Introduction
The United Nations (UN) Security Council recently convened to discuss sustainable peace, as six out of seven countries worldwide report feeling insecure (United Nations 2023). In forming this New Agenda for Peace, Ecuador’s representative to the UN emphasized conflict prevention with mechanisms to overcome organized transnational crime.Footnote 1 Multilateralism was proposed by the representative of Morocco, Omar Kadiri, to combat transnational crime – negotiation among several nations. In this context, we examine the research efforts of the United States (US) and Mexico to reduce crime, funded by the United States Agency for International Development (USAID) and the National Crime Prevention Programme (PRONAPRED) in Mexico.
Mexico has been a leader in rising violence over the last decade (Portillo Reference Portillo2021) in one of the most violent regions (Bergman Reference Bergman2018). Latin American countries with high levels of criminality, like Mexico, maintain an unstable equilibrium where violence spirals out of control after a breakdown of the status quo (Bergman Reference Bergman2018). In Mexico’s case, the well-known war against organized crime (specifically cartel-related crime) beginning in 2007 represents that breakdown. The rise in homicides associated with this war has captured the attention of society (Martinez-Palomares Reference Martinez-Palomares2019) and given rise to our research into homicides over time and national efforts to reduce homicides like PRONAPRED.
President Enrique Peña Nieto launched PRONAPRED in 2011, and criticism for PRONAPRED and similar large-scale crime prevention programmes have been ever-present in journalistic and academic literature alike (Banerjee, Banerjee, and Duflo Reference Banerjee, Banerjee and Duflo2011; Easterly Reference Easterly2009; Sumano Rodríguez Reference Sumano Rodríguez2018). However, Esquivel Hernandez (Reference Esquivel Hernandez2015) suggests that large-scale spending projects in Mexico focused on basic infrastructure can have significant economic and social impacts. In a global context, Sachs (Reference Sachs2006) highlights successful international development programmes that have eradicated disease and fought poverty. On the point of crime reduction, Bondurant, Lindo, and Swensen (Reference Bondurant, Lindo and Swensen2018) have shown that substance abuse treatment facilities reduce violence, with a particularly pronounced reduction in serious crimes. In this paper, we argue that there has been a beneficial but unsustainable reduction in homicides due to national crime prevention spending in Mexico.
Figure 1 shows average homicides in Mexican states over time and a significant downturn associated with PRONAPRED spending in 2013. The upward trend in homicides after 2006 relates to a crackdown on drug cartels by the President Calderón administration.Footnote 2 The vertical line in Figure 1 indicates a noticeable downturn in homicides, visible in and after 2012, which we argue corresponds to the decline in PRONAPRED and USAID spending.Footnote 3
Figure 1. Average number of homicides per year, 1990–2020.
Previous research into the effect of PRONAPRED spending has not shown such beneficial results. Ramirez-de Garay and Diaz Roman (Reference Ramirez-de Garay and Diaz Roman2017) studied the evolution of violence prevention policies in Mexico and the effects of PRONAPRED funding on homicide rates and non-fatal shootings from 2012 to 2014.Footnote 4 By applying propensity score matching to data at the municipal level, they found that PRONAPRED funding significantly affected only the homicide rate but in an unexpected direction, such that a rise in programme funding was associated with a rise in homicides. Hernández (Reference Hernández2013) analysed the PRONAPRED programme based on the analytical framework of governance proposed by Hufty (Reference Hufty2009) in the municipality of Ecatepec de Morelos and concluded that PRONAPRED did not contribute to the reduction of the crime rates.Footnote 5 However, these studies did not consider variation in homicides explained by differences between Mexican states. We show that a large amount of variation in homicides is simply due to geographic location in addition to homicides from the previous year. Our analysis includes a fixed-effect regression technique to account for any unexplained variation in homicides between states.
In addition to failures in outcome, many studies have identified technical issues with PRONAPRED, including difficulties in identifying locations to receive funding, PRONAPRED’s duplication of services provided by other government institutions and the evaluation of programme effectivenessFootnote 6 (Merino and Torreblanca Reference Merino and Torreblanca2017; México Evalúa 2014; Ramirez-de Garay and Diaz Roman Reference Ramirez-de Garay and Diaz Roman2017; Román and González Reference Román and González2019). Román and González (Reference Román and González2019) showed design and evaluation deficiencies within PRONAPRED during the presidential administration of Enrique Peña Nieto from 2012 to 2018. It is also necessary to point out that legally and administratively within PRONAPRED, there were issues such as the duplication of PRONAPRED-provided functionsFootnote 7 with the National Center for Crime Prevention (Ramirez-de Garay and Diaz Roman Reference Ramirez-de Garay and Diaz Roman2017). Additionally, there was a lack of clear guidelines to define PRONAPRED beneficiaries, which remained uncorrected (Ramirez-de Garay and Diaz Roman Reference Ramirez-de Garay and Diaz Roman2017). Finally, according to Ramirez-de Garay and Diaz Roman (Reference Ramirez-de Garay and Diaz Roman2017), PRONAPRED requires greater theoretical elaboration for spending decisions and the design of empirical observables. Although we find beneficial effects from PRONAPRED spending, we highlight the many failures to provide a more balanced picture of crime prevention.
We estimate a model of homicides based on evidence of a natural rate of crime and the influence of economic factors. Narayan, Nielsen, and Smyth (Reference Narayan, Nielsen and Smyth2010) have suggested a natural rate of crime based on the existence of unit roots in crime series for the United Kingdom, the US and other G7 countries.Footnote 8 Marché (Reference Marché1994) modelled a production function for homicide solutions, such that factors like the availability of evidence and investigator experience determine the number of solved homicides. Ferguson and Smith (Reference Ferguson and Smith2021), Ouimet (Reference Ouimet2012), Baron and Straus (Reference Baron and Straus1988), Pratt and Lowenkamp (Reference Pratt and Lowenkamp2002), Rosenfeld (Reference Rosenfeld2009) and Juárez, Urdal, and Vadlamannati (Reference Juárez, Urdal and Vadlamannati2020) have suggested a relationship between homicides and economic factors; however, results typically depend on the form of analysis (cross-lagged panel v. time-series or developed v. developing country analyses). We justify our estimation of homicides in Mexico by state over time in greater detail in the Methods section.
In addition to testing the hypothesis of fewer homicides due to programme spending, we test a common hypothesis in criminology literature of gender inequality and homicides. Moore, Heirigs, and Barnes (Reference Moore, Heirigs and Barnes2021) suggested that increased gender inequality in the US predicts greater homicide numbers. Heirigs and Moore (Reference Heirigs and Moore2018) found a similar relationship between homicides and a gender inequality index across 94 countries. Conversely, Santos, Jacobson, and Georgiev (Reference Santos, Jacobson and Georgiev2021) found a linear, positive relationship between labour market integration and homicide rates in low-income countries.Footnote 9 We find that the positive relationship between gender equality (as measured by rates of male and female labour participation) and homicides is consistent in Mexico.
The main contributions of this paper are measuring the effectiveness of a national crime prevention programme and identifying the relationship between male and female labour participation and homicides. Our analysis builds on previous work by Abt and Winship (Reference Abt and Winship2016), Abt et al. (Reference Abt, Blattman, Magaloni and Tobón2018) and Hernandez Rodriguez (Reference Hernandez Rodriguez2020) to suggest the effectiveness of community-based prevention programmes, which have been scaled up for individual Mexican states and Mexico as a whole. However, as discovered by Donoso Jiménez and Olivera Salado (Reference Donoso Jiménez, Olivera Salado and Valenzuela Aguilera2019), Román and González (Reference Román and González2019), México Evalúa (2014), Ramirez-de Garay and Díaz Roman (Reference Ramirez-de Garay and Diaz Roman2017), Merino and Torreblanca (Reference Merino and Torreblanca2017) noted that the impacts of such programmes in Mexico were also unexpected and unclear. Furthermore, this work contributes to the ongoing discussion by Sachs (Reference Sachs2006), Easterly (Reference Easterly2009) and Banerjee et al. (Reference Banerjee, Banerjee and Duflo2011) regarding the most effective approach to development.
Data
The two main independent variables of interest are PRONAPRED and USAID programme spending. PREPARED spending is allocated to 69 municipalities, seven metropolitan areas, five sections of Mexico City and one conurbation area in Mexico.Footnote 10 USAID spending is similarly allocated to states, municipalities and cities.Footnote 11 Although spending flowed to municipalities, metropolitan areas, cities and sections of a city, the data for this project are organized by state and year for ease of analysis with other variables that are not organized by such locations.
We allocate USAID programme spending according to the programme’s length, location and funding amount specified in the spending schedules provided by the US consulate in Mexico. The actual year-to-year spending per programme per location is not known. Therefore, to construct the USAID spending variable, we take the total amount of money allocated to the programme, divide it by the total number of states receiving the money, and then divide it by the programme’s duration in years. This value is the average USAID spending per Mexican state per year.Footnote 12 Moreover, we repeat the process for all of the USAID programmes listed in the included Excel file.
The two spending variables are adjusted for inflation, and then PRONAPRED spending is converted from Mexican pesos to US dollars (USD). To adjust the spending variables for inflation, we use the US and Mexican Consumer Price Index (CPI) data from the Organization for Economic Co-operation and Development (OECD 2022). We use 2015 as the base year (CPI = 100) for both spending variables. Then, we multiply the spending variable by 100/CPI to adjust for inflation. The PRONAPRED spending is then adjusted to the USD using Federal Reserve Economic Data for the Mexican peso to the USD spot exchange rate (Federal Reserve Bank of St. Louis 2022). These data are the average of daily spot exchange rates for each year from 1990 to 2020 (Federal Reserve Bank of St. Louis 2022). We find the inverse of the Mexican peso to USD (1/(USD/Mex)) to convert USD to Mexican pesos. Then, we multiply by the inflation-adjusted PRONAPRED spending. Both spending variables are in USD and divided by 1 million to put the units in millions of USD. Finally, the spending variables are lagged by one year to address causality with the dependent variable – the number of homicides per state per year.
The homicide data are from the National Institute of Statistics and Geography in Mexico, also known as the Instituto Nacional de Estadística y Geografía (INEGI) (Instituto Nacional de Estadística y Geografía 2022). The data consist of the number of homicides as a total, male homicides and female homicides. The main analysis for this paper uses the total number of homicides to build upon previous research; however, we also consider an analysis by gender in Tables D.8 and D.9 in the Appendix. The dependent variable is the number of homicides, and we include an independent variable of homicides in all other Mexican states to account for trending waves of crime across the country over time.
The summary statistics in Table 1 show greater variation in homicides between Mexican states than over time. The wave of homicides across Mexico is measured by the variable “Homicides in other states”, and it is an average of homicides in the other Mexican states (not in state i). A time indicator measures the rise in homicides within each state over time. An interaction between the time indicator and a dummy variable for years after 2007 identifies the specific rise in homicides related to President Calderón’s order for military and police intervention to break up drug cartels throughout Mexico. We also use homicides and population growth in other states as external variables in order to address possible endogeneity between homicides and programme spending in the Appendix (Table F.11).
Table 1. Summary statistics
Note: PRONAPRED, National Crime Prevention Programme; USD, United States dollars; USAID, United States Agency for International Development.
Methods
The main fixed-effects estimation for homicides by Mexican state and year is:
${H_{i,t}} = \alpha + {\gamma _1}{t_i} + {\gamma _2}{(t \gt 2007)_i} + {\beta _1}Spen{d_{i,t - 1}} + {\beta _2}{H_{i,t - 1}} + {\beta _3}H_{i,t - 1}^j + {\varepsilon _{i,t}}$
The dependent variable is the number of intentional homicides in each state, i, and year, t, using homicide data from INEGI in Mexico. We include time indicators, t and (t>2007), to predict homicide changes per year for the entire sample and specifically during President Calderón’s national crackdown on drug cartels after 2007. The main independent variable, Spend, is either PRONAPRED or USAID spending by Mexican state and year.Footnote 13
We show that the results from fixed-effects ordinary least squares (OLS) are preferable to pooled OLS because of a large degree of variation in homicides that is explained by variation in state-specific characteristics.Footnote 14 The control variables include the number of homicides in all other Mexican states and time indicators for the entire sample and years after 2007 by Mexican state. Our secondary analysis utilizes economic indicators of male and female labour market participation, limited to the years 2005–2020.
Two assumptions of the least squares regression model are that: (1) residuals are independent of each other; and (2) residuals vary constantly from their mean over time. Suppose homicides in a Mexican state are related over time. In that case, an estimation of homicides will not produce random errors with a constant mean of 0 and a constant standard deviation of 1. This relationship over time is known as autocorrelation, and the autocorrelation in homicides will cause the mean and variance to change over time rather than remain constant. If our regression model does not account for the autocorrelation, then the residuals will not be independent of each other nor have constant variance. In this case, the assumptions of linear regression would be broken, and our results would not necessarily be reliable.
Therefore, we test for the condition of stationarity in our dependent variable, the number of homicides in a Mexican state in a given year, and determine that we cannot reject the presence of unit roots. These findings are consistent with Narayan et al. (Reference Narayan, Nielsen and Smyth2010) and suggest a natural rate of crime over time. Therefore, we include a lag of homicides as an independent variable to capture prior explanatory information on homicides. We then estimate equation (1) and finally test the predicted residuals for stationarity. The predicted residuals from equation (1) are stationary as we reject the null hypothesis of a unit root.Footnote 15
Finally, a likely criticism of our model is that programme spending endogenously relates to homicide levels. For instance, USAID or PRONAPRED coordinators may see that homicides are high or rising in a Mexican state and increase the programme spending in that state. While there is no specific evidence to suggest this,Footnote 16 we still address this potential endogeneity in the Appendix (Table F.11) using two-stage least-squares (2SLS) analysis. Ultimately, the root mean squared error found in the 2SLS model was larger than the OLS fixed-effects model highlighted above. Consequently, we relegate the in-depth analysis of endogeneity to the Appendix and acknowledge that it is an excellent avenue for future research.
In Mexico, we hypothesize a relationship between homicides and economic indicators simply because Mexico is a middle-income country. Previous research has yielded mixed results in the US (Baron and Straus Reference Baron and Straus1988; Pratt and Lowenkamp Reference Pratt and Lowenkamp2002; Reinfurt et al. Reference Reinfurt, Stewart and Weaver1991; Rosenfeld Reference Rosenfeld2009) and poor results in developing countries (Ouimet Reference Ouimet2012). However, economic factors are significant predictors of homicides in general (Ferguson and Smith Reference Ferguson and Smith2021) and in countries with medium human development (Ouimet Reference Ouimet2012). Therefore, we hypothesize a significant relationship between homicides and a specific economic indicator of labour force participation in Mexico.
A relationship between homicides and economic indicators is evident across US states (Baron and Straus Reference Baron and Straus1988) and internationally across countries (Ferguson and Smith Reference Ferguson and Smith2021; Ouimet Reference Ouimet2012); however, there is mixed evidence for a relationship between homicides and economic indicators over time (Pratt and Lowenkamp Reference Pratt and Lowenkamp2002; Reinfurt et al. Reference Reinfurt, Stewart and Weaver1991; Rosenfeld Reference Rosenfeld2009). Past empirical research has not shown a relationship between the economy and homicides over time (Reinfurt et al. Reference Reinfurt, Stewart and Weaver1991), although relatively recent research has shown a strong relationship when using different measures of economic conditions (Pratt and Lowenkamp Reference Pratt and Lowenkamp2002) and when using an instrument of perceptions of economic conditions on acquisitive crime (Rosenfeld Reference Rosenfeld2009). Therefore, we hypothesize a strong relationship between homicides and economic indicators across Mexican states.
Results
The results in Table 2 show that 1 million USD of PRONAPRED spending in the previous year is associated with 12 fewer homicides per Mexican state per year, while spending from USAID did not significantly reduce homicides. The number of homicides per year in each Mexican state is related to the number of homicides in the previous year in the same state and all other states.Footnote 17 Finally, a particularly interesting result lies in the R 2 values, as the fixed-effects OLS models exhibit high R 2 values between Mexican states. This result from equation (1) may explain why previous research typically supports panel regression analysis, as regional differences account for a large component of variation in homicides.Footnote 18
Table 2. Homicide levels and crime prevention spending
Notes: National Crime Prevention Programme (PRONAPRED) and United States Agency for International Development (USAID) spending are measured in millions of United States dollars (USD), lagged by one year and adjusted for inflation. Time indicators, control variables and constants are not shown to save space. Standard errors are shown in square brackets.* p < 0.10; ** p < 0.05; *** p< 0.01.
Table 2, column (3) shows the main result that a 1 million USD increase in PRONAPRED spending relates to 12 fewer homicides per state per year. The population coefficient is significantly different from 0, with a p-value less than 0.05. Column (1) shows the base fixed-effects regression model without any spending variables. The coefficient on homicides in the previous year is 0.874, which means that 10 homicides in the past year relate to 8.74 homicides this year. This coefficient is relatively consistent across model specifications and suggests a natural rate of crime (Narayan et al. Reference Narayan, Nielsen and Smyth2010).
Column (2) shows an insignificant relationship between USAID spending and homicides per Mexican state per year. The sample coefficient of –27.22 suggests that the population coefficient is not significantly different from 0, with a p-value greater than 0.10 (or level of confidence less than 90%).
However, it is notable that any three of the R 2 statistics (R 2 within, R 2 between and R 2 overall) remain similar when adding the spending variables to each model. Therefore, the additional variation in homicides explained by USAID or PRONAPRED spending is relatively minor. This result is also important, as we aim to identify factors that explain changes in homicides in addition to assessing programme effectiveness. One such factor is estimated using fixed effects, as indicated by a high R 2 value between groups (Mexican states). Therefore, a good indicator for the number of homicides in Mexico is simply the Mexican state in consideration.
Geographic Evidence for Fixed Effects
Stamatel (Reference Stamatel2015) suggests that data visualization can show new patterns that were not originally apparent and encourage further theoretical thinking in cross-national crime research. In this spirit, we show maps of average homicides by state over time to highlight the explanatory nature of state-specific characteristics estimated by fixed-effects OLS regression.
Figure 2 illustrates that homicides exhibit significant variations between Mexican states, and this trend remains fairly consistent over time. We plot the average number of homicides per state over a three-decade period. The homicides are z-scores based on Mexico’s country-wide mean and standard deviation for each of the three decades. For each figure, we plot six gradients of colour corresponding to six intervals of z-scores. The darker the colour gradient, the greater the number of homicides. The similarities between plots over time in Figure 2 are striking.
Figure 2. Plots of average homicides by state and fixed effects, 1990–2020.
Estimation of Labour Participation
Figure 3 shows male and female labour participation over time, including the fact that male participation declines over time as female participation rises. One possible explanation for the decline in male labour participation is a rise in crime and violence in Mexico, which in turn provides a more lucrative alternative to traditional employment. Therefore, crime prevention programme spending that increases male labour participation may also lower crime in the form of homicide.
Figure 3. Labour participation in Mexico, 2005–2020.
Table 3 shows that variation in male and female labour participation significantly explains variation in homicide levels. Table 3 column (1) shows the coefficient on PRONAPRED spending (–26.90) for the main estimation model from 2005 to 2020. Columns (2) and (3) highlight the significant influence of female and male labour participation on homicide levels – as labour participation increases, homicides decrease. Column (2) shows that a coefficient of –9,378.9 with a 1% (0.01) increase in female labour participation yields 93.79 fewer deaths per state per year. Similarly, column (3) shows a coefficient on male labour participation of –21,871.2, and a 1% increase in male labour participation yields 218.71 fewer deaths per state per year. If we relate 93.79 and 218.71 deaths to the homicide mean of 572, then these are significant declines in homicides.
Table 3. Homicide levels and labour participation, 2005–2020
Notes: The main independent variables are lagged by one year. National Crime Prevention Programme (PRONAPRED) spending is measured in millions of United States dollars (USD) and adjusted for inflation. PRONAPRED spending was given in Mexican pesos, but transformed to USD with averages of daily spot exchange rates. The time indicators by state, other control variables and constants are not shown to save space, but included in regressions. Standard errors are shown in square brackets.* p < 0.10; ** p < 0.05; *** p< 0.01.
Discussion and Conclusion
Programme Effectiveness
Can programme spending reduce homicides? Yes, PRONAPRED was seemingly effective initially, but the reduction in homicides was not sustainable. For instance, consider average PRONAPRED spending per state per year at $680,000 and the reduction in homicides at 12 deaths per 1 million USD. Regardless of whether this is effective or not, the reality is that homicides continue to rise over time in Mexico, even after the initial reduction related to PRONAPRED spending. Furthermore, changes in homicides are largely determined by homicides in the previous year and the state in which they occurred. Therefore, care must be given in applying the programme funding in terms of location, duration and use.
Labour Participation
While programme spending on the face of it does not sustainably lower homicides, more labour participation does relate to fewer homicides. Therefore, programme spending to increase labour participation may be particularly beneficial for crime reduction. Our explanation for this is that traditional employment typically requires giving up criminal opportunities. The choice is a tradeoff, and increasing the benefit of employment causes more labour participation and less crime. Also, a focus on raising male labour participation may have a larger marginal benefit as male homicides comprise more of all homicides.
Future Project Implementation
Should a national programme like PRONAPRED be implemented again in the future? Perhaps, but with changes. We show that a national programme to reduce crime can be beneficial in the short term, but there are issues with cost and longevity of effect. Previous research by Román and González (Reference Román and González2019), Martinez-Palomares (Reference Martinez-Palomares2019), Ramirez-de Garay and Diaz Roman (Reference Ramirez-de Garay and Diaz Roman2017), Merino and Torreblanca (Reference Merino and Torreblanca2017), Chapa Koloffo and Ley (Reference Chapa Koloffo and Ley2015), México Evalúa (2014) and Hernández (Reference Hernández2013) should be consulted to identify the issues with design, implementation and evaluation.
While many are critical of PRONAPRED’s design and implementation, we can also criticize USAID programme spending. One unknown aspect of USAID spending is the precise amount of spending per location per year. We infer constant, equitable spending from a programme schedule, but a detailed outline of programme spending per location per year would be beneficial in determining programme effects over time.
Acknowledgements
The authors wish to express sincere gratitude to the Program on Conflict and Development at Texas A&M University for the guidance at the beginning of this project.
Competing interests
The authors declare no competing interests.
Jorge C. Martinez-Palomares was born and raised in Mexico, where he earned his bachelor’s degree in International Relations and his first master’s degree in Agricultural Economics, including a research exchange semester at Oklahoma State University. He attended Texas A&M University to pursue his second master’s degree in Agricultural Economics. He completed his PhD programme in Agricultural and Applied Economics at the University of Missouri. In his free time, he balances his academic pursuits with a passion for reading philosophy and staying active at the gym.
Eric Lenz earned his PhD in Economics at Southern Illinois University in 2015 and the Certified Business Economist (CBE) designation in 2024. He has taught economics and statistics for several years. He also researches international, monetary and financial economics.
Appendix
Expanded Results for Tables 2 and 3
The expanded main table of results (Table A.4) shows the coefficients on the control and main variables in the fixed-effects regressions. The time-trend coefficient for the sample period 1990–2020 is typically significant and negative while the time-trend coefficient for years 2008 and after is positive and highly significant. Homicides are typically increasing over time after the police and military crackdown on cartels beginning in 2008. Homicides before 2008 are typically fewer than homicides after 2008 – a fact that is also evident in Figure 1.
Notes: The main independent variables are lagged by one year. National Crime Prevention Programme (PRONAPRED) and United States Agency for International Development (USAID) spending are measured in millions of United States dollars (USD and adjusted for inflation. PRONAPRED spending was given in Mexican pesos, but transformed to USD with averages of daily spot exchange rates. The time indicators by state, other control variables and constants are not shown to save space, but included in regressions. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
“Homicides in other states” is another variable in our fixed-effects regression with a significant and negative coefficient. Our hypothesis is a coefficient that is positively related to homicides; therefore, some explanation is due. One reason for a negative coefficient is that the effect of a wave of crime across Mexico is captured by the time trends and differences in homicides among Mexican states are captured by the fixed-effects estimator. Another possible explanation is a highly uneven distribution of homicides across Mexican states. If homicides in other states are driven by a very large increase in homicides in one location (e.g. the variable is skewed by increasing homicides in Mexico City), then the variable may be increasing over time while homicides in most other Mexican states are relatively constant or even decreasing. There may be a negative coefficient on the variable in any of these cases.
Pooled OLS with Additional Independent Variables
We test hypotheses that changes in population measures, cartel presence and peace/violence relate to changes in homicide. Also, Table B.6 shows pooled-OLS regression coefficients because some variables may explain state-specific characteristics (fixed effects). The dummy variables for cartel presence are dropped in the fixed-effects OLS regressions due to collinearity. Therefore, we estimate pooled-OLS regression coefficients to better understand the relationship between cartel-presence and homicides.
Table B.5. Summary statistics for additional variables
Note: DEA, Drug Enforcement Agency.
Table B.6. Pooled OLS with additional variables
Note: USAID, United States Agency for International Development; USD, United States dollars; PRONAPRED, National Crime Prevention Programme. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
First, notice that the control variable coefficients are still significant with the same signs as in the fixed-effects regression in column (1). We also have a high R 2 (0.884) and most of the variation in homicides is still captured by the variation in control variables as we add variables to the base model in (1).
Next, notice that the estimated USAID and PRONAPRED spending coefficients in columns (2) and (3) suggest that the population coefficients are not significantly different from 0. This may suggest that the two variables are not significantly explaining differences in homicides by Mexican state. In this case, the funding allocation would not be related to more violent states (higher homicide numbers) or less violent states (lower homicide numbers). Otherwise, we might expect a more significant effect from the spending.
The population variable comes from the National Population Council or El Consejo Nacional de Población (CONAPO) with population projections beginning in 2016 (CONAPO 2018). We use the population for a state and population growth in other states as control variables in pooled regressions.Footnote 19
The estimated coefficients on the population measures are not significant either; however, the population data are limited to actual values between 1990 and 2015 followed by projections thereafter. Therefore, changes in homicides may not relate to changes in population estimates after 2015.
Finally, we see a significant relationship between Mexican states identified by cartel presence and homicides. The US Drug Enforcement Agency (DEA) cartel presence dummy variable simply identifies Mexican states with cartel presence by a 1 and non-cartel states with a 0. We construct the El Universal cartel presence similarly, but with identification by the Mexico City newspaper El Universal. The Mexico Peace Index identifies the most and least peaceful states in Mexico; however, the homicide information is taken into account when constructing the index. Therefore, we are left only with an important relationship between homicides and DEA-identified cartel presence.
We interpret the significant DEA coefficient in column (6) simply with two possible variable outcomes: a variable value of 0 for no cartel presence (and no effect on homicides) or a value of 1 for cartel presence and an effect of 27 more homicides than otherwise. While this DEA variable is dropped from a fixed-effects OLS regression due to collinearity, it clearly relates to more homicides in Mexico.
Fixed-Effects OLS with Additional Variables
The results in Table C.7 show that OLS fixed-effects coefficients are similar to pooled OLS coefficients in Table B.6; however, we must reject the null hypothesis of equal coefficients from a Hausman test. Also, the positive or non-significant effects from programme spending become significantly negative. Of greatest interest is the significant estimated coefficient on PRONAPRED spending.
Table C.7. Fixed-effects OLS with additional variables
Note: USAID, United States Agency for International Development; USD, United States dollars; PRONAPRED, National Crime Prevention Programme. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
An increase of 1 million USD of PRONAPRED spending (in columns (3)–(8)) relates to about 12.16 fewer homicides per Mexican state per year. The population coefficient is also significant in column (4) with a p-value less than 0.10. Therefore, an increase of about 1 million people relates to 57 more homicides per Mexican state per year. However, we do not include population in the main regressions because it represents actual population numbers from 1990–2015 and forecasts for 2016 and later (CONAPO 2018). Additionally, it only marginally improves the model fit.
Finally, the dummy variables for cartel-presence and peacefulness are dropped for collinearity with the fixed-effects estimator in columns (6)–(8). While we cannot include the DEA variable in a fixed-effects OLS estimation, it is important to note that homicides are greater in cartel-controlled areas.
Male and Female Homicide Victims and Spending
An interesting aspect of estimating homicides is the fact that homicides are not evenly distributed by gender. Table D.8 shows that men suffer disproportionately more from homicides than women. Also, while men suffer from more homicides, crime prevention spending reduces more male homicides than female homicides. Table D.9 shows that for each 1 million USD spent through PRONAPRED, male homicides fell by about 11 and female homicides fell by 1 (according to PRONAPRED coefficients in columns (1) and (3)).
PRONAPRED coefficients are consistently significant when explaining male and female homicides while USAID spending is not. Table D.9 shows that the estimated coefficients on USAID spending in columns (2) and (4), for male and female homicides, are not hypothesized to be significantly different from 0 for the population regression. It is not clear why national programme spending is significantly related to fewer homicides while foreign (USAID) spending is not. PRONAPRED was significantly larger in terms of funding, geographic spread and duration than USAID; however, it is still not entirely clear why one programme was more effective than the other.
Finally, female homicides are increasing more year-to-year than male homicides. The summary statistics in Table D.8 show an average percentage change in homicides of 8.3% for men and 15.5% for women. Beyond some anecdotal evidence from a USAID partner in Monterrey and the metropolitan area that women are getting more involved with criminal activities because of loved ones getting involved, we do not have an explanation for this.
Different Lag Numbers in Spending
Sustainability is an important concern related to effects of programme spending. The results in Table E.10 show that the reductive effect on homicides from PRONAPRED spending lasts only one year. Column (1) shows the main result from 1 million USD in PRONAPRED spending found in the main paper. Columns (2) to (8) show that programme spending from two to four years in the past does not significantly relate to reduced homicides. In fact, columns (2)–(4) and (6)–(8) show an unintended significant and positive relationship between programme spending and homicides.
Table D.8. Summary statistics for male and female homicides
Table D.9. OLS fixed effects for male and female homicides
Note: USAID, United States Agency for International Development; USD, United States dollars; PRONAPRED, National Crime Prevention Programme; M, male; F, female. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
Table E.10. Homicide estimations with different spending lags
Note: USAID, United States Agency for International Development; USD, United States dollars; PRONAPRED, National Crime Prevention Programme. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
The results in Table E.10 should be viewed critically because of the many issues with PRONAPRED identified by other researchers in the Introduction. Additionally, the programme spending schedule by USAID, such as the amount of spending per year per location, is largely unknown. For instance, we infer evenly distributed spending per state per year for what USAID calls “national” programme spending; then, a double-share of spending per state per year when the location is specifically identified by the programme. Also, it may be that much of the spending occurred at the onset of the programme (in late 2012 and 2013) rather than evenly over time. We suspect that this is the case due to the decrease in homicides at the onset of the programmes, but subsequent rise in homicides over several years.
Endogeneity between Programme Spending and Homicides
One potential issue with programme spending is funding allocation. Consider a programme coordinator who determines that funds should go to locations that need it most – locations with high or increasing homicides. This allocation would likely reveal a significant, positive relationship between programme spending and homicides. Therefore, we create a model to address endogeneity using 2SLS regression analysis.
The model below shows a first-stage regression to estimate PRONAPRED and USAID spending. Then, the first-stage estimates of programme spending determine homicides in the second stage.
$$\widehat {Spen{d_{i,t}}} = \alpha + {\gamma _1}{t_i} + {\gamma _2}{(t \gt 2007)_i} + {\beta _1}{X_{i,t - 1}} + {\beta _2}{Z_{i,t - 1}}$$
${H_{i,t}} = \lambda + {\delta _1}{t_i} + {\delta _2}{(t \gt 2007)_i} + {\theta _1}\widehat {Spen{_{i,t}}} + {\upsilon _1}{X_{i,t - 1}} + {\varepsilon _{i,t}}$
The external variable, Z, is the sum of homicide numbers in all other Mexican states. One condition of the 2SLS estimation is that the external variable, Z, is related to programme spending, Spend, but not related to homicides, H. The main results in Table F.11 show that variation in homicides in other states directly relates to programme spending and indirectly relates to variation in homicides. We also include tests for under-identification and instrument validity in the table of results.
The coefficient
$${\theta _1}$$
is compared to the main spending coefficient
$${\beta _1}$$
in the main model. If the coefficient from 2SLS estimation is much larger or smaller and a different sign from the coefficient from fixed-effects OLS, then variation in the external variable (the sum of homicides in other states, Z) may explain previously unaccounted for variation in the endogenous variable (programme spending).
Results
Table F.11 shows that PRONAPRED and USAID spending are significantly determined by the external factor of homicides in other states which in turn greatly influence homicides in Mexico. Columns (1) and (4) show the results from the OLS fixed-effects models, the main models of this paper, while columns (3) and (6) show results from 2SLS. We show the main models without the “homicides in other states” variable in columns (2) and (5) in order to compare coefficients from the 2SLS regressions. The results in the first-stage regressions show that “homicides in other states” influences both USAID and PRONAPRED spending; however, there are inconsistencies with the sign on coefficients and the size of effect from spending in the second-stage regressions with coefficients in columns (3) and (6).
The second-stage regression in column (3) shows that estimated USAID spending from the first-stage regression negatively relates to homicides. The estimated coefficient is –2,242.7 with a highly significant p-value less than 0.05 for a two-tailed hypothesis test of the population beta. However, the second-stage regression in column (6) shows that predicted PRONAPRED spending positively relates to homicides. According to this model, USAID spending was more effective than PRONAPRED spending; however, the root mean-squared errors suggest that the original fixed-effects models better explain variation in the dependent variable.Footnote 20
Table F.11. PRONAPRED and USAID spending and homicides
Note: USAID, United States Agency for International Development; USD, United States dollars; PRONAPRED, National Crime Prevention Programme; MSE, mean squared error; K-P rk LM, Kleibergen-Paap rk Langrange multiplier; K-P rk F, Kleibergen-Paap rk Wald F statistic. Standard errors are shown in square brackets.
* p < 0.10; ** p < 0.05, *** p< 0.01.
Open access
