The Decade of Immigrant Dispersion and Growth: A Cohort Analysis of Children of Immigrants‘ Educational Experiences 1990-2002

Abstract

The 1990s marked the beginning of a new era of immigration in terms of volume and settlement patterns and also witnessed significant changes in the social contexts confronting immigrants. These changes could have significant repercussions for immigrant youth. While previous research on high school dropout behavior suggests immigrant youth are faring better in US schools, our research provides a less optimistic outlook. Using the National Educational Longitudinal Study (1988) and Educational Longitudinal Study (2002), we use multivariate analysis, regression decomposition and fixed effect models to examine how reading and math test scores of children of immigrants changed during the 1990s.

Beginning in the 1990s, US immigration trends changed dramatically both in terms of size and geographic dispersion. Approximately 16 million immigrants entered the US during the 1990s, a significant increase from the previous two decades (10 million in 1980 and 7 million in 1970) and the highest in the nation's history (Capps et al., 2005). Changing the geography of immigrant settlement, many of these immigrants and their families moved to new metropolitan areas and rural locations throughout the West, Southeast, and Midwest (Fortuny et al., 2009; Massey, 2008; Singer, 2004). By 2006, more than one in five children across the nation had at least on foreign-born parent, and in 29 states more than 10% of the total child population had one-foreign born parent (Fortuny et al., 2009).

The extent to which these immigrant youth succeed in the US school system will determine their future economic and social opportunities (Heckman, Stixrud and Urzua, 2006) and shape the path of immigrant incorporation in the US for generations to come (Portes and Rumbaut, 2001). Classical immigration theory (Gordon, 1964) emphasized education as the key variable determining the upward trajectory of children of immigrants. The segmented assimilation framework (Portes and Rumbaut, 2001) still sees education as crucial to assimilation, but argues that due to a combination of differences in government reception of immigrants, different family resources, and different school and neighborhood contexts, not all immigrant groups can be expected to have equal assimilation outcomes in areas such as education. Understanding how the children of immigrants are doing in school is thus crucial to inform heated policy debates on the desired size and type of immigrants entering the US as well as policies thought to affect educational performance in general, such as neighborhood segregation and school funding.

Hence, in this paper, we focus on factors that help or hinder the educational performance of recent immigrants, thereby contributing important descriptive data to ongoing debates about immigration policy and immigrant assimilation. We focus on changes during the 1990s because this decade marked the beginning of a new era of immigration in terms of volume and settlement patterns but also witnessed significant changes in the political and economic contexts confronting new arrivals. For example, what does the educational performance of children of immigrants look like in the Southeast and West, regions that either have fewer resources in schools or simply are not used to dealing with immigrants? Do more recent immigrant families, in particular the increasing number of Asian and Hispanic families, have fewer resources to pass on to their children to help their education than previous immigrants? Has the rise in immigration created segregated neighborhoods that negatively impact the educational performance of children living in those neighborhoods? How has the increase in unauthorized immigrants since 1990 impacted educational performance of children of immigrants? Did the economic upswing of the 1990s lift the children of immigrants through expanded economic resources in their families? How has the rise in anti-immigrant sentiment and the adoption of restrictive immigrant policies affected children of immigrants?

Part of the reason for the current impasse on immigration policy is due to debates about the relative importance of the quality of resources immigrants bring with them and the political and economic conditions shaping the outcomes and assimilation of children of immigrants. Like other important work, such as Fry (2007), Glick and White (2003) and Bean and Stevens (2003), that looked at population-level data to contribute to these heated debates, our paper analyzes national-level educational data at two time-points to try to establish more clearly what has happened to the educational performance of children of immigrants.

Our study thus advances beyond current research on children of immigrants in US schools, which has thus far produced mixed results. In his assessment of educational trends during the 1990s, Fry (2007) found that foreign-born teens had lower high school dropout rates in 2000 than 1990, but that they were also more likely to be classified as limited English proficient (LEP). Thus, while immigrant youth may be more likely to stay in school, they may not be acquiring the basic skills associated with improved educational prospects and labor market outcomes (Bleakley and Chin, 2004). Other research on the Asian-White test score gap further suggests that immigrant youth are not faring better in US schools. Galindo and Pong (2011) found that between 1990 and 2002 the 10^th grade math test scores of Asians, particularly those with foreign-born parents, decreased in comparison to the test scores of non-Latino white youth born in the US to US born parents (i.e. 3^rd generation or higher). Consequently, by 2002 Asian youth who have been promoted as the “model minority” were no longer academically outperforming their non-Latino white counterparts.

This study builds on prior research by examining whether immigrant youth across all ethnic/racial groups enrolled in US schools are faring better than previous cohorts in terms of academic performance in both reading and math. Moreover, by including school characteristics and school fixed effects, our study is the first to examine whether US schools are more effectively educating immigrant youth. Taking advantage of the coincidental timing of data collection for the National Educational Longitudinal Study conducted in 1988 (NELS:88) and the Educational Longitudinal Study conducted in 2002 (ELS:2002), we address the following research questions: 1) How did the academic performance of 1^st, 2^nd, and 3^rd generation immigrants change during the 1990s? 2) How did changes in socio-demographic, family, school, and neighborhood characteristics contribute to differing cohort performance patterns? To answer these questions, we examine math and reading test scores of 10^th grade youth because early high school performance is a strong determinant of high school completion (Driscoll, 1999).

The Role of Selection and Context across Immigrant Cohorts

Studies that compare immigrants over time focus on two main factors to explain observed differences across cohorts. The first is change in immigrant selectivity that stems from the fact that later waves of immigrants are likely to have lower levels of education and job skills than earlier waves (Borjas, 1995). According to network migration theories (Bean and Stevens, 2003), this reduction in immigrant ‘selectivity’ reflects the dynamic nature of migration streams. In the beginning, migration streams are headed by the more ambitious, risk-taking immigrant pioneers who have the resolve to establish a new life for themselves in a foreign-land and by doing so establish support systems that reduce the costs (economic and psychological) of future migration. Once a critical mass of immigrants develops and social networks expand, the flow of migrants increases and the selectivity of migrants decreases.

While not negating the importance of immigrant selection, the second explanation focuses more on the context of reception and how economic and political conditions shape the opportunity structure once immigrants arrive to the US. According to this perspective, the economic and social barriers (particularly ethnic/racial discrimination) that immigrants encounter in the US limit their ability to capitalize on their own resources and the support systems of their ethnic community (Portes and Rumbaut, 2001).

The few studies that have examined cohort changes in the educational outcomes of children of immigrants provide support for both of these explanations. Examining baseline test scores and the trajectory of achievement of immigrant youth in 1980 and 1990, Glick and White (2003) found that while the starting point of immigrant youth may differ across points in time, their trajectory of achievement is shaped more by structural measures, particularly ethnic/racial status, that reflect the broader social context they encounter. Thus, while variation in immigrant selectivity may create an initial advantage or disadvantage for different waves of immigrants, Glick and White's results suggest that these different waves will become more similar over time and reflect broader ethnic/racial trends in achievement rather than immigrant trends per se.

The work by Galindo and Pong (2011) on the Asian-White test score gap, however, suggests that changes in immigrant selectivity during the 1990s may be markedly different from previous cohorts and result in new patterns of immigrant academic performance. The authors caution that the decline in Asian immigrant test scores (compared to whites) they observed may worsen as the Asian immigrant population grows more diverse. Adding to this concern, Fry's (2007) research suggests that though immigrant youth were less likely to drop out of high school in 2000 than in 1990, they did not experience the same decline in the likelihood of dropping out as their native peers. Moreover, his results suggest that during this decade schools have been less effective at promoting English language proficiency among immigrant youth.

The unprecedented changes in both the social environment and background characteristics of immigrants that occurred during the 1990s could have significant repercussions for the academic adaptation of immigrant youth. While our paper will not be able to resolve the debate between immigrant context and selection, we provide evidence on how both of these factors have changed during the 1990s and the educational implications of these changes.

The 1990s: A Decade of Change

The economic boom of the 1990s and the accompanying growth and dispersion of the immigrant population have both positive and negative implications for the social environment that shapes children of immigrant's educational experience. On the one hand, the economic gains immigrant families made during the 1990s most likely advanced the academic performance of children of immigrants. Examining poverty trends among children of immigrants, Van Hook, Brown and Kwenda (2004) found that while poverty rates had been increasing since the 1970s they peaked between 1997 and 1998 and by 2000 had declined to levels lower than in 1990. They also found that parental work effort was higher in 2000 than in 1990, a factor that likely contributed to observed gains in household median income among immigrant households. In fact, between 1994 and 2000 median family income grew faster among immigrant households than native households with a percent change of 26.3% compared to 13.1% (Chapman and Bernstein, 2003). Extant research has shown that familial economic resources are strongly associated with children's academic well-being (Sirin, 2005).

On the other hand, the growth and dispersion of immigrant families during the 1990s created significant challenges for the US educational system that may have hindered academic performance. Between 1990 and 2000, the size of the children of immigrant population grew from 13% of all public school children to over 20% (Passel, 2011). As a consequence, US schools have been overwhelmed by the dramatic influx of a minority population with limited English fluency, few economic resources, and varying educational backgrounds (Van Hook and Fix, 2000). Moreover, prior research indicates that many schools and educational services, particularly those in new settlement states, lack the infrastructure, social resources, and institutional support systems that promote the academic adaptation of immigrant youth (Capps et al., 2005; Consentino de Cohen, Deterding, and Clewell, 2005; Gozdziak and Martin, 2005; Massey 2008; Perreira, Chapman and Levis-Stein, 2006).

Adding to these educational challenges, the political backlash that arose in response to the extraordinary growth in immigration during the 1990s may have also hindered academic performance. Responding to anti-immigrant sentiment, federal, state and local policymakers have adopted exclusionary policies that limit immigrants’ rights. Among the more notable policies occurred in 1996 when the federal government severely restricted the economic and educational resources of immigrant families by: 1) excluding legal noncitizen immigrants (i.e. green card holders) from receiving federally-funded safety net programs (under the Personal Responsibility and Work Opportunity Reconciliation Act; Fix and Passel, 2002); and 2) prohibiting states from providing in-state resident tuition benefits to undocumented immigrants (under the Illegal Immigration Reform and Immigrant Responsibility Act; Flores, 2009). Research suggests both of these provisions increased immigrant youths’ risk for school failure by reducing familial resources (Kalil and Crosby, 2010; Borjas, 2004; Van Hook and Balistreri, 2006) and discouraging high school completion (Potochnick 2014) and post-secondary enrollment (Flores, 2009; Kaushal, 2008). Moreover, these studies suggest that the negative effects of these policies were not limited to undocumented and non-citizen immigrants but had spillover effects on all children of immigrants.

The Changing Characteristics of Immigrants

In addition to learning to adapt to a growing immigrant population, schools in the 1990s were also challenged with educating a changing immigrant citizenry. The composition of immigrants and familial resources of immigrant families in 2000 were distinct from that in 1990. Some of these changes are positively associated with educational performance, but the majority of the changes create additional educational challenges.

On the positive side, compared to the previous decade, immigrant youth in 2000 benefited from some higher socio-economic indicators, including improved English language ability, higher income, and smaller family size (Fry, 2007; Van Hook, Brown, and Kwenda, 2004). The strong influence families, both immigrant and nonimmigrant, have over their children's educational aspirations and performance is well documented (Goyette and Xie, 1999; Glick and White, 2003; Kao and Tienda, 1995; Fuligni, 1997; Fuligni and Fuligni, 2007; Perreira, Harris and Lee, 2006; Rumbaut, 1999). Of all the familial characteristics, research suggests that parental socio-economic status (SES), which incorporates elements of both financial and human capital, is the strongest predictor of academic performance (Glick and White, 2003; Sirin, 2005). For immigrant families, English language usage is another important human capital resource. Research indicates that English language ability of both the parent and child as well as the usage of English in the home can have a positive impact on academic performance (Glick and White, 2003; Fuligni and Fuligni, 2007; Perreira, Harris, and Lee, 2006).

There is mixed evidence, however, as to whether parental education, a key aspect of SES was higher or lower in 2000. In his assessment of foreign-born teens (15-17 years old), Fry (2007) found that more foreign-born heads of households had a high school degree and had completed some college in 2000 than in 1990. When examining both foreign- and US-born children of immigrants, however, Van Hook, Brown and Kwenda (2004) found that parental education was lower in 2000 than 1990. Given the strong influence of parental education, lower levels of parental education in 2000 (if that occurred) may offset the other socio-economic gains immigrant families made since 1990. Moreover, the growth in single-parent households during the 1990s may further offset SES gains (Van Hook, Brown and Kwenda, 2004; NCES, 2006).

Another notable change during the 1990s was the growth of the minority population driven by immigrant youth who were more likely to be of minority status than non-immigrant youth (Passel, 2011). This growth was especially pronounced among Latinos and Asians, the vast majority of whom (60% and 85%, respectively) are children of immigrants (Passel, 2011). Extant research has found significant variation in academic performance across racial/ethnic groups. Of all ethnic groups, Asian American immigrants (with notable exceptions such as Cambodians and Laotians) perform the best academically on multiple educational outcomes (e.g., grades, test scores, high school completion rates, and college enrollment levels) while Latino immigrants, especially Mexicans and Puerto Ricans, perform the worst (Glick and White, 2004; Kao and Thompson, 2003; Kao and Tienda, 1998; Keller and Tillman, 2008; Perreira, Harris and Lee, 2006; Rumbaut, 1999) and white and black immigrant youth fall somewhere in between (Kao and Tienda, 1998; Keller and Tillman, 2008; Perreira, Harris and Lee, 2006). Given Glick and White's (2003) findings that immigrant youth's educational achievement is shaped more by discrimination barriers associated with their ethnic/racial background than immigration status, this rise in the minority immigrant population may significantly influence cohort achievement patterns.

A particularly challenging (but often un-measurable) compositional change in immigration between 1990 and 2000 was the rise in the share of immigrants who were unauthorized. By the late 1990s, more unauthorized migrants entered the US each year than legal immigrants, which resulted in a larger share of unauthorized immigrant youth and youth living in mixed status families (Passel, 2011; Passel and Suro, 2005). By 2009, these youth made up about 6.9% of the total child population and 30% of the immigrant youth population (Passel, 2011). Unauthorized children and children living in mixed status families face several significant educational challenges, including high rates of poverty (40% compared to 17% of US born children; Gonzales, 2009), full or partial (i.e. only some family members qualify) exclusion from vital social services and education programs, and the potential deportation of family members (Suàrez-Orozco et al., 2011). The educational challenges associated with being unauthorized or living in a mixed status family should be most apparent among Latinos, who make up approximately 80% of the unauthorized population (Passel, 2011).

Changes in School and Neighborhood Context

As the national landscape and immigrant population have changed, so have the nation's schools and neighborhoods. In an era of both record high immigration flows and school re-segregation levels, there is significant concern that schools will be able to successfully foster the academic adaptation of immigrant children. Creating a triple disadvantage for many children of immigrants, U.S. schools have re-segregated across racial/ethnic, linguistic, and economic divisions (Orfield and Lee, 2005; Murray, Batalova and Fix, 2000; Van Hook and Fix, 2000). Segregation literature indicates that the lower school quality—measured by poverty rates, school resources, and teacher characteristics—at segregated schools and their concentration of disadvantaged students and prevalence of negative teacher perceptions detracts from students’ aspirations and achievement (Hanushek, Kain and Rivkin, 2009; Mickleson, 2006; Orfield and Lee, 2005). Thus, to assess whether our nation's schools have made progress on educating children of immigrants, this paper examines how the composition and resources (measured by school quality proxies and school fixed effects) of the schools immigrant youth attend changed during the 1990s and the implications of these changes for test scores.

We also examine the educational implications of any changes that may have occurred in the types of neighborhoods where immigrant youth live. While results are not uniform, evidence suggests that compared to adolescents living in more advantaged neighborhoods (i.e., wealthier, racially integrated, and more educated), adolescents residing in disadvantaged neighborhoods are more likely to drop out of high school (Crowder and South, 2003; Perreira, Harris and Lee, 2006), have lower grades and test scores (Ainsworth, 2002; Pong and Hao, 2007), and complete fewer years of schooling (Mayer, 2002).

Study Design

Data and Sample

This analysis utilizes data from the first follow-up of the National Educational Longitudinal Study of 1988 (NELS:88) and from the base year of the Educational Longitudinal study of 2002 (ELS:2002) both of which are sponsored by the National Center for Education Statistics (NCES). For the NELS:88 survey, data were collected on a cohort of approximately 25,000 eighth graders from a sample of 1,052 schools beginning in the spring of 1988 with follow-ups conducted in 1990, 1992, 1994, and 2000 (NCES, 2002). In the 1990 follow-up, NCES freshened the sample by adding new respondents to ensure that the sample was nationally representative of all 10^th graders. To compare the sophomore cohorts of NELS:88 and ELS:2002, we follow the practice of other researchers by using the freshened sample of NELS:88 and eliminating base-year respondents who were no longer in school or were not in 10^th grade in 1990 (Glick and White, 2003; Goyette, 2008).

For ELS:2002, which was designed to allow cross-cohort comparisons with NELS:88 (and other NCES datasets), data were collected on a cohort of over 15,000 tenth graders from a sample of 750 schools beginning in the spring of 2002 with follow-ups conducted in 2004 and 2006. Though there are differences across the two studies, NCES and other researchers have concluded that the similarities in the question wording and sampling strategies allow the two studies to be used for cross-cohort comparisons (Dumais, 2009; Goyette, 2008; NCES, 2005). A main difference between the two datasets is that missing data have been statistically imputed in ELS:2002 for some variables (e.g., test scores, SES, and race/ethnicity) but not in NELS:88.

Providing rich contextual information, both datasets collected information from students, parents, teachers, and school administrators, and the restricted datasets (for which we have licensed access) can be connected to zip code level census data (1990 for NELS:88 and 2000 for ELS:2002) to identify neighborhood characteristics. As with most national-level data, the NELS:88 and ELS:2002 studies do not contain information on neighborhood boundaries. Instead, we follow the work of other researchers and use the smallest ecological unit available (i.e. zip codes) in order to reduce measurement error (Ainsworth, 2002; Goldsmith, 2003). Lastly, the large sample sizes and the over-sampling of minority students in both datasets make it possible to adequately assess the influences of ethnic/racial differences and immigrant status among youth in both cohorts.

We include all self-identified Latino, and non-Latino white, black, Asian, and other race (i.e. Native Americans and multi-racial youth) students in the sample (N=16,590¹ in NELS:88; N=15,890 in ELS:2002). No students had missing values on the dependent variables in ELS:2002 because NCES imputed missing test scores as well as some (e.g., SES and race/ethnicity) but not all (e.g., generational status and school variables) of the independent variables used. In contrast, approximately 1,370 students had missing values on one or more of the dependent variables in NELS:88. Thus, to reduce missing observations in both ELS:2002 and NELS:88 and to better match NELS:88 test scores with the NCES imputed ELS:2002 test scores, we correct for missing data on the dependent and independent variables by imputing missing data using multiple imputation by multivariate normal regression (the MVN command in STATA).²

Measures

Academic Performance

We use reading and math test scores as our indicator for student academic performance for two reasons. First, states have increasingly relied on standardized tests in both math and reading to measure school performance and to serve as requirements for high school graduation (Hanushek and Raymond, 2005). Thus, these measures provide an indication of how schools are faring. Second, math and reading ability are key cognitive indicators that have been shown to affect future labor market outcomes (Farkas, 2003). The written and oral communication skills associated with reading tests and the problem solving skills associated with math tests increase productivity in the labor force (Chiswick and Miller, 2003; Farkas, 2003). For children of immigrants, research has largely focused on reading skills under the assumption that English language ability is directly tied to performance in reading. Recent research, however, suggests that English language ability is as important for performance in math (Beal, Niall and Cohan, 2010). Thus, reading and math are both important indicators of children of immigrants’ cognitive ability and future labor market potential.

To allow for cross-cohort comparisons, we follow prior research (Glick and White, 2003; Goyette, 2008) and use standardized test scores.³ Created by NCES, the standardized test score provides an indicator of academic performance relative to the 10^th grade population and has a mean value of 50 and standard deviation of 10. Thus, across cohorts we measure the relative performance of groups and not absolute cognitive gains. To interpret the effect sizes of coefficients (or mean differences) one simply divides the coefficient by 10 to calculate the difference as a proportion of a standard deviation. For instance, a coefficient size of three indicates about a quarter of a standard deviation difference in academic performance and a coefficient of one indicates a tenth of a standard deviation difference. Because we use two cross-sectional measures of test scores our results provide a baseline assessment of academic performance, not a longitudinal assessment of academic achievement or trajectory over time (cf. Glick and White 2013).

Student Background

Research has shown that educational outcomes vary across a variety of demographic characteristics, including age, gender, ethnic/racial group, and generational status (Kao and Thompson, 2003). To control for these differences, we include a dummy female gender variable to determine the influence of gender on student test score performance and use a month-based age variable to control for the influence of age. To control for variations in academic performance among different ethnic groups, we create five mutually exclusive race/ethnic categories: Latino and non-Latino white (reference category), black, Asian, and other race. It is important to note that the make-up of the other race category differs slightly between ELS:2002 and NELS:88. Unlike NELS:88, ELS:2002 allowed students to identify as multi-racial. Thus, while the number of youth identifying as multi-racial is small, the other race category in ELS:2002 captures these youth as well as Native Americans. Lastly, we use a three-category classification of generational status: 1^st generation (both child and parents were foreign-born), 2^nd generation (child was US-born and at least one parent was foreign-born) and 3^rd generation and higher (child and both parents were US born).⁴

Family Context

Levels of human capital in immigrant families depend on the economic, educational, and linguistic resources of parents as well as the structure of the family. To measure the family's economic and educational well-being, we use the standardized scale of socioeconomic states (SES; range: −3.29 to 2.76) created by NCES, which is a composite measure combining information on the mother's and father's education, income, and occupation.

To measure the linguistic resources of immigrant families, we rely on a proxy measure—student's English language ability.⁵ We create the measure by averaging the self-reported scores students gave about their reading, writing, listening, and/or speaking ability on a scale from 1=“very well” to 4=“not very well.” We reverse code the scale so a higher score indicates stronger English language ability and code native English language speakers as 6 (Goldsmith, 2003).

To control for differences in family structure, we follow the work of Glick and White (2003) and create five dummy variables: 1) respondent lived with both biological parents, 2) respondent lived with one biological parent and that parent's partner, 3) respondent lived with a single mother, 4) respondent lived with a single father, and 5) respondent lived with neither parent (typically lives with grandparents or another relative). Because the sample size of single-parent fathers is small and because research suggests that single mothers and fathers face similar time constraints and parental pressures (Amato, 2005), we combine single parent mothers and fathers into one category. Thus, we have four dummy indicators for family structure: biological parent family (reference group), stepparent family, single parent family, and other family.

School Context

To measure the social context in schools, we include information about the student body and school resources. We tease out the unique influence of class, racial, and linguistic composition in schools. First, we include an indicator for the proportion of students on free and reduced lunch in the school as a measure of the school's poverty level (Orfield and Lee, 2005). Second, we include an indicator of the proportion of minority students in the school to assess the influence of racial composition. Lastly, we account for the proportion of students who are limited English proficient (LEP) to measure linguistic composition. Since proportion LEP is highly skewed, we classify proportion LEP into three dummy variables: low (prop. LEP=0), mid (prop. LEP is >0 and ≤0.10), and high (prop. LEP> 0.10). All school indicators are based on the principal survey and supplemented with information from external school-level data (e.g., Common Core Data) provided by NCES.

To measure school quality we use several proxy measures commonly used (Ainsworth, 2002; Ferguson, 1998; Goldsmith, 2003; Krueger, 2003). Since students are found to perform better in schools with a smaller student-teacher ratio (Ferguson, 1998), we create a control for the number of students per teacher. Additionally, given the varying resources associated with school type, we control for whether the student attended a public or private school. Lastly, we control for differences in urbanicity—urban, rural, and suburban—given that school resources and the characteristics of migrants settling in these areas vary (Parrado and Kandel, 2008).

Neighborhood Context

We measure the social context of neighborhoods by including information on the economic and ethnic/racial make-up of the zip code in which the student lived and by assessing the neighborhood's experience with immigrant populations.⁶ To measure the economic well-being within neighborhoods, we include an indicator of the proportion of households living below the poverty level. To measure the influence of ethnic/racial composition, we include an indicator of the proportion of minorities residing in the zip code by subtracting the proportion of non-Latino white from one. To capture the effect of living near other immigrant groups, we include a measure of the proportion of zip code residents that were foreign-born (Pong and Hao, 2007).

Analytical Approach

To understand how the academic performance of children of immigrants compared in 1990 and 2002, we first estimate chi-square tests and t-tests to evaluate proportion and mean differences in academic performance as well as key socio-demographic, family, school and neighborhood characteristics between the 1990 and 2002 cohorts for each generation.

In order to examine the association between cohort and academic performance, we use several regression strategies. First, we estimate OLS regression models and include two-way interactions between two sets of dummy variables: immigrant status (1^st, 2^nd, and 3^rd generation) and cohort (1990 and 2002). These interactions allow us to compare how the academic performance of 1^st generation (and 2^nd and 3^rd generation) differed between 1990 and 2002. A baseline model that includes only the main effects and interaction variables of the two-way interaction terms indicates the total difference in cohort academic performance for each generation. We then subsequently add blocks of variables representing each of the theoretical constructs (i.e. individual, family, school, and neighborhood characteristics) to assess how differences in each of these constructs contribute to the differing cohort academic performance patterns. All models correct for design effects by using sample weights from each cohort and a correction for the clustering of students within their schools (Goyette, 2008).⁷

Second, utilizing the final model from the OLS regressions, we use regression decomposition to assess the share of the cohort academic performance gap that can be explained by each of the demographic, family, school, and neighborhood constructs. We run the regression decompositions separately for 1^st and 2^nd generation youth, since we are most interested in children of immigrants. Often referred to as the Oaxaca-Blinder decomposition, regression decomposition is a common technique used in the wage discrimination literature but has been applied by education and migration scholars as well. The decomposition technique separates academic performance differences into two components: 1) the explained portion which reflects differences in the characteristics of each immigrant generation across cohorts (i.e. mean/level differences), and 2) the unexplained portion which reflects differences in the associations of these characteristics and academic performance for each cohort (i.e. coefficient/return differences; Jann, 2008). To run the analysis one must choose a set of coefficients to use (e.g. 1990 or 2002). Because there is no compelling argument to use one set of coefficients over another, we follow Elder, Goddeeris, and Haider's (2010) suggestion and use common coefficients estimated from a pooled regression of 1990 and 2002 cohorts with the inclusion of a group specific intercept:

$${\stackrel{‒}{y}}^1-{\stackrel{‒}{y}}^0=({\stackrel{‒}{X}}^1-{\stackrel{‒}{X}}^0){\widehat{\beta }}^p+{\stackrel{‒}{X}}^1({\widehat{\beta }}^1-{\widehat{\beta }}^p)+{\stackrel{‒}{X}}^0({\widehat{\beta }}^p-{\widehat{\beta }}^0)$$

${\widehat{\beta }}^p$ indicates the coefficient vector produced from the pooled regression. The first term on the right hand side is the explained component and the sum of the next two terms is the unexplained component.

Next, because some of the unexplained portion in the regression decomposition may reflect unobserved consequences of immigrant dispersion during the 1990s, we then use state and school fixed effect models to control for immigrant dispersion to new settlement states and for unobserved heterogeneity in schools. Lastly, we assess how cohort academic performance patterns affect different racial/ethnic groups by running three way interactions between three sets of dummy variables: immigrant status, cohort, and racial/ethnic group.

Changes in the Academic Performance and Educational Resources of Children of Immigrants

Reflecting the demographic shifts of the 1990s, the immigrant population increased from 5.4% to 9.3% for 1^st generation youth (a 71% increase) and from 12.4% to 16.4% for 2^nd generation youth (a 32% increase) between the 1990 and 2002 cohorts (Table 1). As children of immigrants are changing the face of the nation's youth, they appear to be faring worse in the US school system. We find that average test scores in reading and math were lower in 2002 than in 1990 for both 1^st (M_Read02=44.91 vs. M_Read90=48.68; M_Math02=46.28 vs. M_Math90=50.81) and 2^nd (M_Read02=48.42 vs. M_Read90=49.44; M_Math02=48.88 vs. M_Math90=50.23) generation youth, though the reduction in math for 2^nd generation youth is only marginally significant. In comparison, test scores for 3^rd generation youth remained relatively unchanged between 1990 and 2002.

Table 1.

Weighted Characteristics of High School Sophomore in 1990 and 2002 by Immigrant Status

	Full Sample			1st Generation			2nd Generation			3rd Generation
	1990	2002	Diff1	1990	2002	Diff1	1990	2002	Diff1	1990	2002	Diff1
Test Score
Reading	50.52 (9.39)	50.02 (10.00)	*	−48.68 (10.19)	44.91 (10.01)	***	49.44 (10.25)	48.42 (9.92)	†	50.74 (9.82)	50.78 (9.34)
Math	50.75 (9.92)	50.02 (10.00)	**	50.81 (10.62)	46.28 (10.63)	***	50.23 (10.31)	48.88 (10.14)	*	50.81 (9.84)	50.57 (9.32)
Demographics
Immigrant Status
1st generation	0.04	0.07	***	--	--		--	--		--	--
2nd generation	0.11	0.14	***	--	--		--	--		--	--
3rd generation	0.86	0.79	***	--	--		--	--		--	--
Female	0.50	0.50		0.50	0.53		0.50	0.47		0.50	0.50
Age	16.12 (.54)	16.18 (.57)	***	16.29 (.63)	16.28 (.80)		16.09 (.54)	16.11 (.58)		16.12 (.54)	16.18 (.54)	***
Race
White (ref.)	0.74	0.66	***	0.15	0.16		0.37	0.30	***	0.81	0.77	***
Black	0.12	0.14	*	006	0.07		0.05	0.08	**	0.13	0.16	**
Asian	0.04	0.04		0.47	0.23	***	0.15	0.14		0.01	0.01	†
Latino	0.10	0.16	***	0.32	0.54	****	0.43	0.48		0.05	0.07	*
Other race	0.02	0.05	***	0.01	0.03	**	0.02	0.07	***	0.02	0.05	***
Family Characteristics
SES	0.00 (.77)	−0.01 (.72)		−0.24 (.93)	−0.42 (.82)	*	−0.17 (.89)	−0.17 (.78)		0.03 (.75)	0.06 (.68)
Family Stucture
Biological parent family (ref.)	0.69	0.57	***	0.76	0.60	***	0.78	0.59	***	0.67	0.56	***
Stepparent family	0.14	0.17	***	0.09	0.14	**	0.07	0.15	***	0.15	0.17	***
Single parent	0.17	0.24	***	0.15	0.22	**	0.15	0.22	**	0.17	0.24	***
Other family	0.01	0.03	***	0.01	0.04	***	0.00	0.03	***	0.01	0.03	***
English language ability	5.60 (.88)	5.66 (.87)	*	3.92 (.95)	3.83 (1.11)		4.33 (1.03)	5.07 (1.16)	***	5.83 (.59)	5.94 (.39)	***
School Characteristics
Prop. free-reduced lunch	0.19 (.21)	0.22 (.20)	*	0.28 (.27)	0.32 (.24)		0.23 (.25)	0.28 (.22)	*	0.13 (.20)	0.20 (.18)
Prop. minority	0.28 (.31)	0.35 (.31)	***	0.56 (.34)	0.59 (.30)		0.48 (.36)	0.54 (.32)	†	0.24 (.29)	0.29 (.28)	**
Prop. LEP population--low (ref.)	0.32	0.34		0.07	0.09		0.12	0.18	*	0.36	0.39
Prop. LEP population--mid	0.52	0.49		0.39	0.42		0.47	0.44		0.53	0.51
Prop. LEP population--high	0.16	0.17		053	0.49		0.41	0.38		0.12	0.10
Student-teacher ratio	16.29 (4.76)	17.16 (4.59)	***	18.19 (4.49)	19.00 (4.40)	†	17.88 (6.36)	18.68 (4.52)		16.01 (4.47)	16.71 (4.52)	**
Urbanicity
Urban	0.27	0.30		0.45	0.49		0.46	0.41		0.24	0.26
Rural	0.32	0.20	***	0.12	0.07		0.15	0.09		0.35	0.23	***
Suburban	0.41	0.50	**	0.43	0.44		0.39	0.50	*	0.41	0.51	**
Public (vs. private)	0.91	0.92		0.90	0.95	*	0.86	0.93	**	0.91	0.92
Neighborhood Characteristics
Prop. zipcode in poverty	0.13 (.10)	0.12 (.09)	*	0.16 (.13)	0.16 (.10)		0.15 (.11)	0.14 (.10)		0.13 (.10)	0.11 (.08)	***
Prop. zipcode is minority	0.23 (.27)	0.31 (.28)	***	0.47 (.31)	0.53 (.29)		0.40 (.32)	0.48 (.31)	**	0.20 (.24)	0.25 (.25)	***
Prop. zipcode is foreign-born	0.07 (.11)	0.10 (.12)	***	0.23 (.17)	0.25 (.17)		0.19 (.15)	0.21 (.15)		0.05 (.07)	0.07 (.08)	***
N2=	16590	15890		900	1470		2060	2610		13630	11820

¹Chi-square tests for proportions and T-tests for means were used to calculate statistical significance of cohort differences.

^†p<.10

^*p<.05

^**p<.01

^***p<.001

²N's are rounded to the nearest 10 as required by NCES.

The lower academic performance of 1^st and 2^nd generation youth may be explained by compositional changes of the immigrant population between the two cohorts. Many of these compositional changes, however, also affected 3^rd generation youth. For example, all three generations experienced a significant increase in the share of their minority populations. For the 1^st generation, the increase in the minority population was driven by growth in the Latino population, which made up about a third (32%) of the 1^st generation population in 1990 but a majority (54%) in 2002. For the 2^nd generation, growth was driven by an increase in the share of the black and other race populations.

In terms of family characteristics, fewer youth, no matter their immigrant status, lived with both biological parents in 2002 than in 1990. By 2002, almost a quarter of youth from all generations lived in single-parent families, compared to about 15% in 1990. First generation youth in 2002 also lived in families with fewer economic, educational, and occupational resources than their counterparts in 1990. Familial SES levels declined between 1990 and 2002 for 1^st generation youth (M_SES90=−0.24 vs. M_SES02=−0.42) but remained stagnant for 2^nd generation youth and possibly increased slightly (marginally significant with a p-value of .11) for 3^rd generation youth. Lastly, both 2^nd and 3^rd generation youth demonstrated a stronger English language background in 2002 than in 1990 (M_EngAbil02=5.07 vs. M_EngAbil90=4.33 and M_EngAbil02=5.94 vs. M_EngAbil90=5.83, respectively), but there was no change in the English language abilities of 1^st generation youth.⁸

Cohort changes in the characteristics of the schools youth attended and the neighborhoods where they lived largely reflected demographic changes and were most apparent among 2^nd and 3^rd generation youth. Given the national rise in the minority population, 2^nd and 3^rd generation youth in 2002 compared to 1990 attended schools with a larger minority population (2^nd generation: 54% vs. 48%; 3^rd generation: 29% vs. 24%) and lived in neighborhoods with more minorities (2^nd generation: 48% vs. 40%; 3^rd generation: 25% vs. 20%). While no similar trend increase occurred for 1^st generation youth, by 2002 1^st generation youth were still more likely to attend schools and live in neighborhoods with a higher minority concentration than either 2^nd or 3^rd generation youth.

For all generations of youth, our results suggest that overall school quality may have declined between 1990 and 2002. We find that 1^st and 3^rd generation youth attended schools with a higher teacher-student ratio in 2002 than in 1990, and 2^nd generation youth attended schools with a larger share of free and reduced lunch students. Moreover, private school enrollment declined among 1^st and 2^nd generation youth between 1990 and 2002.

Do Demographic Changes and Changes in Family, School, and Neighborhood Resources Explain the Decline in Student Academic Performance among Children of Immigrants?

We use multiple regression and two-way interactions to assess the extent to which differences in demographic, family, school and neighborhood characteristics account for observed differences in reading (Table 2) and math (Table 3) test scores between the 1990 and 2002 cohorts across the generations. To interpret the substantive meaning and statistical significance of the interaction terms, we follow Brambor, Clark, and Golder's (2005) guidelines on multiplicative interaction models. We calculate the marginal effect of X (cohort) on Y (test score) by adding the main effect of X and the interactive effect between X and Z (generational status). We report these marginal effect estimates in Table 4.

Table 2.

OLS Regression of Reading Test Scores for High School Sophomores by Immigrant Status, 1990 and 2002 (Data Weighted)

	Model 1 Baseline			Model 2 Demog.			Model 3 Family			Model 4 School			Model 5 Full
	b	(SE)		b	(SE)		b	(SE)		b	(SE)		b	(SE)
Main Effects
2002	0.00	(.25)		0.71	(.22)	**	0.32	(.19)	†	0.36	(.19)	*	0.39	(.19)	*
1st generation	−2.35	(.65)	***	0.03	(.63)		2.16	(.55)	***	2.22	(.55)	***	2.06	(.55)	***
2nd generation	−1.22	(.51)	*	0.77	(46)	†	1.99	(.39)	***	1.86	(.39)	***	1.74	(.40)	***
Two-Way Interaction
1st generation*2002	−3.54	(.77)	***	−3.10	(.70)	***	−2.39	(.59)	***	−2.42	(.59)	***	−2.42	(.58)	***
2nd generation*2002	−1.09	(.63)	†	−1.09	(.52)	*	−1.66	(.45)	***	−1.46	(.44)	**	−1.46	(.44)	***
Demographics
Female				1.25	(.14)	***	1.52	(.13)	***	1.51	(.13)	***	1.51	(.13)	***
Age				−3.71	(.14)	***	−2.76	(.13)	***	−2.73	(.13)	***	−2.72	(.13)	***
Black				−6.49	(.26)	***	−4.62	(.25)	***	−4.29	(.24)	***	−4.12	(.25)	***
Asian				−1.15	(.41)	**	−0.41	(.36)		−0.36	(−36)		−0.32	(.36)
Latino				−6.14	(.30)	***	−2.71	(.27)	***	−2.22	(.27)	***	−2.20	(.28)	***
Other race				−4.26	(.43)	***	−2.81	(.37)	***	−2.58	(.36)	***	−2.51	(.36)	***
SES							4.12	(.10)	***	3.83	(.11)	***	3.84	(.11)	***
Stepparent family							−0.90	(.21)	***	−0 84	(.21)	***	−0 84	(20)	***
Single parent							−0.63	(.19)	**	−0.60	(.18)	**	−0.61	(.18)	**
Other family							−1.85	(.45)	***	−1.78	(.45)	***	−1.78	(.45)	***
English language ability							1.21	(.12)	***	1.11	(.12)	***	1.12	(.12)	***
Prop. free-reduced lunch										−3.01	(.57)	***	−3.24	(.63)	***
Prop. minority										0.30	(.45)		0.64	(.57)
Prop. LEP population--mid										−0.13	(.20)		−0.14	(.20)
Prop. LEP population--high										−0.83	(.34)	***	−1.01	(.36)	**
Student-teacher ratio										−0.04	(.02)	*	−0.04	(.02)	*
Rural										−0.41	(.27)		−0.40	(.27)
Suburban										−0.54	(.24)	*	−0.52	(.24)	*
Public										−0.97	(.31)	*	−0.95	(.32)	**
Prop. zipcode in poverty													1.88	(1.37)
Prop. zipcode is minority													−1.28	(.69)	†
Prop. zipcode is foreign-born													2.60	(1.11)	*
Constant	50.77	(.16)	***	111.22	(2.23)	***	88.37	(2.22)	***	91.04	(2.20)	***	90.65	(2.21)	***

N=32480 (rounded to the nearest 10 as required by NCES)

^†p<.10

^*p<.05

^**p<.01

^***p<.001

Table 3.

OTS Regression of Math Test Scores for High School Sophomores by Immigrant Status. 1990 and 2002 (Data Weighted)

	Model 1 Baseline			Model 2 Demog.			Model 3 Family			Model 4 School			Model 5 Full
	b	(SE)		b	(SE)		b	(SE)		b	(SE)		b	(SE)
Main Effects
2002	−0.27	(.27)		0.55	(.22)	*	0.24	(.13)		0.33	(.13)	†	0.33	(.19)	†
1st generation	−0.45	(.70)		1.13	(.66)	†	2.56	(.56)	***	2.55	(.59)	***	2.35	(.59)	***
2nd generation	−0.44	(.56)		1.43	(.52)	**	2.06	(.48)	***	1.93	(.48)	***	1.79	(.48)	**
Two-Way Interaction
1st generation*2002	−3.85	(.85)	***	−2.78	(.74)	***	−2.14	(.65)	**	−2.13	(.67)	**	−2.14	(.67)	**
2nd generation*2002	−1.23	(.68)	†	−1.32	(.54)	*	−1.53	(.52)	**	−1.33	(.52)	*	−1.34	(.52)	*
Demographics
Female				−1.25	(.14)	***	−0.97	(.13)	***	−0.97	(.13)	***	−0.97	(.13)	***
Age				−4.25	(.13)	***	−3.29	(.12)	***	−3.24	(.12)	***	−3.23	(.12)	***
Black				−7.63	(.25)	***	−5.74	(.24)	***	−4.95	(.24)	***	−4.77	(.26)	***
Asian				1.02	(.47)	*	1.39	(.42)	**	1.61	(.41)	***	1.65	(.41)	***
Latino				−6.84	(.28)	***	−3.74	(.26)	***	−3.03	(.28)	***	−3.01	(.28)	***
Other race				−4.86	(.42)	***	−3.47	(.37)	***	−3.10	(.36)	***	−3.01	(.36)	***
SES							4.25	(.12)	***	3.95	(.12)	***	3.95	(.12)	***
Stepparent family							−1.28	(.20)	***	−1.25	(.20)	***	−1.24	(.20)	***
Single parent							−0.71	(.19)	***	−0.66	(.19)	**	−0.66	(.19)	**
Other family							−2.41	(.45)	***	−2.35	(.49)	***	−2.35	(.49)	***
English language ability							0.70	(.11)	***	0.58	(.11)	***	0.59	(.11)	***
Prop. free-reduced lunch										−3.88	(.58)		−3.88	(.66)
Prop. minority										−0.25	(.46)		−0.07	(.53)
Prop. LEP population--mid										0.23	(.20)		0.20	(.20)
Prop. LEP population--high										0.07	(.32)		−0.16	(.34)
Student-teacher ratio										−0.07	(.02)	**	−0.07	(.02)	***
Rural										0.01	(.28)		0.02	(.28)
Suburban										−0.10	(.24)		−0.10	(.24)
Public										−0.42	(.30)		−0.39	(.32)
Prop. zipcode in poverty													0.87	(1.37)
Prop. zipcode is minority													−1.06	(.69)
Prop. zipcode is foreign-born													2.94	(1.10)	**
Constant	50.84	(.18)	***	121.32	(2.04)	***	101.44	(2.07)	***	103.48	(2.11)	***	103.09	(2.12)	***

N=32480 (rounded to the nearest 10 as required by NCES)

^†p<.10

^*p<.05

^**p<.01

^***p<.001

T4.

Total Marginal Effect of Cohort on Reading and Math Test Scores by Immigrant Generation Based on Corresponding Models in Table 2 and 3 (Data Weighted)

	Model 1 Baseline			Model 2 Demog.			Model 3 Family			Model 4 School			Model 5 Full
	ME	(SE)	Diff1	ME	(SE)	Diff	ME	(SE)	Diff	ME	(SE)	Dif	ME	(SE)	Diff
	A. Reading
Cohort Difference: 2002 vs. 1990
1st generation	−3.54	(.76)	***	−2.39	(.68)	**	−2.07	(.58)	***	−2.06	(.57)	***	−2.04	(.56)	***
2nd generation	−1.08	(.64)	*	−0.37	(.52)		−1.34	(.45)	**	−1.11	(.44)	*	−1.08	(.44)	*
3rd generation	0.00	(.25)		0.71	(.22)	**	0.32	(.19)	†	0.36	(.19)	*	0.39	(.19)	*
	A. Math
Cohort Difference: 2002 vs. 1990
1st generation	−4.13	(.86)	***	−2.23	(.73)	**	−1.90	(.64)	**	−1.81	(.66)	**	−1.81	(.66)	**
2nd generation	−1.50	(.68)	*	−0.77	(.53)		−1.29	(.50)	*	−1.00	(.50)	*	−1.00	(.49)	*
3rd generation	−0.27	(.27)		0.55	(.22)	*	0.24	(.19)		0.33	(.18)	†	0.33	(.19)	†

N=32480 in reading (rounded to the nearest 10 as required by NCES)

¹Indicates marginal effect in 2002 is statistically different from the marginal effect in 1990

^†p<10

^*p<.05

^**p<.01

^***p<.001

Note: Models include the same control variables as the corresponding models in Table 2 and 3.

Replicating the descriptive results, the baseline models in reading and math (Table 4; Model 1) indicate that reading and math test scores were lower in 2002 than in 1990 for both 1^st (ME_Read=−3.54, p<.001; ME_Math=−4.13, p<.001) and 2^nd (ME_Read=−1.08, p<.05; ME_Math=−1.50, p<.05) generation youth and no change for 3^rd generation youth. Demographic changes (Model 2) across cohorts, particularly the rise in the minority population, account for the entire decline in reading and math performance among 2^nd generation youth and some of the decline among 1^st generation youth. Once we control for the fact that black and other race youth (both of whom have lower test scores than white youth) made up a larger share of the 2^nd generation population in 2002 than in 1990, the marginal effect of cohort for 2^nd generation youth decreases and is no longer significant in reading and math. Similarly we find that demographic characteristics (particularly the rise in the Latino population) accounts for about 32% [(3.54-2.39)/3.54*100] and 46% [(4.13-2.23)/4.13*100] of the respective decline in reading and math among 1^st generation youth.

Conversely, we find that the rise in the minority population masked test score gains among the 3^rd generation. Once we account for racial/ethnic composition differences across cohorts, test scores for the 3^rd generation youth are slightly higher in 2002 than in 1990 (ME_Read=0.71, p<.01; ME_Math=.55, p<.05). This test score advantage decreases but remains relatively robust once we control for cohort differences in family, school, and neighborhood characteristics (Models 3-5).

For 1^st and 2^nd generation youth we are unable to explain the remaining academic disadvantage (which reappears for 2^nd generation youth) across cohorts. We find that changes in family and to lesser extent school characteristics contribute to the test score gap across cohorts but do not fully explain the gap. For 1^st generation youth, the greater familial hardships (e.g., lower SES and more single parent families) observed in 2002 than in 1990 partially contributed to the cohort test score gap as seen by the decrease in the marginal effect (Model 2 to Model 3) from −2.39 to −2.07 (13%) in reading and −2.23 to −1.90 (15%) in math once we add family characteristics to the model. This cohort test score gap, however, remains relatively unchanged with the addition of school (Model 4) and neighborhood (Model 5) characteristics.

For 2^nd generation youth, we find that the cohort test score decline reappears once we control for the stronger English language background observed in 2002 than in 1990. The marginal effect of cohort for 2^nd generation youth is negative and significant (ME_Read=−1.34, p<.01; ME_Math=−1.29, p<.05) in the family model. Part of this decline in academic performance reflects changes in the characteristics of the schools 2^nd generation youth attended (particularly the increase in the free/reduced lunch composition and attendance at public schools) across the cohorts. Once we control for school characteristics, the cohort test score gap declines by 21% [(1.34-1.11)/1.34*100] in reading and 29% [(1.29-1.00)/1.29*100] in math, but remains sizeable (about a tenth of a standard deviation) and robust to neighborhood controls.

Decomposing the Relative Role of Family, School, and Neighborhood Context

Now that we have examined which mediating factors contribute to the overall difference in test scores across cohorts, we use regression decomposition to identify the share of the gap explained by each of these factors. For this assessment, we focus on 1^st and 2^nd generation youth, and provide the total share of the difference explained by each of the main constructs as well as the specific measures within these constructs.

For 1^st generation youth, the results indicate that differences in demographics and family resources explain the largest share of the cohort academic performance gap, but that the relative importance of each of these constructs differs for reading and math (Table 5). In reading, changes in family characteristics account for the largest share of the decline in test scores between 1990 and 2002, while changes in demographic characteristics account for the largest share of the decline in math. If 1^st generation youth in 2002 had the same family resources as those in 1990 their test scores would have been 28.71% higher in reading and 18.12% higher in math. Similarly, if they had the same demographic characteristics in 2002 as in 1990 their test scores would have been 10.74% higher in reading and 33.17% higher in math.

Table 5.

Regression Decomposition Showing the Contributions of Demographic, Family, School, and Neighborhood Chacterstics to the 2002-1990 Test Score Performance Gap for 1st and 2nd Generation Youth

	1st Generation				2nd Generation
	Reading		Math		Reading		Math
	Components of Change	% of Total Gap Explained	Components of Change	% of Total Gap Explained	Components of Change	% of Total Gap Explained	Components of Change	% of Total Gap Explained
Demographic Measures Total	−0.44* (.22)	10.72	−1.49*** (.33)	33.17	−.51** (.18)	12.35	−.45* (.20)	10.00
Gender	.01 (.02)	−0.20	−0.01 (.04)	0.30	−.08† (.04)	1.87	0.04 (.03)	−0.90
Age	−.04 (.11)	1.07	−0.04 (.10)	0.88	−0.13 (.09)	3.11	−0.15 (.10)	3.32
Race/Ethnicity	−.40* (.18)	9.86	−1.44*** (.32)	32.00	−.30* (.14)	7.36	−.34* (.16)	7.58
Family Measures Total	−1.18** (.34)	28.71	−.81** (.27)	18.12	−0.06 (.26)	1.46	−.53* (.26)	11.91
SES	−.60* (.23)	14.68	−.54* (.21)	12.02	−0.22 (.20)	5.25	−0.23 (.21)	5.16
Family structure	−.24* (.11)	5.90	−0.17 (.12)	3.68	−.22* (.09)	5.41	−.34*** (.09)	7.56
English language ability	−.33* (.16)	8.13	−0.11 (.07)	2.41	.38** (.12)	−9.20	0.04 (.11)	−0.81
School Measures Total	−0.29 (.25)	7.14	−0.35 (.28)	7.85	−.38* (.19)	9.24	−.50* (.22)	11.09
Prop. free and reduced lunch	−0.22 (.17)	5.42	−0.38 (.25)	8.56	−.21† (.12)	5.08	−0.17 (.11)	3.80
Prop. minority	0.03 (.09)	−0.72	0.00 (.08)	−0.11	0.05 (.10)	−1.21	−0.09 (.12)	1.92
Prop. LEP	0.06 (.10)	−1.38	0.01 (.03)	−0.21	0.03 (.08)	−0.63	0.00 (.07)	−0.03
Student-teacher ratio	−0.15 (.10)	3.74	−0.12 (.09)	2.57	0.01 (.06)	−0.18	−0 07 (.07)	1.66
Urbanicitv	0.00 (.06)	−0.09	0.04 (.06)	−0.93	−.13† (.08)	3.16	−0.09 (.07)	2.04
Public	−0.01 (.07)	0.17	0.09 (.08)	−2.03	−.12† (.07)	3.02	−0.08 (.06)	1.71
Neighborhood Measures Total	0.02 (.14)	−0.44	0.00 (.17)	−0.08	−0.16 (.16)	3.98	−0.01 (.16)	0.12
Prop. zipcode in poverty	0.02 (.07)	−0.46	0.03 (.10)	−0.70	−0.01 (.05)	0.30	−0.01 (.02)	0.12
Prop. zipcode is minority	−0.09 (.15)	2.09	−0.16 (.16)	3.66	−0.21 (.17)	5.19	−0 07 (.16)	1.46
Prop. zipcode is foreign-born	0.09 (.09)	−2.07	0.14 (.13)	−3.04	0.06 (.06)	−1.52	0.07 (.07)	−1.46
Total Explained	−1.89** (.57)	46.04	−2.65*** (.60)	59.06	−1.11* (.48)	64.42	−1.49** (.52)	75.34
Unexplained	−2.21*** (.58)	53.84	−1.84** (.60)	40.94	−0.61 (.43)	35.58	−0.49 (.42)	24.66
Total Change	−4.11*** (.78)		−4.49*** (.78)		−1.72*** (.60)		−1.98** (.62)

N=3450 (rounded to the nearest 10 as required by NCES)

Notes: Regression decompositions are based on pooled coefficients from a pooled regression (including a group-specific intercept) from both new and established states.

^†p<.10

^*p<.05

^**p<.01

^***p<.001

For 2^nd generation youth, there is no dominant compositional change that explains the decrease in test scores from 1990 to 2002. Overall, changes in the demographic composition between cohorts accounts for about a tenth of the decline in both reading and math. In math, however, changes in family and school composition each account for a similar share of the decline (about 10%) as well. For reading, changes in family composition between 1990 and 2002 have been both beneficial and detrimental. For example, if 2^nd generation youth in 2002 had the same family structure as their peers in 1990, their test scores would have been 5.4% higher in reading, but if they had the same English language background, their test scores would have been 9.2% lower. Consequently, the overall net effect of family composition changes across the cohorts is small.

For both 1^st and 2^nd generation youth, a relatively large share of the test score gap across cohorts remains unexplained. Changes in mean characteristics explain about half (46% in reading and 59% in math) of total decline in academic performance for 1^st generation youth and between two-thirds and three-quarters (64% in reading and 75% in math) of the decline for 2^nd generation youth. The remainder of the academic performance gap (ranging from 54% to 11%) is unexplained and may reflect changes in the rates of return (i.e. coefficient effects) of these different resources.

Does Geographic Dispersion Across States and Schools Explain the Decline in Achievement among Children of Immigrants?

The unexplained portion of the test score gap may also reflect unobserved differences in immigrant settlement patterns across cohorts. As noted, children of immigrants lived in different states in 2002 than in 1990 and the schools in newer areas of immigrant settlement were challenged with educating children of immigrants for the first time. To assess how this dispersion affected children of immigrants we add state and school fixed effects to the final model (Table 2 and 3, Model 5). The state fixed effect model controls for stagnant state-level resources and policies associated with immigrant education, while the school fixed effect model controls for unobserved resources at the school level and any district or state-level resources or policies affecting the school. Because school fixed effects are perfectly collinear with year (each school was observed once either in 1990 or 2002) we cannot estimate year, but we can interact year with variables that change within schools (e.g., immigrant status; Wooldridge, 2003). Thus, while we cannot determine how overall academic performance differed in 2002 and 1990 (i.e. the main effect of year), we can determine whether academic performance for 1^st and 2^nd generation youth differed in 2002 and 1990 (i.e. the interactive effect). In other words, we can compare the partial effects of the interaction.

The results (Table 6) indicate that dispersion to new growth states and schools explains only a small portion of the remaining decrease in academic performance observed among children of immigrants. For example, in reading the negative effect of the interaction term between cohort and 1^st generation immigrant from Model 5 in Table 2 (b=−2.42, p<.001) remains significant and relatively unchanged with the addition of state fixed effects (Table 6, Model 6; b=−2.46, p<.001). With the addition of school fixed effects, however, the coefficient on this interaction term decreases slightly to −1.94 but remains significant (Table 6, Model 7). Similar results are found in math and for 2^nd generation immigrants.

Table 6.

State and School Fixed Effect Models of Reading and Math Test Scores for 10th Grade Children of Immigrants and Children of Natives, 1990 and 2002 (Data Weighted)

	Reading						Math
	Model 6 State FE			Model 7 School FE			Model 6 State FE			Model 7 School FE
	b	(SE)		b	(SE)		b	(SE)		b	(SE)
Main Effects
2002	0.33	(.19)	†	--	--		0.42	(.18)	*	--	--
1st generation	2.04	(.54)	***	1.62	(.54)	**	2.37	(.60)	***	2.28	(.62)	***
2nd generation	1.67	(.37)	***	1.33	(.37)	***	1.79	(.47)	**	1.60	(.47)	**
Two-Way Interaction
1st generation*2002	−2.46	(.58)	***	−1.94	(.62)	**	−2.22	(.66)	**	−2.08	(.71)	**
2nd generation*2002	−1.37	(.42)	**	−1.02	(.44)	*	−1.36	(.51)	*	−1.25	(.55)	*
Demographics
Female	1.48	(.13)	***	1.36	(.13)	***	−0.98	(.13)	***	−1.12	(.13)	***
Age	−2.79	(.13)	***	−2.90	(.13)	***	−3.31	(.12)	***	−3.41	(.12)	***
Black	−4.16	(.25)	***	−4.26	(.25)	***	−4.80	(.25)	***	−4.95	(.26)	***
Asian	0.06	(.37)		−0.11	(.36)		1.98	(.41)	***	1.59	(.39)	***
Latino	−2.20	(.28)	***	−2.30	(.28)	***	−3.11	(.29)	***	−3.16	(.30)	***
Other race	−2.42	(.36)	***	−2.29	(.37)	***	−2.89	(.37)	***	−2.54	(.40)	***
SES	3.79	(.11)	***	3.33	(.11)	***	3.92	(.12)	***	3.33	(.11)	***
Stepparent family	−0.82	(.20)	***	−0.70	(.19)	***	−1.22	(.20)	***	−1.04	(.19)	***
Single parent	−0.60	(.18)	**	−0.52	(.17)	**	−0.64	(.19)	**	−0.57	(.17)	**
Other family	−1.80	(.45)	***	−1.51	(.47)	**	−2.32	(.48)	***	−1.85	(.50)	***
English language ability	1.14	(.11)	***	1.13	(.12)	***	0.60	(.11)	***	0.58	(.11)	***
Prop. free/reduced lunch	−3.27	(.63)	***	--	--		−3.88	(.65)	***	--	--
Prop. minority	0.89	(.57)		--			0.22	(.54)		--	--
Prop. LEP population--mid	−0.29	(.20)		--	--		0.09	(.20)		--	--
Prop. LEP population--high	−0.98	(.35)	**	--	--		−0.17	(.33)		--	--
Student-teacher ratio	0.01	(.02)		--	--		−0.03	(.02)		--	--
Rural	−0.27	(.27)		--	--		0.11	(.27)		--	--
Suburban	−0.42	(.24)	†	--	--		0.01	(.22)		--	--
Public	−1.09	(.33)	**	--	--		−0.59	(.32)	†	--	--
Prop. zipcode in poverty	0.49	(1.38)		−0.84	(1.76)		0.36	(1.29)		1.15	(1.66)
Prop. zipcode is minority	−0.97	(.69)		−0.30	(.84)		−1.15	(.69)	†	−1.41	(.82)	†
Prop. zipcode is foreign-born	3.05	(1.12)	**	−0.75	(1.78)		2.93	(1.15)	†	−1.18	(1.57)
Constant	90.12	(2.35)		91.37	(2.19)		103.38	(2.17)		104.39	(2.11)

N=32480 (rounded to the nearest 10 as required by NCES)

^†p<. 10

^*p<.05

^**p<.01

^***p<.001

Is there Variation in Academic Performance Trends for Children of Immigrants across Ethnic/Racial Groups?

Lastly, we assess whether academic performance trends among immigrant generations are heterogeneous across race/ethnicity by including three-way interactions between cohort, immigrant status, and race to the OLS regression models (Models 1-5). Because the cell sizes for some of the ethnic/racial groups were small we combined 1^st and 2^nd generation youth into one category for black and white youth. Thus, the three-way interaction for these groups compare 1^st and 2^nd generation youth (i.e. children of immigrants) to 3^rd generation youth (i.e. children of natives). For the other race category we only included the main effect and did not examine interactive effects. For ease of interpretation, we present the marginal effects (not coefficients) using an adjusted version of the multiplicative interaction term equation noted above.

The results (Table 7) provide some evidence that the decline in reading and math test scores is driven more by race/ethnicity than generational status. Both Latino and black youth, no matter their generational status, had lower reading and math scores in 2002 than in 1990 (though the decline for Blacks is often only marginally significant). For Asian and white youth, the test scores of 1^st and 2^nd generation youth remained stagnant over this time period, while test scores for 3^rd generation youth increased slightly. Thus, Asian and white children of immigrants experienced a relative (compared to children of natives) but not absolute decline in academic performance.

Table 7.

Total Marginal Effect of Cohort on Reading and Math Test Scores for High School Sophomores of Each Ethnic/Racial Group by Immigrant Status (Data Weighted)

	Model 1 Baseline			Model 2 Demog.			Model 3 Family			Model 4 School			Model 5 Full
	ME	(SE)	Diff1	ME	(SE)	Diff	ME	(SE)	Diff	ME	(SE)	Diff	ME	(SE)	Diff
	A. Reading
Latino
1st gen. 2002 vs. 1990	−3.53	(.90)	***	−3.81	(.95)	***	−3.84	(.89)	***	−3.89	(.87)	***	−3.82	(.87)	***
2nd gen. 2002 vs. 1990	−0 90	(.65)		−0.32	(.67)		−2.43	(.62)	***	−2.22	(.61)	***	−2.16	(.60)	***
3rd gen. 2 002 vs. 1990	−0.48	(.68)		−0.47	(.68)		−2.24	(.64)	**	−2.30	(.63)	***	−2.31	(.64)	***
Asian
1st gen. 2002 vs. 1990	−0.96	(1.09)		−1.08	(.98)		−0.52	(.87)		−0.56	(.85)		−0.60	(.84)
2nd gen. 2002 vs. 1990	−1.23	(1.09)		−1.25	(1.05)		−1.07	(.88)		−0.73	(.86)		−0.80	(.86)
3rd gen. 2002 vs. 1990	3.08	(1.63)	†	3.17	(1.54)	*	3.43	(1.41)	*	3.56	(1.44)	*	3.49	(1.44)	*
Black2
1st & 2nd gen. 2002 vs. 1990	−4.55	(2.26)	*	−4.14	(2.05)	*	−4.81	(1.69)	**	−4.27	(1.69)	*	−4.20	(1.69)	*
3rd gen. 2002 vs. 1990	−0.88	(.55)		−0.79	(.52)		−1.08	(.48)	*	−0.83	(.48)	†	−0.83	(.48)	†
Whine2
1st & 2nd gen. 2002 vs. 1990	−0.82	(2.51)		−1.01	(2.37)		−1.66	(1.79)		−1.45	(1.79)		−1.52	(1.79)
3rd gen. 2002 vs. 1990	0.82	(.24)	**	1.03	(.24)	***	0.70	(.20)	***	0.70	(.20)	**	0.72	(.20)	***
	A. Math
Latino
1st gen. 2002 vs. 1990	−3.32	(.91)	***	−3.41	(.92)	***	−3.42	(.91)	***	−3.38	(.90)	***	−3.33	(.90)	***
2nd gen. 2002 vs. 1990	−1.38	(.69)	*	−1.44	(.65)	*	−2.48	(.63)	***	−2.19	(.63)	**	−2.17	(.62)	**
3rd gen. 2002 vs. 1990	−0.01	(.68)		0.06	(.67)		−1.13	(.69)		−1.12	(.68)		−1.17	(.67)	†
Asian
1st gen. 2002 vs. 1990	−0.41	(1.15)		−0.60	(1.12)		0.08	(1.05)		0.10	(1.07)		0.05	(1.06)
2nd gen. 2002 vs. 1990	−1.51	(1.33)		−1.40	(1.31)		−0.81	(1.12)		−0.39	(1.11)		−0.49	(1.10)
3rd gen. 2002 vs. 1990	3.07	(2.07)		2.98	(1.94)		3.27	(1.86)	†	3.45	(1.84)	†	3.34	(1.85)	†
Black2
1st & 2nd gen. 2002 vs. 1990	−3.13	(2.30)		−3.11	(2.07)		−3.60	(1.77)	*	−3.07	(1.77)	†	−3.03	(1.78)	†
3rd gen. 2002 vs. 1990	−0.80	(.49)		−0.76	(.47)		−1.02	(.44)	*	−0.71	(.43)		−0.75	(.44)	†
White2
1st & 2nd gen. 2002 vs. 1990	−1.62	(2.47)		−2.03	(2.38)		−2.34	(1.85)		−1.95	(1.84)		−2.05	(1.83)
3rd gen. 2002 vs. 1990	0.54	(.25)	*	0.78	(.25)	**	0.50	(.20)	*	0.55	(.20)	**	0.56	(.20)	**

N=32480 (rounded to the nearest 10 as required by NCES)

¹Indicates marginal effect in 2002 is statistically different from the marginal effect in 1990

^†p<.10

^*p<.05

^**p<.01

^***p<.001

²Due to smaller sample sizes, 1st and 2nd generation white and black youth are combined into one category identifying children of immigrants.

Notes: Models include the same control variables as the corresponding models in Table 3 and 4.

Discussion

The unprecedented flow and dispersion of international migrants during the 1990s creates significant challenges for the US educational system learning to adapt to a changing student body. Given the growth and changes in the character of immigration, the educational experiences and resources of children of immigrants are likely to diverge from that of previous decades. While previous research on high school dropout behavior suggests that children of immigrants, particularly foreign-born teens, may be faring better in US schools (Fry, 2007), our research suggests a less optimistic outlook.

Our study examined how the academic performance of children of immigrants enrolled in US schools has changed since the 1990s. We found that 10^th grade test scores in both reading and math were markedly lower among 1^st and 2^nd generation youth in 1990 than in 2002. This overall decline in academic performance remained relatively robust to changes in demographics, family, school, and neighborhood characteristics. In contrast, test scores (particularly in reading) of 3^rd generation youth increased slightly during this same time period once we adjusted for the rise in the minority population.

So why aren't children of immigrants faring better in US schools? Part of the challenge is that schools are educating a different population in general. We found that the share of the minority population increased among all generations and that academic performance among minority youth lagged behind their white peers, with the exception of Asian youth in math (but not reading). These results align with current research on the minority test score gap (Kao and Thompson, 2003) and highlight the policy significance of closing the minority test score gap, especially in an era when the minority will become the majority.

In addition to key demographic changes, the families of children of immigrants in 2002 had different resources than immigrant families in 1990. Some of these resource changes differed for 1^st and 2^nd generation youth. While the English language skills of 2^nd generation youth were stronger in 2002 than in 1990, this human capital advancement was limited by other familial changes. Children of immigrants are known to benefit from a strong family structure and the support systems associated with living with both biological parents (Landale, Thomas and Van Hook, 2011). This family support system, however, declined during the 1990s for all generations following national trends that witnessed a rise in single-parent families. For 1^st generation youth, who also experienced a decline in familial socio-economic resources, the loss of family supports was even greater. Overall, these familial resources were important predictors of academic performance, and cohort changes in these resources explained a portion (10%-30%) of the decline in reading and math for 1^st and 2^nd generation youth.

Changes in the composition and resources of schools and neighborhoods did not explain the test score decline among 1^st generation youth and had a small effect on 2^nd generation youth (in math). While previous research suggests that minority and immigrant youth are increasingly attending more economically, linguistically, and racially/ethnically isolated schools (Orfield and Lee, 2005), we found that these changes only affected 2^nd generation youth. For all generations, we found evidence that school quality as measured by several proxy measures (student-teacher ratio and public vs. private) declined across the cohorts. Though troubling, these school quality and compositional changes only explained a small portion (about 10%) of the decline in math test scores for 2^nd generation youth. In terms of neighborhood conditions, we found little change in the composition of the neighborhoods where children of immigrants reside.

There is some evidence that the geographic dispersion of immigrants to new destinations, particularly new destination schools, may have hindered the academic performance of children of immigrants during the 1990s. Our results from the school fixed effect models, which controlled for unobserved school heterogeneity, suggest that changes in the schools attended by children of immigrants explained a small portion of the observed decline in academic performance. Thus, while dispersion to new states may not have had a direct effect on the academic performance of children of immigrants (as suggested by the state fixed effect models which did not explain the test score decline), the associated dispersion to new schools may have hindered performance. Though our models cannot determine whether the unobserved school heterogeneity identified is related to immigrant dispersion specifically, current research indicates that schools in new immigrant destinations lack the resources and expertise to support the needs of their first cohorts of children of immigrants (Gozdziak and Martin, 2005; Massey 2008; Perreira, Chapman and Levis-Stein, 2006; Wainer, 2006). Our results likely reflect some of those challenges.

Nonetheless, despite controlling for key changes in demographic, family, school, and neighborhood characteristics, we still find that 1^st and 2^nd generation youth had lower test scores in math and reading in 2002 than in 1990. This decline in test scores, however, varied across ethnic/racial groups and was most notable among Latino and black youth. Test scores of Latino and black youth, no matter their generational status, declined across the cohorts, while test scores of white and Asian immigrant youth remained stagnant. It is important to note, however, that white and Asian immigrant youth experienced a relative decline in test scores because test scores of their 3^rd generation ethnic peers increased during this time period.

These ethnic/racial results suggest that the rise of the unauthorized population and mixed status families does not likely explain the decline (at least not all of the decline) in test scores among children of immigrants. While the decline was most notable among the group most likely to be undocumented or live in a mixed status family (i.e. 1^st and 2^nd generation Latinos), the observed decline for other groups (i.e. 3^rd generation Latinos and black youth) suggests that other factors are at play.

Instead, the ethnic/racial results align with Glick and White's (2003) prior research that emphasizes the relative importance of ethnic/racial background over immigrant generational status. Their study suggests that declines in immigrant student achievement are tied more to ethnic/racial background and the challenges of racial/ethnic inequality than to nativity status and the migration process. Our results re-emphasize the importance of ethnic/racial background and warrant concern that Latino and black youth, the two most ‘racialized’ disadvantaged minority groups (Bean and Stevens, 2003), are becoming even more disadvantaged over time. Our results also raise concerns that immigrant Asian youth, which have been championed as the ‘model minority,’ are losing their relative academic advantage. This decline in advantage may reflect the growing economic diversification of Asian migration streams, but may also reflect changes in the racial/ethnic barriers a more diverse Asian immigrant population encounters (Galindo and Pong, 2011).

Despite these ethnic/racial differences, however, we still find evidence to suggest that academic performance declined (relatively or absolutely) for children of immigrants in general. Declines were greater among 1^st and 2^nd generation black and Latino youth than their 3^rd generation peers, and 1^st and 2^nd generation Asian and white youth experienced a relative decline in performance compared to their 3^rd generation peers. One possible explanation for the decline in performance is that academic performance may be more heterogeneous, given that more children of immigrants who would have dropped out of school in previous cohorts are now staying in school (Fry, 2007). Thus, while there is optimism that children of immigrants are more likely to stay in school (as suggested by other research), schools may now need to develop better support systems to help those youth succeed academically.

The remaining decline in academic performance may also stem from unobserved differences in immigrant selectivity and context of reception. As suggested by Galindo and Pong (2011), the ‘selectivity’ of migrants in 2002 may be lower as migration streams expand and attract a more diverse group of immigrants. Upon arrival, though, research suggests that the broader economic and political context in the US in the 2000s compared to the previous decades may make it more difficult for these children of immigrants to succeed academically. Compared to previous decades, immigrant families in the 2000s must work longer hours, acquire more educational skills, and spend more time in the US in order to prevent their families from falling into poverty (Van Hook, Brown, and Kwenda, 2004). At the same time, they have fewer state and federal safety net support systems to aid them in this effort (Massey and Bartley, 2005). These added familial challenges are likely to limit the extent to which immigrant parents can invest in their child's education.

Though our study has many strengths—the sample is national and the data have more detail on family, school, and neighborhood characteristics than the US Census—the results should be read with some caveats in mind. First, the analysis uses a cross-section of the panel data available in ELS:2002 and NELS:88. Thus, we identify important associations that need to be further evaluated using longitudinal data. Additionally, while we are able to examine broad ethnic/racial differences among children of immigrants the sample sizes were not large enough to examine within ethnic/racial differences. Given that extant research finds significant pan-ethnic variation in student performance for Asians and Latinos (Kao and Thompson, 2003), future research should examine how the growth and dispersion of immigrant youth in the 1990s affected the academic performance of the different subgroups of Asians (e.g., Chinese, Filipino, etc.) and Latinos (e.g., Mexican, Cuban, etc.). Lastly, given that English language skills are strongly related to math and reading performance (Beal, Niall and Cohan, 2010), future research should examine changes in immigrants’ English language acquisition across cohorts. This research should focus on the role of schools given that schools (under No Child Left Behind) are now held directly accountable for immigrant youth's English language acquisition skills.

Overall, our study highlights that US schools are challenged with educating a growing minority immigrant population that faces more familial (i.e. growth in single parent families) and economic (i.e. lower SES) stressors today than in decades past. Moreover, these familial stressors are compounded by the economic and racial isolation of the schools some children of immigrants attend. In order to develop effective educational policies for immigrant youth, policymakers need to recognize that not only has the children of immigrant population grown and dispersed to new areas but that the characteristics of children of immigrants are changing as well. The past policy lessons and experiences from previous decades may be less applicable today because the background characteristics of immigrants as well as the social environment has changed.