ARTICLE IN PRESS

International Journal of
Educational Research 45 (2006) 202–215
www.elsevier.com/locate/ijedures

Quasi-experimental estimates of class size effect

in primary schools in Poland
Maciej Jakubowski, Pawel Sakowski
Faculty of Economics, Warsaw University, ul. D!uga 44/50, 00-241, Warsaw, Poland

Abstract

In this paper we analyze class size effects in the case of primary schools in Poland. We use two
empirical strategies to avoid endogeneity bias. First, we use average class size in a grade as an
instrumental variable for actual class size. This allows us to control for within school selection of
pupils with different abilities to classes of different sizes. Additionally, we estimate fixed effects for
schools to control for differences between them. Second, we exploit the fact that there is an informal
maximum class size rule. We estimate class size effect only for those enrollment levels where some
schools decide to add a new class and thus dramatically lower class sizes. For such enrollment levels
variance of class size is mainly exogenous and we argue that this allows estimation of quasi-
experimental class size effects. In this case we again use average class size as an instrument with
enrollment as a key control variable. Using both strategies we obtain similar findings. We found that
the positive effects observed with OLS regression disappear when we use instrumental variables. If we
avoid endogeneity bias, then class size negatively affects student achievement. However, this effect is
rather small. We discuss methodology, possible bias of results and the importance of our findings to
current policy issues in Poland.
r 2006 Elsevier Ltd. All rights reserved.

Keywords: Class size; Educational achievement; Student sorting; School fixed effects; Instrumental variables;
Regression discontinuity design

Corresponding author. Fax: +48 22 8312846.

E-mail addresses: [email protected] (M. Jakubowski), [email protected] (P. Sakowski).

0883-0355/$ - see front matter r 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.ijer.2006.11.003
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 203

1. Introduction

Class size reduction is one of the most heavily discussed and controversial issues
in educational debates all over the world. It is a popular policy proposal probably because
of its simplicity. The common intuition is that in smaller classes teachers can devote
more time to each pupil and it is easier for them to maintain order. Thus, intuitively,
reduction in the average class size should improve student achievement. This makes class
size reduction attractive to parents and policy makers. Additionally, teachers may prefer
smaller classes because of a lower workload and a higher demand for their work. In
countries like the United States or France the average class size is now much smaller than
30 yr ago. Some people argue that this expensive policy has no effect on student
achievement, but others still regard class size reduction as cost-effective (Hanushek, 2003;
Krueger, 2003).
However, in many countries class size reduction is not an issue. This is because of scarce
resources or a greater focus on more politically relevant matters. In fact, one can easily
imagine a situation where the average class size is growing because of budgetary cuts. It is
also possible that governments prefer to invest in other school resources like teacher
training, computers etc. Governments can decide to save funds for these investments by
raising the pupils per teacher ratio. It is also possible that reforms of the educational
systems can affect average class size without explicit intention to do so. The problem here
is that when we do not know the effect of a change in class size on teaching quality, then
the overall result of reform is hard to predict.
In Poland, class size reduction is not a widely recognized policy issue. Other problems
and questions are more heavily discussed, especially the financial problems of local
governments and their growing independence. The Polish educational system was
decentralized in the 1990s and increasingly local governments more independently use
their power to organize local school systems. Local authorities share the financial burden
of the provision of primary and secondary education with the central government. In some
places subsidies from the central budget are sufficient, but in others they are not. Many
‘‘gmina’’ (the lowest level of local government, which is responsible for primary education)
spend as much as 50% of their budget on education and try to cut growing costs without
any proper investigation or knowledge of how these actions influence the quality of
teaching. In rural areas many local authorities close smaller schools to save money for
other expenses.
In the beginning of the reform period in the 1990s, the Ministry of Education was
supposed to establish so-called ‘‘educational standards’’ which were expected to regulate
issues like maximum or minimum class size and the pupils per teachers ratio. However,
such regulations have never been introduced. Additionally, in Poland there is no research
that can be used as a reference point for decision makers. Discussion about educational
standards is heavily influenced by the political battle between teachers’ trade unions and
representatives of local governments. Only a few people concentrate on the quality of
education rather than on the financial controversy alone. The aim of this study is to
analyze the effect of class size to support public discussion with some data. This seems to
be important because in the near future average class size could grow because of the
financial problems of local governments.
The causal effect of class size on pupil achievement is empirically difficult to measure.
Actual class size considerably varies between regions, cities, rural and urban areas.
 ARTICLE IN PRESS
204 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

Different schools with different characteristics have different class sizes. There are also
visible differences in class sizes within schools. These differences are very often correlated
with important characteristics of pupils, parents, teachers, schools and local communities
that heavily influence student achievement. In the case of free choice, parents who care
more about their children’s education will send them to better schools. One can also expect
better schools to have bigger classes because of higher demand. Different class sizes within
schools are sometimes correlated with student ability. For example, school principals may
decide to place insubordinate pupils in smaller classes or they may place pupils whose
parents financially support schools in smaller classes with additional teaching equipment.
Local governments can additionally support chosen schools or establish rules for
maximum or minimum class sizes and so one can expect to find smaller classes in richer
communities. All these examples suggest that the actual class size is a result of many
independent and unobserved choices and that it is almost impossible to control for all
variables that could be correlated with student achievement and class size (for a thorough
discussion of potential bias see Hoxby (2000)).
In other words, the problem of endogeneity of class size casts doubt on using simple
OLS regression to estimate class size effect. That is why some scholars dispute earlier
research that claims to have shown no relationship between school resources and student
achievement (see discussion between Hanushek (2003) and Krueger (2003)). They believe
that one needs sophisticated methodology or experimental data to find unbiased estimates
of class size effects.
One way to avoid problems caused by class size endogeneity is to use a proper
instrumental variable. In the case of class size many scholars used as an instrument the
grade-average class size in a school. This allow them to focus on the class size variation
that is not caused by within-school selection of students with different abilities among
classes of different size (Akerhielm, 1995). However, average class size is often correlated
with other variables that could affect student achievement. For example, average class size
can be higher in more popular schools or in cities (important in the Polish case). This
makes estimates based on this instrument suspicious even when one uses a full set of
control variables. Having this in mind, scholars use other techniques to allow for bias
caused by between-school sorting.
One way to deal with this issue is to estimate fixed effects of schools or communities.
This should give less-biased estimates but one needs an appropriate data set (for example
Hoxby (2000) uses longitudinal data on enrollment and achievement and WöXmann
(2003) uses scores from different grades). Another way to deal with a between-school
bias is to find more credible sources of exogenous variation in the class size. Caroline
Hoxby exploits random changes in cohort sizes to estimate class size effects (Hoxby,
2000). She also exploits the fact that random changes in enrollment cause dramatic
change in a class size when a new class is added. Joshua Angrist and Victor Lavy use
this idea to construct an instrument based on the maximum class size rule for the
Israeli case and to obtain quasi-experimental estimates of class size effect (Angrist & Lavy,
1999).
We adopt similar ideas to examine class size effect in primary schools in Poland. We use
2002–2004 test scores from the obligatory primary school leavers’ exam. We focus on two
empirical strategies which we believe give unbiased results. The first strategy uses schools’
grade-average class size as an instrument for an actual class size. It allows us to avoid bias
caused by the within-school segregation of pupils with different abilities to classes of
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 205

different sizes. As we have a 3-yr data set, we are able to control for schools’ fixed effects.
Basically, we look at how change in the average class size affects achievement in
each school between years. In the second strategy, we use the fact that average class size
changes abruptly when the higher enrollment level forces school principals and
local authorities to increase the number of classes in a school. The idea is that the
change in the number of classes depends mainly on the cohort size and is believed to be
exogenous. Thus, we estimate class size effect only for those enrollment levels where
the variance in class size can, to a large extent, be attributed to the change in the number
of classes. We use two instruments in this case. One is identical to that developed
by Angrist and Lavy. The second one is again the average class size. We discuss the
difference between these two instruments in this case. We argue that exploring different
methods and sources of exogenous variation give us a valuable cross checking on empirical
approaches.
The rest of the paper is organized as follows. In Section 2 we depict the educational
system and discuss former research on similar issues in Poland. In Sections 3 and 4 the data
set and methodology are described. In Section 5 results and simple application are
presented. Section 6 summarizes our work.

2. Background of the study

There are two levels of compulsory education in Poland: 6-yr primary school for 7–12 yr
of age and gymnasium (3-yr lower secondary school) for 13–15 yr of age. In this work we
focus on public primary schools. These schools are free of charge and the only admission
criterion is the age limit. Parents have a right to register their children in the nearest school
but they can also try to find a place in another public or private institution. However,
neither of these types of institutions are obliged to offer admission and private schools will
usually charge tuition. For most Polish families private schools are too expensive and only
a small proportion of pupils use them (less than 2% in 2004).
During the 1990s, the Polish educational system was decentralized. Local governments
receive from the central government special grants calculated on a per pupil basis. These
monies can be spent on schools, but also on other expenditures. Some local authorities
spend additional money on schools from their own revenues while others are able to lower
costs and to spend resources on other expenditures. However, 80% of school costs are
attributed to teachers’ salaries which are defined centrally. The easiest way to lower cost is
to close smaller schools and to make bigger classes. Thus, local authorities have a great
financial incentive to increase average class size and some of them do their best to do this.
Between 1990 and 2004, 5189 primary schools were closed. This is partly due to the
decline in the overall number of pupils and the introduction of 6 yr primary schools and
3 yr gymnasiums (instead of 8 yr primary) but also to the ‘rationalization’ of the school
network, which was promoted by the financial incentives described above. The average
class size in newly established gymnasiums is much higher than in primary schools
(25 comparing to 21) and these schools are bigger (on average 242 pupils, compared to
186 pupils; data from 2004). In cities these numbers are higher and there are regions where
the average class size is around 30. Some scholars believe that this is the proper way to
save money for investment in more important resources (Levitas, Golinowska, &
Herczynski, 2001). However, they do not analyze quantitatively how resources are related
to teaching quality.
 ARTICLE IN PRESS
206 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

The fact is that while average class size in primary schools in Poland remains quite
stable, during the last years the discrepancies between local governments have been
substantial. While there are about 200 local governments out of 2622 with an average class
size equal to or lower than 15, there is a similar amount with more than 30 students in class
on average. Differences at the school level are even bigger.
Despite political discussion about teaching quality in Poland there are only a few papers
that analyze the effects of school resources using quantitative methods. A few papers have
touched the problem of class size effects, however none deal with the endogeneity issues.
They use simple correlation or regression analysis to find positive or statistically
insignificant relations between class size and achievement (Bialecki & Haman, 2003;
Herczynski & Herbst, 2002; Sleszynski, 2002). While positive effects can be explained by
peer effects (see Dobbelsteen et al., 2002) we do not trust results from these studies because
of their poor methodology.
In our earlier work we analyzed data sets with much more detailed information on
schools than these that were available in the Polish research mentioned above (Jakubowski
& Sakowski, 2004). To avoid endogeneity we also used average class size as an
instrumental variable. We observed that the initial positive effect of class size became
smaller when we used the instrumental variable. However, we were sure that in the Polish
case this instrument alone is not sufficient because the average class size is also correlated
with important characteristics of local governments, communities and schools. We were
deeply unsatisfied with these results and believed that our empirical strategy and data does
not allow us to estimate unbiased class size effect. This is the reason why we conducted the
study described below.

3. Data

In our study we use data on student achievement on standardized exams conducted in

Poland at the end of primary school. Starting from 2002, every pupil in the sixth grade of
primary school has to solve a test designed to check his/her knowledge and skills needed to
continue education at a higher level. The exam is obligatory and exceptions are very rare.
In this study we concentrate on the biggest region in Poland—Mazowieckie Voivod-
ship—where about 13% of all Polish students live. This region contains economically and
socially differentiated areas. Warsaw, the capital of Poland, and its suburban area are the
most prosperous regions of the country, while at the same time rural areas of Mazowieckie
are among the poorest areas in the European Union. One has to keep in mind that these
differences are probably correlated with class sizes in schools.
We received data on exam results for all primary schools in Mazowieckie from the
Regional Examination Board in Warsaw. The data were aggregated on the class level with
variables describing class mean score, the number of students in the class and indicators for
schools. Enrollment was calculated as a sum of class sizes in each school. Because the exam
is obligatory these numbers are very close to the real number of students in the class or
school. At the time the research was carried out data from the years 2002–2004 were
available.
In the original data set we had 9462 classes. We omitted privately run schools (304
classes), public schools for children with special needs, as well as schools with music and
sport classes, those run in hospitals or other unusual places and finally schools for adults
(in total 324 classes). These classes were irrelevant because our focus was on regular public
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 207

Table 1
Descriptive statistics

Pooled sample, Discontinuity sample (1) Discontinuity sample (2)

2002–2004 2975a 2974a

Class size 21.05 24.03 24.18

(6.55) (5.12) (5.23)
Enrollment 63.49 56.59 50.38
(49.17) (23.47) (23.27)
Number of classes 8664 1867 1511
Number of schools 1657 658 561

2002 sample 2003 sample 2004 sample

Whole Urban Rural Whole Urban Rural Whole Urban Rural

sample without only sample without only sample without only
Warsaw Warsaw Warsaw

Class size 21.02 24.95 17.77 21.15 25.17 17.93 20.98 23.96 18.01
(6.59) (3.93) (6.91) (6.65) (4.02) (6.99) (6.38) (5.15) (6.73)
Mean 30.36 30.41 29.76 27.56 28.13 26.48 25.92 26.15 24.68
score (2.89) (2.83) (2.88) (3.00) (2.55) (2.87) (3.43) (3.06) (3.09)
Number 2891 651 1206 2903 648 1218 2852 700 1124
of classes

Standard deviations are reported in parentheses.

a
Discontinuity samples were constructed for following grade enrollment levels: (1) 24–34, 53–63, 82–92,
(2) 25–33, 54–62, 83–91.

schools. Additionally, from 8834 classes in public schools we excluded 143 classes with 40
or more students, because they cannot be run legally, so we treated them as an error1. We
also decided to eliminate schools which have classes with 10 or less pupils for enrollment
levels greater than 20 (27 cases). Visual inspection of eliminated data shows that omitted
schools were either incorrectly coded or have classes for pupils with special needs. Thus,
our final data set contains 8664 classes (in total we eliminated 8.4% of the whole sample).
However, we repeated all estimations on the full sample to find similar results. In Table 1
we show descriptive statistics of variables used in the analysis.

4. Methodology

Our methodology focuses on empirical strategies that allow us to avoid endogeneity

bias. The dependent variable in each model is the class mean score in the primary school
sixth graders exam. We start with a general model of the following form:
T kst ¼ a þ bC ckst þ K0kst bK þ S0st bS þ Y0t h þ kst , (1)

1
There is no direct regulation of class size in public schools in Poland, however, there are some indirect
restrictions on the students’ group size. For example, Kuratoria (administrative bodies that supervise public
schools) very often claim that sport classes cannot exceed some size for security reasons. Officials in Kuratoria
declared that classes bigger than 40 are not possible in practice. That is the reason why we excluded them from the
analysis. Most of excluded cases were very big classes with 60 or even 80 students which were probably not class
sizes but mistakenly coded cohort sizes.
 ARTICLE IN PRESS
208 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

where the dependent variable Tkst is a class average score, variable ckst denotes the
logarithm of actual class size, the column vectors Kkst and Sst contain, respectively, class-
level (k) and school-level variables (s) and finally Yt is a set of dummy variables for years of
observation (t). We are interested in estimation of the coefficient bc. Similar to Hoxby
(2000) we measure class size in natural logs because a change of class size by 1 student is
proportionally larger for smaller classes. Reducing a 15-student class by 5 students could
have a stronger effect than reducing a 30-student class by the same amount.
As we said earlier such specification of the education production function and its
estimation using the ordinary least-squares method results in biased estimates of bC. This is
because class size is endogenous, which means that it is correlated with many unmeasured,
unobserved or simply omitted factors. Possible reasons for such endogeneity are the
between-schools and within-school sorting of students into classes of different sizes.
Including large set of control variables describing students, parents and community
characteristics in Eq. (1) is not a satisfactory solution. One also needs to incorporate a
complete set of lagged inputs—for example class sizes that students of a particular class
experienced or resources they experienced in their pre-school education. In practice
this is almost impossible and one will always come across some sort of bias using OLS
(Todd & Wolpin, 2003). The common problem in the production function approach is that
students with different abilities, which are very difficult to measure, cluster in classes of
different sizes.
One of the methods to avoid endogeneity bias is to apply the instrumental variable
technique using the two-stage least-squares method. However, this is restricted by the
narrow set of possible instruments2. In this paper we discuss two instruments already used
in several studies.
Our first approach is based on using as an instrument for ckst the natural logarithm of
the schools’ grade-average class size c̄st (Akerhielm, 1995; WöXmann, 2003). This way we
avoid bias from the selection of students of different abilities to classes of different sizes. As
far as there is no segregation between schools or one fully controls for it, the average class
size is believed to be exogenous and can be used to obtain unbiased estimates.
The first step to implement 2SLS method is to estimate the equation:
ckst ¼ $ þ bc̄ c̄st þ K0kst bK þ S0st bS þ Y0t h þ Zksgt , (2)
where the class size logarithm ckst is explained by the logarithm of the average class size in
a grade c̄st and all other exogenous variables. Secondly, having substituted theoretical
values c^kst ¼ ckst  Zkst from the first equation we estimate the following regression:
T kst ¼ a þ bc^ c^kst þ K0kst bK þ S0st bS þ Y0t h þ kst , (3)
where bc^ is believed to be an unbiased estimate of the class size.
However, as we said already this approach results in unbiased estimates only when one
fully controls for differences between schools and there are several reasons for between-
school sorting to be strongly correlated with class size. We mentioned some reasons earlier,
but just to reiterate a few of them: parents’ decisions concerning residential place are linked
to their children’s abilities and school resources. With school choice, better schools will
have higher enrollment levels etc. In Poland schools in rural areas are much smaller and at
2
An instrument has to be chosen very carefully. Using wrong instruments can result in even more biased
estimates (Angrist & Krueger, 2001).
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 209

the same time they scored lower on external exams, which is a very important source of
bias. One way to avoid such bias is to estimate the fixed effects of schools.
To apply this idea we include in the model equation the instrumented class size and
complete set of school dummies Ssu:
T kst ¼ a þ bc^ c^kst þ S0s u þ Y0t h þ kst . (4)
Use of the 2SLS method, with the average class size and school dummies, gives us the
opportunity to focus on class variation that is caused by fluctuations between years in
enrollment levels in the same school. In this approach we follow research by WöXmann
and West, who claim this source of variation to be exogenous (WöXmann & West, 2002;
WöXmann, 2003). However, as Hoxby (2000) points out in this method one has to exclude
class size changes that are unusual, too big or too small, because they could be caused by
external factors (e.g. high migration). Thus, we also estimate school-fixed effects 2SLS in a
subsample with limited magnitude of class size changes.
Our second approach exploits the idea that there is a dramatic change in a given school’s
average class size when a new class is added. Such variation in class size is believed to be
exogenously caused if there exists a rule that defines maximum class size in public schools.
One way to apply this idea is to calculate the function
E st
C st ¼ f ðE st Þ ¼ E st 1

(5)
int C max þ1
and to use a logarithm of this function as an instrument of the class size Ckst:
cst ¼ ln ðC st Þ, (6)
where C st denotes theoretical class size given enrollment in a particular school Est and Cmax
is the maximum class size. Following Angrist and Lavy, we call Eq. (6) a theoretical class
size function (Angrist & Lavy, 1999). Using this function as an instrumental variable gives
us a setting similar to the regression-discontinuity design in its fuzzy version since some
schools add new classes earlier and some do it later (Lee, 2005).
This instrument is based on the assumption that a formal or commonly obeyed rule that
defines maximum class size exists. Angrist and Lavy used a so-called ‘‘Maimonides Rule’’
that in Israel limits the maximum number of students in a class to 40. As we mentioned
earlier, there is no formal regulation of class size in Poland. However, there is a strong
relationship between grade enrollment and class size. In Fig. 1 one can easily notice
enrollment levels for which in some schools new classes are being created. It seems that
most schools in Poland add new classes when the average class size is about 28 or 293.
Except threshold levels, this graph is very similar to that presented for Israel in Angrist and
Lavy paper. Similarly, only in the first two or three discontinuity samples the relationship
between actual class size and theoretical class size function (with 29 as the maximum class
size in this example) is strong and visible. This fact, and the relatively high correlation
coefficients in the whole sample, let us believe that the theoretical class size determined by
function (6) could be a valid instrument for actual class size in our case4.
3
There is no formal regulation on class size in Poland, however, class sizes are restricted in an indirect way by
supervisory bodies, parents and teachers (see footnote 1).
4
Theoretical class size functions constructed for maximum class sizes of 28 and 29 explain about 60% of class
size variation in our data set. This number is very similar to that obtained for Israel by Angrist and Lavy (1999).
 ARTICLE IN PRESS
210 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

Fig. 1. Actual class size, theoretical class size function, grade-average class size and discontinuity samples. Pooled
sample, 2002–2004. Comments: Thick solid line represents theoretical class size function (with threshold set to 29).
Thin solid line stands for grade-average class size for given enrollment level. Each dot represents one class. In
order to avoid dots overlapping random noise was added to the data. Discontinuity samples were constructed for
following grade enrollment levels: 24–34, 53–63, 82–92.

It has to be noticed that such theoretical class size function is identical to the average
class size when the number of classes in each school is similar. Due to this, one needs to
fully control for differences between schools with different enrollment levels. Where this is
not the case, this instrument provides us with unbiased quasi-experimental estimates only
in the neighborhood of the enrollment threshold level where new classes are added5.
It is worth mentioning here the difference between the two instruments described above.
When the number of classes in schools is the same then these instruments are identical. In
other case, where there is a discontinuity between enrollment and class sizes, they work
very differently. When one uses grade-average class size in the discontinuity sample then
one has to control for the differences between schools with different numbers of classes or
to assume that they are negligible. The wider the discontinuity sample, the more critical are
such differences. Schools that ‘too early’ create a new class are probably different than
schools that do not. They could be placed in richer or more influential communities or
simply have more motivated parents and principals.
The crucial point here is the choice of discontinuity sample. If it is chosen properly then
schools with unusually large and small class sizes are excluded. So, one can assume that
differences in class sizes are mostly due to bureaucratic decisions and not to parental or
school characteristics. When this is the case, using theoretical class size function with
arbitrary chosen maximum class sizes seems suspicious, because we treat similar schools
differently. For example, we have three public schools with similar enrollments near the
threshold level. In Poland these could be schools with enrollments like 56, 57 and 58
5
The point is that one needs a really large sample to obtain precise estimates when one wishes to conduct an
analysis only for such ‘‘discontinuity samples’’ (see Hoxby, 2000).
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 211

students. Now, assume that the smallest school creates a new class and two others do not.
The reason could be that in the smallest school there were 3 students who initially enrolled
but then decided to move to another place. Applying theoretical class size function in this
case seems implausible.
Instead one can use average class size as an instrument with enrollment as a key control
variable. We believe that this is a proper instrument in discontinuity sample where the
variance in class sizes is mainly exogenous. The benefit derived from such an approach is
that in our case the average class size is more closely correlated with actual class size than
theoretical class size. The cost is that schools which add new classes for lower enrollment
levels can be different than schools that add new classes for higher enrollment levels and in
our case we do not have any additional data that can be used to check this. Instead we use
enrollment as a key control variable and test robustness of our results for different
discontinuity samples.

5. Results

We start with results calculated separately for each year (Table 2). OLS estimates
without any control variables show that class size is positively correlated with student
achievement. Adding enrollment considerably lowers class size coefficients, although, they
still remain positive. Note that in all regressions enrollment is positively correlated with

Table 2
Estimation results: yearly samples (2002–2004)

Regressors: k 2002 2003 2004

(1) (2) (3) (4) (5) (6)

OLS estimates
Log. of class size 0.475*** 0.164 1.315*** 0.515*** 1.560*** 0.478**
(0.131) (0.159) (0.125) (0.147) (0.157) (0.188)
Enrollment — 0.004*** — 0.012*** — 0.016***
(0.001) (0.001) (0.001)
R2 0.0041 0.0078 0.0364 0.0670 0.0328 0.0660
2SLS Estimates instrumental variable: grade-average class size
Log. of class size 0.218 0.246 1.167*** 0.2348 1.396*** 0.145
(0.134) (0.164) (0.127) (0.152) (0.161) (0.194)
Enrollment — 0.006*** — 0.014*** — 0.017***
(0.001) (0.001) (0.001)
R2 0.0028 0.0055 0.0359 0.0658 0.0324 0.0650
2SLS Estimates instrumental variable: theoretical function of class size
Log. of class size 0.212 0.354 1.347*** 0.234 1.724*** 0.349*
(0.142) (0.181) (0.140) (0.176) (0.173) (0.217)
Enrollment — 0.006*** — 0.014*** — 0.017***
(0.0013) (0.001) (0.001)
R2 0.0028 0.0042 0.0364 0.0658 0.0324 0.0658
N 2891 2903 2852

Clustering-robust standard errors are reported in parentheses. One, two, and three asterisks denote significance at
the 0.10, 0.05, and 0.01 levels, respectively. The unit of observation is the average score in the class.
 ARTICLE IN PRESS
212 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

Table 3
Estimation results: pooled sample (2002–2004)

Regressors: k Full sample Restricted samplea

(1) (2) (3) (4)

OLS estimates
Log. of class size 1.119*** 0.919*** 1.113*** 0.797***
(0.079) (0.150) (0.125) (0.247)
School dummies NO YES NO YES
R2 0.2736 0.5628 0.2703 0.6522
2SLS Estimates instrumental variable: grade-average class size
Log. of class size 0.930*** 0.583*** 0.941*** 0.644***
(0.081) (0.173) (0.127) (0.234)
School dummies NO YES NO YES
R2 0.2732 0.5566 0.2699 0.6477
2SLS Estimates instrumental variable: theoretical function of class size
Log. of class size 1.086*** 0.755*** 1.104*** 0.746***
(0.087) (0.227) (0.141) (0.297)
School dummies NO YES NO YES
R2 0.2736 0.5551 0.2703 0.6470
N 8664 8414

Clustering-robust standard errors are reported in parentheses. One, two, and three asterisks denote significance at
the 0.10, 0.05, and 0.01 levels, respectively. The unit of observation is the average score in the class.
a
In the restricted sample schools were excluded where the number of classes changed by more than one (from
year to year).

achievement. Results obtained for each year with 2SLS are quite similar. Clearly,
estimated coefficients of class size remain positive or insignificant.
To obtain unbiased estimates we conduct regressions with data pooled over 3 yr. Results
are presented in Table 36. The first column shows that OLS and IV methods indicate
positive effects, which repeats observations from 1-yr samples. However, in this case we
were able to estimate school-fixed effects to focus on the effect of changes in average class
size between years in one school. As we said earlier, such changes should give a credible
source of exogenous variation. These crucial results are shown in the second column of
Table 3. We see that both IV methods with school-fixed effects evidence a negative
relationship between class size and achievement. Additionally we repeated analysis for
schools for which a between-years change in the number of classes was 0 or 71. This was
made to assure that we look at a class size variation that is really exogenous7. Results in
column (4) show that this way we obtained similar results.
Our second empirical strategy was to use IV method only for discontinuity samples. In
Table 4 we present results obtained for 29 as the maximum class size and 75 neighbor-
hood (columns 1 and 2) and 74 neighborhood (columns 3 and 4). We also repeated our

6
In all regressions with pooled data we introduced dummies for years. We did not report them, however, there
were significant and correctly reflect year after year decline in average scores.
7
We also tried with limits set to 20% and 30% change of enrollment to find similar results.
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 213

Table 4
Estimation results: discontinuity samples

Regressors: k 2975 2974

24–34, 53–63, 82–92 25–33, 54–62, 83–91

(1) (2) (3) (4)

OLS Estimates
Log. of class size 0.478 0.013 0.346 0.0204
(0.304) (0.299) (0.326) (0.320)
Enrollment — 0.0302*** — 0.0289***
(0.003) (0.003)
R2 0.2669 0.3053 0.2973 0.3317
2SLS Estimates, Instrumental variable: theoretical function of class size
Log. of class size 2.107** .469 2.251* .242
(0.978) (1.044) (1.202) (1.268)
Enrollment — 0.031*** — 0.0292***
(0.003) (0.004)
R2 0.2557 0.3044 0.2815 0.3315

2SLS Estimates, Instrumental variable: grade-average class size

Log. of class size 0.073 0.632** 0.139 0.576*
(0.321) (0.317) (0.344) (0.338)
Enrollment — 0.031*** — 0.029***
(0.003) (0.003)
R2 0.2656 0.3036 0.2963 0.3304
N 1867 1511

Clustering-robust standard errors are reported in parentheses. One, two, and three asterisks denote significance at
the 0.10, 0.05, and 0.01 levels, respectively. The unit of observation is the average score in the class.

investigations for different samples to obtain similar results8. We analyzed only the first 3
discontinuity samples, for thresholds of 29, 58 and 87, because for higher enrollment levels
the patterns of class size change were less clear.
Table 4 shows that in the case of the OLS method all class size coefficients are
insignificant. Crucial results are shown in the second and fourth columns where results
from IV methods with control for enrollment are presented. Thus, we concentrate on
differences between schools with the same enrollment levels but different numbers of
classes. We see that, regardless of the method, without controlling for enrollment all
coefficients are positive. When we control for enrollment all IV coefficients become
negative, however, in the case of theoretical class size function they are insignificant.
One has to notice that in the latter case standard errors are considerably bigger than in
earlier regressions. Note also that in all cases enrollment is positively correlated with
achievement.
The results presented above indicate negative class size effects when we use credible
sources for variation. They indicate that a class size reduction by 10% improves class mean
score by no more than 0.08 points. This is less than 5% of the standard deviation of class

8
Additional results can be found in the working paper available at /www.wne.uw.edu.pl/mjakubowskiS or can
be requested by email: [email protected].
 ARTICLE IN PRESS
214 M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215

mean scores. Compared to the results of other experimental and quasi-experimental

studies, this number is smaller (Angrist & Lavy, 1999; Krueger, 2003; Urquiola, 2000)9.
One can say that this effect is negligible from the practical point of view. Therefore, we
applied our results to one of the most controversial issues in educational policy in Poland,
namely, to the widening gap between student achievement in urban and rural areas.
As we said at the beginning of this paper, some scholars believe that rural local
governments should increase class sizes to save money for investments in other resources
(Levitas et al., 2001). No one shows quantitatively what will be the effect of these ‘‘other’’
investments, however, we can apply our results to calculate how raising average class size
in rural areas to a level similar to that found in urban areas can influence student
achievement.
Estimates presented above are the average treatment effects for all urban and rural
classes. We had to repeat analysis to see whether similar effects are observed in the
subpopulation of rural schools with class sizes between 17 and 25 which is relevant for the
proposed application (see Table 1 for descriptive statistics). We tried several restrictions on
the change in the class size in rural schools to find that in these subsamples effects are even
stronger than in the full sample. The estimated class size effects by the 2SLS IV regressions
with the grade-average class size as the instrument and school-fixed effects were between
0.182 and 0.088 exam points for a 10% class size change. These estimates were based
on much smaller sample sizes and standard errors were 6–10 times bigger. They were not
statistically significant and further interpretation needs to be used cautiously10.
According to these estimates increasing class sizes in rural schools would widen the gap
between exam scores in urban and rural areas by 20–113% depending on the year and
estimate we used11. This numbers are not negligible and suggest that even when effects
obtained in this research seem to be small when measured by exam points, they are
important for policy concerns.

6. Conclusion

This paper analyzed class size effects in the case of primary schools in Poland. In
addition to OLS regressions, 2SLS method with instrumental variables was applied. Two
different instruments were used: the school grade-average class size and the theoretical
class size function based on a maximum class size rule. The differences between these
instruments were discussed.
9
Hoxby (2000), using similar methods, shows that there is no class size effect. However, she uses data from
Connecticut, USA, where class sizes are considerably smaller.
10
We repeated analysis with splines based on class sizes to find some evidence on nonlinear class size effects. For
bigger classes effects were much stronger, however, the sample size was too small to obtain significant results. We
do not report them assuming that further investigation is needed here. One can obtain these results by contacting
authors.
11
We calculated how increasing class sizes in rural schools will affect the difference in achievement between
rural and urban schools. We excluded Warsaw, which each year scores 2–3 points higher than the rest of the
country and has exceptional economic and social characteristics. Calculations were based on class size means for
each year separately. In all cases differences and changes in differences were statistically significant. One can easily
repeat these calculations using data presented in the lower part of Table 1 and estimates of class size effects based
on subsample of rural classes bigger than 16 and smaller than 26 which are cited in the preceding paragraph.
When we used class size effect estimates based on the full sample the policy effect was smaller: increasing rural
class sizes would widen the gap by 13–47%.
 ARTICLE IN PRESS
M. Jakubowski, P. Sakowski / Int. J. Educ. Res. 45 (2006) 202–215 215

Two empirical strategies were applied. The first strategy was to estimate class size effects
using 2SLS method with instrumental variables and school fixed effects. Results indicate
that a between-years variation in class sizes in schools negatively affects student
achievement. The claim was made that such variation is mainly exogenous. The second
strategy focused on an analysis of discontinuity samples. Again, 2SLS method with
instrumental variables was applied. Obtained estimates again indicated that class size
negatively affects student achievement.
The obtained results are of a practical importance. In recent years average class size has
not changed in Poland. However, many local governments try to increase class sizes,
especially in rural areas where they experience financial problems caused by educational
expenses. Thus, in many rural schools class size might grow considerably in the near
future. This paper has shown that increasing class size in rural schools to the level of urban
schools might further widen the achievement gap between them by 20–113%. These results
show that class size effects should be widely recognized by scholars and policy makers,
because they significantly influence achievement.

References

Akerhielm, K. (1995). Does class size matter? Economics of Education Review, XIV, 229–241.
Angrist, J. D., & Krueger, A. B. (2001). Instrumental variables and the search for identification: From supply and
demand to natural experiments. Princeton University, Working paper 455.
Angrist, J. D., & Lavy, V. (1999). Using Maimonides rule to estimate the effect of class size on scholastic
achievement. Quarterly Journal of Economics, 533–575.
Bialecki, I., & Haman, J. (2003). Program Miedzynarodowej Oceny Umiejetnosci Uczniow OECD/PISA. Wyniki
polskie-raport z badan. Warsaw: IFiS PAN.
Dobbelsteen, S., Levin, J., & Oosterbeek, H. (2002). The causal effect of class size on scholastic achievement:
Distinguishing the pure class size effect from the effect of changes in class composition. Oxford Bulletin of
Economics and Statistics, 64, 17–38.
Hanushek, E. A. (2003). The failure of input-based schooling policies. Economic Journal, 113(485), F64–F98.
Herczynski, J., & Herbst, M. (2002). Pierwsza odslona. Fundacja Klub Obywatelski.
Hoxby, C. M. (2000). The effects of class size and composition on student achievement: New evidence from
natural population variation. Quarterly Journal of Economics, 115(4), 1239–1285.
Jakubowski, M., & Sakowski, P. (2004). Wplyw wielkosci klasy na wyniki uczniow w szkolach podstawowych.
Mimeo: Department of Economics, Warsaw University.
Krueger, A. B. (2003). Economic considerations and class size. Economic Journal, 113(485), F34–F63.
Lee, M. (2005). Micro-econometrics for policy, program and treatment effects. In Advanced Texts in
Econometrics. Oxford: Oxford University Press.
Levitas, A., Golinowska, S., & Herczynski, J. (2001). Improving rural education in Poland. Report prepared for the
Warsaw Delegation of the European Commission. CASE-Center for Social and Economic Research, Warsaw,
/www.case.com.plS.
Sleszynski, P. (2002). Ekonomiczne uwarunkowania wynikow sprawdzianu szostoklasistow i egzaminu gimnazjalnego
przeprowadzonego wiosna 2002 roku. Mimeo: Ministry of Education.
Todd, P., & Wolpin, K. (2003). On the specification and estimation of the production function of cognitive
achievement. Economic Journal, 113, 485.
Urquiola, M. (2000). Identifying class size effects in developing countries: Evidence from rural schools in Bolivia.
Development Research Group, The World Bank.
WöXmann, L. (2003). European education production functions: What makes a difference for student achievement in
Europe? European economy, economic papers no. 190, European Commission, Brussels.
WöXmann, L., & West, M. R. (2002). Class-size effects in school systems around the world: Evidence from between-
grade variation in TIMSS. Harvard University, Program on Education Policy and Governance Research
paper, PEPG/02-02.