with women as the head of the household. All
these results are consistent with out migration of the more educated and wealthier members
of these communities. The tables that show all the difference-in-difference coefficients for
the different specifications can be found in Appendix A.6.3. To understand the magnitude
of these coefficients I will compared a municipality in the 75 percentile with one in the 25
percentile of suitability, when heroin prices double in the United States. A municipality
41
The other variables analyzed are percentage of indigenous population, percentage of people with social
security, adult illiteracy, and the mean number of occupants per dwelling. These four variables do not
show significant changes during the time of the analysis.
42
The other variables analyzed were percentage of households that own a TV, refrigerator, washer, car,
phone, and/or a computer. These variables have an increasing trend during the whole sample or are
completely flat during the period analyzed.

29
 Table 3: Homicides per 100,000

Homicides per 100,000 inhabitants

Post-shock 24.53∗∗∗ 20.60∗∗∗ 22.19∗∗∗ 15.61∗∗∗ 11.13∗∗∗
(8.72) (2.87) (2.94) (2.84) (2.54)
Observations 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 12.06 12.06 12.06 12.06 12.06
Municipalities FE 3 3 3 3 3
Time FE 3 3 3 3 3
Covariates 3 3 3
Baseline trends 3 3 3
State trends 3

Notes: This table presents the results of the difference-in-difference model for the dependent vari-
ables: homicide per 100,000 inhabitants. This results are estimated using the the same specification
as Equation 1, but with the event dummy years replace by a post shock dummy interacted with suit-
ability and heroin prices. Column (1) presents the results with municipality and year fixed effects.
Columns (2) adds controls for police presence, military presence, and political party of the mayor
and the governor. Columns (3) controls for baseline characteristics interacted by year such as pop-
ulation, drug activity in 2000 and marginalization rate. Columns (4) adds to the baseline trends,
the set of controls from Column (2) & (7). Columns (5) adds state specific time trends. Standard
errors are clustered using a 500 Km radius. Significant at ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

30
 0.0

0.00

Mean Years of Education

Log Population

−0.5
−0.03

−1.0
−0.06

Post(Average): −0.042 Post(Average):0.456

Post(DD):−0.058*** Post(DD):0.418***
Pre−shock mean,dep.var.:9.37 −1.5 Pre−shock mean,dep.var.:6.24
−0.09
2004 2006 2008 2010 2012 2014 2016 2004 2006 2008 2010 2012 2014 2016
Year Year

(a) Log of Population (b) Mean Years of Education

Figure 12: Demographic outcomes

Post(Average):0.019
0.04 Post(DD):0.015***
Pre−shock mean,dep.var.:0.214
% HH with Women as a Head

0.03

0.02

0.01

0.00

2004 2006 2008 2010 2012 2014 2016

Year

(a) Percentage of households with women as a head

Figure 13: Demographic outcomes

in the 75 percentile of suitability will have a 2.7% decrease in the log of population and
a 19.58% decrease in the mean years of education compare to a municipality in the 25
percentile of suitability. The percentage of households with a women as a head will be
0.7% higher after the shock in the 75 municipality compare to one in the 15 percentile one.
Economic Outcomes: The next set of results suggest that the shock might have had a
positive impact in some socio-economic outcomes. Figure 14 shows the coefficients from
the event study specification (1). From Figure 14a there was an increase in the percentage
of houses that had floors made of materials different from dirt. Figure 14b shows a steady

31
 Post(Average):0.039
Post(DD):0.025***
Pre−shock mean,dep.var.:0.93
0.04
0.10

% HH with Basic Services

% HH with Non Dirt Floor

0.00
0.05

−0.04

Post(Average):0.023 0.00

Post(DD):0.073***
Pre−shock mean,dep.var.:0.54
2004 2006 2008 2010 2012 2014 2016 2004 2006 2008 2010 2012 2014 2016
Year Year

(a) Households with Non Dirt Floor (b) Households with Basic Services
(water/sewer/electricity)

Figure 14: Economic outcomes

increase in the percentage of households with basic services. These services are access
to piped water, access to sewers, and access to electricity. The percentage of households
without dirt floor after the shock will be 3.4% higher in a municipality in the 75 percentile
compare with one in the 25 percentile of suitability when the prices of heroin in the United
States double. The percentage of households with basic services will increase by 1.17%
in a 75 percentile municipality compared to one in the 25 percentile one. The difference-
in-difference coefficients for the different specifications can be found in Appendix A.6.3.
These results suggest that there was an income boost from the increase in heroin prices
and that it is probably increasing the quality of the houses of the poorest members of
these communities. These results are in line with anecdotal evidence from opium poppy
farmers that saw an increase in the amount pay by the cartels during this period of time.
A kilogram of raw opium could be sell at around 275 dollars in 2011 and went up to 900
by 2014 Stevenson (2015).
Military Eradication: In this section I use the eradication data from the Mexican mili-
tary to explore if the shock affected the amounts of opium and marijuana seized. I assume
that the military did not became better at eradicating one particular crop during this pe-
riod and that eradication can be seen as a proxy of total production. Figure 15 shows the
coefficients for specification (1) for the log hectares eradicated of marijuana and opium.
There is a clear jump in the amount of eradication of opium and a slowly decline in the
eradication of marijuana after 2010. There does not seem to be any particular pre-trend
for any of the crops. Table 4 shows the difference-in-difference coefficients with the differ-
ent controls and fixed effects, the results are robust to all of these different specifications.
These results confirm that after the shock the value of cultivating opium increases relative

32
to the value of marijuana, which lead to the farmers to switch crops.43

Figure 15: Log of Hectares eradicated

0.2

0.0
Hectares Eradicated

−0.2

−0.4
MARIJUANA
OPIUM

−0.6

2004 2006 2008 2010 2012 2014 2016

Year

6 Cartel Competition and Concentration in the Heroin Market

In this section I provide evidence of the expansion of several cartels into the heroin mar-
ket after the increase demand in the United States. In the background section I already
mention that historically Mexico has produced opium and exported it to the United States
but it was never their main activity. Until 2006, there was just one cartel, the Sinaloa
Cartel; in areas where opium poppy was eradicated by the Mexican government. By 2016,
the nine main organizations can be found in municipalities where opium poppy was eradi-
cated by the military. Figure 16 shows the Herfindahl–Hirschman Index for the raw opium
market. The index was calculated using the cartel presence and the eradication data set.
I used the military eradication as a proxy for total production and the cartels present in
those municipalities to assign market shares. In municipalities where there was more than
one active cartel, I equally split the production between them. From the image the HHI
drops from 1 in 2004 to around 0.3 by 2016. The graph also shows that the biggest drop
occurs in 2010, after the reformulation of OxyContin, this measure confirms that when the
demand for heroin increased in the United States the cartels decided to enter or expand
43
Figures for seizures of cocaine and methampethamines can be found in the Appendix A.6.4.

33
 Table 4: Opium and Marijuana Eradication

Eradication
log(Hectares Opium Poppy) log(Hectares Marijuana)
Post-shock 0.198∗∗∗ 0.157∗∗∗ 0.088∗ 0.033∗ 0.040∗∗∗ −0.220∗∗∗ −0.192∗∗∗ −0.235∗∗∗ −0.210∗∗∗ −0.166∗∗
Radius 500 Km (0.001) (0.045) (0.038) (0.008) (0.004) (0.002) (0.048) (0.047) (0.047) (0.054)
Observations 32,202 31,752 32,202 31,752 31,752 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 0.130 0.130 0.130 0.130 0.130 0.295 0.295 0.295 0.295 0.295
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
F-test pre-trends 2.51 2.40 0.516 0.256 0.476 0.027 0.848 1.13 0.782 0.124

Notes: This table presents the results of the difference-in-difference model for the dependent vari-
ables: log of hectares of opium poppy and Marijuana eradicated by the Mexican military.These
results are estimated using the the same specification as Equation 1, but with the event dummy
years replace by a post shock dummy interacted with suitability and heroin prices. The first four
columns show the results for the probability of having more than two cartels active. The next four
columns show the coefficients for the number of cartels. Columns (1) and (6) present the results
with municipality and year fixed effects. Columns (2) and (7) add controls for police presence, mil-
itary presence, and political party of the mayor and the governor. Columns (3) and (8) control for
baseline characteristics interacted by year such as population, drug activity in 2000 and marginal-
ization rate. Columns (4) and (9) add to the baseline trends, the set of controls from Column (2)
& (7). Columns (5) and (10) add state specific time trends. Standard errors are clustered using a
500 Km radius. Significant at ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

34
their operations into this market.

Figure 16: HHI Opium

1.0

0.8
HHI for opium

0.6

0.4

2004 2007 2010 2013 2016

Year

6.1 Identifying Expanding and Competitive Cartels

This paper considers nine main Drug Trafficking Organizations recognized by the Mex-
ican Military and the DEA as the major actors. The interactions between these orga-
nizations are complex with them expanding, fragmenting, and disappearing. The last
decade saw some of them fragmenting due to the government strategy of capturing king-
pings. These strategies left power gaps that lead to internal disputes in these organizations
Phillips (2015). Less explored is fragmentation due to market pressures. The increase
in the market value of certain territories might incentivize drug traffickers to break from
their main organizations or stop previously pacific agreements with former allies to capture
higher market shares. Around the time of the shock the Jalisco New Generation Cartel
split from the Sinaloa Cartel InSightCrime (2019) and took some of its former parent or-
ganization’s territory suitable for growing opium poppy. The second fragmentation that
happened was the Knights Templar Cartel splitting from La Familia Michoacana. The
group that started as a self-defense against criminal organizations rapidly became a drug
cartel. These two groups have been fighting with each other for the control of the state of
Michoacan. I used the specification below to quantify the effect that the shock had on the
presence of each of the nine organizations.

35
 Econometric Specification:
2016
X
c
Ymt = βt Suitm ∗ P riceHert + ψXmt + αm + γt + σe t + mt (3)
t=2004

where Ymt c is the probability of having the cartel c active in municipality m and time

t. Suitm is the measure of suitability in the municipality m, P riceHert is the retail price
of heroin in the United States per milligram of heroin adjust by purity normalized to 2002
prices, Xmt is a set of controls that includes: police presence, military presence, political
party of the mayor and political party of the governor of the state. αm are municipality
fixed effects, γt year fixed effects and σe t are state specific time trends.

Results The nine major organizations studied in this paper show at least some pres-
ence in municipalities well suited to grow opium poppy. Figure 17 shows the coefficients
from equation (3) with municipality and year fixed effects, controls for police presence,
military presence, and political affiliation of the mayor and governor of the state, baseline
characteristics interacted with year and state specific time trends. The standard errors are
clustered using a 500Km radius. The dependent variable is a dummy that is one when
cartel c is present and zero otherwise. The graph presents the results for the five main car-
tels that have activity in the heroin market. There are two cartels expanding, the Sinaloa
Cartel and Los Zetas, two cartels splitting from existing ones and immediately entering the
heroin market, Jalisco New Generation and the Templar Knights and one cartel loosing
presence, La Familia Michoacana. The difference-in-difference coefficients for the different
specifications can be found in the Appendix A.6.5. To interpret these results lets compare
a high-suitable municipality with a low-suitable one when the price of heroin increased by
30%, this was the actual price increase between 2009 and 2010, in the United States. The
probability of the Sinaloa Cartel being present in a high suitable municipality is 3.12%
higher each year after the shock compare to a low suitable municipality. The probability
of the Jalisco New Generation or the Templar Knights being present increases by 4.42%
each year after the shock and the probability of Los Zetas increases by 6.5%. Finally, the
probability of La Familia being present decreases by −2.47% each year after the shock.
The other fours cartels: Tijuana, Golf, Juarez, and the Beltran-Leyva Organization do not
show a significant increase in the probability of being present in high-suitable municipalities
despite the fact they do enter and expand during this period of time to opium producing
territories. A detailed analysis of what happened to this other four cartels and the co-
efficients can be found in Appendix A.6.5. These results suggests that market pressures
can also lead criminal organizations to split and fight with previous allies to get control of
valuable territory.

36
 Figure 17: Effect of Demand Shock on Individual Cartel Entry

Post−shock DD
Sinaloa:0.024
CJNG:0.034
Templar−Knights: 0.034
0.10 Zetas:0.050
La Familia:−0.019
Probability of Presence

CJNG
La Familia
0.05 Sinaloa
Templar−Knights
Zetas

0.00

−0.05

2004 2007 2010 2013 2016

Year

7 Conclusion
This paper provides evidence of the relationship between market structure and violence
in illegal markets. I contribute to the literature by adding market pressures and external
demand shocks as factors that increase violence and the number of Drug Trafficking Or-
ganizations. The results from this paper suggest that a shift in the demand for drugs in
consuming countries have direct effects in producing and trafficking countries. Particularly,
I emphasize that criminal organizations are profit-maximizing players that will decide to
enter and expand into profitable markets, and in the absence of a strong institutional set-
ting they will use violence to win market power. Though the Mexican context and the
interaction with the opioid crises is unique in some dimensions, there are many other ex-
ample of illegal markets which might have similar unintended consequences from shifts in
demand. Although, the levels of violence and drug crime related homicides in Mexico are
an exception and not the rule, some other local illegal markets might experience similar
market pressures and lead to spurs of violence or the surge of numerous criminal organiza-
tions. Understanding how criminal organizations organize and how their market structures
interact with violence is key to implementing policies that aim not just to reduce illegal

37
activities like drug trafficking but the negative externalities associated with them.
This paper also provides two novel data sets that can be used to answer other ques-
tions related to how Mexican Drug Trafficking Organizations operate and react to different
policies. I also provide a set of techniques that can be replicated elsewhere to generate
new data sets. The use of online content and deep learning should become a more common
practice that would allow us to measure otherwise difficult to quantify phenomena.
More generally, this paper provides evidence of how the market structure of organized
criminal activities determines the amount of violence, motivating future research on under-
standing their structure, competition practices, and relationships with the legal economy.

38
References
Abby, A., Powell, D. and Pacula, R. L. (2018), ‘Supply-Side Drug Policy in the Presence
of Substitutes: Evidence from the Introduction of Abuse-Deterrent Opioids’, American
Economic Journal: Economic Policy 10(4), 1–34.

Agencias (April 2016), ‘Cae ‘El Gringo Payán’, presunto operador de El Mayo Zambada’,
Zocalo.
URL: https: // www .zocalo .com .mx/ new _site/ articulo/ cae-el-gringo-payan-
presunto-operador-de-el-mayo-zambada-1435099245

Ahmed, A. (August 2015), ‘Young Hands in Mexico Feed Growing U.S. Demand for Heroin’,
The New York Times.
URL: https: // www .nytimes .com/ 2015/ 08/ 30/ world/ americas/ mexican-opium-
production-rises-to-meet-heroin-demand-in-us .html

Atuesta, L. H. and Ponce, A. F. (2017), ‘Meet the Narco: increased competition among
criminal organisations and the explosion of violence in Mexico’, Global Crime 18(4), 375–
402.

Biderman, C., Mello, J. M., De Lima, R. S. D. and Schneider, A. (2018), ‘Pax Monopolista
and Crime: The Case of the Emergence of the Primeiro Comando da Capital in São
Paulo’, Journal of Quantitative Criminology 35(3), 573–605.

Bounadi, M. E. (2018), ‘Weather and Conflicts in Afghanistan’, EBA Working Paper,Expert

Group for Aid Studies, Sweden. .

Calderón, G., Robles, G., Díaz-Cayeros, A. and Magaloni, B. (2015), ‘The Beheading of
Criminal Organizations and the Dynamics of Violence in Mexico’, Journal of Conflict
Resolution 59(8), 1455–1485.

Castillo, J. C. and Kronick, D. (2019), ‘The logic of violence in turf war’, Working paper .

Conley, T. G. (1999), ‘GMM estimation with cross sectional dependence’, Journal of Econo-
metrics 92(1), 1 – 45.

Coscia, M. and Rios, V. (2017), ‘Knowing where and how Criminal Organizations Operate
using Web Content’, Proceedings of the 21st ACM international conference on Informa-
tion and knowledge management pp. pp. 1412–1421.

DEA (2018), 2018 National Drug Threat Assessment, Technical report, U.S. Department
of Justice Drug Enforcement Administration.

Debate, E. (Febrero 2017), ‘Así era el H9, heredero del Cártel de los Beltrán Leyva’, El
Debate.

39
 URL: https: // www .debate .com .mx/ mexico/ Asi-era-el-H9-heredero-del-
Cartel-de-los-Beltran-Leyva-20170212-0068 .html

Dell, M. (2015), ‘Trafficking Networks and the Mexican Drug War’, American Economic
Review 105(6), 1738–1779. &nbsp;.

Dell, M., Feigenberg, B. and Teshima, K. (Forthcoming), ‘The Violent Consequences of

Trade-Induced Worker Displacement in Mexico’, American Economic Review: Insights .

Dube, O., García-Ponce, O. and Thom, K. (2016), ‘From Maize to Haze: Agricultural
Shocks and the Growth of the Mexican Drug Sector’, Journal of the European Economic
Association 14(5), 1181–1224.

Dube, O. and Vargas, J. F. (2013), ‘Commodity Price Shocks and Civil Conflict: Evidence
from Colombia’, The Review of Economic Studies 80(4), 1384–421.

Feuer, A. (December 2018), ‘El Chapo Trial Shows That Mexico’s Corruption Is Even
Worse Than You Think’, The New York Times.
URL: https: // www .businessinsider .com/ mexico-catch-cjng-jalisco-cartel-
gunmen-who-shot-down-army-helicopter-2018-8

FOROtv (May 2017), ‘Vinculan a proceso a ‘El Garo’, operador del Cártel del Golfo en
NL’, Noticieron Televisa.
URL: https: // noticieros .televisa .com/ ultimas-noticias/ vinculan-
proceso-el-garo-operador-cartel-golfo/

Frick, S., Kramell, R., Schmidt, J., Fist, A. J. and Kutchan, T. M. (2005), ‘Comparative
Qualitative and Quantitative Determination of Alkaloids in Narcotic and Condiment
Papaver somniferum Cultivars’, Journal of Natural Products 68(5), 666–673.

Gehring, K., Langlotz, S. and Stefan, K. (2018), ‘Stimulant or Depressant?: Resource-

Related Income Shocks and Conflict’, Household in Conflict (HiCN) Working Paper
(286).

Gentzkow, M., Kelly, B. and Taddy, M. (2017), ‘Text as Data’, (23276).

Grillo, I. (2011), El Narco: Inside Mexico’s Criminal Insurgency, Bloomsbury Publishing.

InSightCrime (2019), Jalisco Cartel New Generation, Technical report, InSight Crime.
URL: https: // www .insightcrime .org/ mexico-organized-crime-news/ jalisco-
cartel-new-generation/

Jones, N. (2013), ‘The unintended consequences of kingpin strategies: kidnap rates and
the Arellano-Félix Organization’, Trends in Organized Crime 16(2), 156–176.

40
Karina (May 2015), ‘Declaran tres presuntos delincuentes detenidos tras balacera Tanhu-
ato’, Sdpnoticias.
URL: https: // www .sdpnoticias .com/ estados/ delincuentes-detenidos-
presuntos-declaran-tres .html

Kienberger, S., Spiekermann, R., Tiede, D., Zeiler, I. and Bussink, C. (2017), ‘Spatial risk
assessment of opium poppy cultivation in Afghanistan: integrating environmental and
socio-economic drivers’, International Journal of Digital Earth 10(7), 719–736.

Kim, Y. (2014), ‘Convolutional neural networks for sentence classification’.

Mavrelli, C. (2017), Transnational Crime and the Developing World, Technical report,
Global Financial Integrity.

Mejía, D. and Restrepo, P. (2013), ‘Bushes and Bullets: Illegal Cocaine Markets and
Violence in Colombia’, (011934).

Mesquita, E. B. d. (2018), ‘Territorial Conflict over Endogenous Rents’, The Journal of

Politics 0(ja), null.

Millán-Quijano, J. (2019), ‘Internal Cocaine Trafficking and Armed Violence in Colombia’,

Economic Inquiry 0(0).

Mosso, R. (April 2015), ‘Detienen a operador de ‘Los Zetas’ en Culiacán’, Milenio.

URL: https: // www .milenio .com/ policia/ detienen-a-operador-de-los-zetas-
en-culiacan

Muhuri, P., Gfroerer, J. and Davies, C. (2013), ‘Associations of Nonmedical Pain Reliever
Use and Initiation of Heroin Use in the United States’, CBHSQ Data Review .

O’Neil, S. (2009), ‘The Real War in Mexico: How Democracy can Defeat the Drug Cartels’,
Foreign Affairs .

Osorio, J. (2015), ‘The Contagion of Drug Violence: Spatiotemporal Dynamics of the

Mexican War on Drug’, The Journal of Conflict Resolution 59(8), 1403–432.

Phillips, B. J. (2015), ‘How Does Leadership Decapitation Affect Violence? The Case of
Drug Trafficking Organizations in Mexico’, The Journal of Politics 77(2), 324–336.

Rios, V. (2013), ‘Why did Mexico become so violent? A self-reinforcing violent equilibrium
caused by competition and enforcement’, Trends in Organized Crime 16(2), 138–155.

Schmalensee, R. (1992), ‘Sunk Costs and Market Structure: A Review Article’, The Journal
of Industrial Economics .

41
Shaver, A., Carter, D. and Shawa, T. W. (2019), ‘Terrain Ruggedness and Land Cover:
Improved Data for Most Research Designs’, Conflict Management and Peace Science
36(2), 191–218.

Signoret, P. (2018), ‘Do Kingpin Removals Cause Violence? Testing Mechanisms and
Reassessing Evidence in Mexico’.

Snyder, R. and Duran-Martinez, A. (2009), ‘Does illegality breed violence? Drug trafficking
and state-sponsored protection rackets’, Crime, Law and Social Change 52(3), 253–273.

Sonin, K., Wilson, J. and Wright, A. (2019), ‘Rebel Capacity and Combat Tactics’, Becker
Friedman Institute , working paper No. 2018-74 .

Stevenson, M. (2015), ‘Narco Mexicano produce más opio’, Dallas News en Español.
URL: https: // www .dallasnews .com/ espanol/ al-dia/ mexico/ 2015/ 10/ 05/
narco-mexicano-produce-mas-opio/

Sviatschi, M. M. (2018), ‘Making a Narco: Childhood Exposure to Illegal Labor Markets

and Criminal Life Paths’, (2018-03).

Sánchez de Tagle, O. (January 2014), ‘Las autodefensas nos dieron un respiro’, Animal
Político.
URL: https: // www .animalpolitico .com/ 2014/ 01/ michoacan-entre-balazos-
la-pgr-y-detenidos-por-incendio-una-farmacia/

Sánchez Valdés, V. M. (2015), ‘La nueva configuración del crimen en Michoacán’, Animal
Político.
URL: https: // www .animalpolitico .com/ el-blog-de-causa-en-comun/ la-
nueva-configuracion-del-crimen-en-michoacan/

Sánchez Valdés, V. M. (2017), ‘Los cárteles que operan en el centro de México’, Animal
Político.
URL: https: // www .animalpolitico .com/ el-blog-de-causa-en-comun/ las-
organizaciones-criminales-operan-centro-mexico/

Woody, C. (November 2017), ‘Mexican heroin is flooding the U.S., and the Sinaloa cartel
is steering the flow’, Business Insider.
URL: https: // www .businessinsider .com/ sinaloa-cartel-sending-mexican-
herion-to-the-us-2017-11

Young, T., Hazarika, D., Poria, S. and Cambria, E. (2017), ‘Recent trends in deep learning
based natural language processing’, CoRR .

Zou, H. and Hastie, T. (2005), ‘Regularization and Variable Selection via the Elastic Net’,
Journal of the Royal Statistical Society. Series B (Statistical Methodology) 67(2), 301–20.

42
A Appendix
A.1 Background
A.1.1 Drug Trade in Mexico
Mexico’s location has made it a key country for transporting goods between Latin
America and the United States, including narcotics and contraband. The origin of the
Drug Trafficking Organizations of today can be traced back to the Prohibition era in the
United States, when the first criminal organizations established several routes from Mexico
to the US border in order to smuggle alcohol (Grillo 2011).
At the beginning of the twentieth century the US and Mexican governments slowly
started prohibiting the production and consumption of some substances that included
marijuana, opium, and cocaine. Mexican traffickers saw this as an opportunity, and started
smuggling illegal drugs through the same routes they used to smuggle alcohol. Networks
are a key element for illegal drug trafficking. The least cost drug transportation routes
are probably different from the ones used by legal goods. Traffickers need not just to
minimize distance and transport costs, also the probability of government interdictions
and turf wars. Once a route has been proven effective to smuggle an illegal good it is
likely that traffickers will use it for other illegal activities. Timidly the drug traffickers
expanded across Mexico and the United States. The increased demand of marijuana during
the sixties in the United States, combined with lax laws regarding cultivation in Mexico,
drove the Mexican production of marijuana up and the consolidation of big trafficking
organizations begun. The United States government did not like the careless attitude
of their Mexican counterparts and launched the first big anti-drugs campaign know as
operation Intercept. This operation nearly shut down the US-Mexican border to stop
marijuana shipments but did not had any significant effect on the amount of marijuana
crossing the border. In 1975, the Mexican government launched its first big anti-drug
operation, known as operation Condor which used the military to eradicate illegal crops
in the Golden Triangle.44 These two operations reduced the drug traffickers revenue from
marijuana and made them shift into the cocaine market. During the late 1980s, the US
authorities started breaking the Colombian Drug Trafficking Organizations and closed the
Caribbean route between Colombia and Miami. The Mexican traffickers slowly took over
the transportation of cocaine between Colombia and the United Stated. They already had
established routes for transporting marijuana and used them to cross cocaine across the
border. The share of cocaine arriving to the US moved through Mexico grew from 50% in
the early 1990s to almost 100% by 2000 (O’Neil 2009).
The origin of most of the current Drug Trafficking Organizations in Mexico can be
trace back to the Guadalajara Cartel, consolidated during the 1980s by Miguel Ángel Félix

44
Region known for its high production of marijuana and opium located where the states of Sinaloa, Durango
and Chihuahua come together.

43
Gallardo, a former police officer. He established the connection between the Colombian and
Mexican organizations and expanded the routes between Mexico and the United States.
Since their origins Mexican Drug Trafficking Organizations have retained close relation-
ships with local authorities. These strong ties survived for decades while the hegemonic
party PRI was in power. During the late 1980s the PRI started loosing elections across the
country. This led to a crack in the previous agreements between the government and crim-
inal organizations. In 2000, the dominant party PRI lost for the first time the presidential
election. Violence across the country remained low but without the previous state spon-
sored protection cartels fought for valuable territory (Snyder and Duran-Martinez 2009).
Violence between several organizations escalated quickly to levels not seen before.

A.1.2 The War on Drugs

Former Mexican President Felipe Calderón, in office between 2006-2012 from the conser-
vative party PAN, declared war on drugs in 2006. The Drug Trafficking Organizations vio-
lently fought back. Violence between drug traffickers escalated particularly in Michoacán,
where the government decided to deployed military and federal police in order to reduce
the cartel violence. At the beginning of the twentieth first century there were just four
major Drug Trafficking Organizations; sixteen years later the DEA and the Mexican mil-
itary identified at least nine major actors. Drug Trafficking Organizations (DTOs) have
displayed their violence with the public beheading of corpses, car bombs, and the murders
of journalists and public officers.
Violence spread quickly beyond the US-Mexican border and into the whole country.
Since 2006, there have been an estimated of 250,547 homicides related to organized crime.
These account for 50% of the country’s homicides. The number of missing people is up
to 37,400, and the number of displaced people due to violence is estimated to be around
380,000. The number of missing or killed media workers exceeds 200. Despite President
Enrique Peña Nieto’s (2012-2018) efforts to reduce violence, his strategy towards Drug
Trafficking Organizations did not significantly change from the one established by Calderon.
Violence decreased a little during the first year of his administration but went up again
immediately afterwards. The government strategy of beheading organizations by targeting
high rank kingpings generated a lot of instability inside the cartels which led to more
violence (Jones 2013; Calderón et al. 2015; Phillips 2015; Signoret 2018). This strategy
has led to cartel fragmentation, when the main kingpin is either detained or killed. This
leaves an open position at the top of the organization, cartel fractions that used to operate
peacefully will splintered from the parent organization and form new organizations.
Corruption and political instability in Mexico adds to the difficulty of fighting these
criminal organizations. The list of former governors, majors, members of congress, police-
men, and military authorities that had some relationship with drug traffickers is extensive
and shows how difficult fighting these organizations can be.45 The War on Drugs led to a
45
Feuer (December 2018)

44
spike in homicides and increased violence particularly between 2008 and 2010. The period
after 2010 has been studied less. This paper adds to the discussion of what generated the
increase in violence.

A.1.3 Mexican Cartels

The Gulf Cartel: founded in the 1930s as an alcohol smuggling organization. After
Prohibition era it continued to operate other illegal activities like prostitution and gambling
rings. By 1980s, it was one of the most powerful cartels in Mexico and had established
connections with the Cali Cartel in Colombia. After the arrest of García-Ábrego in 1996,
the cartel kingpin, they lost the Colombian connection and started to loose power. During
the increased turf wars in the early 2000s, the cartel recruited former military special forces
to form the cartel’s armed wing, Los Zetas. This group will eventually separate from its
parent organization and become an independent cartel. Currently, the Gulf Cartel is still
fighting with Los Zetas to control Mexico east coast.
The Juarez Cartel: originated during the 1970s in the city of Juarez. During the
1980s, the Juarez Cartel worked with the Guadalajara Cartel and after this was disman-
tled Amado Carillo Fuentes, alias ‘Lord of the Skies’ assumed control. The organization
grew exponentially under Carillo Fuentes. Eventually, he controlled half of the Mexican
trafficking and expanded to South America. After his death in 1997 his brothers took over
the organization. The Juarez Cartel has debilitated due to constant fights with the Sinaloa
Cartel over the Juarez corridor. Despite this, it remains a powerful cartel because of its
large and longstanding transportation, storage, and security networks through the country.
La Familia Michoacana: emerged in the late 1980s as a self defense group against
the drug dealers that operated in the state of Michoacán. Since then, they transformed
into a criminal organization. In the early 2000s they were able to drove Los Zetas out of
Michoacán and expanded their operations to neighboring states. They kept the hegemony
of the region and were one of the most violent organizations. In 2011, the cartel fragmented
into two organizations: La Familia and the Knights Templar. La Familia lost almost all
its territory to the Knights Templar Cartel.
The Tijuana Cartel: after Miguel Félix Gallardo was arrested in 1989 he divided his
former territory and gave it to former allies. His nephews the Arellano-Félix brothers got
Tijuana. The cartel was in constant turf with the Sinaloa Cartel for the control of the
border crossing in the city of Tijuana. By 2013, all the Arellano-Félix brothers were death
or in prison. Enedina, the last Arellano-Félix that remains has now the command of the
organization. The Tijuana Cartel loose the control of Tijuana to the Sinaloa Cartel around
2010. Since 2015, the remains of the Tijuana Cartel had been trying to regain control of
the city.
The Sinaloa Cartel: described as the largest and most powerful criminal organization in
the Western Hemisphere. Joaquín, ‘El Chapo’ Guzman a former employee of the Gudala-
jara Cartel kept the Sinaloa territory after Félix Gallardo was captured. El Chapo fought

45
with his former allies in Tijuana and expand his cartel operations across the country. Since
2012, the cartel has been involved in several turf wars with all the other cartels. Despite
El Chapo’s extradition the Sinaloa Cartel has remained one of Mexico’s most powerful
criminal organizations.
Beltrán-Leyva Organization: former allies of the Sinaloa Cartel. The Beltrán-Leyva
worked closely with El Chapo until 2008 when one of the brothers was arrested and the
other four blame their former boss for the arrest. They decided to leave the Sinaloa Cartel
and founded their own organization. Since then they have been engage in several turf wars
with the Sinaloa Cartel.
Jalisco New Generation Cartel(CJNG): emerged from the Sinaloa Cartel in 2010. The
criminal group used to moved drug shipments and managed the finances of the Sinaloa
Cartel. The cartel has been fighting Los Zetas, the Knights Templar, and the police with
extreme violence. The organization expanded rapidly and became one of the most powerful
cartels.
Los Zetas: former armed wing of the Gulf Cartel. Originally composed of commandos
of the Mexican Army that deserted and joined the Gulf Cartel. Los Zetas won power
over time and eventually outnumbered their former parent organization. They started
a violent turf with the remains of the Gulf Cartel. They also started to expand to the
territories dominated by other cartels and brought violence with them. Despite several
internal disputes and that their main leaders have been captured or killed some cells of the
organization are still active and have control of some parts of the country.
The Knights Templar: emerged in 2011 as a splinter of La Familia. They took control
over La Familia operations and territory, including Mexico’s second biggest port Lazaro
Cárdenas. In recent years they have been fighting with the CJNG over the control of the
states of Jalisco and Michoacán.

A.2 Using Google News and Natural Language Processing to Create a Dataset
of Cartel Presence
The algorithm I used to generate the cartel presence data set is described here:
• Web scraping: I used the nine cartels that the DEA and the Mexican persecutor’s of-
fice recognize as major Drug Trafficking Organizations. With these organizations and
all the Mexican municipalities I created a unique set of key words that were used to
search in the Google News interface. To reduce ambiguities regarding municipalities
and cartel names I followed the same rules as Coscia and Rios (2017). These unique
words combinations were passed to the crawler that looked for any information source
containing the municipality-cartel pair between 1990 and 2016. The algorithm found
2,249,561 news articles.46
The next step consists on dropping all the articles that are not talking about a
municipality-cartel pair. The search engine will sometimes give back links that does
46
I restricted the search for articles in Spanish and appearing in Google News Mexico interface.

46
 (a) Example 1 (b) Example 2

Figure A.1: Web Pages Examples

not contained the query in the main article. It may give back as a valid hit of a
particular query a news article that talks about one cartel in the core and mentions
in the side of the article a municipality. Figure A.1a shows an example of this. The
main article talks about the Sinaloa Cartel being present in the city of La Paz but
one of the side articles talks about the municipality of Los Cabos. The search engine
will gave a valid hit on this article for the pairs La Paz-Sinaloa Cartel and Los Cabos-
Sinaloa Cartel, when the main article is just talking about presence in the first case.
The other kind of invalid hits that this step eliminates is the ones where the name of
a city or municipality is part of the newspaper name. Figure A.1b shows an example
of this. The main article describes a police operation in Coyoacán in Mexico City
were they capture several members of the Beltran-Leyva organization. The problem
is that the newspaper name is "El Siglo de Torreón" and Torreón is a municipality.
Then the web crawler will return this article as the pair Coyoacán-Beltran Leyva and
Torreón-Beltran Leyva but just the first pair describes presence of the organization
in that municipality.

• Natural Language Processing: in order to only keep the articles that talk about a
municipality-cartel pair I first used several python libraries that extract the main
core from the whole html. Keeping just the articles with the municipality-cartel pair
will prevent the kind of errors described above. The number of remaining articles
after this process was 1,201,483. I used a sentence extraction algorithm to keep each
sentence of the articles that mention any municipality-cartel pair. This process found
2,802,224 of these sentences.
Next, I manually classified 5,000 sentences as either presence or not to train the
algorithm. The following sentences are examples classified as presence. Most of this
sentences are journalist reporting on police reports, these sentences confirms that
journalist are still reporting on Drug Trafficking Organizations but mostly after the
police has made a report.

47
 Examples classify as presence:
-The attorney general in Nayarit started some operations that end up with the capture
of two Beltran-Leyva operators in Tepic Debate (Febrero 2017).
-Apatzingan, the most dangerous place for the self defense groups is still controlled
by Knights Templar Cartel Sánchez de Tagle (January 2014).
-Mexican Marines caught Hugo Cesar Roman Chavarria, alleged Zetas operator in
charge of trafficking drugs through Coahuila and Nuevo Leon Mosso (April 2015).
-Gerardo Payan alleged operator of the Sinaloa Cartel was caught by marines las
Thursday in Mocorito, Sinaloa Agencias (April 2016).
-El Garo, operator of the Gulf Cartel in the municipalities of Apodaca, Garcia and
Santa Catarina is waiting for trial FOROtv (May 2017).
-Yesterday, while investigating a trespassing case federal police had a confrontation
with members of the Jalico New Generation Cartel in Tanhuato Karina (May 2015).

To classify each of the sentences as the cartel c being in municipality m I used a

semi-supervised convolutional neural network Young et al. (2017). The input that
the CNN uses are the sentences. First, I used a pre-trained word embedding in
Spanish, trained using Wikipedia articles to transformed each of the words in the
sentence into a vector. A word embedding maps meaning into a geometric space,
transforming each sentence into a matrix. Let wi ∈ Rd be the word embedding
for the ith word in a sentence, the particular embedding has d = 100. A sentence
with n words is then represented as a matrix W ∈ Rnxd . The convolution will be
apply to this matrix. The convolution involves a filter k ∈ Rhd which is applied
to a window of h words to produce a new feature. A number of different filters is
then applied to the embedding matrix with different windows sizes that slide over
the entire word embedding matrix. Each of them extracts a particular pattern from
the n-gram. The particular CNN used here is a one dimensional convolution with
a window size of five, 128 filters, and with a rectifier47 activation function. Each
convolution layer is followed by a max-pooling operation that will map the results
of each filter into a fixed dimensionality output. The neural network has 10 hidden
layers that used the rectifier activation function and there is a final layer that uses
the sigmoid function to classify each sentence as a 1, cartel presence or a 0 non cartel
presence. All the parameters for this CNN where chosen by a grid-search of 100
points in each dimension and these combination of parameters gave the higher out
of sample predictability.
To train this model I manually classified 5,000 sentences and then use 4,000 as the
training set and 1,000 as the test set. The CNN used here is semi-supervised because
it uses unlabeled data to understand the general population data. This method is
useful when, like in these case you have a small label subset and a large unlabeled
47
An activation function defined as the positive part of its argument.

48
 data. The algorithm works as follows, after using the small train data and getting
a good prediction, use the unlabeled data to generate predictions (pseudo-labels).
Concatenate the labels and the features of the training and test set. Then, train the
model again using this new training data. The objective of this kind of training is
that the network is able to learn more of the general structure of the data.
The chosen CNN has an out of sample classification accuracy of 0.86, this means in
the 1,000 sentences that where never include in neither the original data set nor the
semi-supervised training. Once each sentence is classify as a 1, presence or a 0 no
presence. I identify the municipality-cartel pair in each sentence and assigned to that
pair. The years are extracted from the whole article, if there is a year in or near the
sentence then the presence is assigned to that year. If there is not year in the article
I used the publication date as the year.
• Alternative to using a CNN I could have used a bag of words algorithm to generate
this data set. One main reason not to do this is because in the context of Drug
Trafficking Organizations in Mexico context and meaning is key. There are several
municipalities with the same name, drug dealers, politicians and military officers
that share last name with the name of a municipality, and also DTOs, states, and
municipalities that share the same name. To provide evidence that the CNN performs
better than a bag of words algorithm I use the same train and test set as in the CNN
to train a bag of words using a logistic classifier. The out of sample classification
accuracy of this algorithm is 0.63 compare to the out of sample accuracy of the CNN
0.86. The accuracy goes down which was expected but in order to know how good it is
at overall predicting cartel presence I compare this alternative data set with the same
data sets used to validate the CNN one. The correlation between the bag of words
data set and Coscia and Rios (2017) data goes down to 0.26, the correlation with the
data sets created by Professor Victor Sanchez goes down to 0.41, and the correlation
between this data and the DEA aggregate data at the state levels decreases to 0.14.
This gives me confidence that the CNN and that including more complex mechanism
of classifying sentences is a good idea specially in cases like this where context and
meaning seem particularly relevant to correctly classify an article.
• Table A.1 shows when and where each cartel appears for the first time in the data
set created by the algorithm above. The first Column shows the year the cartels first
appeared in the data set, some of this cartels are older but the algorithm did not
search before 1990. The second Column shows the first municipality where each cartel
appears in the data set and the next four columns show in how many municipalities
each cartel was active in 1990, 2000, 2010, and 2016. Table A.2 shows the summary
statistics for this data set between 2004 and 2016, the time period used for this paper.
Active is a dummy that is one if there is at least one cartel active that year in the
municipality. Competitive is another dummy that is one when there are at least two
cartels active in a municipality.

49
 Table A.1: Summary of cartel presence
Cartel First Year Appear First Mun Mun Active 1990 Mun Active 2000 Mun Active 2010 Mun Active 2016
Juarez 1990 Chihuahua/Juarez 3 5 42 79
Tijuana 1990 Tijuana/Mexicali 9 2 22 42
Sinaloa 1990 Sinaloa/El Fuerte 9 4 88 160
Beltran Leyva 2004 Sinaloa 0 0 85 118
Fam Mich 1990 Mujica/La Piedad 3 2 47 101
Golf 1990 Matamoros 1 2 36 89
Zetas 1999 Matamoros 0 9 100 227
CJNG 2010 San Juan de los Lagos 0 0 12 211
Templar Knights 2011 Acuitzio 0 0 0 107

Table A.2: Cartel Presence Summary Statistics

Statistic N Mean St. Dev. Min Pctl(25) Pctl(75) Max

Beltran-Leyva 41,752 0.017 0.128 0 0 0 1
Gulf 41,752 0.011 0.103 0 0 0 1
CJNG 41,752 0.014 0.117 0 0 0 1
Juarez 41,752 0.014 0.119 0 0 0 1
Sinaloa 41,752 0.023 0.148 0 0 0 1
Knights-Templar 41,752 0.012 0.107 0 0 0 1
Tijuana 41,752 0.007 0.083 0 0 0 1
La Familia 41,752 0.011 0.106 0 0 0 1
Zetas 41,752 0.033 0.178 0 0 0 1
Number of Cartels 41,752 0.140 0.628 0 0 0 9
Active 41,752 0.072 0.259 0 0 0 1
Competitive 41,752 0.034 0.180 0 0 0 1

50
A.3 Agro-ecological data
From opium to heroin: papaver somniferum is an annual crop with a short productive
cycle that can be harvested 4 times a year. The main product extract from it is raw opium.
Raw opium contains between 8−91.2% of morphine depending on the plant variation (Frick
et al. 2005). Opium is a low cost crop that does not requires many inputs but it is land
intensive. Usually an hectare of poppy flowers will produce between 8 to 15 kilograms of
raw opium. Yields of heroin from raw opium are between 6 to 10 percent. To transform
raw opium into heroin first the morphine needs to be extracted from the opium paste and
dried. Once the morphine is dry acetic anhydride is added to create brown tar heroin,
this can be smoked, inhaled or injected. To produce high purity white heroin: ammonia,
hydrochloric acid, and acetone is added.48
The variables used to construct the suitability index and the data sets they came from are
the following:

• Temperature and Precipitation from the WorldClim data set. This data set includes:
mean annual temperature, mean diurnal range, temperature seasonality, maximum
temperature during the warmest month, minimum temperature during the coldest
month, mean temperature of the wettest quarter, mean temperature of driest quarter,
mean temperature of warmest quarter, mean temperature of coldest quarter, annual
precipitation, precipitation of wettest month, precipitation of driest month, precip-
itation seasonality, precipitation of wettest quarter, precipitation of driest quarter,
precipitation of warmer quarter and precipitation of coldest quarter.

• Elevation data from Shuttle Radar Topography Mission (SRTM).

• Terrain ruggedness from Shaver et al. (2019).

• Land Cover from Globcover 2009.

• Soil quality from the FAO Harmonized World Soil Database.

• River density: diva-gis data-sets weighted by Strahler stream order.

• Soil climatic characteristics from the yearly Copernicus data set. This data set in-
cludes: mean monthly precipitation, mean evaporation, soil temperature, soil water,
heat flux, and rain.

The suitability index was built by regressing the log of opium yields by year and district
in Afghanistan between 1990 and 2018 on the 45 geo-climatic characteristics and all its
interactions. To reduce the number of regressors I use an elastic net. An elastic net is a
penalized OLS that combines lasso and ridge methods. I choose the best model using a
10 fold validation. The best model has parameters α = .4 and λ = 0.076. Here α is the
48
https://www.unodc.org/pdf/research/Bulletin07/bulletin_on_narcotics_2007_Zerell.pdf

51
degree of mixing between the ridge and the LASSO regression and the λ is the shrinkage
parameter. The number of variables reduces from 903 to 50.
Variables selected by the elastic net:
Soil temperature, mean annual evaporation interacted with temperature seasonality, mean
annual precipitation interacted with: average monthly rain, minimum temperature during
the coldest month, mean temperature during the driest quarter and terrain ruggedness.
Soil temperature interacted with seasonal precipitation, soil water interacted with : river
density and topsoil organic carbon, ruggedness interacted with: mean temperature dur-
ing the wettest quarter, topsoil grave, topsoil organic content and topsoil clay, heat flux
interacted with temporal seasonality, precipitation seasonality, river density and topsoil
clay, mean temperature interacted with: minimum temperature during the coldest month
and ruggedness, mean diurnal range interacted with mean temperature during the driest
quarter, precipitation seasonality and ruggedness, minimum temperature during the coldest
month interacted with: mean temperature during the warmest month, precipitation during
the wettest quarter, precipitation during the driest quarter, river density, topsoil calcium
carbonate and topsoil salinity, mean temperature during the wettest quarter interacted
with precipitation during the wettest quarter, precipitation during the warmest quarter,
altitude and topsoil sodicity, mean temperature during the driest quarter interacted with
mean temperature during the coldest quarter, precipitation during the wettest quarter and
ruggedness, mean temperature during the coldest quarter interacted with precipitation
during the warmest quarter, precipitation during the warmest month interacted with pre-
cipitation during the wettest quarter and ruggedness, precipitation during the driest month
interacted with: precipitation during the warmest quarter and river density,precipitation
seasonality interacted with precipitation during the wettest quarter, precipitation during
the wettest quarter interacted with precipitation during the warmest quarter and rugged-
ness,altitude interacted with topsoil teb,river density interacted with topsoil organic car-
bon, topsoil gypsum and topsoil salinity , topsoil sand interacted with topsoil calcium
carbonate, topsoil clay interacted with topsoil teb and topsoil cec soil interacted with top-
soil teb.

A.4 Theoretical Framework:

Under Cournot competition a vector [M1∗ , ..., MN ∗ ] will be an equilibrium if M ∗ max-
i
imizes πi with the strategies of all the other cartels treated as fixed M¯−i ∗ . Symmetric

equilibria will be such that Mi∗ = M ∗ for i = 1, ..., N . In this section I present conditions
that ensure the existence of these kind of equilibria.
A cartel i will choose its military capacity in order to maximize its profits πi = Rsi −
Mi − F taking as given all the other cartels military capacity.
The first order condition will be given by:
RηMiη−1 ReMi2η−1
PN η − η 2 −1=0 (A.1)
( N
P
j=1 Mj j=1 Mj )

52
Symmetric equilibria will imply Mi = M ∗ for all i so that equation (A.1) becomes:

N −1 M∗
 
η = (A.2)
N R

Two additional conditions most be satisfied at equilibrium. First, equation (A.1) must
∂ 2 πi
indicate a local maximum for the profit function. It is sufficient to have ∂M 2 < 0. This
i
second derivative when the military capacity is the same for all cartels become:

η(N − 1)[N (η − 1) + 2η] < 0 (A.3)

Equation (A.3) implies that N ≥ 2 in order for η > 0.

Second, πi at the equilibrium need to be non-negative, otherwise the cartels will be
better off staying out of the market. Profits need to be such that πi = Rsi − M − F > 0.
From equation (A.2) the profits being positive can be rewritten as N (η − 1) − η ≤ 0.
Rearranging this equation:
η
N≤ (A.4)
η−1
Like n ≥ 2 this implies that there will exist symmetric equilibria as long as 0 ≤ η ≤ 2.

A.5 Disappeared and killed journalists data

Mexico is one of the most dangerous countries to be a journalist. Official sources
report 207 disappeared or killed media workers since 2000. Most of these crimes remained
unsolved, improperly investigated, and with few perpetrators arrested and convicted. The
victims are spread across 23 states and 109 different municipalities. The type of media
they worked includes freelancers, government, international, local, and national reporters,
and also TV, and radio presenters.
In order to test if there is any particular bias towards misreporting a particular cartel I
regress the number of killed or disappeared journalists on each of the nine cartels presence.

Yit = βCitd + ψXmt + αm + γt + σs t + mt (A.5)

Yit is the number of disappeared or killed journalists in municipality m in year t. Citd is
a dummy variable that indicates whether the cartel d is present in municipality m and
year t. The αm are municipality fixed effects, γt year fixed effects and σs t state specific
time trends. These fixed effects control for invariant differences between municipalities
and states that might made them more dangerous for journalists. Xmt is a set of controls
that includes; police, and military presence, a dummy if the mayor is from the conservative
party PAN and a dummy if both the mayor an the state governor are from the conservative
party PAN. Standard errors are clustered at the municipal level.
Table A.3 shows the results of this regression for each cartel. All the coefficients are
near zero and none of them are significant. This confirms that there does not seem to be a

53
 Table A.3: Relationship between death and disappeared journalist and cartel presence

Killed or disappeared media workers

killed
Sinaloa 0.016
(0.009)

Beltran-Leyva 0.008
(0.009)

Gulf −0.013
(0.009)

CJNG 0.019
(0.012)

Juarez −0.007
(0.012)

Knights-Templar −0.015
(0.012)

Tijuana −0.007
(0.028)

La Familia −0.004
(0.011)

Los Zetas 0.012

(0.009)
Observations 32,012 32,012 32,012 32,012 32,012 32,012 32,012 32,012 32,012
Municipality FE 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3 3 3 3
State trends 3 3 3 3 3 3 3 3 3

Notes: This table presents the results from regressiong the number of disappeared or killed media
workers on the presence of the nine cartels. Significant at ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

54
particular cartel more lethal to media workers. The correlation between the journalists data
set and the presence of each cartel is also low. The highest correlation is with Los Zetas and
it is 0.17. There does not seem to be a high correlation between the total number of cartels
active in a municipality and the number of disappeared or killed journalists either. This
correlation is 0.13. This results and the map show in the threats to identification section
provides evidence that despite the fact that is dangerous for the reporters to inform on
drug cartels, there does not seem to be a particulae cartel that is more dangerous than the
other ones. This implies that using news articles to measure cartel presence is a good and
is probably close to data that local authorities have.

A.6 Robustness checks

A.6.1 Cartel Activity
Table A.4 shows the results of the difference-in-difference model for the dependent
variables: more than one cartel active and number of cartels. These results are estimated
using the the same specification as Equation (1), but with the event dummy years replace
by a post shock dummy interacted with suitability and heroin prices. Standard errors
are clustered here at the municipality level. The first four columns show the results for
the probability of having more than two cartels active. The next four columns show
the coefficients for the number of cartels. Columns (1) and (6) present the results with
municipality and year fixed effects. Columns (2) and (7) add controls for police presence,
military presence, and political party of the mayor and the governor. Columns (3) and (8)
control for baseline characteristics interacted by year such as population, drug activity in
2000 and marginalization rate. Columns (4) and (9) add to the baseline trends, the set of
controls from Column (2) & (7). Columns (5) and (10) add state specific time trends.

A.6.2 Cartel Entry and Exit

The next tables A.5, A.6, A.7, and A.8 show the coefficients for equation (2) but with
the event time year dummies replaced by a dummy variable P ostEntrymt , indicating that
in year t municipality m experienced the entry or exit of a cartel. The dependent variable
is the number of homicides per 100,000 inhabitants. The tables present the difference-
in-difference model for two set of independent variables. The first four columns show
the results for the first variable and the next four for the second one. Columns (1) and
(6) present the results with municipality and year fixed effects.Columns (2) and (7) add
controls for police presence, military presence, and political party of the mayor and the
governor. Columns (3) and (8) control for baseline characteristics interacted by year such
as population, drug activity in 2000 and marginalization rate. Columns (4) and (9) add
to the baseline trends, the set of controls from Column (2) and (7). Columns (5) and
(10) add state specific time trends. Standard errors are clustered using a 500 Km radius.
Significant at ∗p < 0.1; ∗ ∗ p < 0.05; ∗ ∗ ∗p < 0.01. The sample selection is conditional on

55
 Table A.4: The effect of the reformulation on cartel activity

Cartel Activity
Competitive Number of Cartels
PshockInst(dd) 0.108∗∗∗ 0.121∗∗∗ 0.064∗∗∗ 0.044∗∗ 0.049∗∗ 0.466∗∗∗ 0.537∗∗∗ 0.292∗∗∗ 0.227∗∗∗ 0.241∗∗∗
(0.02) (0.021) (.018) (.017) (0.017) (0.072) (0.075) (.058) (0.057) (0.061)

Observations 32,202 31,752 32,202 31,752 31,752 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 0.012 0.012 0.012 0.012 0.012 0.057 0.057 0.057 0.057 0.057
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3

Notes: This table presents the results of the difference-in-difference model for the dependent vari-
ables: more than one cartel active and number of cartels. These results are estimated using the the
same specification as Equation 1, but with the event dummy years replace by a post shock dummy
interacted with suitability and heroin prices. The first four columns show the results for the prob-
ability of having more than two cartels active. The next four columns show the coefficients for the
number of cartels. Columns (1) and (6) present the results with municipality and year fixed effects.
Columns (2) and (7) add controls for police presence, military presence, and political party of the
mayor and the governor. Columns (3) and (8) control for baseline characteristics interacted by year
such as population, drug activity in 2000 and marginalization rate. Columns (4) and (9) add to
the baseline trends, the set of controls from Column (2) and (7). Columns (5) and (10) add state
specific time trends. Standard errors are clustered by municipality. Significant at ∗ p<0.1; ∗∗ p<0.05;
∗∗∗
p<0.01

56
having n cartels in t − 1. There is selection in the sample so that the comparison group is
municipalities that stay with n cartels versus the ones that go to n + 1 or n − 1. Tables
A.5 to A.7 present the entry events and Table A.8 present the exit ones.

Table A.5: Entry of the First and Second Cartel

Homicides per 100,000 inhabitants

First Carte Entry Second Cartel Entry
Post-entry 1.64 2.05 2.07 2.12 1.43 7.08∗∗∗ 8.33∗∗ 6.41∗∗ 7.61∗∗ 5.28∗
(4.23) (2.14) (1.98) (2.18) (2.05) (2.71) (2.71) (2.80) (2.68) (2.39)
Observations 42,624 42,624 42,624 42,624 42,624 13,389 13,389 13,389 13,389 13,389
Pre-entry mean, dep. var. 15.43 15.43 15.43 15.43 15.43 15.38 15.38 15.38 15.38 15.38
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.6: Entry of the Third and Fourth Cartel

Homicides per 100,000 inhabitants

Third Cartel Fourth Cartel
Post-entry 13.46∗∗∗ 15.62∗∗∗ 14.70∗∗ 14.66∗∗ 12.76∗∗ 13.71∗∗ 15.59∗∗∗ 15.55∗∗ 15.39∗∗ 13.05∗∗∗
(4.90) (4.55) (5.07) (5.05) (4.28) (6.34) (4.42) (4.90) (4.85) (3.69)

Observations 7,719 7,719 7,719 7,719 7,719 4,409 4,409 4,409 4,409 4,409
Pre-entry mean, dep. var. 16.58 16.58 16.58 16.58 16.58 19.35 19.35 19.35 19.35 19.35
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

The results presented in the core of the paper are robust to different controls and fixed
effects. The number of observations is decreasing because the number of municipalities
that have n cartels is decreases with n. There are not significant results of the entry of the
seventh cartel into a municipality that already has six active cartels. There are also not
significant effects of the exit of a cartel when there are more than three organizations in a
municipality.

57
 Table A.7: Entry of the Fifth and Sixth Cartel

Homicides per 100,000 inhabitants

Fifth Cartel Sixth Cartel
Post-entry 17.53∗∗∗ 21.08∗∗∗ 19.01∗∗∗ 19.84∗∗ 16.41∗∗∗ 10.81∗ 14.48∗∗ 10.53∗ 11.62∗ 6.57∗
(7.52) (6.80) (7.19) (7.28) (4.34) (5.06) (4.52) (4.92) (5.95) (3.29)

Observations 2,493 2,493 2,493 2,493 2,493 1,451 1,451 1,451 1,451 1,451
Pre-entry mean, dep. var. 17.31 17.31 17.31 17.31 17.31 18.90 18.90 18.90 18.90 18.90
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.8: Cartel Exit; from one to zero and from two to one

Homicides per 100,000 inhabitants

Exit from 1 to 0 Exit from 2 to 1
Post-exit −2.01 −2.45 −1.85 −2.52 −3.62∗ −2.70∗ −3.27∗∗ −3.35∗∗ −3.23∗ −6.07∗∗
(2.33) (1.55) (1.49) (1.52) (1.64) (1.27) (1.18) (1.28) (1.28) (2.24)

Observations 12,289 12,289 12,289 12,289 12,289 2,556 2,556 2,556 2,556 2,556
Pre-entry mean, dep. var. 18.27 18.27 18.27 18.27 18.27 17.75 17.75 17.75 17.75 17.75
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

58
A.6.3 The effect of the shock in socioeconomic outcomes
Tables A.9 to A.11 show the coefficients for equation (1) but with the event time year
dummies replaced by a dummy variable P ostShockmt interacted with the suitability index
and the heroin price in the United States. The relevant event is again the 2010 OxyContin
reformulation. Each table presents the difference-in-difference model for two dependent
variables. Table A.9, log of the population and mean years of education. Table A.10,
percentage of households with a woman as the head. Finally, Table A.11 the percent-
age of households without dirt floor and the percentage of household with access to basic
services (electricity, water, and sewage). Columns (1) and (6) present the results with
municipality and year fixed effects. Columns (2) and (7) add controls for police presence,
military presence, and political party of the mayor and the governor. Columns (3) and (8)
control for baseline characteristics interacted by year such as population, drug activity in
2000, and marginalization rate. Columns (4) and (9) add to the baseline trends, the set of
controls from Column (2) and (7). Columns (5) and (10) add state specific time trends.
Standard errors are clustered using a 500 Km radius. Significant at ∗p < 0.1; ∗ ∗ p < 0.05;
∗ ∗ ∗p < 0.01. The results for all the other outcomes that did not change significantly
during this period are available upon request. The results in the main part of the paper
are robust to all the controls and fixed effects added in here.

Table A.9: Population and Education

Demographics:
log(population) Years of Education
Post2010 −0.058∗∗∗ −0.014∗∗∗ −0.069∗∗∗ −0.026∗∗ −0.040∗∗∗ −0.418∗∗∗ −0.815∗∗∗ −0.085∗ −0.111∗∗ −0.366∗∗∗
(0.0001) (0.004) (0.013) (0.012) (0.009) (0.008) (0.095) (0.034) 0.035) (0.089)
Observations 32,202 31,752 32,202 31,752 31,752 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 9.37 9.37 9.37 9.37 9.37 6.24 6.24 6.24 6.24 6.24
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

A.6.4 Military Seizures

Throughout the whole paper I argue that is the expansion of the demand for heroin led
to cartel entry and subsequence violence. In order to rule out other drugs expanding at
the same time or in the same municipalities as opium poppy I use specification (1) for the
amount of kilograms of cocaine and methampethamines seized by the Mexican military.
There does not seem to be any increase in the seizures of any of these two drugs in suitable
municipalities to grow opium poppy and there is also not an effect after 2010. Figure
A.2 shows the coefficients from the event study specification, the quantity of cocaine and

59
 Table A.10: Women as Head of Household

Demographics:
% Women Head of Households
Post2010 0.015∗∗∗ 0.011∗ 0.011∗ 0.010∗∗ 0.021∗∗∗
(0.002) (0.005) (0.005) (0.005) (0.006)
Observations 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 0.21 0.21 0.21 0.21 0.21
Municipalities FE 3 3 3 3 3
Time FE 3 3 3 3 3
Covariates 3 3 3
Baseline trends 3 3 3
State trends 3
Note:
∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.11: No-Dirt Floor and No-Basic Services

Economic outcomes:
% without Dirt Floor % of Households with basic Services
Post2010 0.073∗∗∗ 0.038∗ 0.081∗∗∗ 0.081∗∗∗ 0.056∗∗∗ 0.025∗∗∗ 0.014∗ 0.021∗∗ 0.013∗∗ 0.015∗∗∗
(0.009) (0.017) (0.013) (0.013) (0.016) (0.000) (0.008) (0.008) (0.002) (0.001)

Observations 32,202 31,752 32,202 31,752 31,752 32,202 31,752 32,202 31,752 31,752
Pre-shock mean, dep. var. 0.54 0.54 0.54 0.54 0.54 0.93 0.93 0.93 0.93 0.93
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

60
methampethamines is constant and not affected by how suitable a municipality is. This
provides evidence that the increase in cartel entry into high-suitable municipalities after
2010 is entirely related to the shift in the heroin market and not to any other drug market.

Figure A.2: Kilograms seized by the Mexican military

100

COCAINE
Kgs seized

METHAMPETHAMINES
50

−50
2004 2006 2008 2010 2012 2014 2016
Year

A.6.5 Cartel Competition

The next tables A.12 to A.16 show the coefficients for equation (1) but with the event
time year dummies replaced by a dummy variable P ostShockmt interacted with the suit-
ability index and the heroin price in the United States. The relevant event is again the
2010 OxyContin reformulation. Each table presents the difference-in-difference model for
two dependent variables. The dependent variables are dummies for cartel presence of each
of the nine major organizations. Columns (1) and (6) present the results with municipality
and year fixed effects.Columns (2) and (7) add controls for police presence, military pres-
ence, and political party of the mayor and the governor. Columns (3) and (8) control for
baseline characteristics interacted by year such as population, drug activity in 2000 and
marginalization rate. Columns (4) and (9) add to the baseline trends, the set of controls
from Column (2) and (7). Columns (5) and (10) add state specific time trends. Stan-
dard errors are clustered using a 500 Km radius. Significant at ∗p < 0.1; ∗ ∗ p < 0.05;
∗ ∗ ∗p < 0.01. The results for the cartels expanding, entering and shrinking are robust to
all the controls and fixed effects. The coefficients for the other four cartels are not sig-
nificant for all the specifications, but all the coefficients are positive across specifications.
These results and analysis of entry patterns of the cartels suggest that all of them except
for La Familia expanded into territory well-suited to cultivate opium poppy. Maps of the

61
evolution of each cartel are available upon request.

Table A.12: Sinaloa and Jalisco New Generation

Cartel Activity by Cartel

Sinaloa Cartel CJNG
Post2010 0.107∗∗∗ 0.102∗∗∗ 0.0714∗∗∗ 0.041∗∗∗ 0.03∗∗∗ 4 0.080∗∗∗ 0.089∗∗∗ 0.059∗∗∗ 0.052∗∗∗ 0.039∗∗∗
(0.010) (0.011) (0.001) (0.009) (0.008) (0.001) (0.010) (0.001) (0.008) (0.001)
Observations 32,012 31,562 32,012 31,562 31,562 32,012 31,562 32,012 31,562 31,562
Pre-shock mean, dep. var. 0.0104 0.0104 0.0104 0.0104 0.0104 0.0004 0.0004 0.0004 0.0004 0.0004
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.13: Templar-Knights and La Familia Michoacana

Cartel Activity by Cartel

Knights Templar Cartel La Familia
Post2010 0.022∗∗∗ 0.034∗∗∗ 0.006∗∗ 0.006∗∗ 0.032∗∗∗ −0.004∗∗∗ −0.004∗∗ −0.019∗∗ −0.023∗∗ −0.012∗∗
(0.000) (0.008) (0.0009) (0.0008) (0.007) (0.000) (0.0007) (0.0008) (0.007) (0.004)
Observations 32,012 31,562 32,012 31,562 31,562 32,012 31,562 32,012 31,562 31,562
Pre-shock mean, dep. var. 0 0 0 0 0 0.004 0.004 0.004 0.004 0.004
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.14: Los Zetas

Cartel Activity by Cartel

Zetas
Post2010 0.125∗∗∗ 0.138∗∗∗ 0.104∗∗∗ 0.090∗∗∗ 0.054∗∗∗
(0.0001) (0.014) (0.001) ( 0.012) (0.011)
Observations 32,012 31,562 32,012 31,562 31,562
Pre-shock mean, dep. var. 0.014 0.014 0.014 0.014 0.014
Municipalities FE 3 3 3 3 3
Time FE 3 3 3 3 3
Covariates 3 3 3
Baseline trends 3 3 3
State trends 3
Note:
∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

62
 Table A.15: Juarez and Beltran Leyva Organizations

Cartel Activity by Cartel

Juarez Cartel Beltran-Leyva
Post2010 0.041∗∗∗ 0.046∗∗∗ 0.027∗∗ 0.021 0.008 0.046∗∗∗ 0.063∗∗∗ 0.014∗∗∗ 0.012 0.032
(0.008) (0.009) (0.001) (0.009) (0.009) (0.006) (0.008) (0.001) (0.007) (0.017)
Observations 32,012 31,562 32,012 31,562 31,562 32,012 31,562 32,012 31,562 31,562
Pre-shock mean, dep. var. 0.009 0.009 0.009 0.009 0.009 0.007 0.007 0.007 0.007 0.007
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

Table A.16: Tijuana and Gulf Cartels

Cartel Activity by Cartel

Tijuana Cartel Gulf Cartel
Post2010 0.018 0.019 0.009 0.003 0.011 0.029∗∗ 0.040∗∗ 0.019 0.021∗ 0.023∗∗
(0.009) (0.009) (0.005) (0.004) (0.014) (0.001) (0.007) (0.000) (0.006) (0.007)
Observations 32,012 31,562 32,012 31,562 31,562 32,012 31,562 32,012 31,562 31,562
Pre-shock mean, dep. var. 0.004 0.004 0.004 0.004 0.004 0.005 0.005 0.005 0.005 0.005
Municipalities FE 3 3 3 3 3 3 3 3 3 3
Time FE 3 3 3 3 3 3 3 3 3 3
Covariates 3 3 3 3 3 3
Baseline trends 3 3 3 3 3 3
State trends 3 3
Note: ∗ p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01

63