Journal of Economic Perspectives—Volume 30, Number 3—Summer 2016—Pages 3–32

The Importance of School Systems:

Evidence from International Differences
in Student Achievement
Ludger Woessmann

A
verage achievement levels of students differ markedly across countries. On
the most recent international achievement tests in math and science, the
average 15 year-old student in Singapore, Hong Kong, Korea, Japan, and
Taiwan is more than half a standard deviation ahead of the average student of the
same age in the United States (Hanushek and Woessmann 2015b). Following the
rule of thumb that average student learning in a year is equal to about one-quarter
to one-third of a standard deviation, these differences are roughly equivalent to
what students learn during 1.5–2 years of schooling. Similarly, the average student
in Finland and Estonia is 40 percent of a standard deviation ahead of the United
States, and the average Canadian student is about one-third of a standard deviation
ahead. On the other hand, the average student in Peru and Indonesia is more than
1.1 standard deviations behind the United States, and achievement in Ghana, South
Africa, and Honduras lags more than 1.5 standard deviations behind the United
States. Overall, average achievement levels among 15 year-olds between the top- and
bottom-performing countries easily differ by more than two standard deviations, or
the equivalent of 6–8 years of learning.
We will present evidence that the considerable differences in student achievement
across countries are systematically related to differences in the organization and gover-
nance of school systems. For example, students in many high-performing countries
such as Korea and Finland, as well as in some provinces of Canada, face external exit
exams at the end of high school. Most schools in Hong Kong and the United Kingdom

■ Ludger Woessmann is Director of the Center for the Economics of Education, ifo Institute
for Economic Research, and Professor of Economics, University of Munich, both in Munich,
Germany. His email address is woessmann@ifo.de.
†
For supplementary materials such as appendices, datasets, and author disclosure statements, see the
article page at
http://dx.doi.org/10.1257/jep.30.3.3 doi=10.1257/jep.30.3.3
4 Journal of Economic Perspectives

have considerable autonomy in deciding which courses to offer and which teachers to
hire, whereas virtually no schools have this autonomy in Greece. More than half the
students in the Netherlands, Belgium, Ireland, and Korea attend privately operated
schools, while hardly any students in Norway and Poland do so. Students in Austria and
Germany are tracked into different-ability schools at age 10, while two-thirds of OECD
countries have comprehensive school systems at least until age 15.
Educational institutions such as national accountability systems or tracking
regimes often vary only slightly or not at all within countries, but an international
perspective provides an opportunity for comparisons. As the director of the first pilot
project comparing student achievement across countries remarked, “If custom and
law define what is educationally allowable within a nation, the educational systems
beyond one’s national boundaries suggest what is educationally possible” (Foshay
1962). Studies done within or across schools of a particular country can also suffer
from the risk of selection bias, if students with specific backgrounds are more likely
to attend certain schools or to become involved in certain programs, while studying
national-level variation circumvents some of these selection biases.
However, these advantages of the cross-country comparative approach come with
some built-in limitations. Identification of causal effects raises particular challenges
in an international setting. Countries may differ in a variety of hard-to-observe ways
such as cultural traits, valuation of achievement, and other preferences that are associ-
ated with both institutional choices and achievement levels. Such unobserved country
heterogeneity gives rise to omitted variable bias in cross-country analyses. Moreover,
only a limited number of country-level observations are available in the test data.
This essay first describes the size and cross-test consistency of international differ-
ences in student achievement. It then uses the framework of an education production
function to describe how different factors of the school system, as well as factors
beyond the school system, are associated with cross-country achievement differences.
The discussion then focuses on research that attempts to go beyond conditional
correlations by addressing some sources of potential bias in cross-country analysis.
This discussion suggests that the role of resource inputs seems limited; indeed, all of
the above-mentioned high-achieving countries spend considerably less on schools per
student than the United States (OECD 2013). But differences in instruction time and
teacher quality do matter. In addition, institutional features including external exams,
school autonomy, private competition, and tracking affect the level and distribution
of student achievement across countries and account for a substantial part of the
cross-country achievement variation. The conclusion points out some major implica-
tions of educational achievement for the prosperity of individuals and nations.

How Large and Consistent Are International Differences in Student

Achievement?

Large-scale international testing of student achievement has more than half a

century of history, and many studies provide evidence on international differences
in student achievement and how they have evolved over time.
 Ludger Woessmann 5

International Rankings and the Size of Cross-Country Differences

A crucial role in the emergence and continuation of comparative testing has
been played by the International Association for the Evaluation of Educational
Achievement (IEA), an independent cooperative of national research institu-
tions and government agencies (IEA 2016; Mullis, Martin, Foy, and Arora 2012).
Following a pilot project in 1959–61, the IEA conducted its first international math
study of eleven countries in 1964. The first science and reading studies occurred
with 12–16 countries in the early 1970s, and a second round in each subject was
performed in the 1980s and early 1990s. Since 1995, the Trends in International
Mathematics and Science Study (TIMSS) has tested math and science achievement
mostly in fourth and eighth grade every four years in between 38 and 52 voluntarily
participating countries. In addition, the Progress in International Reading Literacy
Study (PIRLS) has tested fourth-grade reading achievement every five years since
2001, with 48 countries participating in the most recent wave.
In 2000, the Organisation for Economic Co-operation and Development
(OECD) entered international testing as a second major player. Since then, its
Programme for International Student Assessment (PISA) tests representative
samples of 15 year-old students in math, science, and reading every three years. In
both 2009 and 2012, 65 countries participated, and 71 countries have signed up to
participate in the most recent PISA installment in 2015.1
All these tests draw random samples of students to ensure representativeness for
the national target populations. In particular, the three ongoing studies have a two-
stage sampling design. At a first stage, they draw a random sample of schools in each
country. Within those schools, they then randomly draw one classroom per grade
(TIMSS, PIRLS) or a random sample of 15 year-old students (PISA), respectively.
Each of these tests uses a common set of questions in all participating countries based
on a particular effort to achieve cross-country comparability. PISA, TIMSS, and PIRLS
each link their own tests over time, too. But there is no direct link between the scales
of the three testing regimes or across time with the older tests.
Table 1 shows the performance of the 81 countries that have participated in the
most recent installments of the PISA (2012) and TIMSS (2011) international math
and science tests. Achievement is expressed on the PISA scale, which is standardized
to have a mean of 500 and a standard deviation of 100 among all students in OECD
countries. This standardization was done in 2003 in math and in 2006 in science.
On average across OECD countries, the actual within-country standard deviation is
92 in math and 93 in science (OECD 2013). The transformation of the TIMSS 2011
data to the PISA scale follows the method suggested in Hanushek and Woessmann
(2015b, Annex B).

1
In addition, there are a couple of separate international tests whose items are aligned to the US school
curriculum (which may limit international comparability), a number of regional tests in Latin America
and sub-Saharan Africa, and adult literacy tests (for discussion, see Hanushek and Woessmann 2011a,
table 2; Hanushek and Woessmann 2015b, chapter 4; Hanushek, Schwerdt, Wiederhold, and Woessmann
2015). The International Association for the Evaluation of Educational Achievement has also conducted
studies in other subjects such as foreign languages, civic education, and computer literacy.
6 Journal of Economic Perspectives

Table 1
Performance on Recent International Student Achievement Tests, 2011–2012

Country Score Country Score Country Score

Shanghai–China 596 Norway 492 Malaysia 420

Singapore 562 Luxembourg 491 Costa Rica 418
Hong Kong–China 558 Spain 490 Mexico 414
Korea 546 Italy 489 Uruguay 413
Japan 542 United States 489 Montenegro 410
Chinese Taipei 542 Portugal 488 Bahrain 408
Finland 532 Lithuania 487 Lebanon 403
Estonia 531 Hungary 486 Georgia 401
Liechtenstein 530 Iceland 485 Brazil 398
Macao–China 529 Russian Federation 484 Jordan 397
Switzerland 523 Sweden 482 Argentina 397
Netherlands 523 Croatia 481 Albania 396
Canada 522 Slovak Republic 476 Tunisia 393
Poland 522 Ukraine 468 Macedonia 392
Vietnam 520 Israel 468 Saudi Arabia 391
Germany 519 Greece 460 Palestine 388
Australia 513 Turkey 456 Colombia 388
Ireland 512 Serbia 447 Qatar 380
Belgium 510 Bulgaria 443 Syria 379
New Zealand 508 Romania 442 Indonesia 379
Slovenia 508 United Arab Emirates 441 Botswana 376
Austria 506 Cyprus 439 Peru 371
United Kingdom 504 Thailand 435 Oman 369
Czech Republic 504 Chile 434 Morocco 348
Denmark 499 Kazakhstan 428 Honduras 328
France 497 Armenia 428 South Africa 315
Latvia 496 Iran 422 Ghana 291

Notes: The table gives the average of the scores on international math and science tests. Black: PISA 2012,
15-year-olds. Grey: TIMSS 2011, 8th grade, transformed to PISA scale as in Hanushek and Woessmann
(2015b).

The cross-country differences in knowledge among same-aged students are in

some cases extremely large. Remember, as a rule of thumb, learning gains on most
national and international tests during one year are equal to between one-quarter
and one-third of a standard deviation, which is 25–30 points on the PISA scale. Thus,
the achievement difference between the average 15 year-old in the United States and
in the PISA top performers—Singapore, Hong Kong, Korea, Japan, Taiwan, Finland,
and Estonia—is roughly twice what students usually learn during one year. At the
other end, the average difference of US achievement to the PISA bottom performers
(Peru and Indonesia) amounts to the equivalent of three to four years of learning,
and to five to six years to the TIMSS bottom performers (Ghana and South Africa).
In looking at lists like Table 1, it is important to focus on scores, not just ranks.
For example, in the PISA 2012 math test, the achievement levels of most countries
are not statistically significantly different from their closest 1–3 neighbors above and
below. Where the scores are closely bunched, Portugal’s achievement at rank 31 does
not differ significantly from ranks 25–37 in the PISA 2012 math test (OECD 2013).
 The Importance of School Systems 7

Figure 1
Distribution of Student Achievement in Selected Countries on PISA Math Test,
Compared to All OECD Countries

A: United States B: Belgium

.005 .005
.004 .004
.003 .003
.002 .002
.001 .001
0 0
200 400 600 800 200 400 600 800
Score Score

C: Finland D: Korea
.005 .005
.004 .004
.003 .003
.002 .002
.001 .001
0 0
200 400 600 800 200 400 600 800
Score Score

Notes: Kernel densities of student achievement on the PISA 2012 math test. Bold solid line: specified
country. Dotted line: OECD countries.

The means in Table 1 may hide important differences in the shape of the
overall distribution of achievement in a country. Figure 1 displays the achievement
distribution on the PISA 2012 math test for the United States and three selected
countries with relatively high performance, comparing each to the overall distribu-
tion in OECD countries. The US distribution is shifted to the left and slightly more
left-steep compared to the OCED distribution, but it does not have a particularly
strong left or right tail. As the three example countries show, it is possible to achieve
above-average mean performance with a relatively equitable distribution (Finland),
with a distribution that is mostly just shifted to the right of the OECD distribution
(Korea), or with a relatively unequal distribution (Belgium).
The relatively low performance of the United States compared to many OECD
countries cannot be attributed to the particularly poor performance of a small group
of students or of students from disadvantaged backgrounds. For example, the 25th,
50th, and 75th percentiles of the US distribution on the PISA 2012 math test are all
between 13 and 15 points below the OECD average of the respective percentiles. In
Hanushek, Peterson, and Woessmann (2013), my coathors and I document that both
the proportion of students who achieve at a basic proficient level and the proportion
of students who achieve at an advanced level in the United States are comparatively
low in an international perspective. In addition, in Hanushek, Peterson, and Woess-
mann (2014), we show that the ranking of US students from better-educated families
when compared to students from better-educated families in other countries is not
8 Journal of Economic Perspectives

much different from the ranking of US students from less-well-educated families

when compared to students from less-well-educated families in other countries.

Consistency across Different Tests

The measurement of educational achievement is subject to many psychometric
and measurement choices. For example, the target population of the TIMSS test
is eighth graders. Also, TIMSS has a strong curricular focus and is based on an
assessment framework developed in a collaborative process with participating coun-
tries, with a test-curriculum matching analysis describing how the test matches each
participating country’s school curriculum. On the other hand, the target popula-
tion of the PISA test is 15 year-olds, and PISA aims to assess the knowledge and skills
essential for full participation in modern society, including the extrapolation and
application of learned knowledge to new real-life situations.
How sensitive are international comparisons to specific measurement choices?
We can compare the achievement of the 28 countries that participated in the most
recent installments of both tests: PISA 2012 and TIMSS 2011. Despite the differences in
timing, target populations, and conceptual approaches, the correlation across the 28
countries participating in both tests is 0.944 in math and 0.930 in science (Hanushek
and Woessmann 2015b). This high correlation suggests that when it comes to interna-
tional comparisons, specific test designs are of secondary importance.
Another potential issue with international achievement tests is cross-country
differences in sample selectivity due to different rates of enrollment, exclusion, and
nonresponse. While sampling was devised to be representative of the student population
in each participating country, some countries do not have universal enrollment at age
15, when students are tested in PISA. In addition, nonrandom differences in patterns
of sample exclusions (for example, for handicapped children) and nonresponse can
compromise comparability across countries. However, the working paper version of
Hanushek and Woessmann (2011c) shows that although these factors are related to
average country scores, controlling for these rates does not affect the qualitative results
on institutional effects in international education production functions presented later
in this paper. The variation in the extent to which countries adequately sample their
entire student populations appears orthogonal to the associations analyzed here.

Changes over Time

While an assessment of countries at a point in time is reasonably straightforward,
assessing changes in country performance over time is harder. The early interna-
tional tests, in particular, constitute separate testing incidents without links across
different tests. In Hanushek and Woessmann (2012, 2015a), we use an empirical
calibration method to put all international tests from 1964 to 2003 on a common
standardized scale. Our analysis shows that 73 percent of the variance across the 693
separate test observations in 50 countries occurs between countries. The remaining
27 percent combines true changes over time in countries’ scores and any measure-
ment error in the testing. That is, most of the variation in the available panel data
of countries over time is across rather than within countries, implying that a large
share of the country differences are consistent over time.
 Ludger Woessmann 9

Figure 2
Long-Run Test Score Trends in Selected Countries, 1964–2012

560
Korea

Japan
540

Finland
520
Test score

Canada
Germany
United Kingdom
500 France
United States

Sweden
480
Norway

460
1960 1970 1980 1990 2000 2010

Source: Extended from Hanushek and Woessmann (2015a).

Notes: Stylized depiction of standardized data from international tests 1964–2012. The figure is based on
age-group- and subject-specific standardized scores from all international tests in 1964–2003 extended
with the subsequently available TIMSS, PIRLS, and PISA data to 2012. It takes out age-group- and subject-
specific trends in each country, smooths available test observations with locally weighted regressions, and
linearly interpolates between available test observations; see Hanushek and Woessmann (2015a) for details.

Still, several countries do show either significant improvements or declines over

time. Figure 2 depicts achievement trends observed in selected example countries
from 1964 to 2012. The more limited variation in early decades likely reflects the lower
frequency of testing before 2000. The figure shows substantial cross-sectional differ-
ences across countries. But some countries do show noteworthy changes over time.
Ripley (2013) acknowledges that a previous version of this figure motivated her work
on the widely acclaimed New York Times bestseller The Smartest Kids In The World—
And How They Got That Way. While the United States was rather typical compared to
most other countries, she wrote there were a few countries where “virtually all kids
were learning critical thinking skills in math, science, and reading” (p. 4). While some
countries such as Canada and Finland over the 1980s and 1990s and Germany and
Japan more recently did improve substantially over time, other well-off countries dete-
riorated, such as Norway during the 1990s, Sweden during the 2000s, and Finland in
recent years. Educational achievement levels of countries seem generally consistent
over time, but they are not set in stone and can be mutable.

Descriptive Patterns Using an Education Production Function

This section uses the framework of an international education production func-

tion to document the extent to which, on a purely descriptive basis, differences in
family background, school resources, and institutions can account for cross-country
10 Journal of Economic Perspectives

differences in student achievement. These inputs are probably not exogenous to

student achievement, so correlations between the inputs and test scores are very
likely to be biased by omitted variables, selection, and reverse causation. While these
descriptive patterns must be interpreted cautiously, they can serve as a useful guide
to the more explicit discussions of causality that follow.

International Education Production Functions

An education production function models the output of education as a func-
tion of different inputs (for example, Hanushek 1986, 2002). We combine the input
factors into three groups: family background factors, school resources, and insti-
tutional structures of school systems. The first group is mostly outside the control
of school systems. The other two groups of factors reflect the quantity of resource
inputs in the systems and the institutional structures. The basic model can be
extended to include interactions between input factors.
A substantial literature has estimated such international education production
functions using cross-sectional data (for an extensive review, see Hanushek and
Woessmann 2011a). Early studies used aggregate country-level data to study the
country-level variation in achievement scores (for example, Bishop 1997; Hanushek
and Kimko 2000; Lee and Barro 2001). More recent studies also use country-level
data to study, for example, the correlates of gender equality in achievement (Guiso,
Monte, Sapienza, and Zingales 2008; Fryer and Levitt 2010).
However, starting with Woessmann (2003b), a number of studies have used
the data from international achievement tests at the student level to estimate cross-
country education production functions. Examples include Woessmann (2005b),
Fuchs and Woessmann (2007), Brunello and Checchi (2007), Woessmann, Luede-
mann, Schuetz, and West (2009), Schneeweis (2011), and Ammermueller (2013).
Because these studies use data on individual students, they can hold constant a large
set of observable factors usually unavailable in national datasets. In effect, they can
compare “observationally equivalent” students across countries.
For concreteness, Table 2 provides an example of a basic cross-sectional esti-
mation of an international education production function.2 The table shows the
categories of data that are available. The dependent variable is the score from
the PISA 2003 math test, with the sample restricted to the 29 participating OECD
countries to provide greater comparability. The model includes a large number of
explanatory variables in the three groups of input factors: family background, school
resources, and institutions. The individual-level measures of family b ­ ackground

2
This is a simplified version of the model used in Woessmann et al. (2009) and Hanushek and Woessmann
(2011a). To allow for a more meaningful accounting analysis below, it drops the GDP per capita of the country
(which is correlated with educational spending at 0.93 and yields a counterintuitive negative estimate), class
size (which has a counterintuitive positive estimate), and the imputation dummies and their interactions with
the main variables contained in those models. Qualitative results are similar with those variables included.
Qualitative results are also unaffected when adding the country-average value of the Index of Economic,
Social and Cultural Status (ESCS), the average share of students with an immigrant background in a country,
or continental fixed effects to the model. Country-average ESCS in fact enters marginally significantly nega-
tively and the migrant share insignificantly. Reported standard errors are clustered at the country level, which
may be overly conservative for variables that vary at the school or student level.
 The Importance of School Systems 11

are taken from student background questionnaires that students complete in the
PISA study; the measures of school resources and institutions are mostly taken
from school background questionnaires that the principals of participating schools
complete; these measures are combined with country-level data on expenditure per
student and external exit exams that come from outside sources (for details, see
Appendix A of Woessmann et al. 2009). Descriptively, this model accounts for 34
percent of the achievement variance at the individual student level.

Factors beyond the School System: Family, Socioeconomic, and Cultural Background
Some of the personal characteristics that have meaningful and statistically signif-
icant magnitudes in Table 2 include student characteristics such as age, gender, and
participation in early childhood education, along with indicators for family status,
parental education, parental work status and occupation, the number of books at
home, immigration background, and the language spoken at home. For example,
the achievement difference between students in the highest category of more than
200 books at home versus the lowest category of fewer than 10 books at home—a
proxy for aspects of educational, social, and economic background—amounts to
more than half a standard deviation in the PISA test score.
There are two main types of analysis in the literature analyzing socioeconomic
backgrounds in the international tests. The first type looks at how much socioeco-
nomic background contributes to country-level differences in educational outcomes.
The second type of analysis compares the within-country association of socioeconomic
factors with student achievement, sometimes referred to as socioeconomic gradients,
across countries. For example, in Schütz, Ursprung, and Woessmann (2008), we esti-
mate the associations of family background with student achievement—interpreted as
measures of the inequality of educational opportunity—in different countries using
TIMSS data and relate them to measures of institutions of the school systems. We
show that family background effects are systematically larger in countries with early
tracking and less-extensive pre-primary education systems.3 Jerrim and Micklewright
(2014) use PISA and PIRLS data to analyze the extent to which cross-country compari-
sons of socioconomic gradients are affected by differences in reporting errors.
Several studies have focused on the achievement of children with an immigra-
tion background, looking at both socioeconomic and institutional characteristics.
For example, Dustmann, Frattini, and Lanzara (2012) show that in many countries,
observed differences in parental background (including parental education and occu-
pation and the language spoken at home) can account for much of the lower PISA
achievement of children of immigrants compared to native children. They also find
that children of Turkish immigrants perform better in most host countries than Turkish
children in Turkey. Also using PISA data, Cobb-Clark, Sinning, and Stillman (2012)
show that the migrant–native achievement gap is significantly associated with institu-
tional features of the host countries such as school starting age, ability tracking, private

3
Applying a similar approach to outcomes beyond school age, Brunello and Checchi (2007) find that
early tracking is related to larger effects of family background on educational attainment and earnings
in the labor market, but not on on-the-job training and adult literacy.
12 Journal of Economic Perspectives

Table 2
A Simple International Education Production Function: A Least-
Squares Regression
(dependent variable is student’s mathematics test score)

Coefficient Standard error

Family Background
Age (years) 17.825*** (3.160)
Female −14.733*** (1.639)
Preprimary education (more than 1 year) 6.832*** (2.428)
School starting age −3.869* (2.030)
Grade repetition in primary school −54.579*** (4.734)
Grade repetition in secondary school −33.726*** (6.702)
Grade
7th grade −47.003*** (10.051)
8th grade −19.213* (10.242)
9th grade −6.772 (6.896)
11th grade −3.275 (5.236)
12th grade 11.949* (6.398)
Living with
Single mother or father 20.045*** (3.949)
Patchwork family 22.678*** (4.286)
Both parents 29.524*** (3.956)
Parents’ working status
Both full-time −2.071 (2.911)
One full-time, one half-time 8.820*** (2.327)
At least one full time 15.926*** (2.891)
At least one half time 10.531*** (2.278)
Parents’ job
Blue collar, high skilled 1.481 (2.365)
White collar, low skilled 3.743* (1.870)
White collar, high skilled 8.189** (3.144)
Books at home
11–25 books 6.760*** (2.290)
26–100 books 24.749*** (2.789)
101–200 books 34.232*** (3.161)
201–500 books 54.400*** (3.238)
More than 500 books 54.166*** (3.703)
Immigration background
First-generation student −11.447** (4.442)
Nonnative student −13.776** (5.375)
Language spoken at home
Other national dialect or language −17.689** (7.064)
Foreign language −7.887*** (2.677)
Index of Economic, Social and Cultural Status (ESCS) 19.926*** (2.153)
s
Community location
Town (3,000–100,000) 9.101** (3.323)
City (100,000–1,000,000) 16.951*** (3.989)
Large city with > 1 million people 13.939*** (4.929)

Continued on next page

 Ludger Woessmann 13

Table 2 (continued)

Coefficient Standard error

School Resources
Cumulative educational expenditure per student 0.270** (0.103)
(1,000 $)c
Shortage of instructional materials s
Large shortage −8.737** (3.514)
No shortage 8.678*** (2.015)
Instruction time (minutes per week) 0.044*** (0.015)
Teacher education (share at school)s
Fully certified teachers 7.699 (8.588)
Tertiary degree in pedagogy 10.211 (6.547)

Institutions
Competition c
Private operation (country share) 56.941*** (9.758)
Government funding (country share) 57.847*** (19.486)
Accountability
External exit examsc 9.433 (9.055)
Assessments used for student retention/promotions 11.744** (4.320)
Monitoring of teacher lessons by principals 6.785* (3.442)
Monitoring of teacher lessons by external 4.842* (2.816)
  inspectorss
Assessments used to compare school to 4.188 (2.870)
  district/nations
Assessments used to group studentss −8.261** (3.021)
Autonomy and its interaction with external exit exams s
Autonomy in establishing starting salaries −15.769*** (5.229)
External exit exams × Autonomy in establishing 14.550* (8.104)
  starting salaries
Autonomy in formulating budget −9.624 (6.901)
External exit exams × Autonomy in formulating 7.882 (8.478)
  budget
Autonomy in determining course content −2.053 (5.435)
External exit exams × Autonomy in determining 11.504 (7.262)
  course content
Autonomy in hiring teachers 18.349* (10.436)
External exit exams × Autonomy in hiring teachers −24.723** (11.796)
Constant 116.126** (51.774)
Students 219,794
Schools 8,245
Countries 29
R 2 (at student level) 0.340

Source: Own calculations on the basis of Woessmann et al. (2009) using data from the
Programme for International Student Assessment (PISA) 2003; the sample is OECD
countries.
Notes: The table presents results from a least-squares regression weighted by students’
sampling probability. The dependent variable is student’s mathematics test score.
Measures vary at the student level unless noted otherwise. Robust standard errors
adjusted for clustering at the country level in parentheses.
s
Observed at school level.
c
Observed at country level.
***, **, and * represent significance levels of 1, 5, and 10 percent, respectively.
14 Journal of Economic Perspectives

schools, and teacher evaluation in a cross-sectional model. In a country-level analysis of

the PISA data, Brunello and Rocco (2013) find that an increased share of immigrant
students has a small negative effect on the achievement level of native students.
Overall, socioeconomic factors contribute substantially to the cross-country
variation in test scores.4 These factors, however, are largely outside the influence of
school systems—although not necessarily beyond the effects of other family, social,
and redistributive policies.

Factors of the School System: Inputs and Institutions

Measures of school resources often fail to achieve economic and statistical
significance in international education production functions, and sometimes
even show counterintuitive coefficients. In Table 2, the point estimate on school
spending is very small: An increase in cumulative educational expenditure per
student until age 15 by $25,000, or one standard deviation, is associated with an
increase in student achievement of less than 7 percent of a standard deviation. If
class size as observed at the individual student level is added to the model, it has a
counterintuitive positive coefficient—purportedly indicating that students achieve
at higher levels in larger classes. Other variables have a more intuitive interpre-
tation: for example, students perform worse in schools whose principal reports
that the school’s capacity to provide instruction is hindered by a shortage or inad-
equacy of instructional materials such as textbooks. Both weekly instruction time
and measures of teacher education are positively associated with student achieve-
ment. Evidence from TIMSS, which provides more detailed teacher information
from individual teacher background questionnaires, shows similar results (Woess-
mann 2003b). To the extent that schools with more resources in the tested grade
also tended to have more resources in earlier grades, the coefficient estimates on
resources capture not just the contemporaneous effect of resources in the specific
grade, but the cumulative effect of resources over the previous grades.
In contrast, institutional features of school systems are strongly associated with
student achievement in studies of this sort. Table 2 offers some examples, as do
Woessmann (2003b, 2005b), Fuchs and Woessmann (2007), and Woessmann et al.
(2009). In particular, measures of the extent of private school operation, govern-
ment funding of schools, and different features of school accountability such as
external exit exams, the use of assessments, and monitoring of lessons are positively
related to student outcomes.5 In addition, there is a tendency for school autonomy
in different decision-making areas to be negatively related to student achievement
in systems without external exit exams but to be unrelated or positively related in
systems where external exit exams promote accountability (Woessmann 2005b). In

4
Additional factors analyzed with international achievement data include gender differences (for
example, Guiso et al. 2008; Fryer and Levitt 2010), relative age at school entry (for example, Bedard and
Dhuey 2006), and peer effects (for example, Ammermueller and Pischke 2009).
5
External exit exams reach statistical significance in a specification of the model of Table 2 that excludes
the interactions with school autonomy. Results on the country-level variables in Table 2 are qualita-
tively the same in a two-step specification that first estimates Table 2 with country fixed effects and then
regresses the coefficients captured on these fixed effects on the country-level variables.
 The Importance of School Systems 15

a study of a variable not included in Table 2, Edwards and Garcia Marin (2015) find
no significant association of country-aggregate student achievement in the PISA test
with whether the right to education is included in a country’s political constitution.
The results on instruction time, teacher education, and institutional effects
provide a prima facie case for the relevance of school systems. Another piece of
evidence for this relevance arises from adding school fixed effects to the estima-
tion of an international education production function. Using PISA data, Freeman
and Viarengo (2014) show that estimated school fixed effects are associated with
observable school policies and teaching practices as well as with socioeconomic
gradients. While they do not rule out nonrandom selection into schools as playing a
role here, they interpret these results as indications of the potential importance of
what schools do, as opposed to national or individual traits.
While most of the international achievement datasets are cross-sectional, Singh
(2015) uses a longitudinal dataset that observes individual students at ages 5 and 8 in
four developing countries. The findings show that the large cross-country learning
gaps between low-performing Peru and high-performing Vietnam (apparent earlier
in Table 1) are virtually nonexistent at school-entry age. They emerge over the first
few school years in a way that is most consistent with large cross-country differences
in the productivity of a school year (estimated from discontinuities in completed
grades emerging from birth months in combination with enrollment thresholds),
rather than with observed differences in socioeconomic background and time use.
Again, these findings suggest that school systems have important effects.

Accounting for the Cross-Country Variation in Test Scores

As indicated, the model in Table 2 accounts for about one-third of the total student-
level variation in the international model. This variation includes within-country
variation as well as cross-country variation. The former is likely to include a component
of random measurement error because of idiosyncrasies in individual performance on
the testing day, a component that would cancel out at the national level.
So to what extent can family background factors, school resources, and institu-
tions account for differences in student achievement across countries? To answer this
question, we have to combine the large number of explanatory variables into a smaller
number of factors. The student-level estimation of Table 2 provides one coefficient
per variable: that is, it effectively forces the between-country associations of student
achievement with the input factors to be the same as the within-country associations.
We use these coefficient estimates on the individual variables in the model of Table 2 to
combine the family background variables into one factor. That is, we simply calculate a
linear combination that is the sum of the products of the individual variables times their
respective coefficient estimates. We do the same for the school resource variables and
the institutional variables. We then collapse the three combined input factors to the level
of the 29 OECD country observations to obtain three aggregate country-level variables.
For descriptive purposes, we regress aggregate academic achievement on these
three composite inputs for the 29 country-level observations. The share of the cross-
country variance in achievement accounted for by the three input factors is 83
percent. That is, using the student-level model to additively and linearly combine
16 Journal of Economic Perspectives

Table 3
Accounting for the Achievement Variance at the Country Level

Family School All three

background resources Institutions factors
Accounted variance when only this factor 0.504 0.181 0.533 0.834
is included in the model
Change in accounted variance when this 0.208 0.045 0.259
factor is added to a model that
already includes the other two factors

Source: Author using data from the PISA 2003.

Notes: The table shows the share of the country-level variance in PISA 2003 mathematics test scores
accounted for by the respective factor. Each factor represents a linear combination of individual variables
using coefficient estimates from the student-level regression shown in Table 2, collapsed to the country level.

the input variables into three factors that can be collapsed to the country level,
our simple international education production function descriptively accounts for
more than four-fifths of the total cross-country variation in student achievement.
Table 3 breaks this explained variance in the country-level model down into
components accounted for by the three groups of input factors. As in any regres-
sion analysis, the contribution of each factor depends on the other variables in the
model. However, the role of family background factors appears substantial, contrib-
uting between 21 and 50 percent to the total cross-country variance in student
achievement. By contrast, the contribution of school resources is much smaller, at
4 to 18 percent. Institutional differences again contribute importantly to the cross-
country achievement variation, at 26 to 53 percent.6
Details of the extent to which the simple model accounts for the achievement
of individual countries are shown in Table 4. For each country, the table shows how
much of the country’s difference from the international mean can be accounted
for by each set of input factors.7 For 14 of the 29 countries, the unaccounted-for
residual achievement is less than 10 percent of a standard deviation. But for some
examples, the model does not perform very well: in top-performing Finland, only
12.9 of the 44.5 percentage points of superior achievement (in standard deviations)
are accounted for by the model. Differences from the international mean in family

6
Compared to the models in Woessmann et al. (2009) and Hanushek and Woessmann (2011a), the
model here excludes GDP per capita and class size, whose counterintuitive coefficients would hamper
the interpretation of the accounting analysis. Including them would, in fact, reduce the separate contri-
butions accounted for by the family background and school resource factors at the country level. Results
are similar when including the imputation dummies contained in those models. It is debatable whether
the model should include grade levels, individual grade repetition, and school starting age; however,
results are similar when excluding these variables. The family background factor includes both individual
student characteristics and genuine family factors; when separating the two, most of the country-level
contribution goes to the genuine family factors and little to the student characteristics.
7
To estimate the contribution of each input factor, we first run the country-level model on demeaned
variables and then multiply the respective coefficient estimates with each country’s value of the respec-
tive input factor. The contributions of the three input factors then sum to the predicted value (shown as
“accounted difference” in Table 4) in this model.
 Ludger Woessmann 17

Table 4
Accounting for Each Country’s Difference from the International Mean

Of which: accounted for by

Observed Unaccounted Accounted Family School
difference difference difference background resources Institutions
(1) (2) (3) (4) (5) (6)
Finland 44.5 31.7 12.9 2.7 −1.3 11.5
Korea 42.0 14.3 27.7 13.0 5.6 9.1
Netherlands 38.4 −8.0 46.4 −3.4 −0.3 50.1
Japan 34.0 4.4 29.6 17.5 2.9 9.2
Canada 33.0 17.4 15.6 15.9 3.2 −3.5
Belgium 29.5 −11.8 41.3 −1.2 1.4 41.0
Switzerland 26.5 27.3 −0.8 −13.2 9.5 2.9
Australia 24.5 2.1 22.4 14.0 6.6 1.7
New Zealand 24.5 17.8 6.7 16.2 −3.0 −6.4
Czech Republic 16.4 2.1 14.3 16.1 −9.0 7.2
Iceland 15.1 −11.6 26.7 29.7 4.9 −7.9
Denmark 14.1 6.0 8.1 0.4 6.5 1.2
Sweden 10.0 5.5 4.5 5.9 −1.0 −0.4
United Kingdom 8.4 −9.1 17.5 13.0 2.7 1.8
Austria 5.5 5.7 −0.2 2.1 6.1 −8.5
Ireland 3.9 −15.0 18.8 −3.3 1.6 20.5
Germany 3.5 5.4 −1.9 −4.0 −0.8 2.8
Slovak Republic −1.0 6.3 −7.3 4.2 −18.0 6.5
Norway −4.3 −26.4 22.1 22.1 2.1 −2.1
Luxembourg −6.3 −10.7 4.4 −25.5 19.3 10.6
Hungary −9.3 −18.7 9.4 4.5 −5.4 10.4
Poland −9.5 2.5 −12.0 −11.5 −8.1 7.6
Spain −14.1 −2.7 −11.4 −4.8 −5.4 −1.2
United States −16.1 −14.7 −1.4 2.3 9.1 −12.9
Portugal −33.5 23.0 −56.5 −27.0 −2.8 −26.7
Italy −33.9 −5.5 −28.3 2.7 3.6 −34.7
Greece −55.1 −22.1 −33.0 −4.1 −3.0 −26.0
Turkey −75.8 −4.4 −71.5 −31.7 −17.5 −22.3
Mexico −114.8 −10.6 −104.2 −52.7 −9.9 −41.6

Notes: Each entry shows the country’s test score difference from the international mean on the PISA 2003
mathematics test, expressed in student-level standard deviations. Column 1: actual difference. Column 2:
difference not accounted for by a country-level regression of the actual test score difference on the three
combined input factors (family background, school resources, institutions), each of which is measured as
a linear combination of individual variables using coefficient estimates from the student-level regression of
Table 2, collapsed to the country level. Column 3: difference accounted for by this country-level regression.
Columns 4–6: difference accounted for by family background, school resources, and institutions,
respectively. By constructions, columns 2 and 3 sum to column 1, and columns 4–6 sum to column 3.

background and school resources hardly contribute to this, but 11.5 percentage
points are contributed by differences in the institutional setting, which in Finland
include the existence of external exams, almost universal use of assessments for
student retention, and widespread school autonomy over course content. For
Korea, about two-thirds of the high relative achievement is accounted for by the
model, and all three groups of input factors contribute to this, including a large
share of privately operated schools, external exams, widespread monitoring of
18 Journal of Economic Perspectives

teacher lessons, and universal course-content autonomy. For third-achieving Neth-

erlands, the model in fact over-predicts its high achievement, and all of this is due to
superior institutions—in particular, the largest share of privately operated schools,
external exams, widespread course-content autonomy, and use of assessments for
retention. At the lower end, most of the poor performance of Mexico and Turkey
is accounted for by the model, in particular detrimental family background and
institutions. The model does not do well at predicting US performance; institutions
such as salary autonomy without external exit exams would predict the lower-
than-average achievement level, but better family background and, in particular,
abundant school resources would point the other way.

Inputs to the School System: Explorations into Causal Effects

Inputs are clearly not exogenous to the education process. There may be reverse
causation, for example, if educational systems assign additional resources to schools
that serve low-achieving students, or if schools with poor student outcomes are induced
to implement specific reforms. There may be bias from selection in that parents from
low-achieving (or high-achieving) students tend to select into schools that offer specific
resources for their children, or if high-performing schools have some ability to select
high-achieving students. There may be omitted variables correlated with both inputs
and outcomes, including country-level factors such as culture and valuation of educa-
tion that may drive both inputs and learning effort, and also differences in preferences
for high-quality education among parents or differences in motivation or ability of
students. The direction of the net bias from such factors is not always obvious.
As a straightforward first step to exclude certain sources of bias when analyzing
the possible effect of expenditure per student, one can ignore differences in the
levels of expenditure and only use changes in average country expenditure over time
as an explanatory variable in first-difference or differences-in-differences panel-type
models. To the extent that sources of bias such as countries’ cultures and parental
background do not change significantly over time, they will no longer bias estimates
based on changes in expenditure. In this spirit, in Gundlach, Woessmann, and
Gmelin (2001), we calculated changes in expenditure and changes in test perfor-
mance in several OECD countries over a 25-year period (1970–1994), finding that
even substantial increases in real expenditure per student did not go hand in hand
with improvements in student achievement.
More recently, the linking of the PISA tests over time allows for a direct compar-
ison of spending changes to changes in achievement. As is directly obvious from
Figure 3, changes in PISA performance from 2000 to 2012 are not systematically
related to concurrent changes in expenditure per student. Countries with large
spending increases do not show different achievement trends from countries that
spend only little more.8 While this analysis may be attenuated by the fact that changes

8
The coefficient estimate on expenditure in the simple underlying first-differenced regression is insig-
nificant. Similarly, using data from the first three PISA waves, the working-paper version of Hanushek
 The Importance of School Systems 19

Figure 3
Changes in Educational Spending and in Student Achievement across Countries

40 POL
Change in PISA reading score 2000–2012 ISR

20 PRT
JPN
CHE
KOR
HUN
ITA BELMEX
0 FRADNK NOR CZE
AUT
ESP IRL
USA
CAN
AUS
−20 FIN
ISL

SWE

−40
100 150 200
Expenditure per student in 2010 (2000 = 100)

Source: Hanushek and Woessmann (2015a) based on OECD data.

Notes: Scatter plot of the expenditure per student in 2010 relative to 2000 (constant prices, 2000 = 100)
against change in PISA reading score, 2000–2012.

in expenditure may take some time to translate into actual inputs and then to affect
student outcomes, the 25-year time horizon of the previous analysis should be able
to reflect major effects. Of course, if other factors changed in a way correlated with
both spending and test outcomes, looking for correlations between them—whether
in levels or in differences—would still suffer from bias in these aggregate analyses.
Several studies have sought to use arguably exogenous variation in particular
inputs by applying more elaborate identification methods. Here, we will discuss
some of the evidence suggesting that smaller class sizes do not make much differ-
ence to educational outcomes but that more instruction time and higher teacher
quality do make a difference.

Class Size
Most of the efforts that seek to uncover a causal effect of class size on test
outcomes using international data turn to within-country variation. For example, in
each school, natural cohort fluctuations in enrollment give rise to random class-size
variation between adjacent grades (Hoxby 2000). In Woessmann and West (2006), we
combine school fixed effects—which seek to eliminate between-school variation—
with an approach that uses average class size in the school’s grade as an instrumental
variable, thus eliminating bias from sorting within a grade in a school. Applying this
identification strategy to TIMSS data in 18 countries, we find significant beneficial

and Woessmann (2011b) reports insignificant negative coefficient estimates on expenditure per student
in first-differenced and fixed-effects models.
20 Journal of Economic Perspectives

effects of smaller classes in only two countries and can rule out large class-size effects
in the majority of countries. Our estimates using this approach suggest that conven-
tional cross-sectional estimates of class-size effects are substantially biased.9
These results are in line with results from a second quasi-experimental
identification strategy suggested by Angrist and Lavy (1999) that exploits the exis-
tence of maximum class-size rules in many countries. Say that the maximum class
size is 40, and that a certain grade has 120 students divided into three classes of
40 students each. If the grade rises to 121 students, the group is then divided into
four classes—three of 30 students and one of 31 students. In this way, the rules give
rise to discontinuous jumps in average class sizes whenever the enrollment in a
grade in a school passes multiples of the maximum class size. Exploiting the induced
class-size variation for ten European countries in a regression discontinuity design
using TIMSS data, the results in Woessmann (2005a) rule out large causal class-
size effects in all countries, with statistically significant but small effects in only two
countries. Furthermore, the cross-country variation in estimated class-size effects in
both studies is consistent with an interpretation that smaller classes have beneficial
effects only in countries with relatively low teacher quality, as measured by relative
teacher salary and teacher education.
The latter result is also confirmed in Altinok and Kingdon (2012), who apply
yet another identification strategy to estimating class-size effects. To avoid bias
from nonrandom sorting of students into schools and from unobserved student
and family characteristics, they exploit the fact that the same students are tested
in different subjects in TIMSS—math and science (sometimes in several specific
domains). Using student fixed effects, they identify class-size effects from variation
in class size between the two subjects for the same students (in countries where such
variation exists). They find significant class-size effects in only 14 of 47 countries and
even these are mostly small, confirming that class sizes play a limited role at best in
understanding achievement differences in the international data.

Instruction Time
The length of school instruction time seems to play a more important role in
educational achievement. For example, in an attempt to address omitted variable
bias from unobserved individual subject-invariant characteristics such as under-
lying ability, motivation, or parental support, Lavy (2015) applies the within-student
between-subject identification approach to estimate the effect of instruction time
in the PISA 2006 data. The approach exploits the fact that different students have
different instruction times in math, language, and science. He finds that instruction
time has a significant positive effect on student achievement that is modest to large,
suggesting that increasing instruction time by one hour per week would increase
achievement by 6 percent of a standard deviation in OECD and Eastern European

9
Applying the same instrumental variable strategy combined with school fixed effects—as well as an
identification strategy based on restrictions placed on higher moments of the error distribution—to the
PISA math data for the United States and the United Kingdom, Denny and Oppedisano (2013) find
positive effects of larger classes, significant in the United Kingdom.
 Ludger Woessmann 21

countries. This effect is only about half as large in developing countries. Further-
more, the effect of instruction time is larger in schools that have accountability
measures such as using achievement data for evaluation, as well as in schools that
have budgetary and personnel autonomy.
Rivkin and Schiman (2015) replicate the main finding of positive effects of instruc-
tion time in the within-student between-subject approach using the PISA 2009 data
and confirm it in a model that uses within-subject variation in instructional time across
grades within schools for identification. Furthermore, their results indicate that there
are diminishing returns to instruction time and its effect is larger in classrooms with
better environments as indicated by survey responses on questions about disruption,
bullying, attendance, and other indicators of the quality of classroom environments.
Positive effects of instruction time are also confirmed in the setting of a specific
education reform in Germany. The reform, which was implemented across German
states at different times in the 2000s, reduced the length of the academic-track high
school from nine to eight years. The reform did not change the curriculum require-
ments or the minimum required instruction time, so that the weekly instruction time
increased in each grade. Pooling the 9th-grade samples of the extended PISA test in
Germany from 2000 to 2009, Andrietti (2015) estimates the effects of the reform in a
differences-in-differences framework that exploits the differing implementation years
across states. Results suggest that an increase in weekly instruction time by one hour
in both 8th and 9th grade increases achievement in the different subjects by between
2 and 3 percent of a standard deviation. Results are also confirmed in a “triple-
difference” model that includes students in school types not affected by the reform as
an additional control group.
A couple of studies have also shown that additional instruction time is related
to smaller achievement gaps between different socioeconomic groups. Pooling
several waves of TIMSS and PISA data, Schneeweis (2011) finds that instruction
time is positively associated with the integration of immigrant students, with some
models including country fixed effects so that effects are effectively identified from
within-country changes over time. Pooling data from PISA and PIRLS for a differ-
ences-in-differences estimation with country fixed effects, Ammermueller (2013)
finds that the achievement difference between students with different numbers of
books at home is lower when instruction time is longer. There is also descriptive
evidence that enrollment in early childhood education—that is, additional time
before school—is related to reduced socioeconomic gradients and to better inte-
gration of migrant children (Schütz, Ursprung, and Woessmann 2008; Schneeweis
2011). Taken together, the results indicate that school instruction time can increase
educational opportunities for students from disadvantaged backgrounds.

Teacher Quality
Teacher quality can be measured in a variety of ways. For example, Hanushek,
Piopiunik, and Wiederhold (2014) use occupation-specific data on adult skills from
the Programme for the International Assessment of Adult Competencies (PIAAC)
to measure teacher skills in numeracy and literacy in 23 countries. Combining these
aggregate measures of teacher skills with student-level PISA data, they estimate the
22 Journal of Economic Perspectives

effect of teacher cognitive skills on international differences in student achievement,

controlling among other factors for PIAAC-based estimates of parents’ cognitive
skills. Models with student fixed effects that exploit within-country variation between
subjects suggest that teacher skills increase student achievement. Constructing a
pseudo-panel from the PIAAC data using teachers’ year of birth, they also exploit
cross-country differences in how alternative job opportunities for women over time
have attracted people with different skills into teaching. Bietenbeck, Piopiunik, and
Wiederhold (2015) apply a within-student between-subject approach to a regional
achievement test of 13 sub-Saharan African countries that includes subject-specific
tests of teachers. They find a significant positive effect of teacher subject knowl-
edge on student achievement that is complementary to access to subject-specific
textbooks.
Measuring teacher quality by both absolute teacher salary and teachers’ relative
salary position in a country’s income distribution, Dolton and Marcenaro-Gutierrez
(2011) find that higher teacher quality is related to better student achievement
using data from several TIMSS and PISA waves. The results are consistent with
positive effects of recruiting higher ability individuals into teaching. Results are
confirmed when adding country fixed effects, so that estimates are identified from
(relatively short-term) fluctuations in teacher pay within countries.
Apart from studies of direct measures of teacher quality, recent evidence also
indicates the relevance of teaching practices. Again applying within-student between-
subject identification to circumvent bias from unobserved student characteristics,
Schwerdt and Wuppermann (2011) show in the US TIMSS sample that for given
levels of teaching methods, traditional lecture-style teaching is related to better
student achievement compared to classroom problem-solving. Using the same esti-
mation strategy on TIMSS data for the United States and nine advanced countries,
Bietenbeck (2014) finds that traditional teaching practices are related to better
overall skills, factual knowledge, and solving of routine problems, whereas modern
teaching practices are related to better reasoning skills. After showing cross-country
correlations of teaching practices with measures of social capital, Algan, Cahuc, and
Shleifer (2013) apply a cross-sectional model with school fixed effects to TIMSS and
PIRLS data to show that progressive practices of having students work in groups are
positively related to student beliefs about cooperation and to student self-confidence.
Despite the result that resource inputs overall play a limited role, instruc-
tion time and certain dimensions of teacher quality do seem to matter for student
achievement. More broadly, these findings suggest ways in which what school systems
do are relevant for educational achievement. Moreover, looking into determinants
of instruction time and teacher quality leads naturally to questions about the institu-
tional framework of school systems that may frame how resources are used.

Institutional Structures of School Systems: Explorations into Causal

Effects

An international comparative approach promises to be fruitful in studying the

effects of educational institutions because institutional structures often do not vary
 The Importance of School Systems 23

nearly as much within countries as they do across countries. Specific institutional

features that have been found to matter for cross-country differences in student achieve-
ment include external exams, school autonomy, private competition, and tracking.

External Exams
In some countries, learning outcomes are assessed by curriculum-based
external exit exams that have real consequences for students (Bishop 1997). A large
literature has shown consistent positive associations between external exams and
student achievement (Hanushek and Woessmann 2011a). However, such cross-
country associations may be biased by unobserved country characteristics such as
specific cultures. For example, a society that favors high educational achievement
might both introduce external exams and also make efforts to induce students to
study, and a positive correlation between external exams and student achievement
does not show that the former has a causal effect on the latter.
There are several ways to explore whether these cultural effects are important
in explaining the connection from exit exams to test scores. One approach is to
look at variation in test scores and exams only within continents. If the international
variation in test scores would have been biased by features more relevant in some
continents than in others—say, if countries in Asia place a higher value on educa-
tional success than countries in other regions—then the coefficient on external
exams will decline in such a model. However, in Woessmann (2003a), I find that the
association between external exams and student achievement in the first two TIMSS
waves is robust to the inclusion of continental fixed effects. Another approach
looks at evidence across states within Germany and compares this with other OECD
countries. German states differ in whether they have external exams or not, but are
otherwise much more similar than OECD countries. However, in this mixture of
PISA data on German states and other countries, students in systems with external
exams have around 20 percent of a standard deviation higher achievement, and
this association is statistically indistinguishable between the OECD country sample
and the German state sample (Woessmann 2010). This result corroborates that
the international association is unlikely to be driven by fundamental differences in
culture, language, or other institutional settings that do not vary within Germany.
In yet another approach, Jürges, Schneider, and Büchel (2005) use the German
TIMSS 1995 data in a differences-in-differences approach that exploits variation
across subjects: specifically, in the relevant school tracks, most German states that
have external exams have them in math but not in science. The identifying assump-
tion of this model is that cross-state achievement differences would not differ
between subjects in the absence of the external exam treatment. While smaller
than their cross-sectional estimates, their differences-in-differences estimates are
significant and substantial at between 13 and 26 percent of a standard deviation.
If there are spillovers between subjects—for example, improved math knowledge
due to external exams also facilitates students’ learning in science—these estimates
provide a lower bound for the full effect of external exams. Until the early 2000s,
only seven of the 16 German states had external exams, but all but one have intro-
duced them over the course of the 2000s. Lüdemann (2011) exploits the different
24 Journal of Economic Perspectives

timing of the introduction of external exams across states and school types in a
differences-in-differences approach using the German extended PISA waves from
2000 to 2006. The identifying assumption is that there would have been common
trends in the absence of the external exam treatment. Results indicate significant
positive effects of the introduction of central exit exams even in the short run.
While external exams direct incentives particularly on students, a way to incentivize
teachers to focus on student outcomes is performance-related pay. Apart from showing
a positive association of teacher pay with student achievement in PISA data, in Woess-
mann (2011), I find that teacher salary adjustments for outstanding performance are
positively associated with student achievement across countries. The use of a country-
level measure of teacher performance pay avoids bias from within-country selection,
and results are robust to including continental fixed effects and to controlling for other
forms of teacher salary adjustments that are not based on performance. An advantage
of the cross-country approach is that it captures general-equilibrium effects such as
sorting into the teaching profession and other long-run incentive effects, whereas short-
term merit pay experiments capture only incentive effects, not selection effects.

School Autonomy
On the one hand, school autonomy may be conducive to student achievement
in school systems with strong surrounding structures that ensure high common
standards; on the other hand, school-based decision-making could hurt student
achievement in low-performing systems that lack basic standards and local capacity.
Cross-sectional evidence from international achievement tests concerning school
autonomy has been mixed (Hanushek and Woessmann 2011a), but these studies
may also be particularly plagued by identification issues.
To avoid bias from unobserved cross-country differences such as those arising
from culture and other government institutions, in Hanushek, Link, and Woessmann
(2013), we introduce the analytical approach of country panel analysis with country
fixed effects. Because many countries have reformed their school systems to become
more or less autonomous over time, we can exploit country-level variation over time
by including country fixed effects that control for systematic, time-invariant differences
across countries. While such panel analysis does not necessarily identify random varia-
tion, we show that prior achievement and prior GDP do not predict autonomy reforms.
To avoid bias from within-country selection of students into autonomous schools and
of schools to become autonomous, we aggregate the school autonomy measure to the
country level, reflecting the average share of autonomous schools in a country.
Pooling the individual data of over one million students in 42 countries in the four
PISA waves from 2000 to 2009, we find that school autonomy has a significant effect
on student achievement, but this effect varies systematically with the level of economic
and educational development: The effect is strongly positive in developed and high-
performing countries but strongly negative in developing and low-performing
countries. The estimates suggest that going from no to full autonomy over academic
content would increase student achievement by 53 percent of a standard deviation in
the highest-income country (Norway) and reduce student achievement by 55 percent
of a standard deviation in the lowest-income country (Indonesia).
 Ludger Woessmann 25

If part of the negative effect of school autonomy stems from a lack of account-
ability, these negative aspects should be eased in school systems where external
exams provide comparative information on ultimate performance. Indeed, in
Hanushek, Link, and Woessmann (2013), we find a significant positive interaction
between changes in school autonomy and (initial) external exit exams—that is,
introducing autonomy is more beneficial in school systems that have accountability
through external exams.
The effects of school autonomy may also be interrelated with the management
capacity of schools. Collecting data on school management practices in operations,
monitoring, target setting, and people management in eight countries, Bloom,
Lemos, Sadun, and Van Reenan (2015) find higher management quality to be
related to better student achievement. While mostly focusing on specific national
achievement datasets, they also report a positive correlation with average PISA
scores across Italian and German regions. Furthermore, autonomous public schools
score highly in terms of management quality. Interestingly, while their previous
work suggested that most of the variation in management quality in other sectors
is within-country, about half of the variance in management quality in the school
sector is between countries, underlining the importance of cross-country analysis of
institutional environments in school systems.

Private Competition
The extent to which schools are operated by public or private entities differs
markedly across countries. For example, more than three-quarters of 15 year-old
students in the Netherlands attend privately operated schools and more than
60 percent in Belgium and Ireland, but this share is below 10 percent in many other
countries. Private school operation is largely independent of the funding of schools;
for example, the average share of government funding of Dutch privately operated
schools is the same (at 95 percent) as in public schools, a feature going back to
constitutional provisions. Private school operation may be related to the extent of
school autonomy, but again these are conceptually different issues: public schools
can have substantial autonomy, and private schools can have limited autonomy.
A key point is that competition from private alternatives may improve the perfor-
mance of public schools as well, which may lift the achievement level systemwide.
Cross-country evidence indeed suggests a strong association of achievement
levels with the share of privately operated schools (for example, Woessmann 2009),
but identification issues are again obvious in cross-country analyses: low quality of
the public school system may induce a political system to encourage private alterna-
tives or parents to choose private alternatives, and other country features related to
the supply of or demand for private schools may introduce omitted variable bias.
To identify exogenous variation in the share of private schools across countries,
in West and Woessmann (2010), my coauthor and I argue that historical differences
in Catholic versus Protestant denominations provide a natural experiment. In late
19th century, Catholic doctrine resisted the emerging nondenominational public
school systems and spurred efforts to establish private schools in many countries.
These efforts were most successful in countries with substantial shares of Catholic
26 Journal of Economic Perspectives

populations but without a Catholic state religion. Therefore, the share of Catho-
lics in a country’s population in 1900 (interacted with an indicator as to whether
Catholicism was the state religion) can be used as an instrumental variable for the
share of privately operated schools in the 2003 PISA data.10 The results suggest
that a 10 percentage point increase in private school shares, induced by historical
Catholic resistance to state schooling, leads to an increase in math achievement by
at least 9 percent of a standard deviation. Much of this effect accrues to students in
public schools, suggesting that most of the overall effect reflects benefits of private
competition and parental choice. In addition to increasing achievement, private
competition is also estimated to reduce total educational expenditure per student.

Tracking
Countries vary in the extent to which students are tracked into different school
types by ability. No country has differing-ability schools in the early grades of primary
school. Some countries such as Austria and Germany track students into different-
ability schools as early as age 10. Many other countries have a comprehensive school
system (although perhaps with some streaming within schools) through the end of
high school. A common concern is that early tracking may increase inequality as lower-
achieving groups are tracked into lower-ability schools, perhaps because of peer effects.
In Hanushek and Woessmann (2006), we suggest an identification strategy that
compares achievement changes from primary to later schooling across tracked and
untracked countries. Using country-level data for several pairs of PIRLS, TIMSS, and
PISA achievement tests administered at the primary and secondary school levels in
the context of a differences-in-differences model, we find that early tracking signifi-
cantly increases the inequality in countries’ achievement outcomes. We do not find
a consistent effect of early tracking on the level of achievement, although most
­estimates tend to be negative. Interestingly, simple cross-sectional estimations do
not indicate an association of tracking with educational inequality.
A variety of other results suggest that earlier tracking tends to raise the inequality
of educational outcomes. Applying the same estimation strategy across grades to
student-level PIRLS and PISA data, Ammermueller (2013) finds that early tracking
and the number of tracked school types increase the effect of parental education on
student achievement. Again using the same identification strategy to estimate the
effect of tracking on the migrant–native achievement gap in a pooled micro dataset
of all PIRLS, TIMSS, and PISA waves from 1995 to 2012, Ruhose and Schwerdt
(2016) do not find that early tracking affects native and migrant students differently
in general. However, they find a detrimental effect of early tracking on the rela-
tive achievement of first-generation migrants and the presumably less-integrated
subgroup of second-generation migrant students who do not speak the host-country
language at home. Piopiunik (2014) exploits a school reform in Bavaria that

10
There is ample evidence that historically, Catholics have placed less emphasis on education than
­ rotestants (for example, Becker and Woessmann 2009), which would bias the instrumental-variable
P
model against finding beneficial effects of competition. Indeed, the current share of Catholics enters
negatively in the second-stage model.
 The Importance of School Systems 27

lowered the age of tracking between the two lowest-ability school types to estimate a
“triple-difference” model using variation across three German PISA waves that allow
a comparison of outcomes in the reformed system to pre-reform outcomes, to other
German states, and to the non-treated highest-ability school type. Results suggest
that earlier tracking reduced achievement in both low- and middle-track schools.

Conclusions

What explains the large international differences in student achievement? On

a descriptive basis, a simple model of three combined factors of family background,
school resources, and institutions is able to account for more than four-fifths of the
total cross-country variation in student achievement. Family background and insti-
tutions contribute roughly equally to this exercise, whereas the contribution of
school resources is quite limited—although the predictive power of the model varies
across countries. Beyond these descriptive patterns, a growing literature uses quasi-
experimental methods in an attempt to identify causal effects of school systems in
the international test data, as well as different types of fixed effects models that aim
to avoid certain sources of bias. Some patterns emerge from this literature. First, this
work tends to confirm that resource inputs such as expenditure per student or class size
appear to have limited effects on student achievement. Second, instruction time and
measures of teacher quality do play a role. Third, a number of institutional features of
school systems seem to contribute to the cross-country differences in student achieve-
ment. For example, external exit exams and competition from privately operated
schools positively affect achievement levels. School autonomy has positive effects in
developed countries and where external exit exams introduce accountability, but
negative effects in developing countries. Early tracking into differing-ability schools
seems to increase inequality in achievement without increasing achievement levels.
Clearly, the exploitation of the potential of international differences in student
achievement to improve our understanding of educational processes is work
in progress. In the future, increasing numbers of participating countries and an
expanding number of waves of available international achievement tests will raise
the scope of possible investigations. A useful direction for international testing
efforts would be to conduct studies in many countries that are longitudinal at the
student level. Existing causal identification strategies will be sharpened and new
approaches developed. It may be especially useful to focus on interactions between
the kinds of factors examined here: for example, little is known about the particular
institutional settings that may strengthen the effectiveness of resource use.
As this work proceeds, it is perhaps useful to remember what is at stake. Levels
and changes in educational achievement are a powerful determinant of output
levels and economic growth. It has long been common to use average years of
schooling in regressions that seek to explain economic growth. But average years
of schooling may be a very noisy measure of actual educational achievement as
measured by test scores. Thus, in Hanushek and Woessmann (2012, 2015a), we
show that a model that includes only countries’ average years of schooling and
28 Journal of Economic Perspectives

their initial level of GDP per capita as predictors accounts for one-quarter of the
total cross-country variation in growth rates in GDP per capita from 1960 to 2000
(or 2009). However, adding average scores on the international achievement tests
between 1964 and 2003 to the model accounts for more than three-quarters of
the variation in long-term growth rates of per-capita GDP—indeed, it renders the
commonly used quantitative measure of years of schooling insignificant. Differences
in math and science achievement can fully account both for the stunning growth
performance of the East Asian miracle countries and for the disheartening
growth performance of Latin American countries (Hanushek and Woessmann 2016).
In Hanushek and Woessmann (2012, 2015a), we report several econometric
analyses that provide a prima facie case that the close and robust association of educa-
tional achievement with countries’ long-run economic growth reflects a causal effect
of population skills. To preclude simple reverse causation, we show that achievement
tests before 1985 predict subsequent growth. To address potential bias from omitted
factors such as differing economic institutions or cultures, we present instrumental-
variable models that use only part of the skill variation that can be predicted from
institutional differences in school systems; show that changes in test scores predict
changes in growth; perform development accounting analyses that take param-
eter values from the micro literature; and report differences-in-differences models
showing that immigrants educated in their home countries receive returns to their
home-country cognitive skills on the US labor market, whereas immigrants from the
same home countries schooled in the United States do not. Additional recent work
on student achievement and international income differences is given in Kaarsen
(2014), and we provide reviews in Hanushek and Woessmann (2008, 2011a). Within
the United States, in Hanushek, Ruhose, and Woessmann (2015), we confirm an
important role for educational achievement in explaining differences in GDP per
capita across US states. At the individual level, performance on adult achievement
tests is strongly associated with employment and earnings in each of the 23 countries
analyzed in Hanushek, Schwerdt, Wiederhold, and Woessmann (2015). Murnane,
Willett, and Levy (1995) and Chetty et al. (2011), among others, provide additional
evidence on individual returns to educational achievement in the United States.
Of course, the implications of improved educational achievement go well beyond
individual earnings and macroeconomic growth rates. Education is important for
economic inequality and the transmission of inequality across generations (for example,
Black and Devereux 2011). Education affects the education and health of children, own
health, crime, and citizenship (for example, Lochner 2011). More broadly, a “capa-
bilities approach” to welfare analysis in the style of Sen and Nussbaum (for example,
Nussbaum and Sen 1993) emphasizes that education is an important determinant of
the ability of people to develop their own capacities and in that sense to be able to exer-
cise autonomy and choice in all aspects of life.

■ Helpful comments from Mark Gertler, Gordon Hanson, Eric Hanushek, Enrico Moretti,
Marc Piopiunik, Jens Ruhose, Timothy Taylor, Martin West, and Simon Wiederhold are
gratefully acknowledged.
 Ludger Woessmann 29

References

Algan, Yann, Pierre Cahuc, and Andrei Shleifer. Brunello, Giorgio, and Daniele Checchi. 2007.
2013. “Teaching Practices and Social Capital.” “Does School Tracking Affect Equality of Opportu-
American Economic Journal: Applied Economics 5(3): nity? New International Evidence.” Economic Policy
189–210. 22(52): 781–861.
Altinok, Nadir, and Geeta Kingdon. 2012. “New Brunello, Giorgio, and Lorenzo Rocco. 2013.
Evidence on Class Size Effects: A Pupil Fixed “The Effect of Immigration on the School Perfor-
Effects Approach.” Oxford Bulletin of Economics and mance of Natives: Cross Country Evidence Using
Statistics 74(2): 203–34. PISA Test Scores.” Economics of Education Review
Ammermueller, Andreas. 2013. “Institutional 32(1): 234–46.
Features of Schooling Systems and Educational Chetty, Raj, John N. Friedman, Nathaniel
Inequality: Cross-country Evidence from PIRLS Hilger, Emmanuel Saez, Diane Whitmore Schan-
and PISA.” German Economic Review 14(2): zenbach, and Danny Yagan. 2011. “How Does Your
190–213. Kindergarten Classroom Affect Your Earnings?
Ammermueller, Andreas, and Jörn-Steffen Evidence from Project STAR.” Quarterly Journal of
Pischke. 2009. “Peer Effects in European Primary Economics 126(4): 1593–1660.
Schools: Evidence from the Progress in Interna- Cobb-Clark, Deborah A., Mathias Sinning, and
tional Reading Literacy Study.” Journal of Labor Steven Stillman. 2012. “Migrant Youths’ Educa-
Economics 27(3): 315–48. tional Achievement: The Role of Institutions.”
Andrietti, Vincenzo. 2015. “The Causal Effects Annals of the American Academy of Political and Social
of Increased Learning Intensity on Student Science 643(1): 18–45.
Achievement: Evidence from a Natural Experi- Denny, Kevin, and Veruska Oppedisano. 2013.
ment.” Universidad Carlos III de Madrid, Working “The Surprising Effect of Larger Class Sizes:
Paper, Economic Series 15-06. Madrid: Universidad Evidence Using Two Identification Strategies.”
Carlos III. Labour Economics 23: 57–65.
Angrist, Joshua D., and Victor Lavy. 1999. Dolton, Peter, and Oscar D. Marcenaro-
“Using Maimonides’ Rule to Estimate the Effect Gutierrez. 2011. “If You Pay Peanuts Do You Get
of Class Size on Scholastic Achievement.” Quarterly Monkeys? A Cross-country Analysis of Teacher Pay
Journal of Economics 114(2): 533–75. and Pupil Performance.” Economic Policy 26(65):
Becker, Sascha O., and Ludger Woessmann. 5–55.
2009. “Was Weber Wrong? A Human Capital Dustmann, Christian, Tommaso Frattini, and
Theory of Protestant Economic History.” Quarterly Gianandrea Lanzara. 2012. “Educational Achieve-
Journal of Economics 124(2): 531–96. ment of Second-Generation Immigrants: An
Bedard, Kelly, and Elizabeth Dhuey. 2006. International Comparison.” Economic Policy 27(69):
“The Persistence of Early Childhood Maturity: 143–85.
International Evidence of Long-Run Age Effects.” Edwards, Sebastian, and Alvaro Garcia Marin.
Quarterly Journal of Economics 121(4): 1437–72. 2015. “Constitutional Rights and Education:
Bietenbeck, Jan. 2014. “Teaching Practices and An International Comparative Study.” Journal of
Cognitive Skills.” Labour Economics 30: 143–53. Comparative Economics 43(4): 938–55.
Bietenbeck, Jan, Marc Piopiunik, and Simon Foshay, Arthur W. 1962. “The Background
Wiederhold. 2015. “Africa’s Skill Tragedy: Does and the Procedures of the Twelve-Country Study.”
Teachers’ Lack of Knowledge Lead to Low Student In Educational Achievement of Thirteen-year-olds in
Performance?” CESifo Working Paper 5470. Twelve Countries: Results of an International Research
Bishop, John H. 1997. “The Effect of National Project, 1959–61, edited by Arthur W. Foshay,
Standards and Curriculum-based Examinations Robert L. Thorndike, Fernand Hotyat, Douglas A.
on Achievement.” American Economic Review 87(2): Pidgeon, and David A. Walker. Hamburg: Unesco
260–64. Institute for Education.
Black, Sandra E., and Paul J. Devereux. 2011. Freeman, Richard B., and Martina Viarengo.
“Recent Developments in Intergenerational 2014. “School and Family Effects on Educational
Mobility.” In Handbook of Labor Economics, Vol. Outcomes across Countries.” Economic Policy
4B, edited by Orley Ashenfelter and David Card, 29(79): 395–446.
1487–1541. Amsterdam: North Holland. Fryer, Roland G., Jr., and Steven D. Levitt.
Bloom, Nicholas, Renata Lemos, Raffaella 2010. “An Empirical Analysis of the Gender Gap in
Sadun, and John Van Reenen. 2015. “Does Mathematics.” American Economic Journal: Applied
Management Matter in Schools?” Economic Journal Economics 2(2): 210–240.
125(584): 647–74. Fuchs, Thomas, and Ludger Woessmann. 2007.
30 Journal of Economic Perspectives

“What Accounts for International Differences in 2011a. “The Economics of International Differ-
Student Performance? A Re-examination Using ences in Educational Achievement.” In Handbook
PISA Data.” Empirical Economics 32(2–3): 433–64. of the Economics of Education, Vol. 3, edited by Eric
Guiso, Luigi, Ferdinando Monte, Paola Sapi- A. Hanushek, Stephen Machin, and Ludger Woess-
enza, and Luigi Zingales. 2008. “Culture, Math, mann, 89–200. Amsterdam: North Holland.
and Gender.” Science 320(5880): 1164–65. Hanushek, Eric A., and Ludger Woessmann.
Gundlach, Erich, Ludger Woessmann, and 2011b. “How Much Do Educational Outcomes
Jens Gmelin. 2001. “The Decline of Schooling Matter in OECD Countries?” Economic Policy
Productivity in OECD Countries.” Economic Journal 26(67): 427–91.
111(471): C135–C47. Hanushek, Eric A., and Ludger Woessmann.
Hanushek, Eric A. 1986. “The Economics of 2011c. “Sample Selectivity and the Validity of Inter-
Schooling: Production and Efficiency in Public national Student Achievement Tests in Economic
Schools.” Journal of Economic Literature 24(3): Research.” Economics Letters 110(2): 79–82.
1141–77. Hanushek, Eric A., and Ludger Woessmann.
Hanushek, Eric A. 2002. “Publicly Provided 2012. “Do Better Schools Lead to More Growth?
Education.” In Handbook of Public Economics, Vol. 4, Cognitive Skills, Economic Outcomes, and
edited by Alan J. Auerbach and Martin Feldstein, ­Causation.” Journal of Economic Growth 17(4):
2045–2141. Amsterdam: North Holland. 267–321.
Hanushek, Eric A., and Dennis D. Kimko. 2000. Hanushek, Eric A., and Ludger Woessmann.
“Schooling, Labor Force Quality, and the Growth 2015a. The Knowledge Capital of Nations: Education
of Nations.” American Economic Review 90(5): and the Economics of Growth. Cambridge, MA: MIT
1184–1208. Press.
Hanushek, Eric A., Susanne Link, and Ludger Hanushek, Eric A., and Ludger Woessmann.
Woessmann. 2013. “Does School Autonomy Make 2015b. Universal Basic Skills: What Countries
Sense Everywhere? Panel Estimates from PISA.” Stand to Gain. Paris: Organisation for Economic
Journal of Development Economics 104: 212–32. Co-operation and Development.
Hanushek, Eric A., Paul E. Peterson, and Hanushek, Eric A., and Ludger Woessmann.
Ludger Woessmann. 2013. Endangering Prosperity: 2016. “Knowledge Capital, Growth, and the East
A Global View of the American School. Washington, Asian Miracle.” Science 351(6271): 344–45.
DC: Brookings Institution Press. Hoxby, Caroline M. 2000. “The Effects of Class
Hanushek, Eric A., Paul E. Peterson, and Size on Student Achievement: New Evidence
Ludger Woessmann. 2014. “U.S. Students from from Population Variation.” Quarterly Journal of
Educated Families Lag in International Tests.” Economics 115(3): 1239–85.
Education Next 14(4): 8–18. IEA. 2016. “Brief History of IEA: 55 Years of
Hanushek, Eric A., Marc Piopiunik, and Simon Educational Research.” Amsterdam: International
Wiederhold. 2014. “The Value of Smarter Teachers: Association for the Evaluation of Educational
International Evidence on Teacher Cognitive Skills Achievement. http://www.iea.nl/brief_history.html.
and Student Performance.” NBER Working Paper Jerrim, John, and John Micklewright. 2014.
20727. “Socio-economic Gradients in Children’s Cogni-
Hanushek, Eric A., Jens Ruhose, and Ludger tive Skills: Are Cross-country Comparisons Robust
Woessmann. 2015. “Human Capital Quality and to Who Reports Family Background?” European
Aggregate Income Differences: Development Sociological Review 30(6): 766–81.
Accounting for U.S. States.” NBER Working Paper Jürges, Hendrik, Kerstin Schneider, and Felix
21295. Büchel. 2005. “The Effect of Central Exit Examina-
Hanushek, Eric A., Guido Schwerdt, Simon tions on Student Achievement: Quasi-experimental
Wiederhold, and Ludger Woessmann. 2015. Evidence from TIMSS Germany.” Journal of the
“Returns to Skills around the World: Evidence from European Economic Association 3(5): 1134–55.
PIAAC.” European Economic Review 73: 103–130. Kaarsen, Nicolai. 2014. “Cross-Country Differ-
Hanushek, Eric A., and Ludger Woessmann. ences in the Quality of Schooling.” Journal of
2006. “Does Educational Tracking Affect Perfor- Development Economics 107: 215–24.
mance and Inequality? Differences-in-differences Lavy, Victor. 2015. “Do Differences in Schools’
Evidence across Countries.” Economic Journal Instruction Time Explain International Achieve-
116(510): C63–C76. ment Gaps? Evidence from Developed and
Hanushek, Eric A., and Ludger Woessmann. Developing Countries.” Economic Journal 125(588):
2008. “The Role of Cognitive Skills in Economic F397–F424.
Development.” Journal of Economic Literature 46(3): Lee, Jong-Wha, and Robert J. Barro. 2001.
607–68. “Schooling Quality in a Cross-section of Countries.”
Hanushek, Eric A., and Ludger Woessmann. Economica 68(272): 465–88.
 The Importance of School Systems 31

Lochner, Lance. 2011. “Nonproduction Benefits Singh, Abhijeet. 2015. “Learning More with
of Education: Crime, Health, and Good Citizen- Every Year: School Year Productivity and Inter-
ship.” In Handbook of the Economics of Education, Vol. national Learning Divergence.” Presented at the
4, edited by Eric A. Hanushek, Stephen Machin, CESifo Area Conference on the Economics of
and Ludger Woessmann, 183–282. Amsterdam: Education, September 11–12, 2015. Available
North Holland. at: http://www.cesifo-group.de/de/ifoHome/
Lüdemann, Elke. 2011. “Intended and Unin- events/Archive/conferences/2015/09/2015-09-
tended Short-run Effects of the Introduction of 11-ee15-Hanushek/Programme.html.
Central Exit Exams: Evidence from Germany.” West, Martin R., and Ludger Woessmann.
In Elke Lüdemann, Schooling and the Formation of 2010. “   ‘Every Catholic Child in a Catholic
Cognitive and Non-cognitive Outcomes. ifo Beiträge School’: Historical Resistance to State Schooling,
zur Wirtschaftsforschung 39. Munich: ifo Institut. Contemporary Private Competition and Student
Mullis, Ina V. S., Michael O. Martin, Pierre Foy, Achievement across Countries.” Economic Journal
and Alka Arora. 2012. TIMSS 2011 International 120 (546): F229–F255.
Results in Mathematics. Chestnut Hill, MA: TIMSS & Woessmann, Ludger. 2003a. “Central Exit
PIRLS International Study Center, Boston College. Exams and Student Achievement: International
Murnane, Richard J., John B. Willett, and Frank Evidence.” In No Child Left Behind? The Politics
Levy. 1995. “The Growing Importance of Cognitive and Practice of School Accountability, edited by Paul
Skills in Wage Determination.” Review of Economics E. Peterson and Martin R. West, 292–323. Wash-
and Statistics 77(2): 251–66. ington, D.C.: Brookings Institution Press.
Nussbaum, Martha C., and Amartya Sen, eds. Woessmann, Ludger. 2003b. “Schooling
1993. The Quality of Life. Oxford University Press. Resources, Educational Institutions, and Student
OECD. 2013. PISA 2012 Results: What Students Performance: The International Evidence.” Oxford
Know and Can Do—Student Performance in Bulletin of Economics and Statistics 65(2): 117–70.
Mathematics, Reading and Science, Vol 1. Paris: Woessmann, Ludger. 2005a. “Educational
Organisation for Economic Co-operation and Production in Europe.” Economic Policy 20(43):
Development. 446–504.
Piopiunik, Marc. 2014. “The Effects of Early Woessmann, Ludger. 2005b. “The Effect
Tracking on Student Performance: Evidence from Heterogeneity of Central Exams: Evidence from
a School Reform in Bavaria.” Economics of Education TIMSS, TIMSS-Repeat and PISA.” Education
Review 42: 12–33. Economics 13(2): 143–169.
Ripley, Amanda. 2013. The Smartest Kids in the Woessmann, Ludger. 2009. “Public–Private
World—And How They Got That Way. New York: Partnerships and Student Achievement: A Cross-
Simon & Schuster. Country Analysis.” In School Choice International:
Rivkin, Steven G., and Jeffrey C. Schiman. Exploring Public–Private Partnerships, edited by
2015. “Instruction Time, Classroom Quality, Rajashri Chakrabarti and Paul E. Peterson, 13–45.
and Academic Achievement.” Economic Journal Cambridge, MA: MIT Press.
125(588): F425–F448. Woessmann, Ludger. 2010. “Institutional Deter-
Ruhose, Jens, and Guido Schwerdt. 2016. minants of School Efficiency and Equity: German
“Does Early Educational Tracking Increase States as a Microcosm for OECD Countries.”
Migrant–Native Achievement Gaps? Differences-in- Journal of Economics and Statistics 230(2): 234–70.
Differences Evidence across Countries.” Economics Woessmann, Ludger. 2011. “Cross-Country
of Education Review 52: 134–54. Evidence on Teacher Performance Pay.” Economics
Schneeweis, Nicole. 2011. “Educational Institu- of Education Review 30(3): 404–18.
tions and the Integration of Migrants.” Journal of Woessmann, Ludger, Elke Luedemann,
Population Economics 24(4): 1281–1308. Gabriela Schuetz, and Martin R. West. 2009. School
Schütz, Gabriela, Heinrich W. Ursprung, and Accountability, Autonomy, and Choice around the
Ludger Woessmann. 2008. “Education Policy and World. Cheltenham, UK: Edward Elgar.
Equality of Opportunity.” Kyklos 61(2): 279–308. Woessmann, Ludger, and Martin R. West.
Schwerdt, Guido, and Amelie C. Wuppermann. 2006. “Class-Size Effects in School Systems
2011. “Is Traditional Teaching Really All That Bad? around the World: Evidence from Between-
A Within-Student Between-Subject Approach.” Grade Variation in TIMSS.” European Economic
Economics of Education Review 30(2): 365–79. Review 50(3): 695–736.
32 Journal of Economic Perspectives