Вы находитесь на странице: 1из 13

Economics of Education Review, Vol. 14, No. 3, pp.

229-241, 1995

,~De~smon Jt 0272--7757(95)00004--6

Elsevier Science Ltd Printed in Great Britain 0272-7757/95 $9.50+0.00

Does Class Size Matter?


KAREN AKERHIELM Mathtech, Inc., 202 Carnegie Center, Suite 111, Princeton, NJ 08540, U.S.A.

Abstract--This paper analyzes pupil-specific public school data unavailable in previous studies and applies instrumental variable econometric methods to account for nomrandom allocation of students to different class sizes and the endogeneity of the class size variable. By using better data and improved statistical techniques, this paper shows that there are returns to investing in smaller classes for certain students and it provides some evidence on why past literature has produced such inconsistent findings on the effects of class size. [JEL 121] I. INTRODUCTION 2. OVERVIEW OF THE LITERATURE

WHETHER OR NOT to reduce class size is one of the most contentious aspects of the debate on educational reform, mainly because past research has failed to show any consistent relationship between class size and student performance. Of a review of 112 studies on class size, only 23 found a statistically significant relationship between class size and student achievement and only fourteen of these exhibited a negative relation; the other nine studies showed that as class size increased, students performed better on tests, not worse. 'There is little apparent merit for schools to pursue their ubiquitous quest for lowered class sizes', (Hanushek, 1986, p. 1167). This paper contributes to the debate on class size by analyzing pupil-specific public school data unavailable in previous studies and by applying econometric methods that account for non-random allocation of students to different class sizes. By using better data and improved statistical techniques, this paper shows that there are returns to investing in smaller classes for certain students and it provides some evidence on why the past literature has produced such inconsistent findings.

While there is not consistent empirical evidence on the link between class size and student achievement, some studies have shown that small class size is important for certain types of students, such as lowachieving students, elementary school students, and students from low socio-economic backgrounds (Summers and Wolfe, 1977; ERS, 1986; Dolan and Schmidt, 1987). Other studies have analyzed the link between class size and future earnings, bypassing the immediate effects on student achievement. One recent study found that males who went to schools with low pupi]teacher ratios ended up with higher wages and that the closing over time of the class size gap between blacks and whites explains about one-fifth of the closure in the black-white wage gap between 1960 and 1980 (Card and Kreuger, 1992). There are two major reasons why past research has revealed such inconsistent results. First, due to missing data on a student's actual class size, past research often used the pupil-teacher ratio for an entire school as the indicator of class size (Coleman et al., 1966: Coleman et al., 1982; Chubb and Moe, 1990; Card

[Manuscript received 16 March 1994; revision accepted for publication 6 December 1994.] 229

230

Economics of Education Review

and Kreuger, 1992 all used aggregate ratios). To the extent that an aggregate ratio differs from the class size that a student was actually exposed to, measurement error exists, biasing the coefficient of the class size variable toward zero. In addition, pupil-teacher ratios are defined as the number of students in the school divided by the number of full-time teachers for an entire school and may have nothing to do with actual class size. This ratio often includes guidance counselors, principals, and special education teachers in the count of teachers, and thus the lower the ratio, the higher the non-teaching staff, regardless of actual class size. Moreover, even if the ratio only includes actual teachers, schools with the same pupil-teacher ratio may have significantly different class sizes depending on the average number of hours of teaching required (Bowles and Levin, 1968). The second problem is that, as opposed to using experimental (or randomized) data collected as part of a specific class size evaluation, past studies have often relied on data that were part of a larger national survey implemented for other purposes. As a result, previous research has used data in which student allocation to different class sizes may not be a random process. If a school has a deliberate policy to assign difficult or less able students to smaller classes, then any positive effect of small class size on student performance may be disguised because such students may tend to score lower on tests. Likewise, if a school has a policy of assigning advanced students to the smaller classes, then a positive relation between small class size and student performance may be overstated. In these examples, class size (and the allocation of students to various class sizes) is a choice variable and thus the estimation of the class size-student achievement relationship using ordinary least squares (OLS) will provide biased coefficients, Experimental studies in Indiana and Tennessee, in which students were randomly assigned to different class sizes and followed over time, found significant, positive effects of small classes on elementary school student achievement (McGiverin et al., 1989; Word et al., 1990). By showing the importance of small classes when there exists a random allocation of students to different class sizes, the results from experimental studies suggest that a non-random allocation may mask the true relationship between class size and student achievement,

3. THE MODEL The specification of the educational process used in this paper is based on the theoretical framework of an educational production function model in which the achievement (or output) of student i in subject j (as measured by test scores) is related to various family background, community, and school resource inputs (Hanushek, 1979; Link and Mulligan, 1991). The model can thus be written as: Y~j = et + 13~X~+ 132Wij+ 133Zij+ij (1)

where Yij is the test score for child i in subject j, X~ is a vector containing family and community characteristics for child i, Wij is characteristics of child i's teacher in subject j, Zij is child i's class size in subject j, and % is an unmeasured error component which includes inputs such as innate ability and motivation. The error term can also be thought of as 'unobserved test-taking ability'. The question this paper asks is, what is the effect Of class size on a student's achievement, holding various family, community, and teacher characteristics constant? Class size may affect achievement if student participation and satisfaction, student attendance, teaching practices and/or the amount of individualized attention received by the student varies for different class sizes (Smith and Glass, 1980; Cahen
et al., 1983). However, students may not be allocated randomly to different class sizes. Suppose a school has a delib-

erate policy to assign less able students to smaller classes. In such a situation, the assumption of Cov (Z,~) = 0 is violated when using OLS because the relationship between Z (class size) and ~ (unobserved test taking ability) is positive. As a result, 133 will be upwardly biased, falsely suggesting that as class size increases, students' cognitive skills increase as well. Similarly, if a school has a systematic policy to assign advanced students to the smaller classes, then a negative relationship between Z and ~ will downwardly bias the OLS relationship between class size and student performance, implying a stronger association between small classes and student achievement than is actually the case. The solution entails instrumenting class size, using a two-stage least squares (2SLS) approach, based on exogenous school-level variables as the instruments for class size. Thus, student achievement is estimated

Does Class Size Matter?

231

conditional on the process of non-random allocation of students to different class sizes as follows: Y~j = ot + 13tXi+ 132Wij+ 1337.ij+ % (2)

where Z is endogenous and thus must be predicted using exogenous variables that are related to Z and not with e. The major problem with this model specification is using data from one year only (in this case the eighth grade) which may not provide a sufficient control for initial innate family cation, ability. However, the correlation between abilities and class size, after controlling for background factors such as income and eduis likely to be small (Hanushek, 1979). 4. NELS DATA SET The data used in this analysis are from the U.S. Department of Education's National Education Longitudinal Study (NELS) of 1988. In this nationally representative sample, over 24,000 eighth graders from public and private schools throughout the U.S. were surveyed in 1988 and tested in four subjects (mathematics, English, science, and history). One parent and two teachers of each student (representing two of the four subjects) were also surveyed as was an administrator from each school. On average, each of the 1,052 participating schools was represented by 24 students and five teachers. The cognitive tests contained a total of 116 items to be completed in 85 minutes, 5. ESTIMATION RESULTS All estimation and testing was conducted for the four different subjects to see if class size effects differ by subject matter and to take advantage of NELS classroom-specific data. As Table 1 illustrates, the average class size and the pupil-teacher ratio for all four subjects, is larger for non-white students than for white students (not controlling for other characteristics). This is inconsistent with recent research that found similar pupilteacher ratios for whites and blacks (Boozer et al., 1992; Card and Kreuger, 1992). Urban non-whites had significantly larger classes than urban whites for all subjects except science. Overall, the average pupil-teacher ratio is about 18 or 19, consistently and substantially lower than average class size.

Table 1 also shows that lower ability students (as defined later in this paper) have smaller classes on average. When t-tests were constructed that compared the differences in class size means among each pair of subgroup of students, for all possible comparisons (except for urban non-white and urban white science students), the null hypothesis of similar class size means for the two subgroups was rejected.

OLS Regression Results Table 2 displays the results of running the OLS regression for all four subjects, using the model described by equation (1). This table also shows the means and standard deviations of each of the variables used. In all four subjects, the pupil's actual class size has a positive influence on achievement (as measured by test scores) while for three of the four st~bjects (all except science), the effect is significant (at the five percent level for math and English and at the ten percent level for history). Overall, using OLS shows that the relationship between class size and student achievement, after controlling for family, community, and teacher inputs, is positive and significant, suggesting that as class size increases, student performance increases as well. Spline Techniques Splines were used to determine if there are certain ranges for which changes in class size may not have an impact on test scores. Reviews of the class size literature (ERS, 1986; Glass and Smith, 1979) have often concluded that changes in class sizes have no effect on student achievement in the mid-ranges of class size (i.e., 20 to 30 students). Since there have been several definitions of the 'mid-range' of class size in the literature, this section examines two possible scenarios as shown in Table 3. In both scenarios, three class size 'spline' variables were included in the regression to represent the low, middle and high ranges of class size. While actual class size was the variable that represented the small range of class size, two new variables were created to represent the middle and large ranges. ~ As a result of this specification, the coefficient of the small-range variable (b~p~i,ej) represents the effect of each additional student in the small range of class size while the coefficients added together from the smallrange and the mid-range (bspllnel +bspline2)represent the effect of an additional student in the mid-range. Likewise bspJine I + bspline2 + bspline3 equals the effect on

232

Economics of Education Review


Table 1. Class size by type of public school student (N=4741) Average math class size 25.33 23.71 8.16 25.70 23.73 9.72 26.t9 25.02 3.33 24.78 21.39 13.19 (N=4741) Type of public school student Non-white White t-Test for differences in means Urban non-white Urban white t-Test for differences in means Average PIT Ratio 19.18 17.45 9.67 19.55 18.21 5.13 (N=4973) (N=4645) (N=4850) A v e r a g e A v e r a g e Average English science history class size class size class size 25.05 23.95 6.15 25.36 23.97 7.37 25.87 24.72 3.50 24.94 21.01 16.35 (N=4973) Average P/T Ratio 18.63 17.67 5.77 18.88 18.37 1.95 25.36 24.57 4.46 25.60 24.55 6.00 25.52 25.67 0.48 25.05 23.17 7.57 (N=4645) Average PtT Ratio 18.53 17.79 4.14 18.01 17.74 1.09 26.09 24.53 8.14 26.65 24.49 10.89 27.06 26.22 2.40 25.25 23.15 7.83 (N=4850) Average P/I" Ratio 18.73 17.73 6.08 19.14 18.21 3.55

Type of public school student Non-white White t-Test for differences in means Urban Non-urban t-Test for differences in means Urban non-white Urban white t-Test for differences in means Non-lower ability Lower ability t-Test for differences in means

Note: The t-tests are based on the assumption of separate variances. Assuming pooled variances yielded "t-tests of similar magnitudes. P/T = Pupil/Teacher.

achievement of one additional student in the large range of class size. Table 4 displays the spline results. For clarity's sake, the coefficients from the other variables are omitted from the table (there were no sign changes or significant differences of magnitude of the coefficients of the other variables as compared to Table 2). As shown at the bottom of the table, the effect of the mid-range is negative for both scenarios for history and science, while the positive coefficients in the lower ranges for math and English achievement swamp the negative coefficients in the mid-range to yield an overall positive effect. 2

Finally, the null hypothesis to be tested is: Ho "bspline2=bspline3= 0.

If Ho is accepted, this implies that the slope of the class size-achievement relation is constant and can be represented by the class size variable itself without the additional spline variables. In all eight cases (four subjects across two scenarios), the null hypothesis was rejected, implying that the effect of class size is not independent of the level of class size and that there are nonlinearities in the effects of class size.

Does Class Size Matter?


Table 2, OLS regression results by subject area Math subject (N = 4741) F=74R2=0.18 Std. err. of est. = 7.82 Mean of dep. var. = 22.64 English subject (N = 4973) F=71 R2=0.17 Std. err. of est. = 4.26 Mean of dep. var. = 12.95

233

Dependent variable = Test score (number right):'

Variable Constant Family income ($10,000's) Family income squared Family composition (1 = married) Parent educ. (1 =beyond H.S.) Sex of student (I = male) Race of student (1 = white) Preschool (1 = attended) % of single parent families in school Urban (1 = Yes) Race of teacher (1 = White)" Sex of teacher (1 = Male)" Years of experience* Years of experience squared* Class size"

Coeff. 8.83 1.00 -0.04 1.06


2.31 0.60 2.55 0.89 -2.09 -0.81 1.90 0.40 0.32 -0.01 0.12

t-Stat 10.86 9.27 -5.53 4.16


9.21 2.63 9.02 3.67 -3.21 -2.56 5.08 1.61 4.87 -3.85 6.25

Mean
3.71 0.65 0.61 0.49 0.72 0.50 0.29 0.23 0.87 0.44 14.62 24.17

Std. dev.
2.98 0.48 0.49 0.50 0.45 0.50 0.18 0.42 0.34 0.50 7.31 6.10

Coeff. 6.80 0.59 -0.02 0.41


1.25 -0.85 1.39 0.46 -1.66 -0.01 0.85 -0.37 0.06 -0.001 0.08

t-Stat 15.54 10.18 -6.73 3.03


9.49 -7.01 9.50 3.53 -4.95 -0.06 4.20 -2.61 1.52 -0.90 7.71

Mean
3.65 0.64 0.60 0.49 0.69 0.50 0.30 0.23 0.88 0.25 15.07 24.29

Std. dev.
2.83 0.48 0.49 0.50 0.46 0.50 0.19 0.42. 0.32 0.43 6.92 5.68

Dependent variable = Test score (number right)"

Science subject (N = 4645) F = 5 0 R2=0.13 Std. err. of est. = 4.20 Mean of dep. var. = 13.82

History subject (N = 4850) F = 5 5 Rz=o.14 Std. err. of est. = 5.02 Mean of dep. var. = 19.44

Variable Constant Family income ($10,000's) Family income squared Family composition (1 = married) Parent educ. ( I = beyond H.S.) Sex of student (I = male) Race of student (1 = white) Preschool (1 = attended) % of single parent families in school Urban (1 = Yes) Race of teacher (1 = white)" Sex of teacher (1 = male)" Years of experience" Years of experience squared" Class size"

Coeff. 8.52 0.47 -0.02 0.31


1.22 0.61 1.35 0.60 -1.31 0.04 1.06 0.19 0.10 -0.003 0.002

t-Stat 16.59 8.08 -5.33 2.25


9.01 4.97 8.90 4.53 -3.46 0.27 3.41 1.46 2.79 -2.48 0.13

Mean
3.89 0.64 0.62 0.48 0.74 0.51 0.28 0.21 0.96 0.57 13.85 24.78

Std. dev.
3.10 0.48 049 0.50 0.44 0.50 0.17 0.41 0.20 0.50 7.74 5.22

Coeff. 12.90 0.60 -0.02 0.48


1.59 0.39 1.17 0.77 0.55 -0.66 0.97 -0.11 0.06 -0.001 0.03

t-Stat 24.66 8.94 -5.82 2.92


9.86 2.67 6.50 4.97 1.20 -3.33 3.71 -0.71 1.50 -0.92 1.91

Mean
3.95 0.64 0.62 0.49 0.74 0.52 0.28 0.21 0.91 0.61 15.29 24.93

Std. dev.
3.29 0.48 0.49 0.50 0.44 0.50 0.17 0.40 0.29 0.49 7.54 5.73

* = These variables are subject specific. For example, in the math student regression, the race, sex and experience of the student's math teacher and the student's math class size are the relevant variables. Likewise, the dependent variable is the student's math test score.

234

Economics o f Education Review

Table 3. Defining the class size spline variables


Splines Spline 1 (small-range) Scenario 1 0-15 Students Scenario 2 0-20 Students Variable definition

Distribution of Students by Ability into Class Sizes


The positive OLS impact of class size on test scores, and the positive effects of the lower ranges of class size as illustrated by the spline results, could be a reflection that students who are doing poorly (or who are tracked into special education classes) are placed into very small classes in order to receive special attention. To explore this line of reasoning, the following variable was analyzed: ability level of the class (high, middle, or low) for all four subjects. The ability level of the student's class is from a question directed to the teacher and is expressed as a variable where one equals being in a lower ability class and zero equals otherwise. About 14 to 18 percent of the students are characterized as being in lower ability classes, as defined by the teacher. Figures 1 through

Actual class size If cl. size >15(20), var = Spline 2 16-30 21-30 CI. size-15(20); (mid-range) Students Students else zero If cl. size >30, var = Spline 3 C1. size-30; else (large-range) 31+ Students 31+ Students zero

Table 4. Results from OLS regressions with splines


1. Regression results for class size variables Math (N = 4741) Coeff. t-Stat. Scenario 1

English (N = 4973) Coeff. t-Stat.

History (N = 4850) Coeff. t-Stat.

Science (N = 4645) Coeff. t-Stat.

0-15 Students 16-30 Students 31+ Students


Scenario 2

0.44 -0.38 0.16 0.24 -0.22 0.22

4.27 -3.26 1.52 4.55 -2.71 1.84

0.31 -0.25 -0.06 0.20 -0.17 -0.00 8.51 8.81

5.11 -3.61 -0.92 6.53 -3.69 -0.03

0.40 -0.42 0.08 0.17 -0.21 0.12 10.33 7.52

4.78 -4.54 1.28 4.21 -3.87 1.75

0.32 -0.33 -0.03 0.17 -0.23 0.05 7.21 10.49

3.67 -3.59 -0.38 4.38 -4.41 0.69

0-20 Students 21-30 Students 31+ Students

2. Joint test of b(mid) = b(large) = 0

Scenario 1 F-test 5.46 Scenario 2 F-test 3.82 Critical value at 5% significance = 3

3. Impact of an additional student in class on achievement in each of the three ranges Formula Math English History Scenario 1

Science

0-15 Students 16-30 Students 31+ Students


Scenario 2

b(small) b(small) + b(mid) b(small) + b(mid) +b(large) b(small) b(small) +b(mid) b(small) + b(mid) +b(large)

0.44 0.06 0.22

0.31 0.06 0.00

0.40 -0.02 0.06

0.32 -0.01 -0.04

0-20 Students 21-30 Students 31+ Students

0.24 0.02 0.24

0.20 0.03 0.03

0.17 -0.04 0.08

0.17 -0.06 -0.01

Does Class Size Matter? 4 depict class size distributions by ability for all four subjects. As shown, lower ability students are disproportionately represented in the lower ranges of class size? Instrumental Variable Approach The evidence presented above suggests that lower ability students are not randomly allocated to different class sizes and are disproportionately found in the smaller classes. One way to correct for the correlation between class size and unobserved test-taking ability would entail an instrumental variable procedure that predicts class size to account for the endogeneity of class size. Based on a theoretical knowledge of how schools work, this paper uses the average class size for a given subject in the student's school and the eighth grade student enrollment in the school as instruments for class size. The average class size variable is exogenous to an individual student but is correlated strongly with an individual student's actual class size. This variable is constructed by taking the average class size of all the students in the school in a given subject that responded to the NELS survey. By aggregating the class size variable, the within-school non-random student allocation problem is bypassed while at the same time avoiding the potential problems of using a pupil-teacher ratio, which may not have any relation to actual class size in a school, A school's eighth grade enrollment is also exogenous to an individual student but may identify which schools are more likely to allocate their students in a non-random way. Larger schools may have the scale and resources to offer specialized small classes for lower ability students; it may not be cost-effective for smaller schools to offer these services. Overall, school size is an important predictor of a school's curriculum and of the prevalence of tracking students by ability (Monk and Hailer, 1993; Monk, 1987; Oakes, 1985; Powell et al., 1985). Table 5 shows the results of regressing class size on all of the exogenous variables in the system as well as the two-stage least squares (2SLS) estimation with the predicted value of class size entered into the achievement equation, as described in equation (2). These are contrasted with the OLS regression results from Table 2. The 2SLS results show that once the endogeneity of class size is taken into account, the effect of class size on student achievement changes from positive to negative and significant for science and history

235

achievement and to negative and insignificant for math and English achievement, suggesting returns to investing in smaller classes for some students. Overall, class size effects are significant for science and history but small, as compared to family background effects. For example, a one unit decrease in a student's science class size will increase that student's science test score by .08 points (based on a 25 point test with a mean of 13.82 and a standard deviation of 4.5). Likewise, decreasing a student's history class size by one person would only increase the student's test score by 0.07 points (based on a 30 point test with a mean of 19.44 and a standard deviation of 5.4). Decreasing class size from the current average of 25 to 15 would improve the student's science and history test scores by 0.8 and 0.7 of a point, respectively. Thus, the incremental benefits may not surpass the incremental costs of decreasing class s i z e . 4 Finally, as the spline results suggested above, the effect of class size, even after correcting for non-random allocation of students, may not be independent of the level of class size. However, it is problematic to create spline covariates which include an endogenous component. One approach may be to recreate the splines as defined above, but using the average class size variable instead of actual class size. The results of this specification yielded a negative and significant effect of class size in the following categories: math classes in the small-range of both scenarios (scenario 1 effect = -0.47, scenario 2 effect = -0.34), science and history classes in the mid-ranges of scenario 2 (science effect = -0.14; history effect = -0.14). These results uphold the previous findings of: (1) nonlinearities in class size effects, and (2) returns to investing in smaller science and history classes. 6. EXTENSIONS O F THE M O D E L In order to examine class size effects more closely, the low-achieving students that are grouped in the small classes were dropped for analysis (defined as low-ability students in classes fewer than 18 students). When using OLS on this subsample of students, the effect of class size was negative and significant only for science and history achievement, just as was found from using the 2SLS technique on the full sample of students, but with even lower magnitudes (the coefficients of the class size variable were - 0 . 0 4 and - 0 . 0 3 for science and history achievement, respectively). The similarity of these findings suggests that the 2SLS technique appropriately con-

to T a b l e 5. 2 S L S vs. O L S r e g r e s s i o n results b y subject a r e a o~

Math achievement Hausman test (OLS vs. 2SLS) = 46

N = 4741 Variable Constant Family income Family income squared Family comp, (1 = married)' Parent educ. (1 = beyond H.S.) Sex of student (1 = male) Race of student (1 = white) Preschool ( 1 = attended) % of Single parent families Urban (1 = Yes) Race of teacher (1 = white) Sex of teacher (1 = male) Years of experience Years of experience squared Class size Eighth grade enrollment Avg math class size in school Coeff. 8.83 1.00 -0.04 1.06 2.31 0.60 2.55 0.89 -2.09 -0.81 1.90 0,40 0.32 -0.01 0.12 0.0004 1.02 t-Stat 12.80 9.68 -5.90 4.13 9.47 2.36 8.39 3.98 -2.98 -2.01 4.74 1.72 4.51 -3.55 -0.74

Dep. vat. = test score (no. right) OLS 2SLS R 2 : 0.18 R 2 = 0.17 F = 74 F = 70 t- Stat Coeff. 10,86 12.30 9.27 1.05 -5.53 -0.04 4.16 1.05 9.21 2.39 2.63 0,54 9.02 2.39 3.67 0.97 -3.21 -1.95 -2.56 -0.64 5.08 1.77 1.61 0.42 4.87 0.30 -3.85 -0.01 6.25 -0.02 0.63 59.64

Dep. var. = math class size OLS R 2 = 0.49 F = 302 Coeff. t-Star - 1.11 - 2.02 0.06 1.02 -0.002 -0.59 0.10 0.69 0,20 1.46 -0.25 - 1.93 0.32 2.00 0.06 0,41 0.69 1.88 0.05 0.30 0.31 1.50 -0.25 - 1.83 -0.04 - 1.09 0,002 1.74

_~ ~" ~,~

English achievement Hausman test (OLS vs. 2SLS) = 83

N = 4973 Variable Constant Family income Family income squared Family comp. ( l = married) Parent educ. (1 = beyond H.S.) Sex of student (1 = male) Race of student (1 = white) Preschool ( I = attended) % of single parent families Urban (1 = yes) Race of teacher (1 = white) Sex of teacher (1 = male) Years of experience Years of experience squared Class size Eighth grade enrollment Avg English class size in school

Dep. var. = test score (no. right) OLS 2SLS R 2 = 0.17 R 2 = 0,15 F = 71 F = 63 Coeff. t-Stat Coeff. 6.80 15.54 9.28 0.59 10.18 0.62 -0.02 -6.73 -0.03 0.41 3,03 0.40 1.25 9.49 1.29 -0.85 -7.01 -0.91 1.39 9.50 1.31 0,46 3.53 0.49 -1.66 -4.95 -1.41 -0.01 -0.06 0.08 0.85 4.20 0.77 -0.37 -2,61 -0.33 0.06 1.52 0,07 -0.001 -0.90 -0.002 0.08 7.71 -0.02

t- Stat 17.93 10.63 -7.18 2.94 9.71 -7.42 8.83 3.79 -4.15 0.52 3.78 -2.28 1.88 - 1.29 - 1.50

Dep. var. = eng. class size OLS R 2 = 0.48 F = 308 Coeff. t- Stat - 0.70 - 1.40 0.16 2.82 -O.Ol -2.46 0.12 0.91 0.40 3.16 -0.54 -4.65 -0.03 -0.20 0.10 0.80 0.27 0.85 -0.05 -0.33 -0,04 -0.20 0.17 1,28 0.06 1.75 -0.002 - 1.38 0.O001 1.00 0.09 62.21

~"

Science achievement Hausman test (OLS vs. 2SLS) = 74

N = 4645 Variable Constant Family income Family income squared Family comp. (1 = married) Parent educ. (1 = beyond H.S.) Sex of student (1 = male) Race of student (1 = white) Preschool (l = attended) % of single parent families Urban (1 = yes) Race of teacher ( 1 = white) Sex of teacher (I = male) Years of eKDerience Years of experience squared Class size Eighth grade enrollment Avg science class size in school
t- Stat 18.63 8.35 -5.52 2.07 9.00 4.81 8.45 4.89 -3.15 0.68 3.14 1.39 3.09 -2.68 -5.40

Dep. var. = test score (no. right) OLS 2SLS R 2 = 0.13 R 2 = 0.12 F = 50 F = 45 C o e f f. t- Stat Coeff. 8.52 16.59 10.58 0.47 8.08 0.49 -0.02 -5.33 -0.02 0.31 2.25 0.29 1.22 9.01 1.23 0.61 4.97 0.60 1.35 8.90 1.28 0.60 4.53 0.65 -1.31 -3.46 -1.19 0.04 0.27 0.11 1.06 3.41 0.98 0.19 1.46 0.18 2.79 0.11 0.10 -0.003 -2.48 -0.003 0.002 0.13 -0.08 -0.0002 1.01 -0.42 76.71

Dep. var. = sci. class size OLS R 2 = 0.61 F = 474 Coeff. t- Stat -0.40 -0.89 0.04 0.90 -0.002 -0.78 0.06 0.56 -0.12 - 1.11 -0.11 - 1.15 0.11 0.88 0.28 2.68 -0.14 -0.48 0.04 0.31 0.17 0.70 0.26 2.55 -0.04 - 1.35 0.001 1.50 e',, ~-~ ,~.r'~

History achievement Hausman test (OLS vs. 2SLS) = 68

N = 4850 Variable Constant Family income Family income squared Family comp. ( 1 = married) Parent educ. (1 = beyond H.S.) Sex of student (1 = male) Race of student (1 = white) Preschool (1 = attended) % of single parent families Urban (1 = yes) Race of teacher (1 = white) Sex of teacher (1 = male) Years of experience Years of experience squared Class size Eighth grade enrollment Avg history class size in school

Dep. var. = test score (no. right) OLS 2SLS Rz =0.14 R 2 =0.13 F = 55 F = 51 Coeff. t-Stat Coeff. 12.90 24.66 15.20 0.60 8.94 0.63 -0.02 5.82 -0.02 0.48 2.92 0.45 1.59 9.86 1.62 0.39 2.67 0.36 1.17 6.50 1.05 0.77 4.97 0.82 0~55 1.20 0.68 -0.66 -3.33 -0.50 0.97 3.71 0.92 -0.11 -0.71 -0.10 0.06 1.50 0.06 -0.001 -0.92 -0.001 0.03 1.91 -0.07

t-Stat 25.60 9.40 -6.27 2.77 10.01 2.46 5.78 5.30 1.50 -2.52 3.48 -0.68 1.53 -0.78 -4.15

Dep. var. = hist. class size OLS R z = 0.58 F = 440 Coeff. t-Star - 1.03 - 2.29 0.11 2.29 -0.01 -2.03 0.17 1.43 0.11 0.90 -0.27 -2.53 0.15 1.13 O. 18 1.56 0.34 0.99 0.01 0.05 0.03 0.14 0.19 1.73 -0.02 -0.68 0.001 1.07 0.0004 1.02 -0.75 73.59

'-,~

..-..1

238

Economics of Education Review


1o

a 6

J,

,I
2

i i 3 4 5 8 ? 0 9 I t 13 i l l t7 II) 21 23 21 :rr 29 $1 ~ N ~ ~) 42 1@ 1= 14 14 tll 20 ;12 24 ;m 28 ~0:12 34 ~ U 40 44 o -* ~ - - - - ' ' 4 e I 10 t2 14 111 t8 20 2;I 24 24 20 SO S2 4 Sll ~1 40 $ 7 0 11 t3 18 1? 19 2t ~ 28 27 2 t $t 33 :18 37 3JD

Figure l, Distribution of students by ability to math class sizes,

Figure 3. Distribution of students by ability to science class sizes.

t0 10

o o

J
4

]-4 $ 8 I 1' 0

JI II 10 12 14 l e 18 ~0 22 24 N 28 3Q 3~ 34 :MI H 40 11 13 IS 17 19 21 ~ 28 27 N 31 ~ U $7 N 41 E ~ C m b

I. d S 0

. 8 ? 0

. . . . . . I I 10 12 14 16 18 ~r0 n .24 N S ~10 12 114 M M 40 4S 441 1t IS 48 17 10 21 ~ 211 27 m 31 3:1 M ~ N 42 M 40 HImyCme m--

IBumd on Mo~ng A ~ m l e of FmJr C h i ~ ~ m ) (h~d on Mo~InO A ~ ~ Fmf C~ S~wl

Figure 2. Distribution of students by ability to English class sizes.

Figure 4. Distribution of students by ability to history class sizes.

Does Class Size Matter?


Table 6. Interactions of class size and student characteristics Coefficients (and t-stats.) of interaction term Math English Science History (N = 4741) (N = 4973) (N = 4645) (N = 4850) -0.20 (-3.32) 0.07 (!.23) -0.01 ( - 1.47) 0.14 ( 1.96) 0.06 (1.92) 0.01 (0.20) -0.003 (-0.57) -0.05 ( - 1.13) -0.01 (-0.22) -0.02 (-0.59) 0.01 (1.34) 0.03 (0.64) -0.08 (-2.04) 0.01 (0.28) -0.01 ( - 1.49) 0.16 (3.57)

239

Interaction term Student race* class size Student sex* class size Family income* class size Urban school* class size

It is the average predicted class size variable that is used to create these interaction variables. Interactions were also tried with the average class size variable, leading to similar results. t-Statistics of the interaction variables are in parentheses.

trois for the problem of lower ability students being disproportionately found in the smaller classes. When the sample was further stratified into various groups of students (i.e., by ability, gender, race, etc.), the magnitude of class size effects remained relatively small for all subgroups, Returning to the original sample of students, a number of interaction variables were created as another method to identify whether class size effects differed by type of student. The fitted value of class size (and alternatively, the average class size variable) was multiplied by the student's race, sex, family income, and the urbanicity of the school to construct four interaction variables for each of the four subject areas. As shown in Table 6, non-white students did better in smaller English classes while white students did better in smaller math and history classes. In addition, non-urban students did better in smaller math and history classes. Under no scenario did sex of student or family income interact with class size in a significant manner. 7. CONCLUSIONS AND P O L I C Y IMPLICATIONS

effect of class size on achievement changes from positive to negative for all four subjects and to significantly negative for science and history achievement. This suggests that if students are assigned randomly to different class sizes, investing in smaller classes in certain subjects may yield positive returns, as measured by increases in student performance. However, these effects on achievement are small in size as compared to family background effects, and the incremental benefits of decreasing class size may not exceed the associated costs. Further research is needed to identify other instruments, for class size that may more accurately gauge the magnitude of effect, and to determine which types of students benefit the most from smaller class sizes. In closing, instead of relying on past studies that claim throwing money at the schools will do no good, educational policymakers should question how the school resource variables were measured in these studies and should focus either on experimental (randomized) assessments of school resources or on statistical techniques that purge the analysis of the endogeneity of school resource inputs. Only by using statistically unbiased measures of educational inputs can the debate on educational reform in the United States be informative, relevant and productive. Acknowledgments--I am grateful to T. Paul Schultz, Duncan Thomas, Jim Akerhielm, Susan Parker, and two anonymous referees of this journal for their comments and suggestions. Any errors are solely attributable to myself.

This paper found that public school students are not randomly allocated to different class sizes, with the lower ability students disproportionately found in the smaller classes. When this is taken into account using instrumental variable regression methods, the

240

Economics o f Education Review

NOTES 1. For example, for scenario I, if class size was greater than 15 students, the created mid-range variable was set equal to actual class size minus 15; if class size was 15 or fewer students, the created midrange variable was set equal to zero. If actual class size was also greater than 30, the created largerange variable was set equal to actual class size minus 30; if not, it was set to zero. 2. For example, in interpreting the spline coefficients for science achievement in scenario 2, increasing .science class size by one in a class of 0 to 20 students will also increase science achievement by 0.17 of a test point. However, if class size increases by one in a class of 21 to 30 students, science achievement will drop 0.06 of a test point (0.17 - 0.23). Likewise, increasing class size by one in a class with more than 30 students will decrease student achievement by 0.01 of a test point (0.17 - 0.23 + 0.05), although this latter change is not statistically significant. 3. Comparing the two distributions of students for all four subjects yielded a Pearson chi-square test statistic that tests the null hypothesis that the two distributions are random samples from the same population. This chi-square test statistic (with degrees of freedom equal to the number of different class sizes minus one) exceeded the critical value of 64 at the 99th percentile for all four subjects. Thus, the hypothesis of equal class size distributions for lower and non-lower ability students was rejected (the chi-square statistics were: math - 477, English - 497, science - 219, history - 309). 4. A Hausman specification test was employed to compare the OLS and 2SLS estimators and to test for the exogeneity of the class size variable. As seen in Table 5, for all four subjects, the chi-square critical value of 25 (five percent significance level with 15 degrees of freedom) is exceeded, rejecting the null hypothesis of equality between the OLS and 2SLS coefficients. That is, although the class size coefficients are small, the OLS and 2SLS estimates can be distinguished from one another on statistical grounds. These results suggest that class size is not exogenous and that the OLS model may be misspecifled.

REFERENCES' BOOZER, M.A., KRUEGER, A.B. and WOLK1N, S. (1992) Race and school quality since Brown v. Board of Education. Brookings Papers on Economic Activity: Microeconomics 1992. Washington DC: Brookings Institution, 269-338. BOWLES, S. and LEVIN, H. M. (1968) The determinants of scholastic achievement - an appraisal of some recent evidence. The Journal of Human Resources 3, 3-24. CAHEN, L.S., FILBY, N., MCCUTCtlEON, G. and KYLE, D.W.'(1983) Class Size and Instruction: A Field Study. New York: Longman. CARD, D. and KREUGER, A. (1992) School quality and black-white relative earnings: A direct assessment. The Quarterly Journal of Economics 107, 151-200. CHUBB, J.E. and MOE, T.M. (1990) Politics, Markets & America's Schools. Washington, D.C.: The Brookings Institution. COLEMAN, J.S., HOFFER, T. and KILOORE, S. (1982) High School Achievement: Public, Catholic and Private Schools Compared. New York: Basic Books. COLEMAN, J.S., CAMPBELL, E.Q., HOBSON, C.J. MCPARTLAND, J., MOOD, A.M., WEINFELD, F.D. and YORK, R.L. (1966) Equality of Educational Opportunity. Washington, D.C.: National Center for Education Statistics. DOLAN, R.C. and SCHr~IDT, R.M. (1987) Assessing the impact of expenditure on achievement: some methodological and policy considerations. Economics of Education Review 6, 285-299. ERS, Ir~c. (1986) Class Size Research: A Related Cluster Analysis for Decision Making. Arlington, VA: ERS, Inc. GLASS, G.V. and SMITH, M.L. (1979) Meta-analysis of research on class size and achievement. Educational Evaluation and Policy Analysis 1, 2-16. HANUSHEK, E.A. (1986) The economics of schooling: Production and efficiency in public schools. Journal of Economic Literature 24, 1141-1177. HANUSnEK, E.A. (1979) Conceptual and empirical issues in the estimation of educational production functions. The Journal of Human Resources 14, 351-388. LINK, C.R. and MULLIGAN, J.G. (1991) Classmates' effects on black student achievement in public school classrooms. Economics of Education Review 10, 297-310. McGIVERIN, J., GILMAN, D. and TILLITSKI, C. (1989). A Meta-analysis of the relationship between class size and achievement. The Elementary School Journal 90, 47-56.

Does Class Size Matter?

241

MONK, D. (1987) Secondary school size and curriculum comprehensiveness. Economics of Education Review 6, 137-150. MONK, D.H. AND HALLER,E.J, (1993) Predictors of High School academic course offerings: The role of school size. American Educational Research Journal 30, 3--21. OAKES, J. (1985) Keeping Track: How Schools Structure Inequality. New Haven, CT: Yale University Press. POWELL, A.G., FARRAR, E. and COHEN, D.K. (1985) The Shopping Mall High School: Winners and Losers in the Educational Marketplace. Boston: Houghton Mifflin. SMITH, M.L. and GLASS, G.V. (1980) Meta-analysis of research on class size and its relationship to attitudes and instruction. American Educational Research Journal 17, 419--433. SUMMERS,A.A. and WOLFE, B.L. (1977) Do schools make a difference? American Economic Review 67, 639---652. WORD, E. et al. (1990) Student~Teacher Achievement Ratio (STAR) - Tennessee's K-3 Class Size Study. Tennessee State Department of Education.

Вам также может понравиться