Employment

Firm employment variables are available quarterly from Mexico’s ENOE, which covers both subordinated formal and informal employment. The dependent variables used in this paper are total firm employment in a locality divided by firm size as self-employment, micro, small, medium and large firms. According to ENOE, self-employment represents around 25% of the total employment in Mexico and subordinated firm employment¹ around 65%. This implies that our study captures around 90% of total firm employment in the country. The remaining 10% of jobs not considered in this paper correspond to workers who do not receive a fixed salary, for example, workers whose payment is based on commissions. On the other hand, metropolitan and non-metropolitan areas are defined following the National Council of Demographic Planning (CONAPO)².

ENOE categorizes subordinated firm employment according to the number of employees and divide firm size into micro (1-10), small (11-50), medium (51-250) and large (250 and more workers). All these estimations are presented by state and metropolitan levels. The labor force in non-metropolitan areas in each state are calculated by subtracting metropolitan values from the total state values.

Table 1 shows how the different categories of firm employment used in this paper are distributed in metropolitan and non-metropolitan areas throughout the country. Total firm employment (non-agricultural) in the country in 2016 was around 42 million people, with approximately 43% (18,210,858) and 57% (23,538,858) in non-metropolitan and metropolitan areas, respectively.

Total firm employment increased by 18% from 2011-Q1 to 2016-Q1. In addition, the table indicates that total firm employment is more concentrated in non-metropolitan areas. Self-employment, micro-firm and small-firm employment each captures 23% of total employment in metropolitan areas, while medium firms and large firms represent 17% and 13% of the total employment, respectively. Medium and large firms are more concentrated in metropolitan than in non-metropolitan areas. In addition, self-employment is the most important labor activity in non-metropolitan areas where it represents 35% of total employment, whereas in metropolitan areas it represents only 23% of the total labor market.

High-impact crime

The high-impact crime variables are collected from the Executive Office of National Public Security (Secretariado Ejecutivo del Sistema Nacional de Seguridad Pública, SESNSP). These official crime statistics are published monthly and then aggregated to create the quarterly observations. Each of the independent variable rates (kidnapping and homicide) measures the number of felonies registered in a locality per 100,000 people during the relevant quarter.

Table 2 shows how crimes are distributed in metropolitan and non-metropolitan areas in the country and how such distribution changed during the period analyzed. Results show that the total crime³ rate decreased by 11.6% and 5.8% in metropolitan and non-metropolitan areas, respectively, during the 2011-Q1 to 2016-Q1 period. While total employment is more concentrated in non-metropolitan areas, the total number of recorded crimes shows an uneven distribution, with metropolitan areas capturing around 60% of the total crime and non-metropolitan areas only 40%. This supports the idea that crime is more concentrated in urban areas. Nevertheless, when homicides are isolated and analyzed, the result is the opposite: 60% of recorded homicides are concentrated in non-metropolitan areas, whereas the remaining 40% are in metropolitan areas. As for kidnappings, the distribution is even, with 50% recorded in each of the metropolitan and non-metropolitan areas in 2011, but, in 2016, it is more concentrated in non-metropolitan areas.

Figure 1 gives more detail about how homicides are distributed throughout the country. It shows that in 2011-Q1, homicides were highly concentrated in the north-western part of the country (dark red), which corresponds to the states of Chihuahua, Durango, and Sinaloa.

Note: Dark red represents two or more times the country average crime rate, red represents above the country average crime rate, but less than two times such crime rate. Yellow refers to below the country average crime rate and white to zero crimes registered in that period.

Source: CONAPO, INEGI Mexican National Public Security Executive Office, Mexican Interior Ministry.

Figure 1 Homicide and Kidnapping Rates, Mexico 2011-Q1 - 2016-Q1 

Furthermore, the state of Tamaulipas, located in the north-eastern part of the country, shows a similarly high concentration of homicides. These dark areas on the map show that homicide rates in those places are twice as higher than the average rate in the country. When homicide rates are compared between 2011 and 2016 maps, 2016 homicides are still concentrated in the same states as in 2011, but with less intensity and were dispersed to other states such as Nuevo León and Baja California Sur. Also in both maps homicides are concentrated in the northern part of the country.

Figure 1 shows that in 2011 kidnapping rates were widely distributed across the country, but slightly more concentrated in the north. The yellow areas represent places where kidnapping rates were below the country average for that period. Then, in 2016, there are few municipalities with values two or more times higher than the country average. It is worth noting that Tamaulipas show the highest kidnapping rates in the country.

Regarding the use of official public crime data, ^{Guerrero-Gutiérrez (2011)} recognizes that an analysis based exclusively on official crime figures poses several shortcomings. These figures are usually published in comprehensive databases only after a long delay; moreover, the figures account exclusively for crimes reported to the authorities. Given Mexico’s law enforcement institutions’ reputation for low performance and pervasive corruption, many citizens choose not to report crimes. Moreover, fear of retaliation from criminal organizations may increasingly prevent victims from filing reports. Due to the uneven performance of law enforcement institutions as well as uneven criminal organization presence throughout the country, these factors may also bias cross-sectional analyses.

Also according to ^{Soares (2004)}, the fraction of the total number of crimes reported to the police varies widely across countries and across different types of crimes. Reported crimes are likely to suffer from measurement error, which thereby introduces a source of bias.

Similarly, ^{Magaloni et al. (2017)} point out this matter for Mexico, arguing that government data might suffer bias from underestimation of case misclassification. In fact, for the crime incidents reported to local public attorney offices and victimization surveys in 2011 the authors estimate that underreporting bias is as high as 87.2% for common crime and 93.2% for extortion.

Additionally, for Mexico, ^{González Andrade (2014)} mentions that the underestimation of crimes is generated by victims who do not report the crime to the authorities or the authorities have not started a preliminary investigation.

While we recognize these limitations, and criminal prosecution and quantification is a complex problem, we use these datasets as our first source of information in keeping with other previous works in this same line of research.

Other variables

Other control variables included in our model are the alternative wage and exchange rate measured in real terms. The alternative wage was is a weighted average of all wages outside the considered locality. Wages are exclusively for the formal sector obtained from INEGI, capturing data from IMSS (Mexican Social Security Institution). The quarterly exchange rate is the average for the analyzed period as reported by the Banco de México website.

The state control variables include a proxy for the real state GDP dynamics variable, real direct foreign investment (FDI), average state education level, and state population. The quarterly indicator for State Economic Activity (ITAEE, 2008=100), available from INEGI, is useful for measuring economic conditions as a proxy for GDP. Following Cabral et al. (2015), we include new FDI reported by the Mexican Ministry of Economy website and available at the state level on a quarterly basis. The average number of school years completed for each state’s labor force, reported by the Ministry of Public Education (SEP), is also included following ^{Partridge and Rickman (2005)} and ^{Mollick and Cabral (2015)}. State population comes from the National Council of Demographic Planning (CONAPO).

Finally, economic activity in Mexico differs greatly across the country’s geographic regions. ^{Cuevas et al. (2002)}, ^{Delajara (2012)}, and ^{Fullerton Jr. (2001)} find that economic performance in northern Mexico is more correlated with US economic activity than with that of southern and central Mexico. In addition, NAFTA has brought Mexico more economic independence between regions and a reduction in the importance of Mexico City (^{Rodríguez-Pose and Sánchez-Reaza, 2005}). To account for these, we add dummy variables quantify fixed factors (north, center and south) and consider the proximity to the US (border).

Unitary changes in crime rate

Our unit of analysis is a locality, defined by metropolitan and non-metropolitan areas in each state in Mexico, i.e. 61 localities in total; we follow employment, crime rates and other variables for each locality quarterly from 2011 to 2016. The selection of state level metropolitan and non-metropolitan areas was in line with previous literature conducted for Mexico and is selected to maximize the information and variance we can extract from the statistics on crime and employment.

The models are estimated in logarithms for most of the control variables, except crime rates. Therefore, the parameters measure elasticities for most of the control variables, whereas the crime rate estimated parameter should be considered as a semi-elasticity.

In each model, we have employment by locality and firm size aggregators, and use the traditional explanatory variables and the different crime rates variable. The model estimations first include all localities (whole country), and are then divided into metropolitan and non-metropolitan, and the dependent variable total firm employment is broken down into firm employment, according to the number of employees per firm.

Endogeneity of crime rates at the local level in a given time period represents one of the biggest challenges for the correct identification of its effect on labor employment. To solve this issue, we followed a two-stage Instrumental Variable (IV) approach where we build a contemporary crime rate (kidnapping and homicide) instrument using its lagged value and the set of exogenous independent variables. Our selection of an instrument is consistent with other works, such as ^{Grossman (1982)}, ^{Hanson’s (2001)} and more recently ^{Coronado and Saucedo (2018)}.

^{Fisher (2010)} describes ^{Wooldridge’s (1995)} alternative endogeneity test, suggested for testing endogeneity of an explanatory variable when using two-stage least squares with robust error model estimation. This test provides both the endogeneity test related to the key hypothesis variable (in our case contemporary crime rates), and provides the properties of the instrument as measured by the R-square values of the first stage regression and the joint F-test of the two equations.

We proceed to estimate Wooldridge’s tests on crime for each model analyzed and decompose the models at aggregate employment, metropolitan areas, and non-metropolitan areas. We report OLS and IV estimations of different versions of equation (1), identifying the effect of homicide and kidnapping rates in Tables 4 and 5, respectively. It is worth noticing that when we analyze total employment, Wooldridge’s test suggests that there is consistent evidence of crime rate endogeneity for both homicide and kidnapping. However, our instrument regression selection provides a valid measure, as shown in both Tables 4 and 5, with adjusted R-square values of first-stage regression of 0.8833 and 0.6102 for homicide and kidnapping, respectively, and joint model F-test with high significance p-values in all cases.

When we open our analysis by metropolitan and non-metropolitan areas, the results suggest endogeneity in the contemporary crime rate, except for homicide in non-metropolitan areas. We followed the same approach instrumenting crime with its lagged value and other control variables, which resulted in a valid instrument, as shown in Tables 4 and 5. For the rest of our analysis, we report the IV model results as the joint F-tests of the two-stage model suggest there is no loss of generality in using this approach, and it provides a direct comparison with the other cases.

Results in Table 3 indicate that no matter which model is used, the homicide rates have a statistically robust negative effect on total firm employment at a country level, but also in metropolitan and non-metropolitan areas. Estimations show that a one-unit increase in homicide reduces total firm employment between 0.05 to 0.15% in all localities. Nevertheless, when the analysis is focused on metropolitan and non-metropolitan areas separately, the IV estimation implies that an increase of 1 unit in homicide would lead to a loss in total firm employment of 0.15% and 0.22%, respectively.

Results also indicate that the alternative wage has a positive effect on state employment. One of the potential reasons for this is that employment data reported in this model come from ENOE, while wages included in the regression come from a government database, which only captures wages in the formal sector.

The real exchange rate, has a positive and statistically significant result in all models, reflecting Mexico’s greater attractiveness to foreign investment and increase in exports; these results are in line with authors such as ^{Faria and León-Ledesma (2005)}, ^{Hua (2007)} and ^{Alexandre et al. (2010)}.

Real aggregate market dynamics (ITAEE) is positive and directly correlated. Results are consistent in all localities and in non-metropolitan areas, as in Gosh (2009), ^{Chiang, Tao and Wong (2015)}, and ^{Klinger and Weber (2015)}.

The new FDI variable has a positive and statistically significant effect on employment only in non-metropolitan areas. Literature that supports this positive correlation for the Mexican economy is ^{Mendoza J. (2011)}, the positive correlation between FDI is with high-skilled employment in the manufacturing industry, but not with all employment in this sector.

In addition, the fixed effects included in the model are statistically significant and indicate the existence of unobserved heterogeneity across geographic regions in the country.

Table 4 follows table 3 regarding the independent variables, but using kidnapping as the main dependent variable. Estimates show that kidnapping has a negative and statistically significant effect on total firm employment when all localities are analyzed together. A one-unit increase in the kidnapping rate decreases total employment between 0.36% and 3.45%. For metropolitan and non-metropolitan areas, the impact of an increase in 1 unit in the kidnaping rate reduces employment by 3.17% and 10.84%, respectively.

Our purpose is to analyze not only the effect of homicides and kidnappings on all firm employment, in metropolitan and non-metropolitan areas, but also the effect of such crimes in different firm sizes. Each of the following regressions has the same independent variables as in previous tables but the dependent variable corresponds to self-employment and employment in micro, small, medium and large firms in all localities, as well as in metropolitan and non-metropolitan areas⁴. All regressions on Table 5 include locality and time fixed-effect control variables.

When firm employment is divided according to work size, the effects of homicide rates on employment are heterogeneous across firm size. On the one hand, findings indicate that when the homicide rate increases by 1 unit, in all localities self-employment, micro and small firm employment drop by 0.12%, 0.23%, and 0.23%, respectively. On the other hand, results in metropolitan areas show that when the homicide rate increases by 1 unit, self-employment decreases by 0.29% and micro firm employment drops by 0.14%. Results also indicate that changes in the homicide rate do not have a statistically significant impact on the employment rate of small, medium and large firms in metropolitan areas. In the case of non-metropolitan areas, estimations show that employment in small firms decreases by 0.51%, whereas all other employment by firm size is not statistically significant.

Regarding kidnapping rate, results indicate that when firm employment is divided according to the number of employees, firms employment responds more sharply to increases in the kidnapping rate than to a similar increase in the homicide rate.

Using IV estimations, results for all localities indicate that when there is a 1-unit increase in the kidnapping rate, small firm employment decreases by 6.82%. In metropolitan areas, the unitary increase in the kidnapping rate would lead to a decrease of 4.99% in self-employment; all these results support ^{BenYishay and Pearlman (2013)} findings. Finally, for non-metropolitan areas, an increase in one unit of this offense induces a 17.61% employment loss in small firms. These large increases due to kidnapping are related to the low probability of a simultaneous kidnapping unitary increase along all the localities given the current variance in each state and locality. Theory dictates that self-employment or small firms are more reactive than medium or large firms when violence increases, as supported by evidence in ^{Amin (2009)}.

Our results on medium and large firms show employment on these firms’ size do not respond to crime rates changes, both for kidnap and homicide. This result is consistent previous research that demonstrates large firms have huge investments in fixed capital and other important costs, which makes it difficult for them to respond elastically when violence increases, in comparison to self-employed or small firms. For instance, ^{Granovetter (1973)} and more recently ^{Brouwer et al. (2004)} find that firms’ mobility incentives decrease with the size of the firm. In addition, ^{McCann (2001)} mentions that relocation costs may be quite significant as firms need to consider, among other factors, the cost of the real-estate site search and acquisition, dismantling, moving and reconstruction of existing facilities, and hiring and training new employees.

^{Kuratko et al. (2000)} mention that small business are 35 times more likely to suffer from business crime than larger firms.

Finally, ^{Perrone (2000)}, in a study for small firms in Australia, finds that certain occupational risks, such as working with the public, working in isolation, working late at night or in the early hours of the morning, being involved in the exchange of money, increase the vulnerability of small firms.

The results of this research are in line with previous literature capturing not only the effect of homicides and kidnappings on total firm employment in Mexico, but also identifying the relative vulnerability of smaller firms to crime.

Employment losses

Our next exercise considers estimating the effect of an increase of 1 standard deviation (SD) in the relevant crime rate by locality. Under normality of a random variable, approximately 68% of its total values are within one standard deviation (higher or lower) from the mean. Therefore, if the crime rate is treated as an exogenous variable, then the range μ ± σ represents where almost 70% where its values occur.

Estimates in Table 6 provide more intuitive results about the effect of different crimes on firm employment. Such numbers are obtained from the homicide and kidnapping semi-elasticities obtained previously. The estimation is performed for each locality and state, 61 in total, and for each type of employment according to firm size and are presented in an average of the percentage point losses in employment, for all localities, metropolitan, and non-metropolitan areas, and across firm sizes.

Using the IV estimation and focusing on the statistically significant coefficients, our results show that an increase in one standard deviation in homicide reduces total employment for all localities by 0.61% and, when estimated by firm size, the employment losses are 0.89% in micro and 0.94% in small firms. The increase in one SD for homicide rates for metropolitan areas induces a reduction in total employment of around 0.59%, but estimations for the self-employment sector and for micro firms are statistically significant, with an average reduction in employment of 0.37% and 0.52%, respectively. Finally, for non-metropolitan areas, an increase in one SD for the homicide rate in each locality reduces total employment by 0.43%, but the effect is mainly on small firms, with a 1.00% reduction in employment for this sector.

On the other hand, IV estimates for the effect of kidnapping on employment is less significant across localities and firm size. In particular, IV estimations show that 1 SD increase in kidnaping crime rate for each locality reduces employment for all localities and all firms by 1.31%, while the effect between metropolitan and non-metropolitan areas is -1.40% and -1.66%, respectively. In this case, the effect for all localities is only significant in small firms, with a decrease of -2.24% in employment. For metropolitan areas, the effect is a reduction of 2.34% in self-employment, while non-metropolitan areas displayed a decrease of 2.69% in small firm employment.

For the second measure of job destruction due to crime, we use these percentage estimations in Table 6 and apply them to the latest employment observations, allowing us to estimate the number of employment losses.

Table 7 shows the point estimations of employment losses due to a one-SD increase in crime rate, divided by crime type, locality and firm size.

Using the IV estimation, we observe that an increase in one SD in the homicide rate in each locality reduces total employment by 560,705, while the reduction is around 256,969 jobs in metropolitan areas and 387,547 jobs in non-metropolitan areas. The estimation of employment losses due to this same increase in homicide crime rates for all localities is 106,307 in micro-firms and 78,532 in small-firms; 64,418 in micro-firms and 21,280 in small firms for metropolitan areas; while, for non-metropolitan areas, the job losses are significant for small firms, at approximately 31,152.

Finally, an increase of one SD in kidnapping rates in each locality has a weaker significant effect on employment losses. In particular, the effect is a reduction in 101,744 job losses for the self-employed in metropolitan areas, and 78,475 jobs in small firms for non-metropolitan areas.

Our main conclusion of this section is, once we control for other relevant variables, employment in medium- and large-sized firms does not respond to crime rates, whereas employment in small, micro, and self-employed are relatively more responsive to increases in these felonies.