The Impact of Intensive Vocation-Related Course Taking on Employment Outcomes for Individuals Who Are Deaf

Dataset

Study Participants

Descriptive statistics for the sample, both in weighted percentages and in frequencies, are included in Table 1. With the exception of degree of hearing loss, all of these variables were used as covariates in the propensity score analysis. Just over 80% of the sample was enrolled in a general education school in Wave 1 of NLTS2 data collection. In the first wave of data collection, parents gave a description of the school type their child attended in the past year. This question included a variety of alternative settings such as whether or not the child attended a charter school, a magnet school, or a juvenile justice facility. In this study, we provide data on the prevalence of intervention use in the first two placement options listed on the parent questionnaire: (a) a regular school that serves a wide variety of students (i.e., general education school) and (b) a school that serves only students with disabilities (i.e., special education school). The general education school category thus includes students who are in special education settings such as pull out or support services for part of the day but who also spend part of their instructional time with students without disabilities. The special education school category includes institutions that primarily serve students with disabilities in a separate environment, including schools for the deaf. Because the NLTS2 instrument was designed for a broad demographic sample, the survey did not allow parents to designate whether their child attended a school for the deaf, specifically, or to make distinctions between residential and day schools.

OPEN IN VIEWER

Table 1.

Covariates and Descriptive Statistics.

Variable	Weighted percentage	Raw frequencies
Covariates
Female	51.6	410
Age (as of 2002)
14	14.3	140
15	26.0	210
16	17.9	170
17	31.9	220
18	9.92	80
Has additional disability	35.7	320
Parents highest education level
No GED/high school diploma	18.3	140
GED/high school diploma only	34.2	250
Some postsecondary education	15.1	110
Vocational degree	3.67	30
Associate’s degree	10.6	90
Bachelor’s degree	9.90	30
Some graduate work	1.96	40
Graduate degree	6.34	10
Yearly household income (in US$)
20,000 or less	19.29	150
20,001–30,000	18.4	110
30,001–40,000	15.2	90
40,001–50,000	7.41	70
50,001–60,000	12.1	70
60,000–70,000	9.35	70
70,001 or more	18.06	120
Ethnicity
White	60.4	490
African American	16.5	150
Hispanic	18.0	130
Asian	4.12	30
Attends general education school	81.3	520
Is hard of hearing (not deaf)	33.4	270
Youth mainly uses sign language to communicate	57.0	500
ASL main language at home	23.2	220
Student’s freshman GPA is . . .
<0.99	3.42	30
0.99–1.99	15.8	90
1.99–2.99	47.3	310
2.99–4.00	33.4	200
Descriptive statistics
Degree of hearing loss
Mild	11.1	70
Moderate	26.7	160
Profound	62.2	500

Note. Because NLTS-2 provides individually identifiable data, the raw frequencies reported here are rounded to the nearest tens place. This is in accordance with IES policy. Note that the weighted percentages and the raw frequencies may not correspond due to the weighting scheme employed by NLTS2. GED = General Educational Development; ASL = American Sign Language; GPA = grade point average; NLTS2 = Second National Longitudinal Transition Study.

The sample included 11.1% of youth with a mild hearing loss, 26.7% with a moderate hearing loss, and 62.2% with a profound hearing loss. Although the inclusion criterion for “hearing impairment” was a general one, this analysis also included three covariates in the dataset related to deafness and language use: “is hard of hearing (vs. deaf),” “mainly uses sign language to communicate,” and “American Sign Language (ASL) is the main language at home.” In this sample, 33.4% of youth were considered hard of hearing, 57% primarily used sign language to communicate, and 23% had ASL as the main language at home.

Table 1 also includes information on additional demographic variables used as covariates in this analysis. For example, just over 60% of participants were Caucasian. Ethnicity data were taken as indicator variables for African American, Asian, and Hispanic students, against a Caucasian comparison group. Although the literature indicates a need for more careful exploration of racial categories and overrepresentation of racial minority groups within special education research (Sullivan & Artiles, 2011), within the context of propensity score matching, the choice of Caucasian as a comparison group has a specific analytical function. The purpose of propensity score matching is that the treated and control groups need to have a similar demographic makeup. If all but one demographic covariate is similar between these groups, then the last covariate, of being Caucasian, must be similar also. This allows for unbiased causal inference.

A variety of variables for a range of co-occurring disabilities were included in the propensity score analysis, with separate indicators for students who had an attention deficit disorder, students who were deaf-blind, students who had a learning disability, students with intellectual disabilities, students with orthopedic impairments, and students with other health impairments. A total of 35.7% of sample participants were identified as having some sort of additional disability. The freshman-year GPA, drawn from the student transcripts, corrects for student achievement in this analysis. Although other achievement data are available, such as the Woodcock Johnson III score, it is possible that those data were collected after students participated in the independent variable. In propensity matching analysis, it is important that the covariates be exogenous to the independent variable. Therefore, no other achievement variables were used as covariates.

Variables

Independent variables and covariates for this analysis were entirely drawn from the first two waves of NLTS2, which were collected in 2001 and 2003. The dependent variables were drawn from across the last three waves, ending with the final wave in 2009.

Propensity Score Analysis Plan

In contrast with regression analysis, propensity score analysis corrects for the effect covariates have on the treatment assignment process (see Note 2). Rather than modeling the relationship between the treatment and the outcome directly, propensity score analysis begins by modeling the treatment assignment as a logistic regression, dependent on a variety of covariates. The ultimate objective of this logistic regression is to define treated and untreated groups that are similar on the set of identified covariates. To do this, the propensity score, defined as the probability of being treated, is obtained for each participant from the logistic regression. Next, treated and untreated participants are matched so that their propensity scores are similar. Participants with the same propensity scores must have the same distribution of covariates (Rosenbaum & Rubin, 1983), which removes this source of bias in the relationship between the treatment and the outcome.

Finally, once treated and untreated participants are matched on the basis of this propensity score, regression analysis or t tests may be used to estimate the treatment effects. The t tests may be preferred for their simplicity in some cases, but this study uses regression analysis to correct for whatever bias remains in the covariates. Theoretically, a perfectly matched dataset on the propensity score would ensure that no bias would exist, but in practice, perfect matches are extremely unlikely. As such, regression analyses correct for the remaining difference in the covariates.

The analysis proceeded in three stages to estimate the average treatment effect on the treated (ATT). First, a propensity score model was fitted for each program, using the study covariates. To handle missing data, we used multiple imputation with chained equations (van Buuren & Groothuis-Oudshoorn, 2011) in R (R Core Team, 2013), and took the propensity score as the average value across imputations. This multiple imputation process did not include any treatment or dependent variables. Once the propensity score had been estimated for a treatment, a caliper matching procedure was conducted, matching students within 0.25 standard deviations on the propensity score, without replacement, using the “Matching” package in R (Sekhon, 2011). Next, balance was assessed for each covariate (postbalance statistics available upon request). Adequate balance was defined as a less than 0.25 standardized mean difference (Ho, Imai, King, & Stuart, 2007). These balance assessments were completed using the weighted data. When inadequate balance was obtained on more than two covariates, interaction terms and squared terms were added to the propensity score model. We repeated this process until we had adequate balance on at least all but two covariates for each intervention. To correct for the remaining imbalance postmatching, all the covariates were used in ordinary least squares (OLS) regression on the postmatching dataset to estimate the effect of each treatment on the outcomes. The treatment variable was allowed to interact with each covariate in the latter regressions, and we used a type I error rate of 0.01. Multiple imputation with chained equations was used again in these latter regression analyses (van Buuren & Groothuis-Oudshoorn, 2011). For each multiple imputation, we used 20 imputations and 10 iterations. Every covariate in this analysis had at most a 20% rate of missingness. No variable of substantive interest was eliminated due to a high rate of missingness.

The use of multiple imputation requires the assumption of data being missing at random (MAR), meaning that, once covariates are controlled for, the probability of a variable being missing does not depend on the value of that variable (Allison, 2001). In short, the assumption is that there is no response bias after correcting for every variable. While it is not possible to evaluate this assumption directly, having a large number of variables in the missing data model makes the assumption more credible (Collins, Schafer, & Kam, 2001). The large number of covariates in this study, in combination with the relatively low rate of missingness, helps to ensure the feasibility of this assumption.

Throughout the analysis, unless otherwise noted, we used the survey package in R to account for the survey sampling design (Lumley, 2004). Because NLTS2 was intended to be nationally representative, surveyors employed a stratified weighted sampling scheme. As such, both the propensity score matching and the later regression analyses were adjusted using the weighting, cluster, and stratification data provided in the dataset, using the weights calculated for any wave and any data source.

Descriptive Statistics

The first purpose of this study was to describe vocation-related course taking utilized by students who are deaf (see Table 2). Because of the historical emphasis on educational placement within deaf education, we disaggregated these data between general education schools and special education schools. First, there were no differences in the proportion of deaf students who had a main academic IEP goal to develop vocational/prevocational skills (27% for both general education and special education schools). Given what we know about the complexity and potentially overlapping characteristics of students who are deaf across school type in secondary grades, this is perhaps not a surprising result (Shaver et al., 2014).

OPEN IN VIEWER

Table 2.

Comparing Intervention Participation Between School Type Within the First Wave of Data Collection.

Intervention	General education	Special education	Wald statistic	p value
Sequence of vocational classes	39.5%	51.7%	1.92	.167
More than 1 year of work-study courses a	6.12	30.1	8.80	.003
Student’s main academic goal is to develop vocational/prevocational skills	27.0	27.0	<1	.988

a

Indicates that differences in the proportion of students receiving the type of intervention is significant at p < .05. Note that each test has 178 degrees of freedom.

What did implementation of these IEP goals look like in terms of course taking related to vocational skill development? Our analysis focused on intensive experiences, those that had the potential to build skills over a long period of time and thus more likely to have an impact on later employment outcomes. Students in general education schools (39.5%) and special education schools (51.7%) were equally likely to take a sequence of vocational classes (four or more courses on a specific occupational area) as part of their high school curriculum, χ²(178) = 1.92, p = .162. This finding shows that a substantial proportion of students who are deaf receive coursework that could help prepare them for more advanced work in a content area as part of their transition into a postsecondary training or employment experience. However, a smaller proportion of students, 6% and 30% for general education and special education schools, respectively, enrolled in more than 1 year’s worth of work-study courses while in high school (χ² = 8.80, p < .01).

Propensity Score Models

First, we report two-way relationships for each intervention variable (e.g., course taking) and each measured employment outcome variable. For continuous outcomes, we report naive point-biserial correlation coefficients. The relationship between identified independent variables and our outcome variables appear to be weak. More specifically, every correlation between coursework enrollment and long-term employment outcomes were less than 0.20. Also, for the binary outcome of promotion opportunities, no two-by-two chi-square test with coursework taken and the binary outcome was statistically significant. Although the naïve estimates of the treatment effects are low, it is possible that this is due to heterogeneity in the treatment effect, which the later analysis accounts for.

Following the procedure outlined in the analysis plan section, above, certain interaction terms were included to improve balance on the specified covariates. One out of the three independent variables achieved adequate balance without adding any of these terms. Namely, the sequence of vocational courses variable had adequate balance without adding any interaction terms. The other two independent variables each had separate specifications. First, the independent variable stating whether the student’s main IEP goal was to improve vocational skills had three interaction terms. Freshman-year GPA was allowed to interact with parent-ratings of social skill and parental expectations about their child’s postsecondary attendance. Also, parent ratings of social skill interacted with parent expectations about postsecondary attendance. It was difficult to obtain proper balance on the final independent variable, spending more than 1 year’s worth of coursework in a work-study program. Every continuous covariate except age had a two-way interaction with every other continuous covariate. In a typical regression approach, this sort of specification might be thought of as overfit. However, the only purpose of the propensity score analysis is to achieve balance on the covariates. Essentially, whatever specification achieves this is admissible in a propensity score approach (Ho et al., 2007; Sekhon, 2011).

Postmatching regressions

The purpose of the propensity score analyses was to estimate the effect of vocation-related course taking on employment outcomes approximately 8 years after their completion. With an alpha level of .01, we failed to find significant results on any outcomes for any intervention. Most treatment effects on the treated had standard errors that were greater than their average value. These findings were robust to matching with or without replacement. Equal numbers of this effect size were positive and negative, with standard errors sometimes twice the size of the ATT (Table 3).

OPEN IN VIEWER

Table 3.

Average Treatment Effects, per Program and Dependent Variable.

Program	ATT for hourly wages (SE)	ATT for job score (SE)	ATT for number of jobs since HS (SE)	ATT for promotion, raise, job opportunity (SE)
Sequence of vocational courses	−1.274 (1.891)	1.678 (3.084)	2.261 (2.112)	1.592 (1.060)
IEP goal to improve vocational skills	−3.293 (3.091)	4.006 (3.193)	−0.2462 (4.281)	−1.259 (1.631)
More than 1 year work-study programs	2.688 (3.198)	3.025 (5.134)	−3.792 (3.047)	−0.4773 (1.927)

Note. ATT = average treatment effect on the treated; HS = high school; IEP = Individualized Education Program.

Prevalence

Student enrollment in intensive vocation-related coursework varied across types, with higher rates for opportunities already integrated into the school’s instructional structure than those that operated outside of it. An intensive sequence of courses in a related area was more prevalent than work-study coursework. The former category is a kind of instructional approach to transition support, one that is perhaps more reasonable in a school setting because it aligns more closely with the other resources available in a school than internships or other off-site opportunities. Work-study course enrollment levels for more than a year, in contrast were utilized by less than 10% of students in the general education school sample and no more than a third of students in the special education school sample. It is possible that there is not time in the schedule or opportunities for additional work-study beyond 1 year due to other constraints on the course schedule, particularly in general education schools where the focus is potentially on broader range of students than at special education schools. Yet these findings are higher than those found for average work-study enrollment for participants in general education settings, where work-study credits ranged on average from .2 credits to .8 credits (one credit = 1 year), depending on the disability category. Work-study courses require greater coordination between school resources and external agencies or institutions, and thus are likely more difficult to easily incorporate into a student’s transition preparation experience. Note that this analysis did not include employment experience that was separate from school-based initiatives, such as having a job on the weekends that was not facilitated by a school transition program, and that students in this sample likely did have some experiences that were outside the structure of the school setting.

When there were differences in program involvement by school type, students in general education schools were less likely to enroll in a sequence of four or more vocation-related courses than students in special education schools. The descriptive analysis of course enrollment did not include consideration of covariates, such as whether or not a student had a co-occurring disability, and thus only provides a surface view of the fit between the school type and intensive vocation-related coursework. In an overview of student characteristics drawn from the same NLTS2 sample, deaf students in special secondary schools were more likely to have profound hearing loss and a disability classification of deafness, and less likely to use spoken language, than deaf students in general secondary schools (Shaver et al., 2014). However, it should be noted that even when there were statistically significant differences in course enrollment between placements, the split between proportions of students enrolled was no more than 25%. This is a relatively small margin given the estimating procedures in the analysis. Thus, even if there were significant interactions in the data based on student characteristics, it is unlikely to provide a substantively different picture of what interventions deaf students are participating in across different types of educational placements than what is provided here.

Impact on Employment

This study also examined the potential impact of vocation-related coursework on long-term employment outcomes for deaf students. The research literature implies that we might see some positive effects of school-based transition interventions, particularly if we are able (as we were) to control for other variables that are also predictive of these types of outcomes (Baer et al., 2003; Benz, Lindstrom, & Yovanoff, 2000; Roessler, Brolin, & Johnson, 1990; Shandra & Hogan, 2008; Solberg et al., 2012). This analysis did not demonstrate a significant relationship between participation in NTLS2 identified vocation-related coursework and employment-related outcomes by the end of the study period.

Before interpreting the meaning of the above findings in relation to the broader NLTS2 literature, it is important to consider the context of analytical approach. The majority of published secondary data analyses using the NLTS2 rely on regression analysis to look at significant factors associated with transition outcomes, such as academic and vocational programs and experiences, social competence, parent expectations, and academic performance (Bouck, 2012; Carter et al., 2012; Cawthon, Caemmerer, et al., 2015; Chiang, Cheung, Hickson, Xiang, & Tsai, 2012; Coyle, 2012). Although covariates do help to filter out effects from factors other than the intervention or program of interest, regression analyses are limited in their capacity to mimic a randomized control trial because they are heavily dependent on the way that they are specified, and analysts have the capability to respecify the model to obtain significant results. In contrast, propensity score matching, the method used in the current analysis, depends less on how exactly the models are specified (see, for example, Cook, Shadish, & Wong, 2008; Morgan & Winship, 2007) than for regression analyses. Propensity score matching is thus more conservative than regression analysis. This difference in analytical approach, as well as our inclusion of several covariates in our analysis, resulted in reduced variability left to explain effects of our study variables and may help to explain our noneffects.

What does it mean to have a nonsignificant effect in this kind of analysis? The remainder of this discussion will address potential limitations related to the dataset and the interventions studied, what it means to look at these variables for deaf individuals, and the relationship between these interventions and other factors that play an important role in opportunities for deaf students.

Variable Definitions

As with all secondary data analyses, there are limitations related to variable selection and construct identification that go beyond the limitations related to self-report often found in survey data collection. More specifically, although the NLTS2 design captured many important variables related to transition for students with disabilities, it was necessarily designed to capture data for a relatively broad swath of students with disabilities. Although the dependent variables in this study are items that can be relatively reliably measured (e.g., salary or employment status), vocation-related courses have a great deal of variability in how they are defined and what is included in their implementation. There is no set curriculum, for example, on what skills or opportunities students need to have when they participate in specific occupational courses or work-study course sequences. Thus, there is wide variation from setting to setting and even from class to class, how the instructor conceptualizes the work, what opportunities are available in the work setting, and how the student engages in the content. Without course syllabi and specific examples of class objectives or work contexts, as well as information about the quality of how these experiences are integrated into other components of transition for deaf students, it is difficult for a secondary analysis to meaningfully assess noneffects of these treatments on employment outcomes.

Moreover, this analysis was unable to obtain appropriate balance on another potential covariate, namely, whether or not the participants attended a general education school or a special education school. We made an attempt to include this covariate in the analysis, and, even after adding several interaction and squared terms, we were unable to specify a propensity score model which, after matching, would obtain adequate balance on this binary variable. State sponsored special schools, including schools for the deaf, were oversampled in the NLTS2 data collection (Shaver et al., 2014), which may be corrected by using survey weights. As such, we have attempted to use survey weights wherever possible in this research.

Outcomes for Deaf Individuals

Even with potential limitations in construct and population definition, the NTLS2 has been utilized by other researchers who have found a positive relationship between transition planning activities and outcomes for students with disabilities (Carter et al., 2012; Chiang et al., 2012; Grigal, Hart, & Migliore, 2011). Why might there be different results in this analysis of interventions for deaf students than for students with other disabilities (or for an analysis of students with disabilities as a whole)? One of the great challenges in this area of research is the inability to randomly assign students to different interventions, and thus the need to use other analytic strategies to control for the inherent differences between students who obtain services and those who do not. By definition, a student receives transition services because there is a perceived need for support beyond what is available to all students. If we can assume there is an appropriate match of services to student needs, such that those who receive services stand to benefit from them and those who do not are reasonably expected to succeed without them, then a nonsignificant result between students who do and do not receive services may be an indication that the intervention is successful in raising outcomes of students who need the intervention to the level of those who do not.

However, a nonsignificant result could also be an indication that there is something different about how deaf individuals experience these interventions compared with other students. It is possible, though not testable using these data, that some of the courses may be limited in their efficacy with deaf individuals. It may be that the vocation-related opportunities are not as fully accessible as they could be, or students need more foundational training before taking vocation-related coursework (Qi & Mitchell, 2012), limiting the amount of information or employment-related skills that students who are deaf receive from those experiences. For work-study opportunities, in particular, it may be that emphasis is solely on entry-level skills, even when students work in those settings for more than a year (which a third of deaf students in special education schools reported doing). That additional time may not be used effectively as an opportunity to develop high-level skills, mitigating the potential long-term impact of a more intensive work-study experience in high school. Finally, it may be that despite training and support provided at the high school level, that employment outcomes for deaf individuals are less about education and training they receive in high school and more about what happens after they leave school and enter the workforce (Kelly, 2013; Winn, 2007). This would be consistent with findings that postsecondary enrollment for deaf individuals is among the highest across all disability groups, and thus experiences in those training environments may have more of an impact than high school and transition services. These findings certainly place vocation-related course taking, even at the intensive levels measured in this study, as one part of what is necessarily a multifaceted approach to career and vocational education. The onus here is thus not only on the individual and the school system to provide training and support skill development but also on the need for larger systemic changes in the accessibility, both physically and psychologically, of the workplace environment.

Limitations and Considerations for Future Research

The purpose of the current study was to describe the types of vocation-related course-taking deaf students enroll in and to understand the potential impact of service use on long-term employment outcomes. Transition services, however, have broader goals than simply supporting individuals as they seek to start a career. Transition planning includes elements such as independent living and meaningful engagement in social environments (Kohler, DeStefano, Wermuth, Grayson, & McGinty, 1994; Sitlington, 1996). This embedded scope is represented in classic definition of transition put forward by Halpern (1994):

Transition refers to a change in status from behaving primarily as a student to assuming emergent adult roles in the community. These roles include employment, participating in post-secondary education, maintaining a home, becoming appropriately involved in the community, and experiencing satisfactory personal and social relationships. (p. 116)

Although these personal and social outcomes are often related to successful employment and gainful careers, isolating work-related variables from the broader context of transition may limit the interpretability of the findings in our current analysis. Our focus on employment outcomes in this study is thus perhaps too distant from the measurable effects of transition programming; identifying mediating variables and experiences, particularly those that are relevant to the linguistic, cultural, and social context of individuals who are deaf, is an important area for future research. For instance, other literature in the field suggests that self-determination and autonomy may make an important contribution to employment outcomes (Martorell, Gutierrez-Recacha, Pereda, & Ayuso-Mateos, 2008; Shogren et al., 2015).

Second, evidence-based transition planning focuses on several key components, including parent involvement, inter-agency collaboration, direct services, and student development (Kohler et al., 1994). In the latter domain, both academic skills and “soft skills” (such as self-determination and social skills) are needed to successfully navigate workplace environments. Some of these soft skill prerequisites are captured by the variables in Waves 1 and 2 of NLTS2 and served as important covariates in the current analysis. Yet the NLTS2 design does not track the development of these soft skills through the transition process. In future study designs, it might be useful to look at development of individual factors typically used as predictors of transition outcomes, and see how these factors develop throughout the transition process. Achieving this goal would require data collection in multiple waves (beyond Waves 1 and 2). With most study participants no longer in a school setting, collecting data via direct administration of measures would prove to be a challenge. However, we cannot assume these skills are static and unaffected by transition services, especially when many of those classes or developmental experiences are designed to directly address those areas. Future models of the impact of transition programming would benefit from the inclusion of student development variables at multiple time points in addition to behavioral benchmarks such as enrolling in postsecondary training or obtaining sustainable employment.

Also, because certain models were overfit, it is possible that the variance of the treatment effect was overestimated. Certain literature suggests that when propensity score models are overfit, the variance of the treatment effect may be overestimated (Lechner & Smith, 2002). This would lead to the statistical tests failing to reject the null hypothesis, even when the null hypothesis is untrue. Possibly the nonsignificant results shown here are a result of this technicality. However, more recent work suggests that the treatment effect itself is well-estimated even when the propensity score itself is misspecified, given that balance exists in the matched dataset (Zhao, 2008). Half of the estimated treatment effects are actually negative, further suggesting null findings.

Third, there were a number of decisions made in the variable selection process that may limit the generalizability of our findings to students in different educational contexts. More specifically, we used the school type designation that was provided during the first wave of data collection rather than to consider possible shifts in school type contexts that students enrolled in across the multiple waves of data collection. Because we did not delete students who may have experienced both general education and special school settings, our designations do not allow for that complexity in enrollment across waves and resultant impact on study outcomes. We also did not exclude individuals who may have still been in a postsecondary setting during the final wave of the study from our analysis. These individuals would have had fewer opportunities, perhaps, to attain employment goals, at least within the timeframe of the data collection process. The current analysis is not able to delineate longer term impact of vocational course taking for individuals who were either in school or in combined work-school environments in the last wave of the study.

Finally, as in all studies of this type, we were unable to account for unobserved factors that may have an effect on measured outcomes. The propensity score approach adjusts for observed covariates, but does not balance on unobserved factors. In this study, we were limited to the variables available in the NLTS2 dataset, and thus there may be some unobserved confounds. For instance, the NLTS2 dataset did not capture individuals’ occupational aspirations beyond the expectation to “have a job,” and these aspirations about the type of employment that they hope to pursue are significantly related to future employment attainments (Schoon & Parsons, 2002). Some other potential confounds that we did not include in this study, and have emerged in other work as predictors of employment, include family background characteristics such as type of job held by the parent, whether parent receives governmental benefits, and the child’s household living conditions (ratio of people and rooms in the house, ownership of household; Schoon & Parsons, 2002). These findings must be viewed with caution and in the context of both the observed and unobserved factors included in the analysis.