Content uploaded by Thomas Hyclak
Author content
All content in this area was uploaded by Thomas Hyclak on Nov 23, 2020
Content may be subject to copyright.
1 23
Journal of Labor Research
ISSN 0195-3613
J Labor Res
DOI 10.1007/s12122-019-09291-2
Wage Returns to English Proficiency in
Poland
Vera A.Adamchik, Thomas J.Hyclak,
Piotr Sedlak & Larry W.Taylor
1 23
Your article is protected by copyright and
all rights are held exclusively by Springer
Science+Business Media, LLC, part of
Springer Nature. This e-offprint is for personal
use only and shall not be self-archived in
electronic repositories. If you wish to self-
archive your article, please use the accepted
manuscript version for posting on your own
website. You may further deposit the accepted
manuscript version in any repository,
provided it is only made publicly available 12
months after official publication or later and
provided acknowledgement is given to the
original source of publication and a link is
inserted to the published article on Springer's
website. The link must be accompanied by
the following text: "The final publication is
available at link.springer.com”.
Wage Returns to English Proficiency in Poland
Vera A. Adamchik
1
&Thomas J. Hyclak
2
&Piotr Sedlak
3
&Larry W. Taylor
2
#Springer Science+Business Media, LLC, part of Springer Nature 2019
Abstract
We investigate the effect of English language proficiency on the wages of native full-time
employees in Poland. Using a unique data set with information on over 600,000 survey
respondents polled over the five-year period from 2013 to 2017, we employ an IVapproach
founded on a natural experiment - namely, the reform of foreign language instruction in
Polish schools. Our preferred estimates indicate that monthly wages for those individuals
with ‘good’or ‘very good’knowledge of English exceeded the wages of those with ‘no
English’(or those with just a conversational proficiency) by nearly 60% for men and more
than 50% for women. The estimates are statistically significant for both genders, and
suggest quantitatively relevant wage returns for proficiency in English.
Keywords Wage differentials .Human capital .English language competence .
Transitional economies
JEL Classification I26 .J24 .J31 .P23
Introduction
How does proficiency in English affect the earnings of Europeans for whom English is not
their native tongue? From survey responses in the multi-lingual European Union (EU), large
majorities express the belief that knowledge of languages other than their mother tongue is
very useful for attaining a better job. Moreover, this belief has influenced the educational
decisions of Europeans and their governments. Consider that in year 2014, 83.8% of all
primary education students across the EU studied at least one foreign language, with English
as the most common language to learn (EC/EACEA/Eurydice 2017,p.11).
Journal of Labor Research
https://doi.org/10.1007/s12122-019-09291-2
*Thomas J. Hyclak
tjh7@lehigh.edu
1
University of Houston –Victoria, Victoria, TX, USA
2
Lehigh University, Bethlehem, PA, USA
3
Cracow University of Economics, Kraków, Poland
Author's personal copy
Previous studies of the wage returns to knowing English for native workers in
European and other countries provide evidence for a significant positive correlation
between earnings and language ability. Much of this evidence, however, is based on
data for the late 1990s and early 2000s –well before the rapid increase in global
economic integration could be expected to enhance the economic value of knowing
English. Furthermore, many of the past studies draw conclusions from potentially
biased ordinary least-squares (OLS) estimation on cross-section data.
In this paper, we examine the impact of English proficiency on wages earned by men
and women in the Polish labor market during the five-year period from 2013 to 2017.
We have access to proprietary data from ongoing surveys conducted by a major Polish
human resource management consulting firm that yields observations on over 330,000
men and 250,000 women who were full-time employees. Most importantly, we employ
an identification strategy that capitalizes on a natural experiment that occurred in year
1989. At that point in time, a national educational reform substantially changed
mandatory foreign-language instruction in Polish schools.
Using the 1989 educational reform to construct an instrumental variable for knowledge
of English allows us to infer causal estimates of the wage returns to English proficiency. Our
approach is that of fuzzy-regression discontinuity design. As a robustness check, we relax
the frequently questionable exclusion restriction assumption in IV estimation by using the
recently proposed quasi-Bayesian Local-to-Zero (LTZ) method from Conley et al. (2012),
coupled with an implementation strategy from van Kippersluis and Rietveld (2018).
Regardless of the econometric method employed, however, we find substantial
economic returns to fluency in English. Our preferred estimates indicate that male
employees with a self-described ‘good’or ‘very good’understanding of English have
monthly earnings of about 60% greater than the group consisting of those individuals
with either ‘no knowledge’of English or just a ‘conversational proficiency’in English.
For women, the estimated wage premium for having a ‘good’or ‘very good’under-
standing of English is just above 50%. These estimates are generally much higher than
those found in previous studies, and support the strongly-held belief by many Euro-
peans in the economic value of knowing English.
The outline of our paper is as follows. Second section presents an overview of the
related literature, and third section provides information about our proprietary data set
and the English language skills of the men and women in the sample. Fourth section
presents our empirical methods, and fifth section presents our estimation results. Sixth
section then presents robustness-check estimates based on the Conley et al. (2012)
approach. We then conclude in seventh section with a discussion of our findings.
Related Literature
This section offers a brief overview of the empirical contributions that discuss rela-
tionships between linguistic skills and earnings. We limit our review to studies that
estimate the effect of English skills on earnings for the native-born population in
countries where English is not demolinguistically dominant.
1
Thus, we deliberately
1
For an extensive bibliography of literature in language economics, the reader is referred to Gazzola et al.
(2016).
Journal of Labor Research
Author's personal copy
do not cover papers analyzing the wage returns to English language skills of immigrant
workers. Indeed, Araújo et al. (2015) point out that “investigating the differential
association between language skills and labour market outcomes for natives and for
migrants requires applying different conceptual frameworks and models”(p.65, foot-
note 32). Additionally, we leave out studies on the wage returns to English language
skills of the native-born population in countries where English is one of the official
languages (e.g., Canada, India, South Africa). Finally, we omit studies on the returns to
using English at the workplace.
2
Hundreds of scholarly contributions have been published about the economic value
of foreign languages. However, we could find only a handful of papers closely related
to our study. These papers investigate the extent to which, ceteris paribus, a better
general knowledge of English among the native-born population (in total, or by
gender) in a non-English-speaking country is associated with higher earnings. In what
follows, we present brief summaries of those analyses. If a study dealt with several
foreign languages, we only report the findings for English.
The most important of the early studies is that by Grin (2001). Grin examines the
value of English in the Swiss labor market. In Switzerland, English is neither an official
language, nor a majority language. Grin draws data from the 1994/95 “Foreign
Language Competence in Switzerland”survey in which 2,400 Swiss residents indicat-
ed their level of proficiency in foreign languages (none, basic, good, and fluent). Using
OLS, Grin finds that English language skills are associated with remarkably high and
statistically robust wage premia: 14% (basic), 19% (good), and 31% (fluent) for men,
and 12%, 28%, and 22% for women, respectively. This early study also uncovered
significant gender differences in converting English language proficiency into earnings.
In fact, there was about a 40% wage gap (for full-time equivalent earnings) between
males and females at all levels of proficiency in English.
In a second relevant study, Lang and Siniver (2009) use the Workplace Occupational
Survey conducted in Israel in 1998–2000. Among other questions, both native Israelis
and immigrants were asked to classify their ability to speak English as ‘not well,’
‘well,’or ‘very well.’For the whole sample of 1,243 native Israelis and 1,483
immigrants, the authors estimate cross-section and longitudinal models of the benefits
of knowing English ‘very well.’For a selected sub-sample of the native Israelis,
however, only cross-section OLS estimates are reported. English fluency among
natives is associated with a 10% wage premium, with as high as 14% for the highly-
educated group and as low as 8% for the low-educated group.
Ginsburgh and Prieto Rodriguez (2011) use the European Community Household
Panel Surveys in 1994–2001 to examine the wage premia for several languages for
native male workers in nine EU countries: Austria, Denmark, Finland, France, Germa-
ny, Greece, Italy, Portugal, and Spain. The number of observations per country varies
from 928 for Finland to 2,401 for Spain. Although the authors generally prefer using
the disenfranchisement rate instead of representing language skills by dummy variables
or by the level of proficiency, they do estimate one model comparable to those
presented in this paper. They define a dummy variable equal to 1 when self-reported
2
DiNardo and Pischke (1997) point out that inferences based on ‘usage at work’of some tools (i.e., skills) can
be misleading if one wants to measure their effect on wages. Grin (2003, pp. 19-20) also states that language
use is a fairly vague and unreliable definition of language competence.
Journal of Labor Research
Author's personal copy
knowledge of English is good enough to converse in most social contexts and to read
complex information, and equal to 0 otherwise. The authors run OLS Mincer-type
regressions, using individual fixed effects to eliminate omitted variables bias, and find
that in all eight countries (there are no estimates for Germany), knowledge of the
English language has a statistically significant positive effect on earnings, with results
varying from 10.7% in Denmark to 48.9% in Spain (with no occupational controls),
and from essentially zero for Austria to 23.5% in Portugal (with occupational controls).
Wang et al. (2017) examine economic returns to proficiency in English in China
using the 2012 and 2014 waves of the China Labor-Force Dynamics Survey, with
9,567 individuals surveyed in both years. The authors use self-assessed English
language proficiency measured on a 0-to-4 Likert-type scale (0 = cannot speak to 4 =
very fluent). They employ panel and cross-section data. For the panel estimation with
random effects, the authors find that overall proficiency in English is associated with an
increase of 4.6% in hourly wages regardless of gender, but only 3.3% for the subsample
of males, and 5.4% for the subsample of females. For single-year 2012, the OLS
estimates are 7.3%, 7.1%, and 7.5%, respectively; and the 2SLS estimates (using lead
English proficiency and whether the individual smoked at age 18 as IVs) are 16.4%,
11.7%, and 19.4%, respectively.
Too me t (2011) analyzes the return to knowing English for Russian-speaking minor-
ity men ages 20 to 60 in Estonia and Latvia. The author uses the 2000–2010 Estonian
Labor Force Surveys (about 8,600 observations) and the 1997 Paths of a Generation
study available for both Estonia and Latvia (about 400 observations for each country).
Self-reported fluency in English is coded on a linear scale from 0 to 1 (0 = no
knowledge, 0.33 = understanding, 0.66 = speaking, and 1 = writing). Toomet then ap-
plies a random effects model to the 2000–2010 Estonian LFS data and finds a 14.4%
wage premium for men who speak fluent English. Using the 1997 wave of the Paths of
a Generation study, the estimates are much higher–41.5% for those living in Estonia
and 62.4% for those in Latvia.
In another Estonian study, Bormann et al. (2019) use the 2000–2012 Estonian Labor
Force Surveys. The authors limit their sample to those 25–55 years of age and consider
four groups of workers: ethnic Estonian men (21,785 observations) and women (26,673
observations), and Russian-speaking minority men (7,978 observations) and women
(9,697 observations). English language proficiency is measured by a dummy variable,
equal to 1 if one can speak or write, and 0 otherwise. The authors apply OLS and find
that English language fluency is associated with a large and statistically significant
wage premia for all analyzed categories of workers. Depending on the specification of
the wage equation, the wage premium is 10.8%–34.3% for ethnic Estonian men, 7.5%–
31.1% for ethnic Estonian women, 10.3%–24.8% for Russian-speaking minority men,
and 5.4%–22.3% for Russian-speaking minority women.
In a recent working paper, Liwiński (2018) uses individual-level data from the
2012–2014 Human Capital Balance surveys to assess the effects of foreign language
skills on wages of the working-age population in Poland. In the surveys, respondents
were asked to list all languages they knew and to assess the level of proficiency for the
three languages they knew best (using a six-grade scale) in four linguistic competences:
reading, writing, speaking, and listening comprehension. For the whole sample of
14,145 observations, Liwiński first calculates the average score for the four aforemen-
tioned linguistic competencies and then uses it to construct three dummy variables for
Journal of Labor Research
Author's personal copy
language proficiency: elementary, intermediate, and advanced. Using OLS, the author
reports wage returns of 7.5% for advanced, 4.4% for intermediate, and −6.2% for an
elementary level of fluency in English.
Di Paolo and Tansel (2015) draw data from the 2007 Adult Education Survey in
Turkey and consider adult (aged 25–65) male wage-earners, totaling 6,018 observa-
tions. In the survey, individuals reported how many foreign languages they knew and
then provided detailed information about the two foreign languages they knew best.
Three dummy variables for language proficiency (basic, regular, advanced) for each
language are used in the estimations. Using OLS, for English the authors find a positive
wage premium that increases with the level of competence: 3.0% for basic, 19.8% for
regular, and 45.4% for advanced levels of knowledge. To control for endogeneity, the
authors apply an IV method by employing information on the frequency of English use
for leisure (as the exclusion restriction). The estimated wage premia are higher than
those obtained without controlling for endogeneity: 9.1%–11.7% for basic, 23.6%–
31.2% for regular, and 48.1%–61.0% for advanced levels of proficiency.
In a follow-up study, Di Paolo and Tansel (2019) use the same survey but consider
only women. They employ two dummy variables for English proficiency (basic and
advanced). OLS estimates of wage premia for 13,491 Turkish women range from
statistically-insignificant 3.0% to 5.4% for basic knowledge, and 19.4% to 24.0% for
advanced knowledge. Di Paolo and Tansel then adopt a bivariate equation framework
and jointly model the effect of English skills on labor market status and, conditional on
being a wage earner, on monthly earnings. After controlling for selection into paid
employment, they find that the wage premia associated with English skills are higher
than those obtained by OLS: marginally-significant 5.5%–7.0% for basic knowledge,
and 25.1%–26.7% for advanced knowledge.
To summarize, empirical studies on the language wage premia almost always rely on
self-reported survey data. For the purpose of estimation, proficiency in languages is
typically represented by dummy variables or is coded on a linear scale. There is a quite
large variation among the reported estimates of the English-language wage premia.
Such differences can be attributed to the quality and comprehensiveness of data, the
model specification, and the empirical estimation strategy adopted by the various
authors. Nevertheless, all studies reviewed in this section converge on the same
conclusion: the effect of English knowledge on earnings is positive and quantitatively
important. Finally, with the exception of the study of Chinese workers by Wang et al.
(2017), estimates of wage effects for women tend to be smaller than those estimated for
men.
Data and Measures of Language Proficiency
In this paper, we employ proprietary data provided by the Sedlak & Sedlak (S&S)
company. S&S was founded in 1990 and is Poland’s oldest Human Resources advisory
firm. They implement salary surveys and provide consulting related to compensation.
S&S collects salary data both from companies and the Polish population at large via
web-based interviews. An S&S survey of special interest to us is the web-based Polish
General Salary Survey (in Polish - Ogólnopolskie Badanie Wynagrodzeń, OBW).
Launched in year 2004 as the largest non-governmental salary survey in Poland, the
Journal of Labor Research
Author's personal copy
survey remains activated and ready to receive responses all-year-round. Annual data-
bases are created by combining the survey responses submitted from January 1 through
December 31 in any given year.
The OBW questionnaire is located on the ‘wynagrodzenia.pl’website (the domain
name can be translated as ‘salary.pl’). Invitations to the survey are distributed through
email campaigns, text links connected to various Internet articles published by S&S,
and through cooperation with partner companies, web pages, and paper magazines.
Unfortunately, this mixed method of soliciting responses makes it difficult to calculate a
true response rate to the survey. It is known, however, that nearly 6 million individual
users visit the website each year, with more than 100,000 participating in the survey
annually. To ensure data reliability and quality, S&S employs a number of quantitative
and qualitative checks of the survey responses. For example, a plausibility analysis is
employed to check for inconsistent and/or conflicting answers to the survey, to
determine whether questionnaire completion time is reasonable, and to check for
outliers. This plausibility analysis typically results in about 5% of annual survey
responses to be excluded from the final database.
We use pooled data from the 2013–2017 OBW surveys. Excluded are employers
and self-employed individuals, and we limit our sample to full-time male employees
ages 18–64 and full-time female employees ages 18–59. The different upper-age limits
correspond to different normal retirement ages by gender. The total number of re-
sponses with complete data is 335,739 for men, and 254,221 for women.
Among other questions, respondents were asked to self-evaluate their proficiency in
English and other languages by using the following pre-defined scale: no knowledge,
conversational ability, good ability, or very-good ability. The last three levels corre-
spond to A1/A2, B1/B2 and C1/C2 levels in the self-assessment grid (CEFR) used in
the European Language Portfolio.
3
For the 5 years from 2013 to 2017, Table 1presents
descriptive statistics on the distribution of self-described English proficiency in our
samples. The importance of English is very clear. Well over 80% of women and men in
the full sample have some knowledge of English. By comparison, data from the survey
(not reported in Table 1) show that about 32% of women and men had some knowledge
of German, 23% had some knowledge of Russian and fewer than 10% had proficiency
at any level in French, Italian or Spanish.
Tab le 1also shows the distribution of self-described English proficiency among
those who started compulsory language instruction in primary school either ‘prior to’or
‘after’the 1989 educational reform. With the demise of the socialist system, the non-
communist government launched numerous political and economic structural reforms
as well as profound reforms of the educational system. A major educational priority
was “a large-scale introduction of West European languages, first of all English, to
schools”(Janowski 1992, p. 50). During the socialist era, the Russian language was
mandatory for all pupils from the age of 11 years, that is, grade 5 of primary school. A
second foreign language –English, German, French or Latin –was introduced only in
secondary general schools. In vocational schools, a second foreign language was rarely
taught. Consequently, at the end of the 1980s Poland had 18,000 teachers of Russian
and only 1,200 teachers of English, and in the 1988/89 school year 100% of pupils
studied Russian (Komorowska 2014).
3
http://www.coe.int/en/web/portfolio/self-assessment-grid
Journal of Labor Research
Author's personal copy
As a result of the 1989 educational reform, starting the 1989/90 school year, Russian
was no longer mandatory and students could elect other languages, including English,
as the first foreign language required in primary school. In addition, the starting age for
languages was lowered from 11 to 10. The reform also broadened the list of languages
for secondary education by introducing Spanish, Italian, Japanese, Hungarian, and
Swedish into the curriculum. Russian continued to dominate foreign languages studied
in Poland until the 1996/97 school year when English moved to first place, followed by
German, Russian and French. This lineup of foreign languages generally remains
unchanged to these days.
It is clear from Table 1that the reform of compulsory language instruction in
primary school is associated with a marked increase in English proficiency among
Polish workers. Both men and women who attended primary school after the reform
reported significantly higher levels of ‘very good’or ‘good’knowledge of English, and
only 7% of both men and women in that group said that they had ‘no English
proficiency’at all.
Empirical Specification and Method
OLS estimates of the economic returns to knowing English may be hampered by
different sources of distortion, typically referred to as endogeneity bias (Card 1999,
2001; Gunderson and Oreopoulos 2010). The literature identifies three sources of this
bias: omitted variables, simultaneity, and measurement error. Of the three, omitted-
variables bias is potentially the most severe.
Consider that people who are more proficient in foreign languages may also have
greater innate abilities, as well as more favorable socio-economic and family
Table 1 English language proficiency: full-time employees, 2013–2017
Level of proficiency Full sample Pre-reform Post-reform
Mean Standard deviation Mean Standard deviation Mean Standard deviation
Women
Very Good 0.20 0.40 0.12 0.32 0.24 0.43
Good 0.39 0.49 0.25 0.43 0.47 0.50
Conversational 0.24 0.42 0.28 0.45 0.22 0.41
None 0.17 0.38 0.35 0.48 0.07 0.26
Observations 254,221 87,110 167,111
Men
Very Good 0.21 0.41 0.14 0.35 0.25 0.43
Good 0.39 0.49 0.26 0.44 0.46 0.50
Conversational 0.24 0.43 0.28 0.45 0.21 0.41
None 0.15 0.36 0.32 0.47 0.07 0.25
Observations 335,739 117,396 218,343
Pre-reform (post-reform) identifies those who started compulsory language instruction in primary school prior
to (during or after) 1989. Proportions may not add to 1.0 due to rounding
Journal of Labor Research
Author's personal copy
background that would allow them to earn more even without knowing a foreign
language. In such cases, the cross-sectional correlation between foreign-language
proficiency and earnings may differ from the true causal effect because the OLS
estimate will reflect a return to these unobserved factors, as well as to foreign language
proficiency itself.
Although panel data can sometimes be employed to eliminate the bias in OLS due to
omitted variables that do not vary over time, the same is not true for pooled cross-
sections like our own. Nonetheless, it is also possible that the decision to learn a foreign
language is uncorrelated with greater general ability. If so, OLS estimation will not be
impacted by the omission.
The problem of misclassification error from self-reported knowledge of foreign
languages is documented by Dustmann and van Soest (2001,2002), Bleakley and
Chin (2004), Hamilton et al. (2008), Akresh and Frank (2011). Although measurement
error can bias OLS estimates in either direction, empirical studies show that the
probability of over-reporting is larger than the probability of under-reporting, thus
leading to an underestimation of the returns to knowing English.
4
To resolve the mis-
measurement problem, Dustmann and van Soest (2001,2002) used leads and lags of
the language measure from their panel data as instruments, and Bleakley and Chin
(2004) used the literacy test score as the true measure of language skills. Hamilton et al.
(2008), Akresh and Frank (2011) used both self- and interviewer-assessed measures of
English language proficiency.
Unfortunately, due to data limitations, none of these approaches is feasible in this
study. Nonetheless, similar to omitted-variables bias, our own IV approach should
address the mis-measurement problem and allow for (asymptotically) unbiased and
consistent estimates of the returns to knowing English.
Indeed, as pointed out by Card (1999, p. 1817), “a standard solution to the problem
of causal inference is instrumental variables (IV).”We hence employ the model below
where our instrument (that we refer to simply as ‘IV’) affects log-wages through
competency in the English language, but is otherwise excluded from the wage equation.
In particular, we posit IV to be uncorrelated with unobserved ability and other
characteristics that are associated with higher earnings.
Formally, our model is:
lnWi¼αLiþβXiþεið1Þ
Li¼θIViþηXiþξið2Þ
where Wiis individual i’s total monthly earnings for individuals i=1,2,…,N; English
proficiency is indicated by Li, an index with non-negative support (with 0 indicating
4
There is also some evidence that self-evaluations may be a generally reliable way of measuring the level of
language skills. Oscarson (1984), Blanche and Merino (1989) and Ross (1998) all conclude that self-
assessments are highly correlated with the outcomes from formal tests of language ability. In a more recent
study, Dragemark Oscarson (2009) reports the results from the Swedish National Evaluation where correla-
tions between self-assessed and formal scores were high (about 0.7) and where 85% of students received the
score they had estimated.
Journal of Labor Research
Author's personal copy
‘no ability’); Xiis a matrix of strictly exogenous variables that control for birth cohort
and age; and IV refers to our instrumental variable.
We use two alternative definitions for Li. The first is a straightforward 0/1
dummy variable, Good English, that identifies those individuals who claim either
one of ‘good knowledge’or ‘very good knowledge’of English, but with those
with ‘no knowledge’or a ‘conversational knowledge’of English serving as the
reference group. Second, in an attempt to incorporate more information about the
distribution of language skills in the analysis, we construct an ordinal ability scale
for English by following the method employed by Bleakley and Chin (2004),
Too me t (2011),andWangetal.(2017). The scale (dubbed the English Index)
assumes a linear hierarchical structure for English proficiency and is set equal to 3
for ‘very good knowledge’,2for‘good knowledge’,1for‘conversational knowl-
edge’and0for‘no knowledge’of English.
The notable difference between the pre-reform and post-reform sub-samples in
Tab le 1suggests the strong possibility that the educational reform implemented in
1989 by the Polish government might serve as an instrumental variable for English
proficiency. Our instrument is equal to one for those individuals who began compulsory
language instruction in primary schools at age 10 starting from year 1989 inclusive, and
is equal to zero otherwise. Since those who began language instruction post-reform had
a greater opportunity to study English in school, we anticipate that our reform IV will
be positively correlated with individual English proficiency. Note that the vast majority
of EU and Polish respondents to language surveys identify formal language instruction
in school as the way they learned a foreign language (EC 2012). In particular, learning
English on the job is rare.
Our empirical approach is similar to that of Oreopoulos (2006) who estimates the
local average treatment effects of education from a change in compulsory schooling
laws. As demonstrated by Angrist and Pischke (2009), our IV estimation is also one of
a fuzzy regression discontinuity (RD) research design. Our fuzzy RD capitalizes on the
discontinuity in the probability of learning English at a young age. Indeed, the design is
fuzzy rather than sharp because the probability of learning the English language is not
identically equal to zero prior to the reform, nor identically equal to one right after the
reform. Rather, there is a jump in the probability of having higher levels of English
proficiency for women and men who were 10-years-old right after the implementation
of the language instruction reform in 1989, and this is why we have identified IVas a
potentially strong instrument.
Figures 1and 2display adult English proficiency for men and women who were
born in the 10 years before and after 1979. Since the educational reform changed
language instruction for 10-year-old pupils beginning in 1989, those born in 1979 were
the first to be affected. The graphs display mean values of our variable Good English
for each cohort. Fitted values were obtained from the full sample by regressing Good
English on the IV dummy variable, birth cohort and birth-cohort squared. For those
born in 1978, about 55% of both men and women report ‘good’or ‘very good’
proficiency in English. For the cohort born in 1979, the percentage rises to 60% for
women and 62% for men –a substantial jump that we attribute to language reform in
Poland. We believe that the jump in English proficiency for those born in year 1979
provides us with an opportunity to formally quantify the impact of language skill on
wages.
Journal of Labor Research
Author's personal copy
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1965 1970 1975 1980 1985 1990 1995
Good English for Men by Birth Year
Fig. 1 Good English identifies individuals with self-assessed good or very good proficiency in English. Dots
plot the mean value of Good English in 2013–2017 for each cohort by birth year. Lines trace fitted values from
a regression of Good English on the educational reform IV, cohort and cohort squared for the full sample
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1965 1970 1975 1980 1985 1990 1995
Good English for Women by Birth Year
Fig. 2 Good English identifies individuals with self-assessed good or very good proficiency in English. Dots
plot the mean value of Good English in 2013–2017 for each cohort by birth year. Lines trace fitted values from
a regression of Good English on the educational reform IV, cohort and cohort squared for the full sample
Journal of Labor Research
Author's personal copy
Estimation Results
In this study, we perform separate analyses for men and women since our review of
previous studies suggests that the economic returns to English skills may differ by
gender. Table 2presents descriptive statistics for the variables in our regression model.
The outcome variable is the log of total monthly wage and salary earnings, adjusted for
inflation by the Consumer Price Index for Poland, with 2015 = 100. In our sample, we
found the educational reform to affect 65% of men and 66% of women. As expected,
our two measures of English proficiency are substantially higher for those who began
compulsory language instruction in primary school at age 10 in the post-reform period.
It is worth noting that average real earnings and average age are lower in the post
reform period, reflecting the typical positive correlation between age and earnings. This
positive correlation strongly suggests the need to control for age in addition to birth
cohort in the regression equation. We can do so because of the multi-year construction
of our data set.
Table 3presents first-stage and reduced-form estimation results for males and
females. Regardless of gender and whether we control for age, our visual interpretation
of Figs. 1and 2is confirmed: the educational reform IV has a statistically significant
Table 2 Characteristics of regression variables: full-time employees, 2013–2017
Full sample Pre-reform Post-reform
Mean Standard
deviation
Mean Standard
deviation
Mean Standard
deviation
Men
Log of Total Monthly
Earnings
8.48 0.62 8.62 0.69 8.41 0.57
Good English 0.61 0.49 0.40 0.49 0.72 0.45
English Index 1.66 0.98 1.22 1.04 1.90 0.86
Birth Cohort 32.3 9.53 21.8 7.31 38.0 4.33
Age 34.79.43 45.17.41 29.14.22
Educational IV 0.65 0.48 0 0 1 0
Observations 335,739 117,396 218,343
Wom en
Log of Total Monthly
Earnings
8.26 0.55 8.37 0.62 8.20 0.51
Good English 0.59 0.49 0.36 0.48 0.71 0.45
English Index 1.62 0.98 1.13 1.03 1.88 0.86
Birth Cohort 27.7 9.03 17.3 6.47 33.0 4.20
Age 34.48.87 44.66.52 29.24.05
Educational IV 0.66 0.47 0 0 1 0
Observations 254,221 87,110 167,111
To ensure a well-conditioned design matrix for the regression equations, Birth Cohort is defined as “Year of
Birth –1948”. The pre-reform (post-reform) identifies those who started compulsory language instruction in
primary school prior to (during or after) 1989
Journal of Labor Research
Author's personal copy
positive effect on both measures of English proficiency. For both men and women, the
estimated fraction of those with Good English skills jumps by just over 10% in the
post-reform period relative to the pre-reform period. Furthermore, the instrumental
variable IV is associated with an increase in the English Index of about 15% for women
and 16% for men (if the English Index is evaluated at the full-sample mean by gender).
The first-stage regressions also allow us to analyze the characteristics of those
individuals likely to learn English because of the educational reform. Such individuals
are called ‘compliers’, and our IV estimates below measure the returns to knowing
English for this group. Importantly, our IV estimates are not relevant to those ‘always
takers’who would choose to learn English even if there was no reform in education.
Likewise, our IV estimates are not relevant to the ‘never-takers’, and (as is commonly
done) we exclude from consideration any ‘defiers’in the group. A defier does the
opposite of what is required –learning English instead of Russian if subject to the
standards in the pre-reform period, but doing the opposite if subject to the standards in
the post-reform period.
Following Imbens and Wooldridge (2007), define πaas the probability of an always-
taker and πcas the probability of a complier. The probability of a never-taker is 1 - πa-
πc.Toobtainanestimateforπa, run the first-stage regression on Good English and
average the predicted (or fitted) values over all individuals for which IV = 0. To obtain
an estimate for πc, average the predicted values over all individuals for which IV = 1,
and then subtract from this value the estimate for πa. The values of πaand πccan be
estimated for various subgroups by restricting the sample to those with the chosen
characteristics.
For the subgroup of men, our estimates are eπa= 0.40 and eπc= 0.32. For the
subgroup of women, eπa=0.37 andeπc= 0.34. This suggests that of the roughly 70%
of the male and female populations with Good English skills in the post reform period,
Table 3 Effect of educational reform IVon English proficiency and log monthly earnings
(1) (2) (3) (4)
Dependent variable Men Women
Good English 0.1114***
(0.0166)
0.1149***
(0.0169)
0.1057***
(0.0183)
0.1079***
(0.0186)
English Index 0.2727***
(0.0382)
0.2657***
(0.0370)
0.2610***
(0.0435)
0.2466***
(0.0415)
Log Monthly Earnings 0.0720***
(0.0199)
0.0950***
(0.0207)
0.0336***
(0.0136)
0.0543***
(0.0142)
Age Polynomial Controls No Yes No Yes
Birth Cohort Polynomial Controls Yes Yes Yes Yes
Observations 335,739 335,739 254,221 254,221
The dependent variables are (first-stage) measures of English proficiency and inflation-adjusted log monthly
earnings. Each reported coefficient on the dummy IV is from a separate regression that controls for a birth
cohort quadratic polynomial. The regression may also control for an age quadratic polynomial –our preferred
specification. Regressions are clustered by birth cohort, and all estimated coefficients are statistically signif-
icant at 1 % (as indicated by ***). Males were born years 1949–1999 and range in age from 18 to 64 at the
time of the survey; females were born years 1954–1999 and range in age from 18 to 59 at the time of the
survey
Journal of Labor Research
Author's personal copy
slightly less than half had learned English in response to the educational reform. For
each gender, we also found that always-takers were much more prevalent among the
highly educated, and compliers were a bit more prevalent among the least educated.
That is, relative to those with less than a college degree, the highly educated (those
with, say, an MBA or Ph.D.) appeared much more self-motivated to learn English
regardless of whether they were subject to the pre-reform or post-reform educational
standards.
For the wage equation, the reduced-form coefficient (from Table 3) on the IV
dummy is highly statistically significant for both women and men. The magnitude of
the point estimate, however, depends on gender. Controlling for age, the estimated
coefficient on IV is about 0.095 for males, but only about 0.054 for females. With
English skills measured by Good English, it follows that the 2SLS estimate of having
good skills in English on log monthly earnings is about 0.83 (= 0.0950 / 0.1149) for
men, but only about 0.50 (= 0.0543/0.1079) for women. In other words, although male
and female full-time employees were similarly affected by the language reform in terms
of English proficiency, the wage returns to having good skills in English seems tilted
towards the males.
Tab le 4presents further empirical evidence. Panel A reports OLS estimation of the
wage returns to our two measures of English proficiency, and Panel B reports the 2SLS
estimates of wage returns with IV as an instrument. All estimations employ clustered
standard errors by birth cohort.
For 2SLS estimation in the context of regression discontinuity design, Card and Lee
(2008) emphasize the need to employ clustered standard errors when cohort births are
measured in discrete time. In our case, births are measured to the nearest year.
Clustering the standard errors accounts for specification uncertainty close to the
threshold of discontinuity for which it is impossible to know the correct functional
form unless time is measured continuously. Nonetheless, for 2SLS estimation, we can
and do employ the IV-RESET functional-form test from Pesaran and Taylor (1999)to
determine whether our chosen specification of the cohort and age polynomials is
acceptable.
The OLS estimates measure the average treatment effect on log monthly
earnings for those trained in English, and these estimates are quite similar regard-
less of whether we control for age. The OLS results suggest that male full-time
employees with ‘good’or ‘very good’English skills earn about 45% more than
those with ‘conversational skills’or ‘no knowledge’of English. For women, the
wage advantage associated with Good English is slightly lower at 37%. For the
OLS estimates of the coefficients on English Index, a male (female) full-time
employee with ‘very good’English skills earns about 80% (69%) more than one
with ‘no knowledge’of English.
Unfortunately, as mentioned earlier, the OLS estimates are potentially biased. It is
for this reason that we are interested in the 2SLS estimates. Of course, the estimand for
2SLS estimation is the local average treatment effect that measures the benefit for
individuals who study English only because of the change in educational policy (that is,
the compliers).
Although the 2SLS estimates share some of the characteristics of the OLS estimates,
they differ in important ways. First, since the functional-form test strongly rejects the
specifications in Columns 1 and 3, the inclusion of age polynomials matters for our
Journal of Labor Research
Author's personal copy
2SLS estimates. We thus focus our attention on Columns 2 and 4. Second, in these
even-numbered columns, the respective coefficients from 2SLS estimation are gener-
ally larger than the OLS estimates.
We expect the 2SLS coefficients to be larger than the respective OLS coefficients if
compliers invest in learning English because they have more to gain than the average
Polish worker. The only minor contradiction to our expectation is for women when
English proficiency is measured by the English Index (in Column 4). Here, the 2SLS
estimate is slightly less than the OLS estimate. Finally, we note that the robust weak-
instrument test of Montiel Olea and Pflueger (2015)suggeststhatIVisanextremely
strong instrument.
Table 4 OLS and IV returns to English proficiency for log monthly earnings
(1) (2) (3) (4)
Men Women
English proficiency Panel A: OLS estimates
Good English 0.4529***
(0.0249)
0.4526***
(0.0253)
0.3694***
(0.0211)
0.3696***
(0.0212)
English Index 0.2742***
(0.0 111)
0.2757***
(0.0108)
0.2290***
(0.0088)
0.2311***
(0.0085)
Panel B: 2SLS estimates
Good English 0.6461***
(0.1745)
0.8266***
(0.1518)
0.3176***
(0.1099)
0.5029***
(0.0894)
Weak IV Test 1,482.04 > > 37.418 1,554.43 > > 37.418 966.02 > > 37.418 990.11 > > 37.418
IV-RESET 8.44*** (0.0037) 1.08 (0.2985) 9.86*** (0.0017) 0.06 (0.8052)
Hausman Test 1.21 (0.2711) 5.20** (0.0225) 0.20 (0.6539) 1.54 (0.2142)
English Index 0.2640***
(0.0659)
0.3576***
(0.0575)
0.1286***
(0.0433)
0.2201***
(0.0357)
Weak IV Test 2,259.96 > > 37.418 2,113.78 > > 37.418 1,511.15 > > 37.418 1,326.82 > > 37.418
IV-RESET 7.43*** (0.0064) 0.47 (0.4916) 9.65*** (0.0019) 0.36 (0.5504)
Hausman Test 0.02 (0.8796) 1.66 (0.1979) 4.99** (0.0254) 0.07 (0.7850)
Age Polynomial
Controls
No Yes No Yes
Observations 335,739 335,739 254,221 254,221
The dependent variable is inflation-adjusted log monthly earnings. Each regression controls for a birth cohort
quadratic polynomial and a measure of English proficiency (instrumented for the 2SLS estimations by an
indicator for educational reform in year 1989). In every 2SLS regression, the Wea k IV Te st indicates a very
strong instrument since the test statistic is much larger than the critical value of 37.418 at a significance level of
5% and a worse-case bias of 5%. Regressions in columns 2 and 4 also control for an age quadratic polynomial
–our preferred specification. In all 2SLS estimations, when age controls are omitted (columns 1 and 3), IV-
RESET rejects the functional form at the 5% significance level (pvalue in parentheses). However, IV-RESET
fails to reject at 5% when age controls are included (columns 2 and 4). The Hausman test for exogeneity is
constructed as the Wald test for whether the residuals from the first-stage regression have a zero coefficient
when augmented to the reduced-form equation for log monthly earnings. We reject the exogeneity of Good
English for men in column 2, but not for women in column 4. Regressions are clustered by birth cohort, and all
estimated coefficients are statistically significant at 1 % (as indicated by ***). Males were born years 1949–
1999 and range in age from 18 to 64 at the time of the survey; females were born years 1954–1999 and range
in age from 18 to 59 at the time of the survey
Journal of Labor Research
Author's personal copy
Overall, the regressions in Table 4indicate large wage effects from English proficiency.
Consider the 2SLS estimates for males and females in (respectively) Columns 2 and 4.
First, observe that the Good English estimates indicate that men (women) with ‘good’or
‘very good’English earn about 83% (50%) more than those with ‘no English’or just
‘conversational knowledge’. Second, the English Index estimates suggest that those with
‘conversational knowledge’of English have wages about 36% (22%) higher than those
with ‘no ability’in English. Moreover, those with ‘very good’knowledge of English enjoy
a wage premium of better than 100% (65%). Finally, the results for both language skill-
indicators suggest that the wage return for women is about two-thirds that for men.
As a caveat, from Column 4, for women the Hausman (1978) test does not reject that
our English proficiency variables are exogenous. This implies that the local average
treatment effects are not statistically significantly different from the average treatment
effects estimated by OLS. As a conservative estimate, this suggests that women with
‘good’or ‘very good’English earn about 37% more than those with ‘no English’or just
‘conversational English’, rather than the 50% obtained from 2SLS. From Column 2, for
men the Hausman test does not reject the English Index to be exogenous, though there
does appear to be some type of endogeneity problem for Good English.
Although more recent studies find wage returns of similar magnitude to our own in
some cases, our estimates are larger than those found in previous studies that focus on
data for the late 1990s and early 2000s. Overall, we find that for workers in Poland in
the years from 2013 to 2017 (nearly a decade and a half following its entry into the
EU), the ability to communicate well in English appears to have had a substantial
monetary return. Conversely, in our sample where a large majority of workers have
good or very good English skills, there is a substantial wage penalty for those with low
or no proficiency in that language.
Robustness Checks
For our reported estimates above, we have assumed that IV is a valid instrument, one
that is uncorrelated with the disturbance term in the wage equation. In this way, the
exclusion restriction is satisfied so that the coefficient on the instrument (IV) in the
wage equation is exactly zero. However, as is well known, a violation of the exclusion
restriction will render both the traditional OLS and 2SLS estimators inconsistent. This
issue is especially relevant here because of the vast array of political and economic
changes associated with the transition to a market economy occurring in Poland after
1989, which corresponds to the period captured by our IV variable. Following van
Kippersluis and Rietveld (2018), we first employ a simple reduced-form test of whether
the exclusion restriction is indeed violated. We also re-estimate the 2SLS regressions
using the local-to-zero (LTZ) procedure developed by Conley et al. (2012) to determine
the effect of English skills on wages when the exclusion restriction is relaxed.
For the reduced-form regressions, we look only at those individuals with no-foreign-
language (NFL) proficiency. We restrict our sample in this way because any statistically
significant correlation between wages and IV in this sample must be due to factors other than
knowing English or any other foreign language. In other words, a statistically significant
correlation is evidence of an alternative channel for IV to affect wages –and thus suggestive
of a violation of the exclusion restriction. For males, the estimated coefficient on IVin such a
Journal of Labor Research
Author's personal copy
reduced form regression is 0.0272 with a standard error of 0.0168. For female workers, the
estimated coefficient is −0.0024 with a standard error of 0.0145.
Neither for males nor females is the estimated coefficient statistically significant. In
fact, the point estimate for females is very close to zero. Moreover, with the NFL
sample size for males (females) equal to 32,071 (25,681), it is unlikely that our lack of
significance is due to power considerations. The implication is that our standard 2SLS
estimator is validated, and the estimates from Table 4can be interpreted as causal. As a
caveat, however, the calculated t-statistic for men is about 1.62, and this relatively high
t-statistic is of nontrivial importance for our LTZ results that follow.
Next, we employ the Bayesian method from Conley et al. (2012) to determine the
robustness of our 2SLS estimator. Conley et al. suggest relaxing the exclusion restric-
tion, so that instead of exactly zero, a range of values is allowed for the coefficient on
IV in the wage equation. Using the (version 15) STATA tool PLAUSEXOG (Clarke
2017; Clarke and Matta 2017), we implement their LTZ procedure, again following van
Kippersluis and Rietveld (2018). With Good English as our measure of English
proficiency, for males the LTZ point estimate of αis 0.5899 with a standard error
(clustered by birth year) of 0.1529. For females, the LTZ point estimate of αis 0.5251
with a clustered standard error of 0.0909. For both males and females, the estimates are
highly significant from zero.
In regard to the effect of English skills on wages, the LTZ results suggest that male
compliers do not differ that much from female compliers. For men, those with ‘good’
and ‘very good’English earn wages nearly 60% greater than the wages of those with
‘no knowledge’of English or just a ‘conversation level’of English. For women, the
LTZ estimated wage advantage is about 52%.
By comparison, from Table 4the 2SLS point estimate for women is very close to the
LTZ estimate. This is not, however, true for men. For men, the 2SLS estimate is 0.8266 with
a standard error of 0.1518. Although the LTZ estimate for men is within 2 standard errors of
the 2SLS estimate (so that LTZ and 2SLS are not totally at odds), the reduced-form NFL
regressions do suggest a marginal, positive direct effect of the post-reform period on male
earnings. In this way, the LTZ estimate may be regarded as a more conservative estimate of
the causal effect of ‘good’and ‘very good’English skills on male earnings.
With English Index as our measure of English proficiency, the LTZ point estimate of
αfor men is 0.2552 with a clustered standard error of 0.0580. For women, the LTZ
point estimate of αis 0.2298 with a clustered standard error of 0.0364. Both estimates
are highly significant from zero. Once again, the LTZ coefficient estimate of αfor
women is very close to the 2SLS estimate reported in Table 4, while the male LTZ
coefficient is about two-thirds of the 2SLS point estimate. As with the results for Good
English, the LTZ estimates of αshow little difference between men and women in
terms of the effect of English skills on wages. Our preferred instrumental variables
results are the LTZ estimates for men and the 2SLS estimates for women.
For both men and women, a conversational knowledge of English leads to a greater
than 20% wage advantage over those with ‘no English’.A‘very good’knowledge of
English translates into a wage advantage of nearly 70% for women and 75% for men.
Overall, even allowing for a direct effect of the educational reform IV on the log of
monthly earnings (through LTZ estimation), we still find evidence of a statistically
significant and quantitatively important causal effect of English proficiency on the
wages of male full-time employees. For women, we prefer the 2SLS estimate.
Journal of Labor Research
Author's personal copy
Discussion and Conclusion
The modern acronym ELF (English as a lingua franca) is used to refer to “any use of
English among speakers of different first languages for whom English is the commu-
nicative medium of choice, and often the only option”(Seidlhofer 2011,p.7).Today,
ELF is prominent in international law, politics, diplomacy, media, tourism, business,
scientific research and tertiary education. A leading linguist, Braj Kachru, observes the
many positive effects of the worldwide spread of English:
Competence in English and the use of this language signify a transmutation: an
added potential for material and social gain and advantage. One sees this attitude
in what the symbol stands for; English is considered a symbol of modernization, a
key to expanded functional roles, and an extra arm for success and mobility in
culturally and linguistically complex and pluralistic societies. As if all this were
not enough, it is also believed that English contributes to yet another type of
transmutation: It internationalizes one’s outlook. In comparison with other lan-
guages of wider communication, knowing English is like possessing the fabled
Aladdin’s lamp, which permits one to open, as it were, the linguistic gates to
international business, technology, science, and travel. In short, English provides
linguistic power. (Kachru 1986,p.1)
In continental Europe, the acquisition of English was not the end result of colonial
history or the large-scale immigration of native English speakers. Mainly for economic
and political reasons, however, “English has become the de facto ‘extraterritorial’
lingua franca throughout Europe”(Seidlhofer 2010, p. 355); and continental Europe
has become “one unified multilingual community dependent on English as the medium
with the most utility when and where people do not share greater proficiency in other
languages”(Modiano 2017,p.325).
In this paper, we confirm the economic value of English proficiency for European
workers. By employing very large samples of male and female Polish employees who
responded to an online wage survey during the five-year period from 2013 to 2017, we
find statistically and materially significant wage-returns to knowing English. For our
study, we take advantage of a natural experiment in the 1988 reform of language
instruction in Polish primary schools to create an instrumental variable (that we refer to
as ‘IV’) for English proficiency.
The study of Russian was required of all primary school students up until year 1988.
Students starting language instruction in year 1989, however, were able to alternatively
elect to study English or a Western European language instead. Our instrumental
variable IV has a pronounced positive effect on an individual’s probability of under-
standing English at a high level of proficiency. We are thus confident that our 2SLS
estimates are not asymptotically biased due to a weak instrument. Moreover, the
robustness of our estimates is confirmed by employing a recent econometric method
that relaxes the standard exclusion restriction in 2SLS estimation.
Relaxing the exclusion restriction is particularly important in our foreign-language
study. Consider that since our IV measure identifies those who began compulsory
language instruction in primary school after year 1999, then IV also identifies those
who experienced the entire period of transition to a market economy in Poland. Hence,
Journal of Labor Research
Author's personal copy
it is natural to question whether this variable satisfies the exclusion restriction. To
examine this issue, we rely on the Bayesian method from Conley et al. (2012)to
determine the robustness of our 2SLS estimates.
For neither men nor women does the analysis suggest that IV has a statistically
significant direct effect on wages, though for men the effect is positive with a t-
statistic of 1.62 (and hence close to statistical significance at the 10% level). Using
the Version-15 STATA tool PLAUSEXOG (Clarke 2017; Clarke and Matta 2017),
we implement their LTZ procedure in the manner suggested by van Kippersluis and
Rietveld (2018). This method yields an estimate of the effect of English proficiency
on wages for women that is very close to our 2SLS estimates. For men, however,
the wage effect of English proficiency is much lower than the 2SLS results –though
still statistically significant and quantitatively important. The male wage effect
estimated via the LTZ procedure is only slightly larger than the estimated effect
for women. This gives intriguing evidence on the possibility that the transition
period has been favorable for male earnings but not for the wages of women –and
that part of the estimated effect of English on wages for men reflects this direct
effect of the transition era.
Overall, our estimates of the wage-return to English proficiency are larger than
previous estimates of this wage-return for workers in Europe, including those living in
the transition economies of Central and Eastern Europe, and higher than our OLS
estimates. We use two variables to measure the self-assessed level of English profi-
ciency of the survey respondents. The first, Good English, is a dummy variable
identifying those who describe their ability to understand English as ‘good’or ‘very
good’in comparison with those who have ‘no knowledge’of English or are limited to a
‘conversational level’of proficiency.
Our preferred IV estimates (based on the LTZ procedure for men and 2SLS for
women) show that men with ‘good’or ‘very good’English ability earn wages that are
nearly 60% higher than the reference group, while the wage advantage for women with
‘good’or ‘very good’English is slightly better than 50%. We found, however, that
these IV estimates (that measure the ‘local’average treatment effects) are not statisti-
cally significant from the even more conservative OLS estimates. This suggests that our
relatively large sub-population of ‘compliers’may not be so different from the general
population of all Polish workers. For men, the OLS estimates show that men with
‘good’or ‘very good’English ability earn wages that are about 45% percent higher
than the reference group, while the comparable advantage for women is about 37%.
Our second measure adopts a convention used in other studies to impose a linear
hierarchy on the four levels of proficiency identified in the data. Our English Index
equals 0 for those with ‘no knowledge’, 1 for those with ‘conversational ability’, 2 for
those claiming ‘good’understanding and 3 for those with ‘very good’English skills.
The coefficient estimates on this variable in our preferred specification is .26 for men
and .23 for women, which implies that men (women) with ‘conversational ability’earn
26% (23%) more than those with ‘no English’, and men (women) with ‘very good’
English proficiency have a wage premium of over 78% (69%) in comparison to those
with ‘no English’. With both measures we see evidence of a substantial wage enhance-
ment for workers with higher levels of English proficiency. Consistent with the findings
of other studies, our estimates of the wage return for men are larger than the wage
return for women.
Journal of Labor Research
Author's personal copy
Acknowledgements We thank two anonymous referees and the editor for their very helpful contributions to
two major revisions of this paper. The authors are listed in alphabetical order and all contributed equally to the
research reported here.
References
Akresh I, Frank R (2011) At the intersection of self and other: English language ability and immigrant labor
market outcomes. Soc Sci Res 40:1362–1370
Angrist J, Pischke J (2009) Mostly harmless econometrics. Princeton University Press, Princeton
Araújo L, Dinis da Costa P, Flisi S, Soto Calvo E (2015) Languages and employability. European Commission
Joint Research Centre science and policy report
Blanche P, Merino B (1989) Self-assessment of foreign language skills: implications for teachers and
researchers. Lang Learn 39:313–340
Bleakley H, Chin A (2004) Language skillsand earnings: evidence from childhood immigrants. Rev Econ Stat
86:481–496
Bormann S, Ridala S, Toomet O (2019) Language skills in an ethnically segmented labour market: Estonia
1989-2012. Int J Manpow 40:304–327
Card D (1999) The causal effect of education on earnings. In: Ashenfelter O, Card D (eds) Handbook of labor
economics, vol 3. Elsevier, Amsterdam, pp 1801–1863
Card D (2001) Estimating the return to schooling: progress on some persistent problems. Econometrica 69:
1137–11 60
Card D, Lee D (2008) Regression discontinuity inference with specification error. J Econ 142:655–674
Clarke D (2017) PLAUSEXOG: version-15 Stata module to implement Conley et al.'s plausibly exogenous
bounds. Persistent link: https://EconPapers.repec.org/RePEc:boc:bocode:s457832
Clarke D, Matta B (2017) Practical considerations for questionable IVs. MPRA Paper No. 79991, University
Library of Munich, Germany
Conley T, Hansen C, Rossi P (2012) Plausibly exogenous. Rev Econ Stat 94:260–272
Di Paolo A, Tansel A (2015) Returns to foreign language skills in a developing country: the case of Turkey. J
Dev Stud 51:407–421
Di Paolo A, Tansel A (2019) English skills, labour market status and earnings of Turkish women. Economic
Research Center (ERC) Working Paper in Economics No. 19/01 (February, 2019). Forthcoming in
Empirica, vol. 46
DiNardo J, Pischke J (1997) The returns to computer use revisited: have pencils changed the wage structure
too? Q J Econ 112:291–303
Dragemark Oscarson A (2009) Self-assessment of writing in learning English as a foreign language.
University of Gothenburg studies in educational sciences No. 277
Dustmann C, van Soest A (2001) Language fluency and earnings: estimation with misclassified language
indicators. Rev Econ Stat 83:663–674
Dustmann C, van Soest A (2002) Language and the earnings of immigrants. Ind Labor Relat Rev 55:473–492
EC [European Commission] (2012) Europeans and Their Languages –Special Eurobarometer 386
EC [European Commission] / EACEA / Eurydice (2017) Key data on teaching Languages as School in
Europe –2017 Edition. Eurydice Report. Luxembourg Publications Office of the European Union
Gazzola M, Grin F, Wickström B (2016) A concise bibliography of language economics. In: Gazzola M,
Wickström B (eds) The economics of language policy. MIT Press, Cambridge
Ginsburgh V, Prieto Rodriguez J (2011) Returns to foreign languages of native workers in the EU. Ind Labor
Relat Rev 64:599–618
Grin F (2001) English as economic value: facts and fallacies. World Englishes 20:65–78
Grin F (2003) Language planning and economics. Curr Issues Lang Plann 4:1–66
Gunderson M, Oreopoulos P (2010) Returns to education in developed countries. In: Brewer D, McEwan P
(eds) Economics of education. Elsevier, New York, pp 37–43
Hamilton D, Goldsmith A, Darity W (2008) Measuring the wage costs of limited English: issues with using
interviewer versus self-re ports in determining Latino wages. Hisp J Behav Sci 30:257–279
Hausman JA (1978) Specification tests in econometrics. Econometrica 46:1251–1271
Imbens G, Wooldridge J (2007) Instrumental variables with treatment effect heterogeneity: Local Average
Treatment Effects. What’s New in Econometrics, NBER Summer Institute Method Lectures, No. 5
Janowski A (1992) Polish education: changes and prospects. Oxf Stud Comp Educ 2:41–55
Journal of Labor Research
Author's personal copy
Kachru B (1986) The alchemy of English: the spread, functions, and models of non-native Englishes.
Pergamon, Oxford
Komorowska H (2014) Analyzing linguistic landscapes: a diachronic study of multilingualism in Poland. In:
Otwinowska A, De Angelis G (eds) Teaching and learning in multilingual contexts: sociolinguistic and
educational perspectives. Multilingual Matters, Buffalo, pp 19–31
Lang K, Siniver E (2009) The return to English in a non-English speaking country: Russian immigrants and
native Israelis in Israel. The B.E. J Econ Anal Policy 9:Article 50
Liwiński J (2018) The wage premium from foreign language skills. GLO Discussion Paper No. 251. Global
Labor Organization, Maastricht
Modiano M (2017) English in a post-Brexit European Union. World Englishes 36:313–327
Montiel Olea J, Pflueger C (2015) A robust test for weak instruments. J Bus Econ Stat 31:358–369
Oreopoulos P (2006) Estimating average and local average treatment effects of education when compulsory
schooling laws really matter. Am Econ Rev 96:152–175
Oscarson [Oscarsson], M (1984) Self-assessment of foreign language skills: a survey of research and
development work. Council of Europe, Council for Cultural Co-operation, Strasbourg
Pesaran M, Taylor L (1999) Diagnostics for IV regressions. Oxf Bull Econ Stat 61:255–281
Ross S (1998) Self-assessment in second language testing: a meta-analysis and analysis of experiential factors.
Lang Test 15:1–20
Seidlhofer B (2010) Lingua franca English: the European context. In: Kirkpatrick A (ed) The Routledge
handbook of world Englishes. Routledge, London, pp 355–371
Seidlhofer B (2011) Understanding English as a lingua Franca. Oxford University Press, Oxford
Toomet O (2011) Learn English, not the local language! Ethnic Russians in the Baltic States. Am Econ Rev
Pap Proc 101:526–531
van Kippersluis H, Rietveld C (2018) Beyond plausibly exogenous. Econ J 21:316–331
Wang H, Smyth R, Cheng Z (2017) The economic returns to proficiency in English in China. China Econ Rev
43:91–104
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Journal of Labor Research
Author's personal copy
A preview of this full-text is provided by Springer Nature.
Content available from Journal of Labor Research
This content is subject to copyright. Terms and conditions apply.
