Explaining Wages; Occupations and Tasks as Base Components
Andrea Jaeger (s1458663)
Master in International Economics and Business
University of Groningen, the Netherlands
April 2011
Abstract
Since the mid 1980‟s, production has become organized at a finer level of disaggregation; the task level. As computerization and offshoring are as well organized on a task level, labor performing non-routine tasks is generally positively affected by computerization and offshoring. Contrary, computers and overseas labor have substituted for routine tasks putting pressure on wages of “routine jobs”. This paper focuses on the extent to which occupations and task variables are able to explain wage variations. Using the US Current Population Survey (CPS), an inequality analysis does not clearly invoke a growing effect of occupations on hourly wages in the period 1983-2002. Form several cross-section OLS wage regressions it stems, however, that besides industry effects which are generally included in wage regressions, occupations and tasks do explain a considerable part of the wage variation. In addition, our results show that in 1983 occupations with a substantial intensity of routine tasks, for instance assembly workers, are panelized while in 2002, non-routine interactive and analytical professions, for instance therapist, are financially rewarded. In contrast to assumptions made in the existing literature, we do find several occupation wage differentials between industries in both 1983 and 2002. The study adds to various fields within international economics as it is related to a variety of research themes as international trade, ICT-developments and economics, and the development of wage inequalities
Supervisors: MSc. S.J. Kok & Prof. dr. J.H. Garretsen
Key words: Occupations, Technological Change, Trade and Labor Market interactions,
Wage Level and Structure, Trends in inequality,
1. Introduction
Since the mid-twentieth century, workers face an increased intensity of competition and uncertainty as the labor market is expected to be considerably affected by computerization and offshoring. In fact, given the rapidity of advancements that technological processes have undergone, technology has taken over a wide variety of tasks that were previously completed by human capital as, for instance, clerical work and telephone assistance (Autor, Katz and Kearny, 2006). Second, an ever-widening sphere of tasks has been replaced to overseas countries, from accounting functions such as tax preparation to architectural design and medical imaging diagnostic interpretation (Krugman, 2008) (Grossman and Rossi-Hansberg, 2006, 2008), (Feenstra and Hanson, 1999). The general believe that offshoring and computerization are expected to affect jobs and wages noticeably, may well lead to fear concerning wage and job security. The anxiety towards offshoring and computerization is likely to be intensified by the uncertainty regards the occupations and industries that are actually vulnerable to these changes (Baldwin, 2006).
The pace and strength with which offshoring currently affects workers, is related to the unbundling of tasks as described by Baldwin (2006). Since the late 19th century, rapidly falling transportation costs have ended the need to produce goods at the place of consumption, also known as the first unbundling. While it might have been less costly to undertake labor-intensive production stages at a low-cost location, production tended to be spatially clustered as the coordination of production activities in remote locations was difficult if not impossible. More recently, a significant decline of coordination costs, due to for instance the rise of internet, has fostered the second unbundling. The second unbundling represents the end of the necessity to perform tasks near each other. In fact, production of a single good or service can be spatially unpacked into various tasks that can be performed at different geographical locations. In other words, production is organized at a finer level of disaggregation.
for routine tasks that can be described with programmed rules as repetitive tasks of clerks and cashiers (Breshnahan, 1999).
Overall, general commotion concerning wage and job security can be clarified by the fact that, recently, competition has been intensified by both individual workers performing similar tasks in different nations, and from computer capital replacing several tasks that were previously performed by human labor. As computerization and offshoring take over various tasks while complement to others, occupations are expected to be influenced differently depending on the task content that is performed. Hence, non-routine tasks are mostly complemented, while routine tasks are substituted by computer capital and foreign labor. Consequently, the wage difference between routine and non-routine jobs is expected to rise. Therefore, here it is hypothesized that, over the between 1983 and 2002, between-occupation wage inequality has increased. In addition, we suggest that growth of the between-occupation wage gap has major implications for wage regressions. In fact, if wage inequality within an occupation is low but the between- occupation wage gap is large, occupations should be of great importance in explaining wages. In this case, individuals with similar occupations have comparable wages, while wage differences between professions are high. Therefore, growth of between-occupation wage inequality1 (ceteris paribus) is associated with a rise of the effect of occupations on wages.
Remarkably, technological advancements, the extensive changes the nature of trade and the subsequent effects seem not to be taken into account in the field of wage structure analyses. Recent studies on wage structures do not include any occupation or task-level related variable but remain estimating industry-specific effects (Blau and Kahn, 2006) (Blau and Devaro, 2007) (Chavalier, 2007) (van der Meer, 2008) (Miller, 2009) (Fransen, Plantinga and Vlasblom, 2009). However, if work is increasingly organized on a task-level, industry-effects are expected to become of less importance. It can be argued that the exclusion of occupation or task-level variables leads to an omitted variable problem in wage structure analyses. By omitting occupation or task-level variables, most studies are likely to overestimate the unexplained component of the wage gap. The aim of this research is to provide evidence for the hypothesis that occupations or task variables do significantly contribute to wage regressions and to test whether occupation wage effects have become of greater importance than industry wage effects.
1Assume we have a world with two occupations, doctors and receptionist. All doctors earn $2 an hour while
By using U.S. data, we take a developed country perspective. Developing countries are likely to be differently affected by for instance offshoring, as they are host country for low cost routine tasks offshoring (Navaretti and Vanables, 2004).
This work is related to a wide range of studies within trade and labor economics. The recent work of Autor et al (2003, 2006), Akcomak, Borghans, Weel (2010), Firpo, Fortin and Lemieux (2010) give an indication of professions that seem to be disappearing and suffer from a drop in wage over time. In addition, in several studies one has tried to explain changes in wage inequality
(Acemogly, 2002) (Firpo et al. 2010). In the recent study of Grossman and Rossi-Hansberg (2008), one tries to establish a linkage between international trade and labor economics, by focusing at the role of improved communication technologies, the break-up of the supply chains and the effect on wages. The theorized wage effects remain, unfortunately, ambiguous. In addition, studies have pointed at the nuanced view of skill-biased technological change, as a result of computerization and offshoring (Goos and Manning, 2007) (Autor and Dorrn, 2009). Finally, a recent body of work focuses on current influences on wage structures, as gender and racial wage differences (Blau and Kahn, 2006) (Blau and Devaro, 2007) (Chevalier, 2007) (Miller, 2009) and returns to education and experience (Mincer and Polachek, 1974) (Jensen, 2010). Trade and labor economists have extensively studied the effect of computerization and offshoring on wages. However, the wage structure stream of literature has entirely neglected the influence of computerization and offshoring on wage regressions.
by showing the highest significance levels and substantially contributing to the explanation of variance in wages. These findings are reinforced by a robustness test.
2. Literature and Theoretical Background
This study relates to three strands in the literature, namely the rising influence of technology on the labor market, international trade economics and wage structure research. All three will briefly be discussed.
2.1 Technological change
The rapid adoption of computer technology has changed the tasks performed by employees and ultimately the demand for human skills. Autor Levy and Murnane (2003) argue that advances in technology substitutes for workers in executing interactive and manual routine tasks that can be easily described in programmed rules, for instance, calculating, repetitive assembly and pricking and sorting tasks. Simultaneously, technological advancement complements workers in performing analytical and interactive non-routine tasks requiring creativity, flexibility and problem-solving capacities as, for example, managing a shop for which face-to-face contact is required. Computer capital only slightly affects non-routine manual tasks as, for instance, truck driving for the reason that driving a vehicle requires visual and motor processing capabilities that can, so far, not been described in a set of programmable rules, and can only be limited complemented by technology (table 1).
Table 1: Predictions of Task Model for the impact of Computerization on Four Categories of Workplace Tasks
Source: Autor, Levy, Murnane (2003) pp. 1286.
Routine tasks Non-routine tasks Analytic and interactive tasks
Examples Computer impact Record keeping Calculating
Repetitive Customer Service Substantial substitution
Managing others Forming, Testing Hypotheses Persuading/Selling Strong complementarities Manual tasks Examples Computer impact Picking or sorting Repetitive assembly Substantial substitution Truck driving Janitorial services
Due to the fact that computer capital and routine manual and routine interactive workers are (perfect) substitutes, a drop of the price of computer capital, decreases the wage paid to routine workers as well. Since own-factor demand curves are downward sloping, demand for routine tasks rises as the price drops. As a result, routine workers will self-select into non-routine manual jobs and the additional demand for routine tasks will be fulfilled by computer capital (Autor, Katz and Kearny, 2006). The final wage effect on non-routine manual workers depends on strength of the counterworking productivity and labor supply effect. As one assumes that greater intensity of routine inputs increases the marginal productivity of non-routine interactive inputs, wage of non-routine interactive workers unambiguously rises.
On of the major implications of the task-based organization of work is that it has made
one reconsider the general distribution of wages between high and low-skilled workers. Autor et al. (2003) and Goos and Manning (2007) developed a nuanced version of skill-biased technological change. The main idea is that ICT developments have not automatically reduced the demand for less-skilled labor as assumed by others (Bermand, Bound and Griliches, 1994). In contrast to the simple SBCT- hypothesis, the task-based approach allows different trends in wage development between low, middle and highly educated groups, explaining the polarization of the labor market. Since the 1980‟s, both wages and employment at the top (non-routine interactive tasks) and the bottom end (non-routine manual tasks) of the wage distribution have grown, at the expense of wage and employment in the middle part of the wage distribution (Autor and Dorn, 2009).
2.2 The Changing Nature of Trade
Philips (2009) and Baumgarten, Geishecker and Görg (2010), have extensively theorized and empirically tested the effect of offshoring on wages and employment. According to the academic literature, the wage effects of offshoring are highly ambiguous. The break-up of production processes may help certain workers in a given firm, while harming others as individual tasks can be replaced. Blinder (2006) argues that certain tasks that are interactive and require face-to-face are unlikely to be offshored (e.g. lawyers and hairdressers) while tasks may be more easily moved abroad (e.g. computer programmers). Due to the offshoring of particular tasks, task-level factor price differences within industries are likely to increase and inter-industry wage premiums are expected to vanish (Baldwin, 2006). Similarly to the developments in labor economics, the imperfect relationship between the offshorability of tasks and the skill-levels required is also acknowledged in trade economics (Grossman and Rossi-Hansberg, 2008). New theories go beyond the simple segregation of low-and high-skilled workers and are based on the idea that various bundles of task have different offshoring cost which suggests that the price and the likelihood of offshoring are only loosely related to skills. In fact, offshoring of routine tasks leads to ambiguous wage effects for routine and non-routine tasks values, depending on the strength of the productivity, labor supply and relative price effect.
2.3 Wage Structure Research
Wage structure research has been used to analyze numerous wage related phenomena as polarization of wage distribution (Goldin and Katz, 2007) (Krugman, 2008), gender and racial wage inequality (Blau and Kahn, 2006) (Blau and Devaro, 2007) (Chavalier, 2007) (van der Meer, 2008) (Miller, 2009) (Fransen, Plantinga and Vlasblom, 2009) and changing returns to education and experience (Mincer and Polachek 1974) (Jensen, 2010). This study is motivated by the unexplained wage components across these research fields.
late 1960s after controlling for individual-level predictors of wages such as education, experience, and demographic categories.
3. Data Selection
As we would like to address the effect of occupations and tasks on wage differences, the empirical strategy in this paper rest on combining individual-level data on wages, occupations and industries, with more aggregate data on the task content of specific jobs. This approach is taken to analyze the development of between- and within-occupation wage gaps and to include occupation and task variables in wage regressions. Here we present the details on the choices which are made for the database.
We use data from the merged outgoing rotation groups (ORG) of the Current Population Survey (CPS) from 1983 to 2002 which is a U.S. household survey including an extensive amount of individual-level data on wage, job, industry, age, education, and union coverage. Hence, the data cannot be used as time series data as respondents are not consistently followed over time. Outgoing rotation groups refers to respondents that are in the CPS panel for four months, rotate off for four months, and then are back on for four more months. The merged outgoing rotation group includes all observation in their fourth and eighth month. For the majority of data decision we have made, we follow the standard practice as shown by Lemieux (2006) and Firpo, Fortin, Lemieux (2010).
3.1 Occupational codes
In order to obtain a consistent series of occupations for the period 1980-20092, we employ the Meyer and Osborne (2005) classification scheme for making occupational groups across the 1980 and 2009 Census occupational coding (COC) schemes comparable. Unlike the IPUMS, Meyer and Osborne‟s system extensively uses the “not elsewhere classified” categories to address emergent or disappearing job titles. They are also investigating ways to split recorded occupations. Rather than allocating all cases into a single, largest category, they propose reallocating data from an earlier year (e.g., 1980) into multiple occupations categories based on trends found in later years (e.g., 1990). To ensure a sufficient number of observations in each occupation, we group occupation by (85) 2-digit occupational groups.
In January 2003, the CPS adopted the 2002 Census industry and occupational classification system derived from 2002 North American Industry Classification System. This
2 The reason that we focus on the entire period 1980-2009 becomes clear in the chapter 5 where it shows
new classification system creates breaks in time series for occupational data at all levels of aggregation. Fortunately, the Census Bureau provides a set of data from 2000 to 2002 that includes the new and old occupation codes. We can show the effect of switching to new occupation codes by estimating each model twice on the combined data from 2000 and 2002, once using the 1998 occupational codes and once using the 2000 occupation codes.
3.2 Industry codes
As we want to compare the influence of occupation-effect compared to the industry-effect on wages, a similar approach should be taken to obtain a consistent series of industries for the period 1980-2009. The fact that in 2000 the CPS has switched from Census Industrial Classification (CIC) to the North American Industry Classification System (NAICS) is the main obstacle. Crosswalks of the Census Industrial Classification (CIC) are provided by Autor et al.(2003) for CIC 80 and CIC 90. Consistent industry crosswalks from 1980 till 2009, however, are not available. Therefore, we limit ourselves to the timeframe 1983-2002 which incorporates both the rise of task-trade and computerization effects on the labor market, and for which a 52 2-digit SIC-based industry codes are available.
3.3 Defining tasks
Like many recent papers (Autor, Levy and Murnane, 2003) (Firpo, Fortin and Lemieux, 2010) (Goos and Manning, 2007), we have used O*NET and the Dictionary of Occupational Titles (DOT) to measure the task content of occupations. Our aim is to produce indexes for all 2-digit occupations as classified by Meyer and Osborne (2005). In the spirit of Autor, Levy and Murnane (2003), we have assigned a value for routine and non-routine manual tasks, non-routine interactive tasks, routine cognitive tasks, and non-routine analytical tasks to each occupation. Routine manual tasks can be defined as repetitive human labor tasks that can easily been described by a set of programmable rules. Tool makers, machinists and, typists hold a large fraction of routine manual tasks. Similarly, routine interactive tasks which can be found in professions as clerk, receptionist, and cashiers are vulnerable to computerization and- or offshoring as the task-content is repetitive information processing that can be taken over by computers or by labor overseas. Noticeably, the share of the labor force employed in occupations intensive in routine manual and routine interactive task has, since the 1980´s, declined significantly (Autor, Levy, Murnane, 2003).
non-routine interactivity are, for instance, work of managers, engineers, and teachers that can be supported by the use of ICT. Examples of non-routine analytical tasks are therapists, legislators, air traffic controllers. These non-routine tasks do not lend themselves to substitution by offshoring and computerization as non-routine tasks demand flexibility, creativity, general problem solving capacities, and complex communication. Contrary to routine tasks, the share of the labor force employed in non-routine interactive and analytical professions has experienced growth over the last four decades (Autor, Levy, Murnane, 2003).
Computer capital only limited affects non-routine manual tasks as, for instance, tasks of the fire police, ship crew and construction workers. The reason for this limited influence of computerization is that these tasks have, so far, not been described in a set of programmable rules, and one can only slightly gain from computer complementarities. Moreover, work of the fire police is bounded to a geographical location and can, therefore, not be offshored. Table 2 provides an overview of the tasks and subsequent occupations.
Table 2: overview of tasks and an illustration of typical occupations
The value of these tasks may be interpreted as levels or changes in task input relative to the 1960 task distribution. To create this match between occupations and the task content, we have used the crosswalks provided by Autor, Levy and Murnane. (2003). This file connects the Censes 1980/1990 codes with the DOT 1977, taking into account differences in tasks content for males and females.
3.4 Top coding
To protect privacy, the CPS decreases wages at the top of the distribution by topcoding. For 1980 till 1988 earnings per hour are topcoded at $99.99. This means that all earnings above that level appear in the CPS public use as $99.99, whatever their actual earnings are. This topcoding can
Task Task description Occupations Effect of computerization
and offshoring
Routine manual Repetitive tasks Tool makers, Machinists,
Typists,
Substitution Routine
interactive
Repetitive information processing tasks Clerks, Receptionists, Cashiers
Substitution Non-routine
interactive
Flexibility, Complex communication tasks Managers, Engineers, Teachers Complementarity Non-routine analytical
Creativity, General problem solving tasks
Therapists, Legislators, Air traffic controllers
Complementarity
Non-routine manual
Tasks that cannot be described in a set of programmable rules, and are
geographically bounded.
Fire police, Ship crew, Construction workers
lead to bias in the measurement of trends in earnings inequality if the proportion of observations affected, changes over time. In particular it will lower the mean and the variance of the wage data relative to the true mean and variance. Moreover, due to topcoding, one cannot observe changes that take place within the group that earns in the top category.
The value of the topcode has changed over time. Irregular and large adjustments to topcodes lead to sudden jumps in the means and variances of wages that are not due to actual changes in the true wage distribution. Moreover, while weekly earnings are topcoded at a threshold that is rarely crossed, monthly earnings are topcoded at much lower thresholds. For 1989 on, the topcode depends on hours worked and is selected so that earning per hour times usual hours is not more than $1.923 per week and in 1998 $2.884 per week. These values are not in constant price dollars and because these are not constant price dollars, one has changed the topcodes over time. Unfortunately, it does not appear that the changing top-codes are handled with sufficient care. Consequently, topcodes are inconsistent over time. Examining the data reveals that the topcode is not uniformly applied. While there is always a density peak at the topcode amount, some observations are generally present at higher wage rates. The number of dropped values fluctuates considerably as for instance between 1988 and 1989 values. Till 1988 earnings per hour are topcoded at $99.99. From 1989, the topcode has been altered and depends on hours worked and is selected so that earning per hour times usual hours is not more than $1.923 per week. It is shown that the number of dropped values decreases from 6748 in 1988 to 749 in 19893. In 1998 the topcoded value is changed once more to $2.884 per week. As a consequence, it is shown that the number of dropped values decreases from 1900 in 1997 to 895 in 1998. We follow the standard practice as shown by Lemieux (2006) and replace topcoded wages with 1.4 the topcoded value.
3.5 Earnings4
Wage data is of great importance in calculating wage inequality and running wage regressions. The Morg file contain information on respondents that are in the CPS panel for four months, rotate off for four months, and then are back on for four more months. The collection of wage information is only restricted to the fourth and the eight month. We convert all wage data to inflation adjusted 1979 dollars. The data shows a clear increase of the mean wage between 1983 and 2002 (appendix A).
3
See appendix A
4 We first change the topcoded values (*1.4) and afterwards, calculate the 1979 constant dollar values.
We adapt the established conventions (Lemieux, 2006) by restricting the sample to workers between the age of 18 and 65. To remove outliers, we also follow the existing literature by trimming very small and very large value of wages. Individuals who reported hourly wages below $1 and above $101. Wage information is collected for workers who are not self-employed. Another point to be made is that prior to 1994, all workers have reported an hourly wage rate. Starting from 1994, workers are first asked for the earnings periodicity. Earnings per hour can be consistently calculated by dividing earnings per week by usual hours.
In different time periods, hourly earnings including overtime tips and commissions have been differently determined. As changes overtime are of primary concern for the research, we could restrict ourselves to data excluding overtime, tips and commissions for hourly workers. The drawback of this approach is that the sample size reduces dramatically from 180310 to 95665 observations in 1987 which is highly undesirable. Therefore, we will use wage information including pay for overtime, tips and commissions
3.6 Missing wage data
For missing wage values we apply a no-imputation approach. The no-imputation method excludes the wages of missing cases but counts them when calculating occupational sizes. Mouw and Kallenberg (2010) show that this approach leads to similar results as a detailed hot-check and a multiple imputation approach.
3.7 Control variables
Table 3: Data overview
Subject Source Year Additional
Occupational codes Meyer and Osborne (2005)
standardized codes
1980-20095 85 2-digit occupational codes
Industry codes Census Industrial Classification 1983-2002 52 2-digit industry codes
Tasks O*NET and the Dictionary of
Occupational Titles (DOT)
- Similar to task definition of Autor, Levy and
Murnane (2003).
Earnings CPS-MORG Files 1983-2009 We focus on wage per hour, including overtime,
tips and commissions.
Top coding Lemieux (2006) - Solution: topcode *1.4
Missing values Mouw and Kallenberg (2010) - No imputation approach
Education CPS-MORG Files 1983-2002 We make four dummy categories: High school
dropout, Completion of the twelfth grade, Some college, Completion of college.
Union -covered CPS-MORG Files 1983-2002 A dummy variable being one for individuals
that are covered by a union.
Non-married CPS-MORG Files 1983-2002 A dummy variable being one non-married
respondents
Non-white CPS-MORG Files 1983-2002 A dummy variable being one for non-white
individuals
Age CPS-MORG Files 1983-2002 Age in years
Female CPS-MORG Files 1983-2002 A dummy variable being one for female
respondents
5 The reason that we focus on the entire period 1983-2009 becomes clear in the next chapter where it
4. Methodology
The aim of this research is to provide evidence for the hypothesis that occupations and task-variables do significantly contribute to wage regressions. The explanatory power of occupations and tasks on wages is expected to be related to the within- and between-occupation wage inequality. In fact, if wage inequality within an occupation is low but the between occupation wage gap is large, the occupation-effect in wage regressions is expected to be high. In this case, individuals with similar occupations have comparable wages, while wage differences between professions are high. Therefore, growth of between-occupation wage inequality is associated with a rise of the effect of occupation on wages (ceteris paribus). Contrary, if within-group wage variance is considerably high, occupations are less likely to have a high level of explanatory power in wage regressions. By using the Theil-index, we estimate how within- and between-occupation inequality has evolved over time.
4.1 Measuring wage inequality: Theil-index
To gain insides into the development of wage inequality we consider the Theil entropy measure and its decomposition as discussed by Steckel and Moehling (2000). The Theil-Index is given by the following equation (1):
T =
n 1 n i i w 1ln
w
i(1)
Where n denotes the number of observations, wi represents the wage of individual i, and μ
represents the sample mean wage. The mean wage is this is the sum of all wi terms, divided by n.
In the case of perfect equality, the Theil measure equals zero. In the case of “perfect inequality” when one individual owns all of the society‟s wealth, the Theil measure equals [ln (n)]. This means that the maximum value of the Theil-index depends on the n the number of observations
.
Ng denotes the number of observations in sub-group g, μ represents the sample mean wage μg
represents the mean wealth of sub-group g, and Tg represents the measure in equation 1 calculated
for sub-group g. The first term on the right-hand side of equation 2 represents the weighted sum of the Theil entropy measures for the subgroup wage distribution where the weights are the sub-group share of total wage. This is equal to wage inequality within sub-groups. The second term of the right- hand side of equation 2 corresponds to the between sub-group inequality. The second term of the right-hand side of equation 2 is equal to equation 1 but here the inequality is measured for a distribution in which each individual is assigned the mean value of the subgroup wage.
As it is mentioned before, the maximum value of the Theil-index depends on the number of observations n. Therefore, we would like to analyze the trend in wage inequality arising from within-group inequality (Tg), between-group inequality (μg/μ) and inequality due to changes in
the population shares of sub-groups (ng/n) (Cowell and Kuga, 1981). The term t corresponds to
time and s represents time t+1.
The contribution of these components can be calculated as follows:
ΔTts within , = G g s s s g s g
n
n
1 (T
gt-T
gs) (3.1) ΔTts between , = G g s s g t t g 1 G g s s g s s g t t g t t g t g t t gT
n
n
1ln
ln
s s gn
n
(3.2) ΔTts n , = G g s s g t t gn
n
n
n
1 G g s s g t t g t g t t g n n n n T 1 t t g t t gln
(3.3)This decomposition provides relevant insights into the sources of wage inequality. The reason to use the Theil-index is a member of the generalized entropy class measures which is desirable for this decomposition (Steckel and Moehling, 2000).
Some occupations have undergone small negative changes in within-group wage inequality as managers, technicians and therapists while others have experienced slight increases of within-group wage inequality, for example, teachers, clerks and receptionist. It must be mentioned, however, that changes of the Theil-index are only very small as, for instance, the highest positive value is 0.00329 (teachers). It is interesting to see that the occupations with the highest wage gains between 1983 and 2002, show mostly positive changes of the Theil-indices while occupations that have experienced negative wage growth present more negative Theil values (table 4).
Table 4: Theil-Indices of occupations with the largest wage changes
Similarly, the changes of industry Theil-indices are only small. The largest positive rise of within- industry wage inequality we find in the aviation sector (0.013) and largest drop of within-industry wage differences have occurred in the communication sector (-0.016).
4.2 Wage regression
It has been mentioned that offshoring and computerization have made it likely that occupations play a major role in explaining wage variation. Thus, as a second set in the analysis, we will estimate the effect of occupations and tasks in a wage regression. The regressions will be done for different years separately and we will compare outcomes for diverse time periods. In fact, we are comparing two cross-sections that differ in time. First we estimate the baseline model:
Log(W)i = β1 + β2union-coveredi + β3educationi+ β4 non-marriedi+ β5agei +
β6non-whitei+ β7femalei+ εi (4)
We have adopted a log-linear model6 where (W) represents the hourly wage in U.S. Dollars. As in Firpo, Fortain and Lemieux (2009), the baseline model includes control variables as union membership, educational level, marital status, age and race. In addition, female is included as a control variable. The subscript i refers to the respondent. β1 is a constant and εi refers to the
error term. In fact, εi captures all influences on hourly wage left unexplained by the explanatory variables. Appendix D provides a table with descriptive statistics. Here we can see that between 1983 and 2002 the average wage, age, and the educational level of the sample has grown. A coefficient of 0.08 estimated for example for β5 indicates that, (ceteris paribus), the hourly wage rate increases with 8% with one additional year of age.
Second, we would like to analyze to what extent occupations are valuable for explaining wage variations. Therefore, we include an occupation fixed-effect7 to the baseline model:
Log(W)i = β1 + β2Occi + δHi+ εi (5)
Occ denotes the 85 2-digit occupation dummy variables and i refers to the respondent. Hiindicates a vector of parameters of union membership, educational level, marital status, age, gender, and race. Next, we add an industry-fixed effect to the baseline model which allows comparing the difference between occupation-fixed effects and industry fixed-effects:
Log(W)i = β1 + β2Indi+ δHi+ εi (6)
Ind denotes the 47 2-digit industry dummy variables and i refers to the respondent. Here, we
expect the explanatory power of the occupational fixed-effect to be larger than the influence of the industry-fixed effect as the occupation controls are expected to account for a larger part of the
6 Here we have taken the natural logarithm instead of the base-10 logarithm.
7 Fixed-effects will be used as the outcomes of the Durbin-Wu-Hausman (DWH) test suggest that the
wage variation. Equation 7 includes and interaction effect between occupation and industry dummy variables.
Log(W) i = β1 + β2 Occi*Indi+ δHi+ εi (7)
This enables us to analyze whether occupational wages have differently evolved in industries or whether hourly wages of professions in various industries developed similarly.
As we have explained in chapter 1 and 2, production has become organized at a finer level of disaggregation; the task level. It is expected that the value of routine manual and routine interactive tasks are negatively influenced by offshoring and computerization while the value non-routine interactive and non-routine analytical tasks are complemented by overseas labor and ICT-technology. Non-routine manual tasks are not expected to be affected by computerization and offshoring. Instead of including occupation dummy variables in a regression, for instance equation 5, we include a value of each task that indicates the importance of that task in an occupation. In fact, for each occupation we assign a value for routine manual, routine interactive, non-routine interactive, non-routine analytical and non-routine manual tasks.
From the descriptive statistics it stems that the average task-content has only slightly changed as the average level of non-routine tasks slightly increased and the task average of routine tasks somewhat decreased (appendix D). Hence, equation 8 is similar to equation 5 simply the occupation dummies are replaced by task measures. The “task content” of an occupation is given by Autor, Levy and Murnane (2003) and is described in a value for non-routine manual tasks, routine manual tasks, non-routine interactive tasks, routine cognitive tasks, and non-routine analytical tasks. The value of these tasks may be interpreted as levels or changes in task input relative to the 1960 task distribution.
We will run a regression including tasks-variables (equation 8).
Log(W) i = β1 +β2Nonroutmani +β3Routmani+ β4Nonroutineraci+β5Routcogi+
β6Nonroutanlyti+ δHi+ εi (8)
Where β2to β6denote non-routine and routine manual tasks, non-routine interactive tasks,
routine cognitive tasks, and non-routine analytical tasks. Finally, White standard errors will be used as error variances are expected to be non-constant8
8 For instance, individuals with an higher educational level have a wider variety of job opportunities than
.2 0 .1 6 .1 8 .1 4 .2 2 T h e il's T -St a ti st ic 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 year Code Change in 2003 Total Occupational Wage Inequality
5. Results
5.1 Occupational wage inequality
Figure I shows the evolution of the trends in wage inequality indicated by the Theil-index for years 1983 to 20099. In the case of perfect equality, the Theil measure equals zero. As the maximum value of the Theil-index varies with sample size, one should focus on the trend instead of absolute values of the Theil entropy measure.
Figure I shows a remarkable sharp and sudden change of wage inequality in 1992 and 2003, potentially due to the implementation of a new occupational classification system. Fortunately, the Census Bureau provides a set of data from 2000 to 2002 that includes the new and old occupational codes. This way, one could show the effect of switching to new occupation codes by estimating each model twice on the combined data from 2000 and 2002, once using the 1998 occupational codes and once using the 2000 occupation codes. Figure II and figure III provide evidence that the Theil-index is highly sensitive to changing occupational classification methods.
Figure I10
Figure II
9
The reason that we focus on the entire period 1983-2009 is that this timeframe is needed to present the effect of code changes on the wage inequality index.
10 A similar wage inequality analysis has been applied by using 52 two-digit industries. However, no clear
trend is found in the evaluation of overall industrial wage inequality (appendix B).
.1 4 .1 6 .1 8 .2 2 .2 0 T h e il's T -St a ti st ic 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 2003 2005 2007 2009 year Theil03 Theil00
Comparison of Code Change Total Occupational Wage Inequality Figure III
The line Theil03, which is similar to figure I, is constructed by using a classification method that alters in 2003. Considerable divergence between the lines occurs, which is a consequence of the occupational classification change in 2000. It can be argued that the Theil-index is highly affected by the inconsistency in occupational codes. In fact, the divergence shown between the both lines is exclusively due to the occupational classification change. Prior to 2000 and after 2002, there is just a single classification system available and for this reason, both lines converge. For 1992 this additional graph can, unfortunately, not be constructed as in the 1992 file, occupations are not coded twice.
To conclude, it is assumed that the sudden shifts of the Theil curve in 1992 and 2003 are caused by the change of the occupational classification system. In the period 1983-2003, the overall occupational wage inequality has experienced slight growth11. This is consistent with the existing literature, for instance Acemoglu (2002), which has shown that overall wage inequality has grown since the 1980‟s.
5.2 Inequality Decomposition
The decomposition of aggregate measured inequality can provide important clues as to the sources of wage inequality. In this section, we extend our analysis to the decomposition of the Theil-index into components representing wage inequality arising from a) differences within a subgroup, b) inequality caused by wage variation between several subgroups and c) a change in the Theil-index due to variation of the number of respondents (n) within subgroups.
Figure IV presents the decompositions of the changes in the Theil entropy measure for occupations from 1983-2002. TB denotes wage inequality between diverse occupational groups,
.1 4 .1 6 .1 8 .2 0 T h e il's T -St a ti st ic 1983 1985 1987 1989 1991 1993 1995 1997 1999 2001 year TB TW TN
Decomposed Occupational Wage Inequality Theil-Index
TW indicates the wage differences of individuals practicing the same profession and TN stands for
the wage variation that is caused by changing population shares within an occupation. It is shown in Figure IV that the increasing inequality was due to shifts in population shares, and increases in between-and within-group inequality as all three lines in figure XIII slightly rise12. The strength of TB, TW and TN has overall remained equal during this period. Population share is the largest and between-occupational wage inequality is the smallest component of total wage inequality. It becomes clear that between-and within-occupation wage inequality are considerably affected by the classification change and further experience slight growth13. The findings are partially supported by Mouw and Kallenberg (2010), which provides evidence that between-occupation inequality has increased over the last two decades.
Figure IV14
5.3 Wage regression
In the previous paragraph, both between-and within-occupational wage inequality have almost evolved similarly. If the within-group inequality of an explanatory variable rises, the explanatory power of this variable drops and, ceteris paribus, the residual increases. As the between-occupational wage variation is almost equal to the development of within-occupation wage inequality, one could argue that the net effect on wages is expected to be zero. However,
11 We have also split the sample in a manufacturing and service sectors and perform separate analyses,
which didn‟t alter the outcomes.
12 We see a sudden and large drop of TB, TW and TN around 1993. We assume here that this is due to a
codification change, similarly as has been shown in figure III.
13 As a robustness test, we have weighted the observations by weekly hours of work as a reasonable
compromise between accepting all workers as equal observations and looking at full-time workers only, irrespective of the number of hours worked. The wage inequality trends do not significantly change.
14 A similar inequality decomposition has been constructed for industries. However, no clear patterns has
other factors are likely to influence within-and between-occupation wage inequality. The literature provides evidence for the fact that within-group wage differences are substantially larger for older and more educated workers than for young and less educated individuals (Lemieux, 2006). Appendix C provides an overview of the rise of average age and education within our sample. Therefore, it can be argued that fraction of the increase of the wage inequality within a professional group, is a consequence of the fact that the work force has grown older and more educated since the early 1980‟s. The effect of age, education and other factors on wage inequality has, so far, not been taken into account in the inequality analyses. Appendix C provides evidence for a larger within-group wage variance for older workers. The Theil-indexes in the previous paragraph have not provided convincing evidence that within-occupation wage inequality has declined. However, due to this composition effect, it has not become evident whether tasks and occupations do deserve a “base position” in the wage regression15. Moreover, we compare the influence of occupation control variables with industry and task-level variables. We now turn to the actual wage regression16. Table 1 and 2 show various regression outcomes for 1983 and 200217. As described earlier, we first estimate the baseline model, and then include an occupation fixed-effect and an industry fixed-effect. Hence, we are comparing cross-sections that differ in time.
Between 1983 and 2002, occupation fixed-effects are expected to grow and industry fixed-effects are predicted to become of less importance. Finally we add task-level variables to the basic model. Non-routine analytical and interactive tasks are likely to have a growing positive influence on wages, while non-routine manual tasks are not expected to have a significant effect on hourly wages. Contrary, routine manual tasks and routine interactive tasks are predicted to provide a growing negative influence on wages.
Table 1 and 2 contain the results of the various wage regressions. The baseline model shows comparable results for both years as union coverage, race, marital status, age and gender are major determinants for hourly wages. The fact that the merits of higher education have grown is in line with the existing literature (Mincer, 1997) (Deschênes, 2001)18. Similarly, the decline of the marriage wage premium and racial discrimination is supported by previous research
15 Mincer (1974) provides evidence that the composition effect is expected to be empirically important. 16 Appendix D provides descriptive statistics
17
The regression results for the years 1984-2001 are in line with the results in table 1 and 2.
18 Mincer (1997) and Deschênes (2001) have shown that log wages have developed as a convex function of
0 2 4 6 8 10 T a s k -C o n te n t C h a n g e Si n c e 1 9 6 0 i n %
Clerks Agricult. Workers Machine Operators Assemblers Significant Occupation Fixed-Effect in 1983
non-routine manual non-routine interactive routine manual routine interact. non-routine analytical
(Korenman, and Neumark, 1991) (Maloney, 1994). Also the drop of the gender wage gap indicated by female is in line with the existing literature (Blau and Kahn, 1994).
Turning to the fixed-effect models, table 1 and 2 present the baseline model including an occupation-fixed effect19. Although, the significance on a 5 percent level is limited, occupation fixed-effects change the explanatory power of the control variables considerably. In particular, the educational variables are highly affected. From 1983 to 2003, the percentage of occupation dummies that is significant, drops from 44 percent to 32 percent. Figure V presents a selection of the occupations for which the fixed-effect is significant in 1983 but turns insignificant in 2002 as clerks, agricultural workers, machine operators and assemblers. These occupations are characterized by a noteworthy negative coefficient with an average of -.22.
For 1983, figure V shows that occupations with a significant and negative fixed-effect are mainly characterized by routine interactive tasks and routine manual tasks, both task that are expected to be taken over by technology or to be offshored in the future. The contribution of technology complementary tasks, as non-routine interactive and non-routine analytical is rather low.
Figure V
Figure VI presents occupations showing a significant fixed-effect in 2002. By contrast, the coefficients of the 2002 occupational coefficients are small (on average 0.06) but positive. These jobs include financial specialists, scientists, therapists and technicians, for which non-routine analytical and non-routine interactive tasks play a major role. Although they do
19 The Hausman test confirms that the coefficients from the fixed-effect models are significantly different
0 2 4 6 8 10 T a s k -C o n te n t, C h a n g e Si n c e 1 9 6 0 i n %
Financial Specialists Scientists Therapists Technicians
Significant Occupation Fixed-Effect in 2002
non-routine manual non-routine interactive routine manual routine interactive non-routine analytical
contribute to the task-content of the job, routine interactive and routine manual tasks are not dominant20.
Figure VI
Appendix E gives an overview of the top five strongest occupation fixed-effects in 1983 and 2002. It shows that the penalty for low-wage service jobs, as parking lot attendants and gardeners has decreased. This is consistent with the existing literature, theorizing that low wage service jobs will only be limited affected by computerization and offshoring (Autor, Katz and Kearny, 2006) (Foritin, Firpo, Lemieux, 2010)
Column 3 in table 1 and 2 shows the results of an industry fixed-effect added to the baseline model. Although the industry fixed-effect alters the control variables coefficient to a small extent, a considerable set of industry dummies provides insignificant results. Contrary to our expectations, a small number of industry dummies, as the apparel and retail trade industry, turns significant in 2002 and shows exclusively negative coefficients with an average of -.20. The wage variation explained by occupations and industries, indicated by the r-squared, is similar.
20 For those occupation-fixed effects that show significant results in 1983 and 2002, the task-content is
Table 1: Regression results for log (wage) 1983
Baseline Occupation Industry Task Model
Model Fixed-effect Fixed-effect (Autor) VARIABLES Union covered 0.21*** 0.23*** 0.17*** 0.23*** (0.00) (0.00) (0.02) 0.00 Non-white -0.07*** -0.03*** -0.06*** -0.04*** (0.00) (0.00) (0.01) (0.00) Non-Married -0.15*** -0.010*** -0.12*** -0.11** (0.00) (0.00) (0.02) (0.00) Education21
(Some college omitted)
HS dropout -0.35*** -0.19*** -0.29*** -0.25*** (0.00) (0.00) (0.02) (0.00) Twelfth grade -0.13*** -0.07*** -0.11*** -0.08*** (0.00) (0.00) (0.02) (0.00) College graduate 0.25*** 0.16*** 0.28*** 0.15*** (0.00) (0.00) (0.00) (0.00) Age 0.01*** 0.01*** 0.01*** 0.01*** (0.00) (0.00) (0.00) (0.00) Female -0.30*** -0.24*** -0.23*** -0.27*** (0.00) (0.00) (0.01) (0.00) Tasks (Non-routineman. omitted) Non-rout interact. 0.02*** (0.00) Routine man. -0.04*** (0.00) Routine interact. 0.03*** (0.00) Non-routine analyt. 0.06*** (0.00) Constant 1.44*** 1.44*** 1.44*** 1.20*** (0.00) (0.02) (0.02) (0.01) Observations 168524 168524 168524 168524 Adj. R-squared 0.36 0.34 0.35 0.43
Notes: We have used OLS and heteroskedasticity consistent standard errors are in parentheses; * Significant at 10%, **significant at 5%; *** significant at 1%
In addition, equation 1 contains an interaction effect of occupation and industries. This interaction effect enables ones to analyze the wage effects of an occupation in different industries and differences in wage effects between several occupations in a single industry. We have 85 2-digit occupations and 47 2-digit industries. Table 2 and 3 do not fit 3995 dummy variables. Consequently, the general findings concerning interaction effects are not included in the table, but will be discussed here22. First, interaction effects between occupations and industry have provided several significant results. For instance typists and safety guards show different occupation effects in different industries. In this, there is no distinction between routine and non-routine analytical and interactive tasks23. In addition, we do not see clear changes of these interaction
21
A well-known problem with including education in a wage regression is that education is not measured in a consistent manner over time. Prior to 1992, individuals reported their highest grade attended while after 1992, the CPS switches to the highest degree completed. Here we have adopted the method of Jaeger (1997) to construct a relatively consistent variable.
22
See appendix F
23 Some interaction effects show significant results, however, if the occupation fixed-effect or industry
terms between 1983 and 2002 and no clear distinction can be made between the manufacturing and the service sector.
Finally, we turn to the baseline model including task-level variables. Results in column 4 indicate that the economic significance of the task-level variables is considerable. Accept from routine interactive tasks, all signs have the expected positive or negative sign and are highly significant24. The values of these routine and non-routine variables are defined as percentage change in task input relative to the 1960 task distribution. This means that, for instance in 2002, a 1 percent increase in the routine analytical tasks in the total task distribution since 1960‟s, increases the wage with 8 percent point, while 1 additional year of age is expected to raise the wage with only 1 percent point (Table 2).
Table 2: Regression results for log (wage) 2002
Baseline Occupation Industry Task Model
Model Fixed-effect Fixed-effect (Autor) VARIABLES Union covered 0.14*** 0.16*** 0.13*** 0.16*** (0.00) (0.00) (0.00) (0.00) Non-white -0.06*** -0.04*** -0.06*** -0.04*** (0.00) (0.00) (0.00) (0.00) Non-Married -0.11*** -0.07*** -0.09*** -0.08*** (0.00) (0.00) (0.00) (0.00) Education
(Some college omitted)
HS dropout -0.38*** -0.21*** -0.31*** -0.27*** (0.00) (0.02) (0.03) (0.00) Twelfth grade -0.13*** -0.05*** -0.11*** -0.07*** (0.00) (0.00) (0.02) (0.00) College graduate 0.40*** 0.27*** 0.40*** 0.28*** (0.02) (0.02) (0.00) Age 0.01*** 0.01*** 0.01*** 0.01*** (0.00) (0.00) (0.00) (0.00) Female -0.20*** -0.17*** -0.17*** -0.20*** (0.00) (0.00) (0.00) (0.00) Tasks
(Non-routine man. omitted)
Non-rout interact. 0.01*** (0.00) Routine man. -0.02*** (0.00) Routine interact. 0.02*** (0.00) Non-routine analyt. 0.08*** (0.00) Constant 1.43*** 1.44*** 1.45*** 1.14*** (0.01) (0.01) (0.00) (0.01) Observations 161349 161349 161349 161349 Adj. R-squared 0.35 0.34 0.35 0.42
Notes: We have used OLS and heteroskedasticity consistent standard errors are in parentheses; * significant at 10%, **significant at 5%; *** significant at 1%
For example, hourly wages of medical scientists, who experienced on average 10 percent growth of the non-routine analytical task in their task distribution will, ceteris paribus, gain 80 percent of wage above apparel pressing operators, for which the share of non-routine analytical task remained constant since the 1960‟s. Here it is shown that, although age is often included in wage regressions, the routine analytical task variable could have a considerable larger effect on wages. Furthermore, while the coefficients of the routine manual task dropped slightly, the economic significance of the non-routine interactive task has grown. These developments are in line with the outcomes of the occupation fixed-effect, which indicated that in 1983 fixed-effect coefficients were negative while in 2002 these became positive. In addition, the adjusted r-squared rises to 0.42 in 2002, which is substantially larger than the wage variance explained by the fixed-effect and baseline model (0.34). This suggests that the routine and non-routine task-level variables make a positive contribution the wage regression.
6. Robustness
The aim of this paper is to provide evidence for the need to include occupation or task variables in wage regressions, caused by computerization and offshoring. In the previous chapter, it is shown that especially task variables do considerably contribute in explaining wages. As predicted, non-routine interactive and analytical tasks positively affect wages while routine manual tasks have the opposite effect. One of the main drawbacks of the task data we have used, is that it is only available for one year. In fact, we know the task content of occupations in 1977. It is highly likely that over time, the task content of a job might have changed to some extent. Here we assume, however, that if for a particular job the most important task in 1977 is routine cognitive, the routine cognitive tasks will remain the most important tasks also in 1983 and 2002. Imagine a sales agent in 1977 carries out a large level of non-routine interactive tasks, as the major role of this sales agent is to assist clients. In 2002 the same sales agent could also carry out another task as for instance, keeping the website updated. We argue here that although the sales agent carries out other tasks, the non-routine interactive tasks remain most important. As a robustness check, we conduct our analysis with task-level dummy variables being one for the highest task value within an occupation.
Table 3: Robustness check
1983 2002 VARIABLES Union covered 0.25*** 0.17*** (0.00) (0.00) Non-white -0.05*** -0.04*** (0.00) (0.00) Non-Married -0.14*** -0.09*** (0.00) (0.00) Education
(Some college omitted)
HS dropout -0.32*** -0.33*** (0.00) (0.00) Twelfth grade -012*** -0.11*** (0.00) (0.00) College graduate 0.20*** 0..31*** (0.00) (0.00) Age 0.01*** 0.01*** (0.00) (0.00) Female -0.31*** -0.34*** (0.00) (0.00) Tasks
(Non-routine man. omitted)
Non-rout interact. 0.25*** 0.23*** (0.01) (0.00) Routine man. -0.01 -0.02 (0.00) (0.00) Routine interact. 0.15*** 0.13*** (0.01) (0.01) Non-routine analyt. 0.30*** 0.35*** (0.01) (0.01) Constant 1..32*** 1.29*** (0.01) (0.01) Observations 168524 161349 Adj. R-squared 0.39 0.38
7. Conclusion, Implications and Limitations
Since the mid 1980‟s, production has become organized at a finer level of disaggregation; the task level. Consequently, labor market competition has been intensified by both individual workers performing similar tasks in different nations, and from computer capital replacing several tasks previously performed by human labor. Within industries, some occupations are expected to be harmed by computerization and offshoring developments while others will gain from it. In this paper we have looked at the contribution of occupations and task-variables in explaining wages by the use of the U.S. CPS. We have found interesting results showing that with occupation one is able to explain a considerable share of the wage variation. In 1983, occupations with a substantial intensity of routine tasks, as assembly workers, are panelized while in 2002, non-routine interactive and analytical professions, for instance therapist, are financially rewarded. In addition, the penalty for low wage service jobs has also decreased which is consistent with the existing literature. Furthermore, we find both statistical and economic significant evidence of task-level variables explaining wages. Especially, non-routine interactive task positively alter wages and this effect has slightly increased between 1983 and 2002. Contrary to the expectations, wage inequality analysis does, however, not provide a clear indication for the need to include occupations in wage regressions as it shows a slight growth of the within-and between-occupation wage gap. In addition, the persistent effect of industries on wage variations was not predicted. The influence of offshoring and computerization on within- and-between occupation wage inequality and wage developments has shown to be less straightforward than hypothesized.
These findings imply that occupations, and especially task variables, explain a significant share of the wage variation. Consequently we argue that occupations or task-variables should be adapted as explanatory variables in wage regressions which are used to analyze a wide range of wage related subjects, as gender wage inequality and racial wage discrimination. By omitting occupations or task-variables, most studies are likely to overestimate the unexplained component of the wage gap. In addition, the occupation industry interaction effect has shown that occupation wage differentials between industries do exist, contrary to the assumption made by, for instance, Akcomak, Borghans, Weel, (2010). A direction for further research is to further examine to what extent these findings are caused by offshoring and computerization as this is not answered satisfactory here.
fixed-effects have shown to be of great influence on wages. However, one should take into account that occupational controls are moderately or highly correlated with many of the other ride-hand-side variables. On the one hand, we argue that excluding occupations in standard wage regression leads to omitted variable bias if the proportion of wage not explained by human capital variables, varies systematically by occupation. On the other hand, one could argue that if occupations are likely to be highly correlated with control variables, occupation fixed-effects are expected to account by far more than can be explained by occupations alone (Levenson and Zoghi, 2010). For instance, in the case of education, occupational sorting will occur as one needs a higher education to become, for instance, a doctor. Therefore it is hard to exactly determine which share of the wage should be attributed to education and which to the occupation fixed-effect. Although the task-variable model suffers from a comparable specification problem as the occupation fixed-effect model, the explanatory power of the control variables remain relatively constant and all task-level variables are significant on a 1 percent level.
References
Acemoglu, D. (2002), “Technological Change, Inequality and The Labor Market” Journal of Economic Literature, 40, pp. 7-72.
Akçomak, I.S., L. Borghans and B. ter Weel (2010), “Measuring and interpreting trends in the division of labour in the Netherlands,” CPB Discussion Paper No. 16.1
Autor, D.H. and Dorn, D., (2009) “Inequality and Specialization: The Growth of Low-Skilled Service Jobs in the United States”, Institute for the Study of Labor, Discussion Paper 4290.
Autor, D.H., Katz, L.F. and Kearny, M.S., (2006) “Trends in The U.S. Wage Inequality: Revising The Revisionists”, Review of Economics and Statistics, 90 pp. 300-323.
Autor, D. H., Levy, F.,and Murnane, R. J., (2003) "The Skill Content of Recent Technological Change: An Empirical Explo- ration." Quarterly Journal of Economics, 118(4), pp. 1279-333. Baldwin, R., (2006) "Globalization: the great unbundling(s)" Economic Council of Finland. Baumgarten, D., Geishecker I., and Görg H., (2010) "Offshoring, tasks, and the skill-wage pattern," Kiel Working Papers 1603, Kiel Institute for the World Economy.
Berman E., Bound, J., Griliches, Z., (1994) "Changes in the Demand for Skilled Labor within U.S. Manufacturing Industries: Evidence from the Annual Survey of Manufacturing," NBER Working Papers 4255, National Bureau of Economic Research, Inc.
Blau, F.D., and Devaro, J.,(2007) “New Evidence on Gender Differences in Promotion Rates: An Empirical Analysis of a Sample of New Hires”, Industrial Relations, 46(3) pp. 511-550.
Blau, F.D., and Kahn, L.M., (1994) “Rising wage Inequality and the US Gender Wage Gap”, American Economic Review, 84 (2).
Blau, F.D., and Kahn, L.M., (2006) “The U.S. Gender Pay gap in the 1990s: Slowing Convergence” Industrial and Labor Relations Review, 60(1) pp. 45-66.
Blinder, A.S., (2006) “Offshoring: The Next Industrial Revolution?” Foreign Affairs, 85(2), pp. 113-128.
Brambor, T., Clark, W.R., Golder, M., (2006) “Understanding Interaction Models: Improving Empirical Analyses”, Political Analysis 14(1), pp. 63-82
Deschênes, O., (2001), “Unobserved Ability, Comparative Advantage and the Rising Return to Education in the United States: A Cohort-Based Approach,” Princeton University Industrial Relations Section Working Paper No. 465.
Dinardo, J., Fortin, N.M., and Lemieux, T., (1996) "Labor Market Institutions and The Distribution of Wages, 1973-1992: A Semiparametric Approach", Econometrica, 64(5), pp. 1001-1044.
Ebenstein, A., Harrison, A., McMillan and M., and Phillips, S., (2009) "Estimating the Impact of Trade and Offshoring on American Workers Using the Current Population Surveys," NBER Working Papers 15107.
Feenstra, R.C., and Hanson G.H., (1999) “The Impact of Outsourcing and High-Technology Capital on Wages: Estimates for the United States, 1979-1990”, Quarterly Journal of Economics, 114(3) pp. 907-940
Firpo, S., Fortin, N., and Lemieux, T., (2010) “Occupational Tasks and Changes in the Wage Structure”, University of British Colombia, Working Paper.
Fransen, E., Plantinga, J., and Vlasblom, J.D., (2009) “Decomposing the Dutch Gender Pay Gap, 1996-2006”, Tjalling C. Koopmans Research Institute.
Goldwin, C., and Katz, L.F., (2007) “Long-Run Changes in the U.S. Wage Structure: Narrowing, Widening, Polarizing”, Harvard University and NBER working paper 13568.
Goos, M., Manning A., (2007) "Louse and Lovely Jobs: The Rising Polarization of The Work in Britain", Review of Economics and Statistics, 89, pp. 118-133.
Grossman, G.M., and Rossi-Hansberg, E., (2006) "The rise of offshoring: it's not wine for cloth anymore," Federal Reserve Bank of Kansas City, pp 59-102.
Grossman, G.M. and Rossi-Hansberg, E., (2008) “Trading tasks: A Simple Theory of Offshoring”, American Economic Review, 98(5).
Helpman, E., (1984) “A simple theory of international trade with multinational corporations.” Journal of Political Economy, 92, pp. 451-71.
Jaeger, D. (1997) „„Reconciling Educational Attainment Questions in the CPS and Census.‟‟ Monthly Labor Review, 121(8), pp. 36–40.
Jensen, D., (2010) “The (Perceived) Returns to Education and the Demand for Schooling” Quarterly Journal of Economics, 125(2), pp. 515-548.
Juhn, C., Murphy, K.M., and Pierce, B., (1991) “Accounting For Slowdown in Black-White Wage Convergence”, in Marvin H. Koster, Workers and Their Wages: Changing Patterns in The United States, AIE Press, Washington.
Kaplan, S. N., and Raugh. J., (2009) “Wall Street and Main Street: What Contributes to the Rise in High Incomes?”, Review of Financial Studies, 23, pp. 1004–1050.
