Migrant Sanctions and Recidivism in Border Apprehensions

Deterring Illegal Entry: Migrant Sanctions and Recidivism in Border Apprehensions Abstract Over 2008 to 2012, the U.S. Border Patrol enacted new sanctions on migrants apprehended attempting to enter the U.S. illegally. Using administrative records on apprehensions of Mexican nationals that include ngerprint-based IDs and other details, we detect if an apprehended migrant is subject to penalties and if he is later re-apprehended. Exploiting plausibly random variation in the roll-out of sanctions, we estimate econometrically that exposure to penalties reduced the 18-month re-apprehension rate for males by 4.6 to 6.1 percentage points o of a baseline rate of 24.2%. These magnitudes imply that sanctions can account for 28 to 44 percent of the observed decline in recidivism in apprehensions. Further results suggest that the drop in recidivism was associated with a reduction in attempted illegal entry. 1 Introduction In this paper, we examine how penalties on illegal border crossings aect the behavior of undocumented immigrants from Mexico. Two-thirds of these migrants enter the U.S. by crossing the U.S.-Mexico land border (Passel and Cohn, 2016). Until 2005, over 95% of Mexican nationals apprehended while trying to cross the border unlawfully were granted voluntary return, they were released into Mexico and subject to no further repercussion. under which In 2008, the U.S. Border Patrol sought to deter new and repeated attempts at illegal entry by replacing voluntary return with sanctions imposed under a Consequence Delivery System (CDS) (Capps et al., 2017). By 2012, the share of apprehended Mexican nationals granted voluntary return was down to 15%. CDS sanctions include administrative consequences, which complicate obtaining a legal U.S. entry visa in the future; programmatic consequences, which disrupt smuggling networks by relocating a migrant far from the point of capture before release into Mexico; and criminal consequences, which entail prosecution in U.S. courts. The CDS rollout was followed by a sharp decline in recidivism: the share of apprehended migrants re-apprehended within the next 18 months fell from 28.1% in 2005 to 17.5% in 2012. We use administrative records from the U.S. Border Patrol on apprehended migrants to estimate the impact of the CDS on recidivism in apprehensions. 1 The CDS rollout capped 20 years of the U.S. intensifying border enforcement (Roberts et al., 2013). Between 1992 and 2007, the U.S. government quadrupled the number of Border Patrol agents (Figure A1). Despite these eorts, the U.S. population of undocumented immigrants grew from 3.5 million in 1990 to 12.2 million in 2007 (Passel and Cohn, 2016). Measuring border enforcement using Border Patrol manpower (Hanson and Spilimbergo, 1999), previous literature nds that this earlier border buildup had at most modest negative impacts on attempted illegal entry (Gathmann, 2008; Orrenius and Zavodny, 2005; Massey et al., 2016), though impacts are larger for the less-skilled (McKenzie and Rapoport, 2010), and tightened border security did induce undocumented migrants already in the U.S. to remain in the country (Angelucci, 2012). 2 Since 2007, the scene at the border has changed. Apprehensions are down sharply (Figure A2), and, after decades of growth, the U.S. population of undocumented immigrants from Mexico fell from These records are maintained by U.S. Customs and Border Protection, the agency of the U.S. Department of Homeland Security that oversees the U.S. Border Patrol. Mexican nationals were long the majority of those apprehended at the border. Since 2013 the share of Central Americans in apprehensions has grown (Figure A2). In strong contrast to Mexican immigrants, Central Americans are likely to seek asylum, which invokes distinct U.S. immigration procedures (Amuedo-Dorantes et al., 2015). 1 2 1 6.9 million in 2007 to 5.8 million in 2014 (Passel and Cohn, 2016). Although the Great Recession and demographic shifts in Mexico explain some of these changes (Hanson et al., 2017), the U.S. recovery has not brought a rebound in illegal entry, and population changes are too gradual to account for the late-2000s immigration dropo. Among recent security measures, Feigenberg (2017) and Allen et al. (2018) nd that a border fence constructed along the land portion of the U.S.-Mexico border between 2006 and 2009 has deterred illegal entry along aected crossing routes, while cross-section surveys of apprehended migrants reveal no connection between exposure to the CDS and migrant 3 plans for illegal entry (Amuedo-Dorantes and Pozo, 2014; Martinez et al., 2018). Because administrative records on apprehensions have been unavailable previously, the literature lacks a longitudinal perspective on how Border Patrol sanctions aect migrant behavior. Our data now permit such an analysis. We face two empirical challenges. A rst is that a migrant's characteristics may be correlated both with the sanctions he receives and his re-entry decision. Exploiting the richness of the apprehension records, we control for interactions of migrant age, birthplace, prior apprehensions, and location and timing of capture. Identication is based on dierences in recidivism among paired groups of migrantsone of which is sanctioned and one of which is notwhere the two groups share the same birthplace, birth cohort, and apprehension history, and were apprehended in the same location on the same day. Our identifying assumption is that during the CDS rollout, the assignment of penalties was as good as random, conditional on the controls. While implementing the CDS, individual Border Patrol sectors had discretion in delivering sanctions. 4 Capacity in sanc- tioning was aected by backlogs of migrants awaiting processing and available space in detention facilities (Capps et al., 2017). We estimate treatment eects by exploiting high-frequency variation 5 in capacity constraints. A second empirical challenge is that, analogous to literature on criminality, we do not observe recidivism in illicit activity (attempted entry) but rather in imprisonment (apprehension). Our estimated impact of the CDS on recidivism thus combines the impact on re-attempted illegal entry and the impact on the probability of capture, conditional on re-attempting entry. We exploit the structure of incentives to avoid capture to help resolve this ambiguity. Our results imply that the CDS can account for one-third of the reduction in re-apprehension Other recent work examines the consequences of changes in immigration enforcement in the U.S. interior (e.g., Orrenius and Zavodny 2009; Bohn et al. 2014; Orrenius and Zavodny 2015; Hoekstra and Orozco-Aleman 2017). Along the Southwestern border, these sectors are: San Diego and El Centro (CA); Yuma and Tucson (AZ); and El Paso, Big Bend, Del Rio, Laredo, and the Rio Grande Valley (TX). For recent work on recidivism in criminal arrests, see Bhuller et al. (2016); Heller et al. (2016); Agan and Makowsky (2018). For a review of the previous literature, see Chaln and McCrary (2017). 3 4 5 2 rates over 2008 to 2012. Less-severe administrative sanctions have similar eects on recidivism as more-severe programmatic and criminal sanctions. Our results are unaected by controlling for individual characteristics, as consistent with our identifying assumption that variation in the application of the treatment was due to Border Patrol capacity constraints and not to the identity of the migrant. The stability in our estimates, despite the large number of controls, suggests limited scope for selection-on-unobservables to explain our ndings (Altonji et al., 2005; Oster, 2017). 2 Data 2.1 CBP Administrative Data on Border Apprehensions Our data cover all apprehensions of individuals attempting to enter the U.S. without authorization between ports of entry along the U.S.-Mexico border. After apprehension, the Border Patrol ngerprints migrants and takes their biographical information (GAO, 2017a). Fingerprint records allow us to track individual migrants over time. Because Border Patrol policy shields women and minors from some sanctions, reserves voluntary return for Mexican nationals, and imposes severe sanctions on the few migrants with many prior apprehensions (GAO, 2017a), we restrict the sample to male Mexican nationals 16 to 50 years of age with six or fewer previous apprehensions. We study 2008 to 2012, which spans the CDS rollout. Our sample contains 0.97 million apprehensions, which represent 79.5% of Mexican nationals apprehended at the border over 2008-2012. As seen in Appendix Table A1, the Tucson sector of the Border Patrol accounts for 53% of sample apprehensions. When border enforcement intensied in the 1990s, many migrants switched from single-day crossings near major border cities to multi-day crossings in rugged eastern Arizona (Massey et al., 2016), which lies within the Tucson sector. Nearly half of apprehensions occur during the rst four months of the year, as migrants arrive for seasonal work in agriculture and construction, and seek to avoid extreme weather. Apprehensions are evenly distributed across days of the week and times of day, reecting randomness in the timing of apprehensions that results from multi-day crossings. Whereas relatively few apprehended migrants are from Mexican states that border the U.S. (11.5%) or states in Mexico's far south (7.4%), a majority are from central Mexico (68.0%), consistent with historical patterns (Massey et al., 1994). Figure 1 plots the fraction of apprehended migrants who were re-apprehended in the following 3 3, 6, 12 or 18 months. While re-apprehension rates are roughly stable between 2005 and 2009, they drop sharply in ensuing years, which span the CDS rollout. The 2005-to-2012 decline in the re-apprehension rate is 8.2 p.p. at the 3-month frequency and 10.6 p.p. at the 18-month frequency. 6 Figure 1: Re-apprehension Rate following Initial Apprehension for Male Mexican Nationals Re-apprehension rate (%) 30 26 22 18 14 2005 2006 2007 2008 2009 Within 3 months Within 12 months Notes : 2010 2011 2012 Within 6 months Within 18 months Data from CBP administrative records showing the re-apprehension rates for the study population of male Mexican nationals 16 to 50 years of age with six or fewer previous apprehensions. 2.2 The Consequence Delivery System A foreign national who enters the U.S. without authorization is in violation of U.S. law. The Border Patrol may refer any apprehended migrant for criminal prosecution. For decades, however, standard practice regarding apprehended Mexican nationals was to oer voluntary return (Roberts et al., 2013), under which a migrant forgoes the right to appear before a judge and agrees to depart the U.S. after transport to the border. He avoids formal removal and thereby escapes legal repercussions from his oense. Historically, the sheer volume of apprehensions, which averaged 1.2 million annually over 1999 to 2007 (DHS, 2008), in part justied voluntary return. Another motivation for the practice was that during its early history an implicit mandate of the Border Patrol was to help regulate the supply of low-skilled labor in the U.S. border region. Calavita (2010) and Roberts et al. (2013) document that log books of U.S. Border Patrol agents in the 1940s and 1950s cite the demand for This decline in recidivism is corroborated by the EMIF-Norte (Survey of Migration in the Northern Border), which surveys apprehended migrants returning to Mexico (Appendix Figure A3). The fraction of apprehended Mexican nationals stating they will re-attempt a border crossing within three months fell from 77% in 2009 to 49% in 2013. 6 4 farm workers in the Rio Grande Valley of Texas as a factor in determining the intensity of border enforcement. 7 Relatedly, Hanson and Spilimbergo (2001) nd that over the period 1970-1997 border enforcement weakened in the months following positive labor demand shocks to U.S. sectors that employed undocumented labor intensively. Under voluntary return, migrants may have interpreted border enforcement as existing to modulate the ow of labor across the border, rather than to prevent it altogether, suggesting that the illegality of crossing the border without authorization may not have been foremost in the minds of many prospective migrants Today, nearly all apprehended migrants are sanctioned under the CDS. Administrative consequences are delivered through a formal removal order (Rosenblum, 2013). Migrants processed for removal are not detained. Instead, they are processed by Border Patrol agents and transferred to the border for release into Mexico. Processing a removal order requires 90 minutes of time, compared to 15 minutes for a voluntary return (Capps et al., 2017). A removal order, which is tied to an individual's ngerprint record, precludes the migrant from applying for a legal entry visa for ve years or more, and counts as a prior infraction when dealing U.S. authorities in the future. Because many undocumented immigrants from Mexico have an application pending for a U.S. green card, this penalty is onerous. In the New Immigrant Survey, Massey and Malone (2002) document that 41% of Mexican nationals who obtained a U.S. green card in 2003 had previously crossed the U.S. border illegally. Administrative consequences thus raise the costs to migrants of entering the U.S. without authorization. The share of apprehended Mexican nationals subject to administrative consequences rose from 15.5% in 2008 to 73.9% in 2012 (Figure 2). Programmatic consequences are used to disrupt smuggling. Given the high probability of apprehension for a Mexican national attempting illegal entry at the border (40-60% during our sample 8 who may assist with several crossings in the event a rst period), most migrants hire a smuggler, try fails (Chávez, 2011). The main programmatic consequence is the Alien Transfer Exit Program (ATEP), under which an apprehended Mexican national, after being subject to other penalties, is repatriated to Mexico at a location far from his entry point. This complicates reconnecting with smugglers, who tend to specialize geographically along the border. Angelucci (2015) nds that the decision to send a migrant to the U.S. among poor households in Mexico is strongly responsive to As evidence of this history, the Border Patrol would frequently refer apprehended migrants to the Bracero Program (1949-1964), through which the U.S. provided temporary work visas to farm workers. See Amuedo-Dorantes and Pozo (2014), Massey et al. (2016), DHS (2017), and Roberts (2015; 2017). 7 8 5 random positive shocks to household income, consistent with nancial constraints limiting the ability of households to aord smuggling fees. The adverse shock of being subject to ATEP would limit the ability of some households to nance further attempts to cross the border by their members. The use of ATEP, which only applies to non-minor males, began in 2009 in the four western-most Border Patrol sectors and is now used in seven (of nine) sectors. Programmatic consequences applied to 14.8% of those apprehended in their initial year of 2009 and to 49.2% in 2012 (Figure 2). Figure 2: Rollout of Consequence Delivery System Apprehensions subject to CDS (%) 80 60 40 20 0 2005 2006 2007 2008 2009 2010 2011 2012 Administrative consequences Programmatic consequences Criminal consequences Notes : Data from CBP administrative records showing the share of apprehended male Mexican nationals 16 to 50 years of age with six or fewer previous apprehensions subject to sanctions. Under criminal consequences, an apprehended migrant is subject to prosecution. Most occur under Operation Streamline, under which a migrant is tried for misdemeanor unlawful entry and appears with a group of migrants for sentencing. Although sentences may be up to 180 days for a rst oense, rst-time oenders are typically sentenced to time served (while awaiting a hearing). If the migrant has many prior apprehensions or is suspected of non-immigration crimes, he may face Standard Prosecution in a U.S. federal district court, which involves sentences of up to two years 9 The imposition of criminal consequences was intended and possibly being tried for a felony oense. to signal the seriousness of the Border Patrol regarding border enforcement (Roberts et al., 2013). Analogous to the broken-windows theory of policing (e.g., Corman and Mocan, 2005), subjecting In the early 2010s, the average sentence under Standard Prosecution was 18 months (USSC, 2015). Long sentences may reduce recidivism mechanically by incapacitating the migrant. However, the use of Standard Prosecution remains uncommon, accounting for only 1.5% of cases in our data. 9 6 migrants to prosecution in court conveys that further apprehensions could bring harsher penalties. Applying criminal consequences requires the participation of the Federal Judiciary, the U.S. Attorney's Oce, the U.S. Marshal's Service, and Immigration and Customs Enforcement (ICE). Criminal consequences were applied to 8.3% of apprehensions in their initial year of use in 2009 and to by 2012 (Figure 2). Operation Streamline accounted for 83.5% 22.6% of these cases. Administrative and criminal consequences, aside from their legal repercussions, may impose additional psychic costs on migrants. The sanctions communicate unequivocally that border crossing is illegal, perhaps changing migrant perceptions of procedural justice surrounding border enforcement (e.g., Sunshine and Tyler, 2003; Tyler, 2004). Whereas the voluntary-return regime did not emphasize the criminality of illegal border crossing, the CDS regime does so strongly. Emphasizing the criminality of the act may have raised the disutility associated with being apprehended. After apprehension, Border Patrol ocers propose the combination of CDS sanctions a migrant receives. Of the 49.0% 35.4% of sample migrants over 2008 to 2012 subject to administrative consequences, were further subject to programmatic or criminal consequences (Table A2). Programmatic consequences, applied in in 51.0% 19.5% of sample apprehensions, also brought administrative consequences of cases. Because programmatic consequences only involve transport before release into Mexico, they do not mandate formal removal. Criminal consequences, applied in 8.8% of appre- hensions, entailed administrative consequences in 94.5% orders for migrants facing criminal prosecution. Not all court districts permit Streamline prose- of cases, as it is standard to issue removal cutions or Standard Prosecution of oenses tied to illegal entry. Because Border Patrol sectors in these districts tend to use programmatic consequences in lieu of criminal consequences, we combine programmatic and criminal consequences when analyzing these penalties. 2.3 Application and Rollout of the CDS Historically, the Border Patrol was a decentralized organization, with sector chiefs having autonomy in setting enforcement strategy (Calavita, 2010). The CDS originated in Border Patrol sectors whose leaders perceived that it might be eective in deterring illegal entry. It was not implemented across all sectors until late 2012 (Simanksi, 2013). During the CDS rollout, variation in its application arose in part from sector-level dierences in capacities for processing migrants. Applying sanctions is time intensive, and stang levels were initially insucient to impose penalties on all those apprehended. 7 Because some sanctions require assistance from other government entities, local resource availability in these entities also aected the delivery of penalties. We examine how these sources of variation in the application of the CDSalong with daily variation in the number of migrants attempting illegal entryhelped generate plausibly exogenous assignment of sanctions to apprehended migrants, conditional on their observable characteristics and the conditions of their apprehension. Discretion in application of consequence programs. During the CDS rollout, decision rules for applying sanctions were based on (i) the origin country of the migrant, (ii) the migrant's apprehension history, and (iii) whether the migrant was traveling with family members (GAO, 2017a). The highest priority for sanctions were migrants from countries other than Mexico with many previous apprehensions (or a record of criminality). The lowest priority were Mexican nationals with no prior apprehensions, and the next lowest priority were Mexican nationals with a few prior apprehensions. Our sample migrants were therefore relatively low priority for sanctions and ones for whom the Border Patrol would have maximal discretion in assigning penalties. Capacity constraints at the level of Border Patrol sectors. Costs in imposing sanctions meant that on high-trac days agents may have been unable to impose sanctions on all apprehendees. Consider ocer time that would have been required to transition fully from a system of voluntary removal to one based on administrative consequences (AC) in each year of the CDS rollout. For each Border Patrol sector, we compute the share of ocer time that would be absorbed by applying AC to all apprehended migrants on each day from 2008 to 2012 (see Appendix B for details). We use sectorday observations on the total number of apprehensions, sectoryear observations on the total number of agents, and an in-depth analysis of time use by Border Patrol agents in the early 2010s (GAO, 2017b). Figure 3 plots the results across Border Patrol sectors during the CDS rollout. In the four busiest sectorsEl Centro, Rio Grande Valley, San Diego, and Tucson, which account for 86.7% of sample apprehensionsmost days would have required close to or more than 100% of agent time to apply AC to all apprehended migrants. Such constraints created substantial day-to-day variation in the fraction of migrants receiving sanctions. Over 2008 to 2012, the standard deviation of the daily fraction of apprehensions subject to AC was 0.19 in El Centro (mean of 0.25), 0.32 in the Rio Grande Valley (mean of 0.38), 0.08 in San Diego (mean of 0.19), and 0.39 in Tucson (mean of 0.51). 8 Figure 3: Capacity Constraints in Switching to Administrative Consequences Big Bend Del Rio El Centro El Paso Laredo Rio Grande Valley San Diego Tucson Yuma 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 2008 2009 2010 2011 2012 0 .2 .4 .6 .8 1 fraction of days requiring more than 100% of agent time to convert all apprehensions from VR to AC Each row shows the fraction of days in which Border Patrol agent time required to impose administrative consequences (AC) on 100% of those apprehended (up from the 2007 baseline of 8%) would exceed all available agent hours for a Border Patrol sector. Estimates are based on equation (4) in Appendix B. Notes: Capacity constraints in partner agencies. While the Border Patrol applies administrative consequences, it relies on other agencies to deliver programmatic and criminal consequences. Applying ATEP may require ICE buses and drivers. With criminal consequences, the Border Patrol requires the U.S. Marshal's Service to transport migrants to court, ICE to hold migrants in detention, the U.S. Attorney's Oce to prosecute cases, and the U.S. federal judiciary to hear cases (GAO, 2017a). Because these agencies face many demands, they sometimes lack resources for matters related to border apprehensions. In the early 2010s, courts requested that the Tucson sector, which rst launched Streamline prosecutions, limit Streamline cases to approximately 70 per day (Capps et al., 2017). Over 2008 to 2012, daily apprehensions in Tucson were above this cap on 99.3% of days. 9 Figure 4: Determinants of which Migrants Receive Sanctions Fixed Effects: Fixed Effects: Age + Birth State + Prior Apprehensions Age + Birth State + Prior Apprehensions Sector + Fiscal Year + Month Sector + Fiscal Year + Month …+ Day of Week + Time of Day …+ Day of Week + Time of Day …+ Age + Birth State …+ Age + Birth State …+ Prior Apprehensions …+ Prior Apprehensions 0 .1 .2 .3 0 R-squared in explaining Administrative Consequences .1 .2 .3 .4 R-squared in explaining Programmatic and Criminal Consequences (A) LPM for AC (B) LPM for PC/CC Each row of these gures reports the R from an OLS regression of a binary indicator for the given CDS sanction on the covariates listed in the row title. Each covariate is an exhaustive set of dummy variables for the given category; the ... indicate the addition of covariates in the given row to covariates in the prior row with the ultimate row including the full set of controls. 2 Notes: If variation in the application of the CDS was due more to prevailing conditions in a given Border Patrol sector than to migrant characteristics, these characteristics should play little role in determining whether a migrant received sanctions. Figure 4 reports R2 from a linear probability model (LPM) regressing an indicator for whether an apprehended migrant is subject to sanctions on controls for the location and the time of the apprehension, the demographic characteristics of the migrant, and the migrant's prior apprehensions. In panel A of Figure 4, the dependent variable indicates whether a migrant is subject to administrative consequences (AC). In row 1, we include only dummies for age, birth state, and prior apprehensions, which yields an R2 of 0.055. 10 We then replace these dummies with indicators for the sector, year, month, day of week, and time of day of the apprehension in row 3, the R2 jumps to 0.276. and prior apprehensions in row 5 raises the R2 Adding back in dummies for migrant birth state, age, modestly to 0.323. Panel B repeats these regressions for programmatic/criminal consequences (PC/CC). Again, age, birth-state, and prior-apprehension dummies have little explanatory power, whereas the sector and timing of the apprehension have substantial explanatory power. During the CDS rollout, where and when a migrant was apprehended played a dominant role in determining whether he was subject to sanctions. Gronau (1998) shows that the R in a LPM equals the dierence between the average predicted probability in the two groups (i.e., how much the covariates dierentiate CDS sanctioned migrants from non-CDS sanctioned migrants) and claries why the R in a LPM is less likely to approach 1 than in the case of a continuous outcome. 10 2 2 10 3 Empirical Specication 3.1 Regression framework We evaluate how being subject to the CDS aects the likelihood that an apprehended migrant is re-apprehended in the future. Our specication is yist+τ = β × CDSist + f (Xit , αs , αt ) + ist where s yist+τ at time t 18 months; is an indicator for whether migrant i who is apprehended in Border Patrol sector is re-apprehended anywhere along the border within CDSist (1) τ periods, for τ = 3, 6, 12, or is dened alternatively as an indicator for whether the migrant was subject to administrative consequences (AC) at apprehension, an indicator for whether the migrant was subject to any consequences at apprehension, or a vector that includes the AC indicator and an indicator for whether the migrant was subject to programmatic/criminal consequences (PC/CC) at apprehension; Xit includes indicators for the migrant's age cohort at apprehension, birth state in Mexico, and number of prior apprehensions; αs indicates the Border Patrol sector of apprehension; αt describes the timing of apprehension (year, month, day of week and time of day); the manner in which we interact the control variables; and ist f (·) characterizes is a disturbance term than captures unobserved variables that aect the likelihood of re-apprehension. To control for variables that may be related both to whether the Border Patrol sanctions a migrant and to whether he re-attempts illegal entry, we dene αs , and αt . f (·) to generate interactions among Xit , We rst interact only the sector, year and month of apprehension, then add interactions for day of week and time of day of apprehension, and then add migrant characteristics. Under the most complete set of interactions, we allow sanctions to be correlated with shocks that aect the migrant's re-entry decision, as long as these can be modeled by sector×calendar-date×age×birthstate×prior-apprehension interactions. Our approach to identication would be invalid if there are additional characteristics of a migrant that aect his decision to re-attempt illegal entry and the decision of the Border Patrol to impose sanctions on him. Because agents decide whether to impose sanctions on a migrant in a matter of minutes, our assumption that, conditional on the controls, the assignment of consequences to a migrant was as good as random may not be unreasonable. Comparing estimates of β across specications reveals the sensitivity of the treatment eect 11 to increasingly more-expansive controls for the characteristics of the migrant. We formally evaluate possible bias due to unobservables using the approach proposed by Altonji et al. (2005) and generalized by Oster (2017). 3.2 Interpreting the CDS treatment eect In (1), we estimate how the imposition of sanctions aects the probability that a migrant is reapprehended in the future. This probability can be written as P (R1 ) = P (E1 |C0 ) × P (A1 |C0 , E1 ) where P (R1 ) P (E1 |C0 ) is the probability that a migrant apprehended at t=0 (2) is re-apprehended at is the probability that the migrant re-attempts illegal entry at having been apprehended and having faced consequence no sanction and apprehended at 1 indicates a sanction; and t = 1, C0 = {0, 1} P (A1 |C0 , E1 ) at t = 1, t = 0, conditional on where t = 1. 0 indicates is the probability that the migrant is conditional on having been apprehended and faced consequence and on re-attempting illegal entry at t = 1; C0 at t=0 The treatment eect that we estimate in (1) is β̂ = P (E1 |1) × P (A1 |1) − P (E1 |0) × P (A1 |0). (3) We (weakly) underestimate the impact of the CDS on the probability that a migrant re-attempts illegal entry if P (A1 |1, E) ≥ P (A1 |0, E1 ). How would sanctions aect the incentive for a migrant to reduce his apprehension probability (e.g., by hiring a higher quality smuggler)? Consider two migrants who each have been apprehended once, where one was subject to sanctions and the other was not. Under administrative or criminal consequences, the sanctioned migrant has lost the ability to seek a legal entry visa for ve years or more (and voided any visa application under review). Because the risk of felony prosecution for a second apprehension is low, he would seem to have less at stake in a subsequent crossing than the non-sanctioned migrant, who can still apply for a legal visa (or pursue a visa application under consideration). After a single apprehension, the sanctioned migrant would thus seem to have a weaker incentive to reduce his probability of apprehension on a subsequent entry attempt, when compared to the non-sanctioned migrant, meaning P (A1 |1, E) ≥ P (A1 |0, E1 ). At some point, accumulating apprehensions may expose a previously sanctioned migrant to risk of felony conviction. Although standard prosecution is rare within our sample, applying to just 12 1.5% of apprehended migrants, a sanctioned migrant with multiple prior apprehensions may have a relatively strong incentive to avoid apprehension in a next crossing attempt, when compared to a non-sanctioned migrant with the same number of prior apprehensions. This discussion suggests that one way to gauge the sensitivity of our results to the confounding eects of changes in the apprehension probability is to allow the treatment eect to vary with the number of prior apprehensions. Estimating the CDS treatment eect for migrants with a single prior apprehensions is where it seems most likely that P (A1 |1, E) ≥ P (A1 |0, E1 ) and that we underestimate the impact of the CDS on the probability that a migrant re-attempts illegal entry. 11 4 Empirical Results 4.1 Administrative Consequences Figure 5 reports estimation results for (1) using our sample of male Mexican nationals with six or fewer previous apprehensions who were apprehended over 2008 to 2012. The outcome variables, organized in panels A-D, are indicators for whether the migrant was re-apprehended in the following 3 or 18 months (with results at 6 and 12 month horizons reported in the Appendix). The rst treatment we consider is whether the migrant was subject to administrative consequences (AC) at apprehension, shown in panels A and B. In row 1, the control variables are complete interactions among dummy variables for the Border Patrol sector, the scal year, and the month in which the apprehension occurred. In row 2, we add interactions with indicators for the time of day and day of week of the apprehension; in row 3, we add interactions with indicators for the migrant's age and birth state; and in row 4, we add interactions with indicators for the migrant's number of prior apprehensions. In row 5, we go further and introduce xed eects for the sector and calendar date of the apprehension occurred (e.g., January 12, 2010 in Tucson); and in row 6, we interact those dummy variables with migrant age, birth state, and number of prior apprehensions. gures include point estimates and 95% condence intervals. The Tables A3 and A4 present the full regression output including sample size, which varies across specications as interactive xed-eect A factor aecting the ability of a migrant to nance multiple border-crossing attemptsand therefore also aecting the impact of the CDS on the apprehensions probabilityis credit constraints (Angelucci, 2012). If successive attempts to cross the border each end in apprehension, a migrant may be progressively less able to marshal the resources to pay smuggling fees on each subsequent crossing. Consistent with this reasoning, data from the EMIF-Norte (see note 6) reveal that the likelihood that a recently apprehended migrant used a coyote on his most recent crossing attempt is negatively correlated with the number of times he has been apprehended in the recent past (Roberts, 2015). 11 13 cells without variation in sanctions across migrants are omitted. We cluster standard errors by the sector-year-month combination (270 clusters), to account for the common exposure of migrants to policies dened at the sector level at a given time. 12 Consider the estimate of the AC treatment on the 3-month re-apprehension rate in row 1 of panel A. The value of −0.064, which is very precisely estimated, indicates that migrants subject to administrative consequences were 6.4 p.p. less likely to be re-apprehended in the next 3 months (compared to a 2008 re-apprehension probability of 0.226). As we allow for more interactions between time and location of apprehension and migrant characteristics, there is essentially no change in the estimated treatment eect. In row 4, with the full set of controls, the estimate is −0.063. The more- exhaustive sector-date xed eects in rows 5 and 6 do not materially change this point estimate, which is −0.066 in row 6. Table A3 explores impacts on the likelihood of re-apprehension at longer time horizons, and again the estimated treatment eects are insensitive to expanding the set of controls. The treatment eect diminishes modestly as we expand the re-apprehension time horizon. For results with exhaustive xed eects in row 6 of Figure 5 (panels A and B), the AC treatment eect falls from 0.269). −0.066 at 3 months to −0.046 at 18 months (2008 re-apprehension probability of This attenuation could indicate that some of the sanction impact is psychological, where the trauma diminishes with time. Alternatively, it may take migrants time to build up the resources to undertake a second crossing, meaning impacts are lower at longer horizons. Overall, the stability of coecients across xed-eect specications provides prima facie evidence against a large role for selection on unobservables to explain our ndings. Consider panel A, where the R2 increases from 0.076 in row 5 with sector-date xed eects to 0.409 in row 6 when adding interactions with age, birthplace, and number of prior apprehensions. 13 Despite this large increase in explanatory power, the treatment eect of AC sanctions remains eectively unchanged, going from -0.064 to -0.066. In other words, adding a large number of observable determinants of re- apprehension does not change the observed impact of AC. This pattern holds across our ndings 14 (see Table A3) and points to very limited selection on unobservables based on the Oster (2017) test. Table A3 also reports more conservative condence intervals and p-values based on clustering at the level of the nine border patrol sectors. Given the small number of clusters, we apply the wild cluster bootstrap of Cameron et al. (2008). All point estimates remain signicant at the 1% or 5% level. This increase in R is understated insomuch as the xed eects fully absorb cells of observations within which there is a single migrant. These singleton cells do not contribute to the identifying variation in the point estimates, are omitted from the sample size in Table A3, and thus do not add  to the R . The Oster (2017) δ-statistics are computed as δ = × , where β is the coecient estimate 12 13 2 2 14 βc βu −βc 14 2 2 Rc −Ru 2 0.3×Rc c Figure 5: Impact of Exposure to CDS Sanctions on Probability of Re-apprehension Administrative Consequences (A) Re-apprehension within 3 months (B) Re-apprehension within 18 months Fixed Effects: Sector-Fiscal Year-Month Fixed Effects: Sector-Fiscal Year-Month …-Day of Week-Time of Day …-Day of Week-Time of Day …-Age Category-Birth State …-Age Category-Birth State …-Prior Apprehensions …-Prior Apprehensions Sector-Date Sector-Date …-Age Cat.-Birth State-Prior Apps. …-Age Cat.-Birth State-Prior Apps. -.1 -.08 -.06 -.04 -.02 0 -.1 βAdmin. Consequences +/- 2 x std. error -.08 -.06 -.04 -.02 0 βAdmin. Consequences +/- 2 x std. error Any Consequences (C) Re-apprehension within 3 months (D) Re-apprehension within 18 months Fixed Effects: Sector-Fiscal Year-Month Fixed Effects: Sector-Fiscal Year-Month …-Day of Week-Time of Day …-Day of Week-Time of Day …-Age Category-Birth State …-Age Category-Birth State …-Prior Apprehensions …-Prior Apprehensions Sector-Date Sector-Date …-Age Cat.-Birth State-Prior Apps. …-Age Cat.-Birth State-Prior Apps. -.1 -.08 -.06 -.04 -.02 βAny Consequences +/- 2 x std. error 0 -.1 -.08 -.06 -.04 Each row of these gures reports the point estimate and 95% condence interval on the dummy variable for administrative consequences, in panels (A) and (B), or any consequences, in panels (C) and (D), in an OLS regression for re-apprehension within the next 3 months (panels A and C) or 18 months (panels B and D). Each row is a separate regression controlling for the xed eects (FE) listed in the row title. These FE enter interactively where the ... indicate the addition of FE in the given row to FE in the prior rows. Standard errors are clustered by sector-scal year-month (270 clusters). Notes: 15 -.02 βAny Consequences +/- 2 x std. error 0 The calculation here (comparing rows 5 and 6) suggests that selection on unobservables would have to be 482 times larger than selection on observables, which far exceeds the rule-of-thumb cuto of 1 for observational studies. A similarly large Oster δ -statistic of 178.6 arises when comparing the specications in rows 4 and 2. In short, selection on unobservables would have to be implausibly large to explain the eects on recidivism that we nd. 4.2 Other Consequence Programs In panels C and D of Figure 5, we repeat the analysis in the upper two panels, redening the treatment as an indicator for any consequence, including administrative (AC), programmatic (PC) or criminal consequences (CC). This broader denition of any CDS sanction implies a larger reduction in recidivism than the AC treatment alone. estimated eect of treatment increases from For the most demanding specication in row 6, the −0.064 for AC alone to (AC, PC, and or CC) at the 3-month horizon and from −0.046 to −0.081 −0.061 for any consequence at the 18-month horizon. Like the AC treatment eect, the any-consequence eect is stable across xed-eect specications. Although the any-consequence treatment implies a larger reduction in recidivism, there is little dierence in the eects of AC versus PC/CC. This can be seen in panel A of Table A5, which shows that when entered as separate indicators the AC and PC/CC treatments have statistically indistinguishable eects across most time horizons, with the distinct eects of each being slightly smaller than the any-consequence treatment. At the 3-month horizon, for example, the AC coecient is −0.066, which we fail to reject being dierent from the PC/CC coecient of −0.060 (p-value of 0.36). We nd analogous patterns for the 18-month horizon (see panel A of Table A6). Given the uneven application of PC and CC across sectors and time (see Section 2.3), we exploit dierent sources of spatial and temporal identifying variation when comparing across consequences. This prevents evaluating the conceptually ideal comparison across AC and PC/CC within-person. These dierent sources of identifying variation may explain why the combined treatment eect of AC and PC/CC is less than two times the AC or the PC/CC treatment alone (see panel B of Tables A5 and A6). Alternatively, the positive interaction between the AC and PC/CC treatments may be evidence of diminishing returns to additional sanctions, such that the combined eects of the full set of sanctions is less than two times the eect of AC or PC/CC alone. with additional controls, β is the reference coecient estimate without those controls, and R and R are the corresponding R from the respective regressions. δ is innite (or undened) when β = β and R > R . u 2 c 16 u 2 c 2 c 2 u 2 u To gauge the economic signicance of the estimates, consider the impact of the CDS on recidivism in apprehensions 18 months after capture, as shown in row 6, panel B of Figure 5 for administrative consequences and row 6, panel D of Figure 5 for any consequences. Between 2008 and 2012, recidivism in apprehensions declined by 9.6 p.p. at the 18-month horizon (Figure 1). the 2008-to-2012 increase in the incidence of the AC treatment of the any-consequence treatment by 69.8 p.p. (85.2 − 15.4) is a reduction in recidivism equivalent to 58.5 p.p. (73.9 − 15.4) and of (see Table A2), the AC treatment eect 28.0% ([.046 × .585]/[.269 − .175)]) of the observed de- cline, and the any-consequence treatment eect is a reduction in recidivism equivalent to ([.061 × .698]/[.269 − .175)]) With 44.4% of the observed decline. CDS sanctions thus account for a substantial share of the observed decline in recidivism in apprehensions. 4.3 Recidivism in Apprehensions versus Recidivism in Attempted Illegal Entry Following Section 3.2, we connect our results on recidivism in apprehensions to recidivism in attempted illegal entry by allowing the CDS treatment eect to vary with the migrant's number of prior apprehensions. A sanctioned migrant with a single prior apprehension would seem to have a weaker incentive to avoid apprehension on a subsequent crossing than a non-sanctioned migrant with a single apprehension, in which case our estimated impact of sanctions on recidivism in apprehensions (weakly) understates the impact of sanctions on recidivism in attempted illegal entry. Table A7 reports extended regression results for the row 4 specications in Figure 5. Panel A shows treatment eects for administrative consequences, and panel B does so for any consequences, where we allow these eects to vary according to whether a migrant has one, two, three, or four to six prior apprehensions. For re-apprehension within 18 months, shown in column 4, the estimated any-consequence treatment eect for migrants with a single prior apprehension (−0.056) is smaller in absolute value than that for migrants with two prior apprehensions (−0.075 = −0.056−0.019), where this dierence is statistically signicant. This pattern holds for both administrative consequences and results are comparable at other time horizons for re-apprehension. Taking the CDS treatment eect for migrants with a single prior apprehension as our most conservative estimate, we obtain magnitudes that are only slightly smaller than in Figure 5. In going from the full sample to the single-apprehension sample, the reduction in re-apprehension rates induced by any consequences falls from −0.079 to −0.074, 17 at the 3-month horizon, and from −0.059 to −0.056, at the 18-month horizon. These results suggest limited scope for the impact of sanctions on the probability of apprehension to confound our estimates, indicating that the CDS treatment eect on recidivism in apprehensions is informative about the (more-policy-relevant) CDS treatment eect on recidivism in attempted illegal entry. 5 Discussion Undocumented immigration is a highly contentious issue in the U.S. Some critique the government's eectiveness in securing borders against illegal entry, while others object to the treatment of immigrants by authorities. To nd viable immigration policies, we need to know the relative eectiveness of alternative approaches for enforcing the border. Early research on border enforcement inspired pessimism about eorts to deter undocumented immigration. the 1990s and early 2000s, so too did illegal entry. As enforcement manpower grew in Since the late 2000s, the Border Patrol has shifted its enforcement strategy from simply sending more ocers to the border to constructing new physical barriers and imposing sanctions on apprehended migrants. Sanctions work by reducing the viability of legal immigration in the future and by raising the risk of incarceration. They represent a potentially lower-cost deterrent than, say, constructing much-more-substantial border barriers, and possibly a less-invasive deterrent than large-scale deportations in the U.S. interior. We nd that sanctions have large negative impacts on recidivism in apprehensions and, plausibly, on recidivism in illegal entry. Further, simple-to-apply administrative sanctions appear to be equally as eective as sanctions that entail costly prosecutions or relocations of migrants. 18