ESSAYS ON LABOR AND DEMOGRAPHIC ECONOMICS
By
Hsiu-Fen Hsu

A DISSERTATION
Submitted to
Michigan State University
in partial fulfillment of the requirements
for the degree of
Economics—Doctor of Philosophy
2015

ABSTRACT
ESSAYS ON LABOR AND DEMOGRAPHIC ECONOMICS
By
Hsiu-Fen Hsu
This dissertation contains three self-contained chapters. The first chapter documents
several changes that occurred in wage distribution in Taiwan between 1978 and 2012. For
men, wage inequality narrowed initially, but then started widening. The declining wage
inequality occurred evenly across the entire male wage distribution before the 1990s,
when the economy was growing rapidly. Since the early 1990s, wage inequality among
male workers has been rising, and the growth in inequality has been mainly due to
expansion in upper-tail inequality. Around the same time, an increase in the college wage
premium for male workers is also observed. Using a hybrid DFL reweighting approach,
this study decomposes the changes in wage inequality into three main components:
changes in the skill composition of the workforce, returns to skill, and residuals. The
results show that for male workers, increases in returns to skill that arise from shifts in
demand for skill play an important role in explaining the rising upper-tail wage inequality
in the 1990s. By contrast, for female workers, changes in the skill composition of the
workforce play an important role in explaining rising upper-tail inequality before the
1990s.
The second chapter investigates how children's educational attainment varies by
birth order. In the literature, high-income and middle- and low-income countries have
been shown to have opposite educational outcomes with regard to birth order. Studies
using data from high-income countries usually find that later-born children have an
educational disadvantage; in contrast, studies using data from middle- and low-income
countries find that later-born children have an educational advantage over earlier-born
children. This study, however, finds that birth order–educational attainment patterns in

high-income countries and Taiwan share some similarities: in smaller Taiwanese families,
both later-born boys and girls have an educational disadvantage compared with their
older siblings, a pattern typically found in high-income countries. This birth order pattern
in smaller families also contradicts previous findings that later-born children receive
more education in Taiwan.
The final chapter explores wage behavior over business cycles in Taiwan. The
results show that real wages during the Great Recession are procyclical, whereas real
wages in the recession of the early 2000s are somewhat acyclical. The finding that real
wages are more procyclical in the Great Recession than in the recession of the early
2000s is consistent with that in the U.K. The analysis also finds that the responses of real
wages to cyclical fluctuations in the 2000s are similar among gender, education, and age
groups.

ACKNOWLEDGMENTS

I would like to sincerely thank my advisor, Gary Solon, for his invaluable guidance and
support, both in terms of comments and suggestions on this dissertation and in my
development as an economist. I will be forever grateful for his mentorship and support
that kept me motivated and upbeat through the toughest moments. I am also grateful to
the other members of my guidance committee: Steven Haider, Todd Elder, and Amita
Chudgar. Each of them has provided helpful feedback on my research, and I am indebted
to all of them for their advice. The papers comprising this dissertation also benefited
greatly from helpful comments from other faculty and graduate students at Michigan
State University including John Goddeeris, Scott Imberman, Leah Lakdawala, Jeff
Wooldridge, and Yu-Wei Luke Chu, among many others. Their combined effort has
helped immensely in the completion of this project. All remaining errors are my
responsibility.

iv

TABLE OF CONTENTS
LIST OF TABLES .......................................................................................................... vii
LIST OF FIGURES ....................................................................................................... viii
CHAPTER 1
DECOMPOSING CHANGES IN THE WAGE DISTRIBUTION IN TAIWAN........1
1.1
Introduction ..........................................................................................................1
1.2
Data and Changes in Wage Inequality .................................................................6
1.2.1 The MS and MUS Data............................................................................6
1.2.2 Changes in Wage Inequality ....................................................................9
1.3
Empirical Methodology .....................................................................................12
1.3.1 Hybrid Version of DFL Reweighting Approach ...................................13
1.3.2 Estimation ..............................................................................................20
1.4
Decomposition Results ......................................................................................21
1.4.1 Composition and Wage Structure Effects ..............................................21
1.4.2 The Role of Returns to Skill and the Role of Residuals ........................23
1.5
Conclusion .........................................................................................................24
APPENDICES ................................................................................................................26
APPENDIX A - FIGURES ................................................................................27
APPENDIX B - TABLES..................................................................................30
APPENDIX C - THE MS/MUS DATA, 1978-2012 .........................................32
APPENDIX D - IMPUTATION ........................................................................35
APPENDIX E - ADDITIONAL FIGURES ......................................................38
APPENDIX F - ADDITIONAL TABLES ........................................................40
BIBLIOGRAPHY ..........................................................................................................42
CHAPTER 2
BIRTH ORDER AND EDUCATIONAL ATTAINMENT: EVIDENCE FROM
TAIWAN .........................................................................................................................45
2.1
Introduction ........................................................................................................45
2.2
Background ........................................................................................................47
2.2.1 Theoretical and Empirical Background .................................................47
2.2.2 The Taiwanese Context..........................................................................50
2.3
Data and Methods ..............................................................................................51
2.3.1 Data ........................................................................................................51
2.3.2 Method ...................................................................................................54
2.4
Results ................................................................................................................55
2.4.1 Overall Birth Order Patterns ..................................................................55
2.4.2 Changing Patterns Across Cohorts ........................................................58
2.5
Sensitivity Analysis ...........................................................................................59
2.6
Conclusion .........................................................................................................61
APPENDICES ................................................................................................................63
APPENDIX A - FIGURE ..................................................................................64
APPENDIX B - TABLES..................................................................................65
APPENDIX C - UNREPRESENTATIVE BIRTH ORDER .............................72
BIBLIOGRAPHY ..........................................................................................................74
CHAPTER 3
WAGE ADJUSTMENT OVER THE BUSINESS CYCLE: EVIDENCE FROM
TAIWAN .........................................................................................................................77
v

3.1
3.2
3.3
3.4

Introduction ........................................................................................................77
Data ....................................................................................................................79
Empirical Methodology .....................................................................................82
Empirical Results ...............................................................................................84
3.4.1 Overall Wage Adjustment over the Business Cycle ..............................84
3.4.2 Group Heterogeneity..............................................................................88
3.5
Conclusion .........................................................................................................88
APPENDICES ................................................................................................................90
APPENDIX A - FIGURES ................................................................................91
APPENDIX B - TABLES..................................................................................97
BIBLIOGRAPHY ........................................................................................................102

vi

LIST OF TABLES

Table 1.1

The Hybrid DFL Decomposition Results for Log Wages, Men ................30

Table 1.2

The Hybrid DFL Decomposition Results for Log Wages, Women ........31

Table 1.3

Descriptive Statistics, Men ......................................................................40

Table 1.4

Descriptive Statistics, Women .................................................................41

Table 2.1

Descriptive Statistics: Means and Proportions ........................................65

Table 2.2

Average of Years of Schooling by Family Size and Birth Order ............66

Table 2.3

Fixed Effects Estimates by Family Size (Full Sample) ...........................67

Table 2.4

Fixed Effects Estimates by Family Size (Girls) .......................................68

Table 2.5

Fixed Effects Estimates by Family Size (Boys) ......................................69

Table 2.6

Fixed-Effects Estimates by Birth Cohort (Families with Four or Less
Children) ....................................................................................................70

Table 2.7

Interval Regression Estimates (Pooled Maximum Likelihood Estimates
with Family Means) .................................................................................71

Table 2.8

Fixed-Effects Estimates (Families with Four or Less Children) ...............73

Table 3.1

Mean Log Real Hourly Wages by Gender and Year ...............................97

Table 3.2

Men's Log Real Wages in the Recessions in the 2000s ...........................98

Table 3.3

Women's Log Real Wages in the Recessions in the 2000s ......................99

Table 3.4

Mean Year-to-Year Changes in Log Real Wage by Gender from
Longitudinally Matched MS/MUS Data ................................................100

Table 3.5

Log Real Wage Changes in the Recessions in the 2000s ......................101

vii

LIST OF FIGURES

Figure 1.1

Changes in Wage Inequality ......................................................................27

Figure 1.2

Changes in Log Wage by Percentile ..........................................................28

Figure 1.3

Changes in Log Wage by Percentile ..........................................................37

Figure 1.4

Evolution of Log Wage ..............................................................................38

Figure 1.5

Cyclicality in Measures of Wage Dispersion: 1978-2012 .........................39

Figure 2.1

Average Years of Schooling by Year of Birth ...........................................64

Figure 3.1

Men's and Women's Mean Log Real Wages over the Business Cycle ......91

Figure 3.2

Log Real Wages in the Recessions in the 2000s........................................92

Figure 3.3

Mean Year-to-Year Changes in Log Real Wage by Gender from
Longitudinally Matched MS/MUS Data....................................................93

Figure 3.4

Log Real Wage Changes in the Recessions in the 2000s ..........................94

Figure 3.5

Log Real Wage Changes in the Recessions in the 2000s by Education,
Age, and Sector, Men.................................................................................95

Figure 3.6

Log Real Wage Changes in the Recessions in the 2000s by Education,
Age, and Sector, Women ...........................................................................96

viii

CHAPTER 1
DECOMPOSING CHANGES IN THE WAGE DISTRIBUTION IN TAIWAN

1.1 Introduction
Changes in wage inequality have been one of the topics that have been extensively
researched in labor economics in recent decades. One reason for the interest for this topic
is that the United States witnessed a steep rise in wage and earnings inequality during the
1980s (e.g., Bound and Johnson, 1992; Levy and Murnane, 1992; Murphy and Welch,
1992; Juhn, Murphy, and Pierce, 1993; Katz and Murphy, 1992; Acemoglu, 2002). 1
The change in the wage gap between more and less educated workers has been
recognized as an important part of the change in overall wage inequality. For instance,
Katz and Murphy (1992) and Bound and Johnson (1992) show that the college-high
school wage gap increased sharply in the 1980s after declining during the 1970s. They
propose a simple supply-and-demand explanation of this phenomenon, which later turned
into the skill-biased technological change (SBTC) hypothesis in the literature. 2 However,
this SBTC hypothesis has been challenged in more recent studies that argue that the surge
of wage inequality in the 1980s is largely explained by nonmarket factors such as
minimum wages (Lee, 1999; Card and DiNardo, 2002; Lemieux, 2006). As a result of the
debate, the literature reached a broad consensus—that is, much of the surge of the U.S.
wage inequality in the 1980s appears to be explained by shifts in the supply of and

1

Other Anglo-Saxon countries, such as the United Kingdom (e.g., Gosling, Machin, and
Meghir, 2000) and Canada (e.g., Boudarbat, Lemieux, and Riddell, 2006), witnessed a
similar expansion in wage and earnings gaps in the 1980s.
2
It is argued in the literature that much of the rise in the U.S. wage inequality evident in
the 1980s reflects an ongoing and secular increase in the demand for skill, which is
primarily due to skill-biased technological change resulting from the onset of the
computer revolution, and slowdown in the growth of the relative supply of
college-equivalent workers during the 1980s (Katz and Murphy, 1992; Autor, Katz, and
Krueger, 1998; Card and Lemieux, 2001; Acemoglu, 2002).
1

demand for skills, combined with nonmarket factors (Autor, Katz, and Kearney, 2008).
Further, recent research has increasingly focused on changes in inequality at
different parts of the wage distribution, in addition to the traditional overall wage
inequality (i.e., the standard deviation and 90-10 gap of wages). For example, Autor, Katz,
and Kearney (2008) show that in the United States, upper-tail inequality, measured as the
90-50 wage gap, increased during the 1990s at roughly the same pace as in the 1980s,
whereas lower-tail inequality (the 50-10 gap) has been falling or flat since the late
1980s. 3 Understanding changes across the entire distribution of wages helps deepen our
understanding of the shifting nature of wage inequality, as we know that changes in
inequality at different parts of the wage distribution tend to be explained by different
factors. 4
While the trend of changes in wage inequality in the United States has been
extensively studied, research on the trend of changes in wage inequality in Taiwan is
relatively limited. Furthermore, most previous studies focus on the evolution of measures
of overall wage inequality, such as the Theil coefficient or wage differentials by
education (Gindling and Sun, 2002; Lin and Orazem, 2003; Vere, 2005; Huang, 2011;
Chen, 2013). 5 No existing works explore how changes in inequality at different parts of
the wage distribution evolve over time, while few studies have investigated the

3

In part motivated by the divergent development of the path of upper-tail and lower-tail
inequality, Acemoglu and Autor (2011) refine the skill-biased technological change
hypothesis, the leading explanation of wage inequality, and propose a task-based model
to explain the change in wage inequality.
4
For example, Lemieux (2008) suggests several explanations for the continuing growth
in upper-tail inequality in the United States. One of the suggested explanations is
performance pay, which is considered to play a role in explaining the top-end wage
inequality. On the other hand, the minimum wage is ruled out as an explanation for
growing top-end inequality, but it is considered to be an important explanation for
lower-tail wage inequality.
5
Chen (2002) examines the impact of minimum wages on the change in wage structure.
The measure of inequality used in his paper is Gini and Theil coefficients, which are the
measure of overall inequality.
2

underlying factors driving changes in wage inequality. 6 In addition, exploration of
changes in the wage distribution after the late 1990s is limited. 7
The case of Taiwan is interesting given the enormous changes in the country over
recent decades (e.g., expansion in higher education, trade liberalization, growth in the
information and communication technology (ICT) industry). The expansion of higher
education is expected to increase the relative supply of skilled workers and mitigate
future inequality that stems from increases in relative demand for skilled workers due to
factors such as changes in industry structure. Although increasing the proportion of
college attendees is considered to be an appropriate public policy response to the
phenomenon of increasing inequality, education may have a secondary effect on wage
inequality. In other words, increasing the supply of high-educated workers exerts a
pressure that decreases the wages of those workers. By contrast, if high-educated workers
experience greater within-group wage dispersion, their increased supply may also lead to
an increase in wage inequality. 8,9 The net result of the so-called composition and price
effects is certainly an empirical question, which constitutes the primary motivation for
this paper.
This paper adds to the literature on wage distribution in several ways. First, to the
best of my knowledge, it is the first paper to document the evolution of wages and
6

The one existing paper that attempts to analyze the underlying factors is Vere (2005).
He proposes an approach to decompose the Theil coefficient, a measure of overall wage
inequality, but he does not analyze how the factors considered affect inequality at
different parts of the wage distribution, and his data only cover the years 1979 to 1998.
7
Vere (2005) studies overall wage inequality, but his data do not cover the 2000s. Huang
(2011) and Chen (2013) extended the time period to 2008 and 2011, respectively.
However, both studies focus on wage differentials by skills, such as education and
experience.
8
Lemieux (2006) points out that without any underlying changes in market prices for
skill, changes in the skill composition of the workforce can raise or lower wage
inequality simply by altering the employment share of workers with higher or lower
within-group wage dispersion.
9
It is also well known that composition effects tend to obscure the procyclicality of the
level of real wages (Solon, Barsky, and Parker, 1994).
3

changes in inequality at different parts of the wage distribution in Taiwan over a long
period (the 35 years from 1978 to 2012). A full understanding of the changes that have
occurred in wage distribution requires disentangling the effect of changes in the skill
composition of the workforce (i.e., composition effects) from the effect of changes in
returns to observed skill (i.e., price effects). As a second and methodological contribution,
this paper extends the hybrid version of the DFL (DiNardo, Fortin, and Lemieux (1996))
reweighting framework proposed by Lemieux (2002) to explore how these composition
and price effects contribute to the observed changes in inequality at different parts of the
wage distribution.
The original DFL reweighting approach proposed by DiNardo, Fortin, and Lemieux
(1996) can be used to semi-parametrically decompose changes in wage inequality into
two components: composition effects and wage structure effects. The limitation of this
approach, however, is that wage structure effects include not only the contribution of
returns to observed skill but also the contribution of residuals. To overcome this
limitation, Lemieux (2002) suggests a simple approach to extend the DFL reweighting
method. The main advantage of the hybrid DFL reweighting approach proposed by
Lemieux (2002) is that it allows one to decompose changes in wage inequality into three
components: composition effects, the role of returns to observed skill and the role of
residuals.
However, one of the limitations of this hybrid approach is that the composition
effects identified depend on a given model of wages and thus they may be sensitive to the
choice of the functional forms of wages. 10 To overcome this issue, I modify Lemieux's
(2002) decomposition procedure. In the first step, I compute the composition effects
semi-parametrically by using the original DFL reweighting approach. In the second step
wage structure effects are further decomposed into the role of returns to observed skill

10

This limitation can be seen clearly in Table 2 in Lemieux (2002).
4

and the role of residuals in the spirit of the hybrid approach.
This paper sets out to document changes in the distribution of wages in Taiwan from
1978 to 2012 for male and female workers using microdata from the Manpower Survey
(MS) and its May supplement, the Manpower Utilization Survey (MUS). The data reveal
that there have been several changes in the wage distribution over these 35 years. First,
for male workers, the wage distribution was narrowing before the 1990s, and the decline
in inequality occurred evenly across the entire wage distribution. Together with the fact
that all parts of the wage distribution grew dramatically during the same period, which
was the fruit of the rapid economic growth, the declining wage inequality suggests that
Taiwan was "growing together." 11
Second, during the 1990s, it is found that the changes in wage inequality reversed
direction. Wage inequality among male workers started rising; the increase in overall
wage inequality was mainly due to an increase in inequality at the upper tail of the male
wage distribution. Around the same time an increase in the college wage premium for
male workers is also observed, suggesting that rising returns to college education may
have contributed to the widening of the upper half of the male wage distribution. Finally,
while the picture of trends in female wage inequality is less clear, the impression is that
there was a substantial increase in overall wage inequality before the 1990s, which was
driven by rising upper-tail inequality.
The decomposition results show that for male workers, returns to skill (price effects)
play an important role in explaining rising upper-tail wage inequality during the 1990s.
They explain about half of the increase in the 90-50 log wage differential for male
workers during this decade. As a substantial increase in returns to college education for
male workers is observed during the same period, the decomposition results thus suggest
that returns to college education play an important role in explaining the rising male
11

A similar "growing together" was once observed in the United States, from the close
of World War II to the 1970s (Goldin and Katz, 2007).
5

upper-tail inequality. Furthermore, together with the fact that the skill of the workforce
increased in the 1990s, rising returns to skill suggest that the shifts in labor market
demand must have outpaced those in supply.
For female workers, the decomposition results show that positive composition
effects (i.e., changes in the skill composition of the workforce) play an important role in
explaining the rise in inequality at the upper tail of the female wage distribution before
the 1990s.
This paper is organized as follows. Section 1.2 provides a brief data description and
documents the evolution of the wage distribution in Taiwan from 1978 to 2012. Section
1.3 describes the hybrid DFL decomposition method. Section 1.4 reports the
decomposition results. The last section summarizes the empirical findings and draws
conclusions.

1.2

Data and Changes in Wage Inequality

1.2.1

The MS and MUS Data

In this section, I briefly describe the data and how they are processed. Data and
measurement issues are discussed in more detail in Appendix C.
The data sets are from the Manpower Survey (MS) and its May supplement, the
Manpower Utilization Survey (MUS). The MS is conducted by the Directorate General of
Budget, Accounting and Statistics (DGBAS) of Taiwan, and its survey design is similar to
the U.S. Current Population Survey (CPS). This paper uses the MS and the MUS
microdata for the years 1978 to 2012.
The MS is given to approximately 20,000 registered households, and there are
nearly 60,000 civilians aged 15 and older in these sampled households. The MS is
conducted monthly in the week right after the reference week covering the 15th of a
month. The MS asks detailed questions on hours worked during the previous week (i.e.,
6

the reference week) and jobs held, in addition to demographic information such as
educational attainment. The MUS is the May supplement to the MS, and it asks about
monthly earnings at an individual's main job, which are a "point-in-time" measure of
earnings in general.
Following most of the literature, this paper uses the hourly wage rate. As discussed
in Lemieux (2006), the hourly wage rate pertains to theories of wage determination and
represents the price of labor; by contrast, weekly, monthly, or annual earnings reflect the
responsiveness of labor supply to changes in the price of labor. Since one of the goals of
this paper is to examine the contribution of returns to skill, the hourly wage rate is the
most appropriate measure. The earnings reported in the MUS, however, are monthly. To
deal with this issue, I follow the procedure used in Card and DiNardo (2002) and
Lemieux (2006) to convert monthly earnings into an hourly wage rate. 12
There are a number of problems in the MS/MUS data, which also occur in many
data sets used in the literature, such as the CPS. For one thing, the earnings of highly paid
workers are top-coded at certain values. A second problem is that some workers refuse to
answer the earnings questions in the MUS. To deal with these earnings issues, I impute
the missing and top-coded earnings. A final problem is that some workers do not report
hours worked in the previous week. To deal with this issue, the missing hours are imputed
as well. The results reported in this paper are based on the imputed earnings obtained by
the censored normal regression imputation method and the imputed hours of work
obtained by the normal linear regression imputation method. Details about the
imputations can be found in Appendix D.
I keep workers aged 15-64 with positive earnings and potential experience.
Following Card and DiNardo (2002), I use data on hourly wages for all workers—men

12

For example, Lemieux (2006) computes an hourly wage rate by dividing usual weekly
earnings by usual hours of work. Details about how the hourly wage rate is constructed in
this paper can be found in Appendix C.
7

and women, full-time and part-time workers. Details about the sample selection criteria
can be found in Appendix C.
Wage measures used for all analyses are log real hourly wages. 13 The MS/MUS
sample weights are used throughout the empirical analysis. In the main analysis, I pool
several years of data together to improve the precision of the estimates. The whole
sample period (1978 to 2012) is divided into three periods. The first, second, and third
periods run from 1978-80 to 1990-92, from 1990-92 to 2000-02, and from 2000-02 to
2010-12, respectively. These three periods are used to capture the changes that occurred
during the 1980s, 1990s, and 2000s.
Tables 1.3 and 1.4 in Appendix F present the summary statistics for male and female
workers, respectively. The average age and potential experience for male workers
increased from 35 and 20 in 1978-80 to 40 and 21 in 2010-12. The changes in age and
experience

are

relatively

limited

for

male

workers—an

increase

of

14%

((40-35)/35≈0.143) in age and 5% in experience. By contrast, there were greater changes
in the female workforce over the 35 years studied. For female workers, the average age
and potential experience increased from 27 and 13 in 1978-80 to 37 and 18 in
2010-12—that is, an increase of 37% in age and 38% in experience.
The average years of schooling for male workers increased from 9 in 1978-80 to 13
in 2010-12; the proportion of workers with at least 16 years of schooling (i.e., college or
above) increased from 0.08 in 1978-80 to 0.28 in 2010-12. For female workers, the
change in the proportion of workers with at least 16 years of schooling is more dramatic.
It increased from 0.05 in 1978-80 to 0.33 in 2010-12.
Tables 1.3 and 1.4 also present the changes in log real hourly wages at the mean. For
both gender groups, average log real hourly wages rose rapidly in the first period
(1978-80 to 1990-92) by about 0.79. The 10th, 50th, and 90th percentiles of the log wage
13

Wages are deflated to 2011 real New Taiwanese dollars (NT$) using the Consumer
Price Index (CPI).
8

distribution also experienced remarkable wage growth, as shown in Figure 1.4 in
Appendix E. In the second period (1990-92 to 2000-02), the wage growth at the means
and the 10th, 50th, and 90th percentiles decelerated for both genders. For example, the
average log wage for men and women increased by about 0.22 and 0.34, respectively. In
the final period (2000-02 to 2010-12), the average log wage for men declined by about
0.07 while that for women increased slightly, by about 0.02.
In summary, there are two main messages from these two tables. First, both male
and female workers became more educated and experienced. In other words, the
workforce became more skilled. The other message is that the female workforce
experienced larger changes in the distribution of observed skills than the male workforce.

1.2.2

Changes in Wage Inequality
In this section, I describe the major changes in wage inequality in Taiwan from 1978

to 2012. I focus on three inequality measures: overall wage inequality, summarized by the
90-10 wage gap; inequality in the upper and lower halves of the wage distribution,
summarized by the 90-50 and 50-10 wage gaps (which are referred to as upper-tail and
lower-tail inequality); and the college-high school wage premium. 14 In order to reduce
measurement error, I use the three-year average of the measure of wage inequality. For
example, the 90-10 gap for 1980 plotted in Figure 1.1 is the arithmetic average of the
gaps for 1979, 1980, and 1981. Each measure is computed separately by gender.
The 90-10 Gap of Log Wages. Figure 1.1 displays the evolution of the 90-10 gap of
log wages from 1978 to 2012. For men, the 90-10 gap of log wages declined in the first
period (1978-80 to 1990-92) and the main decline occurred between the mid-1980s to the

14

The 90-10 gap is the difference between the 90th and 10th percentiles of log real
hourly wage and so on. The college-high school wage gap is the difference between the
log real hourly wage of workers with at least 16 years of schooling and that of workers
with 12 years of schooling conditional on experience.
9

early 1990s. The decline in inequality was followed by growth in inequality in the second
period (1990-92 to 2000-02). The acceleration of inequality, however, did not last. A
moderate decline was observed again in the final period (2000-02 to 2010-12). Over the
35 years studied, the overall wage inequality for male workers experienced ups and
downs, and then reverted to an inequality level close to that of the early 1980s, which was
relatively high. For women, the picture of the changes in the 90-10 gap is less clear. One
impression is that female overall wage inequality rose substantially in the first period,
from 1978-80 to 1990-92.
The Top Versus the Bottom. Figure 1.1 shows that from 1978-80 to 1990-92, both the
male 90-50 and 50-10 log wage differentials declined. As shown in Figure 1.4 in
Appendix E, during the same time period, the 10th, 50th, and 90th percentiles of the log
wage distribution also experienced remarkable wage growth. This suggests that Taiwan
was "growing together" during this period. However, the situation reversed dramatically
in the subsequent period, 1990-92 to 2000-02. It is observed that there was a divergence
in the development of upper-tail and lower-tail inequality—that is, the male 90-50 gap
rose while the male 50-10 gap experienced almost no changes. This finding indicates that
the expansion of the overall wage gap in the 1990s was due to rising upper-tail inequality.
Figure 1.2 illustrates the changes in the development of the lower and upper ends of
the wage distribution in a different manner. It shows the change at each percentile of the
log real hourly wage distribution for the three periods by gender. The changes shown are
those between 3-year averages. The graphs are smoothed in order to minimize the effect
of measurement error and facilitate the visual interpretation.
From 1978-80 to 1990-92, the lower end of the male wage distribution experienced
greater wage growth, but the upper end of the male wage distribution had smaller wage
growth. This difference in wage growth across the wage distribution led to the decline in
male wage inequality. For female workers, the finding is opposite.
The situation reversed dramatically in the subsequent period, 1990-92 to 2000-02. In
10

this period, we can observe that the lower half of the male wage distribution continued to
have a similar degree of wage growth, while the upper half of the male wage distribution
experienced higher wage growth—and the higher the percentile, the greater the wage
growth. As a result, male upper-tail wage inequality expanded sharply during this period,
while lower-tail wage inequality changed little. The lower end of the female wage
distribution (below the 20th percentile) had higher wage growth, leading to declining
overall and lower-tail wage inequality.
The rise in male overall wage inequality observed in 1990-92 to 2000-02 may be in
part due to differences in macroeconomic conditions, since the unemployment rate in
2000-02, 4.2 percent, is greater than the unemployment rate of 1.6 percent in 1990-92.15
However, the fact that the 50-10 log wage differential did not exhibit much change during
this period is somewhat surprising, since recessions are typically believed to have a
particularly adverse impact on the bottom end of the distribution of wages (e.g., the 10th
percentile). As it turns out, wage inequality does not exhibit much of a cyclical pattern.
This can be seen in the case of the 90-50 and 50-10 log wage differentials for both men
and women in Figure 1.5 in Appendix E. This suggests that the role of macroeconomic
conditions in explaining the key inequality changes documented in this paper is relatively
limited.
College-High School Log Wage Gap. Figure 1.1 plots the college/high-school log
wage differential for the years 1978 to 2012 by gender. It plots estimated coefficients of a
college education dummy. These estimates are obtained by regressing log real hourly
wage on a dummy for college education and an unrestricted set of dummies for years of
experience. The regression models are estimated separately by gender and year using
samples of people with either a high school diploma or a college degree (observations of
those with a postcollege degree are also used). Survey sample weights are used in the
15

There were two recessions in the 2000s. One was in 2001-02 and the other was in
2008-09.
11

regressions.
For men, the college/high school log wage differential rose by 0.1 over the 35 years
and the main increase, about 0.13, occurred between the mid-1990s and the early 2000s.
For women, the college wage premium rose slightly over the 35 years. Further, unlike the
United States, where the college/high school log wage gap has been trending upward
since the 1980s, the gap in Taiwan has actually declined since the early 2000s for both
gender groups.
In summary, the data from the MS/MUS reveal several changes in the wage
distribution over the 35 years studied. First, for male workers, wage inequality was
decreasing between 1978-80 and 1990-92, and this decline in wage inequality occurred
evenly across the entire wage distribution. Second, it is found that the decline in male
wage inequality was followed by a sharp rise from 1990-92 to 2000-02, and this growth
in wage inequality was mainly due to an increase in inequality at the upper tail of the
male wage distribution. This rising male upper-tail inequality coincided with an increase
in the college/high school wage gap for male workers (see Figure 1.1), suggesting that
rising returns to college education may have contributed to the widening of the upper half
of the male wage distribution. Finally, for female workers, the patterns of change in wage
distribution are less clear, but an impression is that overall and upper-tail inequality rose
substantially before the 1990s.

1.3 Empirical Methodology
The data discussed in the previous section reveal different patterns of change in wage
distribution in Taiwan over the 35 years studied. They also identify changes in the
distribution of skills, in particular for female workers, and a substantial increase in
returns to education for male workers. Are the changes in wage inequality explained by
changes in workforce composition (i.e., the distribution of skills), or do they reflect
12

changes in skill prices?
Why is it important to account for compositional changes in the workforce?
Lemieux (2006) points out that if within-group wage inequality (i.e., wage inequality
among workers with the same characteristics such as education and experience) is higher
for more educated and experienced workers and the employment share of these workers
also increases over time, these shifts in the workforce (i.e., changes in the distribution of
educational attainment or labor market experience of workers in the workforce) will lead
to a rise in wage dispersion even if skill prices do not change. 16 In other words, without
any underlying change in market prices for skills, changes in the skill composition of the
workforce can raise or lower wage inequality simply by altering the employment share of
workers with higher or lower within-group wage dispersion.

1.3.1

Hybrid Version of DFL Reweighting Approach

One of the influential decomposition methods in the literature is the reweighting method
proposed by DiNardo, Fortin, and Lemieux (1996). The advantage of the DFL
reweighting approach is that it provides a simple way to compute composition effects for
any distributional statistics. The decomposition can be easily computed by running a
single probability model (logit or probit) for group membership and using standard
packages to compute distributional statistics with a reweighting factor as weights.
However, one of the limitations of this approach is that it is unable to calculate the
contribution of returns to observed skill, because the wage structure effects computed
include both returns to observed skills (i.e., regression coefficients on observed skills,
such as education and experience) and returns to unobserved skills (or measurement
error). 17 Lemieux (2002) suggests a simple approach to extend the DFL reweighting
16

Lemieux (2006) shows that composition effects play an important role in the rise in
residual wage inequality between 1973 and 2003 in the United States.
17
The wage structure effects are the difference between overall changes and composition
13

method and in the process overcome the above limitation. The approach suggested by
Lemieux (2002) allows one to further decompose changes in wage inequality into three
components: composition effects, the role of returns to observed skill and the role of
residuals (i.e., returns to unobserved skills and measurement error). For instance, changes
in overall wage inequality, the 90-10 wage differential, between times t and s can be
decomposed into
the role of
returns to observed skills
−10
∆90
o
overall changes in
90 -10 wage gap



=

the role of
residuals



−10
−10
−10
∆90
+
∆90
+ ∆90
β
u
x




.
composition effects
wage structure effects

(1.1)

However, one of the limitations of the approach suggested by Lemieux (2002) is that the
composition effects identified depend on a given model of wages and thus they may be
sensitive to the choice of the functional forms of wages. To overcome this issue, I modify
Lemieux's (2002) decomposition procedure. To compute the composition effects that do
not depend on a given model of wages, I use the DFL reweighting approach in the first
step to obtain composition effects semi-parametrically. For instance, the changes in the
90-10 wage gap between times t and s can be decomposed into
−10
∆90
o

=

overall changes in
90 -10 wage gap

-10
∆90
x

+

composition effects

-10
∆90
w

.

(1.2)

wage structure effects

In the second step, I further decompose the wage structure effects into the role of returns
to skill and the role of residuals in the spirit of the hybrid DFL reweighting approach. For
example, the wage structure effects of changes in the 90-10 wage gap between times t
and s can be decomposed into

effects under the DFL approach.
14

−10
∆90
=
w

−10
∆90
β

−10
+ ∆90
u
the role of
residuals

the role of
returns to observed skills

A.

.

(1.3)

The DFL Reweighting Approach and Composition Effects

The composition effects contributing to changes in wage inequality can be computed by
comparing a reweighted distribution of wages in time s and an actual distribution of
wages in time s . For instance, the composition effects of changes in the 90-10 wage
differential between times t and s are computed as follows. 18

[ Q s,.9 − Q s,.1 ]

reweighted
reweighted
−10
] −
∆90
= [ Q s,.
− Q s,.
x
9 
1 



the 90 -10 gap in a
reweighted distribution of wages



the 90 -10 gap in the
actual distribution of wages

reweighted
reweighted
= [ Q s,.
− Q s,.9 ] − [ Q s,.
− Q s,.1 ]
9 
1 


 




the composition effect
in a decomposition of
the 90th percentile

,

(1.4)

the composition effect
in a decomposition of
the 10th percentile

reweighted
where Qs,.
is the 90th percentile of a reweighted (unconditional) distribution of
9

wages in time s and Qs,.9 is the 90th percentile of a (actual and unconditional)
distribution of wages in time s . 19 The reweighted distribution of wages is asking what
the distribution of wages in time s would look like if the workers' skill distribution in
time s were the same as in time t , holding the conditional distribution of wages in time

s fixed. Thus, the difference between Qsreweighted and Qs is explained by changes in
workers’ characteristics (i.e., composition effects).

18

Autor, Katz, and Kearney (2008) and Dustmann, Ludsteck, and Schonberg (2009) use
this aggregate decomposition method (i.e., the DFL reweighting approach) to analyze the
role of composition effects and wage structure effects. They apply the DFL reweighting
procedure to both actual wages and wage residuals.
19
Various statistics from the counterfactual distribution of wages, such as the 10th, 50th,
and 90th percentile, can be computed using actual wages with the reweighting factor as
weights.
15

Constructing the reweighted distribution of wages in time s simply requires
calculating a reweighting factor, ψ (x ) , defined by DFL. 20 The observed (actual)
densities of wages y at times t and s are:

21 22

f y t ( y ) = ∫ f y | x ( y | x )dFx t ( x )
t t

(1.5)

f y s ( y ) = ∫ f y | x ( y | x )dFx s ( x ) .
s s

(1.6)

As DFL shows, the reweighted density that would prevail in time s if workers in time

s had the observed skill distribution (i.e., the distribution of x ) in time t can be
written as
f reweighted ( y ) = ∫ f y | x ( y | x )dFx t ( x ) .
s s
y

(1.7)

s

By manipulating equation (1.7),

f yreweighted ( y ) = ∫ f y | x ( y | x )ψ ( x )dFx s ( x )
s s
s

where ψ ( x ) =

dFx t ( x )
dFx s ( x )

(1.8)

. In practice, the reweighting factor is usually constructed using

a substitution that follows from Bayes' rule:

ψ (x) =

Pr( Dt = 1 | x ) / Pr( Dt = 1)
Pr( Dt = 0 | x ) / Pr( Dt = 0)

20

(1.9)

Survey sample weights are used throughout the paper.
For simplicity, the individual subscript is left out in density functions.
22
Throughout the paper, time s refers to beginning years and time t refers to end years
for each period. For example, time t stands for 1990-92 and time s stands for 1978-80 for
the period 1978-80 to 1990-92.
21

16

where the dummy variable, Dt , is equal to one if time = t and zero if time
= s ; Pr( Dt = 1) is the unconditional probability that an observation is in time t (the share
of the pooled sample that is in time t ); and Pr( Dt = 1 | x ) is the conditional probability
that an observation is in time t conditional on observed skills, x . 23 The reweighting
factor can be computed by estimating a logit model for Pr( Dt = 1 | x ) applied to the
pooled data from times s and t . 24 Using the predicted probabilities, Pˆr( D = 1 | x ) , we
t

then can compute the reweighting factor ψ (x ) for each observation in time s . This
reweighting procedure follows Firpo et al. (2007) and Fortin et al. (2011). 25
The wage structure effects are then computed as the difference between overall
changes and composition effects,
−10
−10
−10
∆90
= ∆90
− ∆90
.
w
o
x

B.

(1.10)

The Hybrid DFL Reweighting Approach and Decomposing Wage Structure Effects

To decompose the wage structure effects into the contribution of returns to observed skill
and the contribution of unobserved skills (i.e., residuals), we need to model the wage
measure. 26 Consider a regression model, 27
yis = x is β s + uis

(1.11)

where i indexes the individual and s indexes time s ; y is the wage measure, x is
a 1 × k vector of covariates (i.e., observed skills such as education and experience) with
23

Pr( Dt = 0) = 1 − Pr( Dt = 1) and Pr( Dt = 0 | x ) = 1 − Pr( Dt = 1 | x )

24

Survey sample weights are used in regressions.
In the Fortin et al. (2011) decomposition, for example, workers in 1983-85 (time s)
were reweighted to look like workers in 2003-05 (time t).
26
Resorting to a parametric model is necessarily restrictive, but this weakness buys some
additional information.
27
Assuming a linear model is necessarily restrictive. However, in practice we would
never know what the true model is and the linear model is the leading case in the
literature.
25

17

x1,is being unity (i.e., a constant), β s is a k × 1 vector of parameters (i.e., returns to

observed skills such as the return to education), and uis is the error term and assumed to
have a zero conditional mean. 28
Applying least squares regression to equation (1.11), the estimated regression
equation is

yis = x is βˆ s + uˆis ,

(1.12)

where βˆ s is the estimate of β s , and uˆis is the regression residual in time s . 29 Under
the assumption made, the estimates represent the effects of explanatory variables on mean
wages. 30
To measure the role of returns to observed skill, I transform each observation of yis
into a counterfactual wage y c by replacing βˆ with βˆ :
is

s

t

yisc = x is βˆ t + uˆis ,

(1.13)

where βˆ t represents estimated returns to observed skill in time t . The counterfactual
wage yisc can now be used to generate a reweighted distribution of the wages that would
have prevailed in time s if returns to observed skill were the same as in time t . As
before, these counterfactual wages are combined with the reweighting factor ψ (x ) to
control also for changes in the skill distribution. Accordingly, this reweighted distribution
of wages is asking what the distribution of wages in time s would look like if the skill
distribution of and returns to skill for workers were the same as in time t .
Utilizing the reweighted distribution of counterfactual wages described above, the
28

The zero conditional mean assumption is restrictive and certainly does not hold in
practice. However, since decomposition methods are an accounting exercise, not "causal
inference," I follow the decomposition literature and maintain this assumption throughout
this paper.
29
The regression equation is estimated separately by year and gender. Survey sample
weights are used in regressions.
30
There is no heterogeneity in the effects as well.
18

wage structure effects can be decomposed into two components: the role of returns to
observed skill and the role of residuals. For example, the role of returns to observed skill
can be obtained by

90 −10

∆β

reweighted , y c
reweighted
− Qs,.
Qs,.9
9 




reweighted , y c
reweighted
− Qs,.1
− Qs,.
1 




] [

=[

the role of returns to skill
in a decomposition of
the 90th percentile

reweighted , y
where Qs,.
9

]

,

(1.14)

the role of returns to skill
in a decomposition of
the 10th percentile

c

is the 90th percentile of reweighted (unconditional) distribution

of counterfactual wages yisc . In other words, the distributional statistics are computed
using reweighted wages yisc with the reweighting factor ψ (x ) as weights. As
described above, Qsreweighted is obtained by using yis with the reweighting factor

ψ (x ) as weights. Thus, the difference between Qsreweighted , y

c

and Qsreweighted is

explained by changes in returns to skill (i.e., β ).
The role of residuals is defined as

,y
reweighted , y
−10
] − [ Qt,.1 − Qs,.reweighted
],
∆90
= [ Qt,.9 − Qs,.
u
9
1
 
c

c

the role of residuals
in a decomposition of
the 90th percentile

(1.15)

the role of residuals
in a decomposition of
the 10th percentile

where Qt,.9 is the 90th percentile of (actual and unconditional) distribution of wages in
time t . Since Qsreweighted , y

c

is obtained by using yisc with the reweighting factor

ψ (x ) as weights, the difference between Qt and Qsreweighted , y
changes in residuals (i.e., u ).

19

c

is explained by

For example, changes in the 90-10 wage gap between 1990-92 (time s ) and
2000-02 (time t ) can be decomposed into 31

]

] [

[

−10
∆90
= Q00/02,.9 − Q00/02,.1 − Q90/92,.9 − Q90/92,.1
o
 

[

90 −10 gap in 00/02

] [

90 -10 gap in 90/92

]

= Q00/02,.9 − Q90/92,.9 − Q00/02,.1 − Q90/92,.1
 
wage changes
at the 90th percentile

wage changes
at the 10th percentile


reweighted , y c  
reweighted , y c 
Q
Q
= Q00/02,.9 − Q90/92
−
−
9
90/92,.1
  00/02,.1

 ,.



−10
∆90
, the role of residuals
u

 reweighted , y c
reweighted   reweighted , y c
reweighted 
+ Q90/92
− Q90/92
− Q90/92,.1
− Q90/92
,.9
,.9
,.1


 

 



[

−10
∆90
, the role of returns to skill
β

][

]

reweighted
reweighted
+ Q90/92
− Q90/92,.9 − Q90/92
−Q
,.9
1
1.
,.
90/92
,.
−10
∆90
, composition effects
x

1.3.2

Estimation

The same set of covariates is used to estimate the wage equation and the logit model used
to construct the reweighting factor. 33 These covariates include four education dummies
and six potential experience dummies. 34,35 I decompose the changes in wage inequality
during three periods: 1978-80 to 1990-92, 1990-92 to 2000-02, and 2000-02 to 2010-12.
The beginning and end years of each period include observations from three years. For
31

The same decomposition procedure can be implemented for the 90-50 and 50-10 wage differentials.

33

The regression equations are estimated separately by year and gender using WLS with
the survey sample weight. I also weight the observations with the survey sample weight
when estimating the logit model.
34
Tables 1.3 and 1.4 in Appendix F give the details of the education and experience
categories used. High school and 20 to 24 years of experience are the omitted categories.
35
Following Lemieux (2006), I only use education and experience as a proxy for
observed skills. This specification allows me to focus on measures of skills, which are
arguably "pure."
20

instance, for the first period 1978-80 to 1990-92, years 1978 to 1980 are pooled for the
beginning years (i.e., time s mentioned above) and years 1990 to 1992 are pooled for
the end years (i.e., time t ). Pooled observations are employed to improve precision, as
the sample size in each year is relatively small. Standard errors are bootstrap estimates,
and bootstrap samples are taken independently within each town in each year.

1.4

Decomposition Results

1.4.1

Composition and Wage Structure Effects

Tables 1.1 and 1.2 report the hybrid DFL reweighting decomposition results for men and
women, respectively. Panel A of each table provides the decomposition results obtained
with the DFL reweighting approach. Overall changes in wage inequality are decomposed
into composition and wage structure effects. Panel B of each table presents the
decomposition results of wage structure effects. In Panel B, wage structure effects are
decomposed into the role of returns to (observed) skill and the role of residuals using the
WLS coefficient estimates of returns to skill. Further, Tables 1.1 and 1.2 report the
decomposition results for men and women, respectively; each table presents the
decomposition results for changes in wage inequality for each of the three periods
mentioned above. The 90-10 log wage differential is used as a measure of overall wage
inequality; the 90-50 and 50-10 log wage differentials are used as measures of upper-tail
and lower-tail wage inequality, respectively.
Starting with the DFL decomposition results in the first period, 1978-80 to 1990-92,
the first three columns of Table 1.1 show that, consistently with the changes observed in
Figures 1.1 and 1.2, male overall wage inequality (the 90-10 gap) declined by about 0.14.
This decline is due to evenly declining inequality at both the upper and lower ends of the
male wage distribution. The 90-50 and 50-10 log wage differentials declined by about
0.07, respectively. All changes in inequality are statistically significant. The
21

decomposition results indicate that the decrease in the 90-50 gap is entirely attributable to
wage structure effects, which account for more than 100% of the net decline, since
composition effects are negligible and go in the opposite direction. For the 50-10 gap,
both composition and wage structure effects contribute to the decline and they are equally
important in explaining the decrease in inequality.
By contrast, the first three columns of Table 1.2 show that female overall wage
inequality increased by about 0.08. This increase is due to rising upper-tail inequality.36
The decomposition results indicate that the increase in overall and upper-tail inequality is
entirely attributable to composition effects, since wage structure effects go in the opposite
direction or are insignificant.
The middle three columns of Table 1.1 present the results for men for 1990-92 to
2000-02. In contrast to the decline in wage inequality in the previous period, the 90-10
gap in this period increased by about 0.09. Although upper-tail and lower-tail inequality
experienced similar decline from 1978-80 to 1990-92, the changes in upper-tail and
lower-tail inequality diverged in the period 1990-92 to 2000-02. While there is no
significant change in the 50-10 gap, the 90-50 gap increased sharply, by about 0.09, and
this increase accounts for all of the increase in the overall inequality. The decomposition
results indicate that wage structure effects (0.084 with s.e.=0.009) account for over 90%
of the increase in the 90-50 gap. For women, the middle three columns of Table 1.2 show
that during the same period, female overall wage inequality declined, and this decline is
due to the compression in the lower end of the female wage distribution, which is
attributable to wage structure effects.
The last three columns of Table 1.1 show that for the final period, male overall wage
inequality declined slightly between 2000-02 and 2010-12. This moderate decline is
attributable to the declining 50-10 gap, which is entirely due to wage structure effects.

36

Changes in the 50-10 log wage differential are negligible and insignificant.
22

There is no significant change in the 90-50 gap during this period since positive
composition effects are offset by negative wage structure effects. Similarly, female
overall wage inequality declined during this period, and this moderate decline is
attributable to declining upper-tail and lower-tail inequality, which is entirely due to wage
structure effects.

1.4.2

The Role of Returns to Skill and the Role of Residuals

The DFL decomposition results discussed above suggest that most of the observed
changes in the 90-10 and 90-50 wage differentials for male workers before the 2000s are
attributable to wage structure effects. However, the wage structure effects computed
using the DFL reweighting approach express the difference between total changes and
composition effects, and thereby include not only the contribution of returns to skill but
also the contribution of residuals. Utilizing the hybrid version of the DFL reweighting
approach, I separate the role of returns to skill from wage structure effects. The results are
presented in Panel B of Table 1.1. 37
Panel B in Table 1.1 shows that the residuals obtained with the WLS estimates
account for about 68% (-0.047/-0.069) of wage structure effects for the male 90-50 gap in
1978-80 to 1990-92. Therefore, they constitute the most important component in
explaining the declining 90-50 wage differential. In the second period, 1990-92 to
2000-02, the decomposition results show that residuals and returns to skill are equally
important in explaining the rising male 90-50 wage gap and returns to skill account for
37

The picture of changes in wage inequality is less clear for female workers, but an
impression is that there is a substantial rise in female overall wage inequality in the 1980s
and the rise is entirely due to the rising upper-tail inequality that is attributable to
composition effects. Therefore, in this section I focus on the decomposition of wage
structure effects for male workers. The decomposition results for female workers can be
found in Panel B of Table 1.2. One thing worth noting is that returns to observed skill
also contribute to the rising female 90-50 wage gap in a nontrivial way in the 1980s (see
Panel B in Table 1.2), but are offset by the residuals.
23

about half of the rising male upper-tail inequality.
In summary, the decomposition results show that for male workers the decline in
wage inequality between 1978-80 and 1990-92 is accounted for by wage structure effects,
in particular for overall and upper-tail inequality. The role of residuals constitutes the
most important component in explaining the declining 90-50 wage differential. This
decline is followed by a rise in the subsequent period, 1990-92 to 2000-02. This rising
male wage inequality is entirely due to the widening upper end of the wage distribution,
which is attributable to wage structure effects.

Returns to skill explain about half of the

rise in the male 90-50 log wage differential in the 1990s. For female workers, the picture
of changes in wage inequality is less clear, but an impression is that there is a substantial
rise in female overall wage inequality in the 1980s and that this rise is entirely due to
rising upper-tail inequality, which is mostly attributable to composition effects.

1.5

Conclusion

This paper documents several changes in the distribution of wages in Taiwan from 1978
to 2012. The data reveal that for male workers, the wage distribution was narrowing
before the 1990s, and the decline in wage inequality occurred evenly across the entire
wage distribution. The decline, however, was followed by a rise during the 1990s, and the
increase in overall wage inequality was mainly due to the widening of the upper half of
the male wage distribution. It is also observed that around the same time there was an
increase in the college wage premium for male workers. The picture of trends in female
wage inequality is less clear, but an impression is that there was a substantial increase in
overall wage inequality, which was driven by the rising upper-tail inequality before the
1990s.
The decomposition results show that for male workers, returns to (observed) skill
24

play an important role in explaining the rise in the upper-tail wage inequality during the
1990s. They explain about half of the increase in the 90-50 log wage differential for male
workers during the 1990s. As a substantial increase in returns to college education for
male workers is observed during the same time period, the decomposition results suggest
that returns to college education play an important role in explaining the rising male
upper-tail inequality.
Together with the fact that the skill of the workforce increased in the 1990s, the
rising returns to skill suggest that the shifts in labor market demand must have outpaced
the shifts in supply. What caused the increase in the relative demand for more skilled
workers between 1990-92 and 2000-02? While more research is needed on this issue, the
timing implies that it may have been the consequence of growth in the information and
communication technology (ICT) industry, which created demand for engineers (skilled
workers).
For female workers, the decomposition results show that positive composition
effects (i.e., changes in the skill composition of the workforce) play an important role in
explaining the rise in inequality at the upper-tail of the female wage distribution before
the 1990s.
It is well known that decomposition is useful for quantifying the contribution of
various factors to changes in wage inequality, but it may not necessarily pin down the
mechanism that explains the relationship between the factors and changes in wage
inequality. For instance, even if decomposition results suggest that changes in returns to
observable skill account for a large fraction of the change in wage inequality, we still
have no information about what force is driving the changes in returns to observable skill.
That being said, quantifying the importance of factors provides useful indications of
particular explanations to be pursued in more detail.

25

APPENDICES

26

APPENDIX A - FIGURES
Figure 1.1
A. 90-10 Wage Inequality

Changes in Wage Inequality
B. College/High School Wage Gap

C. 90-50 and 50-10 Wage Inequality, Male

D. 90-50 and 50-10 Wage Inequality, Female

Notes: For the college/high school wage gap, the figure plots estimated coefficients of a
college education dummy. These estimates are obtained from regressing log real hourly
wage on a dummy for college education and an unrestricted set of dummies for years of
experience. The regression models are estimated separately by gender and year using
samples of people with either a high school diploma or a college degree (observations
with a postcollege degree are also used). Survey sample weights are used in the
regressions.
27

Figure 1.2

Changes in Log Wage by Percentile
A. 1978/80 to 1990/92

Q1990/92,.1 − Q1978/80,.1

Q1990/92,.9 − Q1978/80,.9

B. 1990/92 to 2000/02

28

Figure 1.2 (cont'd)
C. 2000/02 to 2010/12

Notes: Q1990 / 92,.1 − Q1978 / 80,.1 indicates wage changes at
the 10th percentile between 1978-80 and 1990-92.
Changes

in

the

90-10

log

wage

differential

between1978-80 and 1990-92 can be calculated as follows:
-10
∆90
= (Q1990 / 92,.9 − Q1978 / 80,.9 ) - (Q1990 / 92,.1 − Q1978 / 80,.1 )
o

.

29

APPENDIX B - TABLES

Table 1.1

Inequality Measure:
A. DFL Decomposition
Total Change

The Hybrid DFL Decomposition Results for Log Wages, Men

1978/80 - 1990/92
90-10
90-50
50-10

-0.136
(0.012)
Composition
-0.033
(0.011)
Wage Structure
-0.104
(0.015)
B. Decomposition of Wage Structure
Return to Skill (reg. coef.)
-0.062
(0.011)
Residual
-0.041
(0.009)

1990/92 - 2000/02
90-10
90-50
50-10

2000/02 - 2010/12
90-10 90-50 50-10

-0.068
(0.010)
0.002
(0.009)
-0.069
(0.013)

-0.068
(0.007)
-0.034
(0.009)
-0.034
(0.008)

0.092
(0.010)
0.029
(0.009)
0.063
(0.013)

0.092
(0.008)
0.008
(0.006)
0.084
(0.009)

0.000
(0.006)
0.021
(0.007)
-0.021
(0.010)

-0.023
(0.011)
0.077
(0.008)
-0.100
(0.009)

0.005
(0.007)
0.063
(0.009)
-0.058
(0.008)

-0.023
(0.008)
-0.047
(0.008)

-0.040
(0.007)
0.005
(0.005)

0.040
(0.009)
0.023
(0.011)

0.043
(0.006)
0.041
(0.008)

-0.003
(0.006)
-0.019
(0.008)

0.012
(0.012)
-0.112
(0.011)

-0.016 0.028
(0.011) (0.009)
-0.042 -0.070
(0.010) (0.007)

Notes: Bootstrapped standard errors are in parentheses.

30

-0.028
(0.008)
0.014
(0.009)
-0.042
(0.010)

Table 1.2

Inequality Measure:
A. DFL Decomposition
Total Change

The Hybrid DFL Decomposition Results for Log Wages, Women
1978/80 - 1990/92
90-10
90-50
50-10

0.084
(0.013)
Composition
0.123
(0.015)
Wage Structure
-0.039
(0.018)
B. Decomposition of Wage Structure
Returns to Skill (reg. coef.)
0.031
(0.014)
Residual
-0.070
(0.014)

1990/92 - 2000/02
90-10
90-50
50-10

2000/02 - 2010/12
90-10 90-50 50-10

0.088
(0.010)
0.077
(0.014)
0.011
(0.016)

-0.004
(0.010)
0.046
(0.010)
-0.050
(0.010)

-0.032
(0.011)
0.116
(0.011)
-0.148
(0.012)

0.017
(0.008)
0.030
(0.010)
-0.012
(0.010)

-0.049
(0.008)
0.086
(0.005)
-0.136
(0.008)

-0.035
(0.011)
0.165
(0.011)
-0.200
(0.012)

-0.015
(0.008)
0.015
(0.010)
-0.030
(0.010)

-0.019
(0.008)
0.151
(0.005)
-0.170
(0.008)

0.042
(0.014)
-0.031
(0.015)

-0.012
(0.007)
-0.039
(0.008)

0.015
(0.011)
-0.163
(0.012)

0.026
(0.010)
-0.038
(0.011)

-0.011
(0.006)
-0.125
(0.009)

-0.067
(0.011)
-0.133
(0.012)

-0.022
(0.010)
-0.008
(0.011)

-0.045
(0.006)
-0.125
(0.009)

Notes: Bootstrapped standard errors are in parentheses.

31

APPENDIX C - THE MS/MUS DATA, 1978-2012
C1. Measurement Issues and Variable Description
Earnings
In the MUS, monthly earnings were censored at NT$99,999 from 1978 to 1990, and
at NT$999,999 from 1991 to 2012. Further, the data also have a missing data problem.
For some reason, certain workers refuse to answer the earnings questions in the MUS.
Only a handful of workers, however, have top-coded or missing monthly earnings
throughout the sample period.The

DGBAS

does

not

impute

earnings

for

the

non-respondents based on the response for a sample person with similar demographic
characteristics. Instead, 0 is imputed as the earnings for the non-respondents.
The earnings measure comes from the following question in the MUS:
"How much are your monthly earnings at your main job (excluding the earnings at the
secondary job)?"
According to the Report on the Manpower Utilization Survey (May, 2012), the term
"earnings" indicates the profits earned through industrial or commercial activities, net
income of farm workers, and employees’ regular earnings such as salary, bonus,
commission, overtime pay, tips, etc. Irregular pay such as maternity compensation or
children's education subsidies is excluded.
Once again according to the Report on the Manpower Utilization Survey (May,
2012), the timing of the monthly earnings is defined as follows. First, workers with stable
monthly earnings report their last month's earnings. Since the MUS is conducted in the
week right after the reference week covering May 15th, the last month's earnings indicate
the monthly earnings in April. Second, workers who were newly hired or just transferred
to their current job in May report the estimate of their earnings based on the negotiation
with employers. Finally, workers whose monthly earnings are unstable or seasonally
fluctuating report their average monthly earnings. The reported monthly earnings can be
32

viewed as a point-in-time measure of earnings for the first two types of workers.
Hours of work
The measure of hours of work comes from the question in the MS that asks how
many hours a worker worked last week (the reference week). The survey also asks
workers to report hours of work for their main job and secondary job separately. Hours
worked last week at a main job are used as the hours of work measure. Similarly to
earnings, hours of work also have a missing data problem. That is, for some reason,
certain workers do not report hours worked last week.
The estimated hourly wage
To construct hourly wage measures, I first convert the monthly earnings to an
estimated weekly earnings by dividing by 4.29 (≈ 30/7) and then convert the weekly
earnings to an estimated hourly wage by dividing by hours worked last week at a main
job. Since both hours and earnings measures used are for main jobs, the estimated hourly
wage is the wage for a main job. Hourly wages are deflated by Consumer Price Indices
(base year = 2011). Wage measures used for all analyses are log real hourly wages.
Educational attainment
The survey reports each individual's highest educational qualification but not years
of schooling. Furthermore, before 2007 it did not ask whether a worker graduated. It is
therefore possible that the worker may either be in school, a dropout, or a graduate. To
check if there is a large discrepancy between reported highest educational qualification
and actual educational attainment, I use the data from 2007-2012 surveys to calculate the
graduation rate, the dropout rate, and the in-school rate. 1 For the whole sample (all
individuals in the data sets), the graduation, dropout, and in-school rates are about 82%,
5%, and 13% each year, respectively. The in-school rate seems nontrivial. However,
1

Starting in 2007, the MS provides information on whether an individual is either in
school, a dropout, a graduate, or not schooled. In-school rates, dropout rates, and
graduation rates are calculated by dividing by the sum of dropouts, graduates, and those
in-school.
33

when the data are limited to workers (i.e., those with jobs), the in-school rate declines
substantially from 13% to 2%, and the graduation rate increases from 82% to 95%. Since
the latter sample is of interest and the graduation rate is high in this sample, using the
reported highest educational qualification as a proxy for the actual educational attainment
(and skill) is not likely to be a severe issue. 2
Weighting
For the sake of national representativeness, all the analyses use the provided
sampling weights. Several papers, like DiNardo et al. (1996), Lemieux (2006), and Autor,
Katz and Kearney (2008), also weight the observations by weekly hours of work.
However, that type of weighting is not suitable for this paper since the conceptual object
of interest for this paper is the distribution of wages that workers' skills command in the
labor market. Therefore, following Acemoglu and Autor (2011), I only use the provided
sampling weights to calculate statistics.
C2. Sample Selection Criteria
All samples include individuals between the ages of 15 and 64 with positive
earnings and potential experience. 3 Female and male workers (all the analyses are done
separately by gender) and full- and part-time workers are included in the samples.
Following the literature, all samples use wage and salary workers. 4

2

Since the ratios in 2007-2012 are quite constant over this sample period, it would be reasonable to
assume that the graduation rate is also high before 2007.
3
Potential experience is defined as age minus 6 minus the years of schooling.
4
I exclude one observation that reports an unreasonable number of hours of weekly work. The observation
reported 500 hours for hours worked last week, which is greater than the total number of hours in a week,
168 (=24*7), and clearly misreported.
34

APPENDIX D - IMPUTATION
D1.

Imputation of Earnings

As discussed in Appendix C, some workers' earnings are either top-coded or missing. I
impute the earnings by several imputation methods. Since most of the imputation
methods assume normality, I model the log of (nominal) monthly earnings so that
normality is more closely approximated. For each imputation method, log earnings
regression models are estimated separately by year and gender. Each regression model
includes a set of dummies for age, education, industry, and occupation. 1
The first imputation method, labeled "linear," is the normal linear regression
imputation method. The second method, labeled "pmm," is the predictive mean matching
imputation method. The third method, labeled "cnreg_1," is the censored normal
regression imputation method. The final method, labeled "cnreg_2," imputes log earnings
in a manner similar to "cnreg_1," but the log earnings regression models are further
estimated separately for four education groups: primary, high school, junior college, and
college or above. 2 The first and second methods only deal with missing earnings while
the other methods deal with both top-coded and missing earnings. Except for the second
method, the imputation methods all assume normality, and the imputed log earnings are
simulated from the normal distribution.

1

An unrestricted set of age dummies is included. The category of education varies by
year due to the change in the survey design, so the dummies for these variables also vary
by year. For instance, between 1978 and 1987, there are 8 education categories, while
between 1995 and 2012 there are 10 categories. 12 categories for industry and 10
categories for occupation are used.
2
"Primary" denotes 0 to 9 years of schooling, "high school" denotes 12 years of
schooling, "junior college" denotes 14 years of schooling, and "college or above" denotes
16 or more years of schooling.
35

D2.

Imputation of Hours of Work

As discussed in Appendix C, some workers' hours of work are missing. I impute the
hours by using the normal linear regression imputation method. To ensure that normality
is more closely approximated, I model the log of hours of work. Regression models are
estimated separately by year, gender, and class of work. Each regression model uses the
same set of explanatory variables as in the log earnings regression models described
above.
D3. Comparing Wage Inequality for Different Imputation Methods
After imputing the log monthly earnings and log hours of work, I construct log real
hourly wage using the similar procedure described in Appendix C. In addition to the log
real hourly wages calculated based on the imputed earnings and hours, I also construct
another hourly wage measure using the raw data without imputed earnings and hours of
work. This unadjusted log real hourly wage is labeled "drop."
This section compares changes in wage inequality for the different imputation
methods. These changes are calculated using the selected sample. 3 Figure 1.3 illustrates
the changes in the wage distribution. It shows the change at each percentile of the log real
hourly wage distribution for the three periods by gender. As shown in Figure 1.3, the
change in the wage distribution is very similar (or almost identical) in terms of its trends
and levels across alternative imputation methods. The results are not surprising, as the
top-coding and missing data problems are not severe in the MS/MUS data.

3

The sample selection criteria can be found in Appendix C.
36

Figure 1.3

Changes in Log Wage by Percentile

Note: The changes are between 3-year averages.
37

APPENDIX E - ADDITIONAL FIGURES

Figure 1.4

Evolution of Log Wage
A. Men

B. Women

38

Figure 1.5

Cyclicality in Measures of Wage Dispersion: 1978-2012
A. Men

B. Women

39

APPENDIX F - ADDITIONAL TABLES

Table 1.3

Log Hourly Real Wages
Education
Primary
High School
Junior College
College or Above
Potential Experience
<10
10-14
14-19
20-24
25-29
30+
Age
Years of Education
Years of Potential Experience
No. of Observations

1978/80
Standard
Means
Deviation
4.354
0.527

1990/92
Standard
Means
Deviation
5.142
0.456

Descriptive Statistics, Men
Difference in
Means
(90/92 - 78/80)
0.788

2000/02
Standard
Means
Deviation
5.361
0.472

Difference in
2010/12
Means
Standard
Means
(00/02 - 90/92)
Deviation
0.220
5.289
0.464

Difference in
Means
(10/12 - 00/02)
-0.072

0.648
0.208
0.067
0.076

0.477
0.406
0.250
0.265

0.473
0.313
0.116
0.098

0.499
0.464
0.320
0.297

-0.175
0.105
0.049
0.021

0.318
0.356
0.177
0.149

0.466
0.479
0.382
0.356

-0.155
0.043
0.061
0.052

0.203
0.341
0.179
0.277

0.402
0.474
0.383
0.448

-0.115
-0.015
0.002
0.128

0.266
0.152
0.124
0.093
0.084
0.280
35.145
8.826
20.319
34919

0.442
0.359
0.329
0.291
0.278
0.449
12.306
3.889
13.411

0.224
0.184
0.160
0.118
0.100
0.214
36.169
10.382
19.787
40917

0.417
0.388
0.367
0.323
0.300
0.410
10.795
3.506
12.186

-0.042
0.032
0.037
0.025
0.016
-0.067
1.024
1.556
-0.532

0.218
0.158
0.155
0.145
0.124
0.200
37.399
11.703
19.696
43859

0.413
0.365
0.362
0.352
0.330
0.400
10.111
3.148
11.245

-0.006
-0.026
-0.006
0.027
0.024
-0.013
1.230
1.321
-0.090

0.197
0.147
0.142
0.133
0.123
0.257
39.860
12.860
21.000
40874

0.398
0.354
0.349
0.340
0.328
0.437
10.486
3.049
11.749

-0.020
-0.011
-0.013
-0.012
-0.001
0.056
2.460
1.156
1.304

Note: Statistics are weighted by sample weights.

40

Table 1.4

Log Hourly Real Wages
Education
Primary
High School
Junior College
College or Above
Potential Experience
<10
10-14
14-19
20-24
25-29
30+
Age
Years of Education
Years of Potential Experience
No. of Observations

1978/80
Standard
Means
Deviation
3.913
0.466

1990/92
Standard
Means
Deviation
4.699
0.485

Descriptive Statistics, Women
Difference in
Means
(90/92 - 78/80)
0.786

2000/02
Standard
Means
Deviation
5.043
0.470

Difference in
2010/12
Means
Standard
Means
(00/02 - 90/92 )
Deviation
0.344
5.065
0.444

Difference in
Means
(10/12 - 00/02)
0.022

0.636
0.247
0.064
0.053

0.481
0.431
0.245
0.224

0.426
0.372
0.120
0.082

0.495
0.483
0.324
0.275

-0.210
0.125
0.056
0.029

0.245
0.397
0.206
0.152

0.430
0.489
0.405
0.359

-0.181
0.025
0.087
0.070

0.129
0.336
0.200
0.335

0.335
0.473
0.400
0.472

-0.116
-0.060
-0.006
0.183

0.577
0.119
0.068
0.053
0.050
0.134
27.312
8.553
12.760
17448

0.494
0.324
0.251
0.223
0.218
0.340
10.267
4.098
12.157

0.406
0.155
0.122
0.099
0.082
0.135
31.809
10.312
15.497
24659

0.491
0.362
0.328
0.299
0.274
0.342
9.839
3.713
11.949

-0.171
0.036
0.055
0.047
0.032
0.001
4.497
1.759
2.737

0.343
0.155
0.135
0.118
0.102
0.148
34.441
11.904
16.537
28760

0.475
0.362
0.341
0.323
0.302
0.355
9.782
3.175
11.425

-0.064
0.000
0.012
0.019
0.020
0.013
2.632
1.592
1.040

0.278
0.161
0.143
0.127
0.107
0.184
37.479
13.311
18.168
30544

0.448
0.367
0.350
0.333
0.309
0.388
10.055
2.875
11.442

-0.065
0.006
0.008
0.009
0.006
0.036
3.038
1.407
1.631

Note: Statistics are weighted by sample weights.

41

BIBLIOGRAPHY

42

BIBLOGRAPHY
Acemoglu, Daron. 2002. "Technical Change, Inequality and the Labor Market." Journal
of Economic Literature, 40(1): 7–72.
Acemoglu, Daron, and David H. Autor. 2011. "Skills, Tasks and Technologies:
Implications for Employment and Earnings." In Handbook of Labor Economics,
vol.4B, Orley Ashenfelter and David Card, eds. (Amsterdam: Elsevier-North Holland).
Autor, David H., Lawrence F. Katz, and Alan B. Krueger. 1998. "Computing Inequality:
Have Computers Changed the Labor Market?" Quarterly Journal of Economics,
113(4): 1169–1213.
Autor, David H., Lawrence F. Katz, and Melissa S. Kearney. 2008. "Trends in U.S. Wage
Inequality: Reassessing the Revisionists." Review of Economics and Statistics, 90(2):
300–323.
Boudarbat, Brahim, Thomas Lemieux, and Craig W. Riddell. 2006. "Recent Trends in
Wage Inequality and the Wage Structure in Canada." In Dimensions of Inequality in
Canada, David A. Green and Jonathan R. Kesselman, eds. (Vancouver, BC: UBC
Press).
Bound, John, and George Johnson. 1992. "Changes in the Structure of Wages in the
1980s: An Evaluation of Alternative Explanations." American Economic Review, 82(3):
371–392.
Card, David, and John E. DiNardo. 2002. "Skill-Biased Technological Change and Rising
Wage Inequality: Some Problems and Puzzles." Journal of Labor Economics, 20(4):
733–783.
Card, David, and Thomas Lemieux. 2001. "Can Falling Supply Explain the Rising Return
to College for Young Men?" Quarterly Journal of Economics, 116(2): 705–746.
Chen, Lii-Tarn. 2002. "The Impact of Minimum Wages on the Change in Wage Structure
- Evidence from Taiwan's Data." Taiwan Economic Review, 30(3): 383–420. (In
Chinese) doi:10.6277/ter.2002.303.5
Chen, Uniko. 2013. "The Impact of the Expansion in Post-Secondary Education on the
University Wage Premium in Taiwan." Unpublished manuscript, Department of
Economics, University of Michigan, Ann Arbor.
DiNardo, John E., Nicole Fortin, and Thomas Lemieux. 1996. "Labor Market Institutions
and the Distribution of Wages, 1973-1992: A Semiparametric Approach."
Econometrica, 64(5): 1001–1004.
Dustmann, Christian, Johannes Ludsteck, and Uta Schonberg. 2009. "Revisiting the
German Wage Structure." Quarterly Journal of Economics, 124(2): 843–881.
Firpo, Sergio, Nicole M. Fortin, and Thomas Lemieux. 2007. "Decomposing Wage
Distribution using Recentered Influence Functions Regressions." Mimeo, University
of British Columbia.
Fortin, Nicole, Thomas Lemieux, and Sergio Firpo. 2011. "Decomposition Methods in
43

Economics." Handbook of Labor Economics, vol.4A: 1–102.
Gindling, Thomas H., and Way Sun. 2002. "Higher Education Planning and the Wages of
Workers with Higher Education in Taiwan." Economics of Education Review, 21(2):
153–169.
Goldin, Claudia, and Lawrence F. Katz. 2007. "Long-Run Changes in the Wage Structure:
Narrowing, Widening, Polarizing." Brookings Papers on Economic Activity, 2007(2):
135–165.
Gosling, Amanda, Stephen Machin, and Costas Meghir. 2000. "The Changing
Distribution of Male Wages, 1966-1992." Review of Economic Studies, 67(4):
635–666.
Huang, Fung-Mey. 2011. "Human Capital and Wage Dynamics: Evidence from Taiwan."
Taiwan Economic Forecast and Policy, 42(1): 1–37. (In Chinese) http:
econ.sinica.edu.tw
Juhn, Chinhui, Kevin M. Murphy, and Brooks Pierce. 1993. "Wage Inequality and the
Rise in Returns to Skill." Journal of Political Economy, 101(3): 410–441.
Katz, Lawrence F., and Kevin M. Murphy. 1992. "Changes in Relative Wages, 1963-1987:
Supply and Demand Factors." Quarterly Journal of Economics, 107(1): 35–78.
Lee, David S. 1999. "Wage Inequality in the United States in the 1980s: Rising
Dispersion or Falling Minimum Wage?" Quarterly Journal of Economics, 114(3):
977–1023.
Lemieux, Thomas. 2002. "Decomposing Changes in Wage Distributions: A Unified
Approach." The Canadian Journal of Economics, 35(4): 646–688.
Lemieux, Thomas. 2006. "Increasing Residual Wage Inequality: Composition Effects,
Noisy Data, or Rising Demand for Skill?" American Economic Review, 96(3):
461–498.
Lemieux, Thomas. 2008. "The Changing Nature of Wage Inequality." Journal of
Population Economics, 21(1): 21–48.
Levy, Frank, and Richard J. Murnane. 1992. "U.S. Earnings Levels and Earnings
Inequality: A Review of Recent Trends and Proposed Explanations." Journal of
Economic Literature, 30(3): 1333–1381.
Lin, Chun-Hung A., and Peter F. Orazem. 2003. "Wage Inequality and Returns to Skill in
Taiwan, 1978-96." The Journal of Development Studies, 39(5): 89–108.
Murphy, Kevin M., and Finis Welch. 1992. "The Structure of Wages." Quarterly Journal
of Economics, 107(1): 285–326.
Solon, Gary, Robert Barsky, and Jonathan A. Parker. 1994. "Measuring the Cyclicality of
Real Wages: How Important Is Composition Bias?" Quarterly Journal of Economics,
109(1): 1–25.
Vere, James P. 2005. "Education, Development, and Wage Inequality: The Case of
Taiwan." Economic Development and Cultural Change, 53(3): 711–735.
44

CHAPTER 2
BIRTH ORDER AND EDUCATIONAL ATTAINMENT: EVIDENCE FROM
TAIWAN

2.1

Introduction

The ways in which the family environment affects child outcomes have long fascinated
researchers. A number of arguments suggest that siblings are likely to receive unequal
shares of educational resources allocated by their parents, and birth order is one important
dimension of sibling composition that could play a critical role in determining
within-family resource allocation. 1
Various hypotheses in the literature outline potential reasons for why children's
educational attainment varies by birth order; however, the relationship predicted by these
hypotheses between the two variables is ambiguous. Those predicting that educational
attainment decreases as birth order increases are based on the assumptions of: fixed
parental time endowment and a greater share of time endowment received by children of
a lower birth order, mothers are older when they give birth to children of a higher birth
order, and older mothers are more likely to have children with a lower birth weight. 2
Meanwhile, the hypotheses predicting that educational attainment increases with birth
order are based on the assumptions of: younger siblings benefit from the increase in
family income over the life cycle, and the earlier entry of older siblings into the labor
1

Another focus has been on understanding differences between families, and family size
is one of the factors believed to determine inter-family differences. The economics of the
family suggests that there is a negative correlation between family size and children's
education (Becker and Lewis 1973; Becker and Tomes 1976), and extensive empirical
literature has attempted to identify the causal effect of family size and examine whether a
quantity–quality trade-off exists (Angrist et al. 2010; Black et al. 2005; Lee 2008; Li et al.
2008; Maralani 2008; Rosenzweig and Wolpin 1980; Rosenzweig and Zhang 2009).
2
If differences in parental time endowment and birth weight lead to differences in ability
and thereby the return to education, parents will invest differentially in their children's
education.
45

market increases the resources available for the younger ones. The empirical findings on
the relationship in the literature are mixed: studies using data from high-income countries
find that later-born children fare worse, whereas those using data from middle- and
low-income countries find that they fare better. 3
This discrepancy found between developed and developing countries suggests that
birth order patterns may vary by place and/or stages of development. However, it might
also result from the difference in empirical methodologies employed. One of the most
influential recent birth order studies by Black et al. (2005) aims to explore the
relationship between birth order and educational attainment by estimating a family fixed
effects model, with dummy variables for birth order, using administrative data for the
entire Norwegian population. Similarly, this study employs the fixed effects approach to
explore how children's educational attainment varies by birth order in Taiwan, a
middle-income country. 4,5 As discussed in the literature, controlling for family fixed
effects rules out that estimated birth order patterns are driven by differences in observed
family characteristics (e.g., family size and parental education) and unobserved family
characteristics that are shared by siblings. Furthermore, because of Taiwan’s rapid
economic development and social change, family budgets and norms and values may
have varied over time. Exploring this changing pattern across cohorts therefore provides
a unique opportunity to examine how intra-family educational resource allocation
changes as a society develops.
Contrary to the opposing findings in developed and developing countries presented

3

See de Haan et al. (2014) for a detailed summary of birth order studies in developed
and developing countries.
4
Black et al. (2005) used the fixed effects method (within estimator) to control for
family heterogeneity, and used dummy variables to capture potential nonlinear patterns of
birth order. To overcome the challenge that birth order relates to family size, they also
estimated this pattern by family size.
5
As discussed later, studies using data from Taiwan, which find later-born children fare
better, did not take into account the unobserved family heterogeneity.
46

in the literature, this study finds that high-income countries' and Taiwanese birth order
patterns share some similarities: in smaller Taiwanese families, later-born boys and girls
have an educational disadvantage compared with their older siblings, a pattern typically
found in high-income countries. This finding in smaller families also contradicts the
finding of Yu and Su (2006), who show that later-born children receive more education in
Taiwan. 6 In addition, while later-born children have been shown to fare better in the
developing country studies, I find no evidence that later-born children have an
educational advantage, even in larger families. Moreover, taking advantage of a sample
that contains a wide range of birth cohorts, ranging from 1921 to 1978, I found that the
disadvantage of being born later in smaller families is more evident in recent cohorts.
The remainder of the paper is organized as follows: Section 2.2 provides the
theoretical and empirical background to birth order patterns and briefly describes the
Taiwanese context; Section 2.3 discusses the data and introduces the fixed effects
approach; Section 2.4 presents the results; Section 2.5 tests the sensitivity of the main
results using interval regression; and Section 2.6 provides conclusions.

2.2 Background
2.2.1

Theoretical and Empirical Background

There are a number of hypotheses suggesting that birth order might affect educational
investment. As outlined in Strauss and Thomas (1995), the factors that could bring about
differences between siblings due to birth order are time and financial constraints, as well
as biological factors.
First, given that parents have a fixed time endowment, the firstborn will receive a
greater parental time input than later-born children, who have to compete for parental
6

The household survey data used by Yu and Su (2006) are the same as those in this
study.
47

attention. To the extent that greater parental time input translates into higher educational
achievement, firstborns may fare better than later-born children. However, this argument
also serves to emphasize the role of gaps between children; if the age difference between
siblings (spacing) is large, then the youngest child might benefit more as either older
siblings leave the family nest or through the increase in time input, since both parents and
older siblings spend time with the youngest child (Behrman and Taubman 1986; Birdsall
1991; Hanushek 1992).
Life cycle effects can also be important. The younger parents are at their first birth,
the more financially constrained they may be compared with later in their life cycle;
hence, available resources might be lower for the firstborns of young—and possibly
immature—parents. In contrast, later-born children might benefit through the increase in
family income over the life cycle (Parish and Willis 1993).
In addition, biological factors may be important. By definition, mothers giving birth
to children of a higher birth order are older than when they had those of a lower birth
order. To the extent that older mothers tend to have children with lower birth weight, and
who may be less healthy, later-born children may thus fare worse. On the other hand,
parents may learn with practice and experience, and hence, later-born children might
have an advantage over their older siblings.
Furthermore, other factors can work in both directions. Older children may be
encouraged to leave school early to contribute to the family income, giving an advantage
to their younger siblings with regard to educational attainment. On the other hand, if
older children are expected to assume more responsibility in caring for their younger
siblings, this training may lead them to perform more responsibly at school and become
higher achievers.
Finally, cultural and legal factors may play a part as well. If there is land or an estate
to be passed on, and inheritance customs favor the firstborns, parents may choose to
invest more in the formal education of their younger children to compensate. Cultural
48

factors can also work in the opposite direction where firstborns will assume the
responsibility of caring for elderly parents and inherit paternal authority, and thus receive
higher educational investment. The next subsection will discuss further how cultural
factors influence intra-family educational resource allocation in the Taiwanese context.
In summary, educational achievement might be positively or negatively associated
with birth order, depending on the degree to which the above-mentioned factors affect
children who are otherwise similar. Ultimately, it is an empirical question as to which
dominates.
The empirical findings in the literature are mixed. The bulk of the studies on the
birth order patterns in the developed countries found that later-born children have
educational disadvantage compared with their older siblings. For instance, Behrman and
Taubman (1986) find that children of a higher birth order in the U.S. have lower
educational attainment. A recent study by Kantarevic and Mechoulan (2006) finds that
firstborns in the U.S. have higher education and earnings. Black et al. (2005) use a data
set of the entire population of Norway, finding strong evidence for a negative association
between birth order and children's educational attainment and adult earnings. Using the
British Household Panel Survey, Booth and Kee (2009) find that children's education
decreases with birth order.
Whereas there is evidence in high-income countries that children of a higher birth
order tend to do worse in many dimensions, studies using data from middle- or
low-income countries usually find a reversed pattern, showing that later-born children
have an educational advantage over their older siblings. See Parish and Willis (1993) and
Yu and Su (2006) for evidence in Taiwan, Li et al. (2008) for China, Ejrnaes and Portner
(2004) for the Philippines, Edmonds (2006) for Nepal, Dammert (2010) for Nicaragua
and Guatemala, Tenikue and Verheyden (2010) for 12 African countries, and de Haan et

49

al. (2014) for Ecuador. 7

2.2.2

The Taiwanese Context

To briefly introduce the context for this study, Taiwan industrialization began just before
World War II, and the country experienced rapid economic development in the postwar
decades. In 1955, 61% of the workforce was in the agricultural sector, which declined to
20% by 1980, and nearly 10% in the 1990s. The changing industrial structure not only
increased the demand for more educated workers and returns to education for both men
and women but also affected the government’s educational policies. Primary school
education became mandatory for children born after 1945, while children born in 1956
and after further benefited from the extension of mandatory education for an additional
three years.
The Taiwanese literature on educational attainment suggests the importance of
family background, including characteristics such as socioeconomic status and both
parents’ education (Parish and Willis 1993; Tsai et al. 1994). The association between
sibship characteristics and education might differ for sons and daughters if parental
investment strategies depend on a child’s gender: parents under financial constraints
allocate educational funds according to their conscious assessment of sons’ and
daughters’ relative market opportunities and returns (Brinton 1993; Parish and Willis
1993). The lower return to education for women hence leads to daughters having fewer
educational opportunities compared with sons in East Asia (Brinton 1993). In addition,
son preference and Taiwanese family norms with regard to seniority have been
considered critical factors in determining intra-family resource allocation in Taiwanese
society. Yu and Su (2006) found evidence that parental investment strategies are

7

One exception is that Yu and Su (2006) find first-born sons have an educational
advantage. The privilege for firstborns, however, does not extend to daughters.
50

conditioned not only by family budgets but also by culturally defined family norms. They
showed that male firstborns, who are the ultimate inheritors of paternal authority in
Taiwanese families, have additional leverage in the sibling competition for family
resources. Conversely, this firstborn privilege does not extend to daughters; rather, they
have a general educational disadvantage associated with being older children.
Due to Taiwan’s rapid economic development and social change, family budgets
and norms and values may have changed over time. We therefore might expect
intra-family educational resource allocation to vary across cohorts depending on the
stages of economic and social development; hence, exploring cohort differences provides
a unique opportunity to examine how intra-family educational resource allocation
changes as society develops.

2.3 Data and Methods
2.3.1

Data

I use the 1999, 2000, and 2003 waves of the Taiwan Panel Study of Family Dynamics
(PSFD), in which respondents provided information on themselves, their parents and
spouse's parents, their siblings, and their children. The information about siblings and
children includes year of birth, gender, birth order, and educational attainment;
respondents were also asked how many brothers and sisters and how many sons and
daughters they have (including those deceased), from which family sizes were
calculated. 8
The PSFD reports each individual's highest educational qualification, but not years
of schooling; nor were respondents asked whether their siblings or offspring graduated.
Thus, the typical number of years required to complete each educational qualification
8

In the sample, the respondent, their siblings, and parents form one family unit, while
they and their children form another family.
51

was used: a master's degree was coded as requiring 18 years and a doctorate as 21 years
of schooling. 9 One potential problem with this method is that the presumed number of
years may not reflect the actual years of schooling an individual completed. 10 For
example, it is possible that an individual giving high school as their highest qualification
did not actually graduate from high school; therefore, this individual's actual years of
schooling may be between 9 and 12 years. To check whether the coding for years of
schooling affected the estimation results, the sensitivity of the results was tested using
interval regression (see Section 2.5).
Due to the design of the survey, respondents only provided information on their
living siblings and children, 11 which raises a potential problem that the birth order of the
reported siblings and children may not be representative of their actual position in the
family. 12 For example, in a four-child family where one child had died, if the deceased
child was the third-born, then the fourth-born child will be treated as the third-born when
their parents allocate educational resources between the surviving children. 13 The
robustness check presented in the Appendix C confirms that the estimation results, which
follow, were not affected by this limitation.
In this study, the sample of children provides information on three broad cohorts:
those born in 1921–1955 (Cohort 1), 1956–1969 (Cohort 2), and 1970–1978 (Cohort 3).
Table 2.1 presents descriptive statistics for the full sample and each birth cohort: children
in the full sample, around half of whom were female, received 10.89 years of schooling

9

Eighteen years of schooling include: six years of elementary, three years of junior high,
three years of high school, four years of college, and two years of graduate school.
Twenty-one years of schooling includes an additional three years of graduate school.
10
As the survey did not ask whether respondents graduated, using either educational
qualification or years of schooling encounters the problem of measurement error.
11
Respondents only provided demographic information for their five oldest siblings and
children who were alive at the time of the interview.
12
I thank Steven Haider for pointing out this issue.
13
On the other hand, if the third-born child died after all the children had grown up and
completed their education, their death is not likely to be an issue.
52

on average; in 2003, these children were approximately 43 years old; on average, families
had 4.4 children, with approximately 90% of those surveyed having three or more; and
the educational attainment of respondents’ parents was much lower than that of their
children, with mothers having completed 4.41 and fathers 6.47 years.
The average years of schooling and family size have changed across cohorts. The
former, for both children and their parents, has increased steadily: children's average
schooling was about eight years in the first cohort, increasing to almost 13 years in the
third cohort. Meanwhile, the average family size of 5.76 in the first cohort has declined to
3.67 in the third cohort.
Table 2.2 cross-tabulates the average years of schooling by family size and birth order.
The top panel shows that educational attainment decreases with birth order for smaller
families, but the difference is not substantial. Larger families exhibit a different pattern:
educational attainment increases with birth order, and the educational gap between
earlier-born and later-born children is larger. For example, the firstborn child in families
with six children completes 8.61 years of schooling on average, but the sixth child
completes 11.20 years. This positive relationship between educational attainment and
birth order for larger families may be due to birth cohort effects. As shown in Figure 2.1,
which presents the average years of schooling by birth year, there is an upward trend in
years of schooling. For a given family, younger children (with a higher birth order) are
more likely to benefit from the rapid growth in education, particularly children in larger
families. 14 Therefore, it is important to control for the trend in years of schooling when
investigating birth order patterns.

14

Later-born children may receive more education due to changes in education policy,
such as the extension of compulsory schooling. Compulsory schooling in Taiwan was
extended from six to nine years in 1968, so children born in 1956 and later further
benefited from the extension of mandatory education for an additional three years.
53

2.3.2

Method

Consider the following model in which the educational attainment of each child in a
family can be written as

EDij = β0 + D(birthorder)ij β1 + β 2birthyearij + ci + uij

where

(2.1)

is the educational attainment of the child j in family i as measured by

years of schooling. D(birthorder)ij is a 1 × k vector of dummy variables for the
second, third, fourth, fifth, sixth and higher order of birth, leaving the firstborn child as
the omitted category. The advantage of birth order dummies is that they allow us to
capture potential nonlinear patterns of birth order compared to an absolute birth order
variable that takes the value 1 for the first born, 2 for the second born and so on.
birthyearij is the value obtained by subtracting the mean of year of birth, 1960, from

year of birth, and is used to control for the rapid growth in educational attainment across
birth cohorts observed in Figure 2.1. 15,16 ci is the family-specific effect that may be
correlated with the covariates, and absorbs both observed and unobserved family
characteristics that are constant within a family, such as family size, both parents'
education and birth cohort, and mother's age at first birth, which are important factors that
affect the estimation of birth order patterns, as described in Blake (1989) and related
literature. To examine the pattern of birth order net of family characteristics, I employ the
15

Another way of controlling for the rapid growth in educational attainment is to include
year of birth dummies, which capture cohort effects and education policy changes.
However, as depicted in Figure 2.1, the assumption of a linear birth year trend, captured
by birthyearij , seems consistent with the data. Although not reported in this paper,
regressions using year of birth dummies instead of birthyearij were run and yielded very
similar results.
16
Since birth order is mechanically positively correlated with the year of birth and the
year of birth is used to control for the upward trend of educational attainment, excluding
the year of birth from regressions would result in a misleading (positive) birth
order-education relationship.
54

fixed effects (FE) approach (within estimators) to difference out any family-specific
characteristics (i.e., ci ). 17
As discussed in Section 2.2.2, the relationship between birth order and educational
attainment is likely to be different between girls and boys; hence, the full sample was
divided into two subsamples, a girl sample and a boy sample, throughout the estimation.
Because data on multiple children within a family are used in estimations, I use standard
errors (clustered at family level) that are robust to heteroscedasticity and arbitrary
within-family correlation.

2.4 Results
2.4.1

Overall Birth Order Patterns

Tables 2.3–2.5 provide the fixed effects (FE) estimates for the full sample, girl sample,
and boy sample, respectively. Point estimates in Column 1 of Table 2.3 show that there is
a pattern of falling then rising educational attainment with respect to birth order in the
full sample. The estimates for the second, third, fourth, and fifth children, ranging from
-0.222 to -0.391, are negative, with precise estimates being given for the second and third
child. The estimates for the sixth and seventh or later children (0.092 and 0.376,
respectively) are positive but statistically insignificant.
Columns 2–7 in Table 2.3 present separate regressions for particular family sizes,
and indicate that birth order patterns vary somewhat with family size. For smaller
families, the estimates are all negative and monotonically decreasing with birth order,
while for larger families, the point estimates imply that later-born children receive more
education; however, they are typically not statistically significant, even jointly. To
17

The FE method described in this section denotes the within estimator. In other words,
the (observed and unobserved) family heterogeneity is removed by family demeaning.
Applying least squares with dummy variables for family units (i.e., the so-called family
fixed effects) yields estimates that are numerically identical to within estimates.
55

improve the precision, certain classes of family size were pooled to form two groups:
families with four or less children, and families with five or more children.
Columns 8 and 9 present the results for these two groups. In Column 8, all values for
birth order dummies, ranging from -0.256 to -0.774, are precisely estimated, and suggest
that average educational attainment monotonically decreases with birth order in smaller
families: in other words, later-born children receive less education. Pooling the
observations for families with five or more children improves the precision as well, but
the estimates in Column 9 are still generally not significantly different from zero. 18
Estimates of birth order dummy variables for the sixth child and seventh and later child in
Column 9 are positive, but are not statistically significant, even jointly. Furthermore, the
estimates on birth year are all positive and precisely estimated, accurately capturing the
upward trend of educational attainment depicted in Figure 2.1. 19
Table 2.4 provides the results for girls, and shows that the estimation results are
similar to those for the full sample in Table 2.3. 20 Column 8 of Table 2.4 shows that for
girls in smaller families, there is an educational disadvantage associated with being
later-born: birth order dummies, ranging from -0.015 to -1.124, are precisely estimated,
except for the second child, indicating that later-born girls in smaller families fare worse.
For later-born girls in larger families, the estimates are positive but not statistically
significant, even jointly.
The estimation results for boys are shown in Table 2.5. Similar to the girls in smaller
families (see Column 8 in Table 2.4), Column 8 in Table 2.5 shows a monotonic decline
18

The exception is the estimate for the second child.
The exception is the estimate for birth year in 2-child families, which is positive but
statistically insignificant.
20
The observations in Table 2.4 represent the number of girls in each family. For
example, observations in Column 2 are the total number of girls in 2-child families,
including those from families with one girl and one boy. Girls who have a brother in
2-child families do not contribute to estimation due to the FE method (those observations
drop out). In other words, if there is only one girl observation in each family, this
observation drops out from the FE estimation. The same applies to Table 2.5.
19

56

in average education as birth order increases for boys in smaller families, indicating that
later-born boys in smaller families fare worse. Column 9 presents the estimation results
for boys from larger families, showing that the point estimates for birth order dummies
are all negative but not significantly different from zero, except for the second child.
In summary, although the relationship between educational attainment and birth
order has been found to be opposite for developed and less developed countries in the
empirical literature, I find that their birth order patterns share some similarities: later-born
children in smaller families have an educational disadvantage, regardless of their gender.
In addition, while later-born children have been shown to fare better in the developing
country studies, I find no evidence that later-born children have an educational advantage,
even in larger families. The disadvantage associated with being later-born in smaller
families also contradicts the finding of Yu and Su (2006), who show that later-born
children receive more education in Taiwan. 21 This discrepancy in the findings probably
results from the difference in model specifications, estimation methods, and statistical
inference. Yu and Su (2006) use an absolute birth order variable to capture birth order
patterns, and include dummy variables for firstborn sons and daughters in some
specifications to capture the privilege for firstborns. By contrast, I follow Black et al.
(2005) and use a set of birth order dummies to capture potential nonlinear patterns of
birth order. In addition, I utilize the fixed effects approach (i.e., within estimators) to
remove potential biases resulting from unobserved family heterogeneity, an issue not
taken into account in Yu and Su (2006). Moreover, their statistical inference relies on the
homoscadesticity assumption, which tends to be too strong and usually does not hold in
practice. In contrast, I use standard errors that are robust to both heteroscedasticity and
within-family correlation.

21

The household survey data used in Yu and Su (2006) are the same as those in this
study.
57

2.4.2

Changing Patterns across Cohorts

Because of Taiwan’s rapid economic development and social change, family budgets and
norms and values may have varied over time. Therefore, we might expect intra-family
educational resource allocation to vary across cohorts, depending on the stages of
economic and social development. Taking advantage of the sample that contains a wide
range of birth cohorts, ranging from 1921 to 1978, I explore how birth order patterns in
smaller families have changed across cohorts.
To obtain the birth order pattern for different cohorts, the interaction between birth
order dummies and three birth cohort dummies were examined: Cohort 1 (1 if born
1921–1955 and 0 otherwise), Cohort 2 (1 if born 1956–1969 and 0 otherwise), and
Cohort 3 (1 if born 1970–1978 and 0 otherwise). I then applied the fixed effects method
(within estimators) to a regression of a child's years of schooling on second child*cohort1,
second child*cohort2, second child*cohort3, third child*cohort1, third child*cohort2,
third child*cohort3, fourth child*cohort1, fourth child*cohort2, fourth child*cohort3,
and year of birth. 22,23 Table 2.6 presents the results for girls and boys. 24
The first three columns of Table 2.6 present the results for children from families
with four or less children, with the results for girls in the middle three columns and boys
in the last three columns. The estimates for birth order dummies are generally negative
(significant for most cases) and the magnitude of the estimates become larger for younger
cohorts, implying that being later-born children in a family is more of a disadvantage in
22

The omitted category is firstborns. Using the firstborns in the oldest cohort (Cohort 1,
born 1921–1955) as the base group (i.e., included in the model first child*cohort2 and
first child*cohort3) yielded similar results, but the resulting estimates are less precise. In
addition, the estimates for first child*cohort2 and first child*cohort3 are not significant at
any traditional levels, even jointly.
23
Although not reported in this paper, regressions using year of birth dummies instead of
a linear birth year trend were run and yielded similar but less precisely estimated birth
order patterns across cohorts.
24
The estimate, 0.307, in the first cell of Table 2.6 is the coefficient estimate of second
child*cohort1(born 1921–1955).
58

younger than older cohorts. 25 This pattern might reflect emerging educational inequality
within a family. Furthermore, the changing birth order pattern across cohorts in smaller
families also implies that the behavior of Taiwanese families increasingly resembles that
of families in developed or high-income countries such as the U.S.

2.5

Sensitivity Analysis

The PSFD reports each individual's highest educational qualification, but not years of
schooling; nor were respondents asked whether their siblings or offspring graduated.
Thus, the typical years of schooling required to complete each educational qualification
were used. As discussed in Section 2.3, one potential problem of the presumed years of
schooling is that the presumed years may not reflect the actual years of schooling an
individual completed. In this section, to check whether the coding of years of schooling
affects the estimated birth order patterns already shown in Section 2.4, I conduct a
robustness check using interval regression.
Following Wooldridge (2010, Chapter 19), the Chamberlain-Mundlak device was
employed to control for the family heterogeneity, ci , in equation (2.1):
ci = γ + D(birthorder)i δ + λ birthyeari + ωi

(2.2)

where D(birthorder)i is a vector of family means of birth order dummies and
birthyeari is the family mean of birthyearij . By construction, ωi is uncorrelated with

the explanatory variables. Substituting equation (2.2) into equation (2.1),
25

The joint F test for (second child*cohort1 = second child*cohort2), (third
child*cohort1 = third child*cohort2), (fourth child*cohort1 = fourth child*cohort2)
indicates that the difference in birth order dummy variables between cohorts 1 and 2 is
jointly significant for the full sample but not for the girl and boy samples. Similar joint
tests for the difference in birth order dummy variables between cohorts 1 and 3 suggest
the difference is jointly significant for the full and boy samples but not for the girl
sample.
59

EDij = α0 +1 D(birthorder)ij β + β 2birthyearij + D(birthorder)i δ + λ birthyeari + ηij ,

(2.3)

where ηij = ωi + uij . Interval regression is applied to equation (2.3), and the parameters
on birth order dummies contain the partial effects of interest, which permit a direct
comparison with the FE estimates presented in Section 2.4. 27
As discussed in Section 2.3.1, due to the survey design, the actual years of schooling
completed by an individual were not available, and so the ”usual” number of years of
schooling were used in the estimations. To run interval regression, two variables,
upperij and lowerij , were specified for each individual, which represent the interval

endpoints and thereby determine the interval into which an individual's actual years of
schooling fall.
For an individual whose reported highest educational qualification was elementary
school, lowerij and upperij were set to 0 and 6, respectively; for an individual whose
reported highest educational qualification was junior high school, lowerij and upperij
were set to 6 and 9, respectively, and so on. For an individual who received zero
schooling, lowerij was set to missing and upperij was set to 0. In other words, the
intervals were ED ≤ 0 , 0 < ED ≤ 6 , 6 < ED ≤ 9 , 9 < ED ≤ 12 , 12 < ED ≤ 16 , and 16 < ED .
The results in Table 2.7 show that the magnitude and statistical significance of
interval regression estimates are close to those of the FE estimates reported in Tables
2.3–2.5. The robustness check of the results to alternative interval regression showed that
the birth order patterns were not affected by the presumed number of years of schooling
and the estimator chosen.

27

Note that we are still interested in the regression of years of schooling on birth order
dummies. The structure of interval regression (i.e., the log likelihood) looks a lot like that
of the ordered probit model, but there is an important difference: in ordered probit, the
cut points are parameters to estimate and the coefficient estimators do not measure
interesting partial effects. With interval regression, the interval endpoints are given and
coefficient estimators contains the partial effects of interest.
60

2.6 Conclusion
This study uses data from the Taiwan Panel Study of Family Dynamics to explore how
children's educational attainment varies by birth order. The relationship between
educational attainment and birth order predicted by several hypotheses is ambiguous, and
the empirical findings in the literature are mixed. While empirical studies using data from
high-income countries find that later-born children fare worse, studies using data from
middle- and low-income countries find the opposite is true.
Despite the divergent findings in developed and developing countries, I find that
birth order–education patterns in high-income countries and Taiwan share some
similarities. For instance, in smaller Taiwanese families, once a linear birth year trend is
accounted for, later-born boys and girls have an educational disadvantage compared with
their older siblings, a pattern typically found in high-income countries. 28 Furthermore,
the birth order pattern found in smaller families contradicts the findings of Yu and Su
(2006), and other researchers, that later-born children fare better in Taiwan. In addition,
while later-born children have been shown to fare better in the developing country studies,
I find no evidence that later-born children have an educational advantage, even in larger
families.
Moreover, taking advantage of a sample that contains a wide range of birth cohorts,
ranging from 1921 to 1978, I found that the disadvantage of being later-born children in
smaller families is more evident in recent cohorts, a changing pattern which might reflect
emerging educational inequality within a family. Furthermore, the changing birth order
pattern across cohorts implies that the behavior of Taiwanese families might increasingly
resemble that of families in high-income countries.
As discussed in Section 2.2, educational achievement might be positively or
28

The birth order pattern shown in Black et al. (2005) was estimated by controlling for
indicators for age in 2000. Therefore, the pattern they found is the one net of birth year
trend as well.
61

negatively associated with birth order, depending on the degree to which the various
factors affect children who are otherwise similar; ultimately, it is an empirical question as
to which dominates. For smaller families, the data used in the current study suggest that
the hypotheses predicting the negative relationship between birth order and educational
attainment takes precedence. With limited information, however, we remain less clear on
precisely which factor was driving the birth order pattern observed in smaller families.
Clearly, more research is needed. By providing a description of the different birth
order patterns in smaller and larger families and across cohorts, this study will guide
further research into the underlying factors driving these patterns, particularly in smaller
families, which may be more representative of contemporary household behavior and
consequently merit further attention.

62

APPENDICES

63

APPENDIX A - FIGURE

Figure 2.1

Average Years of Schooling by Year of Birth

Note: Author's calculation based on the sample.

64

APPENDIX B - TABLES

Table 2.1

Descriptive Statistics: Means and Proportions

Years of schooling
Birth order
Female
Year of birth
Age in 2003
Number of children

10.89
2.68
0.50
1960
42.79
13,549

Cohort 1
1921—1955
8.36
2.77
0.50
1947
56.39
4,340

Family size (number of children)
Family with 2 children
Family with 3 children
Family with 4 children
Family with 5 children
Family with 6 children
Family with 7 or more children
Father's age in 2003
Mother's age in 2003
Father's years of schooling
Mother's years of schooling
Number of families

4.41
0.28
0.24
0.16
0.11
0.05
0.12
70.47
66.38
6.47
4.41
3,672

5.76
0.08
0.15
0.22
0.22
0.11
0.30
83.48
80.39
4.51
2.75
1,277

Full Sample

65

Cohort 2
1956—1969
11.60
2.69
0.50
1963
39.82
5,932

Cohort 3
1970—1978
12.97
2.55
0.48
1973
30.13
3,277

4.34
0.27
0.28
0.19
0.13
0.04
0.08
69.97
65.78
6.49
4.80
2,300

3.67
0.39
0.29
0.13
0.05
0.01
0.02
61.35
56.56
7.75
6.90
1,654

Table 2.2

Average Years of Schooling by Family Size and Birth Order

All families

Birth order

Mean
1
2
3
4
5
6
7 or more
N

Mean
1
2
3
4
5
6
7 or more
N

Mean
1
2
3
4
5
6
7 or more
N

10.89
10.92
11.10
11.07
10.87
10.28
10.17
9.51
13,549

10.40
10.36
10.65
10.64
10.33
9.64
9.73
8.82
6,704

11.38
11.47
11.53
11.52
11.36
10.94
10.53
9.99
6,845

2
12.98
13.02
12.92

580

Family size
3
4
All children
12.59 11.94
12.63 11.57
12.60 11.97
12.51 12.12
12.17

2,757

3,228

12.52
12.39
12.70
12.47

Girls
11.80
11.41
11.94
12.01
11.97

221 1,067

1,578

12.54
12.36
12.79

13.25
13.44
13.00

12.63
12.77
12.55
12.54

Boys
12.07
11.76
12.00
12.22
12.30

359

1,690

1,650

Note: N represents the number of children observations.

66

5

6

10.85
10.25
10.32
10.95
11.44
11.62

9.70
8.61
8.86
9.31
10.25
10.55
11.20

2,594

2,138

10.58
10.12
10.03
10.68
11.31
11.31

9.29
8.41
8.77
8.87
9.97
9.94
10.89

1,435

1,153

11.18
10.45
10.73
11.33
11.57
11.88

10.17
8.88
8.99
9.97
10.58
11.15
11.39

1,159

985

7 or more
7.97
6.87
7.31
7.72
8.22
8.50
9.06
9.51
2,252

7.23
5.89
6.44
7.21
7.38
7.96
8.77
8.82
1,250

8.88
8.03
8.56
8.41
9.30
9.29
9.37
9.99
1,002

Table 2.3
Dependent variable:
years of schooling
Second child
Third child
Fourth child
Fifth child
Sixth child
Seventh or later child
Birth year
Constant

F (joint test of birth
order dummies)
R-squared
Observations
Number of families

All
2-child
families
families
(1)
(2)
-0.271***
-0.094
(0.067)
(0.274)
-0.391***
(0.102)
-0.280*
(0.143)
-0.222
(0.195)
0.092
(0.252)
0.376
(0.411)
0.175***
0.132
(0.018)
(0.089)
11.054*** 12.152***
(0.075)
(0.499)
8.96***
0.086
13,549
3,672

0.023
580
325

Fixed Effects Estimates by Family Size (Full Sample)
7+ children Families with 4 or Families with 5 or
families
less children
more children
(7)
(8)
(9)
-0.233
-0.256***
-0.324***
(0.216)
(0.084)
(0.113)
-0.169
-0.568***
-0.216
(0.277)
(0.143)
(0.149)
0.041
-0.774***
0.012
(0.333)
(0.217)
(0.192)
0.056
-0.07
(0.421)
(0.250)
0.289
0.263
(0.505)
(0.316)
0.749
0.567
(0.695)
(0.476)
0.174***
0.155***
0.186***
0.167***
(0.042)
(0.039)
(0.027)
(0.023)
10.480*** 9.640***
11.526***
10.481***
(0.455)
(0.672)
(0.100)
(0.245)

3-child
families
(3)
-0.436***
(0.132)
-0.782***
(0.241)

4-child
families
(4)
-0.066
(0.123)
-0.363*
(0.191)
-0.707**
(0.275)

5-child
families
(5)
-0.404**
(0.162)
-0.207
(0.223)
-0.159
(0.299)
-0.371
(0.398)

6-child
families
(6)
-0.307
(0.222)
-0.308
(0.286)
0.178
(0.371)
0.115
(0.478)
0.377
(0.597)

0.178***
(0.048)
11.718***
(0.247)

0.199***
(0.036)
11.286***
(0.087)

0.176***
(0.037)
11.107***
(0.191)

6.05***

2.97**

2.45**

2.42**

1.06

5.30***

4.30***

0.013
2,757
1,026

0.039
3,228
890

0.085
2,594
594

0.153
2,138
405

0.124
2,252
432

0.025
6,565
2,241

0.121
6,984
1,431

Notes: Firstborns are the omitted category. Robust standard errors (in parentheses) clustered in family. *** p<0.01. **
p<0.05. * p<0.1.

67

Table 2.4
Dependent variable:
years of schooling
Second child
Third child
Fourth child
Fifth child
Sixth child
Seventh or later child
Birth year
Constant

F (joint test of birth
order dummies)
R-squared
Observations
Number of families

All
families
(1)
-0.045
(0.105)
-0.222
(0.151)
-0.049
(0.214)
-0.056
(0.287)
0.425
(0.379)
0.286
(0.586)
0.210***
(0.026)
10.433***
(0.114)

7+ children Families with 4 or Families with 5 or
families
less children
more children
(7)
(8)
(9)
-0.010
-0.015
-0.148
(0.289)
(0.148)
(0.147)
-0.200
-0.619**
-0.114
(0.384)
(0.254)
(0.190)
0.158
-1.124***
0.231
(0.477)
(0.410)
(0.252)
0.262
0.159
(0.607)
(0.330)
0.719
0.696
(0.733)
(0.429)
0.621
0.625
(0.974)
(0.636)
0.190***
0.292***
0.188***
(0.055)
(0.050)
(0.029)
9.090***
10.669***
9.843***
(0.931)
(0.194)
(0.290)

2-child
families
(2)
-0.549
(0.751)

3-child
families
(3)
-0.094
(0.298)
-0.466
(0.514)

4-child
families
(4)
0.047
(0.172)
-0.549*
(0.289)
-1.058**
(0.444)

5-child
families
(5)
-0.396*
(0.208)
-0.104
(0.291)
0.057
(0.393)
-0.317
(0.545)

6-child
families
(6)
0.045
(0.284)
-0.101
(0.320)
0.474
(0.446)
0.473
(0.566)
1.038
(0.742)

0.577**
(0.241)
9.335***
(1.187)

0.234**
(0.106)
11.049***
(0.551)

0.288***
(0.055)
10.633***
(0.158)

0.213***
(0.050)
10.569***
(0.195)

0.164***
(0.048)
9.666***
(0.473)

0.54

4.29***

2.12*

1.54

1.280

4.96***

2.64**

0.071
1,067
798

0.096
1,578
813

0.160
1,435
558

0.202
1,153
390

0.209
1,250
403

0.100
2,866
1,806

0.190
3,838
1,351

2.55**
0.166
6,704
3,157

Fixed Effects Estimates by Family Size (Girls)

0.481
221
195

Notes: Firstborns are the omitted category. Robust standard errors (in parentheses) clustered in family. *** p<0.01. **
p<0.05. * p<0.1.

68

Table 2.5
Dependent variable:
years of schooling
Second child
Third child
Fourth child
Fifth child
Sixth child
Seventh or later child
Birth year
Constant
F (joint test of birth
order dummies)
R-squared
Observations
Number of families

All families
(1)
-0.451***
(0.112)
-0.488***
(0.169)
-0.609**
(0.237)
-0.421
(0.324)
-0.629
(0.390)
-0.446
(0.642)
0.151***
(0.029)
11.692***
(0.121)

2-child
families
(2)
0.136
(0.435)

3-child
families
(3)
-0.593***
(0.194)
-0.832**
(0.331)

4-child
families
(4)
-0.237
(0.203)
-0.376
(0.287)
-0.807**
(0.381)

5-child
families
(5)
-0.538*
(0.283)
-0.061
(0.414)
-0.513
(0.537)
-0.291
(0.693)

-0.18
(0.118)
14.437***
(0.674)

0.157**
(0.065)
11.972***
(0.341)

0.153***
(0.045)
11.806***
(0.118)

0.122**
(0.061)
11.638***
(0.430)

4.77***

1.69

1.85

0.56

0.55

4.30***

0.99

0.015
1,690
951

0.017
1,650
815

0.056
1,159
536

0.110
985
371

0.079
1,002
389

0.012
3,699
2,032

0.078
3,146
1,296

3.99***
0.049
6,845
3,328

Fixed Effects Estimates by Family Size (Boys)

0.032
359
266

6-child
families
(6)
-0.48
(0.375)
-0.559
(0.565)
-0.503
(0.717)
-0.818
(0.952)
-1.167
(1.124)

7+ children Families with 4 or Families with 5 or
more children
families
less children
(7)
(8)
(9)
-0.248
-0.453***
-0.426**
(0.388)
(0.130)
(0.198)
-0.307
-0.593***
-0.265
(0.438)
(0.208)
(0.272)
0.035
-0.905***
-0.323
(0.539)
(0.316)
(0.344)
0.164
-0.289
(0.691)
(0.450)
-0.254
-0.500
(0.805)
(0.535)
0.223
-0.337
(1.130)
(0.804)
0.230***
0.116*
0.143***
0.151***
(0.085)
(0.064)
(0.037)
(0.041)
11.967*** 10.315***
12.014***
11.308***
(1.015)
(1.161)
(0.131)
(0.492)

Notes: Firstborns are the omitted category. Robust standard errors (in parentheses) clustered in family. *** p<0.01. **
p<0.05. * p<0.1.

69

Table 2.6
Dependent variable:
years of schooling
Second child
Third child
Fourth child
Birth year
Constant

R-squared
Observations
Number of families

Fixed Effects Estimates by Birth Cohort (Families with Four or Less Children)

Cohort 1
1921—1955
0.307
(0.220)
0.268
(0.315)
-0.105
(0.670)

Full sample
Cohort 2
1956—1969
-0.292***
(0.102)
-0.482***
(0.154)
-0.568 **
(0.246)
0.177***
(0.027)
11.551***
(0.099)

Cohort 3
1970—1978
-0.319***
(0.111)
-0.705***
(0.164)
-0.847***
(0.223)

Cohort 1
1921—1955
0.345
(0.363)
0.456
(0.551)
0.038
(0.689)

0.029
6,565
2,241

Girls
Cohort 2
1956—1969
-0.110
(0.188)
-0.721***
(0.278)
-0.828*
(0.479)
0.276***
(0.050)
10.724***
(0.193)
0.108
2,866
1,806

Cohort 3
1970—1978
0.072
(0.197)
-0.535*
(0.297)
-1.303***
(0.409)

Cohort 1
1921—1955
0.169
(0.309)
-0.001
(0.390)
-0.423
(1.014)

Boys
Cohort 2
1956—1969
-0.457***
(0.162)
-0.402*
(0.237)
-0.764**
(0.352)
0.136***
(0.037)
12.028***
(0.131)

Cohort 3
1970—1978
-0.626***
(0.181)
-0.836***
(0.243)
-0.890***
(0.325)

0.018
3,699
2,032

Notes: Firstborns are the omitted category. The estimates are obtained by applying the fixed effects method (within
estimators) to a regression of child's years of schooling on second child*cohort1, second child*cohort2, second
child*cohort3, third child*cohort1, third child*cohort2, third child*cohort3, fourth child*cohort1, fourth child*cohort2,
fourth child*cohort3, and year of birth for girls and boys, respectively. For instance, the estimate, 0.307 (SE=0.220), is
the coefficient estimate of second child*cohort 1(1921–1955). Robust standard errors (in parentheses) clustered in
family. *** p<0.01. ** p<0.05. * p<0.1.

70

Table 2.7
Dependent variable:
years of schooling
Second child
Third child
Fourth child
Fifth child
Sixth child
Seventh or later child
Birth year
Constant
Total observations
Left-censored observations
Uncensored observations
Right-censored observations
Interval observations

Interval Regression Estimates (Pooled Maximum Likelihood Estimates with Family Means)

All
(1)
-0.335***
(0.076)
-0.469***
(0.115)
-0.345**
(0.161)
-0.274
(0.221)
0.069
(0.286)
0.364
(0.467)
0.208***
(0.020)
9.156***
(0.270)
13,549
588
0
468
12,493

Full sample
Girls
(2)
-0.094
(0.119)
-0.283
(0.173)
-0.115
(0.243)
-0.112
(0.329)
0.485
(0.436)
0.127
(0.678)
0.248***
(0.030)
9.305***
(0.157)
6,704
413
0
143
6,148

Families with 4 or less children
All
Girls
Boys
(4)
(5)
(6)
-0.310***
-0.038
-0.544***
(0.094)
(0.165)
(0.146)
-0.646*** -0.634** -0.762***
(0.159)
(0.283)
(0.232)
-0.864*** -1.125** -1.154***
(0.241)
(0.451)
(0.353)

Boys
(3)
-0.524***
(0.126)
-0.602***
(0.189)
-0.761***
(0.264)
-0.524
(0.362)
-0.797*
(0.437)
-0.628
(0.718)
0.184***
(0.032)
10.059***
(0.156)
6,845
175
0
325
6,345

0.214***
(0.031)
9.940***
(0.291)
6,565
85
0
347
6,133

0.310***
(0.056)
9.803***
(0.174)
2,866
47
0
100
2,719

0.180***
(0.041)
10.573***
(0.190)
3,699
38
0
247
3,414

Families with 5 or more children
All
Girls
Boys
(7)
(8)
(9)
-0.392***
-0.204
-0.465**
(0.131)
(0.169)
(0.229)
-0.263
-0.167
-0.282
(0.172)
(0.222)
(0.307)
-0.013
0.172
-0.363
(0.220)
(0.292)
(0.386)
-0.108
0.082
-0.296
(0.288)
(0.386)
(0.508)
0.258
0.731
-0.565
(0.364)
(0.502)
(0.604)
0.57
0.435
-0.407
(0.548)
(0.746)
(0.903)
0.204*** 0.233*** 0.183***
(0.026)
(0.035)
(0.046)
6.441*** 8.649*** 8.758***
(0.706)
(0.410)
(0.375)
6,984
3,838
3,146
503
366
137
0
0
0
121
43
78
6,360
3,429
2,931

Notes: Firstborns are the omitted category. Robust standard errors (in parentheses) clustered in family. Interval values are 0, 6,
9, 12, and 16. Each model includes family means of birth order dummies and birth year.*** p<0.01. ** p<0.05. * p<0.1.

71

APPENDIX C - UNREPRESENTATIVE BIRTH ORDER
To check the sensitivity of estimates in respect of unrepresentative birth order, the sample
is divided into two subsamples: one including deceased siblings and the other without
deceased siblings. This sample stratification is conducted for smaller families (families
with four or less children). 1
The estimation results are presented in Table 2.8, in which Column 1 replicates the
results in column 8 in Table 2.3, while Columns 2 and 3 present the results for smaller
families with and without deceased siblings, respectively. The estimates in Column 3
(families without deceased siblings) are very similar to those in Column 1. Furthermore,
the sign of the estimates on birth order dummies are the same for both Columns 2 and 3.
The magnitude of these estimates is smaller in families with deceased siblings (Column
2), but this difference is not statistically significant. Columns 4 through 6 and 7 through 9
present the results for the girl and boy samples, respectively, and are similar to that for
the full sample. Overall, the robustness check shows that the birth order patterns in
smaller families were not affected by the data limitations.

1

Since the survey only asks respondents to report on up to five living siblings or
offspring, it is impossible to define a subsample for larger families that includes all
children in a given family.
72

Table 2.8

Dependent variable:
years of schooling
Second child
Third child
Fourth child
Birth year
Constant
Observations

All
(1)
-0.256***
(0.084)
-0.568***
(0.143)
-0.774***
(0.217)
0.186***
(0.027)
11.526***
(0.100)
6,565

Fixed Effects Estimates (Families with Four or Less Children)
Full sample
With
deceased
sibling
(2)
-0.144
(0.220)
-0.449
(0.416)
-0.335
(1.129)
0.164**
(0.082)
11.175***
(0.726)
992

Without
deceased
sibling
(3)
-0.275***
(0.091)
-0.587***
(0.153)
-0.807***
(0.229)
0.189***
(0.029)
11.619***
(0.082)
5,573

Girls
With
Without
All
deceased deceased
sibling
sibling
(4)
(5)
(6)
-0.015
0.232
-0.039
(0.148)
(0.496)
(0.159)
-0.619**
0.064
-0.681***
(0.254)
(1.085)
(0.263)
-1.124***
-0.620
-1.189***
(0.410)
(1.341)
(0.423)
0.292***
0.190
0.300***
(0.050)
(0.236)
(0.052)
10.669*** 10.570*** 10.852***
(0.194)
(2.145)
(0.152)
2,866
469
2,397

All
(7)
-0.453***
(0.130)
-0.593***
(0.208)
-0.905***
(0.316)
0.143***
(0.037)
12.014***
(0.131)
3,699

Boys
With
Without
deceased deceased
sibling
sibling
(8)
(9)
-0.526
-0.440***
(0.331)
(0.141)
-0.375
-0.592***
(0.518)
(0.223)
-2.649
-0.860***
(2.234)
(0.331)
0.146*
0.141***
(0.084)
(0.040)
11.652*** 12.073***
(0.716)
(0.113)
523
3,176

Notes: Firstborn children is the omitted category. Robust standard errors (in parentheses) clustered in family. ***
p<0.01. ** p<0.05. * p<0.1.

73

BIBLIOGRAPHY

74

BIBLOGRAPHY
Angrist, Joshua, Victor Lavy, and Analia Schlosser. 2010. “Multiple Experiments for the
Causal Link between the Quantity and Quality of Children.” Journal of Labor
Economics 28 (October): 773–824.
Becker, Gary and Gregg Lewis. 1973. “On the Interaction between the Quantity and
Quality of Children.” Journal of Political Economy 81 (March-April): S279–88.
Becker, Gary and Nigel Tomes. 1976. “Child Endowments and the Quantity and Quality
of Children.” Journal of Political Economy 84 (August): S143–62.
Behrman, Jere R., and Paul Taubman. 1986. “Birth Order, Schooling, and Earnings.”
Journal of Labor Economics 4 (July): S121-45.
Birdsall, Nancy. 1991. “Birth Order Effects and Time Allocation. ” In T. P. Schultz (Ed.),
Research in population economics: A research annual (pp. 191–123). Greenwich: JAI
Press.
Black, Sandra, Paul Devereux, and Kjell Salvanes. 2005. “The More the Merrier? The
Effect of Family Size and Birth Order on Children’s Education.” The Quarterly
Journal of Economics 120 (May): 669–700.
Blake, Judith. 1989. Family size and Achievement. University of California Press,
Berkeley and Los Angeles, CA.
Booth, Alison and Hiau Joo Kee. 2009. “Birth Order Matters: the Effect of Family Size
and Birth Order on Educational Attainment.” Journal of Population Economics 22:
367–97.
Brinton, Mary. C. 1993. Women and the Economic Miracle: Gender and Work in
Postwar Japan. Berkeley: University of California Press.
Darmmert, Ana C. 2010. “Siblings, Child Labor, and Schooling in Nicaragua and
Guatenmala.” Journal of Population Economics 23(1): 199–224.
De Haan, Monique, Erik Plug, and Jose Rosero. 2014. “Birth Order and Human Capital
Development: Evidence from Ecuador.” Journal of Human Resources 49(2): 359–92.
Edmonds, Eric V. 2006. “Understanding Sibling Differences in Child Labor.” Journal of
Population Economics 19(4): 795-821.
Ejrnæs, Mette and Claus C. Pörtner. 2004. “Birth-Order and the Intrahousehold
Allocation of Time and Education.” Review of Economics and Statistics 86:
1008–1019.
Hanushek, Eric A. 1992. “The Trade-Off Between Child Quantity and Quality.” Journal
of Political Economics 100(1): 84–117.
Kantarevic, Jasmin and Stephane Mechoulan. 2006. “Birth Order, Educational
Attainment, and Earnings An Investigation Using the PSID.” Journal of Human
Resources 41(4): 755–77.
75

Lee, Jungmin. 2008. “Sibling Size and Investment in Children's Education: An Asian
Instrument.” Journal of Population Economics 21: 855–75.
Li, Hongbin, Junsen Zhang and Yi Zhu. 2008. “The Quantity–Quality Trade–Off of
Children in a Developing Country: Identification Using Chinese Twins.” Demography
45 (1): 223–43.
Maralani, Vida. 2008. “The Changing Relationship between Family Size and Educational
Attainment over the Course of Socioeconomic Development: Evidence from
Indonesia.” Demography 45(3): 693–717.
Parish, William and Robert Willis. 1993. “Daughters, Education, and Family Budgets
Taiwan Experiences.” Journal of Human Resources 28 (4): 863–98.
Rosenzweig, Mark and Kenneth Wolpin. 1980. “Testing the Quantity–Quality Fertility
Model: The Use of Twins as a Natural Experiment.” Econometrica 48 (January):
227–40.
Rosenzweig, Mark and Junsen Zhang. 2009. “Do Population Control Policies Induce
More Human Capital Investment? Twins, Birth Weight, and China's ‘One Child’
Policy.” Review of Economic Studies 76: 1149–74.
Strauss, John and Duncan Thomas. 1995. “Human Resources: Empirical Modeling of
Household and Family Decision.” In J. Behrman, & T. N. Srinivasan (Eds), Handbook
of development economics, Vol. III (pp. 1883–2023). Amsterdam: Elsevier Science.
Tenikue, Michel, and Bertrand Verheyden. 2010. “Birth Order and Schooling: Theory and
Evidence from Twelve Sub-Saharan Countries.” Journal of African Economics 19(4):
459–95.
Tsai, Shu-Ling, Hill Gates and Hei-Yuan Chiu. 1994. “Schooling Taiwan’s Women:
Women’s Educational Attainment in the Mid-Twentieth Century.” Sociology of
Education 67: 243–63.
Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data.
2nd ed. Cambridge, MA: MIT Press.
Yu, Wei-Hsin and Kuo-Hsien Su. 2006. “Gender, Sibship Structure, and Educational
Inequality in Taiwan: Son Preference Revisited.” Journal of Marriage and Family 68
(November): 1057–68.

76

CHAPTER 3
WAGE ADJUSTMENT OVER THE BUSINESS CYCLE: EVIDENCE FROM
TAIWAN

3.1

Introduction

Because aggregate time series on real wages display little cyclicality, macroeconomists
long believed that real wages are nearly unresponsive to the business cycle. The stylized
fact of weak wage cyclicality spawned numerous theories of real wage stickiness. The
theories included efficiency wage models, implicit contract models in which employers
provide real wage insurance to workers, and inside-outsider models. 1
More recently, a series of studies based on longitudinal microdata have
demonstrated that real wages are actually quite procyclical. The true procyclicality of real
wages is obscured in aggregate time series on real wages because of a composition bias:
the aggregate statistics have the tendency to put more weight on low-skill workers during
expansions than recessions. By contrast, studies based on longitudinal microdata have
been able to avoid the composition bias by tracking the same workers over time and
thereby controlling for cyclically changing composition of the workforce.
For example, using the U.S. Panel Study of Income Dynamics data from 1967-68 to
1986-87, Solon, Barsky, and Parker (1994) find when the unemployment rate increases
by an additional percentage point, men's real wage growth tends to decline by 1.4
percentage point. Numerous U.S. studies based on longitudinal data, which are
summarized in Solon, Barsky, and Parker (1992), Abraham and Haltiwanger (1995), and
many follow-up studies, are generally consistent with Solon et al. (1994) in the extent of

1

See Solon and Barsky (1989) and Solon, Barsky, and Parker (1994) for a detailed
summary of the time series evidence and a discussion of how such evidence has
influenced macroeconomic theory.
77

estimated cyclicality of real wages. 23
Turning to other countries, Hart (2006) and Devereux and Hart (2006) address the
issue for the United Kingdom, and Anger (2007) for the German labor market. Peng and
Siebert (2007) address the issue for Germany and the United Kingdom and Peng and
Siebert (2008) for Italy. Martins (2007), Martins, Solon, and Thomas (2012), and
Carneiro, Guimaraes, and Portugal (2012) deal with the Portuguese labor market. 4 While
these studies deal with the European labor market, Shin (2012) provides the first
longitudinal evidence on real wage cyclicality from an Asian economy, South Korea. All
these non-U.S. studies support the U.S. evidence, finding that real wages are procyclical.
The U.S. and non-U.S. microdata-based literature mentioned above, however, is
based on data extending no later than the mid-2000s and did not explore what the cyclical
wage patterns have been during the Great Recession. Recent studies by Elsby, Shin, and
Solon (2013, 2014) update the analysis of real wage cyclicality to 2012 for the U.S. and
U.K. labor markets; Blundell, Crawford, and Jin (2014) and Gregg, Machin, and
Fernandez-Salgado (2014) update the analysis to 2012 for the United Kingdom. To this
date, however, no microdata-based evidence on the cyclical wage patterns during the
Great Recession exists for countries other than the United States and United Kingdom.
This paper provides the first microdata-based evidence on real wage cyclicality that
involves the cyclical wage pattern during the Great Recession from an Asian economy,
Taiwan.
The analysis of real wage cyclicality in Taiwan is based on the Manpower Survey
(MS) and its May supplement, the Manpower Utilization Survey (MUS). I use the

2

Examples from the U.S. literature include Stockman (1983), Bils (1985), Bowlus, Liu,
and Robinson (2002), and Shin and Solon (2007).
3
Substantial procyclicality of real wages is a feature of several macroeconomic theories,
including those of Lucas and Rapping (1969) and Mortensen and Pissarides (1994).
4
See Shin (2012) for a detailed summary of the microdata-based evidence from the
non-U.S. studies.
78

microdata to study cyclical wage patterns over the period 1978 to 2012, with emphasis on
the two recessions of the 2000s. The mean real wage series from repeated cross-sections
show a secular time trend over the sample period, and exhibit no cyclical variation in the
2000s. However, the longitudinal analyses, which allow us to remove a time trend
contaminating cyclical patterns and account for observed and unobserved heterogeneity
(i.e., worker fixed effects) that leads to a composition bias, do reveal some cyclical wage
patterns. The results show that real wages during the Great Recession are procyclical. On
the contrary, real wages in the recession of the early 2000s, which also had a large impact
on the unemployment rate, are somewhat acyclical. The finding that real wages are more
procyclical in the Great Recession than in the recession of the early 2000s is consistent
with that in the United Kingdom. Furthermore, while some literature finds that real wages
tend to be more responsive to the cycle for men than for women (Blank, 1989; Tremblay,
1990; Solon et al., 1994; Park and Shin, 2005; Shin, 2012), I find no heterogeneous
responses of real wages to cyclical fluctuations between gender groups. This finding is
consistent with a U.K. study by Hart (2006) who finds that real wages are equally
procyclical between genders. I also find no group heterogeneity among education and age
groups and among workers in the public and private sectors.
This paper is organized as follows. Sections 3.2 and 3.3 describe the data and
empirical methodology, respectively. Section 3.4 presents evidence on cyclical wage
patterns from both repeated cross-sections and longitudinally matched microdata. Section
3.5 concludes.

3.2

Data

The analyses are based on the Manpower Survey (MS) and its May supplement, the
Manpower Utilization Survey (MUS). The MS is a monthly survey of labor market
activities. The MS is given to approximately 20,000 registered households and there are
79

nearly 60,000 civilians aged 15 and older in the sampled households. The MS asks
detailed questions on hours worked during the previous week (i.e., the reference week)
and jobs held, in addition to demographic information such as gender, educational
attainment, and age. Every MUS asks sample members about their monthly earnings at
their main job, which are a "point-in-time" measure of earnings.
As in most existing studies, my analyses are based on real hourly wages. To
construct hourly wage measures, I first convert the monthly earnings to an estimated
weekly earnings by dividing by 4.29 (≈ 30/7) and then convert the weekly earnings to an
estimated hourly wage by dividing by hours worked last week. 5 Since both hours and
earnings used are for main jobs, the estimated hourly wage is the wage for a main job.
Hourly wages are deflated by the consumer price index (CPI) and expressed as real
wages in 2011 NT dollars. Wage measures used for all analyses are log real hourly wages.
To focus on worker groups with substantial attachment to the labor force, I follow
Elsby, Shin and Solon (2013, 2014) and restrict my samples to workers between the ages
of 25 and 59. I also exclude workers with non-positive earnings and potential experience
(age minus years of education minus 6), and require at least 35 hours of work a week. I
only include in the samples workers from non-agricultural sectors. I then trim the
remaining sample for each year by excluding the cases with the top and bottom 1% of
hourly wages. 6 The resulting sample of men is typically more than 12,000 per year. The
women's sample starts at more than 2,000 in 1978 and exceeds 8,000 in the later years of
the sample period. For the sake of national representativeness, I use the provided MS
sampling weights. 7
5

The reported monthly earnings are for April, so 30 days are used to calculate the
number of weeks.
6
I have verified, however, that the cyclical patterns remain much the same if I use a
15-64 age range, if I do not trim the outliers, and also if I do not require at least 35 hours
of work a week. Requiring at least 35 hours of work a week results in smaller sample
sizes, but the difference is moderate.
7
Unweighted estimates turn out to be similar.
80

To provide context for my analysis, the top panel of Figure 3.1 and Table 3.1 present
the 1978-2012 Taiwanese series for the unemployment rate and mean log real hourly
wages. Wages shown in the table are scaled by the CPI and express real wages in 2011
NT dollars. The annual unemployment rate is used to emphasize which years are
recession years and which are expansion years. 8 The bottom panel of Figure 3.1 displays
the logarithm of the CPI and mean log nominal hourly wages.
As the unemployment rate shows, Taiwan experienced two very severe recessions in
the 2000s. The unemployment rate, which was 2.99% in 2000, was brought to 5.17% in
the recession of the early 2000s, and was again brought to a record high rate, 5.85%, in
the Great Recession. Like the United Kingdom, Taiwan entered the 1980s with high
inflation, reaching nearly 20% on an annual basis. Subsequently, inflation fell rapidly in
the 1980s. Except for an aberration between the late 1980s and mid 1990s (associated
with the boom at that time), inflation remained below 2%. 9 Inflation did rise somewhat
in 2008, though, exceeding 3%.
As displayed in Table 3.1 and the top panel of Figure 3.1, in the first half of the
sample period, the upward trends in men's and women's real wages are dramatic.
However, after experiencing remarkable wage growth, men's and women's real hourly
wages stagnated or even declined in the later years of the sample period. For example,
men's mean log real wage declined from 5.360 in 2000 to 5.264 in 2012, a reduction of
0.096, while women's mean wage experienced a relatively moderate growth of 0.023.
The bottom panel of Figure 3.1 decomposes men's and women's log real wage series

8

The recessions of the 2000s are similarly classified if real GDP growth rates are used as
an alternative indicator. For example, in the recession of the early 2000s, the
unemployment rate increased from 2.99% in 2000 to 4.57% in 2001, while the real GDP
growth rate declined from 5.80% in 2000 to -1.65% in 2001. In the Great Recession, the
unemployment rate rose from 3.91% in 2007 to 5.85% in 2009, while the real GDP
growth rate dropped from 5.98% in 2007 to -1.81% in 2009.
9
The average inflation rate was nearly 4% on an annual basis between 1989 and 1996. It
was below 1% between 2000 and 2005, and nearly 1.5% between 2007 and 2012.
81

already shown in the top panel into the difference between the mean log nominal wage
and the logarithm of the price level. The first thing to notice is that men's nominal wages
grew very little in the 2000s (increased from 5.246 in 2000 to 5.274 in 2012, an increase
of 0.028), and much more slowly than the log price level (increased from -0.110 in 2000
to 0.010 in 2012, an increase of 0.120). The smaller growth in nominal wages meant that
men's real wages underwent a nontrivial decline in the 2000s already shown in Table 3.1
and the top panel of Figure 3.1. On the contrary, women's nominal wage experienced a
larger increase (rose from 4.954 in 2000 to 5.100 in 2012, an increase of 0.146) than
men's, but the growth in women's nominal wage was largely offset by inflation, resulting
in a modest increase in women's real wage.
As depicted in Figure 3.1, men's and women's real wages show no perceptible
cyclical variation in the 2000s, even though Taiwan experienced two severe recessions in
the 2000s (classified by the unemployment rate). All of this, however, is based on wage
measures from repeated cross-sections without correcting for a composition bias
potentially resulting from my sample selection criteria and data limitations. 10 The U.S.
evidence indicates that such measures could be subject to a substantial countercyclical
composition bias. Furthermore, the upward secular trends in men's and women's wages
make it trickier to distill the cyclical patterns. In the next section, I will utilize several
approaches to achieve a partial correction of the composition bias and a time trend
contaminating cyclical patterns.

3.3 Empirical Methodology
Access to the microdata goes a considerable way towards reducing the composition-bias
issues associated with aggregate data. As discussed by Solon, Barsky, and Parker (1994)
and others, low-skill workers' employment is especially responsive to macroeconomic
10

I will discuss this issue in the next section.
82

shocks, so aggregate wage data are substantially countercyclically biased by their
tendency to put more weight on low-skill workers during expansions than recessions.
This generates a countercyclical composition bias in aggregate wage statistics, making
workers' real wages appear less procyclical than they really are.
Using the MUS microdata, I can reduce the composition bias. Unfortunately, I
cannot avoid the composition bias entirely by using the MUS microdata. I am unable to
measure the wage opportunities of individuals with no work and unpaid family workers,
which are excluded from the sample. In addition, because the MUS measures wages only
for those working in the reference week in May, it seems potentially subject to a more
severe composition bias than the U.S. March Current Population Survey (CPS).
I will achieve a partial correction of the resulting composition bias in three ways.
First, following the U.S. and U.K. analyses of real wages in Elsby et al. (2013, 2014), I
regression-adjust wage measures for some observable characteristics (education and
potential experience) of the worker samples. The regression adjustments are conducted
by estimating "regression-adjusted" year effects by applying least squares (weighting by
the provided MS weights) to a regression of individual workers' log real wages on year
dummies and controls for years of education and an unrestricted set of potential
experience dummies. 11 Second, I reweight workers' observable characteristics to be the
same as those in pre-recession years by the reweighting approach proposed by DiNardo,
Fortin, and Lemieux (1996, DFL thereafter), so the observed workforce composition
remains unchanged over the business cycle. 12 Finally, I utilize the rotating panel design
of the MS and conduct longitudinal analyses that not only hold composition constant by
following a portion of workers from one May to the next but also remove a time trend
contaminating cyclical patterns.

11

Using a quartic in potential experience yields similar results.
The regression adjustments and DFL reweighting approach are conducted using
repeated cross-sections.
12

83

3.4

Empirical Results

3.4.1

Overall Wage Adjustment over the Business Cycle

Repeated Cross-sections. I begin by following Elsby et al. (2013, 2014) in
regression-adjusting wage measures, which can partly correct for the composition bias by
controlling for year-to-year changes in the demographic composition of the samples. For
example, as shown in Tables 3.2 and 3.3 and Figure 3.2, in addition to showing mean log
wage for each year in the 2000-2005 period, I also estimate "regression-adjusted" year
effects by applying least squares (again weighting by the provided MS weights) to a
regression of individual workers' log real wages on year dummies for 2001, 2002, 2003,
2004, and 2005 (with 2000 as the omitted reference category) and control for years of
education and an unrestricted set of potential experience dummies.
As expected and consistent with U.S. evidence in Elsby et al. (2013, 2014), the
regression-adjusted year effects show that mean wage patterns in Table 3.1 are affected
by the countercyclical composition bias. Whereas the unadjusted means indicate that
men's mean log real wages were 0.004 (s.e.=0.006) less (but not significantly) in 2003
than in 2000, the adjusted 2003 year effect is 0.034 (s.e.=0.005) less than the 2000
effect. 13,14 Women's unadjusted mean log real wages were 0.037 (s.e.=0.008) higher in
2003 than in 2000, but the adjusted means suggest there is no significant change in mean
wages between 2000 and 2003.
I perform the same exercise for the 2007-2012 period. Whereas the unadjusted
means indicate that, for men, log real wages were 0.025 (s.e.=0.006) lower in 2009 than
in 2007, the adjusted 2009 year effect is 0.056 (s.e.=0.005) less than the 2007 effect. 15
Similar countercyclical composition bias is observed in women's wage patterns between
13

The unadjusted and regression-adjusted 2003 year effects are significantly different
from each other.
14
The unemployment rate is 2.99% in 2000 (the pre-recession year) and 4.99% in 2003.
15
The unemployment is 3.91% in 2007 (the pre-recession year) and 5.85% in 2009.
84

2007 and 2009. Compared with the recession of the early 2000s, the (regression-adjusted)
wage drop in the Great Recession is significantly greater when there is a run-up in the
unemployment rate.
However, as shown in Figure 3.2, both men's and women's mean log wages kept
declining even in the recovery from the two recessions of the 2000s (classified by the
unemployment rate).

This suggests that the cyclical patterns based on these mean wage

series from repeated cross-sections may be contaminated by the downward trend in mean
log real wages. I will return to this issue when I use longitudinally matched data to
correct for the composition bias resulting from workers' observable and unobservable
characteristics and for the time trend.
In addition to the regression-adjusting, I also employ the DFL reweighting approach
to reweight workers' observed characteristics (education and potential experience) in
recession years to be the same as in pre-recession years. 16 This also allows us to hold
worker composition constant and hence accounts for observed heterogeneity. For
example, I construct a reweighting factor, which reweights the composition of the
workforce in 2001, 2002, 2003, 2004, and 2005 to be the same as that in 2000 (the
pre-recession year). After constructing the reweight factor, I estimate the "DFL
reweighted" year effect by applying least squares (weighting by the reweighted factor) to
a regression of individual workers' log real wages on year dummies for 2001, 2002, 2003,
2004, and 2005 (with 2000 as the omitted reference category). I perform a similar
exercise for the 2007-2012 period. As displayed in Tables 3.2 and 3.3, the results are
quite similar to those of regression adjustments.
Longitudinally Matched Microdata. As discussed in the previous subsection, the
cyclical patterns based on the mean wage series from repeated cross-sections may be

16

I estimate a logit model for the group membership. The covariates include an
unrestricted set of education and potential experience dummies and the interaction
between education dummies and years of potential experience.
85

contaminated by the downward trend in mean log real wages in addition to the
composition bias. As displayed in Figure 3.2, regression-adjusted mean log wages
continued to decline even in the recovery from the recessions, suggesting that the year
effects seem to not only capture the cyclical fluctuations but also a time trend.
Presumably, the time trend is linear. 17 The traditional approach to accounting for a linear
time trend is to first-difference wage series and use year-to-year change in wages. Using
the first-differenced wage series can not only remove a linear time trend but also account
for unobserved heterogeneity (i.e., worker fixed effects) that leads to a composition bias.
Fortunately, the rotating panel design of the MS makes it possible to follow a
portion of one May's sample to the next May. However, like the CPS, the longitudinally
matched MS sample is far from ideal panel data. The sample sizes for May-to-May
matches are almost always less than half of the sample sizes for the cross-sections.
Another source of the sample loss is that the MS does not follow residential movers,
resulting in an endogenous sample selection in a study of wage changes. Nevertheless, I
perform year-to-year matches from the samples of workers between adjacent MUS's (the
MS's May supplement). I follow the guidance of Madrian and Lefgren (2000) in
verifying that longitudinal matches on identification numbers are true matches by
requiring that gender also matches, that year-to-year change in reported age is one, and
that years of education is greater than or equal to that in previous years (t-1). Following
Elsby et al. (2013), I weight individuals by the simple average of their weights across the
two years. 18
Table 3.4 shows mean year-to-year change in log real wages by gender for each pair
of years from 1978-79 to 2011-12. Figure 3.3 adds a visual display for each gender. The
17

As depicted in Figure 1, there seems to be a trend break around 1995, so the
assumption of a linear time seems implausible for the whole sample period. However,
since the focus of this paper is on the recessions of the 2000s, the assumption of a linear
time trend seems consistent with wage series in the 2000s.
18
Unweighted estimates turned out to be similar.
86

first thing to notice is that, because of normal life-cycle wage growth, the mean changes
are almost always positive except for 2007-2008, and real wage growth in the 2000s was
much slower than in early years of the sample period. In stark contrast to Figure 3.1
displaying no cyclical variation, Figure 3.3 depicts some cyclical patterns in the 2000s,
suggesting that using the first-differenced wages does control for the downward trend
observed in mean wage series displayed in Figure 3.2. Furthermore, the cyclical patterns
in the recession of the early 2000s and the Great Recession differ. In the Great Recession,
both men's and women's wages followed a procyclical pattern. The year-to-year change in
log wages turned negative as the unemployment rate increased, and then rebounded as the
unemployment rate improved. By contrast, men's and women's wage adjustment seemed
to be disconnected with the cyclical fluctuations in the recession of the early 2000s.
To have a better picture of how year-to-year change in log real wages evolves over
the business cycle, I estimate "regression-adjusted" year effects by applying least squares
(again weighting by the simple average of their MS weights across the two years) to a
regression of individual workers' year-to-year change in log real wages on year dummies
for 2000-2001, 2001-2002, 2002-2003, 2003- 2004, and 2004-2005 (with 1999-2000 as
the omitted reference category) and controls for years of education and an unrestricted set
of potential experience dummies. A similar exercise was also performed for the
2007-2012 period.
Along with the regression-adjusted year effects, Table 3.5 also displays the
unadjusted year effects. As shown in Table 3.5, the variation over time in the estimated
regression-adjusted year effects is virtually identical to that for the unadjusted year
effects. Figure 3.4 provides a visual display by plotting the estimated regression-adjusted
year effects and their 95% confidence interval for the two recessions of the 2000s for
each gender. As shown in Table 3.5 and Figure 3.4, in the Great Recession, men's
year-to-year change in log wages was 0.033 (s.e.=0.005) lower in 2007-2008 than in
2006-2007 (prerecession years), while women's change in log wages was 0.025
87

(s.e.=0.006) lower. Subsequently, both men's and women's change in log wages
rebounded as the unemployment rate improved. In contrast to the procyclical pattern in
the Great Recession, men's and women's wage adjustment seemed to be disconnected
with the cyclical fluctuations in the recession of the early 2000s.

3.4.2

Group Heterogeneity

In this section, I explore between-group heterogeneity. Figure 3.5 displays the estimated
regression-adjusted year effects and their 95% confidence interval for the two recessions
of the 2000s for men, and Figure 3.6 for women. The year effects are estimated using the
year-to-year change in log real wages. I compare two education groups, a high-education
group (defined as some college and above) and a low-education group (defined as high
school and less), and two age groups, a younger group (defined as age<=40) and an older
group (defined as age>40). 19 I also compare workers in the public sector with those in
the private sector. The inspection of these figures suggests that there is no group
heterogeneity in the response of wages to cyclical fluctuations for both genders in the
recessions of the 2000s.

3.5

Conclusion

Using the microdata from the Manpower Survey (MS) and its May supplement, the
Manpower Utilization Survey (MUS), I study cyclical wage patterns over the period 1978
to 2012, with emphasis on the two recessions of the 2000s. The results of longitudinal
analyses show that real wages during the Great Recession are procyclical, whereas real
wages in the recession of the early 2000s are somewhat acyclical. The finding that real
wages are more procyclical in the Great Recession than in the recession of the early

19

I also tried different age cutoffs, such as 35 and 45, and obtained quite similar results.
88

2000s is consistent with that in the United Kingdom. In addition, while most literature
finds that men's real wages tend to be more procyclical than women's, the results suggest
that the response of real wages to cyclical fluctuations among gender groups is quite
similar in the recessions of the 2000s, a finding similar to that in the British labor market.

89

APPENDICES

90

APPENDIX A - FIGURES
Men's and Women's Mean Log Real Wages over the Business Cycle
A. Real Wages

Unemployment Rate (%)

6

5.2

4

5

4.8

3

4.4

2

4

1

3.6

Mean Log Real Hourly Wage

7

Figure 3.1

1978

1980

1985

1990

1995
Year

2000

2005

2010

2005

2010

2012

Men's mean log real wage
Women's mean log real wage
Unemployment rate

.4

5.4

0

5

Log CPI

-.4

4.6

-1.2

-.8

4.2

-1.6

3.8

-2

3.4
3

Mean Log Nominal Hourly Wage

B. Nominal Wages

1978

1980

1985

1990

1995
Year

2000

Men's mean log nominal wage
Women's mean log nominal wage
Log CPI

91

2012

Figure 3.2

Log Real Wages in the Recessions in the 2000s

.04
.02
0
-.08 -.06 -.04 -.02

Unadjusted

Reg-adjusted

-.1

Mean Relative to Pre-recession Year

.06

A. Men

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

.04
-.06

-.04

-.02

0

.02

Unadjusted

-.08

Reg-adjusted

-.1

Mean Relative to Pre-recession Year

.06

B. Women

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Notes: The figures plot the estimation results presented in Tables 3.2 and 3.3 Since the
results from the regression adjustment and DFL reweighting are quite similar, only the
regression adjustment results are plotted here for simplicity. Each point represents the
estimate of mean relative to pre-recession year with its 95% confidence interval.
Pre-recession years are 2000 and 2007, respectively. Means in 2000 and 2007 are
normalized to zero.
92

Year

93
02

07

20

12

20 09
10

20

20

20
05

20

20
00

Men

19
95

19
90

19
85

19 1979
80

Mean Log Real Hourly Wage Changes
-.05 -.03 -.01 .01 .03 .05 .07 .09 .11 .13 .15

Figure 3.3
Mean Year-to-Year Changes in Log Real Wage by Gender from
Longitudinally Matched MS/MUS Data

Women

Figure 3.4

Log Real Wage Changes in the Recessions in the 2000s

.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.04

.06

A. Men

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.04

.06

B. Women

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Notes: The figures plot the estimation results presented in Tables 3.4. Since the
unadjusted and regression-adjusted results are quite similar, only the regression
adjustment results are plotted here for simplicity. Each point represents the estimate of
mean changes relative to pre-recession year with its 95% confidence interval.
Pre-recession years are 1999-2000 and 2006-2007, respectively. Mean changes in
1999-2000 and 2006-2007 are normalized to zero.
94

Figure 3.5

Log Real Wage Changes in the Recessions in the 2000s
by Education, Age, and Sector, Men

A. High school and under
.06
.04
.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.04
.02
0
-.02
-.04
-.06
-.08

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.06

B. Some college and above

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

C. Age<=40
.06
.04
.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.04
.02
0
-.02
-.04
-.06
-.08

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.06

D. Age>40

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

F. Private sector
.04
.02
0
-.02
-.08

-.06

-.04

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.06
.04
.02
0
-.08 -.06 -.04 -.02

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.06

E. Public sector

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Notes: Each point represents the estimate of mean changes relative to pre-recession year
with its 95% confidence interval. Pre-recession years are 1999–2000 and 2006–2007,
respectively. Mean changes in 1999–2000 and 2006–2007 are normalized to zero.
95

Figure 3.6

Log Real Wage Changes in the Recessions in the 2000s
by Education, Age, and Sector, Women
B. Some college and above
.06
.04
.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.04
.02
0
-.02
-.04
-.06
-.08

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.06

A. High school and under

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

D. Age>40
.04
.02
0
-.02
-.04
-.08

-.06

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.04
.02
0
-.02
-.04
-.06
-.08

Mean Log Real Hourly Wage Changes
Relative to Pre-recession Year

.06

.06

C. Age<=40

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

E. Public sector
.04
.02
0
-.02
-.08

-.06

-.04

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.06
.04
.02
0
-.08 -.06 -.04 -.02

Mean Log Real Hourly Wage Changes
Relative to P re-recession Y ear

.06

F. Private sector

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012
Year

Notes: Each point represents the estimate of mean changes relative to pre-recession year
with its 95% confidence interval. Pre-recession years are 1999–2000 and 2006–2007,
respectively. Mean changes in 1999–2000 and 2006–2007 are normalized to zero.

96

APPENDIX B - TABLES
Table 3.1

Mean Log Real Hourly Wages by Gender and Year

1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012

Annual
Unemployment
Rate
1.67
1.27
1.23
1.36
2.14
2.71
2.45
2.91
2.66
1.97
1.69
1.57
1.67
1.51
1.51
1.45
1.56
1.79
2.60
2.72
2.69
2.92
2.99
4.57
5.17
4.99
4.44
4.13
3.91
3.91
4.14
5.85
5.21
4.39
4.24

Notes:

The standard errors (S.E.) are robust to heteroskedasticity.

Year

Men's Log Real Wages
Mean
4.375
4.479
4.522
4.544
4.624
4.650
4.685
4.710
4.786
4.828
4.926
5.030
5.119
5.198
5.221
5.293
5.336
5.331
5.331
5.337
5.351
5.354
5.360
5.377
5.367
5.356
5.348
5.341
5.331
5.324
5.304
5.299
5.289
5.283
5.264

S.E.
0.005
0.005
0.005
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.003
0.004
0.003
0.003
0.003
0.003
0.003
0.003
0.003
0.003
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004

97

Women's Log Real Wages
Mean
4.047
4.124
4.137
4.160
4.236
4.252
4.283
4.314
4.363
4.414
4.518
4.610
4.714
4.794
4.857
4.910
4.959
4.978
4.996
5.024
5.050
5.073
5.068
5.110
5.102
5.106
5.113
5.113
5.106
5.096
5.082
5.094
5.083
5.088
5.091

S.E.
0.011
0.010
0.010
0.009
0.009
0.009
0.008
0.007
0.007
0.007
0.007
0.006
0.006
0.006
0.006
0.006
0.006
0.005
0.005
0.005
0.005
0.005
0.005
0.006
0.006
0.006
0.005
0.005
0.005
0.005
0.005
0.005
0.005
0.005
0.005

Table 3.2

Men's Log Real Wages in the Recessions in the 2000s

0
(normalized)
0.017
(0.006)
0.007
(0.006)
-0.004
(0.006)
-0.012
(0.005)
-0.019
(0.005)

RegressionAdjusted
0
(normalized)
0.006
(0.005)
-0.015
(0.005)
-0.034
(0.005)
-0.048
(0.005)
-0.064
(0.005)

DFL
Reweighted
0
(normalized)
0.009
(0.006)
-0.010
(0.006)
-0.032
(0.006)
-0.046
(0.005)
-0.062
(0.006)

0
(normalized)
-0.020
(0.006)
-0.025
(0.006)
-0.035
(0.006)
-0.041
(0.006)
-0.060
(0.006)

0
(normalized)
-0.032
(0.005)
-0.056
(0.005)
-0.067
(0.005)
-0.077
(0.005)
-0.101
(0.005)

0
(normalized)
-0.033
(0.005)
-0.056
(0.006)
-0.069
(0.006)
-0.079
(0.006)
-0.105
(0.006)

Year

Unadjusted

2000
2001
2002
2003
2004
2005

2007
2008
2009
2010
2011
2012

Note: The standard errors (in parentheses) are robust to heteroskedasticity.

98

Table 3.3

Women's Log Real Wages in the Recessions in the 2000s

0
(normalized)
0.042
(0.008)
0.034
(0.008)
0.037
(0.008)
0.044
(0.008)
0.045
(0.008)

RegressionAdjusted
0
(normalized)
0.024
(0.006)
0.006
(0.006)
-0.008
(0.006)
-0.015
(0.006)
-0.034
(0.006)

DFL
Reweighted
0
(normalized)
0.029
(0.008)
0.009
(0.008)
0.001
(0.008)
0.000
(0.008)
-0.012
(0.008)

0
(normalized)
-0.014
(0.007)
-0.002
(0.007)
-0.013
(0.007)
-0.008
(0.007)
-0.006
(0.007)

0
(normalized)
-0.034
(0.006)
-0.049
(0.006)
-0.067
(0.006)
-0.073
(0.006)
-0.087
(0.006)

0
(normalized)
-0.029
(0.007)
-0.040
(0.007)
-0.055
(0.007)
-0.053
(0.007)
-0.058
(0.007)

Year

Unadjusted

2000
2001
2002
2003
2004
2005

2007
2008
2009
2010
2011
2012

Note: The standard errors (in parentheses) are robust to heteroskedasticity.

99

Table 3.4

Year

1978-1979
1979-1980
1980-1981
1981-1982
1982-1983
1983-1984
1984-1985
1985-1986
1986-1987
1987-1988
1988-1989
1989-1990
1990-1991
1991-1992
1992-1993
1993-1994
1994-1995
1995-1996
1996-1997
1997-1998
1998-1999
1999-2000
2000-2001
2001-2002
2002-2003
2003-2004
2004-2005
2005-2006
2006-2007
2007-2008
2008-2009
2009-2010
2010-2011
2011-2012

Mean Year-to-Year Changes in Log Real Wage by Gender from
Longitudinally Matched MS/MUS Data
Unemploym
ent Rate
(End of
Period)
1.27
1.23
1.36
2.14
2.71
2.45
2.91
2.66
1.97
1.69
1.57
1.67
1.51
1.51
1.45
1.56
1.79
2.60
2.72
2.69
2.92
2.99
4.57
5.17
4.99
4.44
4.13
3.91
3.91
4.14
5.85
5.21
4.39
4.24

Men's Mean Year-to-Year
Changes in Log Real Wages
Mean
0.119
0.072
0.036
0.086
0.048
0.064
0.059
0.094
0.046
0.091
0.090
0.098
0.083
0.035
0.074
0.055
0.018
0.018
0.040
0.041
0.021
0.033
0.041
0.016
0.015
0.015
0.014
0.022
0.021
-0.012
0.018
0.025
0.016
0.017

S.E.
0.008
0.007
0.007
0.006
0.006
0.006
0.006
0.005
0.005
0.005
0.005
0.004
0.005
0.004
0.004
0.004
0.005
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.003
0.003
0.003
0.004
0.003

Women's Mean Year-to-Year
Changes in Log Real Wages
Mean
0.138
0.046
0.037
0.098
0.052
0.089
0.086
0.085
0.075
0.088
0.099
0.108
0.065
0.060
0.076
0.072
0.035
0.031
0.054
0.045
0.032
0.023
0.054
0.032
0.025
0.019
0.011
0.022
0.015
-0.010
0.022
0.014
0.022
0.027

Note: The standard errors (S.E.) are robust to heteroskedasticity.
100

S.E.
0.015
0.012
0.013
0.013
0.011
0.010
0.008
0.008
0.007
0.007
0.007
0.006
0.008
0.006
0.006
0.007
0.006
0.006
0.006
0.005
0.005
0.005
0.005
0.005
0.005
0.005
0.004
0.004
0.004
0.004
0.004
0.004
0.004
0.004

Table 3.5

Log Real Wage Changes in the Recessions in the 2000s

Unemployment
Rate (End of
Period)

Change in
Log CPI

1999-2000

2.99

0.012

2000-2001

4.57

0.004

2001-2002

5.17

0.002

2002-2003

4.99

-0.001

2003-2004

4.44

0.009

2004-2005

4.13

0.016

2006-2007

3.91

0.007

2007-2008

4.14

0.038

2008-2009

5.85

-0.005

2009-2010

5.21

0.013

2010-2011

4.39

0.013

2011-2012

4.24

0.014

Year

Men's Mean Log Real Hourly
Wage Changes

Women's Mean Log Real
Hourly Wage Changes

0
(normalized)
0.008
(0.006)
-0.017
(0.006)
-0.018
(0.005)
-0.019
(0.005)
-0.019
(0.005)

RegressionAdjusted
0
(normalized)
0.008
(0.006)
-0.017
(0.006)
-0.018
(0.005)
-0.019
(0.005)
-0.019
(0.005)

0
(normalized)
0.031
(0.007)
0.008
(0.007)
0.002
(0.007)
-0.004
(0.007)
-0.013
(0.006)

RegressionAdjusted
0
(normalized)
0.031
(0.007)
0.008
(0.007)
0.001
(0.007)
-0.005
(0.007)
-0.014
(0.006)

0
(normalized)
-0.033
(0.005)
-0.003
(0.005)
0.004
(0.005)
-0.005
(0.005)
-0.004
(0.005)

0
(normalized)
-0.033
(0.005)
-0.002
(0.005)
0.005
(0.005)
-0.004
(0.005)
-0.003
(0.005)

0
(normalized)
-0.026
(0.006)
0.006
(0.006)
-0.001
(0.006)
0.007
(0.006)
0.012
(0.006)

0
(normalized)
-0.025
(0.006)
0.007
(0.006)
-0.001
(0.006)
0.007
(0.006)
0.012
(0.006)

Unadjusted

Unadjusted

Note: The standard errors (in parentheses) are robust to heteroskedasticity.

101

BIBLIOGRAPHY

102

BIBLIOGRAPHY
Abraham, Katharine G., and John C. Haltiwanger. 1995. “Real Wages and the Business
Cycle.” Journal of Economic Literature, 33(3): 1215–64.
Anger, Silke. 2007. “The Cyclicality of Effective Wages within Employer-Employee
Matches: Evidence from German Panel Data.” ECB/CEPR Labour Market Workshop
on Wage and Labour Cost Dynamics, Working Paper Series No. 783.
Bils, Mark J. 1985. “Real Wages over the Business Cycle: Evidence from Panel Data.”
Journal of Political Economy, 93(4): 666–89.
Blank, Rebecca M. 1989. “Disaggregating the Effect of the Business Cycle on the
Distribution of Income.” Economica, 56(222): 141–63.
Blundell, Richard, Claire Crawford, and Wenchao Jin. 2014. “What Can Wages and
Employment Tell Us about the UK's. Productivity Puzzle.” Economic Journal,
124(576): 377–407.
Bowlus, Audra, Haoming Liu, and Chris Robinson. 2002. “Business Cycle Models,
Aggregation, and Real Wage Cyclicality.” Journal of Labor Economics, 20(2):
308–35.
Carneiro, Anabela, Paulo Guimarães, and Pedro Portugal. 2012. “Real Wages and the
Business Cycle: Accounting for Worker, Firm, and Job Title Heterogeneity.”
American Economic Journal: Macroeconomics, 4(2): 133–52.
Devereux, Paul J., and Robert A. Hart. 2006. “Real Wage Cyclicality of Job Stayers,
Within- Company Job Movers, and Between-Company Job Movers.” Industrial and
Labor Relations Review, 60(1): 105–19.
DiNardo, John E., Nicole Fortin, and Thomas Lemieux. 1996. “Labor Market Institutions
and the Distribution of Wages, 1973-1992: A Semiparametric Approach. ”
Econometrica, 64(5): 1001–1004.
Elsby, Michael W., Donggyun Shin, and Gary Solon. 2013. “Wage Adjustment in the
Great Recession.” NBER Working Paper No. 19478.
Elsby, Michael W., Donggyun Shin, and Gary Solon. 2014. “Wage Adjustment in the
Great Recession and Other Downturns: Evidence from the United States and Great
Britain.” Forthcoming in Journal of Labor Economics.
Gregg, Paul, Stephen Machin, and Marina Fernandez-Salgado. 2014. “Real Wages and
Unemployment in the Big Squeeze.” Economic Journal, 124(576): 408–32.
Hart, Robert A. 2006. “Worker-Job Matches, Job Mobility and Real Wage Cyclicality.”
Economica, 73(290): 287–98.
Lucas, Robert E., Jr., and Leonard A. Rapping. 1969. “Real Wages, Employment, and
Inflation,” Journal of Political Economy, 77(5): 721–54.
Madrian, Brigitte C., and Lars John Lefgren. 2000. “An Approach to Longitudinally
Matching Current Population Survey (CPS) Respondents.” Journal of Economic and
103

Social Measurement, 26(2000): 31–62.
Martins, Pedro. 2007. “Heterogeneity in Real Wage Cyclicality.” Scottish Journal of
Political Economy, 54(5): 684–98.
Martins, Pedro S., Gary Solon, and Jonathan P. Thomas. 2012. “Measuring What
Employers Do about Entry Wages over the Business Cycle: A New Approach.”
American Economic Journal: Macroeconomics, 4(4): 36–55.
Mortensen, Dale T., and Christopher A. Pissarides. 1994. “Job Creation and Job
Destruction in the Theory of Unemployment.” Review of Economic Studies, 61(3):
397–415.
Park, Seonyoung, and Donggyun Shin. 2005. “Explaining Procyclical Male-Female
Wage Gaps.” Economics Letters, 88(2): 231–55.
Peng, Fei, and W. Stanley Siebert. 2007. “Real Wage Cyclicality in Germany and the UK:
New Results Using Panel Data.” IZA Discussion Paper No. 2688.
Peng, Fei, and W. Stanley Siebert. 2008. “Real Wage Cyclicality in Italy.” Labour, 22(4):
569–91.
Shin, Donggyun. 2012. “Cyclicality of Real Wages in Korea.” B.E. Journal of Economic
Analysis and Policy (Contributions), 12(1): Article 2.
Shin, Donggyun, and Gary Solon. 2007. “New Evidence on Real Wage Cyclicality within
Employer-Employee Matches.” Scottish Journal of Political Economy, 54(5): 648–60.
Solon, Gary, and Robert Barsky. 1989. “Real Wages over the Business Cycle.” NBER
Working Paper No. 2888.
Solon, Gary, Robert Barsky, and Jonathan A. Parker. 1992. “Measuring the Cyclicality of
Real Wages: How Important Is Composition Bias?” NBER Working Paper No. 4202.
Solon, Gary, Robert Barsky, and Jonathan A. Parker. 1994. “Measuring the Cyclicality of
Real Wages: How Important Is Composition Bias?” Quarterly Journal of Economics,
109(1): 1–26.
Stockman, Alan C. 1983. “Aggregation Bias and the Cyclical Behavior of Real Wages.”
Unpublished manuscript.
Tremblay, Carol H. 1990. “Wage Patterns of Women over the Business Cycle.”
Quarterly Review of Economics and Business, 30: 90–101.

104