Fear and the Safety Net: Evidence from Secure Communities

LARGE-SCALE deportation of immigrants may generate fear and insecurity among minority communities, spilling over to co-ethnic citizens (Lopez et al., 2018). If fear of deportation is not limited to noncitizens, can immigration enforcement impede other government objectives?

This study examines the relationship between immigration enforcement and poverty reduction efforts through the maintenance of a robust set of means-tested social insurance (MTSI) programs. Prior work has shown that MTSI programs have large welfare benefits for vulnerable populations (Braun et al., 2017), but such programs still face incomplete take-up, or the nonreceipt of benefits by individuals who are entitled to them (Currie, 2006; Eurofound, 2015).¹ Incomplete take-up varies across ethnic groups (Bertrand et al., 2000; Borjas, 1992) and is higher among groups facing deportation risk, such as Hispanics in the United States (Morin et al., 2012).

We contribute to a deeper understanding of the connection between social and immigration policy by exploring whether noncitizen expulsions influence co-ethnic citizen participation in MTSI programs, even though citizens are not personally at risk of deportation. We focus on such programs because assessment of eligibility and benefits typically require applicants to share household details with the government, including the immigration status of other, nonrecipient household members. As a result, mixed-status families and communities, fearful that disclosing information will increase the likelihood of deportation of network contacts, may forgo participation.

To formally examine this relationship, we study a new and far-reaching immigration enforcement program in the United States that dramatically increased deportations. The program, known as Secure Communities (SC), was administered by the U.S. Immigration and Customs Enforcement Agency (ICE), and we study its nationwide rollout between 2008 to 2014. We exploit the staggered roll-out of SC across counties and its differential impact on Hispanics in a triple-differences design. By interacting race and ethnicity indicators with timing of SC activation, we compare program participation for Hispanic households within a given location to program participation for non-Hispanic white and black households, net of counties that had not yet activated, before versus after SC activation. The triple-differences identification assumption is plausible, requiring that there be no location-specific shocks timed with the staggered SC roll-out and differentially influencing the dynamic path of safety net outcomes for Hispanics.

Our primary outcomes of interest include two of the largest federally funded means-tested safety net programs in the United States—Supplemental Security Income (SSI) and the Supplemental Nutrition Assistance Program (SNAP). In addition to their fiscal importance, we focus on these federal programs because they exclude unauthorized individuals and have uniform eligibility requirements across locations. These two features allow us to estimate indirect effects of enforcement on citizens, while purging our estimates of potentially confounding effects of local programmatic changes that may coincide with SC activation. In addition, both SSI and SNAP include recertification processes, so immediate changes in enrollment among new or prior users are feasible to detect. For both outcomes, we examine responses among the risk set of economically fragile households that are more likely to be linked with Hispanic noncitizens, defined as those in which the head of household earned less than a high school degree.²

We combine data on SNAP and SSI participation from the American Community Survey (ACS) and the Panel Study of Income Dynamics (PSID) with detailed information on the timing of SC implementation and the universe of over two million detainers (“immigration holds”) issued by ICE under SC between 2008 and 2013. These data contain information on the county of issue, crime severity, and country of origin of each arrested individual.

We find that SC activation is associated with substantial declines in safety net participation among Hispanic citizen households in both the ACS and PSID. Hispanic-headed families are 2.1 percentage points less likely to participate in SNAP and 1.7 percentage points less likely to participate in SSI after activation of SC. The estimated spillover effect of deportations on citizens is about 20%–50% as large as the direct effects of deportation on noncitizens reported in the literature (e.g., Watson, 2014; Kaushal & Kaestner, 2005).

A number of findings suggest these results are indeed causal. First, consistent with the parallel trends assumption underlying our triple-differences approach, we find no sharp changes in the evolution of our outcome variables prior to SC activation for Hispanics relative to other groups. Second, our preferred specification includes a full set of race-by-state fixed effects to address the potential concern that states may vary in policies towards minority groups, state-by-year fixed effects to account for changes in state-level immigration enforcement such as the enactment of omnibus enforcement bills (Amuedo-Dorantes & Arenas-Arroyo, 2017), and race-by-year fixed effects, along with race/ethnicity-specific state-level employment changes during the Great Recession (Kochhar et al., 2011; McKernan et al., 2014). While we primarily focus on specifications with state-by-year fixed effects given important state-level variation in immigration policy, our findings are robust to the inclusion of county-by-year and county-by-race fixed effects. Third, we show SC only affected those indirectly treated through their connections to those at greatest risk of deportation: Hispanic Americans. Results on program participation for non-Hispanic blacks or whites are often oppositely signed and not statistically significant.

We next turn to mechanisms. We demonstrate that neither the composition of respondents, employment status or migration of respondents, nor under-reporting of welfare receipt responds to SC activation, eliminating those possible explanations for our results. We also find that information frictions are unlikely to explain the effect of SC on Hispanic Americans. Instead, we report a set of findings that, together, suggest that deportation fear plays a crucial role in explaining our findings.

Our results relate to a vast literature in economics on peer and social network effects, program participation, and immigration enforcement.³ Bertrand et al. (2000) and Borjas and Hilton (1996) document the presence of network effects in welfare program participation among ethnic groups, emphasizing the role of information and destigmatization. Prior research examining the effects of immigration enforcement on program participation primarily focuses on the direct effects among noncitizens (see, e.g., Amuedo et al., 2018; Cascio & Lewis, 2019; Pedraza & Zhu, 2015; Watson, 2014; Yoshikawa, 2011) and immigrant migration decisions (see, e.g., Lessem, 2018). We extend the literature by focusing on the indirect effect of immigration enforcement on citizens who alter their participation in programs for which they qualify, despite lacking personal deportation risk, a chilling effect of a different nature with important policy implications. We also document how the relatively understudied mechanism of fear, like information, can spread through social networks and influence behavior.

Our article is structured as follows. The next section describes the SC program in detail. Section III discusses eligibility rules for public programs in the study. Section IV describes our conceptual framework and section V outlines our data and identification strategy. Section VI reports the results, section VII discusses potential mechanisms, and section VIII concludes.

Secure Communities was an immigration enforcement program administered by ICE from 2008 to 2014, reactivated in 2017, and then revoked in 2021.⁴ The SC program empowered ICE to check the immigration status of anyone arrested by local law enforcement agencies through fingerprint analysis, alerting the Federal Bureau of Investigation (FBI) and Department of Homeland Security (DHS) of potential immigration violators.

Typically, when a person is arrested and booked by a state or local law enforcement agency, his or her fingerprints are taken and submitted to the FBI. The FBI runs these fingerprints in order to conduct a criminal background check, which is forwarded to the state or local authorities. Prior to the implementation of SC, noncitizens in violation of immigration laws were identified by inmate interviews in local jails or prisons, performed by either federal officers under a policy known as the Criminal Alien Program (CAP) or local officers under formal written agreements with DHS, known as 287(g) agreements. These interviews were labor-intensive, such that federal and local officials authorized to conduct these interviews screened less than 15% of local jails and prisons, and in only about 2% of all U.S. counties (Cox & Miles, 2013). In contrast, under SC, fingerprints received by the FBI were automatically and electronically sent to DHS, where they were then compared against its Automated Biometric Identification System. If there was a fingerprint match and probable cause for removability based on other information, ICE then issued a “detainer” on the person. This detainer requested that the state or local law enforcement agency hold the individual for up to 48 hours to allow ICE to assume custody for the initiation of removal proceedings. As a result, individuals who may otherwise be released through the local legal system (such as those whose cases were dismissed or those who were released pretrial pending criminal proceedings) were detained via SC.

Notably, state and local jurisdictions could not easily opt out of SC.⁵ While SC initially required a memorandum of agreement (MOA) between the ICE and State Identification Bureau officials,⁶ governors in Massachusetts, New York, and Illinois ended their respective MOAs in the spring of 2011, leading ICE to determine that MOAs were “not required to activate or operate Secure Communities in any jurisdiction.”⁷ Shortly thereafter, John Morton, former Director of ICE, terminated all MOAs in August 2011 in the “Morton Memo,” stating MOAs had led to “substantial confusion” and that “ICE has determined that an MOA is not required to activate or operate SC for any jurisdiction.”⁸

When SC was activated, it was done so on a county-by-county basis due to various technological constraints. The program began on October 27, 2008, and SC was adopted in most counties by mid-2012 and fully activated across the entire country on January 22, 2013. Cox and Miles (2013) show that the timing of activation across counties is most strongly correlated with the Hispanic population, distance from the Mexican border, and whether a county had a 287(g) agreement with ICE, findings we return to when discussing our identification strategy below.

In response to SC, some jurisdictions known as “sanctuary cities” began to disobey detainer requests from ICE, arguing such detentions were unconstitutional and noting concerns that such practices would discourage immigrant cooperation with local law enforcement.⁹

In this study, we focus on participation in SNAP and SSI, two of the largest means-tested programs in the United States (Daly & Burkhauser, 2003; Duggan et al., 2016). While states have substantial discretion over eligibility rules and benefit levels for programs such as TANF and Medicaid, these features are uniformly set by the federal government for SSI and SNAP.¹⁰ In our context, this uniformity across geographies allows us to estimate our treatment effect using the staggered roll-out of SC. For instance, the interpretation of our results could be complicated if states changed benefit levels contemporaneously with SC activation.

Individuals need to meet various federal guidelines to receive benefits under SNAP.¹¹ Immigrants residing in the country without authorization are ineligible to receive SNAP benefits. In contrast, qualified legal immigrants (e.g., green card holders, parolees, and refugees) are eligible for SNAP if they have lived in the U.S. for five years, if they currently receive disability-related assistance, or if they are children under 18, in addition to the income and resource limits. To apply for benefits, individuals complete an application in-person or online, followed by an interview with a SNAP representative. In our context, immigration enforcement may affect participation because SNAP applications routinely ask for the names, citizenship status, and Social Security numbers of all household members, regardless of whether they are applying for benefits. Some states also ask for country of origin, date of entry, and alien registration number for each household member. Using this information, states verify the immigration status of each household member through DHS to reduce benefit fraud. An example of a state SNAP form is provided in appendix figure A1.

Almost all states assure applicants that their information will only be used to determine eligibility and will not be shared with ICE. For example, the California Department of Social Services states that immigration status “can only be used to determine . . . eligibility, and cannot be used for immigration enforcement.”¹² Nevertheless, anecdotes from advocacy groups and SNAP outreach coordinators suggests that SNAP applications have declined and that this decline has coincided with increased anti-immigration rhetoric.¹³

Supplemental Security Income (SSI) is administered by the Social Security Administration (SSA). To be eligible for SSI, individuals must be aged or disabled and meet federally mandated income and asset limits. Like SNAP, only U.S. citizens, U.S. nationals, or “qualified aliens” are eligible for SSI. Unauthorized individuals are not eligible for SSI. Upon meeting the above requirements, individuals must also meet one of three categorical criteria: age, disability, or blindness.¹⁴

As with SNAP, immigration enforcement may affect participation because the SSI application asks for the names, birthdates, and Social Security numbers of all persons living with the applicant. This portion of the SSI application is provided in appendix figure A2. Questions about household composition are taken into account to determine the contribution of the applicant to household expenses such as rent and utilities and can affect the amount of the SSI federal benefit.¹⁵ The SSA regularly subjects beneficiaries to redeterminations of income, resources, and living arrangements. Anecdotally, being asked to provide this type of information has invoked fear among certain communities given that federal benefits-granting agencies, in particular SSA for the administration of SSI, are required to report to DHS individuals who have undergone a formal determination on a claim who the agencies know are not lawfully present in the United States.¹⁶ However, according to the National Immigration Law Center, these agencies have clarified that this reporting requirement is only triggered for individuals seeking benefits and not relatives or household members.¹⁷

In the online appendix, we adapt Moffitt’s (1983) seminal model of nonparticipation in social programs and formalize how immigration enforcement can lead to spillover effects on Hispanic citizens. In contrast to much of the past literature, we focus on the role of indirect chilling effects on citizens who are not themselves at risk of deportation yet may alter their participation in programs for which they are eligible out of concern for noncitizen household members.

Agents maximize expected household utility placing weights on each set of family members—citizens and noncitizens. Instead of a stigma cost associated with participation, however, we include a deportation cost that is increasing in safety net participation and immigration enforcement. Under this model, the household decision-maker sets the marginal benefit of participation in federal programs, weighted by the welfare importance of citizen household members, equal to the marginal deportation cost induced by participation, weighted by the welfare importance of noncitizen members. Participation thus increases in the size of the program benefit and the weight given to eligible citizens in the household, but declines with immigration enforcement and the weight given to noncitizen household members.

When the household decision-maker is a U.S. citizen, the decision-maker forgoes a personal entitlement benefit to avoid the deportation cost incurred by close noncitizen contacts, even though the decision-maker is not directly affected by immigration enforcement. This is precisely the indirect spillover effect we aim to identify in this paper. By contrast, when the household decision-maker is an ineligible noncitizen, the decision-maker does not forgo any personal benefits. Rather, participation decisions are made on behalf of other household members, such as enrollment in Medicaid or subsidized school lunches for citizen children. In this scenario, a noncitizen decision-maker’s decision to enroll others directly increases the decision-maker’s own risk of deportation.

Our goal is to estimate the causal effect of immigration enforcement on participation in various public services by citizen Hispanic Americans. We now provide an overview of the data sources, before describing our identification strategy and addressing measurement concerns.

Through records available to the public, FOIA requests to ICE, and restricted-use data agreements, we obtained data on the roll-out of SC as well as micro-level data on the universe of detainers issued by ICE from 2002 to 2015 in the United States (US ICE, 2015; TRACFED, 2016). The detailed information includes the reason for the arrest as well as the crime level/severity, the date the detainer was issued, the county the detainer was issued in, the individual’s country of origin, and other individual-level demographics (age and sex). We also have the universe of individuals who were removed (actually deported) from the country due to a fingerprint match under SC from 2008 to 2015, in addition to county-level yearly data on the number of fingerprint submissions and matches under SC from 2008 to 2015.

SC led to a massive increase in immigration enforcement. Figure 1 presents a “first stage” event study plot for the number of detainers issued in a county in the years prior to and following the introduction of SC, controlling for county and state-by-year fixed effects. The figure shows a rapid ramp up in enforcement following SC’s launch in a county. The coefficients in the years before activation are close to zero and mostly insignificant. Effects during and after the year of activation are significantly positive and increasing over time. The figure also demonstrates that the overwhelming majority (91%) of detainers are issued against individuals from Hispanic countries of origin. While not all detainers are honored by local law enforcement agencies or lead to removal from the country, there is a strong positive correlation between detainers and removals under SC. See appendix figure A3.

Appendix table A1 presents difference-in-differences estimates of the impact of SC activation in a county on enforcement. Consistent with Cox and Miles (2014), we find that SC activation had no significant effect on offenses known to law enforcement. In contrast, we find significant increases in the number of fingerprint submissions received by ICE, fingerprint matches, and detainers issued post-SC activation. Additional event study estimates of the impact of SC on detainers issued at the monthly level are presented in appendix figure A4, which shows a sharp 15% increase in the number of detainers issued in the several months post-SC activation with no discernible trend pre-SC activation.

We use publicly available ACS data downloaded from IPUMS-USA at the University of Minnesota (Ruggles et al., 2019). We focus on the 1% ACS samples of the U.S. population over the years 2006–2016 for food stamp and SSI participation.¹⁸ The data include household characteristics such as food stamp receipt in the past 12 months as well as individual characteristics like SSI receipt in the past 12 months, education, nationality, and citizenship status. The most detailed level of geography in the publicly available ACS is the Census-defined Public Use Microdata Areas (PUMA), which contain at least 100 000 people and can cross county but not state lines. Because our activation dates and detainers data are at the county level, we distribute the ACS means at the PUMA level to counties based off the PUMA population in each county.¹⁹

We use household level data on food stamp and SSI participation within the past 12 months, ethnicity and county from the restricted-access Panel Study of Income Dynamics (PSID) from 2005–2015 (PSID 2017) to explore mechanism and conduct robustness checks. Details can be found in appendix B.

To measure awareness and perhaps deportation fear in response to SC, we use data from internet search patterns provided by Google Trends. Google Trends is a publicly available database that provides information on the relative popularity of search terms for 250 metropolitan areas across the United States at the Nielsen DMA media markets level. Details on the measure of popularity reported by Google Trands are discussed in Burchardi et al. (2018).

We use the following commonly searched terms related to the Deportation topic on Google Trends: deportation, deportacion, immigration, inmigracion, immigration lawyer, abogados de inmigracion, undocumented, indocumentado. Following the literature (e.g., Burchardi et al., 2018), we take a simple sum of search intensity across all search terms and normalize it by search terms that are popular in the Hispanic community, such as “deportes” (sports) and “telenovelas” (soap operas). This normalization accounts for differential access to the internet for Hispanics that may vary across geographic units.

Our triple-differences methodology exploits the staggered roll-out of SC activation across counties as well as the disproportionate impact of SC on Hispanics within counties. Specifically, we estimate the change in pre- versus post-SC activation differences in safety net participation by race/ethnicity in counties that have activated compared to counties that have not yet activated.

In the specification above, $IrH$ and $IrB$ are indicators for Hispanic ethnicity and non-Hispanic blacks, respectively. The omitted category is non-Hispanic whites. $Ictpost$ is an indicator equal to one in all county years after the activation of SC. Almost all counties activated between 2008 to 2013, with the majority of counties activating between 2010 to 2012. $Xrcst$ includes average log poverty rate, number of children, the share citizen, and the employment rate that vary across race, county, and time. We control for these characteristics as they are direct determinants of food stamp or SSI eligibility.

The Great Recession differentially impacted minority-headed households. For instance, white families’ wealth fell 26% during the Great Recession, while the wealth of black families and Hispanic families fell by 48% and 44%, respectively (McKernan et al., 2014). These differential wealth effects may have affected food stamp and SSI participation by race and ethnicity (Flores-Lagunes et al., 2018). Indeed, given this pattern, Hispanic Americans are more likely to be eligible for benefits compared to non-Hispanic whites, suggesting our estimates are likely to be a lower bound. We account for this possibility by explicitly including race/ethnicity-specific state-level employment changes during the Great Recession in $Xrcst$⁠.

We also account for county-level crime, $Xcst$⁠, defined as the log number of offenses known to law enforcement. While crime rates are not publicly available disaggregated by race at the county level, they have been shown to have differential effects on minority populations (Sampson & Lauritsen, 1997; Anwar & Fang, 2006; Antonovics & Knight, 2009). To allow for these differences, we interact race/ethnicity indicators with $Xcst$⁠.

State-by-year fixed effects (⁠$δst$⁠) are included in our preferred specification to capture important state-specific policies or economic shocks that might influence participation. These state-by-year fixed effects will incorporate policies such as the enactment of state omnibus immigration bills or mandated use of E-Verify to check the work authorization of new hires (Amuedo-Dorantes & Arenas-Arroyo, 2017). Such fixed effects also account for differential state-level effects of federal immigration reforms. We include state-by-race/ethnicity fixed effects (⁠$θrs$⁠) to control for attitudes and policies in each state that differentially affect minority groups. Race-by-year fixed effects (⁠$κrt$⁠) nonparametrically capture yearly shocks that differentially affect different racial groups, such as changes in economic conditions. Furthermore, we include county fixed effects (⁠$μc$⁠) and their interaction with an indicator for post-Morton Memo (⁠$ItMorton$⁠) to capture unobserved county-level factors that affect participation differentially before and after the 2011 Morton Memo which clarified that county participation in SC was not optional. Additional event studies and estimates from specifications that include county-by-year fixed effects and county-by-race fixed effects are presented in section VI.

For the ACS data, we weight all regressions by the total population in the relevant race-county cell.²⁰ To measure the spillover (indirect) effects of deportation fear we restrict our sample to citizen heads of households, individuals who are not eligible for deportation. We also restrict the sample to those with less than a high school degree as this represents a socioeconomic group that likely meets the income and asset requirements for safety net programs and is more likely to be connected to unauthorized individuals, the relevant risk set.

The coefficient of interest in equation (1) is $β2$⁠, which estimates the impact of SC activation on outcomes of Hispanic households relative to non-Hispanic white households, compared to counties that have not yet activated. $β3$ serves as a placebo test, capturing the effect of SC on black households relative to non-Hispanic white households in counties that have activated versus those that have not yet activated.

In this specification, $Ic,t=n$ is in indicator for each period (other than the year prior to activation $t=-1$⁠), such that the $β2n$ coefficients trace the participation in food stamps for Hispanics in the years before and after SC activation relative to non-Hispanic whites. Leads before five years and lags after four years are coded with the first and last groups, respectively, following McCrary (2007) among others. Similarly, each $β3n$ coefficient traces the participation in food stamps for blacks relative to non-Hispanic whites before and after activation. With this specification, one would expect to see a shift postactivation specifically for Hispanic households, and not black or white households, if we are measuring the causal effect of SC.

The identification assumption underlying our triple-differences approach is that there be no location-specific shocks timed with the staggered SC roll-out and differentially influencing the dynamic path of safety net outcomes for Hispanics. In other words, we assume that any differences in our outcome variables of interest for Hispanic versus white households would have evolved smoothly absent SC activation, conditional on our set of fixed effects and controls.

In addition to testing for pretrends in the event study analysis, we probe the identifying assumption in three ways. First, we test for pre-SC balance in our covariates and outcomes of interest between Hispanics and non-Hispanic whites across SC-activation groups in the spirit of our main specification in equation (1). Specifically, we regress the mean Hispanic-white difference for each dependent variable on fixed effects for each activation year group, state fixed effects, log crime, and the respective black-white difference. Standard errors are clustered at the county level. In table 1, we present F-statistics from a test of the joint significance of the activation year group fixed effects in column 1, with corresponding $p$-values from these tests presented in column 2. In terms of mean level differences, we find minimal evidence of significant differences across each activation group. Most importantly for our identification strategy, we find that there are no significant differences in changes in Hispanic-white food stamp or SSI participation in the pre-SC period across each activation group. These results support the assertion that the roll-out of SC was not correlated with pretrends in the differential participation in safety net programs across groups.²¹

Second, we test for balance by estimating our baseline specification in equation (1) where our dependent variables are county-level demographic controls. Results are presented in appendix table A3 and we can reject that SC activation affected demographic characteristics differentially for Hispanics relative to whites, with a joint F-test for all demographic controls yielding a p-value of 0.99. In appendix figures A5 and A6, we perform event studies with race-specific controls as dependent variables and show that these controls trend smoothly at the time of SC activation, and therefore fail to replicate the dynamics demonstrated for SSI and food stamps participation among Hispanic citizens. We also use all our baseline controls to predict food stamp and SSI participation and then estimate our main specification using predicted program participation as our dependent variable. Appendix figure A7 presents these results and shows no evidence of significant changes or trend breaks in predicted program participation for Hispanics post-SC activation.

Third, we implement a permutation test where we limit our data to preactivation years and randomly permute a pseudo-SC activation year for each county, ensuring that there is at least one year of data pre- and post-“pseudo” activation year. Using these randomly permuted activation years, we then estimate our baseline specification, equation (1), repeating this procedure 500 times. In appendix figure A8, we present the empirical distribution of these placebo effects for $β2$⁠, finding that our actual treatment effects for food stamps and SSI are larger (in absolute value) than 99% to 96% of our placebo estimates, respectively. These results suggest that SC activation had a large and atypical effect on outcomes for Hispanic households.

Despite the fact that our triple-differences design does not rely only on the simple pre- versus post-SC activation differences across counties, it is important to understand the factors that affected the timing of SC activation since nonrandom timing could still introduce bias. For instance, if SC preferentially activated in locations where criminal activity among the unauthorized was on the rise, and such activity decreases program participation,this could lead us to overestimate $β2$ in our main specification [equation (1)]. On the other hand, if locations that activated early were routine targets of immigration enforcement (such as locations close to the Mexican border), Hispanics in these areas may be relatively insensitive to changes in enforcement and thus exhibit small decreases in safety net participation, leading us to underestimate $β2$ in our main specification.

To further understand the timing of SC activation, figure 2 presents maps that show the timing of SC activation across counties, revealing that border counties were the earliest places to activate. These findings are consistent with Cox and Miles (2014), who find that SC activation was not related to crime—though the purported goal of the program was to remove criminal aliens—rather, earlier activation was positively correlated with proximity to the border, the presence of a 287(g) agreement, and the percent Hispanic population.

We take several steps to reduce selection bias that might be generated by the nonrandom timing of SC activation. First, we exclude border areas and three states that actively resisted SC implementation from our analysis since they might be unique in several ways related to both immigration enforcement and program participation. Second, in a robustness check described below, we include county interacted with time to account for features of a county that may affect timing of activation. Third, in another robustness check described below, we explicitly control for the percent of households that are Hispanic at the county-year level using data from the ACS. Fourth, we identified four criteria that likely affected roll-out timing: (1) estimated number of noncitizens, (2) the distance from the Mexican border, (3) crime rates, and (4) prior county relationships with ICE as proxied by the presence of a 287(g) agreement. We use these criteria with their high-level interactions to predict activation year. The corresponding F-statistic for the first-stage of this prediction is 385. Appendix figure A9 presents maps that show the timing of predicted SC activation across counties. In robustness checks, we explore the reduced form relationship between predicted activation and safety net participation, controlling for our preferred set of fixed effects and baseline controls. We note that variation in predicted activation year is driven by high-level interactions between the four criteria, generating plausibly exogenous timing of SC activation. We find nearly identical results when we use predicted activation compared to actual activation (see section VI).

A ubiquitous challenge for enforcement-related research is measurement error associated with missing or falsified responses. Indeed, federal employees with privileged access to linked data sources find that individuals classified as noncitizens in administrative records (AR) are less likely to respond to the citizenship question in the Census (Brown et al., 2018). These researchers also document that citizenship misrepresentation is concentrated among recently arrived foreign-born individuals. Thus, following the Brown et al. (2018) report, we conservatively define an individual as a citizen if he or she was born in the U.S. or naturalized and living in the U.S. for over a decade at the time of response. We verify the accuracy of this designation by comparing our naturalization counts to official figures published by the Office of Immigration Statistics, finding a correlation of 0.94 between the measures (see appendix figure A10). In unreported results, we also take advantage of the panel nature of the PSID and rule out that Hispanic household heads changed their responses to race and nationality questions in the aftermath of SC.

In addition to measurement error associated with reporting of citizenship, individuals may be fearful of reporting safety net participation. Meyer et al. (2020) emphasize that survey data often does not comport with administrative records on program participation, with false negatives and positives both possible, though the former is generally found to be much more important than the latter. Since our main specification relies on the difference between Hispanic and white-headed households, under-reporting will bias our estimates if it changes over time in relationship to SC for one particular group. A prominent explanation for false negatives among Hispanic households is fear related to immigration concerns (Brown, 2015).

In our context, fear could lead to measurement error in the outcome variable that is correlated with the introduction of SC.²² To the extent the implementation of SC heightens deportation fear, our estimates thus capture both the actual reduction in safety net participation as well as an increase in false negatives. Though both effects are of interest, they have very different policy implications.

To gauge whether under-reporting of participation responds to SC, we follow the literature and compare administrative SNAP data where available with survey estimates over time. We use administrative SNAP data from a group of states that provide participation information disaggregated by race and year (California, North Dakota, Oklahoma, Minnesota).²³ Similar to other scholars, we find that ACS estimates of food stamp participation are generally lower than available official yearly state-level estimates. Importantly, however, the ratio between survey and administrative participation does not change significantly post SC for Hispanics (column 1 of appendix table A4). We also find that the difference between the ratio of survey to administrative participation for Hispanics and the ratio for non-Hispanic whites does not change significantly post-SC (column 4 of appendix table A4). These findings suggest that SC had limited effects on reporting of program participation.

In addition to misrepresenting citizenship or program participation, individuals may fail to respond to the survey questionnaire altogether. Nonresponse on a required government survey, like providing false answers, is also subject to penalty by fines. To directly address the concern that SC may affect who is responding at all or to key questions in the ACS, we test and find no evidence that SC activation affected the observable characteristics of respondents in our ACS sample or the share of Hispanics who are U.S. citizens in the ACS (see section VII).

Our estimates may be subject to measurement error associated with apportioning ACS PUMA-level attributes to the county level. We insure this apportionment process does not bias our results by rerunning our analysis at the PUMA-level using the earliest SC activation date in the PUMA. The results are presented below in our robustness checks.

Figure 3 presents our main event study estimates of SC activation on food stamp participation (panel A) and SSI participation (panel B) for non-Hispanic whites, non-Hispanic blacks, and Hispanics using the ACS data, as described in equation (2). For both non-Hispanic whites and blacks, there is no noticeable break in the relative flatness of participation in either safety net program in the five years pre- and four years post-SC activation. In sharp contrast, coefficients on the interaction of time to SC and Hispanic are indistinguishable from zero in the years leading up to activation, but then demonstrate a level shift postactivation, with Hispanic heads greatly decreasing their participation in both food stamps and SSI over time. Specifically, by four years postactivation, Hispanic households reduce participation in food stamps by 5.7 percentage points relative to non-Hispanic whites, a 26% decrease from the preperiod Hispanic mean of 21.8%. Similarly, by four years postactivation, Hispanic households reduce their participation in SSI by 3.8 percentage points relative to non-Hispanic whites, a 72% decrease from the preperiod Hispanic mean of 5.3%.²⁴ Similar event studies comparing Hispanics to all non-Hispanics are shown in appendix figure A11, where both white and black individuals serve as the comparison group.

Table 2 presents our main results on safety net participation in the ACS data. Columns 1 and 2 present results for food stamp participation and columns 3 and 4 report results for SSI participation. In column 1, we find that after SC activation, Hispanic citizen heads of household reduce their participation in food stamps by 2.1 percentage points relative to non-Hispanics, a 10% decrease from the preperiod Hispanic mean. In column 2, we report the same specification as column 1 but add an interaction between our black indicator and post-SC indicator. Our main results are virtually unchanged and we also find a small and insignificant black coefficient post-SC, indicating that SC did not similarly affect the behavior of minority groups less likely to be affected by immigration enforcement. In column 3, we find that Hispanic citizen heads of household reduce their participation in SSI by 1.7 percentage points after SC activation relative to non-Hispanics, a 30% decrease from the preperiod Hispanic mean. Again, these results remain stable with the inclusion of an interaction between our black interacted with post-SC indicator, which is small and statistically insignificant (column 4).²⁵ Our main findings are also qualitatively similar using the PSID data, shown in appendix table A5.

Appendix table A6 presents several robustness checks of our main results for both food stamp participation (panel A) and SSI participation (panel B). Column 1 presents results that include the full sample of localities in the ACS, adding back in Massachusetts, New York, and Illinois (the states that resisted the activation of SC and attempted to opt-out) as well as counties with no known offenses in a given year. Column 2 presents results excluding controls for race-by-state changes in employment during the Great Recession. Column 3 controls for a full set of county-by-year fixed effects with corresponding event studies shown in appendix figure A13. Column 4 presents results using a more balanced panel of counties that activated between 2009 to 2012, dropping early activators in 2008 and late activators in 2013. Column 5 presents results using predicted activation year rather than actual activation year. Column 6 presents results comparing Hispanics to all non-Hispanics (white and black). Column 7 presents results controlling for preactivation trends in program participation following the approach in Freyaldenhoven et al. (2019). Column 8 presents results on a sample of citizen household heads with less than a college degree. Across all alternative specifications and samples, we continue to find economically meaningful and statistically significant reductions in safety net participation for Hispanic households relative to non-Hispanics after SC activation.

Appendix table A2 shows how our result under alternative specifications such as estimating our results in the ACS on a sample of counties that matches the PSID in terms of preperiod participation rates for Hispanics, using a sample of citizen female heads of household or female spouses, excluding citizen heads of households with mixed status family members, dropping cities with highest number of Hispanic immigrants, including spatial lag in SC activation that places lower weights on farther locations from enforced counties, estimating at the PUMA rather than county level, defining control counties as those that activate more than two years in the future to address the issue of negative weights in staggered differences-in-differences following Deshpande and Li (2019), and using a sample of citizen households where the head has some college or more. These robustness checks confirm that our main findings capture a true spillover effect of deportation fear, that Hispanic households are most responsive to enforcement within their own county, and that the effect of SC on program participation is most concentrated among our fragile connected sample.

Finally, appendix table A7 explores alternative weighting schemes and alternative controls. For example, our results are qualitatively similar when we do not use population weights, when including one observation for each family member in a citizen head household in the ACS, when controlling for the share of a county that is Hispanic and for the number of Hispanic noncitizens in each county, and when controlling for a full set of county-by-race fixed effects.

In this section, we explore potential mechanisms for our results. We begin by examining the role fear may have played before turning to other postulated mechanisms, including information and compositional changes. We also consider the role of changes in reporting behavior.

SC increased the number of detainers issued and forcible removals from the interior, which may have increased deportation fear. Indeed, Pew Research Center survey data demonstrate a positive correlation between respondents knowing someone who was detained and being fearful of the same fate befalling a family member or close contact (see panel A of figure 4). This relationship has also been described in anecdotal evidence with regards to SC activation, as detailed in the 2011 Task Force Review on Secure Communities (HSAC Task Force, 2011).

We conducted a series of analyses whose results, together, suggest that fear of deportation is a likely explanatory mechanism. First, we show that following SC activation, Google searches for deportation-related terms such as immigration lawyer and undocumented (in Spanish and English) increased sharply (see panel B of figure 4).²⁶ We find no discernible pretrend, but a sharp 25% increase in normalized deportation-related searches immediately following SC activation, consistent with at least an awareness of the SC program if not fear of its potential consequences. Second, locations where detainers are mostly issued for nonviolent (e.g., misdemeanor) crimes also respond more vigorously, suggesting the failure to target serious noncitizen offenders increases the prevalence of fear for Hispanic communities. In columns 1 and 5 of table 3, we show that in counties where the proportion of nonviolent detainers to all detainers is one, Hispanic households reduce their participation in food stamps and SSI by an additional 0.06 and 0.03 percentage points, respectively. Third, reductions in program participation are higher in areas with increasing deportation fear measured at the Census division level in the Pew survey data (the finest geography available in 2013). We find that a one standard deviation increase in fear is associated with an additional 1.7 percentage point decline in food stamp participation and a 0.8 percentage point decline in SSI participation among Hispanics after SC activation (columns 2 and 6). Fourth, in locations where federal detainers are not uniformly enforced by local authorities (i.e., “sanctuary cities”), we do not detect an effect in response to SC. For Hispanics in sanctuary cities relative to nonsanctuary cities, find a significant and positive effect of SC activation on participation in food stamps and no significant effect of SC activation on participation in SSI (columns 3 and 7). Appendix figure A14 presents corresponding event study estimates for sanctuary versus nonsanctuary jurisdictions. Similarly, in locations where Hispanics are dominated by Puerto Rican and Cuban enclaves, ethnic groups that face zero to minimal risk of deportation, we find more muted effects of SC on Hispanics. In counties with a 10% higher share of Puerto Ricans and Cubans experience significantly smaller reductions in both food stamp (0.3 percentage points) and SSI (0.4 percentage points) participation (columns 4 and 8). Lastly, we find that communities with a higher share of Hispanic noncitizens or mixed immigration-status households have an exaggerated response to SC. In appendix figure A15, we show how Hispanic households from counties with a 10% higher share of mixed-status households decrease participation in SNAP and SSI by an additional 0.1 and 0.3 percentage points after SC activation, respectively.

We next consider an alternative mechanism—the role of information. We partially test for the role of information in driving our results by comparing our estimated effects for Hispanic households that had never previously taken up the relevant public program prior to SC versus Hispanic households that previously took up the program following Aizer and Currie (2004). If a household has previously taken up the program, the household will likely already have information about the program, such as eligibility and how to apply. As a result, if information explains our findings, we would expect to find smaller effects of SC activation for prior use households. Columns 2 and 5 of appendix table A5 presents these results in the PSID sample where we limit the sample exclusively to all individuals in households that have taken up food stamps or SSI prior to SC activation, respectively (“prior users”). Among prior users, SC activation reduced Hispanic heads of household participation in food stamps by 49.6 percentage points and participation in SSI by 112.9 percentage points relative to non-Hispanics. We find that the decline in both food stamp and SSI participation post SC is largely driven by Hispanic heads that have previously taken up these safety net programs. In contrast, SC activation did not significantly affect participation in SNAP or SSI for Hispanic heads of household relative to non-Hispanics among individuals in households that had not participated in the program prior to SC activation (columns 3 and 6). The confidence intervals are wide given the sample size, but the results suggest that our main findings are not due to Hispanic households being less likely to receive information about public programs as their co-ethnics reduce sign up. This finding, along with qualitative evidence suggesting that Hispanic families are not renewing benefits, also lessens the likelihood that explanations like stigma are driving our results.

We also consider the possibility that SC activation may have affected the number or types of Hispanic citizens and households living within a particular county or within the United States, or more subtly, the number or types willing to declare their ethnicity or report program participation in surveys like the ACS. To test this channel, appendix table A3 presents our main specification in the ACS. We find no significant relationship between SC activation and compositional changes in the types of Hispanics relative to non-Hispanics in each county year in terms of number of children, poverty level, or employment rate (columns 1–3).²⁷ We also find no evidence of differential migration of Hispanics, measured by the share that report moving between states in the past year (column 4) or evidence that household weights from the ACS change differentially for Hispanics relative to other groups after SC activation (column 5). Finally, we find no statistically significant change in the percent of Hispanic mixed-status families or share citizen within a county post-SC activation (columns 6–7). Corresponding event study graphs of race-specific observable characteristics and predicted participation using all observables are presented in appendix figures A5, A6, and A7. While there are some smooth linear trends, these event study graphs show no evidence of significant changes or trend breaks for Hispanics post-SC activation, in sharp contrast to our main event studies in figure 3.

In this study, we test the hypothesis that linkages between citizens and noncitizens reduce safety net participation in the presence of enhanced immigration enforcement activity. Leveraging the roll-out of Secure Communities under the Obama administration, we find that citizen Hispanic Americans are indeed sensitive to such enforcement although they themselves are not at risk of removal—an indirect spillover effect. In particular, we find significant reductions in food stamp and SSI participation among Hispanic households. We find evidence that our results may be driven by deportation fear rather than lack of benefit information, measurement error, or stigma. Hispanic citizens residing in areas with a higher degree of connectedness with noncitizens, areas with a higher incidence of detainers issued for low-level arrests, and areas with greater increases in deportation fear exhibit larger decreases in participation in response to SC. In contrast, Hispanic households residing in sanctuary cities and areas with a higher share of Puerto Ricans and Cubans exhibited more muted responses to SC activation.

Our findings are particularly relevant given recent immigration policies. For one, the reactivation of SC in 2017 has substantially increased deportations relative to prior years. In contrast to the operation of SC under the Obama administration, a larger share of deportations under the Trump administration result from arrests for misdemeanor and petty offenses, which may enhance deportation fear. Second, recently proposed changes to the “public charge” determination, a designation that can prevent a noncitizen from adjusting their immigration status to legal permanent resident, could further intensify the spillover effects of fear. In 2018, DHS proposed to substantially expand the definition of a public charge to include any immigrant who “uses or receives one or more public benefits,” including both SSI and food stamps (previously exempt from public-charge determination). Moreover, the new proposal contemplates that use of these programs by U.S.-born citizen spouses and children could also count towards noncitizens’ use of public assistance, with some estimates suggesting that up to one third of U.S.-born citizens could have their use of public benefits considered in the public-charge determination of a family member (Perreira et al., 2018). Our results suggest that these policy changes could lead to further decreases in sign-up of safety net programs as Hispanic citizens may fear that their participation could jeopardize the chances that a family member obtains legal permanent residency.

Ultimately, our results have several implications on health and well-being for Hispanic households. The number of recipients and average benefit size suggests that as a result of SC, Hispanic households forgo over $212 million and $77 million in SNAP and SSI benefits per year, respectively. Extrapolating from the work of other scholars, families could experience adverse long-run consequences from forgoing benefits in response to stricter immigration enforcement. For example, Hoynes (2016) show that food stamp participation reduces the incidence of metabolic syndrome in adulthood, and Tiehen et al. (2012) find that food stamp participation reduced the child poverty rate by 5.6% from 2000 to 2009. Bronchetti et al. (2019) find that higher food stamp purchasing power increases the utilization of preventive medical care for children and reduces days of school missed due to illness. Similarly, Duggan and Kearney (2007) find that child participation in SSI is linked with long-term reductions in child poverty and Deshpande (2016) finds that removing disabled youth from SSI leads to a large decrease in observed lifetime income and exposes youth to greater income volatility. Schmidt et al. (2016) find that SSI program participation leads to a reduction in family food insecurity. These results suggest that reductions in food stamp and SSI usage among Hispanics in response to immigration enforcement could have long-run consequences for health and economic security. Most broadly, our results reveal how a government’s immigration policies may affect social policies aimed at citizens, yielding potentially unexpected results for vulnerable households.