2. DACA program

DACA was introduced by President Obama on June 15, 2012, as a substitute for Dream Act legislation. DACA provides a solution to the long-term residence of millions of undocumented immigrants who had been brought to the US by their parents as a child. It allows recipients to remain in the country with temporary lawful status. DACA recipients may apply for work authorization to legally work in the US. In many states, DACA status additionally enables undocumented immigrants to acquire occupational licenses (Liang, Reference Liang2023). However, DACA does not provide a path to permanent residency, therefore, DACA recipients have to renew their status every 2 years.

To be eligible for DACA, an individual has to qualify for all of the following requirements: (a) they must be undocumented as of June 15, 2012; (b) they entered the US before their 16th birthday; (c) they must be under 31 as of June 15, 2012; (d) they must have constantly resided in the US since June 15, 2007; (e) they must be either enrolled in school, have obtained a high-school diploma, general education development, or be an honorably discharged veteran of the Coast Guard or Armed Forces of the United States; f) they must have no record of a felony or have significant misdemeanors.

Nonetheless, the precise estimate on the number of DACA-eligible population is challenging due to the shortage of administrative data. According to the Migration Policy Institute, there are over 1.3 million DACA-eligible individuals. This estimate does not account for some criteria that are unavailable to researchers, which are criminal records and continuous presence in the US. So, this estimate is on the high end of the range of DACA-eligible population.Footnote ⁸ There were over 800,000 immigrants who had ever been DACA holders, which made up around 60% of the total DACA-eligible population. Of those who did not apply for DACA, 43% of them claimed that they couldn’t afford the application fee, while 22% were missing the required paperwork and 17% were afraid that the DACA application process would expose them to authorities (Watson & Thompson, Reference Watson and Thompson2022). As of March 2024, around 528,000 individuals had active DACA status because a proportion of DACA holders either failed to renew their status or adjusted to long-term legal status.Footnote ⁹

During the 2016 presidential election, DACA was one of the most controversial topics between the candidates. After President Trump announced that his administration would end DACA on September 5, 2017, the Department of Homeland Security began phasing out the program, initially halting new applications and ending renewals for permits that expired after March 5, 2018. This decision faced several legal challenges during the next 2 years; however, it significantly impacted the number of new DACA applicants. On June 18, 2020, the Supreme Court ruled 5-4 that the Trump administration’s attempt to end DACA was unlawful. However, it was not until December 4, 2020, that a federal judge ordered the Department of Homeland Security to fully reinstate DACA, including accepting new applications and granting work permits. Figure 1 shows the total number of DACA recipients as well as the number of initial and renewal recipients from 2012 to 2020. The number of initial DACA recipients peaked in 2013 and began to decline in 2014, reaching almost 0 in 2019 and 2020 as a results of legal battles during the Trump administration. These legal challenges affected not only the number of new DACA applications but also had negative impacts on DACA recipients. For example, self-reported health among DACA recipients and their children worsened (Patler et al., Reference Patler, Hamilton, Meagher and Savinar2019), labor market outcomes also negatively impacted (Zaiour, Reference Zaiour2023). Upon assuming office, the Biden administration has made several attempts to codify DACA into federal regulation. However, these efforts have encountered repeated challenges. Following a court ruling that deemed their actions unlawful, the administration issued a final rule on DACA, only to face another legal challenge in June 2023.

Figure 1:

The number of cumulative, initial, and renewal DACA recipients.

Source: US Citizenship and Immigration Services.

DACA recipients reside in all 50 US states and the District of Columbia. Nonetheless, nearly half of them live in California and Texas. California alone made up for almost 29% of nationwide DACA recipients, while 17% of them name Texas as their home state. Figure 2 illustrates the map of DACA recipient distribution by state as of March, 2021.

Figure 2:

The number of DACA recipients by state.

Source: US Citizenship and Immigration Services.

3. Data and descriptive statistics

3.4. Descriptive statistics

Table 1 displays the summary statistics for people who are non-citizen immigrants under 16 years old upon arrival in the US, entered the US before 2007, and have obtained a high-school diploma. I report the summary in two groups, one is DACA eligibles if individuals are under 31 years old in 2012 and ineligibles otherwise. Panel A represents people who are potentially eligible for DACA, they tend to be younger (28.97 versus 44.35 years of age); have lived in the US for a shorter time (19.94 versus 35.26 years); are less likely to be self-employed (0.07 versus 0.13) and have lower wage income (US$31,200 versus US$42,600) than people who are potentially ineligible for DACA. Panel B shows that in general, people who are potentially eligible for DACA, work in jobs that require lower job skills than people who are not.Footnote ¹²

Table 1.

Summary statistics

Notes: Sample in Panel A includes who arrived in the US before their 16 years old, before 2007, and have obtained high-school diploma. Sample in Panel B is conditional on being employed.

4. Econometric strategies

To identify the effects of DACA as a quasi-experiment on labor market outcomes, this paper exploits a parametric RDD. Two options that are potentially used as running variables: individuals’ age in 2012 and individuals’ age at arrival. Nonetheless, age at arrival is correlated with education (Evans & Fitzgerald, Reference Evans and Fitzgerald2017; Gonzalez, Reference Gonzalez2003) and English proficiency (Bleakley & Chin, Reference Bleakley and Chin2010). Therefore, it is correlated with labor market outcomes. Moreover, the RDD model based on age at arrival with a 16-year-old threshold faces limited support from the right side of the threshold due to another discontinuity at the age of 18. Specifically, individuals who immigrate at the age of 18 are more likely to have already completed high school and decided to move to the U.S. independently, whereas those who immigrate at younger ages may not have finished high school and typically relocate with their families. Consequently, in addition to the 16-year-old threshold, there exists a discontinuity at the age of 18, making it difficult to measure the limit from the right side of the 16-year-old threshold and estimate the RDD model. Thus, I leverage individuals’ age in 2012 as my primary running variable.

As explained in Section 3.1, I restrict my sample to only non-citizen immigrants who meet three out of four observable DACA criteria and then define a treatment group and a control group based on my running variable. Specifically, individuals who are under 31 as of June 15, 2012, are eligible and identified as a treatment group. On the other hand, individuals who are 31 or older are ineligible and classified as a control group in my setting. To simplify my notation and computation, I normalize individuals’ age in 2012. Let ${R_{it}}$ = individual’s age in 2012–31. Then,

$${D_{it}} = \left( {\matrix{ 0 \hfill & {{\rm{if\;}}{R_{it}} \ge 0} \hfill \cr 1 \hfill & {{\rm{if}}\;{R_{it}} \lt 0} \hfill \cr } } \right.$$

is defined as a binary treatment variable.

The main empirical specification has the following form:

(1) $${Y_{it}} = \alpha + \beta {{\rm{D}}_{it}} + \mathop \sum \limits_1^n {\gamma _n}{\rm{R}}_{it}^n + \mathop \sum \limits_1^n {\delta _n}{\rm{R}}_{it}^n{\rm{*}}{{\rm{D}}_{it}} + \lambda {{\rm{X}}_{it}} + {\varepsilon _{it}}$$

in which ${Y_{it}}$ refers to the outcome variables of interest of individual i at time t. In this parametric regression discontinuity, $n$ indicates the order of the polynomial function, where $n = 1,2,3$ are linear, quadratic, and cubic functions respectively. The coefficient of interest $\beta $ measures my RDD intention-to-treat effects.

This model also includes a vector of control variables ${{\rm{X}}_{it}}$ , which controls for sex, years of education, and number of years in the US.Footnote ¹³

In this paper, the running variable is discrete, with 42 mass points (i.e., unique values) and 64,064 observations. To use a continuity-based RDD approach with discrete running variables, one criterion is that the number of mass points must be large (Cattaneo & Titiunik, Reference Cattaneo and Titiunik2022). However, in this setup, due to the very limited number of mass points, the continuity-based RDD approach is not advisable.Footnote ¹⁴ Instead, I use a parametric RDD approach. In this approach, I have to select the functional form and the bandwidth.

Functional form. In this approach, I have to rely more on choosing a functional form to correctly identify the effect of treatment on outcome variables (Lee & Card, Reference Lee and Card2008). That being said, the uncertainty in the selection of functional form would produce specification errors. In other words, the low-order polynomials are going to introduce some biases unless I use an extremely high-order polynomials. However, if I keep increasing polynomial order, estimates will rely heavily on observations far away from the threshold. To minimize the possible bias arising from specification errors, I instead use the conditional expectation function (CEF) among natives as an approximation for the CEF among immigrants, then add additional polynomial adjustment to account for any remaining differences in the CEF between natives and immigrants. This approach is based on two fundamental assumptions. First, for natives, I have a larger sample size and can therefore estimate the CEF precisely. Then, using the estimates from those native sample to eliminate any noises can reduce the bias introduced from mis-specification. Second, this correction would only introduce unknown bias if there were a discontinuity in the labor market outcomes of natives that is unrelated to the potential counterfactual discontinuity for immigrants. In other words, we would need to assume that a discontinuity exists in the labor market outcomes of natives at the age of 31 in 2012. However, there is no empirical or theoretical evidence to support such a discontinuity. To examine these assumptions, I visualize the means of the outcome variables among a sample of natives, which show that the linear functional form fits the data on natives well and there is no discontinuity at the age of 31 among natives. I present these plots in Online Appendix K. Specifically, I follow a 2-step method as described below:

• Step 1: I regress all outcome variables on dummy variables of individuals’ age in 2012 for natives only.

  (2) $${Y_i} = \kappa + \mathop \sum \limits_{m = - 14}^{m = 28} {\nu _m}{\rm{*}}1\left( {{{\rm{R}}_i} = m} \right) + {\tau _i}$$

  in which, ${Y_i}$ is the original outcome for native individual i and m is individuals’ normalized age in 2012.

• Step 2: Then, I use the estimate of ${Y_i}$ ( ${\hat Y_i}$ ) in Equation (2) to adjust for my original outcome variables as follows:

  $${\tilde Y_{it}} \equiv {Y_{it}} - {\hat Y_i}.$$

Specifically, my three models using the parametric regression discontinuity approach are:Footnote ¹⁵

1. 1. ${\tilde Y_{it}}$ = ${\alpha _0}$ + $\beta {{\rm{D}}_{it}}$ + ${\gamma _1}{{\rm{R}}_{it}}$ + ${\delta _1}{{\rm{R}}_{it}}{{\rm{D}}_{it}}$ + $\lambda {{\rm{X}}_{it}}$ + ${\varepsilon _{it}}$ (1a)

2. 2. ${\tilde Y_{it}}$ = ${\alpha _0}$ + $\beta {{\rm{D}}_{it}}$ + ${\gamma _1}{{\rm{R}}_{it}}$ + ${\gamma _2}{\rm{R}}_{it}^2$ + ${\delta _1}{{\rm{R}}_{it}}{{\rm{D}}_{it}}$ + ${\delta _2}{\rm{R}}_{it}^2{{\rm{D}}_{it}}$ + $\lambda {{\rm{X}}_{it}}$ + ${\varepsilon _{it}}$ (1b)

3. 3. ${\tilde Y_{it}}$ = ${\alpha _0}$ + $\beta {{\rm{D}}_{it}}$ + ${\gamma _1}{{\rm{R}}_{it}}$ + ${\gamma _2}{\rm{*R}}_{it}^2$ + ${\gamma _3}{\rm{R}}_{it}^3$ + ${\delta _1}{{\rm{R}}_{it}}{{\rm{D}}_{it}}$ + ${\delta _2}{\rm{R}}_{it}^2{{\rm{D}}_{it}}$ + ${\delta _3}{\rm{R}}_{it}^3{{\rm{D}}_{it}}$ + $\lambda {{\rm{X}}_{it}}$ + ${\varepsilon _{it}}$ (1c)

Bandwidth. One option is to choose the bandwidth based on the researcher’s judgment. The downside of this approach is its subjectivity. The other option is using the data-driven method. In this model setting, where the running variable is discrete with few mass points, it is not ideal to use the bandwidth selection methods proposed by Calonico et al. (Reference Borjas2014, Reference Borjas and Cassidy2018, Reference Calonico, Cattaneo and Farrell2019, Reference Calonico, Cattaneo and Farrell2020). Thus, I handpick a bandwidth of 6 to start. To ensure robustness and eliminate subjective judgment, I also run the model with bandwidths ranging 3 to 7.Footnote ¹⁶ As I show in specification curves in Section 6.1 (Simonsohn et al., Reference Simonsohn, Simmons and Nelson2020), this does not matter which bandwidths I am using because the estimates are all small and insignificant.

Next, the main concern for RDD is the possibility of data manipulation and discontinuity in unobservables around the threshold. In other words, the results will be misleading if people who are close to the threshold, might attempt to manipulate it and sort them into their preferred group. To address that, I perform the density test based on the non-parametric local polynomial density estimator developed by McCrary (Reference McCrary2008). The test statistic is -0.004 with s.e 0.009, which fails to reject the null hypothesis of continuity. I plot the density of the running variable in Figure 3, following McCrary (Reference McCrary2008), which visually confirms the smoothness of the density function of my running variable.

Figure 3:

McCrary (Reference McCrary2008) test.

Notes: This figure shows the formal manipulation test based on a methodology proposed by McCrary (Reference McCrary2008). This supports the reliability of the RDD method that observations near the threshold are comparable and free from manipulation.

I also demonstrate in Figure 4 the graphical version of balance tests, plotting the means of variables in different brackets of age in 2012. Figure 4(a) shows the probability of qualifying for the other three DACA requirements, which are under the age of 16 at arrivals, arrivals before 2007, and high-school diploma or equivalent holder. It is evident that there is no bunching around the threshold. Similarly, Figure 4(b), (c), and (d) illustrate that all plots have smooth transitions at the threshold. Thus, individuals who are adjacent to the threshold are comparable.Footnote ¹⁷

Figure 4:

Balance check of covariates.

Notes: This figure shows the means of four variables to verify the continuities of those variables across the threshold. Data spans from 2013 to 2019.

6. Robustness checks

6.1. Specification curves

In my preferred specification, I utilize a sample comprising non-citizens, employing the linear functional form with a bandwidth of 6. Nonetheless, to enhance the robustness of my analysis, I explore various specifications to evaluate my RDD estimates. I present all estimates in a form of specification curves (Simonsohn et al., Reference Simonsohn, Simmons and Nelson2020).

• • Functional forms: In my specification curves, I present the estimates for the linear, quadratic, and cubic functional forms.

• • Bandwidths: In addition to a bandwidth of 6, I also use a range of bandwidths from 2 to 7.

• • Methods to impute legal status: To focus on potential DACA recipients, I expand the analysis to include Mexican non-citizen immigrants residing in the entire US and those in California and Texas. See Online Appendix B for details.”

• • Quasi-experimental method: In addition to my primary econometric model, I implement a difference-in-discontinuities design following Bae (Reference Bae2020). See Online Appendix B for details.

• • Natives’ CEF: In addition to my primary econometric model, where I use natives to adjust for my immigrant sample, I also perform my main specification without adjusting for the natives’ CEF.

• • Sample: In Figure 8, I also present the shorter sample of 2013 to 2014 as I describe in Section 5.3. For the specification curves for occupational skill usage (Figure 9), I also use a sample excluding the lowest 2% of observations by job skills to estimate the maximum possible effects if all DACA individuals entering employment are low-skilled. See sample selection in Online Appendix B for details.

Figure 8:

Specification curves on employment outcomes.

Notes: This figure presents the RDD estimates along with 95% confidence intervals for 97 different specification choices (functional forms, bandwidths, samples, and econometric models. The lower panel shows the choices made in each specification (e.g., if the dot in the linear is bold, that specific specification uses the linear functional form). The upper panel shows the RDD estimates and confidence intervals. In sub-figures d) and e), I remove several specification curves with extreme value to enhance the visibility. In sub-figure e), weekly working hours are scaled by a factor of 10. In sub-figure f), wage income is scaled by a factor of 10,000.

Figure 9:

Specification curves on occupational skill usage.

Notes: This figure presents the RDD estimates along with 95% confidence intervals for 97 different specification choices (functional forms, bandwidths, samples, and econometric models. The lower panel shows the choices made in each specification (e.g., if the dot in the linear is bold, that specific specification uses the linear functional form). The upper panel shows the RDD estimates and confidence intervals.

The results from the specification curves show three key points. First, DACA eligibility has minimal and statistically insignificant effects on most labor market outcomes. Second, the few extreme values are associated with the cubic functional form, which are not reliable as discussed in Section 5. Third, in Figure 9, after removing the bottom 2% from each job skill distribution, the effects on occupational skill usage tend to be positive but remain small and statistically insignificant. In general, my main results are robust to a variety of specification choices, and do not heavily depend on bandwidths, methods of imputing legal status, or the choice of econometric models.

6.2. Placebo tests

If the only reason for DACA eligibility affecting labor market outcomes is a temporary legal status, then those effects should be null in samples where DACA eligibility is not relevant. In order to confirm that, I run the main specification on naturalized citizens.Footnote ²⁴

However, one of the concerns on naturalized citizens is that individuals who had been DACA recipients and then were naturalized later on, which would contaminate my estimates. To deal with that issue, I restrict my immigrant citizens to only individuals who were naturalized before 2012.Footnote ²⁵ Table 6 presents the results of DACA eligibility on employment outcomes. It is evident that most of coefficients are negative, but nearly 0. Table 7 presents the results of DACA eligibility on the probability of working in high-skilled jobs among immigrant citizens. Like what I find in Table 6, most of coefficients are statistically insignificant and trivial. In general, I find no evidence that DACA eligibility has impacts on employment and working in high-skilled jobs among naturalized citizens upon the launch of DACA.

Table 6:

DACA eligibility and employment outcomes: Naturalized citizens

Confidence intervals are presented in parentheses and calculated with 95% confidence. Standard errors in parentheses are clustered at the state-year level.

Notes: This table shows the placebo tests effects of DACA eligibility on labor market outcomes among naturalized citizens, employing the linear functional form with a bandwidth of 6. Sample includes those who have obtained high-school diploma, have entered the US before their 16th birthday and have immigrated to the US before 2007.

* $p \lt .10$ , ** $p \lt .05$ , *** $p \lt .01$ .

Table 7:

DACA eligibility and occupational skill usage: Naturalized citizens

Confidence intervals are presented in parentheses and calculated with 95% confidence. Standard errors in parentheses are clustered at the state-year level. Coefficients are measured in standard deviation.

Notes. This table shows the placebo tests effects of DACA eligibility on job-related skill indices among naturalized citizens, employing the linear functional form with a bandwidth of 6. Sample includes those who have obtained high-school diploma, have entered the US before their 16th birthday and have immigrated to the US before 2007.

* $p \lt .10$ , ** $p \lt .05$ , *** $p \lt .01$ .

In general, I find no evidence that DACA eligibility has the effects on labor market outcomes among people who are not justified by the policy.