A Deportation Boomerang? Evidence From U.S. Removals to Latin America and the Caribbean Open Access

Abstract

Introduction and Contribution to the Literature

Countries use a wide variety of strategies to prevent, restrict, and decrease the entry of migrants into their states and their labor markets. High-income countries have increased residency requirements, added language tests, and raised educational requirements as barriers to legal migration (e.g., Helbling and Leblang 2019). Likewise, destination countries have developed and deployed a wide range of policies to combat undocumented migration. The construction and fortification of border walls and fences has become an increasingly popular strategy of politicians across the world despite little evidence that these physical barriers deter would-be immigrants (Avdan and Gelpi 2017; Linebarger and Braithwaite 2020; Schon and Leblang 2021).

Even before Donald Trump announced his intention to build a “big, beautiful wall,” the United States had been using various strategies to deter those seeking to enter without documentation. The strategy of “Prevention Through Deterrence” began in the 1990s with policies such as Operation Blockade, which attempted to shift migrant entry away from cities on the U.S. southern border to areas with topography and weather that would put migrants’ safety at risk. These policies were paired with increased deportations, with the premise that making the journey more hazardous and increasing the consequences of apprehension would alter the cost–benefit calculation of those contemplating unauthorized migration, ultimately discouraging them from attempting entry (Goodman 2020; Hiemstra 2017). So ingrained was the emphasis on deportation that President Bill Clinton implemented a policy referred to as “Prevention Through Deterrence and Deterrence Through Deportation” (Spotts 2001), and President Barack Obama was referred to as the “Deporter-in-Chief.” Whereas Clinton emphasized deportation in general, Obama's administration emphasized the deportation of those with criminal convictions.

The Trump and Biden administrations followed suit, combining deportations with other punitive measures to address the ever-increasing demand for undocumented entry across the U.S. southern border. Does this policy work? We answer this question by examining the largest migration and deportation corridor in the world. During our study period of 2003–2021, countries from Latin America and the Caribbean received 5.7 million deportees and were the target of 97% of all deportations from the United States. Total annual deportations to Latin America and the Caribbean peaked in 2013 at 420,000 deportees (Figure 1). During our study period, individuals from Latin American and Caribbean countries accounted for 98% of annual apprehensions at the U.S. southern border and 43% of annual asylum requests. The number of unauthorized migrants from Latin American and Caribbean countries apprehended at the U.S. southern border declined strongly between 2005 and 2015. Apprehensions increased again until 2020, when the COVID-19 pandemic was associated with a decline in border apprehensions. Parallel to a decreasing trend in border apprehensions, the number of asylum applications from Latin American and Caribbean countries increased after 2013, suggesting a shift in forms of entry.

Empirical evidence on the effect of deportation risk on migration is mixed. Earlier studies on the U.S. case found either no effect or only weak effects on overall migration stocks and flows (Donato et al. 1992; Espenshade 1994; Hanson and Spilimbergo 1999; Kossoudji 1992). The explanation from these studies is that stricter border enforcement disrupted patterns of circular migration, leading to longer stays of undocumented migrant populations (Angelucci 2012; Kossoudji 1992; Massey et al. 2016). Moreover, stricter enforcement changed migration routes and crossing points (Gathmann 2008) rather than reducing flows. Other studies have reported evidence consistent with the intuitive idea that stricter immigration enforcement lowers emigration rates. Angelucci (2012) observed that stricter border enforcement in the United States deterred the inflow of Mexican migrants. Separately, Amuedo-Dorantes et al. (2013) confirmed that deportation fears curb deportees’ intent to return to the United States in the near future. Bohn and Pugatch (2015) estimated that Mexican migration to California and Texas would have been 8 percentage points greater had states not increased border enforcement between 1994 and 2011. In a study on Mexico's Southern Border Plan, Martínez Flores (2020) found that increased enforcement in Mexico decreased Central American migrants’ likelihood of attempting repeated unauthorized crossings.

These studies are useful in identifying and assessing individual economic cost–benefit considerations and the question of how immigration enforcement affects optimization strategies, often with unintended effects, such as the interruption of circular migration patterns or a change in forms of entry. Our study differs from these studies by focusing on deportation externalities in migrants’ countries of origin as an additional factor contributing to the emigration decision. Empirically, we hypothesize that the overall effects of immigration enforcement on later emigration will be positive if the unintended effects of deportation externalities outweigh the deterrence effect of deportation threats.

Our argument is rooted in qualitative and quantitative literature finding that the forced return of migrants who committed criminal offenses in the United States leads to an increase in violence in the origin community. One stream of work has traced the spread of violent gangs in Central America to the urban peripheries of U.S. metropoles where children of Central American migrants grew up. Partly as a response to the involvement of the immigrant population in drug trafficking and other illegal gang activities, large-scale deportations of Central Americans began in the mid-1990s. Removals intensified with the passage of the Illegal Immigrant Reform and Immigrant Responsibility Act in 1996 (Seelke 2011). Qualitative research has studied the role of deported migrants in the transformation of gangs and the recruitment of members from among local populations (Cruz 2013). Recent quantitative studies confirmed the effect of deportations on gang-related violence in El Salvador using different identification strategies. Ambrosius (2021) traced the contagion of violence along migration corridors in El Salvador and observed a rise in homicides parallel to the arrival of deportees with prior convictions. Sviatschi (2022) found that incarceration rates are higher among adults who were exposed to the arrival of deportees during childhood, exploring variation across cohorts and over time. Both Kalsi (2018) and Sviatschi (2019) leveraged the timing of deportations and different degrees of exposure across cohorts by comparing individuals born before with those born after the arrival of deportees to estimate the effect of gang-related violence on education outcomes.

Such patterns do not seem to be exclusive to Central America. For the Mexican case, Slack (2019) described how criminal groups specifically target deportees who can easily be coerced into criminal activities because of the economic precariousness and stigma they face upon return. Rozo et al. (2021) diagnosed an increase in violence around repatriation centers in Mexico, the first arrival point of deportees, in years when deportations surged. Similarly, Ambrosius (2025) observed that when deportation threats at migrants’ U.S. destinations increased, violent crime in their Mexican municipalities of origin increased. Ambrosius and Leblang (2020) confirmed relationships between deportations and violence in a cross-country panel, using migrants’ exposure to U.S. migration policies as an instrumental variable. All these studies reported coherent patterns of negative deportation externalities (i.e., increased violence in the deportee's home country) using different empirical strategies. Note that these results need not imply that deportees are violent. It is plausible that deportees fall victim to the very crimes from which they tried to escape—as reported, for instance, for deportees to Central America (Bender 2019)—or that other negative effects of deportation threats (such as a decline in remittances) contribute to an increase in violent crime. What is also important to emphasize is that higher immigration did not increase violent crime in the United States, and that the foreign-born population in the United States has a lower incarceration rate than natives (Abramitzky et al. 2024). Our view is that polices that cause extreme precariousness among deportees led to an increase in violent crime at origin, not traits intrinsic to migrants or deportees themselves.

At the same time, escaping violence has joined economic incentives as an important driver of migration dynamics from Latin America to the United States. In exploring the case of unaccompanied minors from Northern Triangle countries—El Salvador, Guatemala, and Honduras—Clemens (2021) found that the effects of short-term increases in violence are roughly equal to the explanatory power of long-term economic characteristics, such as average income and poverty. In the case of El Salvador, deportations have increased migration to the United States as a result of gang violence (Sviatschi 2022). According to Orozco-Aleman and Gonzalez-Lozano (2018), Mexican municipalities most affected by drug-related violence experienced an increase in emigration to the United States. Therefore, we argue that rather than deterring migration, deportations have a boomerang effect. Deportations increase violence in the returnees’ home country, subsequently increasing emigration. The identification and specification of models to test the boomerang effect of enforcement policies are at the heart of this study. Mediation analysis applied to an annual panel of Latin American and Caribbean countries allows us to decompose the effect of deportations on emigration into three coefficients of interest: (1) a total effect of deportations on later emigration, (2) an effect of deportations on the mediator variable of violence, and (3) an effect of violence on emigration.

Our main empirical challenge is the endogeneity of our key explanatory variables. An increase in undocumented migration will likely also translate into more deportations. The same concern applies to the mediator variable of violence: violence is expected to increase emigration, which might lead to more deportees. In this case, we would measure an underlying increase in emigration rather than an exogenous increase in the arrival of deportees. To address these endogenous feedback loops, we follow several recent studies that have exploited the unequal and staggered implementation of migration policies at the U.S. state level as an opportunity for causal analysis (East et al. 2023; East and Velásquez 2024; Ellis et al. 2016; Hines and Peri 2019; Leerkes et al. 2013; Miles and Cox 2014; Orrenius and Zavodny 2015; Orrenius and Zavodny 2022). Our research differs from these studies by relating the unequal exposure to U.S. enforcement policies to dynamics in migrants’ countries of origin in a shift-share design: if migrants happen to live in U.S. states that adopt anti-immigrant policies—policies that increase deportation risk—their countries of origin will receive a larger number of deportees and, thus, experience more violence. Following the framework of Borusyak et al. (2022), we identify variation as a shift in policy changes over time and not as a change in the shares of migrants’ distribution across the United States before these policy changes. We postulate that U.S. state decisions to implement these policies are not directly related to violence and emigration in migrants’ countries of origin in later periods, other than through the deportation channel. Under this assumption, the shifts provide plausibly exogenous variation in deportation risk conditional on the inclusion of country and year fixed effects and a battery of time-varying covariates.

Our findings support a counterintuitive and unintended effect of pro-deportation policies on later emigration, mediated via an increase in violence. Using exogenous exposure to deportation risk at migrants’ destinations as an instrument, we observe an increase in homicide rates as a result of the arrival of deportees with prior convictions, confirming the results of previous studies. We then use exposure to deportation risk to directly predict violence in what is equivalent to the reduced-form regression of the previous model. Violence predicted from exposure to deportation risk increases migration intentions in later periods. In terms of actual migration, we see roughly two additional asylum requests for every deportee with a record of crime in the United States, a pattern that is strongly driven by dynamics in Central America. Total apprehensions at the U.S. southern border show no clear pattern in the full sample. Among those apprehended, unaccompanied minors and entire family units as a share of total apprehensions increase in response to deportation threats, suggesting a shift in migration strategies and a change in the demographic profile of migrants.

Identification: Exposure to Deportation Risk

In this research, we aim to trace the effect of deportations on emigration rates in subsequent periods, as mediated via violence as an unintended external effect of deportations. An empirical test of these mechanisms poses an identification challenge owing to the endogeneity of key variables. Migration surges will likely translate to more deportations. In this case, a positive correlation between migration indicators and deportations would reflect reverse causality from migration to deportations rather than from deportations to migration. In a situation of such feedback loops, violence would also be endogenous: violence is expected to increase emigration, which might lead to more deportees.

Therefore, to estimate causal effects, we need to obtain exogenous variation on the two key explanatory variables: (1) the “treatment” of deportations and (2) the mediator, violence. Several studies have used the unequal and staggered rollout of migration policies in the U.S. federal immigration system to evaluate their effects on labor markets (East et al. 2023; East and Velásquez 2024; Orrenius and Zavodny 2015), crime (Hines and Peri 2019; Miles and Cox 2014), and settlement patterns of migrants (Ellis et al. 2016; Leerkes et al. 2013; Orrenius and Zavodny 2022). Unlike these studies, we focus on the policies’ effects in migrants’ countries of origin. To this end, we apply a shift-share instrumental strategy (Bartik 1991; cf. Goldsmith-Pinkham et al. 2020), like that applied by Ambrosius (2025), Bucheli and Fontenla (forthcoming) and Caballero (2022) in work focusing on Mexico. At the core of this strategy lies the reasoning that countries of origin receive more deportees if migrants happen to live in U.S. states or counties that adopt policies that increase deportation risk.

Whereas the shifting variable in this case derives from short-term changes in policies at the U.S. state level, the share variable is defined via historically established migration corridors that create spatial variation in migrants’ exposure to these policies. Consistent with the identification framework of Borusyak et al. (2022), exogenous variation in our setting comes from the shifts, not from the shares. Because migration corridors are defined before the policy changes, the instrument is not affected by the short-term sorting of migrants into destinations—in particular, the possibility that migrants avoid states with more hostile policies (Leerkes et al. 2013; Orrenius and Zavodny 2022). Our core identifying assumption is that U.S. state migration policy decisions are not directly related to subsequent trends in violence or emigration rates in countries where migrants residing in specific states originated. This assumption means that selection into these programs is not necessarily random but that the politics of adoption should be unrelated to underlying country-specific trends in migration and violence back home. We return to empirical support for this assumption later.

Over the study period, two programs were adopted across U.S. states with the explicit intention of expanding deportations (for descriptions and years of adoption, see Bernstein et al. 2022). Secure Communities was a federal data-sharing program through which the FBI shared fingerprints it received from local law enforcement agencies with immigration enforcement agencies for checks against immigration databases. Depending on the results, immigration officials decided whether to take enforcement action, such as issuing a detainer request. Second, Section 287(g) of the Immigration and Nationality Act enables state and local law enforcement officers to collaborate with the federal government to enforce federal immigration laws. Over time, three enforcement models have been in place under Section 287(g). First, the jail enforcement model deputizes officers to interrogate the immigration status of arrested foreigners, who might then place immigration detainers for suspected immigration violations. Second, under the warrant service officer model (hereafter, the warrant model), state and local law enforcement officers can perform the arrest functions of an immigration officer in jails and/or correctional facilities. Third, the task force model, which operated only until 2012, allowed deputized officers who encountered alleged foreigners during daily activities to question and arrest individuals they believed had violated federal immigration laws.

Figure 2 plots the share of all states that had adopted at least one of the three models under Section 287(g) or Secure Communities each year on the left axis. Up to 40% of states had applied one of the three models at its highest value in 2009. Secure Communities began in 2008 and became mandatory in 2012. In 2014, the program was temporarily abandoned, but it was reinstated in 2017 and has been mandatory for state and local jurisdictions since then.¹ The right axis of the figure shows the total number of deportations from the United States to Latin American and Caribbean countries. This number increased until 2013, when it peaked at 420,000 total annual deportations to Latin American and Caribbean countries. Of these deported individuals, 195,000 had been convicted of a criminal offense.²

We evaluate the relationship of these programs with the number of deportations at the U.S. state level in Table A4 and describe data and their sources for analysis at the U.S. state level in Table A2 (all tables and figures designated with an “A” appear in the online appendix). Enforcement under Section 287(g)—in particular, the jail model—is a strong predictor of (logged) deportations.³ The effect of Secure Communities is not statistically significant in these regressions, likely because the staggered rollout of Secure Communities covers only 4 of the 18 years included in the estimation sample utilized in Table A4, leaving us with little variation over time from which to estimate an effect. In addition, the program's implementation was unequal when being adopted (NBC News 2014). We therefore rely on exposure to policies of Section 287(g) to predict deportations.

where the percentage share of all migrants (⁠$M$⁠) born in country $i$ is multiplied by a variable that indicates whether policy $P$ was in place in destination state $k$ during year $t$⁠. Values are then summed across all destinations (⁠$K$⁠). The approximate distribution of migrants by their birth country across 50 U.S. states (⁠$Mk,i$⁠) is calculated from the 2005–2009 American Community Survey. $P$ equals 1 if some of the counties with the most immigrants in a state or a statewide agency had any of three policies (the 287(g) jail model, the task force agreement, or the warrant model) in place. Figure A1 plots the variation in exposure to 287(g) policies over time for each of the 30 countries in the sample.

In Eq. (2), $D$ refers to the number of deportees per 100,000 origin country individuals who were convicted of a crime before being returned. We use the share of deportations for unlawful conduct rather than the total number of deportees because the policies under Section 287(g) explicitly target this subgroup and because research has identified the return of deportees with prior convictions (not deportees as a homogeneous class) as a driver of violence (e.g., Ambrosius 2021; Rozo et al. 2021; Sviatschi 2022).

Equation (3) follows the same setup as Eq. (2) but predicts levels of violence (⁠$V$⁠) proxied by homicide rates (per 100,000) instead of deportations. The rationale for the regression in Eq. (3) is that deportation risk also affects violence via deportations. We can think of the latter as the reduced-form version of the instrumented effect of deportations on violence that allows us to isolate the variation in violence that responds to deportation risk. The instrument of exposure to deportation risk (⁠$DepRisk$⁠) is lagged by one period with respect to violence. Because the exact lag structure is unknown, we provide model results for alternative lags and for two-year periods in the section on robustness checks. $β$ are the estimated coefficients, and $μ$ is the usual error term. Year fixed effects (⁠$τ$⁠) control for overall trends that do not vary by country, such as changes in nationwide policies that affect all migrants equally. Country fixed effects (⁠$υ$⁠) control for all time-constant variables, including the location of migration corridors and the time-invariant institutional legacies, geographic conditions, and characteristics of migrant populations. This ensures that identifying information comes from changes in the U.S. state-level immigration policies. $X$ is a vector of time-varying controls, lagged by one year to capture pretreatment conditions. We include the log of population size, the log of per capita GDP, the log of remittances received, and growth rates in country $i$⁠. We include these controls to distinguish an effect of deportations from these general demographic and economic conditions that could also respond to migration dynamics.⁴ In addition to origin country controls, we also include three indicators of exposure to the conditions that migrants face in U.S. states: migrants’ average exposure to unemployment rates, exposure to violent crime rates, and exposure to the growth of the Hispanic population. We calculate these indicators the same way we created the indicator capturing exposure to deportation risk.

As mentioned earlier, our key identifying assumption is that the adoption of 287(g) policies is not related to subsequent trends in violence and migration in migrants’ countries of origin other than through the deportation channel. A threat to this assumption would arise if states adopted enforcement policies in response to migration surges in their specific states. We find no empirical support for this possibility. Table A5 evaluates the likelihood of adopting 287(g) policies in a conditional logit model. The growth of the Hispanic population is not a statistically significant predictor of adoption within states and has a negative sign for the jail model, the main driver of deportation risk. The violent crime rate has a negative sign and is statistically significant for the case of the jail model, suggesting that an increase in crime rates did not motivate the adoption of 287(g) policies. This finding makes the existence of reverse causality between migration, violence, and enforcement policies unlikely and is in line with prior assessments that domestic politics (particularly cooperation between federal and local agencies), not migration-related variables, have been a main driver of policy adoption (Ciancio and García-Jimeno 2024; Magazinnik 2022).⁵

Mediation Analysis

The process we want to trace is visualized in Figure 3, with arrows referring to potential causal relationships between key variables. Our interest lies in the effect of forced returns (deportations) on future emigration, mediated by violence as a negative externality of forced returns. Mediation analysis allows us to decompose the effect of deportations on emigration into three coefficients of interest. First, a total effect of forced return on emigration in subsequent periods, captured by coefficient $βED$⁠, can operate in different ways, and the expected sign of the total effect is therefore ambiguous. Deportations could affect emigration negatively, given that an increase in migration risk might deter some individuals from emigration. On the other hand, deportations could increase posterior migration through their effects on variables that cause more migration.

More specifically, deportations are expected to generate negative externalities in the form of more violence (⁠$V$⁠), the mediator variable. The second coefficient of interest, $βVD$⁠, refers to the effect of deportations (⁠$D$⁠) on violence (⁠$V$⁠). The increase in violence might increase emigration (⁠$E$⁠) in subsequent periods. Lastly, coefficient $βEV$ refers to the effect of violence (⁠$V$⁠) on emigration (⁠$E$⁠). If we obtain estimates on all three coefficients, we can assess the magnitude of the total effect (⁠$βED$⁠) compared with a mediated effect (⁠$βVD$ × $βEV$⁠) of forced return on emigration. This framework allows us to examine whether there is a boomerang effect, whereby deportations increase subsequent immigration to the United States through the violence channel. Endogeneity is captured in an arrow that leads from migration back to deportations. In Figure 3, we therefore include exposure to deportation risk as an exogenous instrument Z that predicts deportations and violence in the two separate first-step regressions. We use predicted deportations (⁠$D^$⁠) and predicted violence (⁠$V^$⁠) from the first-step regressions, Eqs. (2) and (3), to obtain plausibly unbiased estimates for all three coefficients of interest: the total effect (⁠$βED$⁠), the partial effect of deportations on violence (⁠$βVD$⁠), and the mediated effect from violence on emigration (⁠$βEV$⁠). Violation of the assumption of sequential independence forbids using the standard approach to mediation analysis involving controlling for the mediator (violence) in a regression of deportations (treatment) on emigration (outcome) (e.g., Imai et al. 2010).⁶

Equation (4) estimates the total effect of predicted deportations (⁠$D^$⁠) obtained from the first-step regression in Eq. (2) on actual or intended emigration (⁠$E$⁠) (coefficient $βED$⁠). Equation (5) estimates the effect of predicted deportations (⁠$D^$⁠) on the mediator variable violence (⁠$V$⁠) (coefficient $βVD$⁠). Finally, Eq. (6) estimates the effect of predicted violence (⁠$V^$⁠) obtained from the first-step regression in Eq. (3) on actual or intended emigration (coefficient $βEV$⁠). Emigration is lagged by one period with respect to violence and by two periods with respect to deportations. Given the exogeneity assumption, second-step regressions in Eqs. (4), (5), and (6) estimate a local average treatment effect for those countries whose migrant populations respond to deportation risk.

We use three alternative indicators to measure actual or intended migration (⁠$E$⁠). For actual migration, we use the number of asylum requests per 100,000 home country individuals and the number of apprehensions of undocumented migrants at the U.S. southern border per 100,000 home country individuals. Both indicators measure different dimensions of migration flows outside the formal visa system. Whereas asylum requests likely include a larger share of those fleeing violence or persecution in their countries of origin, apprehensions at the U.S. southern border typically involve a larger number of young males migrating for economic purposes. For a smaller sample, we also calculate the share of minors and entire family units relative to total apprehensions to assess the effect of deportation threats on the composition of migrants. For migration intentions, we use survey data and create country-year averages of responses to the Gallup World Poll question, “Ideally, if you had the opportunity, would you like to move permanently to another country?” The latter covers fewer years and countries and is therefore not well identified in some of the specifications. Table A1 summarizes and describes all variables used and the respective sources for the country-level analysis.

Results

Table 1 shows regression results for first-step regressions of exposure to deportation risk on the two endogenous variables: the arrival of deportees with a prior conviction per 100,000 home country individuals and the number of homicides per 100,000 persons. Results are shown for three different samples, depending on the availability of data in the second stage for each of the three migration indicators. All models include country and year fixed effects, and a set of time-varying control variables related to both migrants’ destination country conditions and their origin country conditions. Exposure to deportation risk has a strong and statistically significant effect on the number of deportees received per 100,000 and on homicide rates. The observed country-level patterns are consistent with patterns we find at the U.S. state level: 287(g) policies are a strong predictor of deportations in both cases (cf. Table A4). Exposure to deportation risk has a stronger effect on the arrival of deportees than on homicides in terms of both magnitude and statistical significance. This finding is in line with our expectation of an indirect effect on homicides that operates via the deportation channel. Hence, the lower panel in Table 1 can also be interpreted as a reduced-form effect of deportation threats on homicide rates.

Table 2 provides second-step results of deportations on homicide rates, using migrants’ exposure to deportation threats (the upper panel in Table 1) as an instrumental variable in the first-step regression. Coefficients estimated in Table 2 are equivalent to the partial effect of deportations on homicides (coefficient $βVD$⁠) in Figure 2. We see a large (lagged) effect of the arrival of deportees with prior convictions on homicide rates. For every deportee with prior convictions, homicides increase by more than 0.2 in the larger samples (Samples 2 and 3). These results support previous findings for a cross-country panel (Ambrosius and Leblang 2020) and for country case studies on Mexico (Ambrosius 2025; Rozo et al. 2021) and El Salvador (Ambrosius 2021; Kalsi 2018; Sviatschi 2022) using different strategies. Weak-instrument F statistics from first-step regressions are large (above 40) in Samples 2 and 3 and lie above the classical threshold of 10 (Stock and Yogo 2002) for the smaller sample in the first column. The coefficient for the smaller sample is larger (0.43), indicating that the coefficient size is sensitive to periods and countries covered.

We then use predicted deportations and predicted violence as in Table 1 to estimate the total effect of deportations on emigration (coefficient $βED$⁠) and the effect of violence on emigration (coefficient $βEV$⁠). Results from these second-step models are provided in Table 3. Migration intentions respond positively to the arrival of deportees: for an increase of 10 deportees with prior convictions per 100,000 persons, migration intentions increase by 3 percentage points. Column 2 shows results for asylum requests per 100,000 persons. For every deportee with a prior conviction, asylum requests increase by more than two in later years.

To assess the sensitivity of results with respect to the underlying functional form, we also repeat the model on log-transformed variables instead of per capita rates in column 3. In this case, results can be interpreted in terms of elasticities. For asylum requests, a 1% increase in the arrival of deportees with prior convictions leads to roughly a 0.5% increase in asylum requests; a 1% increase in homicides leads to more than a 1% increase in asylum requests. We find no statistically significant effect of deportations on apprehensions in per capita terms (column 4) or in log terms (column 5). The negative (albeit statistically insignificant) value in column 5 could indicate that deportation threats shift migration strategies from undocumented border crossings to entry via the asylum system, as Schon and Leblang (2021) observed. The effects of deportations on actual and intended emigration are reflected in the coefficients we obtain on violence in the lower panel. Regarding intentions to migrate, an additional 10 predicted homicides per 100,000 persons increases migration intentions by 8 percentage points, and asylum applications increase by almost 10 for every additional homicide per 100,000 persons. For the model in log terms, a 1% increase in homicides is associated with more than a 1% increase in asylum requests (column 3) and more than a 1% decrease in apprehensions (column 5). The results of homicides on apprehensions (column 4) are not statistically significant if measured in per capita terms.

Using estimates of the individual coefficients in Table 3, we calculate the effect on emigration that is mediated through violence by multiplying $βVD$ × $βEV$⁠. We then compare this effect with a total effect that we obtain directly from a regression of deportations on emigration. Figure 4 shows the total and mediated effects on emigration intentions and asylum requests in per capita terms, along with the respective 95% confidence intervals. The figure shows that basically the entirety of the effect of deportations on intended emigration in later periods and the effect on asylum requests are mediated through the effect of deportations on violence. The observed effect of deportations on violence and the effect of deportations on migration intentions and asylum requests are large but in line with magnitudes found in other recent studies that have separately examined the effect of deportations on violence (Ambrosius 2021, 2025; Rozo et al. 2021; Sviatschi 2022) and on the effect of violence on migration (Clemens 2021). A likely explanation for these large effects is that deportations generate externalities in migrants’ countries of origin beyond their effect on individual deportees.

Table 4 evaluates the effect of exposure to 287(g) policies on migration indicators in later periods, irrespective of the channels. In line with our observation for instrumented deportations, exposure to deportation threats has a strong and statistically significant positive effect on migration intentions and on the log of asylum requests. We see a negative and weakly significant effect on the log of apprehensions. For a shorter period of time for which we have data on the demographic composition by origin country, we also evaluate whether deportation threats changed the composition of apprehended migrants. A stronger exposure to deportation risk is associated with a larger share of family units relative to total apprehensions, a larger share of unaccompanied minors, and a lower share of single adults.⁸ These findings suggest that enforcement policies affected both migration strategies and the demographic composition of new entrants.

Heterogeneity Analysis and Robustness Tests

The average effects analyzed in the previous section might hide important differences in strategies based on the migrant's origin country. To test the sensitivity of our results to country samples, we estimate models on three different subsets. First, we exclude the border country of Mexico, which accounts for almost 80% of all apprehensions at the U.S. border during our study period and therefore strongly affects average outcomes. In a second sample, we exclude the three Northern Triangle countries of Central America (El Salvador, Guatemala, and Honduras), which account for much of the overall variation in deportations, homicides, and migration in per capita terms. In a third sample, we exclude 11 countries from South America that are geographically distant from the United States and that receive a relatively small number of U.S. deportees. We summarize the results in coefficient plots with point estimates and confidence intervals of 2 standard errors for three migration indicators: intentions to migrate (Figure A2), asylum requests (Figure A3), and apprehensions (Figure A4). Other than changes in the country sample, all specifications follow those shown in Table 3.

Results for the (instrumented) effect of deportations on homicides hold for different subsets, with some variation in precision and instrument strengths. Similarly, migration intentions show similar patterns across country samples, although statistical significance and instrument strength are affected by a lower number of observations. Excluding South American countries from the sample makes little difference, supporting the expectation that the relationship between deportations, violence, and migration is of little empirical relevance to these countries. On the other hand, the effect of deportations and homicides on asylum requests is strongly driven by dynamics in Central America: when we exclude variation from the three Northern Triangle countries of Central America, the results of the model in log terms and the model in per capita terms become statistically insignificant. With respect to apprehensions, results differ strongly in the subset excluding Mexico and in the subset excluding the Northern Triangle countries. Violence resulting from deportation threats is associated with a decrease in apprehensions when we exclude the three countries of Central America and is associated with an increase in apprehensions when we exclude Mexico. Residual plots for the model on asylum applications and apprehensions in log terms highlight the variation that comes from the three Central American countries and Mexico (Figure A5).

Whereas the pattern we observe for Central American countries is in line with previous case study evidence (Clemens 2021; Sviatschi 2022), research on the effect of violence on international migration has been less clear-cut for Mexico. Although Orozco-Aleman and Gonzalez-Lozano (2018) detected a positive effect of violence on migration to the United States, other studies questioned the relationship between violence in Mexico and international border crossings (Basu and Pearlman 2017; Massey et al. 2020). Further, Rios Contreras (2014) detected higher migration outflows in places with higher drug-related violence and crime. One answer to this apparent puzzle could lie in domestic migration as an alternative to international migration in the presence of increased deportation threats and geographically concentrated violent crime: Massey et al. (2020) found no effect of violence on international migration in Mexico, but they identified an effect on domestic migration. Daniele et al. (2023) found a stronger response to violence in terms of domestic migration compared with international migration. These studies offer a suggestive interpretation of the heterogeneity in international migration responses we observed, but a detailed exploration of country-specific patterns is beyond the scope of this article.

We also evaluate the robustness of results to the choice of instruments and different lag structures. To capture outcomes for a larger number of model specifications, we summarize results in coefficient plots for each of our three main migration indicators (see Figures A6–A8). Because deportations seem to be driven mainly by the jail model (as shown in Table A4), we provide results for second-step specifications on emigration indicators using exposure to the jail enforcement model only. Second, given that 287(g) policies were implemented at the county level and not at the state level, we also create an alternative instrument using pseudo-counties created from the Public Use Microdata Sample (PUMS) used in the American Community Survey.⁹ Third, to assuage concerns about our choice of time lags, we use instruments with an additional lag and provide the main results using two-year averages rather than lags of annual periods. We also exclude the pandemic years 2020 and 2021 from regressions in the coefficient plots. The results of these robustness exercises generally remain unchanged, and the coefficients from different instruments, specifications, and lags generally move within uncertainty bounds.

Results using exposure to the jail model are similar to results using any of the three modes of operation (the task force, jail, or warrant model). The instrument using pseudo-counties is weaker, likely because PUMS boundaries do not match county boundaries in many cases and because the foreign-born population is measured imprecisely at the pseudo-county level. Overall, coherent findings across specifications and variable choices provide strong evidence for the existence of an unintended effect of deportations on later emigration mediated via an increase in violence. At the same time, given the imperfect observational data, we emphasize the existence and direction of an effect rather than a precisely estimated coefficient.

Following previous literature (e.g., Ambrosius 2021; Rozo et al. 2021; Sviatschi 2022), we postulate that the effect of deportations on violence is driven by deportees with a prior conviction. In fact, we cannot disentangle the effect of deportees with convictions from those without. Both deportation types increase in response to 287(g) policy adoption, and the composition of deportees remains largely unchanged (see Table A6). These findings suggest that 287(g) policies increased risk for all undocumented migrants despite being portrayed as targeting deportees with prior convictions. Analogous to Figure 4 in the main text, Figure A9 depicts total and mediated effects for asylum requests and apprehensions when using all deportations rather than the subset of deportees with prior convictions. Overall patterns are unchanged, although coefficients estimated on the larger group of all deportees are naturally smaller.

Conclusions

The causes and consequences of migration continue to captivate academic and public debate. The arguments and evidence we present make several noteworthy contributions. First, the return of deportees is associated with increased violence in migrants’ origin countries, confirming findings from studies using a variety of empirical strategies (Ambrosius 2021, 2025; Ambrosius and Leblang 2020; Kalsi 2018; Rozo et al. 2021; Sviatschi 2022). This increase in violence likely results from multiple direct and indirect mechanisms, including the targeting of deportees and their recruitment by criminal groups (Slack 2019) in a context where deportees face limited economic opportunities paired with strong social stigmas (Brotherton and Barrios 2009; Mojica Madrigal 2017; Schuster and Majidi 2013; Silver 2018) and where states have been unable to control the spread of violence. Our novel contribution lies in the observation that these negative deportation externalities lead to a boomerang effect, increasing emigration intentions and asylum requests in subsequent periods. For apprehensions at the U.S. southern border, we observe substantial heterogeneity in patterns across countries and no statistically significant effect on average. Further, deportation threats changed the demographic profile of migrants, increasing the share of minors and entire family units among apprehended migrants and decreasing the share of single adults.

Three main caveats should be considered when interpreting these effects, given that they are based on imperfect observational data. First, the reported effects occur with time lags that are a priori unknown and hard to model precisely. We therefore emphasize the existence and direction of an effect rather than precisely estimated coefficients. Second, we cannot effectively separate the effect of convicted deportees from that of other deportees because deportation threats seem not to have changed the composition of deportees. Both migrants with convictions and those without convictions were at higher risk of being deported. Third, average effects hide large differences in country-specific dynamics. More research is needed to clarify adaptation of migration strategies in response to deportation risk and violence in country-specific settings; a deeper exploration of context-specific migration responses lies beyond the scope of this article and is left for future research.

One pillar of U.S. immigration policy since the mid-1990s has been a tough stance on immigrants who have committed criminal offenses while in the United States. Thus, our findings have immediate policy implications. If a deterrence effect of deportations exists, it seems to be small compared with the effect of deportations on intended and actual emigration mediated through violence. By emphasizing negative deportation externalities at origin, we complement literature highlighting unintended outcomes of these policies in the United States. Deportations did not make deporting communities safer (Hines and Peri 2019; Miles and Cox 2014). Instead, they had a negative impact on native workers in the United States, who suffered from increased labor costs, lower aggregate consumption, and less job creation (East et al. 2023) and faced the loss of domestic services that undocumented migrants provided to mothers (East and Velásquez 2024). The general message that emerges is that while deportations in the U.S. context have been motivated by an empirically unfounded narrative of migrants as drivers of crime (Abramitzky et al. 2023), these policies created important unintended effects on both sides of the border.

Acknowledgments

This article builds on an earlier manuscript circulating under the title “Violence, Emigration Demand, and the Deportation Boomerang,” which benefited from feedback and comments by participants at seminars and conferences at Facultad Latinoamericana de Ciencias Sociales Mexico, Freie Universität Berlin, El Centro de Investigacíon y Docencia Económicas, Instituto Technológico y de Estudios Superiores de Monterrey, Instituto Technológico Autónomo de México, Colegio de México, the Symposium on Migration in North America at the House of the University of California in Mexico City, the LACEA-LAMES conference in Puebla, and the 13th Migration and Development Conference hosted by the World Bank, as well as several anonymous reviewers. This new version benefited from comments received at the Nordic Conference in Development Economics (Gothenburg, June 2023), the German Development Economics Conference (Dresden, June 2023), and Universidad EAFIT (Medellín, February 2024), as well as from comments by Tobias Heidland, Thomas Goda, and three anonymous referees. David Leblang acknowledges funding from the Department of Homeland Security under grant award 2015-ST-061-BSH001, which was awarded to the Borders, Trade, and Immigration Institute: A DHS Center of Excellence led by the University of Houston and includes support for the project “Modeling International Migrant Flows” awarded to the University of Virginia. The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the Department of Homeland Security or the U.S. government.

Notes

References