Do crowded classrooms crowd out learning? Evidence from the food for education program in Bangladesh

www.elsevier.com/locate/worlddev

World Development Vol. 34, No. 4, pp. 665–684, 2006 Ó 2006 Elsevier Ltd. All rights reserved 0305-750X/$ - see front matter

doi:10.1016/j.worlddev.2005.09.011

Do Crowded Classrooms Crowd Out Learning? Evidence from the Food for Education Program in Bangladesh AKHTER U. AHMED International Food Policy Research Institute, Washington, DC, USA

and MARY ARENDS-KUENNING * University of Illinois at Urbana-Champaign, USA Summary. — Bangladesh’s Food for Education Program (FFE), which provided free food to poor families if their children attended primary school, was successful in increasing children’s school enrollment, especially for girls. However, this success came at a price as class sizes increased. This paper uses a rich data set that includes school achievement test scores, information on schools, and household data to explore the impact of FFE on the quality of education. The analysis focuses on the impact of FFE on the achievement test scores of students who did not receive benefits. We find evidence for a negative impact of FFE on the test scores of non-beneficiary students through peer effects rather than through classroom crowding effects. Ó 2006 Elsevier Ltd. All rights reserved. Key words — South Asia, Bangladesh, food for education, test scores, classroom crowding, peer effect

ration of food grains to poor families in rural areas if their children attended primary school. Thus, the FFE food grain ration became a monthly income entitlement enabling a child from a poor family to go to school. The family could consume the grain, thus reducing its cash outlay for food, or it could sell the grain and use the cash to meet other expenses. The GOB terminated the FFE program in 2002 and initiated the Primary Education Stipend (PES) program. The PES program provides cash assistance to poor families if they send their children to primary school. Recent evaluations of the Bangladesh’s FFE program found that the program fulfilled its objectives of increasing school enrollment, promoting school attendance, and preventing dropouts. The enrollment increase was greater

1. INTRODUCTION It is hard to over-emphasize the importance of quality education for improving the welfare of individuals. In developing countries, providing universal primary education connotes a great opportunity to reduce poverty and to promote economic growth. Quality primary education would equip children from poor families with literacy, numeracy, and basic problemsolving skills to move out of poverty. In Bangladesh, pervasive poverty has kept generations of families from sending their children to school, and without education, their children’s future will be a distressing echo of their own. In an effort to increase primary school enrollment of children from poor families and to retain them in school, the Government of Bangladesh (GOB) launched the Food for Education (FFE) program in 1993. The FFE program provided a free monthly

* Final revision accepted: September 23, 2005.

665

666

WORLD DEVELOPMENT

for girls than for boys (Ahmed, 2000; Ahmed & Billah, 1994; Arends-Kuenning & Amin, forthcoming; BIDS, 1997; DPC, 2000; Khandker, 1996; PMED, 2000; Ravallion & Wodon, 1997). However, because of increased enrollment and class attendance rates, FFE school classrooms were more crowded than non-FFE school classrooms. Consequently, there have been concerns about the deterioration of the quality of education in FFE schools. In this paper, we discuss whether these concerns are valid. We compare the characteristics of FFE and non-FFE schools, including class size, teachers’ education and training, and physical characteristics of schools. FFE schools had significantly larger class sizes than non-FFE schools. As for other school quality characteristics, we show that there were bigger differentials between government and non-government schools than between government schools that had the FFE program and those that did not. We discuss how FFE encouraged the school attendance of poor children, who had socioeconomic characteristics that would lead them to score lower on achievement tests than children who had been in school before the program. The argument that FFE caused school quality to deteriorate implies that FFE brought negative impacts upon children who were enrolled in school before the program started. Therefore, we examine the impact of the FFE program on the achievement test scores of those children who did not receive the benefits. 2. THE FFE PROGRAM The FFE program grew out of the Government of Bangladesh’s Rural Rationing program, a food subsidy program in the public food distribution system, which was designed to provide food security to poor families and was not linked with children’s education. Past research (Ahmed, 1992) found that 70% of the subsidized food grain (mostly rice) distributed by the Rural Rationing program was going to the non-poor. The high cost of subsidy and heavy leakage to the non-poor motivated its abolition. As an alternative to the Rural Rationing program, the Food for Education program was proposed by a working group as a means to get better targeting of food to poor families while increasing children’s educational attainment. The FFE program started in 1993 on a large, nationwide scale. Poor children enrolled in pri-

mary school grades 1–5 were eligible. The program was not available in all areas; it covered 460 unions, one union in each of the 460 rural thanas in Bangladesh. 1 The program expanded to 1,247 unions by 2000. Before the program was terminated in June 2002 and replaced by a cash transfer program (the Primary Education Stipend Program), FFE covered about 27% of all primary schools and enrolled about one-third of all primary school students in Bangladesh. The 2.1 million FFE beneficiary students accounted for about 13% of all students in primary schools. In 1999–2000, the GOB spent Tk 52.39 billion (US$1.03 billion) on education, which accounted for 15% of Bangladesh’s total budget in that year. Government expenditure on primary education was Tk 24.35 billion (US$478 million) or 46% of total education expenditure in 1999–2000 (BANBEIS, 2002). The average recurring costs per year to run a primary school in 1999–2000 were Tk 356,316 ($6,987) per year. The FFE program accounted for a significant share of Bangladesh’s expenditure on primary education, increasing from 4.7% of these expenditures in 1993–94 to 16.1% in 1999– 2000. By 1999–2000, the annual cost of FFE increased to Tk 3.94 billion (US$77 million), which translated into Tk 1,897 (US$37.19) per beneficiary student per year. 2 Expenditure on the program represented a sizable proportion of the Government’s educational expenditures, and if programs that paid children to attend school had negative impacts on school quality, some of the funds now going to the cash benefit program might be better spent on improving school quality. The FFE program used a two-step targeting mechanism, which proved to be fairly successful at reaching the poor (Ahmed & del Ninno, 2002). First, two to three unions that were economically backward and had a low literacy rate were selected from each of the 460 rural thanas. The program covered all government, registered non-government, community (low-cost), and satellite primary schools, and one Ebtedayee Madrasa (religion-based primary school) in these selected unions. Second, within each union, households with primary-schoolage children became eligible for FFE benefits if they met criteria based on landholding, living in a female-headed household, and having a household head in a low-income occupation. Households that were covered under another targeted food-based program of the government were not eligible to receive FFE food grains.

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

Beneficiary households were chosen by local groups, who, based on the targeting criteria, prepared a list of FFE beneficiary households in every union at the beginning of each year. Due to resource constraints, the total number of beneficiary households was identified so that no more than 40% of students received FFE rations. If a household was selected to participate in the FFE program, it was entitled to receive a free ration of up to 20 kilograms (kg) of wheat or 16 kg of rice per month for sending its children to a primary school. 3 If a household had only one primary school-age child (6–10 years) who attended school, then that household was entitled to receive 15 kg of wheat or 12 kg of rice per month. To be eligible for 20 kg of wheat or 16 kg of rice, a household was required to send more than one child, and all primary-school-age children, to school. To maintain their eligibility, children had to attend 85% of total classes in a month. The FFE program was successful at meeting its goal of increasing children’s school attendance. The International Food Policy Research Institute (IFPRI) conducted an evaluation of FFE in 2000, which suggested that the overall rate of class attendance was 70% in FFE schools and only 58% in non-FFE schools (Ahmed & del Ninno, 2002). Arends-Kuenning and Amin (forthcoming) found that enrollment in two villages increased from 68% to 87% for girls aged 6–10 and from 46% to 71% for boys aged 6–10 during 1992–95. However, the government of Bangladesh did not invest in hiring new teachers, building new schools, or increasing the quality of schools at the same rate at which enrollment increased. The government did not plan well for the consequences of the success of the FFE program. Therefore, with the increase in enrollment and class attendance rates, the classrooms of FFE schools became crowded. Class sizes were larger in FFE schools than in non-FFE schools. Increasing children’s school enrollment and attendance rates is certainly an impressive achievement, but now must be accompanied with a focus on improving student learning. 3. THE EFFECT OF CLASS SIZE ON STUDENT LEARNING Given that class size increased dramatically as a result of the FFE program, what impact did the increased class size have on student

667

achievement? The impact of class size on student learning is not a settled research question in the economics of education literature. Hanushek (forthcoming) noted that few published studies in the United States and in developing countries found the expected negative and significant impact of class size on student learning, indicating that school resources were not utilized well. Krueger (2002) argued that most published studies were seriously flawed. If one analyzed the published studies giving more weight to the well-designed studies, class size had a significant and negative impact on achievement test scores. Glewwe (2002) was highly critical of conventional studies of the impact of class size on student achievement test scores. He pointed out that most studies did not properly control for the unobserved characteristics that influenced class size and achievement. For example, governments might provide more resources in areas that have low schooling attainment compared to areas with high schooling attainment. In this case, regression analysis that did not control for the government’s resource allocation might find that larger class sizes were associated with higher school achievement. Glewwe cited two welldesigned studies in South Africa (Case & Deaton, 1999) and in Israel (Angrist & Lavy, 1999) that both found significant positive impacts of smaller class sizes on student achievement. Smaller class sizes might not be associated with higher achievement test scores for other reasons. If classroom performance is negatively affected by disruptions, as class size increases, the probability that a disruption will occur tends to increase, other things being equal. However, disruptive students might be placed in smaller sized classes, in an effort to monitor their behavior, which would tend to attenuate an otherwise positive impact of small class sizes on school achievement (Lazear, 1999). In a study about determinants of primary education quality in Francophone Africa, Michaelowa (2001) found a positive effect of the number of students in each class on student achievement up to a class size of 62 students. Above this size, however, the effect became increasingly negative. Hanushek (1998) and Mingat and Suchaut (1998) also found that, below a certain limit, an increase in class size would not lower student achievement. However, Michaelowa’s study assumed that class size was not related to any unobserved characteristics of children, communities, or countries, so her study was subject to Krueger’s criticisms.

668

WORLD DEVELOPMENT

Studies about FFE can help inform the debate about the impact of class size on student achievement. The FFE program provided exogenous variation in class size as school enrollment increased dramatically in FFE schools relative to non-FFE schools. Because the program was allocated at the union level, we can control for the placement of the FFE program in poorer unions by adding union-level dummy variables. School enrollment increased without much government investment in new schools or in school quality. Most villages had only one school, so students did not have a choice of schools to attend (see Appendix A for further discussion of this point). Parents of children with high ability were not able to select schools with lower class sizes, which would lead to an endogeneity problem. Parents had very low educational levels in the study areas. They lacked political clout and were not able to pressure the government to improve school quality or to hire more teachers in the schools. A detailed explanation of why the endogeneity problems that Glewwe noted are mitigated in rural Bangladesh is found in Appendix A. 4. THE DATA Our data came from school and household surveys conducted in Bangladesh by IFPRI in September–October 2000 for an evaluation of the FFE program. The surveys included primary schools with and without the FFE program, a cross section of households including program beneficiaries and non-beneficiaries, and a survey of community informants. The sample included 600 households in 60 villages in 30 unions in 10 thanas, and 110 schools in the same 30 unions from which the household sample was drawn. First, the sampling process randomly selected 10 thanas with probability proportional to size (PPS), based on thana-level population data from the 1991 census. Second, two FFE unions and one non-FFE union were randomly selected per thana. Third, two villages from each union were randomly selected with PPS using village-level population data from the 1991 census. A complete census of the households was carried out in each of the selected villages. Then, 10 households that had at least one primary-school-age child (6–12 years old) were randomly selected in each village from the census list of households. The sample included program beneficiaries and non-beneficiaries. A child who

attended school could be in the non-beneficiary category for two reasons—if she lived in a union that did not have the program, or if she was not selected to participate in the program although the union had the program. The household questionnaire collected information on a wide variety of topics, such as household composition, occupation, education, school participation, dwelling characteristics, assets, expenditures, and use of the FFE system. In addition, a village census questionnaire collected information on household demography, school enrollment, literacy, and FFE participation from 17,134 households. Only those schools attended by the children in the sample households were selected for the school survey. A total of 110 primary schools (70 FFE and 40 non-FFE schools) were surveyed. The school questionnaire collected information on student enrollment, class attendance, dropout and repetition, teacher qualification, school facilities, physical characteristics, school expenditures, and FFE program participation. The data included children’s academic achievement test scores. Two sets of test scores were available. The test was administered twice—once to 3,369 fourth-grade students attending the 110 surveyed FFE and nonFFE schools, and separately to a sub sample of 288 children in the household. The test was a standard academic achievement test, designed to assess the quality of education received by students. The household sample of test score data was limited because of its small sample size and because children had to agree to take the exam at home and were resistant to doing so, resulting in a possibly non-random sample. Another limitation of using the household-based test score data was that the 288 students who took the exam at home attended only 66 of the 110 schools included in the sample. The school sample test score data had the advantage of a large sample that related to school characteristics from the school survey, but lacked information on students’ socioeconomic background. However, the school test score data could be combined with community data and aggregate data from the household survey to control for socioeconomic background. We focused on the school test data because the benefits from the large sample size and the sample’s randomness outweighed the disadvantages of missing household socioeconomic data. The test included four subjects—Bangla, English, mathematics, and environmental awareness.

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

5. SCHOOL-LEVEL ANALYSIS (a) General information about schools Observations during the school survey suggest that non-government primary school buildings in rural Bangladesh were in much poorer condition than those of government primary schools. Only about 11% of the total sample of non-government schools had a permanent building structure of concrete or tin roofs, brick walls, and cement floors compared to 45% of all surveyed government schools that had such structure. Table 1 indicates that the average size of FFE schools (in terms of number of students per school) was about 27% larger than that of non-FFE schools because the FFE program enticed more children to attend schools. About half of all students were girls. Overall, about three out of 10 teachers were female. Average annual school operating expenses per student (excluding teacher salaries) were generally low (around Tk 40 per student a year), or very low (only Tk 27 per student a year) for nongovernment FFE schools. Both government and non-government schools under the FFE program were more intensively inspected than schools that were not in the program. More teachers in non-government schools were en-

669

gaged in private tutoring compared to government schools, and this was true for both FFE and non-FFE schools. Having many teachers engaged in tutoring is hypothesized to be a marker of low quality schools because teachers dilute their energy from teaching in the classroom when they tutor. Also, they have incentives to teach poorly so that their students will need their services. Not much difference existed between teachers in government FFE and government non-FFE schools, but teachers in government schools were better educated and better paid than teachers in non-government schools (Table 2). About 32% of government schoolteachers had a bachelor’s degree or above. In contrast, only 9.3% of all non-government schoolteachers had a bachelor’s degree. There was almost no difference in teacher salaries between FFE and nonFFE schools. However, the average salary of a government schoolteacher was about 3.5 times higher than that of a non-government schoolteacher. Further, most non-government schoolteachers were not paid regularly. Government schoolteachers were better off than non-government schoolteachers, as reflected by the relative levels of monthly household expenditures. School salary accounted for about threefourths of total income of government schoolteachers, whereas it accounted for only 27%

Table 1. General information about schools in rural Bangladesh, by type of school Information

FFE schools Government Non-government

Number of students per school in 2000 Proportion of girls (% of total) Average number of teachers per school Share of female teachers (% of all teachers) Average operating expenses per student (taka/year)* Inspection made by school inspectors in 1999 (% of schools) Number of inspections in 1999 Teachers engaged in private tutoring (% of teachers)

Non-FFE schools All

Government Non-government

All

350

315

343

286

162

270

50.0

50.0

50.0

50.0

48.3

49.9

4.7

3.9

4.5

4.4

4.0

4.4

28.9

29.3

29.2

33.1

...

...

43

27

40

41

...

...

100.0

92.9

98.6

88.6

80.0

87.5

5.7 14.3

3.4 50.0

5.2 21.4

5.1 25.7

2.4 80.0

4.8 32.5

Source: Based on data from IFPRI’s ‘‘Food for Education Evaluation Survey, 2000: School Survey,’’ Bangladesh. Note: Ellipsis (. . .) indicates information was not available. School operating expenses exclude teacher salaries, and include the costs of stationery and supplies, repair and maintenance, utilities, and communication.

670

Table 2. Information about teachers, rural Bangladesh, by type of school Type of information

Source of income (percent of total income) School salary Agriculture Small business Large business Other

Non-FFE schools

All government

All non-government

36.5 32.0 25.7 4.5 0.6 3,960 90.3

36.2 30.2 27.8 4.3 1.2 4,439 97.1

46.7 42.7 9.3 – – 1,285 32.0

4,265

6,635

6,991

4,072

20.0 75.0 – 5.0 –

63.4 24.6 1.1 1.7 6.9

72.7 14.4 1.7 1.2 5.3

26.7 61.3 5.3 4.0 1.3

Non-government

All

Government

Non-government

All

37.4 29.8 27.5 3.8 1.5 4,519 95.8

43.6 43.6 10.9 – – 1,279 36.4

38.5 32.2 24.6 3.2 1.3 3,960 85.5

34.2 31.0 28.4 5.2 0.6 4,306 99.4

55.0 40.0 5.0 – – 1,300 20.0

7,013

3,996

6,489

6,956

74.8 12.2 1.9 1.1 3.8

29.1 56.4 7.3 3.6 1.8

66.9 19.9 2.8 1.6 3.5

69.0 18.1 1.3 1.3 7.7

Source: Based on data from IFPRI’s ‘‘Food for Education Evaluation Survey, 2000: School Survey,’’ Bangladesh.

WORLD DEVELOPMENT

Educational qualifications (percent of teachers) S.S.C. H.S.C. B.A./B.A. B.Ed. M.A./M.A. M.Ed. Other Monthly salary (taka) Receive salary regularly (percent of teachers) Monthly household expenditure (taka)

FFE schools Government

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

of the total income of non-government schoolteachers. Non-government schoolteachers mainly depended on agriculture for their livelihood and were therefore less likely to devote themselves to teaching full time. When evaluating the impact of FFE on school achievement, we need to keep in mind that FFE expanded the proportion of students in low quality non-government schools from 21% in 1992–93 to 31% in 1998–99, according to government school enrollment statistics (Bangladesh Bureau of Statistics). In our school achievement test sample, we obtained a smaller percentage of children enrolled in non-government schools. Twelve percent of all students in the sample of fourth-grade students who took the test were in non-government schools. Eighteen percent of FFE beneficiaries were in non-government schools. (b) School-level effects on learning The main focus of our study is to assess the effects of FFE on learning. The relative quality of education in FFE and non-FFE schools could be judged on the basis of student/teacher ratios, number of students per classroom, and students’ academic achievement test scores. FFE schools had a higher number of students per teacher than non-FFE schools because during 1993–2000, the number of teachers per school remained virtually constant, while student enrollment increased significantly in FFE

671

schools. On average, whereas there were 55 students per teacher in non-FFE primary schools based on the survey, FFE schools had 67 students per teacher in 2000. Of the non-government schools, FFE schools had 81 students per teacher, whereas those without FFE had only 41 students per teacher in 2000 (calculated from Table 1). In the fourth grade, the focus of our study, there were 61 students per teacher in FFE schools and 48 students per teacher in non-FFE schools. Table 3 uses the results from the achievement test that was administered in the schools to fourth-grade students. In general, high standards were maintained in Bangla, but test outcomes were disappointing for English. Students were intermediate performers in mathematics. Average test scores were lower in FFE schools (49.1% correct answers) than in non-FFE schools (52.0% correct answers), and this difference was statistically significant. Within FFE schools, the average test score of FFE beneficiary students (45.6% correct answers) is statistically significantly less than that of the non-beneficiary students (53.2% correct answers), which brought down the aggregate score in FFE schools. FFE beneficiaries scored lower than non-beneficiaries probably because of their relatively lower socioeconomic status. The difference in test scores was larger between government and non-government schools than that between FFE and non-FFE schools, with government school students

Table 3. Student achievement test scores at the fourth-grade level: school survey results, rural Bangladesh, 2000 Type of school and program participation

All FFE schools Government schools FFE beneficiary students Non-beneficiary students Non-government schools FFE beneficiary students Non-beneficiary students All beneficiary students All non-beneficiaries in FFE schools All non-FFE schools Government schools Non-government schools All non-beneficiary students in all schools

Average rate of correct answers Mean achievement in Bangla 67.6 69.2 65.8 72.7 57.5 53.5 67.8 63.6 72.3 70.2 70.3 68.4 71.4

Mean achievement in English

Mean achievement in Mathematics

(Percent of correct answers) 27.9 47.3 29.0 48.7 27.5 45.5 30.6 52.2 20.4 38.2 17.6 33.9 27.7 49.4 25.7 43.4 30.4 51.9 30.7 50.5 31.3 51.7 23.4 35.6 30.5 51.3

Mean achievement in all subjects 49.1 50.5 47.6 53.5 40.1 36.3 49.9 45.6 53.2 52.0 52.7 43.7 52.7

Source: Based on data from IFPRI’s ‘‘Food for Education Evaluation Survey, 2000: School Survey,’’ Bangladesh.

672

WORLD DEVELOPMENT

performing better than non-government school students (Table 3). Government primary schools had better facilities, had more qualified teachers, and provided higher incentives to teachers compared to non-government primary schools. 6. HOUSEHOLD-LEVEL ANALYSIS (a) Profile of survey households Table 4 presents the characteristics of households living in FFE and non-FFE unions, disaggregated by per capita expenditure quartiles. 4 In the FFE program unions, about half (52%) of all households with primaryschool-age children were program beneficiaries. The results presented in the first two rows in Table 4 indicated that the distribution of FFE beneficiaries was somewhat progressive. About 60% of the households in the poorest quartile (i.e., the bottom 25% of households in the income distribution) were program beneficiaries, compared to 37% of the households in the richest quartile that received FFE benefits. However, this pattern also shows evidence of mistargeting, as many households in the higher income groups were included in the program.

About 45% of all FFE beneficiary households belonged to the richer half of all households. The results suggest that, for households with primary-school-age children in the first two quartiles (the bottom 50% of all households), about 18% in FFE unions and 26% in nonFFE unions did not send their children to school. Overall, about 13% of all households in FFE unions and 19% in non-FFE unions did not send their children to school. This pattern is an indication of the success of FFE in attracting children from poorer families to attend school. The average years of parents’ schooling were very low. Moreover, among the adult household members, over half of all adult males and almost three-quarters of all adult females never attended school. In both FFE and nonFFE unions, educational attainment of parents and other adults was positively correlated with income. (b) Household-level effects on student achievement Based on data for children who attended primary school, Table 5 shows that the average monthly per capita expenditure of non-beneficiary households with children attending FFE

Table 4. Characteristics of respondent households by per capita expenditure quartile: household survey results, rural Bangladesh, 2000 Per capita expenditure quartiles

FFE unions FFE beneficiary households (%) Percent of all beneficiaries Percent of households with primary-school-age children not going to school Years of schooling, father Years of schooling, mother No schooling, adult male (%) No schooling, adult female (%) Per capita monthly expenditure (taka) Non-FFE unions Percent of households with primary-school-age children not going to school Years of schooling, father Years of schooling, mother No schooling, adult male (%) No schooling, adult female (%) Per capita monthly expenditure (taka)

Total

1

2

3

4

60 29.0 19

53 25.6 17

57 27.5 7

37 17.9 8

52 100.0 13

0.9 0.5 66 83 335

1.9 0.9 59 80 498

2.7 1.1 53 77 671

3.9 2.4 36 50 1,474

2.3 1.2 54 73 745

20

32

16

8

19

0.8 0.2 56 92 356

1.6 0.9 68 80 521

3.3 1.5 56 76 728

5.0 3.7 24 36 1,597

2.7 1.6 51 71 800

Source: Based on data from IFPRI’s ‘‘Food for Education Evaluation Survey, 2000: Household Survey,’’ Bangladesh.

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

673

Table 5. Means of household and individual characteristics by beneficiary status, students enrolled in primary school: household survey results, rural Bangladesh, 2000 FFE unions

Mother’s education (years) Father’s education (years) Per capita monthly expenditure (taka)

Non-FFE unions

FFE beneficiaries

Non-beneficiaries

Non-beneficiaries

1.31 2.64 622

1.64 3.12 831

1.89 3.13 788

Source: Based on data from IFPRI’s ‘‘Food for Education Evaluation Survey, 2000: Household Survey,’’ Bangladesh.

school (Tk 831 per month) was 34% higher than that of FFE beneficiary households (Tk 622 per month). 5 This income difference is statistically significant. The average per capita expenditure of non-beneficiary households living in non-FFE unions with school-going children (Tk 788 per month) was about 5% lower than that of non-beneficiary households living in FFE unions. However, this difference was not statistically significant. Children who received FFE had parents with less education than children who lived in an FFE union, but did not receive FFE. However, neither of these differences was statistically significant. The educational levels of parents of children who did not receive FFE but lived in an FFE union were very similar to the educational levels of parents of children who did not live in FFE unions. Our school-level analysis presented in Table 3 suggests that the average test score of FFE beneficiary students was significantly lower than that of the non-beneficiary students in FFE schools. This is also true if we look at achievement test scores in the government schools. Further, the average score of non-beneficiary students in non-FFE union schools was slightly lower than that of the non-beneficiary students in FFE schools. This observation is also true if we only look at government schools, except for English scores, which were slightly higher for non-beneficiary students in non-FFE schools than for non-beneficiary students in FFE schools. One reason why students who received FFE had lower test scores than students who did not receive FFE is that they came from poorer households, and their parents had lower levels of education. We could not test this directly in our multivariate analysis, because the school achievement test data did not include data about the students’ backgrounds. However, the results are suggestive and imply that we need to separate FFE from non-FFE stu-

dents to investigate the impact of FFE on student achievement. 7. MULTIVARIATE ANALYSIS (a) Description of the model To determine the impact of FFE on the quality of education, researchers must take into account the characteristics of the students who decided to enroll in school because of the FFE program. These students were likely to be of lower ability than students who enrolled in school regardless of FFE. Therefore, a naı¨ve comparison between schools with FFE and schools without FFE would show that students in FFE schools had lower average test scores than schools without FFE. The relevant research question is whether children who would have been in school without FFE are now performing worse on tests because lower ability children are enrolled or because classrooms are crowded. These peer effects would be an indirect negative impact of FFE on children’s performance in school. Another relevant research question is whether the students who enrolled in school because of FFE learned ‘‘enough’’ to justify the costs of schooling to the government and the opportunity cost of the children’s time. The data were not detailed enough to answer this question. Arends-Kuenning and Amin (forthcoming) provided some evidence that FFE students spent time outside of school studying. If students were only going to school to get FFE benefits and not to learn, they would not study outside of class. The requirement that students meet a minimum standard on exams gave children incentives to learn class material. In the sample, two types of primary-school students were non-beneficiaries of the FFE program. The first group consisted of children

674

WORLD DEVELOPMENT

whose families were not eligible for benefits, although they lived in unions with the FFE program and sent their children to an FFE school. The second group consisted of primary school students who lived in unions that did not have the program in 2000, so that even if the children were poor, they did not have access to the program. We focus our analysis on children who attended government schools because of the large differences in quality between non-government and government schools. Of the non-beneficiary students in our sample with complete information, 92% attended government schools. The focus on government schools is also appropriate because government controls the inputs into these schools. In our model, student achievement (T), the test score taken from the school survey, is determined by a set of explanatory variables. We estimate the following equation: T ¼ a þ b gender of student þ v FFE school þ d percentage of fourth-grade students receiving FFE þ / classroom crowding þ c school classification þ g teachers’ characteristics þ i school processes þ u physical characteristics of schools þ j family characteristics þ l union control variables þ e;

ð1Þ

where a is a scalar, b, v, d, /, c, g, i, u, j, and l are parameters of corresponding explanatory variables, and e is an error term. The dependent variable, T, is the achievement test score taken from the school survey, which covered 3,369 fourth grade students in 110 schools. Only the non-beneficiary students were chosen for the analysis. With missing data, the selection of children who attend government schools, and the selection of non-beneficiaries, a total of 1,823 observations are used. We estimate separate regressions for math and Bangla test scores, which are the two most important subjects covered in the exam. The school characteristics that affect math scores are different from those that affect Bangla scores. In Appendix Table 1, we present results

for the total test scores including all four subjects of the exam (math, Bangla, English, and the Environment) and for the sum of Bangla and math scores. Combining the scores obscures the different impacts of school characteristics for different subjects. Also, there were only four questions on the exam about the environment, and the subject is not a core course taught in fourth grade. The controls for whether the school has an FFE program, the percentage of children in fourth grade who participate in the FFE program, and for crowding allow for a distinction among different impacts that FFE might have on student achievement. FFE school is a dummy variable equal to one if the school is an FFE school. To continue being eligible to receive FFE benefits, schools were required to meet a set of minimum educational quality standards, which might have had a positive impact on student achievement. 6 On the other hand, schools in poorer unions were more likely to be designated FFE schools than schools in richer unions, so the FFE dummy variable could indicate a poor school, which would tend to lower achievement. The number of students in the fourth-grade classroom measures the crowding effect. When the school survey data were collected, interviewers counted the number of students in each grade’s classroom who were attending school on that day. This information comprises the classroom crowding variable. The inclusion of both the percentage of fourth-grade students who participate in FFE and the number of students present in the fourth-grade classrooms in the regression allows for a distinction between peer effects and resource dilution effects on non-beneficiary students’ achievement. Resource dilution effects result when class resources, such as the teacher’s time, or classroom seating, are divided among increasing numbers of students. In a class of 20, teachers can spend an average of 3 minutes per hour per student, but in a class of 40, teachers can only spend an average of 1.5 minutes per hour on each student. As the percentage of children who receive FFE increases, the percentage of poor, low ability students is also likely to increase. If the achievement of non-beneficiary students decreases as the percentage of children who receive FFE increases, then that is an indication of negative peer effects. If FFE results in more children in the classroom, and therefore resources such as the tea-

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

cher’s time and attention are diluted, then the impact of the number of children in the fourth grade on non-beneficiary student achievement should be negative. Controlling for both FFE school status and the percentage of children who receive FFE distinguishes between the effect of having the FFE program and the extent of the program. The school classification variables are the government’s rating of the school’s quality. These ratings vary from ‘‘A’’ to ‘‘D.’’ Schools rated ‘‘D’’ are not allowed to participate in the FFE program. In our sample, no schools were rated ‘‘D.’’ Teachers’ characteristics include the proportion of teachers in the school who are female, the proportion of teachers in the school who have at least a bachelor’s degree, the average teacher salary in the school, and the proportion of teachers in the school who have other sources of income besides teaching, which is an indicator of the teachers’ dedication to teaching. Although many other teacher characteristic variables are available in the data, such as teachers’ experience, the teacher variables tend to be highly correlated, especially with a variable indicating that the school is a government school. Teacher experience, teacher salary, and working in a government school are positively correlated. Regression results that excluded the teachers’ salary variable were no different from the regression results of Eqn. (1). The school processes variables include the number of inspections per year by government officials, whether parents attend meetings at the school, and whether students are given daily homework assignments. The physical characteristics variables include whether the school has electricity, the ratio of blackboards to teachers, and whether classrooms were classified by the survey interviewers as being in poor condition. To capture family characteristics, the regression includes the average per capita expenditure of non-beneficiary families in the school. This variable is calculated using the household survey and is aggregated at the school level and matched to the school test score data. In cases where no data were available for household per capita expenditure for a school, the mean (796 taka) was assigned to that observation. About 30% of the observations were missing data on school-level per capita expenditure of non-beneficiary families. Controlling for family characteristics was important because the estimated impacts of being in an FFE school on student achievement could be a

675

proxy for household characteristics that differ for non-beneficiary families in FFE schools and non-beneficiary families in non-FFE schools if non-beneficiary families in FFE schools come from families with higher socioeconomic status than all families in non-FFE schools. As McEwan (2003) discusses, families who live in the same geographic area and whose children attend the same school are likely to have common unobserved characteristics. If family characteristics are not controlled for, the estimated peer effects may actually be picking up unobservable family characteristics such as preferences for children’s education. McEwan found that the average of mothers’ education within a classroom had a large impact on children’s school achievement, so we also took the average of fathers’ education and of mothers’ education for the non-beneficiary families. Including the average of parents’ education did not have a significant impact on children’s achievement test scores and did not change the estimates in any meaningful way. Finally, the regressions include dummy variables for the unions where the children lived. The FFE program was targeted at the union level. The government attempted to give priority to poorer unions in the FFE program, so the regression controls for these union characteristics in the most comprehensive manner. If the variables were left out, the estimates of the impact of attending an FFE school would be biased because the error term would be correlated with the FFE school dummy variable. This set of union dummy variables also controls for the fact that students who are nonbeneficiaries in schools that do not have FFE are on average likely to be poorer than students who are non-beneficiaries in schools that have FFE. Some poor children go to school even if they do not receive an FFE ration. Adding the union-level dummy variable is equivalent to controlling for any union-level characteristic, both observed and unobserved. However, because all schools in a union were eligible for FFE, provided that they met the government’s criteria, the interpretation of the FFE school variable is complicated. When controlling for union dummy variables, the regression estimates the FFE school coefficient based on differences between FFE schools and nonFFE schools within a union. Therefore, the coefficient is likely to reflect selection into the FFE program rather than the impact of the program on school quality and student achievement.

676

WORLD DEVELOPMENT

The statistical analysis takes into account the nature of the dependent variable and the survey sampling design so as to make correct statistical inferences. The achievement test score ranges from 0 to 10 for math and from 0 to 9 for Bangla, and therefore the regression is estimated as a tobit model. The standard errors are also corrected for sampling effects—in this case, that the sample was stratified to include both unions that participated in the FFE program and did not participate, and that the random sampling occurred at the union level and not at the individual level. 7 Endogeneity problems could arise in the econometric model specification in Eqn. (1) if school characteristics and school achievement test scores were both caused by characteristics that were not observed by the researcher. In Appendix A, we argue that in the rural Bangladesh setting, endogeneity is not the problem that often arises in other settings. (b) Results and discussion The mean, minimum, and maximum values for the variables used in the regression analysis are presented in Table 6. The results of the tobit

model estimation are presented in Table 7. To aid interpretation, the predicted achievement scores from a series of simulations that use the estimated coefficients from Table 7 are presented in Table 8. For each simulation, we change only the variables that are the focus of the specific analysis. All other variables are held at their actual values. Boys did significantly better in the math achievement tests than girls did. However, for Bangla test scores, there was no significant difference between girls and boys. Although FFE increased the enrollment rates of girls relative to boys, girls lagged behind in achievement in math. For both math and Bangla achievement test scores, the effects of attending an FFE school were positive, but not statistically significant. When the union dummies were dropped from the regression, the effects of attending an FFE school were positive and statistically significant for both math and Bangla. The fact that the effects became statistically insignificant when the union dummy variables were added suggests that those schools that became FFE schools might have been better than schools that did not participate in the program before the pro-

Table 6. Descriptive statistics for variables used in regression analysis, non-beneficiary students, rural Bangladesh, 2000 Variable

Mean

Minimum

Maximum

Math achievement test score Bangla achievement test score Male student School has FFE program Proportion of students who participate in FFE in grade 4 Measure of crowding—total number of children observed in grade 4 classroom Proportion of schools classified as ‘‘good’’ (A) Proportion of schools classified as ‘‘fair’’ (B) Proportion of schools classified as ‘‘poor’’ (C) Proportion of female teachers Proportion of teachers with at least a Bachelors degree Teachers’ average salary per month Proportion of teachers who report other sources of income Number of inspections in last school year Parental participation—parents come to school meetings Children are given daily homework School has electricity Classrooms in poor condition as observed by survey enumerators Number of blackboards per teacher in school Mean per capita expenditure of families of non-beneficiary students in school Proportion of observations where mean per capita expenditure was missing and value was imputed N

5.23 6.47 0.47 0.58 0.25 53.89

0 0 0 0 0 4

10 9 1 1 0.71 104

0.38 0.49 0.13 0.31 0.31 4,647.14 0.15 5.80 0.77 0.85 0.36 0.04 1.00 844.26

0 0 0 0 0 1,300.00 0 0 0 0 0 0 0 293.76

1 1 1 1 1 10,577.33 1 14 1 1 1 1 14.8 2,370.41

0.31

0

1

1,823

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

677

Table 7. Tobit regression analysis of the impact of FFE on the fourth grade achievement test scores of non-beneficiary students in government schools, rural Bangladesh, 2000 Variable

Male student School has FFE program Proportion of students who participate in FFE in grade 4 Measure of crowding—total number of children observed in grade 4 classroom School quality classification (Good is omitted) Fair school quality Poor school quality Teacher characteristics at the school level Proportion of female teachers Proportion of teachers with at least a bachelor’s degree Teachers’ average salary per month (in ‘000s taka) Proportion of teachers who report other sources of income School processes Number of inspections in last school year Parental participation—parents come to school meetings Children are given daily homework Physical characteristics of schools School has electricity Classrooms in poor condition as observed by survey enumerators Number of blackboards in school per teacher Average per capita expenditure for non-beneficiary families in the school (in ‘000s taka) F-test for significance of union dummies F(17, 12) N Log sigma (goodness of fit)

Math scores

Bangla scores

Coefficient

t-Test

Coefficient

t-Test

1.359** 4.128 10.387***

2.38 1.17 2.79

0.347 3.941 8.247***

1.33 1.47 2.73

0.021

1.00

0.001

0.04

2.080* 1.112

2.03 0.89

1.922** 1.328

2.38 1.58

0.174 0.753

0.10 0.70

0.542 0.423

0.48 0.48

0.563** 1.021

2.12 0.62

0.161 0.507

1.20 0.41

0.003 0.155

0.02 0.18

0.027 0.630

0.32 1.14

1.301

1.18

1.670**

2.07

0.205 0.133

0.18 0.08

0.909* 1.719**

1.89 2.09

0.243** 1.093

2.46 0.84

0.093 0.813

1.20 0.83

26.94

683.40*** 1,823 1.356***

23.74

209.39*** 1,823 1.543***

*

Significant at the 10% level. ** Significant at the 5% level. *** Significant at the 1% level.

gram started. Therefore, the positive effects of attending an FFE school were driven by selection effects. The positive, but statistically insignificant effects of attending an FFE school were mitigated by the negative and statistically significant ‘‘peer effects’’ on achievement test scores of non-beneficiary students, which arose from the percentage of students in the classroom who receive FFE benefits. For both math and Bangla scores, the peer effects were negative and highly statistically significant (Table 7). These negative peer effects were robust to excluding union dummy variables and measures of school quality in the regressions. Negative and significant peer effects were also

found when the dependent variable was the combined math and Bangla score and when the dependent variable was the total test score (Appendix Table 1). The simulation is useful to see whether the positive effect (although statistically insignificant) of going to an FFE school is larger in magnitude than the negative impact of having a non-zero percentage of children receiving FFE benefits. The simulations used the coefficients from Table 7 and present different scenarios. For example, the first row of Table 8 presents the predicted achievement test score that would be obtained if no schools had the FFE program. To calculate the predicted score, the variables ‘‘FFE school’’ and ‘‘proportion of

678

WORLD DEVELOPMENT

Table 8. Simulations of predicted values of learning achievement under various scenarios, coefficients from Table 7 Math Predicted percentage of correct answers Actual mean test scores in the sample (Table 6) School has no FFE program, 0% of students receive FFE benefits School has FFE program, 44% of students receive FFE benefits (mean value) School has FFE program, percentage of students who receive FFE benefits increases by 10 percentage points to 54% Teacher’s average salary is 4,647 taka Teacher’s salary is raised by one standard deviation to 5,905 taka School has 0 blackboards School has 1 blackboard per teacher School doubles the number of blackboards to 2 blackboards per teacher Children are not given daily homework Children are given daily homework Classroom is in poor condition Classroom is not in poor condition School has no electricity School has electricity

Bangla

Percentage change from baseline

52.3a

Predicted percentage of correct answers

Percentage change from baseline

71.9a

55.2

–

82.0

–

50.8

8.0

85.6

4.3

40.4

26.8

76.3

6.9

53.2 46.1

– 13.3

– –

– –

50.8 53.2 55.7

– 4.7 9.6

– – –

– – –

– – – – – –

– – – – – –

100.0 81.8 66.2 85.2 80.9 91.0

– 18.5 – 28.9 – 12.5

All variables are held at their actual values in the data, unless specified. a Actual means, not predicted means.

children in fourth grade who receive FFE benefits’’ were set equal to 0 and then the mean score was calculated for all the non-beneficiary students in the sample. All other variables were kept at their actual values. If no schools had FFE, on average students would get 55% of math questions correct and 82% of Bangla questions correct. Although the variable ‘‘attends FFE school’’ was not significant in either the math or Bangla achievement test score regressions, we changed the value of the variable in the simulations. We changed both variables in the simulations because the proportion of children who received FFE was inextricably bound up with whether the school had an FFE program or not. The regression analyses included both variables and if one was dropped, the estimated magnitudes of the other variable changed. Therefore, it is appropriate to change both variables together in the simulations. In the second row of Table 8, the value of ‘‘FFE school’’ was set to 1 and ‘‘the percentage

of children who receive FFE benefits’’ was set to 44, which was the average percentage of children in fourth grade who received FFE benefits in schools that had the FFE program. A comparison of the first and second rows shows that if all schools had FFE and had 44% of children in FFE, then the net impact of FFE would be to decrease math achievement test scores of non-beneficiaries from 55.2% to 50.8%, or by 8%. For Bangla achievement test scores, the net impact of going from having no FFE program to having an FFE program and 44% of students participating was to increase Bangla test scores from 82.0% to 85.6% correct, or an increase of 4.3%. Therefore, the negative peer effect of raising the proportion of children who received FFE in the classroom was stronger for math test scores than for Bangla test scores. The simulations also show what would happen if the percentage of children who received FFE increased in an FFE school. Comparing rows 2 and 3 of Table 8, when the percentage

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

of students receiving FFE increased by 10 percentage points from 44% to 54%, the predicted math test score fell from 50.8% correct answers to 40.4% correct answers, or a decrease of 20.5%. For Bangla, increasing the percentage of students who received FFE from 44% to 54% resulted in a decrease in the percentage of correct answers from 85.6% to 76.3%, or a decrease of 10.9%. The class size or crowding effect was not statistically significant for either math or Bangla test scores. As the number of children in fourth-grade classrooms increased, there was no statistically significant impact on achievement test scores. This finding was consistent with some other findings in the literature, which also found little impact of class size on student achievement (Hanushek, forthcoming; Michaelowa, 2001). In the rural Bangladesh context, the finding was not likely due to disruptive students being placed in smaller classrooms, because most rural primary schools did not have the resources to provide more than one class at a grade level. These schools were not accountable to parent associations and parents had little say in school decisions. Therefore, class size was a more random variable in the rural Bangladesh context than it is in other countries, and results are less likely to be biased. The finding suggests that the negative impacts of FFE on student achievement operated primarily through peer effects, and not through class size. The impacts of school quality on both math and Bangla test scores were positive and significant, with children attending good schools (the omitted category) scoring higher on achievement tests than children who attended fair schools. In the case of math scores, the effect was only marginally statistically significant. Children who attended poor schools did no worse in math and Bangla than children who attended good schools. Of the four teachers’ characteristics variables, only the teachers’ average salary per month had a statistically significant effect on achievement test scores, and this effect was only found for math test scores (Table 7). Children obtained lower scores on achievement tests the higher were their teachers’ salaries. The effect was not negligible in magnitude. Increasing teachers’ salaries from the average of 4,647 taka per month by a standard deviation to 5,905 taka per month would result in a decrease in the predicted percentage of correct math answers from 53.2% to 46.1%, a decrease of

679

13.3% (Table 8). This result is interesting and unexpected. One possible explanation is that government schools with high teachers’ salaries attracted more people who wanted to be in the job because of the pay. In rural Bangladesh, government jobs are highly coveted, and it is likely that personal influence rather than teachers’ quality played a role in obtaining the jobs. In any case, the finding is suggestive that raising teachers’ salaries would not necessarily attract better quality teachers into the school. Of the three school processes variables, the only statistically significant variable was whether the teacher assigned daily homework in the Bangla test score regression. Contrary to expectations, if all children attended schools that assigned daily homework, Bangla test scores would be 18.5% lower than they would be if no children attended schools that assigned daily homework (Table 8). This effect was statistically significant at the 5% level. According to Table 6, 85% of the children in our sample were given daily homework in their schools. Perhaps schools where children were lagging behind in Bangla were especially likely to give daily homework. Given the low percentage of children who were not given daily homework, there was likely something unique about the schools that did not assign daily homework that was not measured by the interviewers. None of the school process variables had a statistically significant impact on children’s math test scores. The physical characteristics of schools that had statistically significant impacts on math test scores were different from those characteristics that impacted Bangla test scores. The differences are interesting and suggest different policy instruments to improve math and Bangla scores. For math test scores, the ratio of blackboards per teacher had a positive and significant impact. The impact was modest in magnitude. When no schools had blackboards, the predicted math score was 50.8%. If all schools had 1 blackboard per teacher, the predicted score would increase to 53.2%, an increase of 4.7%. If all schools had 2 blackboards per teacher, the predicted math score would increase to 55.7% (Table 8). Other studies by Glewwe and Jacoby (1994) in Ghana and Michaelowa (2001) in five Sub-Saharan African countries have also found positive and significant, but small in magnitude, impacts of blackboards on student achievement. Blackboards had significant impact on math scores, but not on Bangla scores, which might indicate

680

WORLD DEVELOPMENT

the importance of working out math problems on a blackboard for children’s learning. Two school characteristics had positive and statistically significant impacts on Bangla scores. If all children attended school in classrooms that were rated as being in poor condition by interviewers, the predicted percentage of correct answers on the Bangla exam would be 66.2. If no classrooms were rated as being in poor condition, the children’s predicted scores would increase by 29% to 85.2. However, only 4% of the children in the sample attended schools in classrooms that were rated as poor, so there is not a great scope for raising test scores by upgrading the classroom conditions from poor. Having electricity had a marginally statistically significant impact on Bangla scores. If no children attended a school with electricity, the children would be predicted to get 80.9% of the Bangla questions correct. If all children attended a school with electricity, they would be predicted to get 91% of the Bangla questions correct, an increase of 12.5%. The result should be interpreted with caution, however, because having electricity could be correlated with other unobserved characteristics of the school, such as the neighborhood where it is located. Providing electricity to schools might also be very expensive compared to other interventions. 8. CONCLUSIONS AND POLICY IMPLICATIONS The evidence is clear from past studies that the Food for Education program in Bangladesh was very successful at getting poor students enrolled in school, especially girls. However, because Bangladesh did not invest in school resources at the same rate at which enrollment increased, class sizes increased. Parents, teachers, and policymakers expressed concern about decreasing quality of FFE schools, specifically about the perceived negative impact of crowding in classrooms on student achievement. The results of our analysis at the school-level reveal that the average test scores were lower in FFE schools than in non-FFE schools. About 40% of students in FFE schools were program beneficiaries and the rest were non-beneficiaries. Within FFE schools, the average score of FFE beneficiaries was less than that of the non-beneficiaries, which brought down the aggregate score in FFE schools. FFE beneficia-

ries had lower scores than non-beneficiaries, probably because of their relatively lower socioeconomic status. To investigate the possible negative impact of FFE on test scores, student achievement test data collected in schools are analyzed in a multivariate framework. We focus on the impact of FFE on non-beneficiary students’ achievement. Our approach allows us to distinguish among the effects of students being enrolled in an FFE school, having a higher percentage of FFE beneficiary children in the classroom, and having a larger class size. The results of our multivariate analysis reveal that the class size had no statistically significant effect on student achievement in rural Bangladesh. This finding negated the assertion that the increased number of students in FFE school classrooms reduced learning. However, as the percentage of students who received FFE grows, test scores of non-beneficiary students in FFE schools decreased, implying that there were negative peer effects of FFE on non-beneficiary students. For example, the FFE beneficiary students were poorer and, if they were inspired by FFE to enroll in school, less academically experienced than the non-beneficiary students; therefore, teachers may have had to give more attention to them than the nonbeneficiary students. The negative impact of the FFE program on learning of non-beneficiary students operated primarily through peer effects, and not through class size. The peer effects result is suggestive and merits further study. Future data collection efforts should include more complete data on children’s socioeconomic backgrounds. This would allow us to say more definitively that the estimated effect is due to peer effects and not due to omitted household characteristics. The results suggest that an investment in blackboards might increase achievement test scores in math. Such an investment would be relatively low cost. The results for Bangla achievement test scores suggest that the few schools that were in poor physical conditions should be upgraded, if the costs are not high. Providing electricity to schools might also modestly improve Bangla test scores, but this is likely to be expensive. Clearly, there is much room for improvement in the quality of Bangladesh’s government schools. The government has been successful in increasing enrollment through its FFE program, but now must focus on improving the nation’s schools. Unlike programs like Progresa

DO CROWDED CLASSROOMS CROWD OUT LEARNING?

in Mexico, FFE did not involve large investments in school quality. The result that class size did not have a significant effect on student achievement but peer effects did suggests that building new schools is not the solution to the problem. It is very difficult to get an estimate of the cost of building a new school in Bangladesh because costs vary widely, but school expansion is likely to be expensive relative to other options. The operating costs per year for a school are about $7,000. Recently, economists and policymakers have been conducting controlled experiments in developing countries to determine which policies improve student learning. An experiment in the urban slums of India hired young women from the community to teach basic literacy and numeracy skills to children who were falling behind in government schools. The children went to be tutored for half of the school day. This program increased average test scores of all children in the treatment schools by 0.14 standard deviations in the first year and 0.28 in the second year. In another experiment, children who were behind spent two hours a week playing computer games that reinforced math skills. This program increased math scores by 0.36 standard deviations in the first year and 0.54 in the second year (Banerjee, Cole, Duflo, & Linden, 2004). Hiring young women to tutor students was especially cost effective. The class size was 15–20 students, and the young women were secondary school graduates who were paid the equivalent of $10–15 per month, a much lower salary than teachers (Banerjee et al., 2004). Hiring 10 tutors for 10 months would cost a total of $1,000 to $1,500, compared to $7,000 in operating costs alone for a new school. Another cost would be the cost of separate space for the tutors to meet with students.

681

The tutoring program tried in India is a form of student tracking by achievement. Low-achieving students are separated for at least part of the school day from high-achieving students. In the US literature, tracking by achievement has fallen into disfavor because of research that suggested that tracking was harmful to disadvantaged students (Figlio & Page, 2002). However, Figlio and Page criticized this literature and showed that once they controlled for the fact that US parents choose schools based on whether or not a school tracks, tracking did not have a negative impact on low-ability students, and in fact, may have had a positive impact (Figlio & Page, 2002). The government of Bangladesh should consider instituting carefully designed experiments to tutor low-achieving students and to track students by achievement. One problem might be the scarcity of young women with secondary school education in rural areas who are willing to work, given Bangladesh’s low levels of female labor force participation, but perhaps the program might work by hiring primary school graduates. If these experiments show positive results, they would be easy to scale up. Finally, setting clear standards for performance is important for all levels of education. In Bangladesh, although performance indicators are embodied in certification examinations at the secondary and higher levels, they are neglected at the primary level. Carefully devised, minimum performance standards might be incorporated in the design of the recently implemented Primary Education Stipend program, a cash-for-education program that has replaced the FFE program. Such standards might encourage both beneficiary and non-beneficiary students to learn more in the classroom and outside the classroom.

NOTES 1. The administrative structure of Bangladesh consists of divisions, districts, thanas, and unions, in decreasing order by size. There are six divisions, 64 districts, 489 thanas (of which 29 are in four city corporations), and 4,451 unions (all rural). The FFE program was implemented in all 460 rural thanas. 2. The official exchange rate for the taka (Tk), the currency of Bangladesh, was Tk 51.00 per US$1.00 in June 2000.

3. Of the total quantity of FFE foodgrain distributed from 1997–98 to 1999–2000, wheat accounted for about 64%, and rice, about 36%. 4. Quartile groups are based on household quartiles ranked by total per capita expenditures. ‘‘Expenditure quartile’’ should be understood to mean households in any of the two strata-FFE unions and non-FFE unions. In this study, we use per capita expenditure as a proxy for income for two reasons. First, expenditures are

682

WORLD DEVELOPMENT

likely to reflect permanent income and are, hence, a better indicator of consumption behavior (Friedman, 1957). Second, data on expenditures are generally more reliable and stable than income data. Because expenditures are intended to proxy for income, the terms ‘‘expenditure’’ and ‘‘income’’ will be used interchangeably.

5. Per capita monthly expenditures of FFE beneficiary households exclude the income transfer from the FFE program.

6. The minimum quality standards are: (1) the school must have a government classification of at least ‘‘C’’ on a scale from ‘‘A’’ to ‘‘D,’’ (2) at least 10% of grade 5 students must qualify for the annual scholarship examination, (3) students in grades 3, 4, and 5 should obtain at least 40% of total points in the previous year’s annual examination, and (4) the FFE ration is suspended for any school in which a random inspection reveals less than 60% attendance, until the attendance record improves. 7. The regressions were estimated using the ‘‘svyintreg’’ command in STATA version 8.2.

REFERENCES Ahmed, A. U. (1992). Operational performance of the rural rationing program in Bangladesh. Working Paper on Bangladesh No. 5, International Food Policy Research Institute, Washington, DC. Ahmed, A. U. (2000). Targeted distribution. In R. Ahmed, S. Haggblade, & T. E. Chowdhury (Eds.), Out of the shadow of famine: Evolving food markets and food policy in Bangladesh (pp. 213–231). John Hopkins University Press. Ahmed, A. U., & Billah, K. (1994). Food for education program in Bangladesh: An early assessment. Bangladesh Food Policy Project Manuscript No. 62, International Food Policy Research Institute, Washington, DC. Ahmed, A.U., & del Ninno, C. (2002). The food for education program in Bangladesh: An evaluation of its impact on educational attainment and food security. Food Consumption and Nutrition Division Discussion Paper No. 138, International Food Policy Research Institute, Washington, DC. Angrist, J., & Lavy, V. (1999). Using Maimonides’ rule to estimate the effect of class size on scholastic achievement. Quarterly Journal of Economics, 114(2), 533–575. Arends-Kuenning, M., & Amin, S. (forthcoming). School incentive programs and children’s activities: The case of Bangladesh. Comparative Education Review. BANBEIS (Bangladesh Bureau of Educational Information and Statistics) (2002). Statistical profile on education in Bangladesh. Dhaka: Bangladesh Bureau of Educational Information and Statistics. Banerjee, A., Cole, S., Duflo, E., & Linden, L. (2004). Remedying education: Evidence from two randomized experiment in India. Working Paper, MIT. BIDS (Bangladesh Institute of Development Studies) (1997). An evaluation of the food for education program: Enhancing accessibility to and retention in primary education for the rural poor in Bangladesh. Dhaka. Case, A., & Deaton, A. (1999). School inputs and educational outcomes in South Africa. Quarterly Journal of Economics, 114(3), 1047–1084. DPC (Development Planners and Consultants) (2000). Comprehensive assessment/evaluation of the food

for education program in Bangladesh. A report prepared for the Primary and Mass Education Division, Food for Education Program Project Implementation Unit, Dhaka. Figlio, D., & Page, M. (2002). School choice and the distributional effects of ability tracking: Does separation increase inequality? Journal of Urban Economics, 51, 497–514. Friedman, M. (1957). A Theory of the Consumption Function. Princeton, NJ: Princeton University Press. Glewwe, P. (2002). Schools and skills in developing countries: Educational policies and socioeconomic outcomes. Journal of Economic Literature, 40(2), 436–482. Glewwe, P., & Jacoby, H. (1994). Student achievement and schooling choice in low-income countries: Evidence from Ghana. Journal of Human Resources, 29(3), 843–864. Hanushek, E. A. (1998). The evidence on class size. Occasional Paper No. 98-1, W. Allen Wallis Institute of Political Economy, University of Rochester. Hanushek, E. A. (forthcoming). The failure of inputbased schooling policies. Economic Journal. Khandker, S. R. (1996). Education achievements and school efficiency in rural Bangladesh. World Bank Discussion Paper 319, World Bank, Washington, DC. Krueger, A. (2002). Economic considerations and class size. NBER Working Paper No. 8875, National Bureau of Economic Research, Cambridge, MA. Lazear, A. P. (1999). Educational production. NBER Working Paper No. 7349, National Bureau of Economic Research, Cambridge, MA. McEwan, P. (2003). Peer effects on student achievement: Evidence from Chile. Economics of Education Review, 22, 131–141. Michaelowa, K. (2001). Primary education quality in Francophone Sub-Saharan Africa: Determinants of learning achievement and efficiency considerations. World Development, 29(10), 1699–1716. Mingat, A., & Suchaut, B. (1998). Une Analyse Economique Comparative Des Syste`mes Educatifs Africaines. Rapport re´alise´ pour le Ministe`re Franc¸ais des Affaires Etrange`res, Coope´ration et Francophonie, Paris.

DO CROWDED CLASSROOMS CROWD OUT LEARNING? PMED (Primary and Mass Education Division) (2000). Project report: Food for education program (in Bangla). Project Implementation Unit, Food for Education Program, Dhaka. Ravallion, M., & Wodon, Q. (1997). Evaluating a targeted social program when placement is decentralized. Washington, DC: World Bank.

APPENDIX A. POTENTIAL SOURCES OF ENDOGENEITY Endogeneity problems could arise in the econometric model employed in this study if school characteristics and school achievement test scores were both caused by characteristics that were not observed by the researcher. For example, if parents who place a high value on their children’s schooling spend more time doing homework with their children and also choose where to live based on school resources, then, the impact of school resources on children’s achievement test scores is overestimated. Glewwe (2002) discusses at length the difficulties of estimating the impact of school resources, especially class size, on student achievement. We argue that in the rural Bangladesh setting, endogeneity is not the problem that often arises in other settings. In the Bangladesh FFE program case, one might imagine that the following endogeneity problems would arise: the government allocates the FFE program and school resources to low-performing areas, leading to an underestimate of the impact of school resources on student achievement; parents choose which schools their children attend, so that motivated parents choose schools with more resources; and parents who care about schooling are able to get more government resources to their children’s schools. We explain in turn why each of these situations does not create an endogeneity problem for our study. The FFE program is targeted to poor rural areas. The district administration gets funds, which are allocated to the thana-level officials, the lowest level of government administration in Bangladesh. At the thana level, unions are chosen to participate based on their socioeconomic levels and their literacy rates. In the econometric model specification in Eqn. (1), we control for this selection by including union dummy variables, which control for all observed and unobserved characteristics of the union, including those that are used to select

683

the unions for the program. So, the targeting of FFE to poor areas is controlled for in the regression, as is any other targeting of school resources that occurs at the union level. All selection of schools into the FFE program occurs at the union level. When a union is chosen for the FFE program, basically all the schools within the union are eligible, provided that they meet the regulations. All the government and registered non-government schools participate. Only one madrasa within the union is eligible, but only about 4% of students in our census attend a madrasa. The Bangladesh Rural advancement Committee (BRAC) is an NGO that provides schools in villages at the early primary level. BRAC opted not to participate in FFE, but NGOs only enroll about 6% of the primary-school students. The next potential source of endogeneity arises from parental choice of their children’s schools. In the rural areas, where the data were collected, often only one school is available in the village. The census data collected for this study show that about 80% of children go to a school within their village, and the remaining 20% go to a school within their union. Seventytwo percent of students attend government schools, and about 13% go to non-government schools. Students in rural Bangladesh do not have much choice in the schools that they attend. In the sample of school test scores, 64% of the children lived in a village where there was only one school, and an additional 20% lived in a village where there were two schools. Therefore, in this setting, the school choice does not create as large of an endogeneity problem as it would in other settings, such as an urban area of Bangladesh. The final source of potential endogeneity arises from parents being able to organize and obtain the political power to improve schools. Even if only one school is available in a village, parents who care about their children’s education might be able to pressure the government to invest more resources in the local school or to build another school when the current school gets crowded. This scenario is unlikely in rural Bangladesh. Parents have little experience with school, and so are unable to judge school quality. Table 4 shows that in FFE unions on average, fathers have 2.3 years of schooling and mothers have 1.2 years of schooling. The people who live in the villages are mostly poor and have little political clout. Schooling resource decisions are made at higher levels of

684

WORLD DEVELOPMENT

Appendix Table 1. Tobit regression analysis of the impact of FFE on the fourth grade achievement test scores of non-beneficiary students in government schools, rural Bangladesh, 2000 Variable

Male student School has FFE program Fraction of students who participate in FFE in grade 4 Measure of crowding—total number of children in grade 4 classroom School quality classification (Good is omitted) Fair school quality Poor school quality Teacher characteristics at the school level Fraction of female teachers Fraction of teachers with at least a bachelor’s degree Teachers’ average salary per month (in ‘000s taka) Fraction of teachers who report other sources of income School processes Number of inspections in last school year Parental participation—parents come to school meetings Children are given daily homework Physical characteristics of schools School has electricity Classrooms in poor condition as observed by survey enumerators Number of blackboards in school per teacher Average per capita expenditure for non-beneficiary families in the school (in ‘000s taka) F-test for significance of union dummies F(17, 12) N Log sigma (goodness of fit) *

Math + Bangla scores

Total scores (Math, Bangla, English and Environment)

Coefficient

t-Test

Coefficient

t-Test

1.211** 5.507 12.880**

2.22 1.25 2.70

1.647** 5.910 15.439**

2.54 1.02 2.26

0.016

0.65

0.015

0.44

2.863** 1.975

2.41 1.46

3.136* 1.590

2.03 0.89

0.749 0.754

0.38 0.60

1.325 1.461

0.43 0.83

0.253 0.440

0.84 0.23

0.408 0.694

0.91 0.27

0.007 0.626

0.06 0.53

0.021 1.497

0.09 1.06

0.088

0.06

0.606

0.31

0.607 1.002

0.53 0.59

1.898 2.272

1.11 1.09

0.247* 0.206

2.02 0.13

0.241 0.446

1.55 0.25

1,823 1.669***

44.45

1,823 1.915***

47.95

Significant at the 10% level. Significant at the 5% level. Significant at the 1% level.

**

***

government, not at the local levels, and the administrators of these programs often have little incentive to respond to village needs. To conclude, endogeneity of school resources and class size is not as much of a problem in our study setting as it would be in other set-

tings. The FFE program created variation across schools in class size, and this variation allows us to estimate the impacts of being in an FFE school, having a high proportion of students who receive FFE, and classroom crowding on school achievement.