Consequences of immigrating during a recession: Evidence from the US Refugee Resettlement program

My primary empirical strategy is based on the assumption that the month and year of arrival for refugees is plausibly exogenous. I used the monthly seasonally-adjusted civilian national unemployment rate each refugee faced at arrival to proxy for initial economic conditions. Since refugees cannot choose to selectively migrate to the US based on economic conditions, percentage point changes in the arrival national unemployment rate measure the changes in outcomes for refugees arriving under different economic conditions. A rich set of controls are also used to ensure demographic characteristics, duration in the US, and contemporaneous economic conditions do not drive my results.

The base specification is

y_it is either the employment status or log wages for each refugee i in survey year t. ue_i is the monthly seasonally adjusted national unemployment rate that corresponds to the date-of-arrival of each refugee. The arrival unemployment rate never varies for a refugee, so it is not possible to measure a scarring effect by comparing an individual refugee to himself across time. Therefore, I control for individual characteristics to create comparisons between individuals with similar characteristics. X_i contains a vector of controls which includes years of education prior to arrival, gender, age, English ability at arrival, disability status, marital status, and parental status. Disability status, marital status, and parental status are questioned in the context of the period being surveyed, so I used only the initial answer given when the refugee first appears in the dataset.

Given that refugees are not surveyed until at least 6 months after entry, these controls could still be endogenous. However, differences between columns 5 through 8 in Tables A1–A2 in Appendix provide evidence that these endogeneity concerns do not seem to drive results.

Calendar-month-of-arrival fixed effects, ϕ<!-- ϕ -->im, $\phi_{i}^{m},$ are used to control for seasonal variation in the monthly unemployment rate.

I do not control for date-of-arrival as the national unemployment rate does not vary within a particular arrival month and year. In Section 8, I alternatively use the arrival placement-state unemployment rate because this provides variation in treatment for a particular arrival date, allowing me to control for both arrival month and year. However, this state treatment specification is inferior because the geographical placement is somewhat endogenous, whereas the timing of resettlement is not.

Finally, the years-since-migration fixed effect variable, ϕ<!-- ϕ -->itk, $\phi_{i t}^{k},$ divides the number of days since each refugee arrived (calculated using the contemporaneous survey date and the documented arrival date) into intervals of 1–5. The earliest a refugee appears in the Annual Survey of Refugees data is 6 months post-arrival. Therefore, a value of 1 for k would represent a refugee who has been in the US between 6 months and 18 months. A value of 5 for k represents a refugee who has been in the US between 4.5 years and 2,175 days, the longest-tenured refugee in the sample. This control ensures that only refugees with the same number of years in the country are being compared with one another. My coefficient of interest, β, therefore measures the average effect of initial economic conditions on subsequent assimilation outcomes that is unexplained by post-arrival economic conditions for refugees of the same nationality, demographic characteristics, and years living in the US.

To measure how this effect might vary over time, I borrowed from Godøy (2017)

Godøy (2017) used immigrant employment rates instead of unemployment rates because Norway measures unemployment based on the number of registered jobseekers. Refugees in Norway have little incentive to register as jobseekers. This is not a concern in the US context because unemployment rates are derived from the randomized sampling of the entire population.

This specification is similar to the base specification, but the coefficient of interest, β_k, stratifies the average effect found in my base specification across years since migration. This specification also allows for full flexibility since I do not make any linearity assumptions regarding the interaction between years since migration and the initial unemployment rate.

In Figure 2, I provided evidence that total refugee immigration is not systematically related to national economic conditions. I used fiscal year refugee arrival totals found in Zong et al. (2017) for the period 1980–2015. These data cover the entire period of the refugee resettlement program. I compared these data with annual new immigrant arrival totals calculated using IPUMS American Community Survey data (Ruggles et al., 2017) for the period 1980–2015. I converted both sets of totals to logs to ease interpretation (immigrant totals are in millions while refugee totals are in tens of thousands) and plotted them across average national unemployment rates for the time periods for which the totals were reported. The graph shows that while total immigration falls as national economic conditions worsen, refugee immigration appears unaffected, or counter-cyclical. For better precision, I regressed both sets of totals on the arrival annual national unemployment rate. I found that total immigration decreases at a statistically significant rate of 9.85% for every one percentage point increase in the national unemployment rate. Total refugee migration, however, shows no statistically significant response to changes in the national unemployment rate.

Figure 2

As I am working with only a sample of refugees, I also need to assess whether the arrival national unemployment rate and arrival placement-state unemployment rate are not systematically related to any of my covariates. It is understood that country of origin will be systematically related to the timing of arrival for refugees because of both push and pull factors. Push factors, including the break out of conflict in a particular country at a particular time, partially determine the number of refugees who are applying to the UNHCR and US Refugee Resettlement program from that particular country. Pull factors, including differential arrival quotas of refugees by region,

https://www.state.gov/j/prm/releases/docsforcongress/261956.htm

In column 1 of Table 3, I tested whether any other covariates might be systematically related to the arrival national unemployment rate after controlling for country of origin. I used the following specification, uei=α<!-- α -->+δ<!-- δ -->Xi+ϕ<!-- ϕ -->ic+ε<!-- ε -->i. $u e_{i}=\alpha+\delta X_{i}+\phi_{i}^{c}+\varepsilon_{i}.$ This regression tests whether any of the covariates, Xi, $X_{i},$ are related to the arrival national unemployment rate, ue_i, after controlling for country of origin fixed effects, ϕ<!-- ϕ -->ic. $\phi_{i}^{c}.$ For comparison purposes, in column 2 of Table 3, I also tested whether any of the covariates, X_i, are related to an alternative treatment, the state unemployment rate at arrival, ueis. $u e_{i}^{s}.$ The specification for this column is ueis=α<!-- α -->+δ<!-- δ -->Xi+ϕ<!-- ϕ -->ic+ϕ<!-- ϕ -->i0+ε<!-- ε -->i. $u e_{i}^{s}=\alpha+\delta X_{i}+\phi_{i}^{c}+\phi_{i}^{0}+\varepsilon_{i}.$This specification controls for both country of origin fixed effects, ϕ<!-- ϕ -->ic, $\phi_{i}^{c},$ and date-of-arrival fixed effects, ϕ<!-- ϕ -->i0. $\phi_{i}^{0}.$Date-of-arrival fixed effects are used to demean state unemployment rates from national economic conditions so that I can test whether covariates are related to states with better or worse economic conditions.

Column 1 of Table 3 shows that the number of years of education has a slight positive relationship with the national unemployment rate. This could provide some indication that more educated refugees are arriving in the US in worse economic conditions. However, the coefficient is very small and only marginally significant. Given that my covariates do not appear to correlate in general with the national unemployment after controlling for country of origin, I am confident that there is a minimal compositional change within nationality groups across arrival years.

However, in column 2 of Table 3, I found that marriage is strongly correlated with the arrival placement-state unemployment rate. It means that states with worse economic conditions than the rest of the country receive more married individuals. This could potentially bias estimates using the placement-state unemployment rate downward as marriage is linked to better labor outcomes and those individuals are placed in states with worse initial economic conditions. Refugees are placed semi-randomly geographically if they do not have family already in the US. Unfortunately, the ASR data do not provide any information on refugees who are placed with family members. Therefore, family placement could be driving estimates using the placement-state unemployment rate treatment. For this reason, I rely solely on the national unemployment rate treatment to provide unbiased estimates of scarring for refugees.

Figure 3 provides a naive comparison of outcomes for refugees arriving under different economic conditions, which will guide the reader on my empirical results. I first divided my sample across the median arrival-national unemployment rate. I then plotted average outcomes across employment, hourly wages, and household utilization of social welfare benefits for the above and below-median groups over the 5 years sampling period. I found that refugees who arrive during an above-median arrival-national unemployment rate (bad economy) on average experience a persistent lower probability of employment, lower hourly wages conditional on employment, and an increased household usage of social welfare programs. The goal of my empirical strategy is to identify the portion of this effect that cannot be explained by demographics or subsequent economic conditions.

Figure 3

In columns 1 and 3 of Table 4, I tested whether initial economic conditions have a general effect on employment and log wages, respectively, after accounting for demographics, duration in the US, and subsequent economic conditions. The regression performed is outlined in Sections 5.2. Employment represents employment status at the time the refugee was surveyed and should be interpreted as percentage point changes in the probability of a refugee being employed. Log wages represent a log transformation of hourly wages of employed individuals

The log wage estimates are based only on those individuals who are employed at the time they are surveyed. This is a classic selection bias issue. To verify results, I estimate the effect of initial economic conditions on hourly wages (with those currently unemployed reporting zero dollars in wages) using a Poisson QMLE model and find results that have the same sign but are larger in magnitude, as expected.

In column 1 of Table 4, I observed that refugees after 5 years in the US, on average, experience a 1.57 percentage point decrease in the probability of current employment for every one percentage point increase in the arrival national unemployment rate. Considering that I control for the contemporaneous economic conditions and years since migration, these estimates represent the effect of labor market conditions at arrival that is unexplained by the persistence of economic conditions or experience. Standard errors are clustered at the date-of-arrival level and statistically significant at the 5% level. In column 3 of Table 4, I found that refugees experience a 1.98% decrease on average in wages for every one percentage point increase in the arrival national unemployment rate. Standard errors are also clustered at the date-of-arrival level and statistically significant at the 1% level.

To get a better understanding of how this effect might vary over time, as presented in columns 2 and 4 of Table 4, I analyzed the results found in columns 1 and 3 of Table 4, respectively, with the years since migration fixed effect. A value of “1 year, ue₀” represents the interaction between the arrival national unemployment rate and refugees who have been in the US between 6 months (the earliest a refugee appears in the data) and 18 months. A value of “5 years, ue₀” represents the interaction between the arrival national unemployment rate and refugees who have been in the US between 4.5 years and 2,175 days, the longest-tenured refugee in the sample.

Curiously, in column 2 of Table 4, I observed a positive relationship between employment probability and the arrival national unemployment for refugees who have been in the US between 6 months and 18 months (1 year). In Table 5, I split the sample by gender and found in column 6 that this initial increase in employment probability is owing to female refugees mostly. This might be related to the family income. In Table 6, I observed a negative relationship between welfare utilization and the arrival national unemployment rate during the first year, providing some evidence that income may be more constrained for refugees in their first year if initial economic conditions are unfavorable. Regardless, this increase in labor force attachment disappears by the second year and turns negative between the third year and fifth year, suggesting these were likely bad matches. In column 4 of Table 4, I observed a wage scarring effect that mostly persists at statistically significant levels for the entire 5 year period. In columns 4 and 8 of Table 5, I showed that this wage scarring effect is observed for both males and females at similar levels over the entire period.

In Table 6, I showed how the arrival national unemployment rate affects utilization of means-tested social welfare programs for refugees. Unlike most immigrants, refugees are an exempt group that is allowed to participate in means-tested social welfare programs during their first 5 years in the country.

https://aspe.hhs.gov/basic-report/overview-immigrants-eligibility-snap-tanf-medicaid-and-chip

In Table 6, row “1 year, ue₀”, I observed that refugees 1 year post-arrival show large decreases in the utilization of each program in response to worsening arrival economic conditions. It’s unclear why this is the case, but it could be related to pro-cyclical delays in scaling services, differential guidelines across states, unobserved income, fear of stigmatization, or other unobserved factors. Refugees might have more trouble in getting approved to receive AFDC/TANF

AFDC/TANF is a cash grant program for families with children, https://www.cbpp.org/research/policy-basics-an-introduction-to-tanf

The other possible explanation is that individual states have a fair amount of latitude in how these benefits are approved and dispersed. For programs like TANF, states set income and work requirements that might make it more difficult for refugees to get approved (LaPalo, 2019). If states react to deteriorating economic conditions by limiting access to these programs, refugees would have a harder time for getting approved. Another unobserved factor is outside income. In addition to VOLAGs, refugees also work with the local community- and religious-based organizations.

https://www.acf.hhs.gov/orr/state-programs-annual-overview

Finally, chilling, or the inhibition to exercise legitimate rights because of fear of stigmatization, might also be a contributing factor. In 1997, the Personal Responsibility and Work Opportunity Reconciliation Act (PRWORA) denied eligibility to most welfare programs for immigrants who had been in the country for <5 years. Despite refugees being exempt from this policy change, utilization of these programs by refugees dropped 37% after the law was passed (Fix and Passel, 1999).

Regardless, welfare utilization levels for refugees who arrived during bust periods are roughly the same as refugees who arrived during boom periods after the first year. There is also some evidence of an increase in the utilization of SNAP benefits after the first year, but a statistically significant effect is only observed in the fourth year post-arrival. On average for the entire 5 years period, I observed no statistically significant change in welfare utilization.

In addition to looking at the entire sample population, I also assessed whether scarring might differ across gender and origin-country educational attainment. As stated in Section 6.3, I showed in Table 5 that male and female refugees have different employment probabilities in the first year, but experience similar employment scarring effects in later years. Wage scarring persists for both male and female refugees throughout the entire 5 years period. In Tables 7 and 8, I further split the sample based on educational attainment. Educational attainment is classified as “No High School” for refugees with <12 years of education in their country of origin. I classified refugees who report between 12 years and 15 years of education in their country of origin as “High School.” Finally, I classified refugees who completed >16 years of education in their country of origin as “College.”

Tables 7 and 8 are divided into two parts. The first part shows the average effect of the arrival national unemployment rate, similar to columns 1 and 3 of Table 4. The second part shows the results of interaction between the arrival national unemployment rate and years since migration, similar to columns 2 and 4 of Table 4. Broadly, it appears that college-educated refugees experience poorer outcomes from entering the US during a recession than less-educated refugees. In column 4 of Table 7, I found that non-high-school-educated female refugees are the primary group driving the initial increase in employment probability. In columns 3 and 6 of Table 7, I found that college-educated male and female refugees are much less likely to enter the job market during the first year if arrival economic conditions are unfavorable. I also observed poorer employment probabilities for less-educated refugees in later periods.

In Table 8, I found strong evidence leading to the conclusion that college-educated male and female refugees experience poorer wage outcomes than their less-educated peers as a result of poor initial economic conditions. This is probably because college-educated refugees have a better chance of finding a job commiserate with their skill level if initial economic conditions are favorable. Non-high-school-educated male refugees have the best outcomes of any gender-education group, but all groups suffer some degree of persistent wage-scarring. Unfortunately, statistical power is not available to make a precise determination.

In Section 4, I provided an overview of the ASR data and described potential attrition issues with the panel data. Since my treatment variable never varies for the individual refugee, I am not comparing individual refugees to themselves over time. I am comparing individuals to similar individuals over time. Therefore, the principal concern with attrition is not the number of panels a particular person appears, but whether there are differences in the underlying composition of individuals over time, as measured by years-since-migration. Composition changes across years-since-migration can create a bias if the trajectory of these changes differs between those who entered the US when there were conditions of high and low unemployment prevailing, respectively. In Table 9, I provided descriptive summary statistics across years-since-migration between those entering the country during bust periods and boom periods, respectively. This table splits the sample by the median monthly national-unemployment-rate-at-arrival (the same methodology used to construct Figure 3). For the above-median group, or those who enter during a busting economy, I found some evidence of composition changes between 1 year and 5 years post-migration. Refugees who enter the US during bust periods and observed 5 years thereafter are more likely to be male, younger, and less likely to be educated, fluent in English at arrival, married at arrival, or have children when they arrive. Refugees who enter the US during boom periods (the below-median group) show fewer changes in education and gender, but I do observe several similarities with the above-median group in regards to the trajectory of these composition changes. Refugees who enter the US during boom periods and observed 5 years thereafter are also younger, less fluent in English at arrival, less likely to be married, and less likely to have children. However, refugees who are disabled at arrival are less likely to appear in later years if they migrate during boom periods.

To test formally how these composition changes might bias my estimates, I used a predicted outcomes test. I first regressed employment probability and log wages, y_it, on all the covariates, X_i: years of education, gender, age, English fluency at arrival, disability at arrival, marriage at arrival, and whether or not you are a parent at arrival. I then regressed the predicted outcomes from the first regression on my original specification, without covariates, to provide a means of comparison with my main results found in Table 4.

Step 1: yit=α<!-- α -->+γ<!-- γ -->Xi+ε<!-- ε -->it $y_{i t}=\alpha+\gamma X_{i}+\varepsilon_{i t}$

Step 2: y^<!-- ^ -->it=α<!-- α -->+β<!-- β -->uei+ϕ<!-- ϕ -->im+ϕ<!-- ϕ -->ic+ϕ<!-- ϕ -->t+δ<!-- δ -->ueits+ϕ<!-- ϕ -->itk+ε<!-- ε -->it $\hat{y}_{i t}=\alpha+\beta u e_{i}+\phi_{i}^{m}+\phi_{i}^{c}+\phi_{t}+\delta u e_{i t}^{s}+\phi_{i t}^{k}+\varepsilon_{i t}$

As explained in Section 5, ue_i is the monthly national-unemployment-rate each refugee faces at arrival. ϕ<!-- ϕ -->im,ϕ<!-- ϕ -->ic,ϕ<!-- ϕ -->t, and ϕ<!-- ϕ -->itk $\phi_{i}^{m}, \phi_{i}^{c}, \phi_{t}, \text { and } \phi_{i t}^{k}$ correspond to calendar-month-of-arrival, country-of-origin, contemporaneous year, and years-since-migration fixed effects, respectively. ueits $u e_{i t}^{s}$ is the contemporaneous state-of-placement unemployment rate. In addition, I have also regressed the predicted outcomes from the regression in Step 1 on my preferred specification outlined in Section 5.3, without covariates that include an interaction between the arrival national unemployment rate and years-since-migration.

Step 2a: y^<!-- ^ -->it=α<!-- α -->+β<!-- β -->kuei×<!-- × -->ϕ<!-- ϕ -->itk+ϕ<!-- ϕ -->im+ϕ<!-- ϕ -->ic+ϕ<!-- ϕ -->t+δ<!-- δ -->ueits+ϕ<!-- ϕ -->itk+ε<!-- ε -->it $\hat{y}_{i t}=\alpha+\beta_{k}\left(u e_{i} \times \phi_{i t}^{k}\right)+\phi_{i}^{m}+\phi_{i}^{c}+\phi_{t}+\delta u e_{i t}^{s}+\phi_{i t}^{k}+\varepsilon_{i t}$

The regressions in Steps 2 and 2a are measuring the portion of my estimated scarring effect that is predicted by changes in composition. If there are no differential changes in composition over time between cohorts which arrived during boom periods and bust periods, respectively, I should observe a zero effect. If I observe a non-zero effect, the sign and magnitude provide an estimate of how much of the observed scarring effect is driven or attenuated by composition changes.

Table 10 shows the results of this analysis and is analogous to Table 4. There is some evidence that composition changes do affect my employment estimates in periods 1, 4, and 5, and my wage estimates in period 5. However, when I compared the estimates found in Table 10 with my main estimates in Table 4, I saw that all the signs of the significant coefficients in Table 10 are opposite to the coefficients laid down in Table 4. This means that composition changes are actually attenuating my results, not driving them. For example, in Table 4, I estimated that refugees entering the US face a 2.08 percentage point reduction in their employment probability for every percentage point increase in the arrival unemployment rate after 5 years. However, the estimate in column 2 of Table 10, row “5 years, ue₀” is positive. This suggests that composition changes are responsible for a 1.55 percentage point increase in this estimate. Therefore the true effect for employment scarring in the fifth year might be closer to a drop of 3.64 percentage points in magnitude. Conversely, the coefficient in row “1 year, ue₀” in Table 10 is negative, while the coefficient in my main results table, Table 4, is a positive estimate of 1.68 percentage points. This suggests that the increase in employment probability in the first period is likely closer to 2.6 percentage points. Overall, the composite effect of 0.52 percentage points found in column 1 of Table 10 suggests that the composite estimate for employment scarring in column 1 of Table 4 is likely closer to a drop of 2.09 percentage points.

Post-arrival interstate mobility is another important outcome that could be affected by initial economic conditions. A refugee placed in a state with poorer economic conditions than neighboring states could move and potentially experience better outcomes. Wozniak (2010) has made the observation that economic improvement in states can drive relocation for highly educated workers. Unfortunately, the Annual Survey of Refugees data do not offer a credible way to test actual mobility. There is no information on a refugee’s current state of residence. There is a question about whether a refugee lived in the same state in the previous year, but a large portion (>40%) of the observations is missing. However, I can use the remaining sample of observed non-movers to gain a better understanding of how post-arrival mobility might affect my estimates. In Table 11, I showed the results of regression on a sub-sample of known non-movers using the national-unemployment-rate-at-arrival treatment. The estimates in Table 11 are larger in magnitude than the main estimates in Table 4, suggesting that post-arrival movement is likely attenuating national unemployment rate estimates. In Section 7.1, I have shown that composition changes are also likely attenuating my estimates. The post-arrival movement will be a driver of changes in composition over time if it is assumed that refugees who move are also less likely to participate in future surveys.

As a robustness check for my preferred estimates in Table 4, I have also included estimates from alternate specifications in Tables A1–A2 in Appendix. Column 1 of Tables A1 and A2 in Appendix shows results without any of the following covariates: years of education at arrival, gender, age, English fluency at arrival, marital status at arrival, disability status at arrival, and parental status at arrival. Columns 2–8 of Tables A1 and A2 in Appendix show how these estimates change as each covariate is added, with column 8 being the preferred specification.

In Table 12, I showed the results of the arrival placement-state unemployment rate treatment on employment probability and log wages. The estimates suggest that refugees experience a slight increase in employment probability in their fifth year, while wage scarring decreases each year. However, I do not have the means to differentiate between refugees who are, respectively, placed with family and placed randomly geographically. Therefore, it is not possible to determine whether these estimates reflect a true decrease in wage scarring or if they are the result of non-random placement in areas with better economic conditions.

Further, to overcome this selection bias problem in my arrival state-placement estimates, I limited my analysis to refugees who are less likely to have family already living in the US. If a refugee is one of the first to be resettled from their home country, it is less likely they have family already living here. To achieve this, I create two different groups of pioneers. The first group, nationality-by-state pioneers, represents refugees who are resettled in a particular state within 2 years of the nationality’s first appearance in that state. I use both the Annual Survey of Refugees data and previous ORR Annual Reports

https://www.acf.hhs.gov/orr/resource/annual-orr-reports-to-congress

In Tables 13 and 14, I showed the estimates of the effect of the arrival placement-state unemployment rate on these two groups of pioneers. In columns 1 and 2 of Table 13, a statistically significant wage scarring effect is observed in the first year for nationality-by-state pioneers. However, the magnitudes of the estimates in columns 1 and 2 of Table 13 are similar to the results found in Table 4 using the arrival national unemployment rate treatment. In Table 14, I gain more precision and find that estimates are larger in magnitude than Table 4, but also follow a similar pattern. This provides evidence that my original arrival placement-state unemployment rate estimates are likely biased toward positive outcomes by an unknown number of sample respondents being placed near family.