Immigration courts and removal proceedings

Immigration courts are not part of the judicial branch but rather, part of the executive branch. IJs are attorneys whom the Attorney General appoints as his or her “delegates” (8 C.F.R. § 1003.10[a]) within an agency called the Executive Office for Immigration Review (EOIR). The EOIR is a component of the US Department of Justice (8 USC § 1101[b][4]), the executive branch that prosecutes immigrants in federal court. Unlike administrative law judges hired by some of the other federal agencies, IJs are subject to performance reviews and lack special protection from removal from the bench afforded by the Administrative Procedures Act (see Reference JainJain, 2019; Reference TaylorTaylor, 2007). This means that IJs can be removed from the bench or reassigned to another position by the Attorney General. Thus, IJs may be “particularly vulnerable to political pressure and interference” (ABA et al., 2019). Over the years, these concerns have led to repeated—and thus far unsuccessful—calls by the American Bar Association, the National Association of Immigration Judges, and other prominent legal organizations, for the establishment of an independent immigration court system under Article I of the US Constitution (Reference Gilbert and CohenGilbert & Cohen, 2017; Reference Slavin and MarksSlavin & Marks, 2011).

Removal proceedings constitute the most common type of proceedings in immigration courts (for other types of proceedings that IJs conduct, see ABA, 2019). At the start of removal proceedings, the IJ must decide whether to sustain the government's charges. If the charging document known as the Notice to Appear does not state a valid ground for removal, the case will be “terminated.” If the charges are sustained, a number of outcomes are possible in the subsequent stage of the removal proceeding. For example, the noncitizen may request (at the beginning or the end of the proceeding), and the IJ may grant, a “voluntary departure,” which means that the noncitizen takes on the financial burden of returning to his or her country of origin and avoids triggering some of the legal bars to future lawful admission associated with removal orders. If the noncitizen does not file an application for relief from removal (e.g., asylum, cancelation of removal), the IJ will order them removed.Footnote ³ If the noncitizen pursues legal relief from removal and is granted such relief, then they will be allowed to remain in the United States.

As noted earlier, government trial attorneys represent the Department of Homeland Security (DHS) in removal proceedings.Footnote ⁴ In contrast, noncitizens in removal proceedings may hire an attorney at their own expense but they lack the right to a court-appointed defense counsel (8 U.S.C. § 1229a[b][4][A]). This means that a noncitizen seeking legal representation typically must find either pro bono representation or pay to retain a private attorney. Non-attorneys working for nonprofit organizations who obtain accreditation through EOIR's Recognition and Accreditation Program may also represent noncitizens in removal proceedings (for a list of accredited representatives by state, see EOIR, 2020).

Access to counsel for noncitizens

According to the Transactional Records Access Clearinghouse (TRAC), only about 44% of cases filed between fiscal years 2001 and 2020 involving noncitizens who were never detained or released from detention had legal representation (TRAC, 2020c). Among detained noncitizens in removal proceedings, the representation rate was even lower: about 17%. These low rates of representation reflect the reality that many noncitizens lack financial means to hire a private attorney. Some noncitizens may attempt to save up funds over time to hire a private attorney, but judges may deny their request for continuances, effectively foreclosing this option (Reference Hausman and SrikantiahHausman & Srikantiah, 2016). In addition, studies have documented a shortage of qualified immigration lawyers who can provide deportation defense (Reference KatzmannKatzmann, 2014). Unsurprisingly, these challenges are magnified for noncitizens who are held in immigration detention, as prolonged detention invariably leads to a job loss and social isolation from the outside world (Reference RyoRyo, 2017).

These access-to-counsel issues have drawn a great deal of public attention because legal representation—particularly representation of high quality—is associated with greater engagement with the legal process for noncitizens and higher chances of successfully fighting removal (Reference Eagly and ShaferEagly & Shafer, 2015; GAO, 2016; Reference HausmanHausman, 2016; Reference Markowitz, Annobil, Caplow, Cobb, Morawetz, Root, Slovinsky, Cheng and NashMarkowitz et al., 2011; Reference Miller, Keith and HolmesMiller et al., 2015a; Reference RyoRyo, 2016, Reference Ryo2018). But largely overlooked in public discourse and scholarly research on access to counsel for immigrants are important questions about whether and to what extent the impact of legal representation might be contingent.

Research on the role of lawyers

Given that lawyers play such a prominent role in the American legal system, the question of whether lawyers matter for case outcomes has been the focus of a large body of scholarship on the legal profession and access to counsel (for a review, see Reference Taylor Poppe and RachlinskiTaylor Poppe & Rachlinski, 2016). In addition, scholars have also examined how lawyers might matter for case outcomes. On that question, Sandefur's meta-analysis (Reference SandefurSandefur, 2015) offers important insights. Sandefur argues that there are two key types of professional expertise through which lawyers might make a difference in civil proceedings: substantive expertise and relational expertise.

Substantive expertise refers to lawyers' knowledge of the law and legal process. An essential aspect of lawyers' substantive expertise that can lead to positive case outcomes for their clients is the lawyers' understanding of the relevant legal rules (statutes, doctrines, case law, etc.), whether they are substantive or procedural. Substantive expertise also encompasses lawyers' knowledge about how to construct and present a set of facts in legally relevant terms (Reference SandefurSandefur, 2015, p. 3). One reason why substantive expertise is so important in court proceedings is because judges have substantial informational needs. Scholars of Supreme Court decision-making, for example, have pointed out: “In making decisions, the justices have certain information needs; they require a clear and faithful focus on the issues presented in a case, an understanding of the relationship of those issues to existing law, a clarification of uncertainties, and a view of the implications of a decision for public policy, tempered with candor” (Reference McGuireMcGuire, 1995).

Relational expertise, by contrast, “reflects skill at negotiating the interpersonal environments in which professional work takes place” (Reference SandefurSandefur, 2015, p. 16). Relational expertise can be real or perceived. On the latter type of relational expertise, Sandefur points out that “lawyer representation may act as an endorsement of lower-status parties that affects how judges and other court staff treat them and evaluate their claims” (Reference SandefurSandefur, 2015, pp. 16–17). Focusing on the flipside of this positive signaling effect of lawyers, Reference Quintanilla, Allen and HirtQuintanilla et al. (2017) have found evidence of negative signaling effects of pro se litigants. Their experimental study involving a realistic case file in a Title VII sex discrimination case showed that law-trained study subjects viewed unrepresented claimants as less competent than represented claimants (despite their identical case facts), and that these stereotypes explained why the study subjects awarded lower settlement awards to unrepresented claimants. Moreover, Reference Quintanilla, Allen and HirtQuintanilla et al. (2017) found this signaling effect among law-trained study subjects (law students and lawyers) but not among members of the lay public, suggesting that this type of a signaling effect may be a function of socialization within the legal profession.

In the immigration adjudication context, there is a growing body of research on the prevalence and the legal consequences of lack of access to counsel, but much less are known about how lawyers make a difference. In a study of asylum cases, Reference Miller, Keith and HolmesMiller et al. (2015a) highlight the importance of dyadic relationships between judges and lawyers. Miller et al. argue, “a large component of attorney success is found in being able to tailor an argument to a specific IJ, perhaps knowing and playing on their proclivities” (Reference Miller, Keith and HolmesMiller et al., 2015a, p. 230). Reference RyoRyo (2018) extends this literature by examining procedural and substantive differences between represented and unrepresented immigration bond hearings using matched samples. This study shows that represented detainees engaged in greater levels of court advocacy (such as submitting documents and making affirmative arguments for release) but finds no evidence that such activities explain the positive effect of representation on hearing outcomes.

Taken together, these studies suggest that a fuller understanding of the role of lawyers requires us to broaden the inquiry beyond an exclusive focus on lawyers to consider how the effect of their professional expertise may be moderated by judges and their decisional environments. As to which judge characteristics and contextual factors might play this moderating role—and why—findings from research on judicial decision-making offer important guidance.

Research on judicial decision-making

The formalistic legal model of judicial decision-making posits that judges make decisions based on the applicable law (see Reference CrossCross, 2003). Yet decades of empirical research across various areas of law show that judicial decisions are often better predicted by personal attributes of judges such as their race, gender, ideology, and length of tenure (see e.g., Reference Epstein, Landes and PosnerEpstein et al., 2013; Reference Harris and SenHarris & Sen, 2019; Reference Rachlinski and WistrichRachlinski & Wistrich, 2017), and contextual factors such as the legal, political, and institutional environment in which the judges are making their decisions (see, e.g., Reference Brace and HallBrace & Hall, 1997; Reference Engel and WeinshallEngel & Weinshall, 2020; Reference George and LeeGeorge & Lee, 1992). This approach to studying judicial decision-making has been influential in shaping the more nascent body of research on judicial decision-making in immigration courts, which has been motivated by concerns over large disparities in grant/denial rates across IJs and immigration courts (see, e.g., Reference HausmanHausman, 2016; Reference Kim and SemetKim & Semet, 2020; Reference RyoRyo, 2016).

In seeking to explain these disparities, studies of immigration adjudication have focused on a number of personal characteristics of IJs. For example, Reference Ramji-Nogales, Schoenholtz and SchragRamji-Nogales et al. (2007, p. 342) found that respondents appearing before female IJs (as opposed to male IJs) had a higher chance of prevailing in asylum claims. Reference Miller, Keith and HolmesMiller et al. (2015b) showed that increasing liberalism of IJs, as measured by their prior work experiences, increased the likelihood of asylum grant rates. The Government Accountability Office (2016, p. 32) found that IJs with more experience on the bench were less likely to grant asylum. Studies of immigration adjudication have also shown that certain contextual factors are significant predictors of IJ decision-making. For example, Reference Kim and SemetKim and Semet (2020) found that the political party of the US president at the time of the proceeding completion is a significant predictor of IJ removal decisions. This finding is generally consistent with Rottman et al.'s argument (Reference Rottman, Fariss and PoeRottman et al., 2009, p. 9) that IJs are “political decision-makers,” embedded in a milieu that “defines opportunities that are open to (them) and affects the probabilities that they will choose particular options over others.”

The foregoing review of research on judicial decision-making suggests that certain IJ characteristics and decisional environments are significant predictors of their decision-making. We argue that these same factors likely also shape IJs' predisposition toward and treatment of lawyers and their legal advocacy. Specifically, we focus on these judge characteristics that are most frequently examined in research on judicial decision-making: race, gender, political ideology, and tenure length. Existing research also points to the importance of examining the following dimensions of the IJs' political, legal, and organizational environment: the political party of the presidential administration in control at the time of the proceeding's completion, the federal judicial circuit in which the proceeding was adjudicated, and the caseload of IJs. Why might these judge characteristics and contextual factors amplify or dampen the potential impact of defense counsels' substantive expertise and/or relational expertise on case outcomes in removal proceedings? We draw on insights from studies of judicial decision-making to consider of number of possible explanations.

Moderating role of judge characteristics and contextual factors

Research that examines the impact of judges' race and gender on case outcomes have theorized that female judges and judges who are members of racial minority groups may vote differently than their white male counterparts, because the former set of judges use different judging styles or because they bring a different knowledge base that reflects the experiences of women or racial minorities (see Reference Harris and SenHarris & Sen, 2019, p. 247). Likewise, Reference Ramji-Nogales, Schoenholtz and SchragRamji-Nogales et al. (2007, p. 344) have argued that female IJs may have a greater inclination toward nonadversarial proceedings in their courtrooms and therefore more likely to “solicit a coherent and complete story” from the respondents (see Reference Ramji-Nogales, Schoenholtz and SchragRamji-Nogales et al., 2007, p. 344). If so, the same disposition may also make female IJs more open to fully considering the evidence and arguments presented by defense counsel appearing before them. The same type of dynamic may also apply to nonwhite IJs to suggest that IJs' race will moderate the effect of legal representation on case outcomes. In short, particular judging styles or personal backgrounds of IJs that vary along racial and gender lines may facilitate or hinder the exercise of substantive and relational expertise of defense counsel in removal proceedings.

Research on judicial decision-making has also emphasized the importance of understanding cognitive processes that impact judicial behavior. We highlight one such cognitive process here that may be directly relevant to this study. Scholars of the attitudinal model of judicial decision-making have argued that law serves primarily as a post hoc justification for decisions that reflect judges' political preferences (Reference Segal and SpaethSegal & Spaeth, 2002). Motivated reasoning—defined as “a biased decision process where decision makers are predisposed to find authority consistent with their attitudes more convincing than cited authority that goes against desired outcomes”—plays a central role in understanding this dynamic (see also Reference Braman and NelsonBraman & Nelson, 2007; Reference BybeeBybee, 2012, p. 74). This insight suggests that IJs' ideology might moderate the effect of legal representation on case outcomes in the following way: IJs who are less favorably disposed to noncitizens and their claims may be more likely to discount the arguments made by defense counsel and more resistant to the positive signaling effects of their presence in their courtrooms.

Another important judge characteristic that we examine is IJs' tenure. On the one hand, more experience on the bench may mean greater substantive legal expertise among judges, which in turn might reduce their informational reliance on defense counsel (see Reference Szmer and GinnSzmer & Ginn, 2014). On the other hand, the specific working environment of IJs suggests that the representation effect may become larger as tenure length increases. According to a survey of IJs by Reference Lustig, Delucchi, Tennakoon, Kaul, Marks and SlavinLustig et al. (2008a), IJs experience levels of stress and burnout that are higher than many other professional, including physicians in busy hospitals and prison wardens. Furthermore, Reference Lustig, Karnik, Delucchi, Tennakoon, Kaul, Marks and SlavinLustig et al. (2008b) has found that some components of this burnout are significantly associated with increased number of years on the bench. This makes sense given that the risk of stress and secondary trauma exposure resulting from adjudicating cases involving extremely vulnerable populations is likely cumulative (Reference AschenbrennerAschenbrenner, 2013).Footnote ⁵ Relatedly, longer years on the bench may also breed indifference or apathy. Under such circumstances, we might expect defense counsel to play a more substantial role for long-tenured IJs than for newly appointed IJs in terms of helping to focus judges' attention on salient legal and factual issues.

Next, we turn to contextual factors that we expect will play an important role in moderating the effect of legal representation on case outcomes. Here too, insights from research on judicial decision-making are instructive. Studies show that judges attend to the preferences of other political actors and are sensitive to public opinion in rendering certain types of decisions (Reference Brace and HallBrace & Hall, 1997; Reference GibsonGibson, 1980; Reference Rogol, Montgomery and KingslandRogol et al., 2018). Similarly, Reference Kim and SemetKim and Semet (2020) suggest that IJs take cues from and align their decision-making in accordance with real and perceived policy directives and preferences of the political leaders in power. In contemporary American politics, Democratic presidential administrations typically embrace a more pro-immigrant stance than Republican presidential administrations. Taken together, these studies suggest that all else being equal, IJs may provide defense counsel greater opportunities to exercise their substantive and relational expertise during Democratic administrations than during Republican administrations. Moreover, this dynamic may be even more pronounced in a legal environment that is protective of immigrant rights such as the Ninth Circuit (see, e.g., Reference NanosNanos, 2021). Lawyers—unlike most pro se respondents—are uniquely positioned to invoke substantive legal arguments and relevant legal rules, given their training and expected role in the courtroom. But their ability to do so is limited by the applicable substantive and procedural laws in effect.

Finally, research on judicial decision-making also suggests that the impact of defense counsels' substantive and relational expertise may be stronger when judges are under time pressure to make decisions. IJs generally face a large docket and rushed deliberation is common in many immigration proceedings. The following statement by an IJ is illustrative of this problem: “In those cases where I would like more time to consider all the facts and weigh what I have heard I rarely have much time to do so simply because of the pressure to complete cases. What is required to meet the case completion is quantity over quality.” (Reference Slavin and MarksSlavin & Marks, 2011, p. 1787). Research suggests that the pressure of unmanageable caseloads may induce judges to rely more heavily on instincts, heuristics, and stereotypes rather than reasoned legal analysis and the facts of each case (see Reference GuptaGupta, 2016; Reference MaroufMarouf, 2011; Reference Rachlinski and WistrichRachlinski & Wistrich, 2017). Under these circumstances, defense counsel's substantive and procedural expertise—or perhaps even just their mere presence in the courtroom that might provide a positive signal that bolsters the legitimacy of the case at hand—may operate as an important check against the activation of cognitive shortcuts and implicit biases among IJs.

In sum, IJs differ in their style of interacting (or not) with lawyers, and in their approaches to processing legal arguments and evidence presented by lawyers under different conditions. These varying levels and modes of engagement with defense counsel can either nullify or amplify the efficacy of the lawyers' substantive and relational expertise. To be clear, we do not and cannot formally test causal mechanisms through which the various IJ characteristics and contextual factors may moderate the effect of legal representation on case outcomes. However, testing these interactive relationships represents an important first step toward understanding both the potential and limits of legal representation in immigration courtrooms.

EOIR Dataset

Constructing the data on removal proceedings for our analysis required multiple steps. First, we combined separate data files released by the EOIR pertaining to cases, proceedings, hearings, judges, and attorney appearances. It is important to note that a given case may have multiple proceedings, and a given proceeding within that case may have multiple hearings. Next, we identified the relevant removal proceedings for inclusion in our analysis. The IIRIRA, which went into effect on April 1, 1997, substantially changed immigration law. To avoid the transition period post IIRIRA, we excluded proceedings that were completed before January 1, 1998. We also excluded proceedings that were completed after February 2020, as they were likely impacted by a substantial change in day-to-day administration and function of the immigration courts resulting from the COVID-19 pandemic. In short, our study period is between January 1, 1998 and February 29, 2020.

We subset the data to include only those proceedings that reached substantive merits decisions during our study period. We excluded stipulated removal proceedings, in absentia proceedings, and rider cases (for a description of these proceedings and cases, see EOIR, 2010; Reference KohKoh, 2017). Given that some cases can have multiple proceedings, we kept only the first proceeding within a given case that reached a merits decision. EOIR sometimes may update information on the assigned IJ in the proceedings data after the completion of a proceeding (this might happen, for example, if a motion to reopen is filed after a proceeding's completion and a different IJ is assigned to handle the motion). Therefore, we excluded any proceedings where the proceeding completion date came before the IJ's appointment date.Footnote ⁶ We also excluded any proceedings where the proceeding completion date preceded the assigned IJ's earliest hearing date in the hearings data file. Finally, we restricted our analysis to IJs who decided 100 or more proceedings during our study period. We describe in greater detail each of these steps and why they were necessary in the Data and Methods Appendix.

The EOIR Dataset contains a variety of information on noncitizens in removal proceedings and their case characteristics, such as the noncitizen's nationality, primary language, custody status (e.g., detained or not), and whether an attorney filed a notice of appearance. The EOIR Dataset also contains unique judge codes for presiding IJs, which allowed us to merge the EOIR Dataset with the Judge Dataset described below.

Judge Dataset

Although in more recent years, the EOIR has made public announcements of newly appointed IJs with their background information, EOIR has not released a comprehensive list of IJs who has ever served, along with their biographical information. Thus, we set out to develop such a database. Although our Judge Dataset does not contain every IJ ever appointed, it contains information on IJs who collectively account for 99% of all removal proceedings that reached a final merits decision between January 1998 and February 2020.

We based our sampling frame of IJs on two lists disclosed by the EOIR. The first is a list of IJ names and corresponding judge codes that the EOIR disclosed to TRAC pursuant to a Freedom of Information Act request (TRAC List). The TRAC List contains 625 unique judge name-judge code combinations pertaining to IJs who were active during or prior to fiscal year 2016. The second EOIR disclosure on which we rely consists of a table of IJs found in a broader set of case data files that the EOIR posted on its website (Judge Lookup Table). The Judge Lookup Table contains 785 unique judge name-judge code combinations. We paired the Judge Lookup Table with the TRAC List to create the master list containing 756 unique judge name-judge code combinations.Footnote ⁷

Finally, using a variety of sources, we coded each IJ's biographical information, such as his or her gender, race, work experience, and appointment year. These sources include official announcements of newly appointed judges that the EOIR makes available online,Footnote ⁸ judge biography files that we obtained from TRAC, and our own independent internet searches.

Measures

Table A1 contains summary descriptions of all of the measures that we discuss below.

Analytical strategy

To test for the existence of contingent effects of legal representation in removal proceedings, we estimated a series of logistic regression models using maximum likelihood. For each of the variables previously described in the Judge Variables and Contextual Variables sections, we estimated a baseline logistic regression model that included a term for the focal variable, a term for legal representation, and a term for the interaction between the focal variable and legal representation.Footnote ¹⁰ For the statistically significant interactions, we then estimated a series of models that included a term for the focal variable, a term for legal representation, a term for the interaction between the focal variable and legal representation, all control variables, and circuit fixed effects.Footnote ¹¹ We refer to this latter set of models as full models, which take the form:

$\text{logit}\left(\pi \right)=\alpha +{\beta }_{1}X+{\beta }_{2}Z+{\beta }_3\mathrm{XZ}+{\beta }_{4}Q,$

where logit(π) represents the log odds of Removed = 1 versus Removed = 0. α and βs are parameters to be estimated. X represents Attorney. Z represents the Judge or Contextual variable of interest. XZ is the product term of Attorney and the Judge or Contextual variable of interest. Q represents the covariates as discussed earlier.

When discussing the estimates from our models, we do not discuss the coefficients, as the coefficients in nonlinear models do not necessarily provide accurate information about the interaction effects of interest (see Reference Ai and NortonAi & Norton, 2003; Reference MizeMize, 2019). Rather, we discuss the results in terms of predicted probabilities, first differences, and second differences. First difference refers to the average marginal effect (AME) of switching from 0 to 1 for Attorney for any given subpopulation of interest (e.g., female IJs). Second difference refers to the difference in AMEs for Attorney between any given pair of subpopulations (e.g., female IJs vs. male IJs). We use Wald tests to assess whether the second differences are statistically distinct from 0. We performed a series of robustness checks using these full models to test the sensitivity of our results to different data management and modeling assumptions. Across these various checks, we found that results were consistent. Thus, we present only the results from our main analysis to conserve space.Footnote ¹²

A note on selection bias is in order. Noncitizens who have legal representation may differ systematically from those who lack representation. Observational studies that seek to examine the effect of legal representation on case outcomes are concerned with such selection issues, because any observed differences in case outcomes between represented and unrepresented cases may be due to this selection effect. This problem is less salient in the current study, as we do not focus on the main effect of legal representation on case outcomes but rather, on the interactional effect between legal representation and judge characteristics, and between legal representation and contextual factors, respectively. More specifically, we have no reasons to believe that the potential selection effects we have described above systematically vary by judge characteristics or decisional environments, given that IJs and lawyers have no control over which individual cases get assigned to which judge.Footnote ¹³ Nonetheless, we re-estimated our regression models using a matched sample of cases that are comparable on various case characteristics but for their represented/unrepresented status. The results of this analysis showed relatively smaller first differences than those we found using the unmatched sample. But our key findings relating to the second differences—that is, the contingent effects of representation—remained substantially the same. The matching procedure that we used and the results of the regression models using the matched sample are described in the Data and Methods Appendix.

Descriptive analyses

Table 1 reports IJ-level descriptive statistics for IJs included in our analytic sample. About 39% of the IJs in the sample are women. Nearly 68% are non-Hispanic white. The median value on IJ Liberalism is −0.325; the highest value (0.446) is about 4/5th of a point higher than the median, and the lowest value (−1.369) is over a whole point lower than the median. The median year of appointment for the IJs in the sample is 2010, with one IJ having been appointed as early as 1966 and dozens of IJs appointed in 2019. Government, INS/DHS, and private practice are the three most common types of work experiences among IJs, with about 62%, 56%, and 46% ever having worked in those settings, respectively, before being appointed to the bench. Experiences in the military, NGO, and academia are less common, with about 14%, 13%, and 3% ever having worked in those settings, respectively, before being appointed to the bench. Finally, about 46% of the IJs were appointed during a Democratic administration.Footnote ¹⁴

Table 1. Descriptive statistics on judges

Note: N = 676 judges.

Abbreviations: DHS, Department of Homeland Security; IJ, immigration judge; INS, Immigration and Naturalization Service.

Table 2 provides proceedings-level descriptive statistics for all variables used in our analysis. Nearly 80% of proceedings in our sample resulted in removal orders. About 40% of the proceedings had legal representation. The rest of our discussion focuses on the judge variables and the contextual variables. About 28% of the proceedings were decided by female IJs, and over 70% of the proceedings were decided by white IJs. The mean value on IJ Liberalism was −0.333, and the average number of years on the bench at the time of case completion was over 9 years. About 46% of proceedings were completed while a Democratic president was in office. Nearly one-third of the hearings took place in the Ninth Circuit Court of Appeals. Finally, the median caseload during the 2-week period immediately preceding the completion of a given proceeding was 1.7 individual hearings per day.

Table 2. Descriptive statistics on variables used in the analyses

Note: N = 1,906,832 proceedings.

Regression analyses

We now turn to our regression analysis results. To conserve space, our discussion focuses only on the full models that have statistically significant interactions, as evidenced by second differences that are statistically distinct from zero (p < 0.05).

Judge characteristics

Figure 1 shows predicted probabilities of removal from the full model by IJ gender and whether the proceeding had legal representation. Table 3 displays these same probabilities in the first column (Pr[removal]). The second column shows the size of the decrease in probability of removal associated with adding legal representation for male IJs and female IJs, respectively (first difference). The last column shows the difference in the first differences between male IJs and female IJs (second difference). The results show that, controlling for all other variables in the model, proceedings with legal representation have a considerably lower probability of removal among both male IJs (Δ = 0.433; p < 0.001) and female IJs (Δ = 0.471; p < 0.001). The reduction in probability of removal associated with legal representation, however, is larger for female IJs than it is for male IJs—by nearly 4 percentage points (Δ = −0.038; p < 0.01).

Figure 1. Probability of removal by IJ gender and attorney (full model)

Table 3. Probability of removal by IJ gender and attorney (full model)

Note: Covariates include all judge variables except Female IJ, all contextual variables, and all control variables. This model includes circuit fixed effects. Standard errors are clustered by IJs in the model. First and second differences are calculated using seven significant figures.

^*p < 0.05, ^**p < 0.01, ^***p < 0.001 (two tailed tests).

Figure 2 uses estimates from the full model to plot the differences in predicted probability of removal for proceedings with legal representation and proceedings without legal representation, across number of years on the bench. Figure 2 shows that, controlling for all other variables in the model, the reduction in probability of removal associated with legal representation becomes larger as the number of years on the bench increases. The first column in Table 4 (Pr[removal]) provides predicted probabilities of removal from the full model for IJs with the median value for Years on Bench (median = 9 years) and a one standard deviation (SD) increase from median value for Years on Bench (median + 1 SD = 15.955 years), by whether a proceeding had legal representation. The second column (first difference) shows the reduction in probability of removal associated with having legal representation for these two values of Years on Bench, respectively. The last column (second difference) shows the difference in the size of this reduction across the two values of Years on Bench.

Figure 2. Average marginal effect of attorney by years on the bench (full model)

Table 4. Probability of removal by years on bench and attorney (full model)

Note: Covariates include all judge variables except Years on Bench, all contextual variables, and all control variables. This model includes circuit fixed effects. Standard errors are clustered by IJs in the model. First and second differences are calculated using seven significant figures.

^*p < 0.05, ^**p < 0.01, ^***p < 0.001 (two tailed tests).

Table 4 shows that at both values for Years on Bench, proceedings with legal representation have significantly lower probabilities of removal (Δ for median = 0.448, Δ for median + 1 SD = 0.467; for both, p < 0.001). Finally, Table 4 indicates that a one SD increase from the median value in Years on Bench contributes to an additional 2 percentage point reduction in the probability of removal for those with legal representation (Δ = −0.020; p < 0.001).

We did not find statistically significant interaction effects (at p < 0.05) for IJs' race (White) and ideology (IJ Liberalism).

Contextual factors

So far, we have discussed how the effect of legal representation on removal proceeding outcomes varies by certain IJ characteristics. Now we turn to our analysis of how the effect of legal representation might vary by key contextual factors. Figure 3 shows predicted probabilities of removal from the full model by political party of the presidential administration at the time of the proceeding's completion, and whether a given proceeding had legal representation. Table 5 displays these same probabilities in the first column (Pr[removal]). The second column shows the size of the decrease in probability of removal associated with adding legal representation for Republican and Democratic administrations, respectively (first difference). The last column shows the difference in the first differences between Republican and Democratic administrations (second difference).

Figure 3. Probability of removal by Democratic administration and attorney (full model)

Table 5. Probability of removal by Democratic administration and attorney (full model)

Note: Covariates include all Judge Variables, all contextual variables except Democratic Administration, and all control variables. This model includes circuit fixed effects. Standard errors are clustered by IJs in the model. First and second differences are calculated using seven significant figures.

^*p < 0.05, ^**p < 0.01, ^***p < 0.001 (two tailed tests).

The results show that, controlling for all other variables in the model, proceedings with legal representation have a considerably lower probability of removal in both Republican (Δ = 0.420; p < 0.001) and Democratic administrations (Δ = 0.491; p < 0.001). Nonetheless, the reduction in probability of removal associated with legal representation is nearly 7 percentage points larger during Democratic administrations than during Republican administrations (Δ = −0.071; p < 0.001).

Figure 4 shows predicted probabilities of removal from the full model by whether a given proceeding was adjudicated in an immigration court located in the Ninth Circuit Court of Appeals, and whether the proceeding had legal representation. Table 6 shows these same probabilities in the first column (Pr[removal]). The second column shows the size of the decrease in probability of removal associated with adding legal representation for all federal judicial circuits other than the Ninth Circuit Court of Appeals and the Ninth Circuit Court of Appeals, respectively (first difference). The last column shows the difference in the first differences between all federal judicial circuits other than the Ninth Circuit Court of Appeals and the Ninth Circuit Court of Appeals (second difference).

Figure 4. Probability of removal by ninth circuit and attorney (full model)

Table 6. Probability of removal by ninth circuit and attorney (full model)

Note: Covariates include all Judge Variables, all contextual variables except Ninth Circuit, and all control variables. Standard errors are clustered by IJs in the model. First and second differences are calculated using seven significant figures.

^*p < 0.05, ^**p < 0.01, ^***p < 0.001 (two tailed tests).

The results show that, controlling for all other variables in the model, proceedings with legal representation have a substantially lower probability of removal across all circuits (Ninth Circuit: Δ = 0.527, other circuits: Δ = 0.423; for both, p < 0.001). However, legal representation reduces the probability of removal by an additional 10 percentage points in the Ninth Circuit Court of Appeals compared to all other federal judicial circuits (Δ = −0.104; p < 0.001).

Finally, Figure 5 uses estimates from the full model to plot the differences in predicted probability of removal for proceedings with legal representation and proceedings without legal representation, across IJs' current caseload. Figure 5 shows that controlling for all other variables in the model, the reduction in probability of removal associated with legal representation becomes larger as the current caseload increases. The first column in Table 7 (Pr[removal]) provides predicted probabilities of removal from the full model for IJs with the median value for Current Caseload (median = 1 individual hearing per day) and a one SD increase from median value for Current Caseload (median + 1 SD = 3.024 individual hearings per day), by whether a proceeding has legal representation. The second column (first difference) shows the reduction in probability of removal associated with having legal representation for the two values of Current Caseload, respectively. The last column (second difference) shows the difference in the size of this reduction across the two values of Current Caseload.

Figure 5. Average marginal effect of attorney by current caseload (full model)

Table 7. Probability of removal by current caseload and attorney (full model)

Note: Covariates include all judge variables, all contextual variables except Current Caseload, and all control variables. This model includes circuit fixed effects. Standard errors are clustered by IJs in the model. First and second differences are calculated using seven significant figures.

^*p < 0.05, ^**p < 0.01, ^***p < 0.001 (two tailed tests).

Table 7 shows that at both values for Current Caseload, proceedings with legal representation have significantly lower probabilities of removal (Δ for median = 0.443, Δ for median + 1 SD = 0.458; for both, p < 0.001). Finally, Table 7 shows that a one SD increase from the median value in Current Caseload contributes to an additional 1.5 percentage point reduction in the probability of removal for those with legal representation (Δ = −0.015, p < 0.001).

It is worth noting that among unrepresented cases, the differences in the probability of removal across various judge characteristics and contextual factors, respectively, are relatively small compared to the same set of differences among represented cases. Specifically, for each of the judge characteristics and contextual factors, we examined the first difference (e.g., the probability of removal among male versus female judges) among unrepresented cases, and we did the same among represented cases. We found that the magnitude of these first differences are consistently and significantly larger among represented cases than unrepresented cases, which suggests that representation exacerbates outcome disparities attributable to varying judge characteristics and contextual factors.

It is also important to note that although some of our findings indicate a relatively small effect size for any given judge characteristic or contextual factor, those findings can translate into a large substantive impact when considered in light of the sheer number of individuals facing removal every year. For example, recall that our analysis showed that the reduction in probability of removal associated with legal representation is 3.8 percentage points larger for female IJs than it is for male IJs. As we noted earlier, the US government initiated over 690,000 removal proceedings in fiscal year 2019. A difference of 3.8 percentage points for this population alone could mean a difference of 26,224 fewer removal orders.