This short blog is replicating the regression discontinuity in: Dell, Melissa. 2015. “Trafficking Networks and the Mexican Drug War.” American Economic Review 105(6): 1738-79.

Introduction

Does cracking down on drug trafficking organization inadvertently cause violence to increase?

In December of 2006, the newly elect President of Mexico Felipe Calderon declared war against organized crime and deployed the military to combat these groups. While the motivation behind this action is contested, the expectation was that a crackdown would fragment and weaken criminal organizations enough to prevent them from undermining political institutions through corruption or violence. However, starting in 2007 Mexico witnessed an unprecedented increase in drug-related violence that still plagues the country. But why did Mexico become so violent?

Prevalent theories about why Mexico became so violent after the crackdown on drug cartels in 2006 point to the crackdown itself. These theories argue that the crackdown did fragment criminal organizations, but did not weaken them sufficiently to usurp their power. This fragmentation, rather than leading to weak criminal organizations trafficking drugs in relative peace, intensified competition in the trafficking market between powerful criminal organizations. This surge in illicit market competition–largely caused by the crackdown–led to drastic increases in violence throughout the country.

Given the partisan nature of Mexican politics, it might be reasonable to assume that local elected officials from the PAN party actively participated in the newly-elected president’s crackdown given the pressure to make their party’s new policy a success. Inversely, local officials from opposition parties had little incentive to cooperate and implement the new policy, especially the PRD, who had just lost the presidency by less than 1% of the votes.

If we accept these assumptions, we should expect that municipalities with PAN elected officials cracked down on organized crime starting in December of 2006, while municipalities ruled by opposition parties did not. Additionally, since we have sufficient evidence that increased enforcement against organized crime led to a rise in violence in Mexico over the past decade, we should expect that municipalities with PAN elected officials experienced higher levels of violence than municipalities controlled by the opposition.

Research Design

To test the hypothesis proposed above, we can use a regression discontinuity design to look at local elections and compare levels of violence in municipalities where PAN won versus where the opposition won. We can use an RD design if we assume that the outcomes of closely contested elections resulted from idiosyncratic factors, which in the case of Mexico is a reasonable assumption. By using an RDD however, we discard elections that were not closely contested since they could differ systematically from those that were closely contested. We therefore cannot generalize the results of this research onto elections in general. For this analysis I define competitive elections as those with a vote spread of 5% or less.

Results

To implement the RDD described above, I use data from the Mexican electoral agency (INE) on 152 municipal elections that occurred between 2007 and 2008 where the PAN won or lost by a 5% margin or less. I also use monthly data on municipal homicide rates (homicides per 100,000) from the Mexican Office of the President. Finally, the covariate data comes from the Mexican National Statistical Agency (INEGI).

I first offer the covariate balance table comparing municipalities where PAN won and where PAN lost. The table shows the mean for treated and control groups, along with the standard mean difference and the corresponding p-value. As can be seen, the covariates are fairly balanced.

Covariate Balance (5% vote spread)
PAN won PAN lost Standard means difference P-value
Detour length (km) 26.146 22.478 2.744 0.847
Surface area (km2) 1,787.360 725.375 16.431 0.180
Average precipitation, cm 1,164.399 1,084.634 9.592 0.516
Average max. temperature, C 22.536 22.897 -8.669 0.594
Average min. temperature, C 7.167 7.594 -7.908 0.643
Elevation (m) 1,403.418 1,367.947 4.279 0.799
Slope (degrees) 3.621 3.186 16.232 0.307
% alternations (1976-2006) 0.310 0.315 -3.006 0.853
PAN incumbent 0.271 0.317 -10.191 0.541
PRD incumbent 0.171 0.134 9.821 0.529
PRI never lost (1976-2006) 0.071 0.073 -0.672 0.967
Marginality index (2005) -0.154 -0.119 -3.865 0.816
Population (2005) 6.026 5.099 4.810 0.727
Income per capita level (2005) 4.285 4.483 -8.281 0.595
Mun. taxes per capita (2005) 59.845 56.746 3.221 0.821
Population density (2005) 220.232 191.051 6.852 0.677
Migrants per capita (2005) 0.016 0.018 -11.533 0.490
Malnutrition (2005) 32.760 31.198 8.735 0.599
Mean years schooling (2005) 6.262 6.186 5.321 0.746
Infant mortality (2005) 22.538 22.263 3.969 0.826
HH w/o access to sewage (2005) 8.505 8.436 0.845 0.958
HH w/o access to water (2005) 16.141 18.221 -10.419 0.535
Distance U.S. (km) 708.270 735.487 -8.412 0.586
Road density (km/km2) 0.151 0.131 14.157 0.329
Observations: Total = 152; PAN won = 72; PAN lost = 80.


Given that the treatment and control groups are balanced, I now analyze the effects of elections on violence.

If local crackdowns are affecting levels of violence, I expect that these effects would only occur after the inauguration of the new municipal president and not before that. I therefore provide results of close elections on homicide rates during four different time periods–before the elections, during the lame duck period, the six months following the inauguration, and the entire mayoral term.

For each time period I present the results of a linear model and a quadratic model, with each observation weighted by the population size of the municipality since homicide reporting has a lower variance in urban areas (this is standard practice in the crime literature). I also provide graphs visualizing these results using the quadratic model.

Results:

Close PAN elections and Homicide rates: 2007-2008 Elections
Homicide rate
Before elections Lame duck period First six months in office Entire term
(1) (2) (3) (4)
PAN win (linear) 3.088 2.457 17.133*** 56.630***
(4.873) (3.044) (5.812) (15.536)
PAN win (quadratic) -1.951 4.442 14.200*** 68.550***
(6.703) (3.770) (5.502) (14.745)
Observations 152 152 152 152
Note: p<0.1; p<0.05; p<0.01

Robustness check

Unfortunately the dataset only includes observations within the 5% spread so I’m unable to test the results with a wider bandwidth. However, below I provide the results of an RD with the smaller bandwidth of 0.03. As can be seen, the results hold.

Close PAN elections and Homicide rates: 2007-2008 Elections (bw of 0.03)
Homicide rate
Before elections Lame duck period First six months in office Entire term
(1) (2) (3) (4)
PAN win (linear) 2.200 3.725 17.237*** 68.026***
(5.190) (2.291) (5.411) (14.229)
PAN win (quadratic) -4.500 0.817 8.725 64.872**
(5.720) (4.071) (5.490) (26.722)
Observations 130 130 130 130
Note: p<0.1; p<0.05; p<0.01


I also conducted two placebo tests using the bandwidth of 0.03, but moving the cutoff 0.02 degrees left or right. Ideally, there should not be jumps at the new cutoffs, but there are, so I don’t report the results. However, this is likely due to the relatively small shifts in cutoffs I’m constricted to with the dataset.

Potential issues

There are two major potential issues that I want to explore:

  1. Close elections might affect homicide rates in municipalities where drug cartels operate, but not in municipalities where cartels do not operate.

  2. If the results are driven by a few observations, and these observations are where the military intervened, municipal elections might not be the source underlying the spike in violence.

Cartel presence

The first issue is that drug cartels do not operate in every municipality in Mexico. How could a close PAN victory at the municipal level increase violence as a result of a crackdown on drug cartels if there are no drug cartels to crack down on?

Thus, close elections might have an effect in municipalities where drug cartels operate, but not where they don’t.

To test this hypothesis, I use a dataset that determines what municipalities cartels operate in using web content (Coscia and Rios 2012). Using this dataset, I run two separate RDs, one with municipalities that had cartel presence in the 2007-2008 period, and another where they did not.

The results, shown below, confirm my intuition: close elections in municipalities where cartels operate have a significant and positive effect on homicide rates, but have no effect in municipalities where cartels do not operate.

Close PAN elections and Homicide rates: 2007-2008 Elections
Homicide rate
Cartel presence No cartel presence
Before inauguration After inauguration Before inauguration After inauguration
(1) (2) (3) (4)
PAN win (linear) 2.573 63.181*** -1.218 -7.031
(5.987) (13.356) (3.310) (8.238)
PAN win (quadratic) -4.648 69.027*** 2.748 9.703
(7.715) (12.228) (5.483) (12.286)
Observations 47 47 105 105
Note: p<0.1; p<0.05; p<0.01

Militarization

A second potential issue might be that the main results are driven by a handful of observations.

Here I explore this possibility:

The table below gives a summary of the homicide rates before and after the inauguration, with the plots under it visualizing the homicide rate densities and homicide rates before versus after inauguration, with the diagonal line representing no change.

The density plots and the first scatterplot show that there’s clearly a small number of observations whose homicide rates are considerably higher than the rest. The second scatterplot re-scales the axes to see the bottom cluster of observations better and reveals homicide rates having no obvious pattern of a partisan divide.

Homicide rate summary
Min. 1st Qu. Median Mean 3rd Qu. Max
Before inauguration 0 3.72 7.82 12.01 14.43 132.34
After inauguration 0 3.89 8.74 26.98 28.4 263.05

Having suggestive evidence that the results might depend on a few observations, I list the top 10% most violent municipalities post-inauguration below. Interestingly, 11 of the 15 most violent municipalities are in the state of Chihuahua.

PAN win Homicide rate Municipality State
1 0 263.05 Aquiles Serdán Chihuahua
2 0 255.08 Matamoros Chihuahua
3 0 200.51 Santa Isabel Chihuahua
4 1 158.83 Moris Chihuahua
5 1 152.29 Satevó Chihuahua
6 1 148.03 Uruachi Chihuahua
7 0 125.21 Guachochi Chihuahua
8 0 124 El Oro Durango
9 1 105.05 Temósachic Chihuahua
10 1 96.54 Cuajinicuilapa Guerrero
11 1 92.59 San Dimas Durango
12 1 90.95 Chihuahua Chihuahua
13 1 89.04 Hidalgo del Parral Chihuahua
14 0 83.02 Bachíniva Chihuahua
15 0 82.63 Otatitlán Veracruz


After further research, I found that the federal government began intense military operations in Chihuahua in 2007 as a response to heavy cartel presence (officially “Operation Chihuahua”). These operations are likely the primary reason why these municipalities experienced high levels of violence–all in the highest quartile–not a local crackdown. Below is a table of Chihuahua’s municipalities in the sample.

Importantly, only seven of the sixteen municipalities were won by the PAN–further evidence that violence was not driven by local PAN crackdowns.

PAN win Homicide rate Municipality State
1 0 263.05 Aquiles Serdán Chihuahua
2 0 255.08 Matamoros Chihuahua
3 0 200.51 Santa Isabel Chihuahua
4 1 158.83 Moris Chihuahua
5 1 152.29 Satevó Chihuahua
6 1 148.03 Uruachi Chihuahua
7 0 125.21 Guachochi Chihuahua
8 1 105.05 Temósachic Chihuahua
9 1 90.95 Chihuahua Chihuahua
10 1 89.04 Hidalgo del Parral Chihuahua
11 0 83.02 Bachíniva Chihuahua
12 0 74.87 Coronado Chihuahua
13 1 72.66 Guerrero Chihuahua
14 0 49.02 Namiquipa Chihuahua
15 0 39.54 Delicias Chihuahua
16 0 28.32 Julimes Chihuahua


Assuming that violence in Chihuahua was caused by the heavy military intervention, I run an RD without these observations (10% of the sample). The effects of close elections becomes insignificant once these observations are excluded. These reults suggest that, at best, Dell overestimates the effects that local crackdowns have on homicide rates.

Close PAN elections and Homicide rates: 2007-2008 Elections
Homicide rate
Before elections Lame duck period First six months in office Entire term
(1) (2) (3) (4)
PAN win (linear) 5.254 2.382 18.117* 36.269
(5.495) (4.747) (10.512) (27.571)
PAN win (quadratic) 7.210 12.089 21.304* 39.858
(7.859) (8.099) (12.690) (29.455)
Observations 136 136 136 136
Note: p<0.1; p<0.05; p<0.01

Conclusion

Did close local elections where the PAN party won lead to an increase in violence because they cracked down on drug cartels?

Dell’s analysis is suggestive that local elections did impact homicide rates. Yet, as I have shown, two factors weaken Dell’s evidence.

While I cannot conclude with certainty that Dell’s results are biased upward, I provide suggestive evidence that they likely are.

References

Coscia, Michele and Viridiana Rios. 2012. “Knowing Where and How Criminal Organizations Operate Using Web Content.” ACM International Conference Proceeding Series. 1412-1421. 10.1145/2396761.2398446.

Dell, Melissa. 2015. “Trafficking Networks and the Mexican Drug War.” American Economic Review 105(6): 1738-79.