My sample consists of all individual between 15 and 80 years old, successfully con-tacted in the three MxFLS waves.9 For this reason, the usual age range of 15 to 65 has been extended to include those up to 80 years old. I allow working and non-working re-spondents in the sample considering that information about their decision to eventually join the labor market can be relevant for the results.10 From the MxFLS1 baseline sam-ple of 35,000 respondents, I excluded individuals who died between the survey waves, individuals who migrated to the United States, and respondents that did not continue the follow-ups for the subsequent surveys. In MxFLS2 and MxFLS3, new household members joined the surveys. I discarded these new members from the sample since they were non-panel respondents in the MxFLS1 or MxFLS2. After deleting observations with missing or dubious information, 10,928 male and female individuals and a total of 32,784 observations remained.
State-level crime rates (total crime, homicide, robbery, and aggravated assault) were assigned to each individual in the MxFLS sample, subject to the individual’s state of residence and the MxFLS interview date. Interviews were carried from mid to end 2002, late 2005 to mid-2006, and mid-2009 to late 2012. Due to expected delays in the response to violence, I assign crime rates from a year before the MxFLS interview date. From a total of 32 states in Mexico, the MxFLS baseline was conducted in 16 of them.11 Including respondents that changed their state of residence for subsequent MxFLS follow-ups, a total of 22 states are covered in my sample. Conducting this study at a more disaggregated level of analysis, such as at the municipality-level, would provide more accuracy about local variations of the impact of violent crime on labor decisions; however, criminal activity data is only available at the state-level of analysis.
4.4 Descriptive Statistics
Table 1 shows the descriptive statistics for the sample. I examine three primary la-bor dependent variables: worked last year, which is a dummy variable equal to 1 if the respondent worked during the 12 months before the interview, and zero otherwise;
the average number of weeks worked during the 12 months before the interview; and the average number of hours worked per week during the 12 months before the
in-9Financial incentives that allow worker’s retirement are an important determinant on the decision to stop participating in the labor force, however, in Mexico the pension system is less generous than in developed countries. Because of the lack of income and financial resources, elderly individuals are more likely to continue working (Van Gameren, 2008).
10In correspondence to the labor force participation definition.
11The 16 states included in the MxFLS baseline are Baja California Sur, Ciudad de Mexico, Coahuila, Durango, Guanajuato, Jalisco, Mexico, Michoacan, Morelos, Nuevo Leon, Oaxaca, Puebla, Sinaloa, Sonora, Veracruz, and Yucatan.
‧
terview. Here, the variable worked last year does not denote labor force participation since respondents who said they did not work in the past 12 months are not necessarily looking for work. From the total sample, about 51% of the individuals living in states with below-average crime rates have worked during the 12 months before the interview.
Likewise, 50% of the respondents living in states with crime rates higher than the mean have been working during the year prior to the survey. The average number of weeks worked during the last year is 24 for individuals in low-crime states and 25 weeks for individuals in high-crime states. The average number of hours worked per week in low-crime states is 22, while in high-crime states is 23.
I include individual, household, and state-level controls in the model. Individual-level controls are age, gender, marital status, years of education, whether or not the respondent is the head of the household, whether or not the head of the household head is female, and whether or not the individual has relatives in the United States.
Table 1 suggests that individuals living in states with a higher total crime rate are, on average, more educated, more likely to have relatives in the U.S. and, less likely to be informal workers (Vel´asquez, 2020). This pattern emphasizes the importance of using longitudinal data that evidences the diverse characteristics of individuals living across different locations. As a macroeconomic independent variable, I use the annual GDP per capita at a state-level to control for any possible effects of the 2007-2008 financial crises on the labor outcomes. Data suggests that states with higher GDP are also states with higher total crime rates.
Additionally, the individual’s perception of violence is registered in a series of ques-tions in the MxFLS. I use as central variables whether or not the respondent feels fear of being assaulted, whether or not they feel fear of being assaulted at night, and whether or not they feel more unsafe going out at night than five years ago. Table 1 shows that individuals’ fear dramatically increased between the second and third waves of the MxFLS in specific states, supporting the notion that the increases in violent crime were unanticipated and unrelated to previous trends.
5 Empirical Strategy
This study aims to understand individuals’ labor responses due to exponential increases in violent crime caused by the Mexican drug war. By using a longitudinal survey com-paring the same respondent over periods of low and high violence intensity, I adopt an individual fixed-effects model as the main empirical strategy and include temporal and geographical crime rate variations between 2002 and 2012. The underlying intu-ition for using an individual fixed-effects strategy is that respondents can act as their
‧
own counterfactual by comparing their labor decisions during a low-violence period to their labor decisions while experiencing high violence levels. Hence, the premise is that each respondent’s labor behavior before the Mexican drug war implementation is a fair counterfactual for what their behavior would have been after 2007 in the absence of an increase in violent crime.
Moreover, identifying the impact of violent crime on labor market outcomes is not straightforward since confounding factors can interfere in a simple analysis of the rela-tionship. For instance, increases in violent crime rates usually coexist with other events that also affect the local economy and the labor market behavior; these events may alter an individual’s labor decisions even in the absence of any actual causal effect of violent crime on labor choices.
Furthermore, an important factor to consider is the omitted variable bias that may arise from not considering the effect of unobserved, heterogeneous characteristics of states and individuals on labor decisions and their correlation with the violence trend.
These unobserved characteristics, such as the economic performance of a state (and other confounding factors), may directly affect an individuals’ labor decisions. At the same time, a states’ violence trend may be dependent on its economic performance.
Such omitted variable bias is an inconvenience since it can lead to underestimating or overestimating the impact of violent crime on labor decisions.
To correctly identify the effects of violent crime on labor outcomes, I use a fixed-effects model as an empirical specification. This strategy ensures cleaner findings by eliminating unobservable state and individual’s time-invariant characteristics that may create endogeneity. I estimate the individual fixed-effects strategy in the following re-gression:
yi jt = βVjt+ γ0Xi jt+ δ0Zjt+ θt+ λj+ µi+ εi jt (1)
Individuals are indexed i = 1, . . . , N. They are observed once per period t = 1, ..., Ti in state j = 1, ..., J. yi jt is the outcome of interest, and it refers to the labor market behavior of individuals who already were of working-age (15-80) during the MxFLS1, expressed as a binary variable which indicates whether or not the respondent worked during the 12 months before the interview. The other two outcomes refer to the number of weeks worked by an individual over the last 12 months, and the average hours worked during a week over the last 12 months.
The violence variable is represented by Vjt which denotes the logarithm of the total violent crime rate at a state-level 12 months before the MxFLS interview date. Xit is a vector of individual time-varying characteristics that includes marital status, years of education, household status, and whether or not the respondent has relatives in the
‧
United States. Zjt is a vector of state time-varying characteristics conformed by the annual GDP at a state level. θt captures the year fixed-effects, λj captures the state fixed-effects, and µicaptures the individual fixed-effects. If the time, state or individual fixed-effects are correlated with the violence covariates Vjt, then omitting θt, λj or µi will lead to a biased estimate of the violence effects, β .
The proposed fixed-effects model allows to eliminate bias from unobservables that change over time but are constant in states and individuals (θt), and it controls for factors that differ across states and individuals but are constant over time (λj and µi, respectively). The year fixed-effects θt captures unobserved temporal situations, such as economic events, that occurred during some but not all of the studied years; these unobservables that change over time would affect all individuals in all states. The state fixed-effects λjcaptures the unobserved factors in a given state that affect its residents’
labor decisions; these factors are time-invariant during the studied years. The individual fixed-effects µicaptures unobserved inherent characteristics of an individual that influ-ence their labor decisions; these do not change over time and also do not change with the state of residence.
The abrupt change in the violence trend after the Mexican drug war across states is crucial to establishing a relationship between violent crime and labor decisions. My identifying assumption proposes that in the absence of the Mexican drug war strategy, each individual’s labor decisions would be constant over time (common trends assump-tion). Therefore, I assume the crime rate trend to be random and not correlated to any unobserved events; following the fixed-effects assumptions, I consider extreme exo-geneity so that the idiosyncratic error εi jt is not correlated with changes in the crime rate or with any of the explanatory variables.
The fixed-effects model appears to be a useful tool if we think that there are unob-served variables such as individuals’ inherent abilities or the states’ characteristics that affect all its residents, and unobserved variables such as economic events on the same date as the MxFLS year. After controlling for individual, state, and year fixed-effects, and observable time-varying individual and state characteristics, I can accurately mea-sure violent crime’s causal effect on labor decisions.
Finally, a factor to consider when estimating the effects of violence on labor out-comes is the possible systematic responses to violent crime. Given the increase in violence in certain states, it is reasonable to think that some individuals will choose to migrate to less dangerous regions. In this sense, non-random migration can be a potential problem since it would affect the estimation of violent crime’s effect on labor decisions. To deal with potential selective migration, I use an intent-to-treat approach.12
12It is difficult to assume that selective migration can be completely ruled out from the analysis. Even if violence escalation affects migration decisions, destination choices may be related to unobserved factors.
‧
For the cases where respondents migrate after the violence escalation in 2007, I assigned the total violence rate based on the state of residence in the first wave of the MxFLS rather than the current state. By fixing respondents to their location before the violence increase, the correlation between increases in the crime rate and migration should not impact the respondent’s assigned total crime rate.
6 Results
In this section, I first estimate the effects of violent crime on labor decisions using an individual fixed-effects model and differentiating by gender and formality status using the same strategy. After that, I repeat the analysis by age group. Following this, I explore the potential mechanisms triggering the causal effect of violent crime on labor market outcomes. Finally, I present a leading-values test to support my empirical design.
6.1 Main Results
Table 2 displays the results of analyzing the impact of the total crime rate on labor market outcomes.13 Panel A includes as a dependent variable the participation of the respondent on the labor market for the 12 months before the interview, Panel B displays the number of weeks worked for the 12 months before the interview, and Panel C shows the average number of hours worked per week for the 12 months before the interview.
To illustrate the relevance of controlling for individual and regional heterogeneity, I first present the results of a model without fixed-effects. Column (1) contains a simple OLS estimates, where the estimate treats λj and µiare part of the error term. Column (2) contains both time and state fixed-effects, and Column (3) includes the full set of state, time, and individual fixed-effects. Estimates for the model with the full set of fixed-effects and socio-demographic controls are shown in Column (4).
The OLS estimates in Column (1) suggest an insignificant effect of the total crime rate on the labor market outcomes for the three panels. Since the strict exogeneity as-sumption for the OLS estimator is different from that on the fixed-effects estimator, these findings are unsurprising. The model probably fails to account for individual and regional heterogeneity, inaccurately estimating the impact of the total crime rate on
la-If this is the case, it will lead to assigning an improper total violence rate to individuals that migrated;
this situation could bias my analysis if I do not use an intent-to-treat approach.
13Estimates for the impact of violent crime on labor market outcomes disaggregated by crime type are shown in Table 3. Overall, results from the effects of homicide, robbery, and aggravated assault rates separately do not show significant impacts on labor outcomes.
‧
bor market outcomes. Columns (2) to (4) include fixed-effects estimations in stages, first adding state and time fixed-effects and then individual fixed-effects. Columns (3) and (4) that include individual fixed-effects share similar results. Both models indicate a negative impact of the total crime rate on the labor market outcomes. However, the intensity of the effects is slightly higher in Column (3), a model without the socioeco-nomic controls.
Column (4) shows the estimates for Equation 1. Results from Panel A indicate that the decision to participate in the labor market is negatively affected by increases in the total crime rate. From 2000 until the drug war implementation in December 2006, there was a reduction in the total crime rate of 14%. After 2006, however, Mexico experienced an increase in the total crime rate of approximately 90%. Column (4) es-timates suggest that an individual experiencing the drug war violence would decrease their probability of joining the labor market by 2.25 percentage points, contrary to the increase of 0.35 percentage points in their likelihood of being employed before the drug war implementation. Similarly, Panel C shows a negative impact of violence on the av-erage hours worked per week. Considering the actual increase of 90% in the total crime rate, individuals will experience a reduction of 1.6 hours worked per week, contrary to the addition of 0.25 hours worked per week that an individual would experience before the drug war.