Which Middle School Model Works Best? Evidence From the Early Childhood Longitudinal Study

Abstract

There are few areas of school organization that reflect more dissatisfaction than how to structure the education of adolescents in the middle grades. This study uses multilevel models on nationally representative data provided by the Early Childhood Longitudinal Study to investigate the relationship between schools’ middle-level grade span and students’ math achievement. Classroom quality was considered as an explanation for any relationships between grade span and achievement. Also examined was whether gender and family structure moderated this relationship. Results indicate that there is no generalizable relationship between grade span configuration and math achievement, but that measures of classroom quality predicted math achievement. The results should give reflective pause to reformers considering whole-scale changes to the ways in which grade spans are organized and sharpen the policy focus on classroom quality.

Keywords

education educational achievement quantitative methods

There are few areas of school organization that reflect more dissatisfaction than how to structure the education of adolescents in the middle-level grades. From the emergence of the junior high school over a century ago, through the peak years of the middle school movement, to the current trend toward K-8 schools, states and school districts across the United States are perplexed about how to best organize schools’ grade spans for adolescents. Moreover, research that has informed the development and adoption of these various models has not generated consensus as to which model works best, nor which types of students are best served by certain configurations.

If there is a consensus in this research literature, it is that early adolescence is a stage during which differences among students’ educational trajectories greatly accelerate (Juvonen, Le, Kaganoff, Augustine, & Constant, 2004). For many students, course grades decline (Barber & Olsen, 2004), anxieties increase (Grills-Taquechel, Norton, & Ollendick, 2010), academic motivation, school interest and sense of belonging decrease (Maehr & Midgley, 1996), and behavioral troubles surface (Theriot, Craun, & Dupper, 2010), all trends that are more pronounced for traditionally marginalized students (French, Seidman, Allen & Aber, 2006). In addition to schools’ grade spans, differences in educational trajectories have also been attributed to school- and classroom-level characteristics, particularly the quality of classroom instruction (Eccles, Lord, & Midgely, 1991). Thus, the middle school years are characterized by a set of negative outcomes that may jeopardize the likelihood of secondary and postsecondary success. In response, states and school districts across the United States are reconsidering the practice of educating adolescents in stand-alone middle schools, which typically span Grades 6 to 8, and replacing them with K-8 schools (Hough, 2005).

The research that has informed these efforts, however, is limited by three factors. First, although a small number of rigorous studies offer strong evidence of the effect of one schooling form versus another, many studies have been hampered by limited external validity (see, e.g., Rockoff & Lockwood, 2010; West & Schwerdt, 2012). Second, many of these studies rely on administrative data and exclude measures of classroom quality. Third, while both of these issues are related to research design, the final issue is that few studies have examined how student characteristics moderate the influence schools’ grade spans have on adolescents’ academic achievement. While attempting to discern the effects of one grade span model versus another, policies based upon this research have resulted in a one-size-fits-all approach to finding the optimal middle-level grade span. These approaches have sought to identify average effects, but not how these effects vary across different types of adolescents.

Background

Middle School Models

The middle school is generally believed to have not lived up to its potential and that, as typically implemented, does not adequately meet the needs of adolescents (Cuban, 1992; Weiss & Kipnes, 2006). Like the junior high, the middle school has been the focus of reform efforts practically since its inception (Juvonen et al., 2004). Based in part on research on middle school effectiveness, several districts have initiated reforms to dismantle their middle schools and educate students in the middle grades through other models (Cook, MacCoun, Muschkin, & Vigdor, 2008).

This shift is motivated by research that asserts that these other models are associated with smaller, more personal instruction, tighter social connections, and as a result, gains in students’ cognitive and noncognitive outcomes. For example, research has attributed to attendance at a middle school a number of negative changes in the middle grades years, including academic achievement (Alspaugh, 2001; Hanushek, Kain, & Rivkin, 2004) and motivation (Rudolph, Lambert, Clark, & Kurlakowsky, 2001). Moreover, others have reported that students in K-8 schools have higher levels of self-esteem and perceptions of school safety (e.g., Way, Reddy, & Rhodes, 2007; Weiss & Kipnes, 2006). Cook, MacCoun, Muschkin, and Vigdor (2008) found that sixth graders attending middle schools are substantially more likely to experience disciplinary trouble at school.

Fewer studies have reported on the cognitive benefits of K-8 schools; however, a few are noteworthy. In Philadelphia, Mac Iver and Mac Iver (2006) found that students in long-established K-8 schools generally outperformed middle school students in math, but these gains were not as large in newly established K-8 schools, a finding later corroborated by Byrnes and Ruby (2007). A study of schools in Cleveland found higher reading and math scores for sixth grade students attending K-8 schools, compared to peers in middle schools (Poncelet & Metis Associates, 2004). Rockoff and Lockwood (2010) reported substantial cognitive benefits for students in New York City who attend K-8 schools. Alternatively, studies that have incorporated measures related to classroom-level phenomena have reached different conclusions. For example, Holas and Huston (2012) report that while the grade span of the middle-level grades matters, its effect depends on classroom quality and other school characteristics. Therefore, the grade span configuration of the middle-level grades may mask a more important and malleable mechanism—classroom quality.

Middle School Effects

Some portion of the dissatisfaction with both the junior high and middle school can be traced to what developmental psychologists have referred to as the “developmental mismatch” between adolescents’ developmental needs and the social and academic configurations of the middle schools they attend. An extension of person-environment fit theory, stage-environment fit focuses on how the myriad changes early adolescents experience are not well supported by the middle school environment (Eccles, Lord, & Midgely, 1991; Seidman, Aber, Allen, & French, 1994).

In addition, early adolescents also experience major changes in relationships with peers, parents, and adults at school, particularly teachers (Eccles & Roeser, 2009). Early adolescents desire more autonomy, greater decision making ability, an increased need for competence and relatedness, and a shift from family-centered orientation to peer culture (Deci & Ryan, 1994; Giordano, 2003). Relations with teachers and other adults at school become less personal, less positive, and more punitive (Feldlaufer, Midgley, & Eccles, 1988).

These changes may be exacerbated by middle schools and the transitions into them (Rudolph et al., 2001). For example, middle school administrations tend to place greater emphasis on control and discipline, giving students less academic autonomy during a critical period where they desire more. Further, in middle schools the quality of student–teacher relationships decrease, becoming less personal and less positive (Eccles, Wigfield, Midgley, & Reuman, 1993). Grading practices also become more stringent, which can undermine student self-efficacy and feelings of competence (Eccles, et al., 1993; Friedel, Cortina, Turner, & Midgley, 2010). These environmental changes, coupled with the normal course of adolescent development, result in a developmental mismatch between the needs of adolescents and the school environment.

While the literature suggests that middle schools are generally challenging environments for all students, a number of studies have reported that boys and girls differ in their reaction, but these differences are inconsistent (Holas & Huston, 2012). For example, several studies (e.g., Akos & Galassi, 2004; Skinner, Marchand, Furrer, & Kindermann, 2008) have reported that girls maintain their engagement with school, while boys’ decline; however, Barber and Olsen (2004) reported that girls experience a decline in achievement from sixth to seventh grade, but no decline from seventh to eighth grade in contrast to boys, who demonstrated a steady decline from sixth through eighth grade. In sum, the evidence that boys and girls differ in their response to middle school has failed to generate a consensus as to which grade span configuration, if any, is best suited for boys or girls (Holas & Huston, 2012).

In addition, there is evidence that family structure influences how adolescents experience middle school transitions (Hines, 2007), although evidence has yet to yield a clear story. For example, as peers play an increasingly important role in the lives of adolescents, single parents may not have the resources to counterbalance their influence, which may result in declines in school performance as well as noncognitive outcomes (Demo & Acock, 1996). Conversely, it may also be that the presence of older peers, especially those who are academically oriented, can serve as a protective factor (Rashmi, Melanson, & Levin, 2007). Given the different ways in which schools with varied middle-level grade spans interact with family dynamics, it worth investigating how family structure moderates the influence schools’ grade spans have on adolescents’ academic achievement.

Estimating the Effects of Middle Schools

A majority of past studies on the effects of middle schools have used two primary strategies: (a) within-cohort designs and (b) cross-sectional designs that have limitations in assessing causal influence. A small but influential number of studies have addressed these limitations. First, Blyth, Simmons, and Bush (1978) found that seventh-grade boys in junior high school do not experience the growth in self-esteem of K-8 seventh graders, nor do they experience the decline in self-esteem of the girls in junior high school. Second, Byrnes and Ruby (2007) used data from over 40,000 eighth-grade students collected over five years in Philadelphia. Controlling for prior achievement, their models revealed that (a) established K-8 schools had higher levels of achievement than newly formed K-8 schools or middle schools and (b) that this advantage decreased when controlling for students’ demographics. Third, also using data from Philadelphia, Weiss and Baker-Smith (2010) found that, controlling for prior achievement, students who attended middle school in eighth grade performed worse in ninth grade, relative to peers who attended K-8 schools; however, a substantial portion of this K-8 advantage is due to differential likelihood of attendance at the district’s magnet high schools. Fourth, Rockoff and Lockwood (2010) examined 10 years of data on New York City school children and concluded that when students move to a middle school, their academic achievement fell substantially relative to that of their counterparts who continued to attend K-8 schools. Finally, West and Schwerdt (2012) used statewide administrative data from Florida to estimate the impact of attending public schools with different middle-level grade configurations on student achievement through Grade 10. Similar to Rockoff and Lockwood, they find that students moving from elementary to middle school suffer a drop in achievement in the transition year. Specifically, they found that middle school entry results in achievement declines by at least 0.12 and 0.09 standard deviation units in math and reading, respectively, for the predominant group of students entering middle schools in Grade 6.

Though all five studies have strengths, their designs are not without shortcomings. First, because these five studies focus on a specific geographic area (school district or state), they are limited in the extent to which their results are generalizable to larger populations. Notably, the studies also rely exclusively on administrative data, confirming achievement differences between students in different middle-level grade span configurations, but lacking information regarding the specific mechanisms. Generally, these studies find middle school students performing more poorly than their same-grade peers in K-8 schools, but few, if any, measures of classroom-level characteristics are included, inhibiting any conclusions to be drawn about the presumed effects of grade span versus classroom quality (Holas & Huston, 2012). Moreover, they identify average effects without paying mind to how these effects vary across different types of students.

Purpose of the Present Study

The purpose of this study was to expand upon the strengths and limitations of prior research by examining the specific mechanisms within schools’ of middle-level grade spans have on students’ achievement outcomes using data from a nationally representative sample of adolescents, the Early Childhood Longitudinal Study–Kindergarten Class (ECLS-K). To do so, three primary research questions were addressed. First, what is the influence middle schools’ grade spans have on eighth-grade adolescent students’ academic achievement? Second, after controlling for relevant student- and school-level characteristics, does classroom quality account for (if any) variation in students’ achievement among the different grade span configurations? Third, do gender and family structure moderate the influence middle schools’ grade spans have on students’ achievement? The idea motivating this last question is that certain types of students, specifically those with social and demographic characteristics that have been associated with variations in school-related outcomes during the middle-level grades may benefit more (or less) from one type of configuration versus another.

Method

Sample

The data used to address these questions are drawn from the Early Childhood Longitudinal Study, Kindergarten Class 1998-1999 (ECLS-K). Developed under the sponsorship of the U.S. Department of Education, the ECLS-K followed a nationally representative cohort of children from kindergarten into middle school. The base-year data were collected in the fall and spring of the 1998-1999 school year, when the sampled children were in kindergarten. The sample was updated in the spring first grade (round four), when first graders who had not been enrolled in kindergarten in 1998-1999—and, therefore, had no chance of being included in the ECLS-K base year—were included. The sample design for spring eighth grade, the seventh and final round of data collection, called for including all 12,129 children eligible after spring fifth grade (round six) and following all “movers” (i.e., those who switched schools) without any subsampling. A complete discussion on the sample design can be found in Tourangeau, Nord, Lê, Sorongon, & Najarian (2009, Chapter 4).

To reduce the complexity of the research questions and to exploit the study’s panel design, data from the spring fifth grade and spring eighth grade waves are used. The final analytic sample was derived from those students who met four criteria: (a) a nonzero longitudinal weight; (b) a valid score on the spring eighth-grade math assessment; (c) attended a comprehensive public school in the spring eighth grade, which excludes students in magnet and charter schools; and (d) whose eighth-grade math teacher completed the teacher questionnaire. This subsampling strategy resulted in a final analytic sample that included 2,729 children nested within 977 schools. The final analytical sample can be generalized to children who attended kindergarten in the United States in the 1998-1999 school year or attended first grade in the United States in the 1999-2000 school year. All analyses include the appropriate longitudinal child-level weight (ECLS-K source variable: C67CW0).

Measures

Grade 8 math achievement

While much research has examined the relationship between schools’ middle-level grade span and adolescents’ noncognitive outcomes, this study exclusively focuses on an important cognitive outcome, specifically mathematics achievement. The dependent variable is students’ scores on the ECLS-K math assessment administered in the spring eighth grade. Achievement in mathematics was chosen for three reasons. First, this is a subject area identified as a priority by the No Child Left Behind Act of 2002. Second, it is particularly sensitive to school-based instruction (Burris, Heubert, & Levin, 2006). Finally, math is a subject area that is widely considered to be a critical “gateway” to secondary and postsecondary success (Matthews & Farmer, 2008).

The ECLS-K math assessment maximized the accuracy of measurement that could be achieved in a limited amount of testing time while minimizing floor and ceiling effects by matching sets of test questions to initial estimates of students’ achievement (Tourangeau et al., 2009). The test’s specifications were derived from national and state performance standards. Broad-based scores using the full set of assessment items were calculated using IRT procedures. The IRT scale scores estimated children’s performance on the whole set of assessment questions, while standardized scores reported children’s performance relative to their peers on the content domains. The standardized score, which is used in these analyses, provides a norm-referenced measurement of achievement; that is, estimates of achievement relative to the population as a whole. Checks on the reliability and validity of the math assessment are reported in Najarian, Pollack, Sorongon, and Hausken (2009).

Middle-level grade span

The primary covariate of interest is the grade span of students’ spring eighth-grade school. These variables were constructed from two items on the spring eighth-grade administrator’s survey, which asked respondents to indicate their school’s lowest and highest grade. Responses on these two items were collapsed into five nonoverlapping categories that reflect the landscape of middle-level grade configurations. These grade span categories include: (a) K-8 schools with at least one grade lower than fifth and no grade higher than eighth; (b) Grade 6 to 8 schools, which includes schools with no grade lower than sixth and not higher than eighth; (c) Grade 7 to 8 schools with no grade lower than seventh and no grade higher than eighth; (d) grades 7 to 12 includes those schools with no grade lower than seventh and at least one grade higher than eighth; and (d) K-12 schools that have at least one grade lower than seventh and one grade higher than tenth. Grades 6 to 8 schools, the modal category, serve as the reference group.

Math classroom quality

We include four student-level measures derived from the students’ math teacher questionnaire that reflect important characteristics related to the quality of students’ spring eighth-grade math classrooms (Mayer, Mullens, Moore, & Ralph, 2000). The first, algebra, is an indicator for whether the student’s math class is taught at or above the level of algebra (1 = yes, 0 = no). Second, class time per week, is a set of three indicators (<3, 3-4.9, and ≥5) that report on the number of hours the class meets for instruction per week, with the category 3.0 to 4.9 hours/week serving as the reference group. Next, class behaves well or extremely well, is an indicator collapsed from a set of four possible responses (1 = yes, 0 = no). Finally, rigor is a standardized composite score (5 items, α = .67, M = 0, SD = 1) that captures the frequency with which students do group work, discuss math, write about math, relate math to “real” life, and encounter math problems with no solutions. Higher scores represent higher rigor.

Covariates

Several student- and school-level measures employed as statistical controls to reduce selection bias inherent to observational studies. At the student-level these measures include socioeconomic status (SES), a composite of parents’ income, education, and occupational prestige (a z-score [M = 0, SD = 1]). The analyses also employ a dummy-coded gender measure (girls = 1, boys = 0) and a measure indicating whether the child was a member of a historically higher performing racial/ethnic group (White and Asian children = 1, Hispanic, African American, Native American and multiracial children = 0). In addition, the models also include a measure of whether the child lived in a single-parent home (yes = 1, no = 0); and whether the student was a nonnative English speaker (yes = 1, no = 0). The final two student-level control measures are related to students’ academic background, indicating whether the child has a disability (yes = 1, no = 0) and is below expected grade level (yes = 1, no = 0).

At the school-level, the models employ measures that control for schools’ sociodemographic and academic characteristics. School size and cohort size are especially noteworthy. First, a series of indicator variables are used to measure schools’ total enrollment: small schools (<150 children); small-medium schools (150-299); medium-sized schools (300-499); medium-large schools (500-749) and; large schools (>750). In the analyses that follow, medium-sized schools are the reference category. Second, cohort size is measured using an indicator variable for whether there are more than 180 students in the eighth grade (yes = 1, no = 0). The school-level models also incorporate sociodemographic controls for high-minority enrollment (a dummy variable indicating Hispanic or African American enrollments at or above 25%) and a standardized measure for the percentage of students receiving free lunch (a z-score [M = 0, SD = 1]).

Analytic Procedure

Multilevel models with random intercepts are used to estimate the relationship among these variables and account for the nested nature of the data. They also include controls for the ECLS-K stratified sampling design and for the probability of selection for individuals. These models include a common set of predictor variables, consisting of students’ and schools’ social and academic characteristics. The models also take advantage of the panel design of the ECLS-K study by controlling for respondents’ status on the dependent variable in the spring fifth grade in predicting their spring eighth-grade outcome. This control not only allows more precise specification of the relationships between grade span and math achievement, but also provides a control for any relationships that might have resulted from previous school experiences. The expectation maximization (EM) algorithm was used to account for missing data (Dempster, Laird, & Rubin, 1977), which estimates missing values by using the otheximum r nonmissing values in the data set (Allison, 2002).

The model building process follows the study’s three research questions. First, descriptive statistics on all covariates were examined to uncover any imbalance in adolescents’ likelihood of attending one middle-level grade span versus another. Second, multilevel models were built from the first level up to predict math achievement from schools’ middle-level grade span. Third, measures related to students’ math classroom quality were incorporated into the models. Fourth, gender and family structure (i.e., single parent) were tested as moderators of the influence different middle school grade spans had on math achievement. Estimates derived from models in steps two through four use maximum likelihood estimation (MLE) to compare the fit across models with varied fixed effects. Model fit was assessed using the BIC and AIC indices, with lower values indicating a better fit. Finally, likelihood-ratio tests were performed to compare nested models fitted with MLE.

Results

Descriptive Results

Descriptive statistics for all analytical variables are presented in Table 1. Three points are of interest. First, differences in average eighth-grade math scores show that students in middle schools score higher than do those in other school forms. Students in K-8 and K-12 schools have the lowest mean scores for both the spring fifth- and eighth-grade assessments in math and reading; however, these differences may be related, at least in part, to compositional differences. Second, compared to either the Grades 6 to 8 or Grades 7 to 8 schools, Grades K-8 schools have more students eligible for free lunch, and are also likely to be smaller in total enrollment. Third, mirroring these school-level descriptive statistics, children in K-8 schools have lower mean SES scores and are more likely to have a disability. Compared to their counterparts in the Grades 6 to 8 or Grades 7 to 8 schools, they are also more likely to come from single parent homes. We emphasize, however, that these three points are not attributable to statistically significant differences.

Table 1.

Student- and School-Level Descriptive Statistics by Middle Level Grade Span (N = 2,729).

	K-8 %/M (SD)	6-8 %/M (SD)	7-8 %/M (SD)	7-12 %/M (SD)	K-12 %/M (SD)	Overall %/M (SD)
N (%)	8.9	58.6	19.0	8.9	4.6
Student characteristics
Math test score Grade 8	49.14 (9.87)	51.39 (9.64)	52.45 (8.99)	52.68 (9.18)	48.66 (10.01)	51.39 (9.60)
Math test score Grade 5 (z)	−0.25 (0.99)	0.01 (1.02)	0.09 (0.95)	0.11 (0.94)	−0.19 (1.04)	—
SES (z)	−0.27 (0.87)	0.04 (1.02)	0.06 (1.02)	−0.01 (0.92)	−0.32 (0.94)	—
Nonnative English	12.3	15.5	15.2	8.8	11.6	14.1
Female	51.6	49.2	51.6	49.5	41.7	49.4
Below expected grade level	14.0	9.5	8.2	4.5	18.3	9.9
Disability	15.4	15.2	12.6	12.1	18.1	14.8
Single Parent	26.8	22.5	23.2	20.4	20.8	22.2
White/Asian	72.4	66.6	64.2	81.4	80.0	69.1
Math classroom quality
Rigor (z)	0.06 (0.94)	−0.04 (0.99)	0.05 (0.10)	−0.02 (0.86)	0.56 (0.95)	—
At or above algebra	40.0	38.8	39.0	38.6	29.2	39.0
Class behaves well/extremely well	56.1	55.9	53.1	57.5	58.9	55.5
Class time per week
< 3 hours/wk.	1.0	1.0	0.0	0.0	0.0	1.0
3-4.9 hours/wk.	74.7	65.0	78.5	93.5	61.1	70.2
≥5 hours/wk.	24.4	34.0	21.1	6.5	38.9	29
School characteristics
Free Lunch (z)	0.25 (0.98)	0.01 (1.06)	−0.07 (0.88)	−0.24 (0.74)	−0.05 (1.00)	—
High Minority	29.9	34.0	39.5	16.0	16.5	31.3
Cohort Size >180	8.8	78.2	88.1	61.0	3.2	66.7
Total Enrollment
<150	7.2	1.0	0.0	2.7	5.3	2.0
150-299	18.6	3.0	4.2	4.5	13.2	6.2
300-499	34.8	17.6	24.3	25.0	18.4	21.0
500-749	29.0	24.8	33.8	5.0	28.9	26.0
>750	10.4	53.9	37.4	62.7	34.2	44.9

Note. Standard deviations reported only for continuous variables.

The descriptive statistics for this nationally representative dataset suggest that the five different grade span configurations serve different types of students and also vary on critical school-level characteristics. Therefore, subsequent multilevel analyses must control for these sociodemographic and academic background characteristics.

Grade Span on Math Achievement

Table 2 presents four different multilevel models examining various predictors influence on eighth-grade students’ math achievement. Model 1 is the unconditional model with no predictors at either the student or school level. The average student in the average school had an average eighth-grade score of 51.10. The derived intraclass correlation from the math model is 16%, which is in line with estimates ranging from 10% to 30% in the school-effects literature and substantiating the use of multilevel models (Bryk & Raudenbush, 1992).

Table 2.

Multilevel Model Predicting Students’ Eighth Grade Math Achievement From Schools’ Middle-Level Grade Span (N = 2,729).

	Model 1	Model 2	Model 3	Model 4
Constant	51.10 (0.24)	51.13 (0.27)	52.10 (0.41)	52.07 (0.45)
Fixed effects
Math test score Grade 5		7.63 (0.13)***	7.58 (0.13)***	7.61 (0.13)***
Grade span
K-8				−0.69 (0.52)
7-8				0.27 (0.33)
7-12				0.11 (0.46)
K-12				−1.62 (0.75)*
Student-level controls^a	No	Yes	Yes	Yes
School-level controls^b	No	No	Yes	Yes
Random effects
SD (school-level)	3.80 (0.28)	1.50 (0.16)	1.41 (0.16)	1.32 (0.18)
SD (student-level)	8.84 (0.14)	4.86 (0.08)	4.85 (0.08)	4.83 (0.08)
AIC	19,987	14,429	13,743	12,556
BIC	20,005	14,493	13,846	12,680
LR χ^2c		5,573.73***	700.58***	1,194.91***

Note. Values in parentheses are standard errors.

Student-level covariates include SES (z) and indicators for nonnative English speaker, disability, single parent, White/Asian, below expected grade level, and female (1 = yes, 0 = no). ^bSchool-level covariates include percentage free lunch (z) and indicators for high minority, cohort size >180 (1 = yes, 0 = no), and total school enrollment (medium-sized schools are the referent category). ^cχ² value for the likelihood-ratio (LR) test. Significant results indicate an improvement in fit from the previous model.

p < .10. ** p < .01. *** p < .001. (two-tailed tests).

Models 2 and 3 incorporate student- and school-level controls, with an emphasis on keeping the models parsimonious and including only those controls that are statistically significant. Model 3 shows that a one standard deviation increase in Grade 5 math test score is associated with a three-quarters standard deviation increase in Grade 8 math test score (b = 7.58, z = 58.20, p < .001).

Model 4 as reported in Table 2 directly addresses the first primary research question and includes student- and school-level controls. Specifically, the results show that including grade span as a predictor (Model 4) had a statistically significant better fit than the previous model containing student- and school-level controls (Model 3), LRχ² (4, N = 2,729) = 1,194.91, p < .001. Aside from the statistically significant negative coefficient for the K-12 grade span indicator, the results (see Table 2) suggest no consistent benefit to eighth-grade students from any grade span configuration compared to the Grade 6 to 8 span configuration.

Math Classroom Quality on Math Achievement

Table 3, Models 5 and 6, incorporates four measures of students’ math classrooms, two of which are significant predictors of students’ Grade 8 math test scores. Model 5 shows that students in a class that is at or above algebra are expected to score 1.48 points higher than students not enrolled in algebra (b = 1.48, z = 0.57, p < 0.001). This model also had a statistically significant better fit than Model 3, LR (5, N = 2,729) = 290.96, p < .001. This estimate for at or above algebra remains significant in the following model, which incorporates the grade span indicator variables. In addition, the indicator for whether students’ math class behaves well or extremely well is a significant predictor in both models. The point estimate for this covariate in Model 6 is associated with about 0.08 standard deviations increase in Grade 8 math test scores (.72/9.60, where 9.60 is the standard deviation of Grade 8 math test score). In addition, Model 6 is a significantly better fit than Model 5, LRχ² (4, N = 2,729) = 1,148.34, p < .001.

Table 3.

Multilevel Model Predicting Students’ Eighth Grade Math Achievement From Schools’ Middle Level Grade Span and Math Class Characteristics (N = 2,729).

	Model 5	Model 6
Constant	51.06 (0.44)	50.98 (0.49)
Fixed effects
Math test score Grade 5	7.18 (0.14)***	7.23 (0.15)***
Math classroom quality
Rigor	−0.20 (0.11)*	−0.19 (0.12)
At or above algebra	1.48 (0.26)***	1.41 (0.27)***
Class behaves well/extremely well	0.74 (0.23) ***	0.72 (0.24)**
Class time per week
< 3 hours/wk.	0.19 (1.26)	0.05 (1.29)
≥5 hours/wk.	0.05 (0.27)	0.20 (0.28)
Grade span
K-8		−0.75 (0.52)
7-8		0.42 (0.34)
7-12		0.27 (0.47)
K-12		−1.39 (0.74)*
Student-level controls^a	Yes	Yes
School-level controls^b	Yes	Yes
Random effects
SD (school-level)	1.58 (0.16)	1.49 (0.17)
SD (student-level)	4.75 (0.08)	4.74 (0.08)
AIC	13,462	12,321
BIC	13,593	12,473
LR χ^2c	290.96***	1,148.34***

Note. Values in parentheses are standard errors.

Student-level covariates include SES, and indicators for nonnative English speaker, disability, single parent, White/Asian, below expected grade level, and female (1 = yes, 0 = no). ^bSchool-level covariates include percentage free lunch (z) and indicators for high minority, cohort size >180 (1 = yes, 0 = no), and total school enrollment (medium-sized schools are the referent category). ^cχ²value for the likelihood ratio (LR) test. Significant result for Model 5 indicates an improvement in fit from Model 3. Result for Model 6 indicates a significant improvement in fit from Model 4.

p < .10. ** p < .01. *** p < .001. (two-tailed tests).

Gender and Family Structure as Moderators

Table 4, Model 7 includes interaction terms between gender and grade span to test whether the one grade span configuration or another is better for females. The results indicate that females in the Grades 7 to 12 configuration are expected to score 1.2 points less than males in Grade 6 to 8 schools (the referent category). However, caution is warranted when interpreting this result, as it is just shy of conventional statistical significance (b = 1.21, z = −1.58, p < 0.115, two-tailed). This finding hints that when females transition at a later point into middle school (Grade 7) and when that school has older students at or above Grade 9, then they may be likely to experience declines in math performance. Both the later transition and the presence of older students may exacerbate the depression and hostility that affects females more so than males at this time (Hirsch & Rapkin, 1987). At the very least, this null result suggests that future research should further evaluate how the presence of older peers may adversely influence females. In addition, the inclusion of these interaction terms did not improve fit from the previous model (Model 6), LRχ² (4, N = 2,729) = 4.30, p = .367.

Table 4.

Multilevel Model Predicting Students’ Eighth Grade Math Achievement From Interactions Between Schools’ Middle Level Grade Span and Select Student Characteristics (N = 2,729).

	Model 7	Model 8
Constant	50.89 (0.50)	50.88 (0.49)
Fixed Effects
Math test score Grade 5	7.22 (0.14)***	7.22 (0.14)***
Female	0.64 (0.29)*	0.51 (0.22)*
Single parent	−0.36 (0.28)	0.11 (0.36)
Math classroom quality
At or above algebra	1.40 (0.27)***	1.41 (0.27)***
Class behaves well/extremely well	0.74 (0.24)**	0.73 (0.24)**
Grade span
K-8	−0.24 (0.68)	−0.25 (0.57)
7-8	0.26 (0.45)	0.68 (0.37)*
7-12	0.90 (0.62)	0.46 (0.51)
K-12	−1.33 (0.90)	−1.46 (0.79)*
Interaction terms
K-8 × female	−0.94 (0.81)
7-8 × female	0.30 (0.57)
7-12 × female	−1.21 (0.77)
K-12 × female	−0.11 (1.11)
K-8 × single parent		−1.96 (0.94)*
7-8 × single parent		−1.16 (0.70)*
7-12 × single parent		−0.84 (0.97)
K-12 × single parent		0.75 (1.47)
Student-level controls^a	Yes	Yes
School-level controls^b	Yes	Yes
Random Effects
SD (school-level)	1.47 (0.17)	1.48 (0.17)
SD (student-level)	4.74 (0.08)	4.73 (0.08)
AIC	12,325	12,322
BIC	12,499	12,497
LR χ^{2 c}	4.30	6.89

Note. Values in parentheses are standard errors.

Student-level covariates include SES, rigor, and indicators for nonnative English speaker, disability, White/Asian, below expected grade level (1 = yes, 0 = no), and class time per week (3.0-4.9 hours/week serving as the referent). ^bSchool-level covariates include percentage free lunch (z) and indicators for high minority, cohort size >180 (1 = yes, 0 = no), and total school enrollment (medium-sized schools are the referent category). ^cχ²value for the likelihood ratio (LR) test. Model 6 nested in Models 7 and 8.

p < .10. ** p < .01. *** p < .001. (two-tailed tests).

Table 4, Model 8 includes an interaction between the single parent and grade span indicators. The math scores of students with single parents in K-8 schools are predicted to decrease (b = −1.96, z = −2.08, p = 0.037), as are those students with single parents in Grades 7 to 8 schools (b = −1.16, z = −1.66, p = 0.096). However, similar to the previous model, Model 8 is not a significantly better fit than Model 6, which does not contain any interaction terms, LRχ² (4, N = 2,729) = 6.89, p = .142. The conclusion to be drawn from both Models 7 and 8 is that, while both offer clues on the moderating effect of family structure and gender, neither model improves upon the main effects presented in Model 6.

Alternative Model Specifications

To determine whether the results were specific to the model’s specifications, we also employed an alternative modeling strategy for the full models (Models 4-8), using a two-stage least squares (2SLS) regression that accounted for the potential endogeneity of attending a K-8 school in the eighth grade. Any endogeneity is likely the result of omitted variables due to self-selection (Woolridge, 2002), the most common problem in the social and behavioral sciences (Vella, 1998). Therefore, we instrumented on this indicator using the grade span of students’ fifth-grade schools as the instrument (an indicator for whether the student attended a combined school, i.e., K-8 or K-12). This instrument was consistently shown to be both strong and relevant (based on the first-stage F-statistic) across the five different models. In addition, this instrument conforms to the assumption that it is correlated with the middle-level grade span of adolescents’ eighth-grade schools with little reason to believe that it also affects eighth-grade math test scores. Using this strategy, the coefficients of the grade span indicators and classroom quality variables were substantively similar to those derived from the HLM estimates reported in Tables 2 to 4.

Discussion

While these results do not corroborate recent findings regarding the benefits of Grade K-8 schools, the study’s results are validated by three elements of its design. First, the sample was both large and nationally representative and included detailed information on students, classrooms, and schools, which contrasts with other studies and their dependence on administrative data or nonrepresentative samples. Second, we compared same-grade adolescents in schools with different middle-level grade configurations while controlling for a number a possible confounds. Third, the data included teacher reports of classroom quality to complement the data provided by the schools’ administrator, students, and their parents. Given the strengths of the design and the confirmatory results from 2SLS models, we are confident that these results question the supposed adverse effects of middle schools and encourage researchers and policy makers to refocus attention on classroom quality.

This research has focused on a small number of indicators related to the quality of students’ subject-specific classroom. By doing so, this research makes two contributions to a broader conversation on middle grades schooling and adolescent development. First, this work reinforces and extends previous discussions (Eccles & Roeser, 1999, 2009) on the ecological components of school systems and their influence on adolescent development. Eccles and Roeser (1999, 2009) parse the complex hierarchy of school systems into the classroom level and the school/district level. At the classroom level, several nuanced components, such as instruction, emotional support, management and motivational climate, teacher beliefs, efficacy, and expectations have proximal predictive power on student outcomes. At the school/district level, attributes such as overall school climate, school size, curricular differentiation, and middle-grade span/transitions are thought to be largely predictive as well. However, in specifically examining middle-level grade span while considering the developmental needs of adolescents, this study suggests that decade-long efforts made to change school/district level components, such as grade span of schools, are likely better spent at the point-of-instruction—the classroom level.

Second, this work encourages researchers to refocus attention on the classroom-level “best practices” that were originally thought to be the main advantage of middle schools. The school and its grade span may be too large of an aggregate category to detect any meaningful effects. Given that the classroom is the primary arena through which students engage subject matter, this is the level where adolescents need to have the most appropriate developmental fit. It is there where factors such as maintaining an intense focus on academic achievement, proactive intervention, and teacher competencies matter most for adolescents. Incidentally, these factors, as well as a small number of others, are those that are associated with the highest performing schools with middle-level grades, regardless of their grade span configuration (Williams, Kirst, & Haertel, 2010). This requires researchers to move beyond school-level factors and toward classroom-level instructional practices that can be adapted to suit the varied developmental needs of adolescents.

While shifting attention away from large-scale structural factors, it is also important to consider just how much an effect, on average, one can expect from such changes. While Rockoff and Lockwood (2010) report sizable positive effects of K-8 schools, Byrnes and Ruby’s (2007) estimates are much more modest. While the results reported here generally found no such advantage, any differences between students had little to do with schools. Rather, much of the difference can be attributed to student-level factors. Specifically, Model 1, which includes no explanatory variables at either level, shows that about 84% of the variation in eight-grade math scores can be attributed to the student-level (3.80² + 8.84² = 92.59 [total variance], 8.84²/92.59 = 84.40%). This finding supports Byrnes and Ruby’s (2007) call for a more reasonable expectation as to what effects school systems and broad school-based reforms can have on student achievement.

While recognizing the contributions of this study, caution is warranted when interpreting results. Two limitations are noteworthy. First, this study focused exclusively on a cognitive outcome. Historically, much of the research on middle schools and adolescent development has focused on noncognitive outcomes such as self-perceptions, motivation, and others. Therefore, it is important to note that the lack of relationship between grade span and the achievement measures does not extend these noncognitive outcomes. For example, the benefits of K-8 schools reported by Weiss and Kipnes (2006) centered on outcomes such as self-esteem and whether one was the target of a threat, which may also be equally important for adolescents from a policy perspective. Here, too, future analyses should consider the relationship between grade span and these important outcomes.

Second, students are not randomly assigned to schools in the ECLS-K, and so these data have the same potential selection bias as all other observational studies. To limit the magnitude of this bias, this study employs the standard strategy of using control variables that have been associated with students’ academic achievement in previous research. In addition, the robustness checks using a 2SLS regression with instrumental variables also addressed this potential bias. As with all analyses based on observational data (and even for some studies based on randomized experimental data), caution must be exercised in interpreting any significant relationships as causal; it is through the accumulation of similar estimates from studies with varying data and alternative methodologies that causal conclusions become substantiated.

These limitations notwithstanding, there are three immediate implications to be derived from this research. First, the results should give pause to reformers who are considering whole-scale changes to the ways in which the middle-level grade spans are organized. Reforms such as these are very costly, and their effects are not uniformly beneficial—nor may they be beneficial in the aggregate. At best, large districts that can provide a number of middle-level grade span configurations for adolescents should do so in a deliberate manner that contributes to a more consistent evidentiary base. The results reported here, however, question whether districts should move beyond these relatively small pilot efforts and adopt the K-8 model on a larger scale.

Second, this study highlights the importance of accounting for classroom-level characteristics when examining school effects. Eccles (2004) posited that classroom and school characteristics serve as proximal factors influencing students’ outcomes, mediating the more distal effects of middle-level grade span configuration. This work corroborates this perspective and contrasts with a number of rigorous studies that find the transition to middle school is responsible for performance declines. However, after accounting for both school and classroom characteristics, these results show that instructional characteristics matter, whereas grade span does not. These results do not stand entirely alone, as prior studies have also reached similar conclusions (Carolan & Chesky, 2012; Eccles et al., 1993; Holas & Huston, 2012).

Finally, these results should further remind reformers that the K-8 model, or other grade span models that serve adolescents, is not a one-size-fits-all solution. School-level reform must be carefully constructed in each locale to reflect both individual capacity and needs. Because of the difficult and different challenges that adolescents must confront, creating the right fit for these students at a developmentally tumultuous time requires flexibility that few districts can afford to provide. While certainly not suggesting that all districts convert their middle school from one configuration to another, it may be that districts that can afford to provide a range of different configurations would be well positioned to offer these options in a way that best match students’ characteristics. This would require research to identify how the effects of schools’ middle-level grade spans vary across different types of students. Regardless, the majority of adolescents in the United States will attend a Grades 6-8 middle school and although the results of this study need to be confirmed and extended, they remind researchers and policy makers of the importance of classroom quality, particularly as they relate to adolescents’ cognitive outcomes.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

Author Biographies

Brian V. Carolan, PhD is an associate professor of quantitative research methods in the College of Education and Human Services at Montclair State University. His research has focused primarily on the social organization of classrooms and schools. Areas of specialty include sociology of education and quantitative research methods.

Christopher C. Weiss, PhD directs the Applied Quantitative Research (AQR) MA Program in the Department of Sociology at New York University, where he is also an affiliate of the Institute of Human Development and Social Change and of the Center for Advanced Social Science Research. His primary research interests focus on three distinct areas: the influence of organizations and institutions on children and adolescents; how features of urban environments influence physical and mental health; and the causes and consequences of criminal justice and incarceration policies in the United States.

Jamaal S. Matthews, PhD is an assistant professor of educational psychology in the College of Education and Human Services at Montclair State University. He received his PhD. in Education and Psychology from the University of Michigan. His scholarly interests involve the development of self-regulatory and socio-emotional skills during childhood, as well as the relationship between academic identity development, achievement motivation and self-regulated learning during adolescence.

References

Allison

(2002). Missing data. Thousand Oaks, CA: SAGE.

Akos

Galassi

J. P.

(2004). Gender and race as variables in psychosocial adjustment to middle and high school. Journal of Educational Research, 98, 102-108. doi:10.3200/JOER.98.2.102-108

Alspaugh

J. W.

(2001). Achievement loss associated with the transition to middle school and high school. Journal of Educational Research, 92, 20-25. doi:10.1080/00220679809597572

Barber

B. K.

Olsen

J. A.

(2004). Assessing the transitions to middle and high school. Journal of Adolescent Research, 19, 3-30. doi:10.1177/0743558403258113

Blyth

D. A.

Simmons

R. G.

Bush

(1978). The transition into early adolescence: A longitudinal comparison of youth in two educational contexts. Sociology of Education, 51(3), 149-162. doi:10.2307/2112661

Bryk

A. S.

Raudenbush

S. W.

(1992). Hierarchical linear models. Newbury Park, CA: SAGE.

Byrnes

Ruby

(2007). Comparing achievement between K-8 and middle schools: A large-scale empirical study. American Journal of Education, 114, 101-135. doi:10.1086/520693

Burris

C. C.

Heubert

J. P.

Levin

H. M.

(2006). Accelerating mathematics achievement using heterogeneous grouping. American Educational Research Journal, 43(1), 105-136. doi:10.3102/00028312043001105

Carolan

B. V.

Chesky

N. Z.

(2012). The relationship among grade configuration, school attachment, and achievement. Middle School Journal, 43(4), 32-39.

10.

Cook

P. J.

MacCoun

Muschkin

Vigdor

(2008). The negative impacts of starting middle school in sixth grade. Journal of Policy Analysis and Mangement, 27, 104-121. doi:10.1002/pam.20309

11.

Cuban

(1992). What happens to reforms that last? The case of the junior high school. American Educational Research Journal, 29, 227-251.

12.

Deci

E. L.

Ryan

R. M.

(1994). Promoting self-determined education. Scandinavian Journal of Educational Research, 38, 3-14. doi:10.1080/0031383940380101

13.

Demo

D. H.

Acock

A. C.

(1996). Family structure, family process, and adolescent well being. Journal of Research on Adolescence, 6, 457-488.

14.

Dempster

A. P.

Laird

N. M.

Rubin

D. B.

(1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society. Series B (Methodological), 39(1), 1-38.

15.

Eccles

J. S.

(2004). Schools, academic motivation, and stage-environment fit. In Lerner

R. M.

Steinberg

(Eds.). Handbook of adolescent psychology (2nd ed., pp. 125-153). Hoboken, NJ: Wiley.

16.

Eccles

Lord

Midgely

(1991). What are we doing to early adolescents? The impact of educational contexts on early adolescents. American Journal of Education, 99, 521-542. doi:10.1086/443996

17.

Eccles

J. S.

Roeser

(1999). School and community influences on human development. In Bornstein

Lamb

(Eds.), Developmental psychology: An advanced textbook (4th ed., pp. 503-554). Mahwah, NJ: Lawrence Erlbaum.

18.

Eccles

J. S.

Roeser

R. W.

(1999/2009). Schools, academic motivation, and stage-environment fit. In Lerner

R. M.

Steinberg

(Eds.), Handbook of adolescent psychology (3rd ed., pp. 404-434). Hoboken, NJ: Wiley.

19.

Eccles

J. S.

Wigfield

Midgley

Reuman

(1993). Negative effects of traditional middle schools on students’ motivation. Elementary School Journal, 93, 553-574. doi:10.1086/461740

20.

Feldlaufer

Midgley

Eccles

J. S.

(1988). Student, teacher, and observer perceptions of the classroom before and after the transition to junior high school. Journal of Early Adolescence, 8, 133-156. doi:10.1177/0272431688082003

21.

French

S. E.

Seidman

Allen

Aber

J. L.

(2006). The development of ethnic identity during adolescence. Developmental Psychology, 42(1), 1-10. doi:10.1037/0012-1649.42.1.1

22.

Friedel

J. M.

Cortina

K. S.

Turner

J. C.

Midgley

(2010). Changes in efficacy beliefs in mathematics across the transition to middle school: Examining the effects of perceived teacher and parent goal emphases. Journal of Educational Psychology, 102, 102-114. doi:10.1037/a0017590

23.

Giordano

P. C.

(2003). Relationships in adolescence. Annual Review of Sociology, 29, 257-281. doi:10.1146/annurev.soc.29.010202.100047

24.

Grills-Taquechel

A. E.

Norton

Ollendick

T. H.

(2010). A longitudinal examination of factors predicting anxiety during the transition to middle school. Anxiety, Stress & Coping: An International Journal, 23, 493-513.

25.

Hanushek

E. A.

Kain

J. F.

Rivkin

S. G.

(2004). Disruption versus tiebout improvement: The costs and benefits of switching schools. Journal of Public Economics, 88, 1721-1746. doi:10.1016/S0047-2727(03)00063-X

26.

Hines

M. T.

III. (2007). Adolescent adjustment to the middle school transition: The intersection of divorce and gender in review. RMLE Online: Research in Middle Level Education, 31(2), 1-15.

27.

Hirsch

B. J.

Rapkin

B. D.

(1987). The transition to junior high school: A longitudinal study of self-esteem, psychological symptomatology, school life, and social support. Child Development, 58, 1235-1243. doi:10.2307/1130617

28.

Holas

Huston

A. C.

(2012). Are middle school harmful? The role of transition timing, classroom quality, and school characteristics. Journal of Youth and Adolescence, 41, 333-345. doi:10.1007/s10964-011-9732-9

29.

Hough

D. L.

(2005). The rise of the “Elemiddle” schools: Not every K-8 school truly applies best middle-level practices and deserves the designation. School Administrator, 62(3), 10-13.

30.

Juvonen

V.-N.

Kaganoff

Augustine

C. H.

Constant

(2004). Focus on the wonder years: Challenges facing the American middle school. Santa Monica, CA: RAND Corporation.

31.

Mac Iver

M. A.

Mac Iver

D. J.

(2006). Which bets paid off? Early findings on the impact of private management and K-8 conversion reforms on the achievement of Philadelphia students. Review of Policy Research, 23, 1077-1093. doi:10.1111/j.1541-1338.2006.00250.x

32.

Maehr

M. L.

Midgley

(1996). Transforming school cultures. Boulder, CO: Westview Press.

33.

Matthews

M. S.

Farmer

J. L.

(2008). Factors affecting the algebra I achievement of academically talented learners. Journal of Advanced Academics, 19, 472-501.

34.

Mayer

D. P.

Mullens

J. E.

Moore

M. T.

Ralph

(2000). Monitoring school quality: An indicators report (NCES 2001–030). Washington, DC: National Center for Education Statistics.

35.

Najarian

Pollack

J. M.

Sorongon

A. G.

Hausken

E. G.

(2009). Early childhood longitudinal study, kindergarten class of 1998-99 (ECLS-K) psychometric report for the eighth grade. Washington, DC: National Center for Education Statistics.

36.

Poncelet

, & Metis Associates (2004). Restructuring schools in Cleveland for the social, emotional, and intellectual development of early adolescents. Journal of Education for Students Placed at Risk, 9(2), 91-96. doi:10.1207/s15327671espr0902_1

37.

Rashmi

Melanson

Levin

(2007). Educational aspirations of male and female adolescents from single-parent and two biological parent families: A comparison of influential factors. Journal of Youth and Adolescence, 36, 1010-1023. doi:10.1007/s10964-006-9137-3

38.

Rockoff

J. E.

Lockwood

(2010). Stuck in the middle: Impacts of grade configurations in public schools. Journal of Public Economics, 94, 1051-1061. doi:10.1016/j.jpubeco.2010.06.017

39.

Rudolph

K. D.

Lambert

S. F.

Clark

A. G.

Kurlakowsky

K. D.

(2001). Negotiating the transition to middle school: the role of self-regulatory processes. Child Development, 72, 929-946. doi:10.1111/1467-8624.00325

40.

Seidman

Aber

J. L.

Allen

French

S. E.

(1994). The impact of school transitions in early adolescence on the self-system and social context of poor urban youth. Child Development, 65, 507-522. doi:10.2307/1131399

41.

Skinner

Marchand

Furrer

Kindermann

(2008). Engagement and disaffection in the classroom: Part of a larger motivational dynamic? Journal of Educational Psychology, 100, 765-781. doi:10.1037/a0012840

42.

Theriot

M. T.

Craun

S. W.

Dupper

D. R.

(2010). Multilevel evaluation of factors predicting school exclusion among middle and high school students. Children and Youth Services Review, 32(1), 13-19. doi:10.1016/j.childyouth.2009.06.009

43.

Tourangeau

Nord

Lê

Sorongon

A. G.

Najarian

(2009). Early Childhood Longitudinal Study, Kindergarten Class of 1998-99 (ECLS-K), combined user’s manual for the ECLS-K eighth-grade and K-8 full sample data files and electronic codebooks. Washington, DC: National Center for Education Statistics, Institute of Education Sciences.

44.

Vella

(1998). Estimating models with sample selection bias: A survey. Journal of Human Resources, 33(1), 127-169. doi:10.2307/146317

45.

Way

Reddy

Rhodes

(2007). Students’ perceptions of school climate during the middle school years: Associations with trajectories of psychological and behavioral adjustment. American Journal of Community Psychology, 40, 194-213. doi:10.1007/s10464-007-9143-y

46.

Weiss

C. C.

Baker-Smith

E. C.

(2010). Eighth grade school form and resilience in the transition to high school: A comparison of middle schools and K-8 schools. Journal of Research on Adolescence, 20, 825-839. doi:10.1111/j.1532-7795.2010.00664.x

47.

Weiss

C. C.

Kipnes

(2006). Reexamining middle school effects: A comparison of middle grades students in middle schools and K-8 schools. American Journal of Education, 112, 239-272. doi:10.1086/498996

48.

West

Schwerdt

(2012). The middle school plunge. Education Next, 12. Retrieved from http://educationnext.org/the-middle-school-plunge/

49.

Williams

Kirst

Haertel

(2010). Gaining ground in the middle grades: Why some schools do better: A large-scale study of middle grades practices and student outcomes Middle Grades Study. Mountain View, CA: EdSource.

50.

Woolridge

J. M.

(2002). Econometric analysis of cross section and panel data (3rd ed.). Mason, OH: Thomson-South Western.