Racial/Ethnic Differences in Effects of Welfare Policies on Early School Readiness and Later Achievement

Abstract

This study examined whether the effects of employment-based policies on children’s math and reading achievement differed for African American, Latino and Caucasian children of welfare receiving parents, and if so, why. Two kinds of employment policies were examined: education-first programs with an emphasis on adult education and job training; and work-first programs with an emphasis on immediate employment. With data from two- and five-year follow-ups in four experimental demonstrations in Grand Rapids, Michigan (N = 591) and Riverside County, California (N = 629), there was evidence of small positive effects of the Grand Rapids and Riverside education-first programs on African American and Latino children’s school readiness and math scores. An opposite pattern of effects emerged among Caucasian children. In one of the two sites, we found that Latino parents’ higher levels of goals for pursuing their own education appeared to explain why their children benefited to a greater degree from the program than their Caucasian counterparts.

Welfare reform legislation in 1996 transformed welfare from a federal entitlement to a system focused on work mandates. As welfare policy changed over time, researchers sought to understand the effects of these programs on child development, particularly on child school performance and achievement. The emphasis on school achievement is supported by developmental research showing that the target of welfare policy changes—parents’ employment, earnings, and income—are related to children’s achievement (Duncan, Yeung, Brooks-Gunn, & Smith, 1998; Perry-Jenkins, Repetti, & Crouter, 2000).

Although prior research has examined welfare programs’ effects on children overall and on children of different ages, questions remain about effects on other important subgroups, particularly children with differing racial or ethnic backgrounds. Race/ethnicity is a social category that continues to mark inequalities in levels of school achievement, even after controlling for levels of income, education, earnings, and occupational status that are correlated with race (Jencks & Phillips, 1998). Racial/ethnic inequalities are evident in reading and math skills at school entry and expand into the early primary grades and beyond (Duncan & Magnuson, 2005; Rouse, Brooks-Gunn, & McLanahan, 2005). Thus, from standpoints of both equity and developmental theory, it is important to examine whether welfare policies affect racial/ethnic groups similarly (Raver, Gershoff, & Aber, 2007). Moreover, race/ethnicity and class are confounded in U.S. society. It is therefore important to examine how policies that solely target indicators of class (i.e., family economic and human capital factors) affect racial/ethnic disparities in achievement.

The current study seeks to fill this gap in the literature by examining whether the effects of two types of welfare policies on children’s school achievement differ by race/ethnicity. Our investigation focuses on two key developmental periods, early and middle childhood, and uses large-scale, random-assignment studies with longitudinal follow-up of children of different racial/ethnic groups.

IMPORTANCE OF EARLY AND MIDDLE CHILDHOOD

The children who are the focus in this study were initially recruited in early childhood and then followed into middle childhood. Across these developmental periods, it is important to note differences in sensitivity to changes in the economic environment. Early childhood, the period during which families in this study were initially exposed to different welfare policies, is a sensitive period, when children may be particularly highly influenced by changes in their family economic and educational environments (Duncan et al., 1998; Morris, Duncan, & Clark-Kauffman, 2005; Magnuson, 2007; Shonkoff & Phillips, 2000; Sroufe, 1979). For example, it appears that increases in adult basic education during this period are causally related to cognitive assessments in the first two years of school (Magnuson, 2007), and programs that boost low-income parents’ employment have stronger positive effects when experienced during the transition from the preschool to early primary grades (Morris et al., 2005). Thus, we might expect the most pronounced impacts of policies targeting parent employment, education, and income on children in early childhood.

Standardized Measures of School Achievement

We focus in this study on standardized measures of children’s school achievement for a number of reasons. First, standardized measures are most often cited in research on racial/ethnic differences in achievement. Concern about the achievement gap is driven by differences in scores on these types of measures. Second, a range of school-level policies, such as considering a particular student for special education or specialized school admissions, are based on standardized achievement scores. Finally, cognitive ability, as measured by standardized reading and math achievement test scores in the early primary grades, can have long-lasting effects on children’s trajectories through the school years and beyond (Card, 1999; Krueger, 2003).

Mandatory Employment Policies for Welfare Recipients

Prior to the 1996 welfare reform, there were two main approaches to moving welfare recipients to work: a work-first approach and an education-first approach. Both approaches shared the same overarching goals: to increase employment and reduce welfare dependence. However, they targeted different pathways to achieving these goals. Founded on the premise that quick entry into employment fosters labor-force attachment, mandatory work-first programs required welfare recipients to start working right away, often with only a very short initial period of job search activities. In contrast, mandatory education-first programs were founded on the premise that welfare recipients need to build educational skills to make sustained gains in employment. They required welfare recipients to engage in educational activities, including adult basic education, higher education, and/or job training, before entering the labor force. Both programs were made mandatory through the use of sanctions; families risked losing some or all of their welfare benefit if they failed to meet program requirements.

Mechanisms Linking Welfare Program Approach to Child School Achievement

Children were not direct targets of the policy interventions examined here; their parents’ employment and education were the targets of policy changes. Thus, any effects on child school achievement must be indirect. There are a number of mediating pathways through which a welfare program approach might ultimately affect child school achievement. The first set of mediators is targeted ones, that is, those mediating mechanisms that the programs were explicitly trying to affect. Targeted mediators include employment, earnings, and education. The resulting changes in family economic circumstances and human capital might affect parents’ ability to invest in children’s learning opportunities, leading to better child cognitive outcomes (Becker & Tomes, 1986; Haveman & Wolfe, 1995; Mayer, 1997). Evidence supports this view, finding among families in poverty that parents with more earnings, more income and higher levels of education have children with higher levels of school achievement (Allhusen et al., 2005; Duncan & Brooks-Gunn, 1997; Gennetian, Magnuson, & Morris, 2008; Linver, Brooks-Gunn, & Kohen, 2002; Magnuson, 2007; Morris & Gennetian, 2003).

Although the different welfare program approaches are likely to affect children’s school achievement through the targeted mediators, there are other mediating mechanisms that might have been affected by changes in the targeted mediators. These non-targeted mediators include maternal well-being, cognitive stimulation, and family structure. Psychological theory posits that increases in parents’ employment, earnings, and education can decrease parental depression and stress by fulfilling life goals in the realms of work and schooling (Chase-Lansdale & Pittman, 2002; McLoyd, 1990, 1998; Weisner, Yoshikawa, Lowe, & Carter, 2006). Economic theory suggests that increases in families’ earnings and education will increase the amount of resources available in the home for supporting children’s development (Becker, 1981; Bergstrom, 1997; Coleman, 1988). In particular, those with higher earnings and education will be able to provide more cognitively stimulating resources that will lead, in turn, to improved child school achievement (Bradley, Corwyn, McAdoo, & Garcia Coll, 2001). Finally, there are a number of theoretical reasons to believe that altering welfare-dependent women’s economic and educational circumstances could change family structure, in particular their likelihood of marrying or cohabiting with a boyfriend or partner. Women who increase their employment and education may be exposed to a new pool of potential partners through workplace or school-related social networks, may experience less economic stress in their relationships, and may be seen as better partners (Edin & Kefalas, 2005; Gassman-Pines & Yoshikawa, 2006; Gennetian & Knox, 2003; Gibson-Davis, Edin, & McLanahan, 2005). Family structure has been linked to child school achievement, with children of married parents performing better than children of single mothers (Garfinkel & McLanahan, 1986; McLanahan & Sandefur, 1994).

When different program approaches to mandating work and education were designed and initially tested, predictions about which program approach would produce better long-term outcomes for families and children were theoretically ambiguous. One initial hypothesis was that the work-first approach would move welfare recipients into employment more quickly than the education-first approach and would, therefore, lead to changes in both targeted and non-targeted mediators more quickly. Short-term impacts on children’s school achievement were thought to be larger in a work-first program. A competing hypothesis, however, was that although the education-first approach was likely to increase employment and decrease welfare receipt more slowly than the work-first approach, the education-first approach would ultimately have larger effects on parents’ human capital, potentially leading to the acquisition of better jobs with higher earnings. The resulting higher earnings would, in turn, have larger effects on the non-targeted mediators and would ultimately produce larger medium- and long-term impacts on children’s school achievement.

Evidence from random-assignment experiments that evaluated the two program approaches suggests that both work-first and education-first programs were effective at increasing the activities they mandated (Hamilton et al., 2001). However, these programs’ effects on other family outcomes and on child school achievement were generally small and inconsistent (McGroder, Zaslow, Moore, & LeMenestrel, 2000). Recent findings from other types of experimentally-evaluated welfare and anti-poverty programs have highlighted variation in program impacts for families with differing circumstances (Alderson, Gennetian, Dowsett, Imes, & Huston, 2008; Gassman-Pines, Godfrey, & Yoshikawa, 2009; Morris, Bloom, Kemple, & Hendra, 2003; Yoshikawa, Magnuson, Bos, & Hsueh, 2003). Thus, overall program effects may be masking important subgroup differences across subsets of families. Families’ race/ethnicity is one such subgroup.

Variation by Race/Ethnicity

We have proposed a model whereby children’s school achievement could be affected by their mothers’ participation in a work-first or education-first welfare program via either targeted or non-targeted mediators (see Figure 1 for the conceptual model for the study). How might these different pathways linking program participation to child achievement vary by race/ethnicity? At each stage in this set of proposed mediating pathways, there is potential for variation in outcomes by race.

FIGURE 1.

Conceptual model.

First, welfare recipients of different races may experience the same program approach differently. Welfare caseworkers are the conduit through which messages about program approach and welfare program rules are transmitted to recipients. Welfare caseworkers can interact differently with recipients with different racial/ethnic backgrounds. In interactions with welfare caseworkers, African Americans report higher levels of sanctioning and lower amounts of discretionary, onetime payments, than their Caucasian counterparts, given similar situations, suggesting some evidence of discrimination (Gooden, 1998). These differences in interactions could affect the ways in which African American and Caucasian welfare recipients are encouraged and supported in their transition to employment or their participation in education.

Second, effects of the programs on welfare recipients’ targeted outcomes might also differ by race/ethnicity, for several reasons. For example, experiences of race-related discrimination on the job by African American parents (Hughes & Dodge, 1997; Mercer, Heacock, & Beck, 1993) could lead to greater barriers to employment among this group of families. Similarly, research on employer demand for welfare recipients in four cities suggests that various factors, including location and preferences, disproportionately limit the employment options of African American and Latino welfare recipients, compared to their Caucasian counterparts (Holzer and Stoll, 2000). Thus, whether using the work-first or education-first approach, welfare policies could have weaker effects in increasing employment and earnings among African American and Latino parents, when compared to Caucasian parents. In addition, these employment-based differences may be especially apparent in work-first programs, when compared to education-first programs, because work-first programs aim to increase employment more quickly.

Third, the effects of these programs on welfare recipients’ non-targeted mediators might also differ by race/ethnicity for several reasons. The direct effects of the programs on the non-targeted mediators could simply be moderated by welfare recipient race/ethnicity. But it is also possible that differential program effects on the targeted mediators might, in turn, lead to differential effects on the non-targeted mediators. For example, Latinos tend to report higher levels of depressive symptoms than Caucasians or African Americans (Surgeon General, 2001a; 2001b), therefore, they may experience greater declines in depression as a result of welfare program-induced increases in employment, earnings, or education. Due to links between earnings and cognitive stimulation, any racial/ethnic differences in the effects of the welfare programs on earnings could lead to racial/ethnic differences in levels of cognitive stimulation. Finally, in terms of family structure, program effects on family structure could differ across African American, Latino, and Caucasian families. Studies suggest that norms of marriage and pressures to marry in low-income Latino families may be stronger than in low-income Caucasian or African American families (Oropesa, 1996; Taub, 1991). Actual marriage rates are indeed generally higher among Latino families (Graefe & Lichter, 2002). Thus, mandatory employment policies that increase employment and education might lead to larger increases in marriage among Latino parents than among Caucasian or African American parents. On the other hand, given lower marriage rates among African American mothers in poverty, even after a nonmarital birth, there may be more “room” for an increase in marriage in this group.

These contrasting findings from the relevant literatures suggest some alternative hypotheses. First, the impacts of welfare policies on child school achievement might ultimately be more favorable for Caucasian welfare recipients, compared to African-Americans or Latinos. Compared to welfare recipients of color, Caucasians are expected to experience fewer negative interactions with welfare caseworkers and less discrimination in the job market or workplace. Thus, any positive effects of the work-first or education-first approaches on child school achievement might be larger for Caucasians. The second, alternate hypothesis is that impacts may be more favorable for Latino welfare recipients, compared to Caucasians or African Americans. Latinos might experience more positive effects on two of the non-targeted mediators—maternal well-being and family structure—that could ultimately result in more benefits to the school achievement of children in Latino families. Third, African American families may experience larger increases in marriage, which may produce larger positive impacts on child achievement.

Pre-Existing Differences Between Racial/Ethnic Groups

Although there are many reasons to believe that the effects of work-first and education-first programs on child school achievement might vary by race/ethnicity, it is possible that differences in program effects by race actually reflect pre-existing differences among racial/ethnic groups in other characteristics. In the United States, race/ethnicity is correlated with a number of different characteristics. If differences in program impacts by race/ethnicity are found, they could be due to these differences among racial/ethnic groups that precede exposure to the policies, and not to race/ethnicity itself. First, individuals from different racial/ethnic groups differ in human capital and income levels (Duncan & Magnuson, 2005; Jencks & Phillips, 1998), with African Americans and Latinos typically attaining lower levels of education and income than Caucasians (these differences may be smaller, however, in low-income samples). Second, rates of single-parent families vary substantially among low-income African American, Latino, and Caucasian families (U.S. Census Bureau, 2007). Third, individuals from different racial/ethnic groups might differ in their goals for work and education. Some research, for example, suggests that immigrant parents differ from native-born parents in the intensity of their goals to pursue their own education (Kao & Tienda, 1995). The fit between pre-existing goals to pursue education vs. work and program focus (education-first vs. work-first) might then explain program differences across groups that vary in ethnicity and immigration status (Gassman-Pines et al., 2009; Moos, 1984). Finally, members of racial/ethnic groups differ in their levels of depression, even controlling for class-related factors (Surgeon General, 2001b). Any of these pre-existing factors, rather than race/ethnicity itself, may be more important than targeted or non-targeted mediators in explaining racial/ethnic differences in welfare program impacts. Both sets of explanations are examined in this study (see Figure 1).

The Current Study

This study examines the following research questions and hypotheses:

1. Do mandatory employment policies’ effects on cognitive school readiness and later achievement differ by race/ethnicity? Mandatory employment policies of two kinds are investigated: those using an education-first approach and those using a work-first approach. Two racial/ethnic group comparisons will be made (Latinos to Caucasians and African Americans to Caucasians, in two separate sites).

2. Do differences in pre-existing characteristics (characteristics preceding exposure to the particular policies) across racial/ethnic groups help explain differential effects of mandatory employment policies on standardized achievement? We hypothesize that differential program effects on achievement will be partially accounted for by pre-existing differences between racial/ethnic groups.

3. Do differential effects on targeted and non-targeted mediators, after exposure to these policies, help explain differential effects of the policies on standardized achievement? Mediators include economic and educational behaviors (earnings, participation in educational activities), maternal well-being, cognitive stimulation, and family structure. We hypothesize that differential program effects on achievement will be partially accounted for by differential effects of mediators.

METHODS

Participants

We selected experimental programs from a larger sample of 11 programs that operated in the 1990s and were evaluated using an experimental design (for information on the 11 welfare and employment policy experiments, see Morris et al., 2005). We selected those that met the following criteria: 1) sufficient numbers of Caucasian and either African American or Latino families; and 2) assessment of school readiness in early childhood and standardized school achievement in middle childhood. These criteria were met by four programs in two sites: Grand Rapids, Michigan and Riverside, California in the National Evaluation of Welfare to Work Strategies (NEWWS).

The NEWWS evaluation was a federally mandated study conducted to evaluate aspects of the 1988 Family Support Act (centrally, the Jobs Opportunity and Basic Skills programs, or JOBS). In particular, the study tested the impact of making employment-related activities mandatory. In three sites of the NEWWS evaluation (out of a total of seven sites), extensive data on children’s development were collected. In two of these three sites—Grand Rapids, Michigan, and Riverside County, California—sufficient numbers of multiple racial/ethnic groups permitted examination of differences in effects. In each of these sites, two welfare policy programs were implemented and tested: a work-first program, in which caseworkers emphasized immediate employment, and an education-first program, in which caseworkers emphasized adult basic education prior to employment. Each site operated one work-first program and one education-first program, resulting in four programs in two sites. In both conditions, the respective activities were mandatory; that is, welfare benefits were reduced for not engaging in the activities. We tested differences in welfare policy effects for African Americans and Caucasians in Grand Rapids, and Latinos and Caucasians in Riverside. There were not sufficient numbers of Latinos in Grand Rapids or African Americans in Riverside to calculate experimental effects for these groups. According to the 1990 Census (closest to the time of enrollment), 88% of low-income Latino parents of young children in Riverside County were Mexican (U.S. Census Bureau, 2010). Unfortunately, we do not have data on immigration status of parents or children.

Respondents were eligible for enrollment in the overall NEWWS evaluation if they had applied for or were receiving welfare (Aid to Families with Dependent Children, or AFDC) at the time of enrollment, and if they were not exempt from participation in the JOBS program (i.e., exempt due to being ill or incapacitated, caring for a household member who was ill, or incapacitated, pregnant past the first trimester, having a child younger than age three in Riverside and younger than age one in Grand Rapids). Families in the overall evaluation were chosen for the Child Outcomes Study if they had a child between age 3 and 5 years at enrollment, and if that child was a biological or adoptive child of the mother.

Enrollment occurred in 1992 and 1993. Using three-way random assignment, welfare recipients were assigned to one of three conditions (education-first program, work-first program, or control). The control condition remained under existing AFDC rules governing welfare, which at that time did not mandate employment for mothers whose youngest child was preschool aged. All recipients had children ages 3 to 5 at random assignment. Our data came from the 2-year follow-up, at which point the children were ages 5 to 7, and the 5-year follow-up, at which point they were ages 8 to 10 (for further details, see McGroder et al., 2000; Hamilton et al., 2001).

In the education-first condition in Riverside, state regulations required that only recipients deemed in need of education could be eligible for that condition. “In-need” status consisted of meeting one of the following criteria: not proficient in English, no high school diploma or GED, or scoring below a cutoff on a basic math/reading skills exam. Thus, mothers in that condition (and in the analyses involving that condition here, in the corresponding control group) in Riverside are more disadvantaged, on average, than those in the other conditions in Riverside and Grand Rapids.

Table 1 presents descriptive statistics for the Grand Rapids and Riverside samples at baseline, both for those entire samples and separately for racial/ethnic subgroups. At baseline, our sample in Grand Rapids included 252 African Americans and 339 Caucasians. In Riverside, our sample included 309 Latinos and 320 Caucasians (each of these numbers includes families from all three experimental conditions at that site).

TABLE 1.

Sample Characteristics at Baseline

	Grand Rapids			Riverside
	All	Caucasian	African American	All	Caucasian	Latino
Caucasian	53.0%			40.1%
African American	38.0%			–
Latino	–			39.2%
On AFDC in prior year	28.5%	19.9%	41.0%	31.9%	24.6%	36.4%
High school diploma	60.9%	61.6%	62.6%	43.4%	51.3%	25.1%
Ever married	40.8%	55.4%	19.4%	55.1%	63.0%	56.0%
Youngest child less than 3	32.4%	31.9%	31.5%	–	–	–
Earnings in prior year	$2016.01 (4221.82)	$2175.12 (4701.62)	$2003.69 (3806.49)	$1355.89 (3419.57)	$1595.69 (3892.46)	$1230.08 (3096.75)
Family first scale	2.35 (0.60)	2.47 (0.60)	2.13 (0.52)	2.34 (0.64)	2.40 (0.62)	2.37 (0.65)
Value of school scale	2.82 (0.52)	2.77 (0.50)	2.85 (0.55)	2.93 (0.50)	2.87 (0.50)	3.08 (0.46)

Note: Percentages given for categorical variables; means (standard deviations) given for continuous variables.

Measures

Baseline Covariates

These were measured prior to random assignment and, therefore, prior to exposure to the work-first or education-first policy conditions. We utilized baseline characteristics in four areas: human capital, family structure, goals, and maternal well-being. Human capital measures included whether the mother had a high school diploma or GED, earnings in the year prior to random assignment, and long-term welfare receipt (greater than 5 years). Family structure variables included whether the parent had ever been married (all parents were unmarried at baseline) and whether the family had a child under age 3 (a factor that is associated with lower levels of employment). Goals measures included a 6-item measure of preference for staying at home with children over working (“family first”; sample item “I do not want a job because I would miss my kids too much”), on a 4-point response scale from disagree a lot to agree a lot. A 5-item scale tapping education goals included items such as “I would like to go to school for reading or math,” with the same response categories. Reliabilities (Cronbach’s alpha) for these measures were tested within ethnic group, for each site; all were above .60. Maternal well-being was represented by four items from the Center for Epidemiological Studies-Depression (CESD) scale (alpha reliability was greater than .85 for each ethnic group). Items were scored on a 0 to 3 point scale and include, “I was bothered by things that don’t usually bother me.”

Mediators

Mediators were measured at the 2-year follow-up, that is, 2 years after initial exposure to policy conditions. Four areas of mediating mechanisms were explored: parent economic and educational behaviors, family structure, maternal well-being, and cognitive stimulation. Economic and educational behaviors, calculated across the first 2 years of follow-up, were obtained from both administrative records data and survey data. Measures collected through administrative data included average yearly earnings and average yearly income. Income was defined as the sum of earnings, AFDC, and Food Stamps. Survey data were used to measure months of participation in education activities (GED, adult basic education, high school, or college classes). Family structure measures were collected through the survey and included whether the respondent was married or cohabiting at the 2-year follow up. Maternal well-being was collected through the survey and was represented by a measure of maternal depression (12 items from the 20-item CESD scale; alphas were greater than or equal to .90 for each ethnic group, in each site) (Radloff, 1977). Finally, a 9-item measure of cognitive stimulation was collected through the survey (alphas .54 and .67 for African Americans and Caucasian parents in Grand Rapids, and .63 and .68 for Latino and Caucasian parents in Riverside). This measure was adapted from the cognitive stimulation subscale of the Home Observation for Measurement of the Environment (Caldwell and Bradley, 1984) and taps stimulation of the child at home, as represented by reading to the child and existence of toys, books, and other materials.

Child Achievement

At the 2-year follow up, the Bracken Basic Concept Scale School Readiness Composite was administered (Bracken, 1984). This standardized direct assessment contains 61 items that assess children’s knowledge of colors, letters, numbers, counting, comparisons, and shapes. Scores represent the number of concepts (out of 61) that children answered correctly.

At the 5-year follow up, subscales from the Woodcock-Johnson Revised achievement test was administered (Woodcock and Johnson, 1989). These assessments included children’s Broad Math (Calculation and Applied Problems subtests) and Reading (Passage Comprehension and Letter-Word Identification tests) scores. The Calculation subtest measures the child’s skills in basic arithmetic. The Applied Problems test includes items that require application of mathematics skills, such as telling time and temperature. In the Passage Comprehension test, the child reads a short passage and identifies a missing word. The Letter-Word Identification test consists of two types of items: 1) matching a representation of a word with a picture of an object and 2) identifying isolated letters and words presented in large type.

Analytic Plan

Using OLS techniques, we estimated a series of regression models predicting children’s achievement outcomes. For clarity, the regression models are summarized in Table 2. To answer Question 1, we conducted OLS regressions predicting child achievement from the experiment (program vs. control group), race/ethnicity and a set of background characteristics (see Model 2 in Table 2). The significance of differences in program effects across racial/ethnic groups was tested by including race/ethnicity by program interaction terms as predictors in the models. Separate regressions were run for each of the four programs (education-first and work-first programs in the Grand Rapids and Riverside sites). For comparison purposes, we also present regressions that do not control for background characteristics (see Model 1 in Table 2). To address Question 2, we added interactions of the background variables with the experimental condition as predictors to the model (see Model 3 in Table 2). To address Question 3, we added the six mediator variables from the two-year follow up as predictors (see Model 4 in Table 2). The significance of mediated associations was tested using methods suggested by Mackinnon, Lockwood, Hoffman, West, and Sheets (2002).

TABLE 2.

Overview of Analysis Plan

Regression Model 1	Regression Model 2	Regression Model 3	Regression Model 4
Program	Program	Program	Program
Race/Ethnicity	Race/Ethnicity	Race/Ethnicity	Race/Ethnicity
Program × Race/Ethnicity Interaction	Program × Race/Ethnicity Interaction	Program ×Race/Ethnicity Interaction	Program ×Race/Ethnicity Interaction
	Baseline Covariates	Baseline Covariates	Baseline Covariates
		Program × Baseline Covariate Interactions	Program × Baseline Covariate Interactions
			2-Year Mediators

Note: “Program” represents either employment-first program or education-first program vs. control. Race/ethnicity represents African American vs. Caucasian in Grand Rapids and Latino vs. Caucasian in Riverside. This table presents predictors only. Each set of models was run separately for all three outcome variables.

Each set of OLS regressions was conducted for each of 3 dependent variables: the 2-year Bracken School Readiness score; 5-year Woodcock-Johnson reading achievement; and 5-year Woodcock-Johnson math achievement. All results reported utilize the alpha level of .05, two-tailed. Results between alpha levels of .05 and .10 are referred to as “marginally significant.”

Preliminary Attrition Analyses

All analyses presented pertain to the sample that was followed up at both 2 and 5 years. The overall retention rates at the 5-year follow-up, in relation to the initial samples at random assignment, were 65% for Riverside and 82% for Grand Rapids. In addition, due to a problem in survey administration, the goals measures were not assessed at baseline for the first 23% of the survey sample recruited in the Riverside site. Percentages not followed up did not vary by more than 5 percentage points across experimental conditions (Hamilton et al., 2001). However, in order to ensure that differences in experimental impacts between racial/ethnic groups were not due to differences in attrition patterns over time, we examined whether there were racial/ethnic differences in attrition rates or predictors of attrition rates across experimental groups. Results indicated no differential effects of race/ethnicity, nor interactions of race/ethnicity with the experimental conditions or baseline covariates above chance, on whether or not a respondent was retained. The variables that predicted whether a respondent was retained included the following. In Grand Rapids, older mothers were less likely to have been followed up at five years, and in Riverside, those who were married or who had higher incomes at baseline were more likely to have been followed up at five years.

RESULTS

Descriptive Analyses of Racial/Ethnic Differences in Baseline Characteristics

Before turning to the analyses for Questions 1 and 2, we ran descriptive analyses to examine racial/ethnic group differences in baseline characteristics. The baseline characteristics (high school diploma, earnings, long-term welfare receipt, ever married, child under age 3, “family first” index, education goals, and depressive symptoms) were entered into logistic regressions, along with the experimental condition variables. For each of the two regressions (one for Grand Rapids, the other for Riverside), the race/ethnicity contrast for that site was the dependent variable (African Americans/Caucasians for Grand Rapids; Latinos/Caucasians for Riverside). In Grand Rapids, African Americans were more likely than Caucasians to be long-term welfare recipients (odds ratio (OR) = 3.77, p < .01), less likely to have ever been married (OR = .23, p < .01), less likely to prefer staying at home with children than working (OR = .32, p < .01), and marginally more likely to report wanting to pursue their own education (OR = 1.42, p < .10). In Riverside, Latinos were less likely than Caucasians to have a high school diploma (OR = .32, p < .01), more likely to report wanting to pursue their own education (OR = 2.40, p < .01), and marginally less likely to prefer staying at home with children than working (OR = .71, p < .10). As expected, in neither regression were the program variables significant, indicating that racial/ethnic groups were equally represented across experimental conditions.

Are There Racial/Ethnic Differences in Experimental Effects of Welfare Policies on Standardized Achievement?

Table 3 presents results for the Bracken score at the 2-year follow-up. Each of the four columns represents one of the four programs (education-first and work-first, in Grand Rapids and Riverside). The first three rows in the first panel present the associations of the program-by-ethnicity interaction and the main effects of program and race/ethnicity, without any covariates in the model. The first three rows of the second panel present the same associations, after entering the baseline covariates (the next seven rows of the second panel).

TABLE 3.

Summary of Regression Results, Models 1 and 2, Predicting 2-Year Bracken Scores

	Grand Rapids Program Site						Riverside Program Site
	Education-First Program			Work-First Program			Education-First Program			Work-First Program
	b	β	SE	b	β	SE	b	β	SE	b	β	SE
Model 1: No baseline covariates
Race/ethnicity* Program	1.13	(0.02)	3.08	−2.12	(−0.05)	2.67	11.03	(0.20)	4.58*	−2.94	(−0.05)	4.10
Program	−0.81	(−0.03)	1.49	0.59	(0.03)	1.33	−0.51	(−0.02)	2.29	1.51	(0.05)	2.03
Race/ethnicity	−0.85	(−0.03)	1.53	−2.42	(−0.10)	1.34†	−0.68	(−0.02)	2.28	−5.00	(−0.19)	1.93*
(Caucasian = 0, African American = 1 in Grand Rapids; Caucasian = 0, Latino = 1 in Riveriside)
Model 2: Including baseline covariates
Race/ethnicity×Program	0.60	(0.01)	3.14	−1.86	(−0.04)	2.74	9.64	(0.17)	4.84*	−2.80	(−0.05)	4.15
Program	−1.00	(−0.04)	1.51	0.91	(0.04)	1.35	−2.28	(−0.08)	2.37	0.86	(0.03)	2.06
Race/ethnicity	−0.94	(−0.04)	1.73	−2.24	(−0.10)	1.57	−0.20	(−0.01)	2.43	−4.31	(−0.16)	2.09*
Earnings in prior year	0.06	(0.02)	0.17	0.26	(0.09)	0.18	0.17	(0.02)	0.62	−0.02	(−0.004)	0.27
On AFDC 5+ years	3.08	(0.11)	1.83†	−0.10	(−0.004)	1.60	−2.89	(−0.10)	2.51	−0.98	(−0.04)	2.16
High school diploma	2.62	(0.10)	1.61	1.63	(0.07)	1.41	2.86	(0.08)	2.92	0.73	(0.03)	2.10
Ever married	1.65	(0.06)	1.63	1.58	(0.07)	1.50	0.58	(0.02)	2.48	−2.06	(−0.08)	2.01
Youngest child < 3	−1.31	(−0.05)	1.68	1.71	(0.07)	1.56	–	–	–	–	–	–
Family first goals	−1.10	(−0.05)	1.31	−0.77	(−0.04)	1.28	−3.00	(−0.14)	1.80†	−2.25	(−0.11)	1.70
Education goals	−1.89	(−0.08)	1.44	−0.55	(−0.03)	1.28	−4.08	(−0.13)	2.75	−2.38	(−0.09)	1.98
Depressive symptoms	−0.42	(−0.12)	0.21*	−0.30	(−0.09)	0.19	0.45	(0.12)	0.33	0.57	(0.14)	0.30†
	n = 287			n = 298			n = 141			n = 184

Note:

^†p < .10;

^*p < .05

A significant racial/ethnic difference was found in the effects of the Riverside education-first program on the Bracken (b = 11.03, p < .05). The magnitude and significance of this interaction does not appreciably decrease after including baseline variables in the equation (first row, third column, bottom panel). Figure 2 presents the plot of this interaction. The education-first program produced a small increase in Bracken scores among Latinos (effect size of +.09; left pair of bars) and a large decrease in Bracken scores among Caucasians (effect size of −.56; right pair of bars). Effect sizes were calculated as the difference between the experimental and control groups for that particular program, divided by the standard deviation of the variable in the control group.

FIGURE 2.

Racial/ethnic differences in effects of Riverside education-first program on 2-year bracken scores.

Results for the 5-year Woodcock-Johnson math scores are presented in Table 4. A significant racial/ethnic difference was found in the effects of the Grand Rapids education-first program on math scores (b = 11.54, p < .05) (see first row, first column, top panel of Table 4). The magnitude and significance of this interaction does not appreciably decrease after including baseline variables in the equation (first row, first column, bottom panel). Figure 3 presents the plot of this interaction. The education-first program produced a small increase in Woodcock- Johnson math scores among African Americans (effect size of +.22; left pair of bars) and a moderate decrease in Woodcock Johnson math scores among Caucasians (effect size of −.42; right pair of bars). Results for the 5-year Woodcock-Johnson reading scores are presented in Table 5. No significant racial/ethnic differences in any of the four programs’ effects on this outcome were found. We also conducted analyses that controlled for the 2-year Bracken score when predicting the Woodcock-Johnson math and reading scores. In that set of analyses, the interaction predicting the 5-year Woodcock-Johnson math score in the Grand Rapids education-first program remained significant (b = 9.33; p < .05).

TABLE 4.

Summary of Regression Results, Models 1 and 2, Predicting 5-Year Woodcock-Johnson Math Scores

	Grand Rapids Program Site						Riverside Program Site
	Education-First Program			Work-First Program			Education-First Program			Work-First Program
	b	β	SE	b	β	SE	b	β	SE	b	β	SE
Model 1: No baseline covariates
Race/ethnicity* Program	11.54	(0.15)	4.62*	3.56	(0.05)	4.08	6.2	(0.09)	4.18	0.01	(0.00)	5.62
Program	−1.39	(−0.04)	2.23	1.70	(0.05)	2.03	0.1	(0.00)	2.07	1.14	(0.03)	2.77
Race/ethnicity	−1.66	(−0.04)	2.30	−5.32	(−0.15)	2.04**	−5.1	(−0.14)	2.08*	−4.77	(−0.13)	2.66†
(Caucasian = 0, African American = 1 in Grand Rapids; Caucasian = 0, Latino = 1 in Riveriside)
Model 2: Including baseline covariates
Race/ethnicity* Program	10.80	(0.14)	4.57*	4.98	(0.07)	3.89	6.67	(0.09)	4.22	−0.90	(−0.01)	5.51
Program	−1.72	(−0.05)	2.19	2.45	(0.07)	1.92	−0.07	(0.00)	2.09	0.15	(0.00)	2.72
Race/ethnicity	−1.09	(−0.03)	2.52	−6.38	(−0.18)	2.23**	−5.03	(−0.14)	2.13*	−3.81	(−0.10)	2.80
Earnings in prior year (in thousands)	0.31	(0.08)	0.25	0.39	(0.08)	0.26	0.25	(0.04)	0.37	0.02	(0.00)	0.35
On AFDC 5+ years	1.78	(0.04)	2.66	−4.27	(−0.11)	2.27†	−1.29	(−0.04)	2.22	−3.16	(−0.08)	2.88
High school diploma	6.33	(0.16)	2.34**	6.55	(0.18)	2.01**	0.28	(0.01)	2.58	4.65	(0.13)	2.78†
Ever married	3.81	(0.10)	2.39	2.26	(0.06)	2.14	−1.90	(−0.05)	2.11	−1.62	(−0.04)	2.69
Youngest child < 3	−1.74	(−0.04)	2.44	6.58	(0.17)	2.22**
Family first goals	−0.67	(−0.02)	1.91	−5.82	(−0.19)	1.83**				−5.64	(−0.19)	2.28*
Education goals	−1.33	(−0.04)	2.10	1.21	(0.04)	1.83				−1.85	(−0.05)	2.60
Depressive symptoms	−0.73	(−0.14)	0.31*	−0.71	(−0.15)	0.27**				0.19	(0.04)	0.40
	n = 289			n = 302			n = 282			n = 188

Note:

^†p < .10;

^*p < .05;

^**p < .01

FIGURE 3.

Racial/ethnic differences in effects of Grand Rapids education-first program on 5-year woodcock-johnson math scores.

TABLE 5.

Summary of Regression Results, Models 1 and 2, Predicting 5-Year Woodcock-Johnson Reading Scores

	Grand Rapids Program Site						Riverside Program Site
	Education-First Program			Work-First Program			Education-First Program			Work-First Program
	b	β	SE	b	β	SE	b	β	SE	b	β	SE
Model 1: No baseline covariates
Race/ethnicity* Program	5.93	(0.08)	4.39	2.09	(0.03)	3.99	−1.88	(−0.03)	4.19	3.79	(0.05)	5.62
Program	−0.56	(−0.02)	2.12	3.10	(0.09)	1.98	1.78	(0.05)	2.07	−0.24	(−0.01)	2.77
Race/ethnicity	−3.67	(−0.10)	2.18†	−5.42	(−0.15)	1.99**	−5.81	(−0.16)	2.09**	−6.02	(−0.16)	2.67*
Model 2: Including baseline covariates
Race/ethnicity* Program	5.20	(0.07)	4.40	3.42	(0.05)	3.82	−0.63	(−0.01)	4.17	1.90	(0.02)	5.39
Program	−0.80	(−0.02)	2.11	3.92	(0.11)	1.88*	1.00	(0.03)	2.07	−1.15	(−0.03)	2.67
Race/ethnicity	−2.72	(−0.07)	2.43	−5.45	(−0.16)	2.19*	−5.91	(−0.17)	2.10**	−5.77	(−0.16)	2.74*
Earnings in prior year (in thousands)	0.19	(0.05)	0.24	0.18	(0.04)	0.25	0.36	(0.06)	0.36	0.10	(0.02)	0.34
On AFDC 5+ years	−1.86	(−0.05)	2.56	−7.98	(−0.21)	2.23**	−3.64	(−0.10)	2.20†	−3.64	(−0.09)	2.82
High school diploma	5.24	(0.14)	2.25*	5.62	(0.16)	1.97**	−3.81	(−0.09)	2.55	0.87	(0.02)	2.72
Ever married	3.10	(0.09)	2.30	1.07	(0.03)	2.10	−4.28	(−0.12)	2.08*	−4.28	(−0.11)	2.63
Youngest child < 3	−0.39	(−0.01)	2.35	4.84	(0.13)	2.18*
Family first goals	−0.45	(−0.02)	1.84	−4.36	(−0.14)	1.80*				−7.78	(−0.3)	2.23**
Education goals	−0.95	(−0.03)	2.03	0.96	(0.03)	1.79				−4.24	(−0.12)	2.55†
Depressive symptoms	−0.38	(−0.08)	0.30	−0.44	(−0.09)	0.27				−0.07	(−0.01)	0.39
	n = 289			n = 302			n = 282			n = 188

Note:

^†p < .10;

^*p < .05;

^**p < .01

Overall, we examined 12 race/ethnicity by program interactions (3 outcomes in 2 programs in 2 sites). We found 2 statistically significant interactions (16% of tests conducted). This is greater than the number of significant interactions we would expect to find by chance.

Are Racial/Ethnic Differences in Policies’ Effects on Achievement Explained by Pre-Existing Differences (Background Characteristics Prior to Exposure to the Policies)?

Questions 2 and 3 were only explored for the two significant race/ethnicity-by-program interactions (predicting 2-year Bracken scores in the Riverside education-first program and 5-year Woodcock-Johnson math scores in the Grand Rapids education-first program). Table 6 presents the results for Bracken scores in the Riverside education-first program. Baseline covariates are not included in this table to conserve space; however, none of the coefficients differed substantially from those reported in Table 3. As shown in the first row of the first column of Table 6, once the baseline by education-first program interactions were included in the regression, the race/ethnicity by program interaction is no longer significant. The magnitude of the coefficient declined from 11.03 to 8.02, a 27 percent decrease. Among those baseline-by-program interactions, the interaction between education goals and the education-first program was marginally significant (b = 10.81, p < .10). This interaction is plotted in Figure 4. Among mothers reporting high levels of education goals at baseline (plotted at one standard deviation above the mean), the education-first program had a very small effect on Bracken score (effect size = −.06; left pair of bars in Figure 4). In contrast, among mothers reporting low levels of education goals, the Riverside education-first program brought about a large decrease in Bracken scores (effect size = −.80; right pair of bars).

TABLE 6.

Summary of Regression Results, Models 3 and 4, Predicting 2-Year Bracken Scores, Riverside Education-First Program

	Model 3			Model 4
	b	β	SE	b	β	SE
Measured at baseline
Race/ethnicity*Program	8.02	(0.25)	5.08	8.89	(0.27)	5.24†
Program	−6.29	(−0.22)	3.62†	−6.41	(−0.23)	3.73†
Race/ethnicity	−4.22	(−0.15)	3.35	−4.12	(−0.15)	3.40
Earnings in prior year*Program	0.84	(0.06)	1.26	1.33	(0.09)	1.31
On AFDC 5+ years*Program	−0.80	(−0.01)	5.15	−0.04	(−0.00)	5.24
High school diploma*Program	3.92	(0.06)	6.22	6.00	(0.09)	6.47
Ever married*Program	−0.05	(0.00)	5.06	−1.23	(−0.02)	5.13
Family first goals*Program	−1.13	(−0.03)	3.72	−0.76	(−0.02)	3.73
Education goals*Program	10.81	(0.17)	5.86†	11.28	(0.17)	5.90†
Depressive symptoms*Program	−1.01	(−0.13)	0.70	−0.87	(−0.11)	0.70
Measured 2 years after random assignment
Average yearly earnings				−0.68	(−0.14)	0.49
Average yearly income				−0.09	(−0.02)	0.36
Months of education				0.21	(0.07)	0.29
Depression				3.80	(0.13)	2.76
Cognitive stimulation				1.50	(0.08)	1.79
Married or cohabiting				−2.50	(−0.08)	3.02
	n = 141

Note:

^†p < .10;

^*p < .05; all analyses included baseline covariates.

FIGURE 4.

Differences in effects of Riverside education-first program on 2-year bracken scores, by baseline education goals.

Table 7 presents the results for Woodcock-Johnson math scores in the Grand Rapids education-first program. The coefficient for the race/ethnicity by program interaction remains significant after entering all of the baseline-by-program interaction terms (b = 13.84, p < .01). Therefore, the race/ethnicity by program interaction is not due to pre-existing differences between African American and Caucasian sample members.

TABLE 7.

Summary of Regression Results, Models 3 and 4, Predicting 5-Year Math Scores in Grand Rapids Education-First Program

	b	β	SE	b	β	SE
Measured at baseline
Race/ethnicity*Program	13.84	(0.18)	5.02**	13.79	(0.17)	5.11**
Program	−1.53	(−0.04)	2.16	−1.62	(−0.04)	2.19
Race/ethnicity	−1.29	(−0.03)	2.51	−1.59	(−0.04)	2.79
Earnings in prior year*Program (in thousands)	−0.07	(−0.01)	0.49	−0.11	(−0.01)	0.55
On AFDC 5+ years*Program	4.60	(0.05)	5.37	4.26	(0.05)	5.46
High school diploma*Program	−5.63	(−0.07)	4.69	−5.03	(−0.06)	4.80
Ever married*Program	−0.07	(0.00)	4.74	−0.04	(0.00)	4.78
Youngest child < 3*Program	−14.93	(−0.18)	4.88**	−14.45	(−0.18)	4.99**
Family first goals*Program	6.75	(0.11)	3.77†	6.88	(0.11)	3.81†
Education goals*Program	−2.23	(−0.03)	4.23	−1.69	(−0.02)	4.29
Depressive symptoms*Program	−0.11	(−0.01)	0.61	0.02	(0.00)	0.64
Measured 2 years after random assignment
Average yearly earnings (in thousands)				−0.31	(−0.08)	0.46
Average yearly income (in thousands)				0.39	(0.10)	0.47
Months of education				0.08	(0.03)	0.19
Depression				−1.77	(−0.04)	2.53
Cognitive stimulation				1.11	(0.04)	1.78
Married or cohabiting				−0.40	(−0.01)	2.71
	n = 289

Note:

^†p < .10;

^**p < .01; all analyses included baseline covariates.

Are Racial/Ethnic Differences in These Policies’ Effects on Achievement Explained by Racial/Ethnic Differences in Their Effects on Mediators Measured After Exposure to the Policies?

This question was addressed by including 2-year mediators in the regressions predicting 2-year Bracken scores in the Riverside education-first program and 5-year Woodcock-Johnson math scores in the Grand Rapids education-first program. The last columns of Tables 6 and 7 display these results. None of the mediators was associated with achievement, and in neither of the two programs was the coefficient representing the race/ethnicity by program interaction substantially changed in magnitude or significance. As further evidence that the hypothesized mediators do not explain the differential program effects by race/ethnicity, it is also the case that the program impacts on the hypothesized mediators do not line up in a consistent pattern with the observed effects on children’s achievement by racial/ethnic group (results available from authors).

DISCUSSION

Among the many studies examining effects of welfare policies on children’s development since the passage of welfare reform legislation in 1996, few have examined whether effects might differ by racial/ethnic group, and if so, why. This study is one of the first to focus on this question. Among the 11 large-scale experiments conducted in the 1990s with information on child developmental outcomes (Morris, Huston, Duncan, Crosby, & Bos, 2001; Morris et al., 2005), we focused on the programs of the National Evaluation of Welfare to Work Strategies because they included the few studies with sufficient numbers of African American, Latino, and Caucasian families to make comparisons. We chose student achievement as our outcome because of its importance in studies of racial/ethnic disparities in the United States and because the primary targets of welfare policies—earnings, employment, and education—have been linked to achievement in many prior studies.

We examined, moreover, two different forms of mandatory employment policies: education-first and work-first programs. We suspected that these programs might have different effects on African American, Latino, and Caucasian families due to either differences in pre-existing characteristics (measured prior to exposure to the policy) or differences in parents’ economic behaviors, family structure, well-being, or cognitive stimulation occurring after exposure to the policy.

For two of the six outcomes examined in the education-first programs, we found significant racial/ethnic differences in effects on child achievement in early childhood and in middle childhood. The education-first programs led to moderate to large decreases in school readiness and math scores among Caucasian children and small increases in these outcomes among Latino and African American children. Specifically, in Riverside, California, the education-first program increased school readiness among Latino children (effect size of +.09) and decreased school readiness among Caucasian children (effect size of −.56). In Grand Rapids, Michigan, the education-first program increased math scores among African American children (effect size of +.22) and decreased math scores among Caucasian children (effect size of −.42). Differences in program effects on reading scores were in the same direction as the effects on math scores, although the effects on reading scores did not reach statistical significance.

These differences in effects across groups are moderate to large in size, while effect sizes within groups are small to moderate in size. Although the effect sizes are not large by Cohen’s (1988) effect size standards, recent economic work from the Tennessee STAR class size experiment (Krueger, 2003) found that effect size impacts on test scores in middle childhood by roughly .2 standard deviation translated into changes in future career earnings of between $5,000 and $50,000, depending on the variation in assumed discount and earnings growth rates. Thus, the effects within groups in these comparisons may have practical as well as statistical significance.

It is interesting that effects of the education-first programs differed by race/ethnicity, but effects of the employment-first programs did not. It is possible that being assigned to an education-first program made education and learning more salient for those families, resulting in positive effects on children’s school achievement scores. Being assigned to an education-first program could have increased African American and Latino parents’ educational expectations for their children more than Caucasian parents’ expectations, resulting in differential effects on the achievement of children in these groups (unfortunately, these data sets did not include measures of educational expectations). Welfare has historically been a program that is more stigmatizing for African Americans and Latinos than Caucasians (Gilens, 1999); encountering education-first messages from caseworkers may have had a more surprising and positive effect on expectations for these groups than for Caucasians. The prior literature on low-wage employment and children’s outcomes suggested that changes in quantity of employment without accompanying changes in quality (e.g., occupational prestige) may have few effects on children (Enchautegui-de-Jesus, Yoshikawa, & McLoyd, 2006; Raver, 2003). The work-first policies examined here did not include components designed to improve the quality of low-wage work; this may be why we did not find racial/ethnic differences in effects on children.

It is also interesting that in the Riverside education-first program we found differential effects for Latinos’ and Caucasians’ 2-year Bracken scores but not 5-year Woodcock-Johnson scores. However, while not statistically significant by conventional standards, the interaction of race/ethnicity by program group status on Woodcock-Johnson math is in the same direction as the one found for the Bracken: Latinos experienced an increase in math scores (effect size of +.19) while Caucasians experienced a decrease in math scores (effect size of −.20). There is likely some stability in the common underlying abilities tapped by both the Bracken and the Woodcock Johnson math. In fact, Bracken scores were predictive of Woodcock Johnson math scores (correlation of .44), and when we examined the Riverside education-first program’s effect on Woodcock Johnson math scores, controlling for 2-year Bracken scores, the coefficient on the interaction term became substantially smaller.

What might have explained the differences in experimental effects on school achievement that we observed? First, we examined whether pre-existing differences among the ethnic groups we studied may have played a role in the pattern of effects in the Riverside and Grand Rapids education-first programs. Differences in background characteristics between Latino and Caucasian mothers in Riverside, and African American and Caucasian mothers in Grand Rapids, might explain why they might have responded to these policy environments differently. We considered four sets of baseline factors: human capital, family structure, goals related to work and education, and maternal well-being. We did find that at baseline parents from these ethnic groups differed in areas of both human capital and their preference for education and work, with Latino and African American parents showing lower levels of human capital but a higher preference for education and work, respectively. In Grand Rapids, African Americans were more likely than Caucasians to be long-term welfare recipients and less likely to have ever been married. In Riverside, Latinos were less likely than Caucasians to have a high school diploma. These patterns are consistent with general national patterns of differences in human capital across these groups (U.S. Census Bureau, 2007). With regard to differences in their preferences for employment and education, African American mothers in Grand Rapids were more likely to report a preference to work than their Caucasian counterparts, and Latino mothers in Riverside reported a stronger preference for pursuing their own education than their Caucasian counterparts. The difference between African American and Caucasian mothers in Grand Rapids in their levels of goals for work is consistent with other literature that finds that African American women report higher levels of perceived benefits from work and lower levels of perceived costs, relative to Caucasian mothers with similar levels of job skills (Granrose & Cunningham, 1989; Johnson, Jaeger, Randolph, Cauce, & Ward, 2003; Murrell, Frieze, & Frost, 1991).

Were these baseline differences related to the racial/ethnic differences in effects of policies on children? In Riverside, but not Grand Rapids, we found some evidence that pre-existing differences between racial/ethnic groups helped explain differences in policy effects. In Riverside, pre-existing differences between Latino and Caucasian parents’ education goals appeared to play a role. As we hypothesized, person-environment fit, as represented by the fit between parents’ goals and policy approach, helped explain effects of these programs on children (Moos, 1984). Recall that in Riverside, the education-first program produced a small increase in Bracken school-readiness scores among Latino children and a large decrease among Caucasian children. When examining the interaction of the Riverside education-first program with parents’ education goals, we observed an interaction of almost identical shape (Figure 4). This program had essentially no effect on Bracken scores for children of mothers with high levels of goals to pursue their own education (effect size −.06). In contrast, for children of mothers with low levels of goals to pursuing their own education, the program had a large negative effect (effect size −.80). Once this interaction was entered, the coefficient on the ethnicity by program interaction was diminished by 27% in magnitude. These findings support a “misfit” hypothesis. That is, a program that strongly encourages educational activities may represent a poor fit for parents with low levels of interest in pursuing their own education. This “misfit” appears to be associated with lower school readiness scores. In our case, Caucasian parents in Riverside were significantly less likely than their Latino counterparts to report wanting to pursue their own schooling, and, when exposed to the education-first program, their children experienced much greater declines in school readiness than Latino children. As we mentioned previously, based on Census data, the vast majority of the Latino families in Riverside were likely to be of Mexican origin during the period of this study. Although data on their immigration status was not included in this data set, our results mirror prior research that shows that foreign-born parents place greater value on their own education than U.S.-born parents (Kao & Tienda, 1995).

Low-income parents vary in the relative strength of goals to pursue work, education, and parent roles (Weisner et al., 2006); this study showed that the degree of fit between goals and policy approach may influence children. In particular, instrumental goals may matter in the implementation of welfare and employment policy. Instrumental goals help to organize attention and efforts around life tasks such as work and parenting (D’Andrade & Strauss, 1992). Parents with low levels of interest in education, and their children, may not benefit from programs that emphasize that pathway to meeting work requirements, because that set of activities does not fit their instrumental goals. Recent findings from another study examining the interaction of welfare policy with instrumental goals support this hypothesis (Gassman-Pines et al., 2009).

While pre-existing differences between racial/ethnic groups helped explain the pattern of effects in Riverside, we were not able to explain the pattern of effects in Grand Rapids through either pre-existing characteristics or post-random-assignment mediators. The difference in program impacts by race in Grand Rapids was similar to the one in Riverside (i.e., an increase in math scores among African American children and a decrease among Caucasian children). However, neither differences in program effects by baseline characteristics, nor differences in program effects on mediators, reduced the magnitude of this interaction (in fact, the coefficient became somewhat larger and more significant as additional variables were added to the basic model). Some unmeasured characteristics distinguishing African American from Caucasian families may explain the difference. We do know from exploratory analyses that African American mothers in Grand Rapids reported lower levels of support from caseworkers than Caucasian parents, suggesting a potential role of differential caseworker treatment by race. However, because these measures were not collected in the control groups, we were not able to use them to help explain differences in experimental impacts. In addition, African American mothers in Grand Rapids reported higher levels of valuing work over family, when compared to their Caucasian counterparts. However, this difference in baseline goals did not reduce the experiment-by-race interaction on math scores in Grand Rapids.

One descriptive question that arises regarding these results is the magnitude of achievement gaps among African American, Latino, and Caucasian children in the absence of these programs. How did these children’s school readiness and math scores, for example, differ in the control groups? Examining Figures 2 and 3, we see gaps in school readiness and math scores predicted by the literature on the achievement gap (Jencks & Phillips, 1998; Lee & Burkham, 2002). In Riverside, Latino children in the control group had Bracken scores that were 5 points (roughly .18 SD) lower than Caucasian children (first and third bars in Figure 2). In Grand Rapids, African American children in the control group had math scores that were 6 points (roughly .35 SD) lower than Caucasian children (first and third bars in Figure 3). As prior work suggests, even within a particular income class (in this case, poverty), and controlling for baseline demographic and human capital characteristics, African American and Latino children perform at lower levels on standardized achievement tests than Caucasian children. Note that the gap no longer appears if one compares the same groups in the experimental conditions (compare second and fourth bars in Figures 2 and 3). Latino children’s school readiness scores are actually higher than those of their Caucasian counterparts in the Riverside experimental condition, and African American children’s math scores are higher than Caucasians in the Grand Rapids experimental condition. These programs effectively eliminated the achievement gap between racial/ethnic minority and Caucasian children in these samples. However, the policy implications of this reduction are limited; the reduction occurred partially through reductions in the achievement of Caucasian children, not a preferred method of addressing achievement gaps (Magnuson & Waldfogel, 2008).

Several limitations of this study should be noted. First, we did not have measures of other potential mediators of the racial/ethnic differences in program effects, such as mothers’ experiences of discrimination in workplaces, educational expectations, or the quality of children’s schools. Second, we were not able to more fully examine the role of ethnicity, culture, and immigration experiences among these groups. For example, we extrapolated from Census data to estimate that the majority of Latino parents in Riverside County were Mexican. Future demonstrations in the field of welfare-to-work must do a much better job of assessing these factors. In addition, no qualitative sub-studies were embedded in these evaluations, to provide richer access to cultural factors that may underlie our findings. Third, the rates of attrition in Riverside were quite high (retention of 65% to the 5-year follow-up). However, our analyses showed no differences above chance in attrition or predictors of attrition across experimental or racial/ethnic groups. Finally, we caution against generalizing these findings to families in other parts of the United States, or at other points in time. Our data are quite specific to welfare recipients in our two sites, in the mid-1990s. The welfare programs in these two states, for example, have changed dramatically since this evaluation was conducted.

In sum, although we found relatively few racial/ethnic differences in effects overall, our results were above chance levels. They suggest that mandatory employment policies can have different effects by race/ethnicity. Most notably, education-first programs appeared to bring about increases in African American and Latino children’s school readiness and math scores, but not for Caucasian children in either site. We found some evidence that the Riverside education-first program may have been a poorer fit with Caucasian parents’ goals for their education than with those of their Latino counterparts. Education-first programs may represent a closer “fit” with some parents’ goals than others. The data provided supports a “mixed” approach to moving welfare recipients to work, allowing more tailoring of activities that meet employment mandates to individual preferences and needs. Activities encouraged in the mixed approach can encompass those emphasized in both the “work-first” and “education-first” programs analyzed in this study. Prior research shows that mandatory employment programs that take this approach produce the largest effects on welfare recipients’ employment and earnings (Hamilton et al., 2001); the current results suggest potential benefits for children.