The Effects of Nationwide Tuition Fee Elimination on Enrollment and Attainment∗ Robert Garlick† September 28, 2017
Abstract I study the effects on education outcomes of nationwide primary and secondary school fee eliminations in South Africa. This policy shifted education financing from a mixed user fee and government transfer system to a pure transfer system. This mirrors policy debates in developed and developing countries about the optimal mix of school fees, (subsidized) loans, and transfers to finance primary, secondary, and tertiary education. I find that fee elimination has a small positive effect on enrollment, a small negative effect on secondary school graduation, and nearzero effects on grade progression, per-student school resources, and the socio-economic profile of the enrolled students. My results are robust to accounting for school-level selection into fee elimination, differential time trends between fee-eliminating and fee-charging schools, and student transfers between fee-charging and fee-eliminating schools. The enrollment effects imply a price elasticity of -0.10 for secondary school enrollment and 0 for primary school enrollment. This price insensitive demand is not explained by ceiling effects on enrollment, capacity constraints in schools, or measurement error in administrative data. I argue that the pattern of results may reflect low valuation of additional years of education by youths living near fee-eliminating schools, potentially due to a weak relationship between enrollment and subsequent attainment. My results show that when education quality is low, demand-side subsidies may have limited ability to increase education participation and can even lower attainment.
JEL codes: I25, O15
∗
I am grateful for helpful suggestions from Manuela Angelucci, John Bound, Jishnu Das, John DiNardo, Erica Field, Deon Filmer, Brian Jacob, David Lam, Jeffrey Smith, Stephen Taylor, Duncan Thomas, and conference and seminar participants at Bristol, CSAE, Duke, ESSA, IFPRI, Michigan, MIEDC, NBER Economics of Education, NEUDC, Toronto, UC Irvine, and UCSD. Rirhandzu Baloyi, Justice Libago, Christo Lombaard, Erna Lubbe, Hersheela Narsee, Ralph Mehl, Siza Shongwe, Stephen Taylor, and Hylton Visagie from South Africa’s Department of Basic Education provided invaluable assistance in obtaining and interpreting the data used in this project. All errors are my own. † Duke University,
[email protected].
1
1
Introduction
Almost all governments around the world provide some form of public education and must choose how to finance this public education. Different countries and regions employ different mixes of user fees, typically called “school fees” or “tuition fees,” and transfers from the national or regional budget. Many countries have shifted this funding mix through time: school fees for tertiary education in Germany, the UK, and US have changed substantially in recent years; tax incentives for education spending in the US have changed over time; and school fees for primary and secondary education have changed substantially in Ghana, Kenya, Nigeria, and South Africa. Different financing mixes may lead to different levels of enrollment in public education, learning, and hence labour force composition. The optimal mix of school fees and transfers is theoretically ambiguous. School fees may promote accountability of public education providers to students and their families, and may screen out students with low returns to enrollment. However, school fees may reduce enrollment for credit-constrained or myopic students with high returns to enrollment. Transfers from government budgets may have distortionary effects on labour supply through the tax system. Given this theoretical ambiguity, empirical evidence on the effects of different education finance mixes is particularly important. I study the effects of a nationwide shift in public education financing in South Africa. Primary and secondary public schools were historically funded by a mix of mandatory school fees and government transfers.1 Between 2007 and 2010, school fees were eliminated in approximately 75 percent of public primary and secondary schools. Fees were first eliminated in neighborhoods with poverty rates above a threshold value, generating both time-series and cross-sectional variation in fee-charging status. I use panel data and regression discontinuity methods with administrative and household survey data to study the effects of fee elimination. I find small increases in enrollment (≈ 1 percent of baseline enrollment), concentrated in the first two grades of secondary school. I find a small but significant fall in the number of students completing secondary school, and negligible effects on school resources (class sizes, socio-economic status of enrolled students) or learning outcomes (grade progression). I find no evidence of transfers between fee-eliminating schools and nearby fee-charging schools. These enrollment effects imply that the price elasticity of demand for enrollment in public schools is at most -0.03, though this rises to -0.1 for secondary school enrollment. I show that the inelastic demand is not explained by ceiling effects on enrollment, capacity constraints in schools, noncompliance with the fee elimination policy, or measurement error in administrative data on enrollment. Instead, I hypothesize that demand is price inelastic in this setting because students in fee-eliminating schools derive little value from additional years of enrollment. They face low 1
School fees for primary and secondary public schools are common in many developing countries, particularly in Africa. Primary school fees have been eliminated in many countries over the past two decades, and several countries are actively considering eliminating secondary school fees at the time of writing.
2
probabilities of grade progression and low probabilities of secondary school graduation. I cannot directly test this hypothesis, but it is consistent with treatment effects being concentrated early in secondary school, negative effects on secondary school graduation, and with other research showing that grade progression is low and uncertain in South African secondary schools in high-poverty neighborhoods (Lam, Ardington, and Leibbrandt, 2010; Van der Berg and Louw, 2007). Low valuation of enrollment for marginal students is perhaps surprising, given that Mincer returns to additional years of secondary education in South Africa exceed 10% (Branson, Ardington, Lam, and Leibbrandt, 2013).2 This contrasts with findings from multiple countries that returns to enrollment in school and university are relatively high for financially constrained individuals, and not lower than Mincer returns (Card, 1999; Duflo, 2001; Heckman, Lochner, and Todd, 2006). I cannot directly determine why my results break this pattern. Results in other countries may be driven by marginal students attending reasonably high quality schools, while my results are driven by marginal students attending low quality schools. This paper relates and contributes to four literatures.3 The first literature examines the effect of demand-side incentives to enroll in (mostly primary) school in developing countries. These incentives include school fee elimination (Deininger, 2003; Fafchamps and Minten, 2007; Lucas and Mbiti, 2012), reductions in non-fee costs of enrollment (Evans, Kremer, and Ngatia, 2009; Hidalgo, Onofa, Oosterbeek, and Ponce, 2013), and conditional cash transfers (Baird, McIntosh, and Ozler, 2011; Filmer and Schady, 2008; Schultz, 2004).4 The literature on conditional cash transfers is particularly large and includes many well-identified studies.5 However, conditional cash transfers are a pure demand-side subsidy, while school fee elimination combines a price ceiling with a supply-side subsidy and may change school revenues, so the two types of policies may have substantially different effects. My paper is most closely related to studies by Barrera-Osorio, Linden, and Urquiola (2007) in Colombia and Borkum (2012) in South Africa, which study the enrollment effects of school fee elimination using natural experiments. I extend this literature by considering a wider range of outcomes, interpreting treatment effects in terms of an economic model of human capital investment, and exploring the relationship between treatment effects and labour market returns to education.6 My results suggest that large demand-side subsidies can be ineffective in increasing enrollment when education quality is low, reinforcing recent concerns that about poor 2 Branson and Leibbrandt (2013) find that these large and convex returns to education persist even in Mincer regressions that control for proxies of school quality in the area where respondents grew up. 3 There is another related literature that examines the effect of user fees on investments in health, including use of public clinics. This literature considers a range of user fees from insurance co-pays to direct clinic fees. Some papers in this literature explicitly examine the role of user fees in screening low- versus high-valuation considers (Dupas, 2014). 4 See Kremer and Holla (2009) for a review of this literature. Most studies of fee elimination find relatively large effects on enrollment, though several rely on before-after comparisons that face identification challenges. 5 See Fizbein and Schady (2009) for a review of this literature. 6 Relative to Borkum’s study of the South African experience, I also use a dataset with wider geographic and longer time coverage, and incorporate multiple sources of household survey data.
3
learning outcomes in developing country schools (Pritchett, 2013). My results also show that school fee elimination can be very expensive relative to the number of marginal students, consistent with a result for conditional cash transfers in Todd and Wolpin (2006). The second literature examines the effect of price variation, including demand-side subsidies, on enrollment in and progression through public and private education institutions in developed countries. Some papers consider the effect of price variation on the choice between free public and fee-charging private primary or secondary schools (Dynarski, Gruber, and Li, 2009). More papers consider the effect of price variation on the choice of whether, where, and for how to enroll in either private or public post-secondary education institutions (Dynarski, 2003; Kane, 1994).7 Unlike these studies, I focus on the margin of public primary/secondary school enrollment versus no enrollment. Like some of these studies, I consider the role of price discrimination in driving decisions about where to enroll in school. Fee elimination in South Africa was geographically targeted and I show that this targeting was relatively effective, in the sense that few students transferred from feecharging to fee-eliminating schools. This contrasts with results from the US showing students are sensitive to geographic variation in the price of university enrollment (Knight and Schiff, 2016). The third literature examines the effect of education financing on education outcomes. There is substantial debate about whether the level of education financing matters for education outcomes (Hanushek, 2003; Hoxby, 2001; Hyman, 2016; Jackson, Johnson, and Persico, 2016). I instead study the effect of the source of education financing on education outcomes, as the loss in school fee revenue was meant to be offset by larger transfers from the national government budget. There are few other papers that directly examine the effects of changing the source of education financing, holding the level of education financing constant. I do not observe school-level transfers and so cannot verify whether these offsetting transfers were received. My analysis instead has an intentionto-treat flavor, examining the effect of fee elimination and intended transfer increases on some measures of school resources and learning outcomes. I find no measureable effects on these outcomes, showing that substantial shifts in the source of education spending need change education outcomes. The fourth literature examines link between political economy and education financing. School fees may promote local accountability of schools to students and their families, as they face a direct loss of revenue if students disenroll. Andrabi, Das, and Khwaja (2014) find that private primary schools’ price and quality are responsive to competition induced by improved information about local education markets. I am aware of no comparable results for fee-charging public schools. My results do not directly speak to this mechanism, but the absence of transfers between feecharging and fee-eliminating schools suggests that threats of school exit by students and their families may not be very credible to schools. Harding and Stasavage (2014) argue that school fees are particularly salient and broad-based relative to other sources of education financing, and so are 7
This literature is too large to discuss in detail here. See Hoxby (2004) for a review and discussion of policy implications.
4
particularly politically unpopular. South Africa’s school fee elimination occurred approximately ten years after the country’s first democratic election and may partly reflect political, rather than economic, considerations. This possibility is reinforced by the government’s rejection of a proposed system of need-based tuition fee exemptions, which are arguably less salient, even though they can effectively increase enrollment by low-income students in some settings (Andrews and Stange, 2016). The paper is organized in six substantive sections. I develop a simple conceptual framework in section 2 that motivates the empirical analysis. Section 3 provides an overview of the South African education system and data sources that I use. Section 4 lays out my research designs and how these build on the design of the fee elimination policy. Section 5 reports the treatment effects of school fee elimination on enrollment levels and rates. I also show that these effects are robust to accounting for observed and unobserved differences between fee-eliminating and fee-charging schools. I explore a number of explanations for these price-insensitive demand estimates in section 6. Section 7 reports treatment effects on high school graduation, grade progression, school resources, and the socio-economic profile of enrolled students. Section 8 concludes.
2
Conceptual Framework
Consider a simple reduced form model of the enrollment decision. Assume that each youth i decides whether to enroll in the local school s at time t. Panel A of figure 1 depicts this decision in a simple demand and supply framework. The vertical axis V shows the value and cost of enrolling for an additional year of schooling in money metric terms. The horizontal axis shows the proportion P ∈ [0, 1] of youth in this school’s neighborhood who are enrolled. The downward-sloping demand curve D captures the idea that the value of enrollment is heterogeneous across individuals within a school.8 This heterogeneity may arise from different levels of academic ability, different outside options, and the fact that students have reached different grade levels at time t. Assume for now that the cost of enrollment is identical for all agents, so that the cost curve C is horizontal. This assumption is relaxed below. Equilibrium enrollment occurs at p1 . Note that v1 is the net value of enrollment for the marginal student (gross value less cost), not a market-clearing price. Eliminating school fees shifts the cost curve downward from C1 to C2 . The new equilibrium enrollment rate is p2 ≥ p1 . Students can be divided into three categories with respect to the fee elimination intervention. p1 of the students are inframarginal and enroll whether fees are charged or not; 1 − p2 of the students are inframarginal and do not enroll even if fees are not charged; p2 − p1 of the students are marginal and enroll if and only if fees are eliminated. This yields two general implications of the framework: 8
Throughout this section I assume that the demand curve has no discontinuities in the interior. This follows from the assumption that the density of individual valuations of enrollment within a school is strictly continuous.
5
Figure 1: Conceptual framework showing the effect of fee elimination C1 → C2 on the proportion of students enrolled P Panel A: Basic framework V
v1
C1
v2
C2 D p1
0
p2
1
P
Panel B: Different elasticities of demand V DY
DX C1 C2
p1X
0
p1Y p2X
p2Y
1
P
Panel C: Relating baseline valuation & enrollment changes V
DZ DY DX
C1 C2
0
p1X
p2X
p1Y
p2Y
p1Z p2Z 1
P
Notes: This figure illustrates the conceptual framework used in the paper. The proportion of students enrolled P is increased by eliminating school fees (panel A), the magnitude of the effect is larger when demand is more elastic (panel B), and the magnitude of the effect need not be correlated with mean baseline valuation (panel C).
6
I1 Eliminating school fees increases total enrollment and the enrollment rate. Larger baseline school fees will lead to a larger effect on enrollment. If the baseline enrollment rate is at or near one, eliminating fees will have a smaller effect on enrollment. I2 The change in the enrollment rate equals the proportion of students whose gross value of enrollment is smaller than the cost of enrollment including fees C1 but larger than the cost of enrollment excluding fees C2 . Panel B of figure 1 considers two potential valuation curves: DX is relatively elastic in the neighborhood of the equilibrium and inelastic DY is relatively inelastic. Eliminating fees increases Y X Y the equilibrium enrollment rates from pX 1 and p1 to p2 and p2 . The former change is clearly larger,
generating another implication of the framework: I3 The effect of fee elimination on enrollment is increasing in the local elasticity of the valuation curve D. Panel C of figure 1 considers three potential demand curves: convex DX , linear DY , and concave DZ . The change in enrollment rates induced by eliminating fees is largest for DY and smallest for DZ . The mean valuation amongst the “always-enrollers” is largest for DZ and smallest for DX . This generates another implication of the framework: I4 There is not a monotonic relationship between the mean valuation amongst students enrolled when school fees are charged and the increase in enrollment rate induced by eliminating fees. I4 is particularly empirically important. It implies that school-level treatment effects of school fee elimination will not necessarily be correlated with proxies for the valuation of enrollment prior to elimination. Hence, enrollment may rise by a smaller or larger margin in “good” than “bad” schools. In this depiction, the proportion of always-enrollers and their mean valuation of enrollment are positively correlated. However, this does not hold more generally. The framework can be extended to allow for a number of more realistic elements. In particular: I5 If baseline costs of enrollment are heterogeneous, the cost curve will be upward-sloping and the effect of fee elimination on enrollment will be attenuated relative to the constant-cost case. I6 If schools face binding capacity constraints, the equilibrium enrollment rate will be constrained to be below 1 and the effect of fee elimination on enrollment may be attenuated. I7 If fee elimination reduces the valuation of enrollment, perhaps due to resource constraints or negative peer effects, the demand curve will shift downward and attenuate or potentially reverse the effect of fee elimination on enrollment.
7
This simple framework does not explicitly model the dynamic path of returns to enrollment, or uncertainty about this path. Some useful insights arise from an informal discussion about dynamics and uncertainty. Youth i’s return to enrolling in school s in grade g in year t can be decomposed into the value of enrollment conditional on advancing a grade, the value conditional on not advancing a grade, and the probability of advancing: E[Vi,s,g,t ] = E[Vi,s,g,t |advance = 1] · P r(advance = 1)i,g,s,t + E[Vi,g,s,t |advance = 0] · P r(advance = 0)i,g,s,t . The value conditional on advancing incorporates psychic benefits from advancing, labour market returns to an additional year of completed education, and the option value of enrolling in grade g + 1 in year t + 1. Hence the value of enrollment will be increasing in the probability of grade advancement. Lam, Ardington, and Leibbrandt (2010) show that grade advancement in South Africa is low in some schools and that some schools are systematically worse at translating measures of student “skill” into grade progression. These observations imply that I8 Fee elimination will increase enrollment by more if the probability of grade advancement is higher for marginal students. I9 Fee elimination may decrease enrollment if it lowers the probability of grade advancement for inframarginal students. The probability of grade progression may be systematically different for marginal and inframarginal students within the same school, reflecting differences in student characteristics. This framework generates several guidelines for the empirical analysis. First, eliminating fees should increase the enrollment rate (I1). Second, if the enrollment rate falls, this must be due to negative effects of fee elimination on the perceived valuation of enrollment (I7), potentially by reducing grade progression rates (I9). Third, if the change in the enrollment rate is small, this may reflect near universal baseline enrollment (I1), a low initial value of fees (I1), inelastic demand (I2/3), heterogeneous costs (I5), binding capacity constraints (I6), negative effects on the valuation of enrollment (I7), and/or low probabilities of grade progression for marginal students (I9). Fourth, the magnitude of the treatment effects need not be explained by the mean valuation of enrollment amongst students enrolled at baseline (I4). In particular, the magnitude of the treatment effects need not be explained by the probability of grade progression for students enrolled at baseline. The framework as written assumes that credit constraints never bind on the enrollment decision. The presence of credit constraints will have an ambiguous impact on the relationship between fee elimination and enrollment. In terms of figure 1 panel A, eliminating fees will induce some creditconstrained students with valuations above v1 to enroll, However, credit-constrained students with valuations between v2 and v1 will not be able to enroll. If the former group is larger than the latter, the treatment effect of fee elimination on enrollment will be increased relative to a world with no credit constraints and vice versa.
8
3
The South African Education System
The South African school system consists of twelve grades, divided into nine years of mandatory ‘basic education’ and three years of ‘further education.’ Most schools are either ‘primary schools’ offering grades 1-7 or ‘secondary schools’ offering grades 8-12. Students are expected to enter grade 1 in the calendar year that they turn 7 so a student completing one grade per year would finish secondary school at age 18. Some students attend formal education before grade 1 but data on this attendance is limited (Van der Berg, Girdwood, Sheperd, Van Wyk, Kruger, Viljoen, Ezeobi, and Ntaka, 2013). Students write secondary school graduation examinations at the end of grade 12. These are content-based exams in at least six subjects (at least seven after 2007) that build on the grade 10-12 curriculum and are set, graded, and moderated by the national Department of Education.9 Higher grade attainment and better performance on the graduation exams are associated with large gains in earnings and the probability of employment (Branson, Ardington, Lam, and Leibbrandt, 2013; Branson and Leibbrandt, 2013). During the period I study South African public schools used a partial school choice system: students were allowed to enroll in any school but schools were allowed to prefer students living nearby for admission. The private sector was historically small and expensive but private schools aimed at poor and middle-class households have become more common in recent years. The share of students enrolled in private schools rose from 2.8 to 3.2 percent between 2006 and 2012, but this measure misses some students in unregistered private schools (Department of Basic Education, 2006, 2012). South Africa’s public education system was racially segregated until the early 1990s, and per capita government expenditure on white schools was orders of magnitude larger than on black schools. Curricula at black schools were deliberately focused on non-academic subjects, reflecting the apartheid government’s insistence on preparing black students for manual employment only. Few black students completed secondary schooling, pass rates on secondary school graduation examinations were low, and even fewer students took mathematics or physical science as secondary school subjects (Fedderke, Luiz, and de Kadt, 2000). State expenditure on black education rose substantially in the 1970s, 1980s, and 1990s and this was associated with rapidly rising enrollment rates (Seekings and Nattrass, 2005). However, the quality of education remained low in historically black schools. The education system was officially desegregated in the early 1990s. There is no single data source that provides a comprehensive description of education in South Africa. I therefore combine four difference data sources, summarized in table 1. The national Department of Education’s Education Management Education System (EMIS) contains school-byyear measures of official fee-charging status enrollment, and some school resources. In 2005 to 9
The Department of Education was split in 2009 into the Department of Basic Education, responsible for grades 1-12 and early childhood education, and the Department of Higher Education, responsible for universities, technical training colleges, and adult literacy. I use the terms ‘Department of Education’ and ‘Department of Basic Education’ interchangeably.
9
Table 1: Data Sources Dataset
Education Management Information System 2003-2012 school × grade X X X
General Household Surveys 2003-2012 individual
Income & Expend-iture Surveys 2005/6 & 2010/1 household
National Income Dynamics Study 2008, 2010, 2012 individual X imputed
Years available Unit of observation Panel dataset Treatment status Enrollment Attendance X X Grade attainment X X X Tuition fee spending in bins X X Other education spending X X Total spending in bins X X Total income X X Notes: This table shows coverage of years and variables in four nationally representative datasets. An ideal dataset would contain individual-level measures of attendance/enrollment, treatment status, education expenditure, and total expenditure or income. Fee elimination begins in 2007, before the first wave of the National Income Dynamics Study. I impute individual-level treatment status in the National Income Dynamics Study based on the treatment status of nearby schools using geocodes for households and schools.
2010 I also observe school-level results on graduation examinations and school-by-grade enrollment breakdowns. I use these data to conduct the primary analyses, relating fee-charging status to enrollment, graduation, and resources. The annual General Household Survey (GHS) provides nationally representative data on school attendance and some measures of household socio-economic status. But there are no geographic identifiers below the province level, so I cannot even approximately infer whether households live near fee-eliminating schools. I use these data to describe attendance rates both before and after fee elimination and as a plausibility check on the administrative enrollment data. The GHS measures current attendance, whereas the EMIS measures enrollment. These are not identical measures of education participation and could in principle show different patterns of education participation through time. In practice, the two measures from the two data sources yield very consistent information. The Income and Expenditure Surveys (IES) in 2005/6 and 2010/1 provide nationally representative data on total expenditure and detailed expenditure on fees and other education costs. But there are no measures of school attendance and no geographic identifiers below the province level. I use these data to describe the cost of fees relative to total expenditure and other education expenditure and as a plausibility check on the administrative fee payment data. The National Income Dynamics Study contains nationally representative data on attendance, fee payment, household socio-economic status, and geographic identifiers that reveal the treatment status of nearby schools. But this study starts in 2008, soon after the fee elimination policy. Future versions of the paper will use these data to examine the effects of the second phase of tuition fee elimination in 2010. Contemporary South African education is characterized by high enrollment, low grade attain-
10
Figure 2: Attendance and Attainment before Fee Elimination
60 40 0
20
Percent
80
100
Panel A: Attendance and Attainment by Age
6
9
12
15
18
21
24
Age Attending
Graduated
Non−participating
60 40 0
20
Percent
80
100
Panel B: Attendance by Grade-Eligibility
1
2
3
4
5
6
7
8
9
10
11
12
Grade Attending conditional on qualifying for grade
Notes: The first panel shows the shares of people who are currently attending primary/secondary education (red dashed line) and have completed secondary education (blue solid line). The remaining people (black dashed line) are still eligible to attend primary/secondary school but are not doing so. This 18% of the population provides an upper bound for the effect of fee elimination on school attendance and completion. The second panel shows the share of people eligible to attend each grade who are attending that grade. Conditional on eligibility, attendance in primary school (grades 1-7) is 93 percent and attendance in secondary school (grades 8-12) is 69 percent. 95% confidence intervals are based on heteroskedasticity-robust standard errors. All calculations use the General Household Surveys for 2003-2006, the four years preceding the fee elimination policy. Individuals are included in the sample if they are aged 5-24 and live in households with monthly total expenditure of less than USD520, approximately the 80th percentile of the expenditure distribution. All calculations use individual-level post-stratification weights constructed by Statistics South Africa, normalized to sum to one in each survey year.
11
ment, and poor learning outcomes. Figure 2A shows that in 2003-2006, 90 percent of people aged 5-18 attended primary or secondary school.10 However, figure 2A also shows only 30 percent of people aged 18-24 in 2003-2006 had completed secondary school. I define the 18 percent of people aged 5-24 who are not attending school and have not graduated from secondary school as nonparticipants. Non-participation is high at young ages due to late primary school starts, is below 3 percent for ages 8-14, and then rises rapidly to 22 percent at age 18 and 43 percent at age 21. Non-participation at older ages is largely explained by drop-out during secondary school. Figure 2B shows that attendance in grade G conditional on completing grade G − 1 is higher than 80 percent in primary school grades 1-7 but drops rapidly through secondary school to 60 percent in grade 12. These low graduation rates are consistent with very poor performance by South African students on international literacy and numeracy assessments, even relative to poorer countries (Reddy, Visser, Winnar, Arends, Juan, Prinsloo, and Isdale, 2016; Van der Berg and Louw, 2007). The population averages shown in figure 2 conceal important heterogeneity.11 Attendance was high across most population groups but black and poor people had higher grade repetition, higher drop-out and hence lower attainment than white and less poor people. The share of 18-24 years olds who have graduated from secondary school is 45 percentage points lower for black than white youths and 23 percentage points lower for people living in households below the median socioeconomic status (SES).12 Attendance varied less by race and SES: school attendance by 5-18 year olds was 1.7 percentage points lower for black than white youths and 2.5 percentage points lower for those living in households with below-median SES. Figure 3A shows how attendance, graduation and hence non-participation vary by household SES. Non-participation by people aged 5-24 falls from 23 percent in the lowest SES bin to 3 percent in the highest SES bin. Non-participation is also 10 percent higher for black than white people in the same age range. The strong negative relationship between SES and participation suggests that tuition fees may have posed a barrier to secondary school attendance and graduation. (Primary school attendance was already nearly universal.) All public schools in South Africa charged tuition fees to fund discretionary spending until 2007 (Pampallis, 2008). Provincial education departments paid for teacher salaries, infrastructure costs, and a per-student subsidy for discretionary spending that was larger for schools in lower-SES neighborhoods. Figure 3B shows that in 2005/6 fees and other education costs accounted for respectively 3 and 5.1 percent of total expenditure in households with 10
All statistics are calculated by pooling the 2003 to 2006 General Household Surveys. Calculations use individuallevel post-stratification weights provided by Statistics South Africa. All differences discussed in the text are statistically significant at the 1 percent level, using heteroskedasticity-robust standard errors. 11 There was limited gender heterogeneity in primary and secondary education outcomes: attendance and secondary school graduation were respectively 0.8 percentage points lower and 1.4 percentage points higher for women than men. 12 I measure household socio-economic status based on reported total monthly household expenditure from the General Household Survey, in bins. This measure should be interpreted as a rough proxy: the General Household Survey does not measure household income, measures total expenditure in bins whose boundaries are not adjusted annually for inflation, and the binned values cannot be easily adjusted for differences in household size.
12
Figure 3: Attendance, Attainment, and Expenditure by Household Expenditure before Fee Elimination
60 50 40 0
10
20
30
Percent
70
80
90
100
Panel A: Attendance and Attainment by Total Household Expenditure
0 − 120
120 − 230
230 − 350
350 − 520
520 − 720
720 − 1450
1450 − 2890
> 2890
Monthly HH expenditure in USD Attending
Graduated
Population share in expenditure bin
0
2
4
Percent
6
8
10
Panel B: Share of Total Household Expenditure Spent on Education
5
6
7
8
9
Log total monthly household expenditure in USD Fees
Other costs
Population share
Notes: The first panel shows the share of people who are currently attending primary/secondary education (red dashed line) and have completed secondary education (blue solid line) by bins of total monthly household expenditure. All calculations in the first panel use the General Household Surveys for 2003-2006, the four years preceding the fee elimination policy. All calculations use individual-level post-stratification weights constructed by Statistics South Africa, normalized to sum to one in each survey year. The second panel shows the share of total annual household expenditure spent on fees (red dashed line) and other education costs (blue solid line) in the 2005/6 Income and Expenditure Survey. All calculations exclude households with no members aged 5-24 who are eligible to enroll in primary or secondary school, exclude households with zero spending on any education category, and use poststratification weights supplied by Statistics South Africa multiplied by the number of household members aged 5-24 who have not completed secondary school. These restrictions and weights are designed to approximate the schoolattending population but will incorrectly include age-eligible non-attenders in households where another member is attending school. Expenditure converted from South 13 African Rands to 2006 USD values using 2006 purchasing power parity-adjusted exchange rates from the World Development Indicators. 95% confidence intervals are based on heteroskedasticity-robust standard errors.
members enrolled in school.13 Even in households in the poorest quintile, fees and other education costs account for respectively 2.6 and 6.2 percent of total expenditure. This is large relative to education spending patterns of poor households reviewed by Banerjee and Duflo (2007). Households directly report that fees are a deterrent to attendance: in 2003-2006, 37 percent of non-participating 5-24 year olds listed tuition fees as a barrier to attendance. Children in households that receive positive, anticipated income shocks from age-linked social welfare grants have significantly higher school attendance rates, consistent with the existence of financial barriers (Eyal and Woolard, 2013). In rural areas specifically, school attendance rises when a household member becomes age-eligible for a state pension (Edmonds, 2006). This indicates that households are unable or unwilling to fund education by borrowing against income they will receive with high probability in a few years. Funding education by borrowing against long-term future earnings is thus unlikely. These factors all suggest that financial barriers can reduce attendance and, by limiting home education inputs and study time, grade progression and secondary school graduation. However, the negative relationship between SES and participation is also consistent with other explanations. In particular, economic returns to participation may be positively correlated with SES because learning is low in schools in low-income neighborhoods. Under this model, eliminating tuition fees will have a limited positive effect on attendance and graduation. If financial barriers are negligible and eliminating tuition fees reduces school revenue and hence learning inputs, eliminating fees can even reduce attendance or graduation.
4
Fee Elimination Policy and Research Design
Tuition fee elimination was announced in 2006 and implemented in 2007. Schools treated by this intervention were required to eliminate all tuition and enrollment fees. These ‘no fee’ schools were selected by a three-stage interaction between provincial and national governments, laid out in guidelines published by the national Department of Education. The selection process was designed to eliminate fees first in schools serving low-SES communities. The Department of Education committed to provide fee-eliminating schools with larger per-student grants to offset the loss of fee revenue. No systematic data appear to exist on the relationship between pre-elimination fees and subsequent transfers. In the first stage, provincial governments assigned each school in their province a “poverty score” based on characteristics of the electoral ward in which it was located.14 These scores ranked all 13
I calculate these shares using the nationally representative Income Expenditure survey from 2005/6. This survey measures detailed expenditure at the household level, including expenditure on tuition fees, other fees paid to the school (which are negligible), transport to school, school uniforms, and books. The survey does not measure attendance or enrollment. I approximate the population of enrolled youths by (1) restricting the sample to households that report positive expenditure on at least one education category and have at least one member aged 5-24 who has not graduate from secondary school and (2) multiplying the household-level post-stratification weights by the number of household members aged 5-24 who have not graduated from secondary school. 14 The electoral ward is not an administrative unit in South Africa. Assignment took place at this level because it
14
schools within the province from least to most poor, with ties permitted. The national Department of Education provided each province with ward-level data on income, employment, education, health, and amenities from the 2001 census as a starting point for the assignment of poverty scores. Provinces were permitted choose their own weighting of these five data series and to make ad hoc adjustments to the resultant score to reflect within-ward heterogeneity. They were not permitted to use any data collected directly from schools, such as administrative data on school’s physical facilities or student-teacher ratios. Wildeman (2008) conducted anonymous interviews with provincial officials responsible for creating the poverty scores and reported that most of the ad hoc adjustments were made for schools near the boundaries of electoral wards, as the socio-economic characteristics of their students may have differed from those of the electoral ward. Wildeman’s interviewees reported no incidents of schools lobbying provincial officials to change their scores, although lobbying may have occurred in the third stage described below. The formulae used to determine the poverty scores were left to the discretion of the provinces and no province has made its formula publicly available.15 In the second stage, the national government divided all schools in the country into five quintiles based on these poverty scores. The number of schools in each quintile was chosen after the poverty scores had already been assigned, so it was not possible for poverty scores to be precisely manipulated in the neighborhood of the cutoffs between quintiles. Each quintile was intended to contain approximately 20% of the students in the country (based on 2006 enrollment data) and each quintile would contain similarly poor schools in each province. In the relatively poor Eastern Cape province, 35% and 6% of all schools were assigned to the first and fifth quintiles respectively; in the relatively low poverty Western Cape province, 7% and 23% of all schools were assigned to the first and fifth quintiles respectively. The choice of how many schools were to be treated in each province was based on province-level data from the 2001 census but the exact algorithm used for this decision is unclear. The original implementation schedule is shown in figure 4A: quintile one and two schools were meant to eliminate fees in 2007, quintile three schools in the two least poor provinces were meant to eliminate fees in 2008, and quintile three schools in the remaining seven provinces were meant to eliminate fees in 2010. In the third stage, provincial governments instructed specific schools to eliminate fees by sending letters to principals. Figure 4B shows the share of schools instructed to eliminate fees by year and quintile. These fee elimination instructions followed the original implementation schedule quite closely, but provincial governments clearly extended the policy to additional schools. I use data is the smallest geographic unit at which census data was available. 15 Each province used a different scale for the poverty scores. I standardize these by recentering them at the cutoff between quintiles 2 and 3 and rescaling them to have standard deviation one within each province. The results are reasonably robust to alternative standardizations: rescaling the range to one within each province or rescaling the variance to minimize the sum of the differences between the quintile 1/2 cutoff and the quintile 3/4 cutoff.
15
from the 2003-2012 General Household Surveys to check school compliance with fee elimination. Figure 4C shows that fee payment was nearly universal up to 2006.16 But by 2011, approximately 60 percent of students attending both primary and secondary schools reported paying zero fees. These data confirm that fees were widely eliminated but that actual elimination was not as widespread as instructed, consistent with case study evidence that some designated no-fee schools continued to demand fees (Thwala, 2010). Given this elimination schedule, I use a research design that compares outcome trends across fee-charging and no-fee schools. I begin with a difference-in-differences model: Yst = αs + βt + δ DD × 1 {Fees = 0}st + s t
(1)
where Yst denotes the outcome for school s in year t, αs and βt are respectively school and time fixed effects, and 1 {Fees = 0}it is an indicator equal to one if and only if school s does not charge fees in year t. Assuming fee-charging and fee-eliminating schools would have experienced the parallel time trends in Y in the absence of fee elimination, δ is the average treatment effect of fee elimination on outcome Y for schools that eliminate fees. There are several reasons to question the parallel trends so I present a variety of robustness checks. First, I instrument fee-charging status assigned by the provincial government with feecharging status from the original implementation schedule in figure 4. This accounts for deviations from the original implementation schedule that are correlated with outcome trends. For example, schools might lobby to eliminate fees because they expect falling enrollment. This would violate the parallel trends assumption by systematically assigning schools with lower enrollment trends to eliminate fees. Using the original implementation schedule to instrument for fee elimination avoids this problem, but does not account for differential trends that are correlated with the original schedule. Second, fee elimination may change the population composition if fee elimination causes schools or their neighbors to close. I therefore restrict the sample to the balanced panel of schools that are open in all years from 2003 to 2012. I also verify that there is no difference in the probability of closing in years t + 1 or t + 2 between schools that do and do not eliminate fees in year t. Third, I replace βt with a vector of province-year and then district-year fixed effects. This accounts for geography-specific time-varying heterogeneity. Education districts in South Africa are administrative units designed to include both high- and low-SES communities, so 80% of schools are in a district that contains both a fee-charging and a no-fee school in at least one year. National, provincial and district education departments are responsible for respectively policy development, funding, and policy implementation. So district- and province-specific time-varying heterogeneity 16
A previous policy introduced by the Department of Education required schools to waive fees for children from households whose total income was less than ten times the value of the fees. But 97 percent of students reported paying fees in 2007, showing that the waiver policy was not widely used.
16
Figure 4: Planned and Reported Fee Elimination Schedule Panel A: Planned Fee Elimination Schedule
Panel B: Actual Fee Elimination Schedule from Provincial Government Records
60 40 0
20
Percent
80
100
Panel C: Actual Fee Elimination Schedule from Household Survey Data
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
Year Primary school
Secondary school
Notes: The first panel shows the share of public schools that were scheduled to eliminate fees according to the plan announced in 2006. The second panel shows the share of public schools that were instructed by provincial governments to eliminate fees in each year. The first two panels use school-level administrative data weighted by baseline enrollment. The third panel shows the share of people aged 5-24 enrolled in primary or school who report paying zero fees. The third panel uses student-level data from the General Household Survey with post-stratification weights constructed by Statistics South Africa, normalized to sum to one in each survey year.
17
is a relevant concern. Fourth, I reweight fee-charging schools to have the same distribution of observed pre-elimination characteristics as no-fee schools following Abadie (2005) and DiNardo, Fortin, and Lemiuex (1996). This accounts for heterogeneity across time-invariant observed characteristics in time trends but does not account for heterogeneity in time trends across time-invariant unobserved characteristics. Fifth, I compare schools with poverty scores near the cutoff between quintiles two and three. Schools with poverty scores just above the cutoff were assigned to eliminate fees in 2007 under the original schedule while schools with poverty scores just below the cutoff were assigned to maintain fees until 2008 or 2010, depending on their province. I use this feature of the policy design to implement a regression discontinuity design. Specifically, I estimate 1 {Fees = 0}s =1{Poverty score ≥ 0}s × f + (Poverty scores ) − 1{Poverty score < 0}s × f − (Poverty scores ) + i
(2)
and ∆Enrollments =1{Poverty score ≥ 0}s × f + (Poverty scores ) − 1{Poverty score < 0}s × f − (Poverty scores ) + i
(3)
where f + , f − , g + , and g − are local linear functions of poverty scores. The idea behind these models is to specify flexibly the relationship between the outcome of interest (respectively, no fee status and the change in enrollment from 2005/6 to 2007/8). I then evaluate δ RD =
lim
Poverty score↓0
fˆ+ (Poverty scores ) −
lim
Poverty score↑0
fˆ− (Poverty scores )
to estimate the magnitude of the change in the outcome of interest that occurs in the neighborhood of the cutoff poverty score that separates intention to treat and control schools. Provided schools are not able to manipulate precisely the poverty score they are assigned by province, untreated schools on one side of the cutoff should be a valid counterfactual for treated schools on the other side of the cutoff. Note that I use a time-differenced outcome variable in equation (3) so the specification already removes any time-invariant observed or unobserved characteristics. This identification strategy is more robust relative than standard regression discontinuity designs that use only crosssection data. This identification strategy generates two testable predictions. First, I verify that none of the observed characteristics listed above have statistically significant or economically meaningful jumps at the cutoff. Second, I verify that there is no evidence that the density of the poverty score variable jumps at the cutoff following McCrary (2008). Note that δˆDD and δˆRD recover different treatment effects, if both their identifying assumptions
18
are satisfied. The former estimator recovers the average treatment effect on the treated schools, i.e. the average effect of eliminating fees for the schools that eliminated them. This is more useful for evaluating the policy as implemented and for studying the economic relationship between education outcomes and fee elimination. The latter estimator recovers the average treatment effect on schools near the poverty score cutoff between quintiles two and three. This is more useful for predicting the effects of marginally lowering the poverty score threshold to eliminate fees in more schools. I focus on δˆDD throughout the paper because I argue this recovers a more economically meaningful parameter.
5
Treatment Effects of Fee Elimination
Table 2 reports estimates from equation 1 for the sample of all public schools in quintiles one to four. I omit schools in quintile 5 from the analysis, which are located in historically white neighborhoods, charged high fees throughout this period, and function as a largely separate education system. Enrollment increases by 3.3 students per school. This treatment effect ranges from 2.6 to 3.3 students when I account for province-by-year and district-by-year fixed effects and omit schools that are open for only part of the period (columns 3-5). The effect is slightly smaller when I reweight fee-charging schools to have the same distribution of baseline observed characteristics as fee-eliminating schools (column 2). The original fee elimination schedule is a strong instrument for actual fee elimination and the instrumented treatment effect is 1.7 (column 2). Taken together, these results point to a robust, positive and statistically significant effect of fee elimination on enrollment. But the effects are also robustly very small. This is an increase of 0.2 to 0.3 percent relatively to mean baseline enrollment of 625 students per school. The attendance rate in the General Household Survey for 2003 to 2006 in households in the bottom four quartiles of the SES distribution was 80.4 percent, so these estimates imply enrollment increased by 0.16 to 0.48 percentage points.17 The high baseline attendance rates for primary school-eligible people shown in figure 2 suggests that fee elimination should only increase enrollment in secondary schools. Table 3 therefore reports estimates of equation (1) for secondary schools only. The effects are substantially larger: fee elimination increases enrollment by 19.6 to 22.9 students per school. This is an increase of 2.3 to 2.6 percent on the mean baseline enrollment rate of 851 students per secondary school. The 2003-2006 attendance rate by secondary school-eligible youths in households in the bottom four quartiles of the SES distribution was 65.6 percent, so these estimates imply enrollment increased by 17 These figures are derived from Effect on Enrollment Count×Baseline Enrollment Rate/Baseline Enrollment Count. All quantities are estimated and the combination of the estimators is consistent for the combination of the estimands but not unbiased. The biased can be calculated using a Taylor series approximation but removing the leading term in the bias requires an estimate of the covariance matrix between the three estimators. This quantity is not identified with the available data but the magnitude of the leading term in the bias is small under a range of assumptions about the magnitude of the covariances.
19
Table 2: Treatment Effects of Fee Elimination on Enrollment Effect on # enrolled students Effect on enrollment rate School fixed effects Year fixed effects Province-year fixed effects District-year fixed effects Inverse probability weights Balanced panel First stage
3.28 (1.12) 0.0031 (0.0022) × ×
1.72 (0.90) 0.0016 (0.0018) × ×
2.63 (1.20) 0.0025 (0.0024) × ×
2.94 (1.24) 0.0028 (0.0025) ×
3.12 (1.80) 0.0030 (0.0036) ×
5.03 (2.32) 0.0048 (0.0046) × ×
× × × ×
0.912 (0.003) # observations 246951 246951 213213 246951 238578 246951 # schools 24652 24652 19383 24652 23248 24652 Baseline mean enrollment 625 625 625 625 625 625 Notes: Sample is all public schools in quintiles 1-4 open in 2003-2012 with non-missing enrollment data. Heteroskedasticity-robust standard errors clustered at the school level are shown in parnetheses. Instrument in column 2 is derived from planned fee elimination schedule shown in figure 4A. Results in column 3 omit the 21 percent of schools that not open in all years 2003-2012. Inverse probability weights in column 6 adjust for differences in baseline enrollment, historical racial classification, location, governance structure.
1.51 to 1.77 percentage points. Figure 5 shows grade-specific treatment effects and verifies that the effects are nearly zero throughout primary school and large only in the first two grades of secondary school. The effects are negative (though imprecisely estimated) in grades 11 and 12, a pattern I discuss further in section 7. These effects can also be scaled by the actual change in education costs to calculate a price elasticity of demand for education enrollment. This exercise faces serious data limitations so the precise numbers should be interpreted with caution. Households in quintiles one to four of the SES distribution in the 2005/6 Income and Expenditure Survey allocated 2.4 and 5.7 percent of their total expenditure to respectively tuition fees and other education costs. Under the assumptions that the ratio of fees to other education costs is constant across the population and that fees are entirely eliminated, these data imply a 29 percent reduction in total education costs. The timeseries of fee payments in figure 4C shows that fees were not fully eliminated, but the share of students from households in quintiles one to four of the SES distribution paying zero fees rose from 3 to 72 percent. Combining this with the ratio of fees to other education costs implies that total education costs fell by 20 percent. The arc price elasticity of demand for enrollment is therefore 0.591 0.202 %∆Enrollment ∈ − ,− = [−0.029, −0.010] = %∆Education spending 20.44 20.44 for all students. Calculating the elasticity for secondary school enrollment is more difficult because the Income and Expenditure Survey does not allow me to calculate the ratio of fees to other education costs separately for primary and secondary school students. If I assume this ratio is the
20
Table 3: Treatment Effects of Fee Elimination on Secondary School Enrollment Effect on # enrolled students
20.61 (2.86) 0.0159 (0.0070) × ×
Effect on enrollment rate School fixed effects Year fixed effects Province-year fixed effects District-year fixed effects Inverse probability weights Balanced panel First stage
22.92 (2.35) 0.0217 (0.0058) × ×
20.38 (3.05) 0.0193 (0.0075) × ×
19.55 (3.17) 0.0185 (0.0078) ×
21.35 (5.24) 0.0202 (0.0128) ×
22.51 (4.29) 0.0213 (0.0105) × ×
× × × ×
0.923 (0.006) # observations 54930 54930 47960 54930 52742 54930 # schools 5389 5389 4360 5389 5013 5389 Baseline mean enrollment 851 851 851 851 851 851 Notes: Sample is all public secondary schools in quintiles 1-4 open in 2003-2012 with non-missing enrollment data. Heteroskedasticity-robust standard errors clustered at the school level are shown in parnetheses. Instrument in column 2 is derived from planned fee elimination schedule shown in figure 4A. Results in column 3 omit the 21 percent of schools that not open in all years 2003-2012. Inverse probability weights in column 6 adjust for differences in baseline enrollment, historical racial classification, location, governance structure.
Figure 5: Treatment effects of fee elimination on enrollment by grade 20
Change in enrollment
15
10
5
0 0
1
2
3
4
5
6
7
8
9
10
11
12
-5
-10
Grade Notes: This figure shows grade-specific treatment effects of fee elimination on enrollment. The dashed lines show 95 percent confidence intervals based on from heteroskedasticity-robust standard errors clustered at the school level.
21
same for both groups of students, then secondary ∈ [−0.132, −0.109] . These estimates of the effect of fee elimination are valid only if the enrollment trends in feecharging and fee-eliminating schools would have been identical in the absence of the elimination policy. The results in tables 2 and 3 provide some reassurance that the results are not driven by time-varying shocks that differentially affect the two groups of schools or time trends that vary across observed pre-elimination characteristics. Figure 6 shows that I obtain a similar result using a regression discontinuity design that compares enrollment changes in schools just above and just below the poverty score cutoff between quintiles two and three. The estimates are calculated from local linear regressions with bandwidth choices following Imbens and Kalyanaraman (2012). Panel A shows that the probabilty of fee elimination rises from approximately 8% for schools just less poor than the cutoff to approximately 86% for schools just more poor than the cutoff. The difference of 76 percentage points is smaller than the difference between all intention to treat and control schools but is still substantial. The second figure shows that enrollment increases from 2005/6 to 2007/8 by 2.8 students more in intention to treat than control schools (confidence interval -0.7 to 6.3). This is slightly smaller than the difference for all schools, though the estimates are not statistically significantly different. Estimates from global linear, quadratic and cubic models and from local linear models with halved and doubled bandwidths yield estimates between 0.8 and 4 students per school. This reinforces the earlier result that fee elimination had a relatively small effect on student enrollment.
6
Explaining Price-Insensitive Participation
This section builds off the conceptual framework developed in section 2 to explore possible reasons for the small enrollment effects that I find. The framework suggests that this may be due to (1) high baseline enrollment or ceiling effects, (2) low baseline fees, (3) inelastic demand, (4) binding capacity constraints, (5) credit constraints, or (6) negative effects on the valuation of enrollment. Small treatment effects are also possible if (7) students transfer from fee-eliminating to fee-charging schools because they fear a fall in school resources or negative peer effects. Figure 3 shows that households allocated an average of 2.4 percent of their total expenditure to fees at baseline, casting doubt on the second explanation. I explore the first, fourth, fifth, and seventh explanations in this section and find that they cannot fully account for the small effects. The sixth explanation is left to the next section. The analysis as a whole suggests an important role for the residual explanation, inelastic demand. Baseline enrollment rates are high enough for primary school-eligible youths that ceiling effects
22
Figure 6: Treatment Effects of Fee Elimination on Enrollment in Schools near The Poverty Score Cutoff 1
Intention to treat schools
Probability of treatment
0.8
0.6
0.4
0.2
Intention to control schools
0 -2 -3
0 -1
2 1
Poverty score 30 25
Change in enrollment
20 15 10
Intention to control schools
5
Intention to treat schools
0 -2 -5
-3
0 -1
2 1
-10
Poverty score Notes: Panel A shows the fitted probability that schools eliminate fees by their assigned poverty score. The estimated difference between the curves at the threshold value is 76 percentage points with a standard error of 1 percentage point. Panel B shows the fitted change in school enrollment from 2005/6 to 2007/8 by schools’ assigned poverty scores. The estimated difference between the curves at the threshold value is 2.8 students with a standard error of 1.8 students. The fitted curves and 95% confidence intervals in both figures are from local linear regressions estimated separately on either side of the cutoff, with bandwidth choices following Imbens and Kalyanaraman (2012).
23
are plausible for this population. Figure 5 confirms that enrollment effects in this population are close to zero. However, enrollment effects on grades 10 to 12 are also small and not robustly positive. This pattern is not consistent with ceiling effects on enrollment. Some schools may have a binding upper limit on the number of students they can accommodate, due to limited personnel, classroom space or other physical facilities. This would lead them to deny enrollment to students who would otherwise be induced to enroll by the policy change. The legal status of such denials is subject to an ongoing court challenge but anecdotal reports suggest that it is uncommon. I formally test for the existence of capacity constraints in two steps. I first calculate the maximum enrollment in each school between 2003 and 2006. This is one measure of each school’s maximum capacity. I then compare the frequency with which 2007 or 2008 enrollment exceeds the previous maximum by quintile. This occurs in 26 percent of fee-charging shools and 29 percent of fee-eliminating schools. A substantial fraction of schools are thus able to accommodate additional students and this proportion is higher in fee-eliminating than fee-charging schools. Furthermore, the change in enrollment from 2006 to 2007 is smaller than at least one previous annual change in enrollment (2003 to 2004, 2004 to 2005 or 2005 to 2006) in over half of the sample. Hence, many schools are able to accommodate larger increases in enrollment than they experience when fees are eliminated. The results do not conclusively rule out capacity constraints but they do not appear to be of central importance. Credit constraints can, as I discuss in section 2, either attenuate or increase the treatment effects of fee elimination. Edmonds (2006) reports some evidence of credit constraints to school enrollment in rural but not urban areas. Lam, Ardington, Branson, Goostrey, and Leibbrandt (2010) study a largely urban population and find little evidence of credit constraints to enrollment in tertiary education, which is typically much more expensive than primary or secondary enrollment. These results motivate estimating treatment effects separately for urban and rural schools.18 The enrollment effects are approximately twice as large in urban areas as in rural areas, though the difference is not statistically significant. If credit constraints are indeed present in rural but not urban areas, these results suggest that there are a substantial number of rural students who would be induced to enroll by fee elimination if not for credit constraints. This is consistent with Edmonds’ results, as the income shock he uses to test for credit constraints is substantially larger than the value of school fees. However, the results could also be due to larger baseline fees in urban areas or more elastic demand in urban areas. Even the larger effects in urban areas imply increases of less than 2 percent of baseline enrollment. Eliminating fees therefore leaves a large fraction of the population unenrolled in both rural and urban areas. My estimation strategy assumes that the school fee elimination policy has no effect on enrollment levels at the control schools that continue to charge fees. This assumption may be violated if 18
This includes all non-rural schools, urban and suburban.
24
−3
Change in enrollment −2 −1 0
1
Figure 7: Changes in enrollment from 2005 to 2006 and 2006 to 2007 for fee-charging schools located at different distances from fee-eliminating schools
0
.1
.2
.3
Distance (degrees) 2007 change 2006 change
95% CI
Notes: This figure reports local linear regression estimates of the relationship between the year-on-year change in enrollment and proximity to a fee-eliminating school. The solid line shows the relationship the enrollment change from for 2006 to 2007, when fees were eliminated, and the dashed line shows the relationship for the enrollment change from 2005 to 2006 as a placebo test. The two curves are neither substantively nor significantly different to each other. This provides evidence against substantial transfers between fee-charging and fee-eliminating schools.
students who attended control schools before the policy change transfered to treatment schools after the policy change to take advantage of their lower cost. Such behavior would result in an upward bias in the estimated treatment effect of the fee elimination intervention. I do not observe student-level data on transfers that would permit a direct test of this hypothesis. I therefore implement an indirect test that examines whether control schools that are geographically closer to treatment schools experience falls in enrollment from 2006 to 2007 relative to farther away control schools and relative to their own enrollment change in previous years. Figure 7 shows a local linear regression of the change in grade-level enrollment from 2006 to 2007 at control schools against the distance from the nearest treatment school offering the same grade.19 Control schools nearer to treatment schools actually experience small gains in enrollment relative to control schools farther away. I cannot reject that this pattern of changes is identical to that observed between 2005 and 2006, before the fee elimination policy was implemented. The result is robust to restricting the sample to control schools within one half standard deviation of the cutoff. I interpret this as strong evidence against the spillover hypothesis. I also estimate a linear regression of change in enrollment by grade at fee-charging schools from 19
I construct this distance measure using GIS codes for every school in the sample.
25
2006 to 2007 on the same measure at the nearest fee-eliminating school. If the treatment effect is driven entirely by transfers, the slope coefficient should be approximately equal to -1. Instead, it equals 0.045 (standard error 0.016). This is not significantly different to the coefficient in the equivalent regression using changes from 2005 to 2006 (0.037, with standard error 0.08). This result is robust to weighting the regression by the inverse distance between treatment and control schools, to restricting the sample to control schools within one half standard deviation of the cutoff, and to excluding control schools that are more than 10 miles from the nearest treatment school. These results strongly suggest that the treatment effects are not driven by transfers between schools. However, I observe only net transfers and not gross inflow and outflow of students into each school. I cannot rule out the possibility that approximately equal numbers of students transfer in each direction between fee-charging and fee-eliminating schools. One interpretation of these results is that geographically targeted variation in the prices of public services may be an effective alternative to individual-level means testing. The absence of spillovers in this context suggests that setting lower prices for public services in poor neighborhoods does not induce people from wealthier neighborhoods to adjust their behavior to take advantage of the lower cost services. Geographic targeting may be a desirable alternative to individual meanstesting when the latter is expensive. However, caution should be exercised in generalizing this result to settings where use of public services is more price-sensitive or transport costs are lower. Taken together, these results demonstrate that the small enrollment effects are not explained by ceiling effects, capacity constraints, credit constraints, or spillovers. The next section explores whether they may be explained by a declining valuation of enrollment.
7
Treatment Effects of Fee Elimination on Equilibrium Education Outcomes
Fee elimination may change the resources available to schools and the composition of their student body. Compensating government transfers may be larger or smaller than the foregone fee revenue. Even if per-student revenues are unaffected, education inputs such as classrooms and teachers may adjust to changes in student numbers with a lag. Student expectations about changes in education resources may in turn affect their enrollment decisions and so attenuate or augment the enrollment response to fee elimination. This section explores the equilibrium effects of fee elimination on graduation, grade repetition, drop-out, school resources (proxied by class size), and the socio-economic profile of the enrolled students (proxied by the proportion of students eligible for means-tested government social grants and the proportion who have had at least one parent die). These effects are equilibrium in that they are conditional on student enrollment decisions. They provide no direct information about the function mapping enrollment to education outcomes or vice versa. The section also explores 26
Table 4: Treatment Effects of Fee Elimination on School Resources, Grade Progression, and Composition Outcome Effect of fee elimination Pre-treatment mean Year fixed effects School fixed effects # clusters # observations
(1) Dropout rate -0.72 (0.09) 2.78 × × 9857 33399
(2) Grade repetition -0.26 (0.19) 9.63 × × 9856 33396
(3) Class size 0.50 (0.23) 39.31 × × 10147 36986
(4) Share receiving social grants -0.03 (0.65) 37.31 × × 10235 37123
(5) Share of student orphans -1.13 (0.28) 20.66 × × 10131 37073
Graduation exams # taking # passing -2.4 -2.3 (0.89) (0.71) 80.4 45.1 × × × × 5371 5371 21309 21309
Notes: Sample is all public schools in quintiles 1-4 open in 2005-2010 with non-missing data on enrollment and the relevant outcome. Heteroskedasticity-robust standard errors clustered at the school level are shown in parnetheses.
the enrollment trends in control schools close to and far from treatment schools in order to address the possibility of spillovers. I estimate treatment effects on the number of students enrolled in grade 12, number of students writing the secondary school graduation exam, and number of students passing the exam and report the results in figure 5 and table 4 columns 6 and 7.20 Fee elimination reduces the number of students enrolled in grade 12, writing the secondary school graduation exam, and passing the exam by respectively 3.2, 2.4, and 2.3 per school. These estimates imply a roughly 3% fall in the number of students enrolled in grade 12 and writing the exam, and a larger fall of roughly 4.5% in the number of candidates passing the exam. These numbers are all significantly different to zero, but not to each other. These results demonstrate that fee elimination is ineffective at increasing secondary school graduation rates, at least over a four year time horizon. There are multiple possible explanations for the negative effect on graduation. Perhaps most obviously, secondary schools may reallocate resources away from grade 11 and 12 classes toward the expanding grade 8 and 9 classes (see figure 5). I cannot reliably test this explanation without observing grade-specific measures of resource allocations such as class sizes and individual teacher assignments. The falls in the number of students attempting and passing the exam are almost identical and the pass rate is not significantly affected by treatment. This shows that the marginal students induced not to enroll in grade 12 by the fee elimination policy were not negatively selected on their probability of passing the exam. The welfare implications of this result are arguably worse than if the treatment reduced enrollment by academically marginal students, whose probability of graduating secondary school was already low. In particular, this implies that school fee elimination does not improve the extent to which 20 Students can receive different types of passes: some simply denote secondary school graduation, some allow entry to post-secondary vocational training programs, and some allow entry to university. I also estimate treatment effects on the number of students who qualify for these “higher” passes. The results are not substantially different for all passes and for different pass types.
27
students with better academic prospects self-select into enrollment. The low graduation rates in treated secondary schools (61 percent at baseline) and negative effects of fee elimination on graduation imply that students may face binding academic constraints that explain their price-insensitive enrollment behaviour. I test this hypothesis by augmenting the difference-in-differences models in section 4 to allow an interaction between treatment and an indicator for schools with higher-than-median graduation rates prior to fee elimination. I find that the effects on total enrollment, grade 12 enrollment, and number of students attempting secondary school graduation exams are all positive in schools with higher baseline pass rates. Effects in schools with lower baseline pass rates are significantly smaller and are negative in some model specifications. However, the effect on the number of students graduating from secondary school is negative, small, and almost identical in both groups of schools. This pattern of results suggests that marginal students are more price-responsive when academic constraints are less binding for inframarginal students. The negative effect of fee elimination on graduation numbers in both groups of schools may occur because rising enrollment strains resources or because academic constraints continue to bind for marginal students. In contrast, table 4 columns 1 and 2 show that fee elimination lowers the dropout rate from 2.8 to 2.1 percent and the rate of grade repetition from 9.6 to 9.2 percent, though the latter effect is not statistically significant. The graduation, drop-out, progression and grade-specific enrollment results together show that fee elimination keeps students in school for longer and speeds up their progress through schooling. But these effects are dominated by the early years of secondary school and the net effect on reaching and successfully completing grade 12 is negative. This patter may be driven by reductions in resources available to schools or negative peer effects on enrolled students. However, I observe only noisy proxies for measures of school resources or learning environments. Average class size increases from 39.3 to 39.8 students (table 4 column 3). The effect of fee elimination on the socio-economic profile of the enrolled student population is not negative: the share of students eligible for means-tested government social grants is unchanged and the fraction of students who have lost at least one parent to death falls from 20.7 to 19.5 percent.21 These results do not support positive selection into enrollment on family socio-economic status, which might in turn generate negative peer effects of fee elimination on inframarginal enrolled students. However, these data are collected as teachers’ reports of students’ self-reports and so are likely to be measured with substantial error. The effects should therefore be interpreted with caution. 21
The effect on the proportion of students who have lost both parents to death falls by 0.2 percentage points from a baseline of 3.8 percent. The patterns for both definition of orphan are therefore consistent.
28
8
Conclusion
Increasing participation in formal schooling is an important thrust of public policy throughout the developing world. A wide range of countries have employed a wide range of demand- and supply-side interventions to increase student enrollment and attendance. Empirical work in microeconomics has found evidence of positive relationships between formal schooling and individual earnings, health and future children’s human capital accumulation, at least some of which seem to capture causal effects of education. The macroeconomic literature points to a potential role of formal schooling in explaining cross-country differences in per capita income and growth. This paper contributes to this literature and policy by studying the effect of geographically targeted school fee eliminations on primary and secondary school enrollment in South Africa. Feereducing and -eliminating interventions have become popular in developing countries over the past two decades. There is, however, relatively little empirical evidence on their effects. I find that eliminating fees to enroll in South African schools in high-poverty neighborhoods increased enrollment by a relatively small margin. My preferred estimates suggest that 2 to 3 additional students were induced to enroll in the average school. This increased the baseline enrollment level by 1 percent or less and the baseline enrollment rate by less than 1 percentage point. Enrollment in secondary schools rose by 2 to 3 percent and back-of-the-envelope calculations suggest a price elasticity of secondary school enrollment of roughly -0.1. However, the number of students graduating from secondary school actually fell, illustrating an important welfare trade-off of fee elimination. I explore a variety of explanations for these patterns. The results are not explained by timevarying shocks that differentially affect fee-charging and fee-eliminating schools, heterogeneous time trends by observed pre-elimination characteristics, overstated enrollment levels, or transfers between fee-eliminating schools and fee charging schools. They are not fully explained by ceiling effects on enrollment, capacity constraints in schools or credit constraints in households. I conclude that demand for schooling is relatively price insensitive in the neighborhoods treated by the fee elimination intervention. This may be due to low returns to enrollment for marginal students in these schools, because enrollment has a low probability of translating into grade progression and secondary school graduation. Mincerian returns to years of completed education and particularly to secondary school graduation are very high in South Africa (Branson, Ardington, Lam, and Leibbrandt, 2013; Branson and Leibbrandt, 2013). Taken together, these two results imply that the labour market returns to enrollment may be lower for marginal students than for the inframarginal students whose education attainment drives the Mincerian results. This result contrasts with a series of findings in other countries that returns to education are typically at least as large for marginal as inframarginal students (Card, 1999; Duflo, 2001; Heckman, Lochner, and Todd, 2006). The difference may arise because students who are marginal with respect to the fee elimination policy I study would be enrolling in relatively low “quality” schools.
29
My results imply that reducing enrollment costs may have limited impact on school participation levels in settings where enrollment does not translate into substantial learning or grade progression. This emphasizes a potential complementarity between cost-reduction and quality-upgrading interventions and points to an important avenue for future research.
30
References Abadie, A. (2005): “Semiparametric Difference-in-Differences Estimators,” Review of Economics and Statistics, 72, 1–19. Andrabi, T., J. Das, and A. Khwaja (2014): “Report Cards: The Impact of Providing School and Child Test Scores on Educational Markets,” Working paper, Pomona College. Andrews, R., and K. Stange (2016): “Price Regulation and Equality of Opportunity in Higher Education: Evidence from Tuition Deregulation in Texas,” Working paper, University of Texas at Dallas. Baird, S., C. McIntosh, and B. Ozler (2011): “Cash or Condition? Evidence from a Cash Transfer Experiment,” Quarterly Journal of Economics, 126(4), 1709–1753. Banerjee, A., and E. Duflo (2007): “The Economic Lives of the Poor,” Journal of Economic Perspectives, 21(1), 141–168. Barrera-Osorio, F., L. Linden, and M. Urquiola (2007): “The Effects of User Fee Reductions on Enrollment: Evidence from a Quasi-experiment,” Working paper, Columbia University. Borkum, E. (2012): “Can Eliminating School Fees in Poor Districts Boost Enrollment? Evidence from South Africa,” Economic Development and Cultural Change, 60(2), 359–398, Forthcoming. Branson, N., C. Ardington, D. Lam, and M. Leibbrandt (2013): “Changes in education, employment and earnings in South Africa – A cohort analysis,” Discussion Paper 105, Southern Africa Labour and Development Research Unit. Branson, N., and M. Leibbrandt (2013): “Education quality and labour market outcomes in South Africa,” Discussion Paper (2013)13, Organisation for Economic Cooperation and Development. Card, D. (1999): “The Causal Effect of Education on Earnings,” in Handbook of Labor Economics Volume 3A, ed. by D. Card, and O. Ashenfelter, pp. 1801–1863. Elsevier. Deininger, K. (2003): “Does the Cost of Schooling Affect Enrollment by the Poor? Universal Primary Education in Uganda,” Economics of Education Review, 22, 291–305. Department of Basic Education (2006): “School Realities 2006,” https://www.education.gov.za/Programmes/EMIS/StatisticalPublications.aspx.
Accessed at
(2012): “School Realities 2012,” Accessed at https://www.education.gov.za/Programmes/EMIS/Statistica DiNardo, J., N. Fortin, and T. Lemiuex (1996): “Labor Market Institutions and the Distribution of Wages, 1973 - 1992: A Semiparametric Approach,” Econometrica, 64(5), 1001–1044. Duflo, E. (2001): “Schooling and Labor Market Consequences of School Construction in Indonesia: Evidence from an Unusual Policy Experiment,” American Economic Review, 91(4), 795–813. Dupas, P. (2014): “Short-Run Subsidies and Long-Run Adoption of New Health Products: Evidence From a Field Experiment,” Econometrica, 82(1), 197–228. Dynarski, S. (2003): “Does Aid Matter? Measuring the Effect of Student Aid on College Attendance and Completion,” American Economic Review, 93(1), 279–288. 31
Dynarski, S., J. Gruber, and D. Li (2009): “Cheaper by the Dozen: Using Sibling Discounts at Catholic Schools to Estimate the Price Elasticity of Private School Attendance,” Discussion Paper 15461, National Bureau of Economic Research. Edmonds, E. (2006): “Child Labor and Schooling Responses to Anticipated Income in South Africa,” Journal of Development Economics, 81(2), 386–414. Evans, D., M. Kremer, and M. Ngatia (2009): “The Impact of Distributing School Uniforms on Children’s Education in Kenya,” Mimeo. Eyal, K., and I. Woolard (2013): “School Enrolment and the Child Support Grant: Evidence from South Africa,” Discussion Paper 2013/07, Southern Africa Labour and Development Research Unit. Fafchamps, M., and B. Minten (2007): “Public service provision, user fees and political turmoil,” Journal of African Economies, 16(3), 485–518. Fedderke, J., J. Luiz, and R. de Kadt (2000): “Uneducating South Africa: The Failure to Address the 1910–1993 Legacy,” International Review of Education, 46(3/4), 257–281. Filmer, D., and N. Schady (2008): “Getting Girls into School: Evidence from a Scholarship Program in Cambodia,” Economic Development and Cultural Change, 56(3), 581–617. Fizbein, A., and N. Schady (2009): Conditional Cash Transfers: Reducing Current and Future Poverty. World Bank. Hanushek, E. (2003): “The Failure of Input-Based Schooling Policies,” Economic Journal, 113, 64–98. Harding, R., and D. Stasavage (2014): “What Democracy Does (and Doesn’t do) for Basic Services: School Fees, School Inputs, and African Elections,” Journal of Politics, 76(1), 229–245. Heckman, J., L. Lochner, and P. Todd (2006): “Earnings functions, rates of return and treatment effects: The Mincer equation and beyond,” in Handbook of the Economics of Education Volume 2, ed. by E. Hanushek, and F. Welch. North-Holland. Hidalgo, D., M. Onofa, H. Oosterbeek, and J. Ponce (2013): “Can Provision of Free School Uniforms Harm Attendance? Evidence from Ecuador,” Journal of Development Economics, 103(3), 43–51. Hoxby, C. (2001): “All School Finance Equalizations Are Not Created Equal,” Quarterly Journal of Economics, 116(4), 1189–1231. (ed.) (2004): College Choices: The Economics of Where to Go, When to Go, and How to Pay For It. National Bureau of Economic Research. Hyman, J. (2016): “Does Money Matter in the Long Run? Effects of School Spending on Educational Attainment,” Working paper, University of Connecticut. Imbens, G., and K. Kalyanaraman (2012): “Optimal Bandwidth Choice for the Regression Discontinuity Estimator,” Review of Economic Studies, 79(3), 933–959.
32
Jackson, K., R. Johnson, and C. Persico (2016): “The Effects of School Spending on Educational and Economic Outcomes: Evidence from School Finance Reforms,” Quarterly Journal of Economics, 131(1), 157–218. Kane, T. (1994): “College Entry by Blacks since 1970: The Role of College Costs, Family Background, and the Returns to Education,” Journal of Political Economy, 102(5), 878–911. Knight, B., and N. Schiff (2016): “The Out of State Tuition Distortion,” Working paper, Brown University. Kremer, M., and A. Holla (2009): “Improving Education in the Developing World: What Have We Learned from Randomized Evaluations?,” Annual Review of Economics, 1, 513–542. Lam, D., C. Ardington, N. Branson, K. Goostrey, and M. Leibbrandt (2010): “Credit Constraints and the Racial Gap in Post-secondary Education in South Africa,” University of Michigan. Lam, D., C. Ardington, and M. Leibbrandt (2010): “Schooling as a Lottery: Racial Differences in School Advancement in Urban South Africa,” Journal of Development Economics, 95(2), 121–136. Lucas, A., and I. Mbiti (2012): “Access, Sorting and Achievement: The Short-Run Effects of Free Primary Education in Kenya,” American Economic Journal: Applied Economics, 4(4), 226–253. McCrary, J. (2008): “Manipulation of the Running Variable in the Regression Discontinuity Design: A Density Test,” Journal of Econometrics, 142(2), 698–714. Pampallis, J. (2008): School Fees. Centre for Education Policy Development, Johannesburg, ZA. Pritchett, L. (2013): The Rebirth of Education: Schooling Ain’t Learning. Center for Global Development, Washington, DC. Reddy, V., M. Visser, L. Winnar, F. Arends, A. Juan, C. Prinsloo, and K. Isdale (2016): “TIMSS 2015: Highlights of Mathematics Achievement amongst Grade 9 South African Learners,” Discussion paper, Human Sciences Research Council. Schultz, P. (2004): “School Subsidies for the Poor: Evaluating the Mexican PROGRESA Poverty Program,” Journal of Development Economics, 74(1), 199–250. Seekings, J., and N. Nattrass (2005): Class, Race and Inequality in South Africa. Yale University Press, New Haven, CT. Thwala, S. (2010): “The Management of “No Fee” Schools in Mpumalanga: A Case Study of Selected Secondary Schools,” Master’s thesis, University of South Africa. Todd, P., and K. Wolpin (2006): “Assessing the Impact of a School Subsidy Program in Mexico: Using a Social Experiment to Validate a Dynamic Behavioral Model of Child Schooling and Fertility,” American Economic Review, 96(5), 1384–1417. Van der Berg, S., E. Girdwood, D. Sheperd, C. Van Wyk, J. Kruger, J. Viljoen, O. Ezeobi, and P. Ntaka (2013): “The Impact of the Introduction of Grade R on Learning Outcomes,” Discussion paper, University of Stellenbosch, Final Report for the Department of 33
Basic Education and the Department of Performance Monitoring and Evaluation in the Presidency. Van der Berg, S., and M. Louw (2007): “Lessons Learnt From SACMEQII: South African Student Performance in Regional Context,” Discussion Paper 16, Stellebosch University. Wildeman, A. (2008): “Reviewing Eight Years of the Implementation of School Funding Norms,” Idasa Economic Government Programme Research Paper.
34
